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 *save_copy_start; /* Provides one level of saving
135 and restoring 'copy_start' */
136 char *copy_start_in_input; /* Position in input string
137 corresponding to copy_start */
138 SSize_t whilem_seen; /* number of WHILEM in this expr */
139 regnode *emit_start; /* Start of emitted-code area */
140 regnode_offset emit; /* Code-emit pointer */
141 I32 naughty; /* How bad is this pattern? */
142 I32 sawback; /* Did we see \1, ...? */
144 SSize_t size; /* Number of regnode equivalents in
147 /* position beyond 'precomp' of the warning message furthest away from
148 * 'precomp'. During the parse, no warnings are raised for any problems
149 * earlier in the parse than this position. This works if warnings are
150 * raised the first time a given spot is parsed, and if only one
151 * independent warning is raised for any given spot */
152 Size_t latest_warn_offset;
154 I32 npar; /* Capture buffer count so far in the
155 parse, (OPEN) plus one. ("par" 0 is
157 I32 total_par; /* During initial parse, is either 0,
158 or -1; the latter indicating a
159 reparse is needed. After that pass,
160 it is what 'npar' became after the
161 pass. Hence, it being > 0 indicates
162 we are in a reparse situation */
163 I32 nestroot; /* root parens we are in - used by
166 regnode_offset *open_parens; /* offsets to open parens */
167 regnode_offset *close_parens; /* offsets to close parens */
168 I32 parens_buf_size; /* #slots malloced open/close_parens */
169 regnode *end_op; /* END node in program */
170 I32 utf8; /* whether the pattern is utf8 or not */
171 I32 orig_utf8; /* whether the pattern was originally in utf8 */
172 /* XXX use this for future optimisation of case
173 * where pattern must be upgraded to utf8. */
174 I32 uni_semantics; /* If a d charset modifier should use unicode
175 rules, even if the pattern is not in
177 HV *paren_names; /* Paren names */
179 regnode **recurse; /* Recurse regops */
180 I32 recurse_count; /* Number of recurse regops we have generated */
181 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
183 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
187 I32 override_recoding;
188 I32 recode_x_to_native;
189 I32 in_multi_char_class;
190 struct reg_code_blocks *code_blocks;/* positions of literal (?{})
192 int code_index; /* next code_blocks[] slot */
193 SSize_t maxlen; /* mininum possible number of chars in string to match */
194 scan_frame *frame_head;
195 scan_frame *frame_last;
199 #ifdef ADD_TO_REGEXEC
200 char *starttry; /* -Dr: where regtry was called. */
201 #define RExC_starttry (pRExC_state->starttry)
203 SV *runtime_code_qr; /* qr with the runtime code blocks */
205 const char *lastparse;
207 AV *paren_name_list; /* idx -> name */
208 U32 study_chunk_recursed_count;
212 #define RExC_lastparse (pRExC_state->lastparse)
213 #define RExC_lastnum (pRExC_state->lastnum)
214 #define RExC_paren_name_list (pRExC_state->paren_name_list)
215 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
216 #define RExC_mysv (pRExC_state->mysv1)
217 #define RExC_mysv1 (pRExC_state->mysv1)
218 #define RExC_mysv2 (pRExC_state->mysv2)
228 #define RExC_flags (pRExC_state->flags)
229 #define RExC_pm_flags (pRExC_state->pm_flags)
230 #define RExC_precomp (pRExC_state->precomp)
231 #define RExC_copy_start_in_input (pRExC_state->copy_start_in_input)
232 #define RExC_copy_start_in_constructed (pRExC_state->copy_start)
233 #define RExC_save_copy_start_in_constructed (pRExC_state->save_copy_start)
234 #define RExC_precomp_end (pRExC_state->precomp_end)
235 #define RExC_rx_sv (pRExC_state->rx_sv)
236 #define RExC_rx (pRExC_state->rx)
237 #define RExC_rxi (pRExC_state->rxi)
238 #define RExC_start (pRExC_state->start)
239 #define RExC_end (pRExC_state->end)
240 #define RExC_parse (pRExC_state->parse)
241 #define RExC_latest_warn_offset (pRExC_state->latest_warn_offset )
242 #define RExC_whilem_seen (pRExC_state->whilem_seen)
243 #define RExC_seen_d_op (pRExC_state->seen_d_op) /* Seen something that differs
244 under /d from /u ? */
246 #ifdef RE_TRACK_PATTERN_OFFSETS
247 # define RExC_offsets (RExC_rxi->u.offsets) /* I am not like the
250 #define RExC_emit (pRExC_state->emit)
251 #define RExC_emit_start (pRExC_state->emit_start)
252 #define RExC_sawback (pRExC_state->sawback)
253 #define RExC_seen (pRExC_state->seen)
254 #define RExC_size (pRExC_state->size)
255 #define RExC_maxlen (pRExC_state->maxlen)
256 #define RExC_npar (pRExC_state->npar)
257 #define RExC_total_parens (pRExC_state->total_par)
258 #define RExC_parens_buf_size (pRExC_state->parens_buf_size)
259 #define RExC_nestroot (pRExC_state->nestroot)
260 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
261 #define RExC_utf8 (pRExC_state->utf8)
262 #define RExC_uni_semantics (pRExC_state->uni_semantics)
263 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
264 #define RExC_open_parens (pRExC_state->open_parens)
265 #define RExC_close_parens (pRExC_state->close_parens)
266 #define RExC_end_op (pRExC_state->end_op)
267 #define RExC_paren_names (pRExC_state->paren_names)
268 #define RExC_recurse (pRExC_state->recurse)
269 #define RExC_recurse_count (pRExC_state->recurse_count)
270 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
271 #define RExC_study_chunk_recursed_bytes \
272 (pRExC_state->study_chunk_recursed_bytes)
273 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
274 #define RExC_in_lookahead (pRExC_state->in_lookahead)
275 #define RExC_contains_locale (pRExC_state->contains_locale)
276 #define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
279 # define SET_recode_x_to_native(x) \
280 STMT_START { RExC_recode_x_to_native = (x); } STMT_END
282 # define SET_recode_x_to_native(x) NOOP
285 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
286 #define RExC_frame_head (pRExC_state->frame_head)
287 #define RExC_frame_last (pRExC_state->frame_last)
288 #define RExC_frame_count (pRExC_state->frame_count)
289 #define RExC_strict (pRExC_state->strict)
290 #define RExC_study_started (pRExC_state->study_started)
291 #define RExC_warn_text (pRExC_state->warn_text)
292 #define RExC_in_script_run (pRExC_state->in_script_run)
293 #define RExC_use_BRANCHJ (pRExC_state->use_BRANCHJ)
294 #define RExC_unlexed_names (pRExC_state->unlexed_names)
296 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
297 * a flag to disable back-off on the fixed/floating substrings - if it's
298 * a high complexity pattern we assume the benefit of avoiding a full match
299 * is worth the cost of checking for the substrings even if they rarely help.
301 #define RExC_naughty (pRExC_state->naughty)
302 #define TOO_NAUGHTY (10)
303 #define MARK_NAUGHTY(add) \
304 if (RExC_naughty < TOO_NAUGHTY) \
305 RExC_naughty += (add)
306 #define MARK_NAUGHTY_EXP(exp, add) \
307 if (RExC_naughty < TOO_NAUGHTY) \
308 RExC_naughty += RExC_naughty / (exp) + (add)
310 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
311 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
312 ((*s) == '{' && regcurly(s)))
315 * Flags to be passed up and down.
317 #define WORST 0 /* Worst case. */
318 #define HASWIDTH 0x01 /* Known to not match null strings, could match
321 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
322 * character. (There needs to be a case: in the switch statement in regexec.c
323 * for any node marked SIMPLE.) Note that this is not the same thing as
326 #define SPSTART 0x04 /* Starts with * or + */
327 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
328 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
329 #define RESTART_PARSE 0x20 /* Need to redo the parse */
330 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PARSE, need to
331 calcuate sizes as UTF-8 */
333 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
335 /* whether trie related optimizations are enabled */
336 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
337 #define TRIE_STUDY_OPT
338 #define FULL_TRIE_STUDY
344 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
345 #define PBITVAL(paren) (1 << ((paren) & 7))
346 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
347 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
348 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
350 #define REQUIRE_UTF8(flagp) STMT_START { \
352 *flagp = RESTART_PARSE|NEED_UTF8; \
357 /* Change from /d into /u rules, and restart the parse. RExC_uni_semantics is
358 * a flag that indicates we need to override /d with /u as a result of
359 * something in the pattern. It should only be used in regards to calling
360 * set_regex_charset() or get_regex_charse() */
361 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
363 if (DEPENDS_SEMANTICS) { \
364 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
365 RExC_uni_semantics = 1; \
366 if (RExC_seen_d_op && LIKELY(! IN_PARENS_PASS)) { \
367 /* No need to restart the parse if we haven't seen \
368 * anything that differs between /u and /d, and no need \
369 * to restart immediately if we're going to reparse \
370 * anyway to count parens */ \
371 *flagp |= RESTART_PARSE; \
372 return restart_retval; \
377 #define REQUIRE_BRANCHJ(flagp, restart_retval) \
379 RExC_use_BRANCHJ = 1; \
380 *flagp |= RESTART_PARSE; \
381 return restart_retval; \
384 /* Until we have completed the parse, we leave RExC_total_parens at 0 or
385 * less. After that, it must always be positive, because the whole re is
386 * considered to be surrounded by virtual parens. Setting it to negative
387 * indicates there is some construct that needs to know the actual number of
388 * parens to be properly handled. And that means an extra pass will be
389 * required after we've counted them all */
390 #define ALL_PARENS_COUNTED (RExC_total_parens > 0)
391 #define REQUIRE_PARENS_PASS \
392 STMT_START { /* No-op if have completed a pass */ \
393 if (! ALL_PARENS_COUNTED) RExC_total_parens = -1; \
395 #define IN_PARENS_PASS (RExC_total_parens < 0)
398 /* This is used to return failure (zero) early from the calling function if
399 * various flags in 'flags' are set. Two flags always cause a return:
400 * 'RESTART_PARSE' and 'NEED_UTF8'. 'extra' can be used to specify any
401 * additional flags that should cause a return; 0 if none. If the return will
402 * be done, '*flagp' is first set to be all of the flags that caused the
404 #define RETURN_FAIL_ON_RESTART_OR_FLAGS(flags,flagp,extra) \
406 if ((flags) & (RESTART_PARSE|NEED_UTF8|(extra))) { \
407 *(flagp) = (flags) & (RESTART_PARSE|NEED_UTF8|(extra)); \
412 #define MUST_RESTART(flags) ((flags) & (RESTART_PARSE))
414 #define RETURN_FAIL_ON_RESTART(flags,flagp) \
415 RETURN_FAIL_ON_RESTART_OR_FLAGS( flags, flagp, 0)
416 #define RETURN_FAIL_ON_RESTART_FLAGP(flagp) \
417 if (MUST_RESTART(*(flagp))) return 0
419 /* This converts the named class defined in regcomp.h to its equivalent class
420 * number defined in handy.h. */
421 #define namedclass_to_classnum(class) ((int) ((class) / 2))
422 #define classnum_to_namedclass(classnum) ((classnum) * 2)
424 #define _invlist_union_complement_2nd(a, b, output) \
425 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
426 #define _invlist_intersection_complement_2nd(a, b, output) \
427 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
429 /* About scan_data_t.
431 During optimisation we recurse through the regexp program performing
432 various inplace (keyhole style) optimisations. In addition study_chunk
433 and scan_commit populate this data structure with information about
434 what strings MUST appear in the pattern. We look for the longest
435 string that must appear at a fixed location, and we look for the
436 longest string that may appear at a floating location. So for instance
441 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
442 strings (because they follow a .* construct). study_chunk will identify
443 both FOO and BAR as being the longest fixed and floating strings respectively.
445 The strings can be composites, for instance
449 will result in a composite fixed substring 'foo'.
451 For each string some basic information is maintained:
454 This is the position the string must appear at, or not before.
455 It also implicitly (when combined with minlenp) tells us how many
456 characters must match before the string we are searching for.
457 Likewise when combined with minlenp and the length of the string it
458 tells us how many characters must appear after the string we have
462 Only used for floating strings. This is the rightmost point that
463 the string can appear at. If set to SSize_t_MAX it indicates that the
464 string can occur infinitely far to the right.
465 For fixed strings, it is equal to min_offset.
468 A pointer to the minimum number of characters of the pattern that the
469 string was found inside. This is important as in the case of positive
470 lookahead or positive lookbehind we can have multiple patterns
475 The minimum length of the pattern overall is 3, the minimum length
476 of the lookahead part is 3, but the minimum length of the part that
477 will actually match is 1. So 'FOO's minimum length is 3, but the
478 minimum length for the F is 1. This is important as the minimum length
479 is used to determine offsets in front of and behind the string being
480 looked for. Since strings can be composites this is the length of the
481 pattern at the time it was committed with a scan_commit. Note that
482 the length is calculated by study_chunk, so that the minimum lengths
483 are not known until the full pattern has been compiled, thus the
484 pointer to the value.
488 In the case of lookbehind the string being searched for can be
489 offset past the start point of the final matching string.
490 If this value was just blithely removed from the min_offset it would
491 invalidate some of the calculations for how many chars must match
492 before or after (as they are derived from min_offset and minlen and
493 the length of the string being searched for).
494 When the final pattern is compiled and the data is moved from the
495 scan_data_t structure into the regexp structure the information
496 about lookbehind is factored in, with the information that would
497 have been lost precalculated in the end_shift field for the
500 The fields pos_min and pos_delta are used to store the minimum offset
501 and the delta to the maximum offset at the current point in the pattern.
505 struct scan_data_substrs {
506 SV *str; /* longest substring found in pattern */
507 SSize_t min_offset; /* earliest point in string it can appear */
508 SSize_t max_offset; /* latest point in string it can appear */
509 SSize_t *minlenp; /* pointer to the minlen relevant to the string */
510 SSize_t lookbehind; /* is the pos of the string modified by LB */
511 I32 flags; /* per substring SF_* and SCF_* flags */
514 typedef struct scan_data_t {
515 /*I32 len_min; unused */
516 /*I32 len_delta; unused */
520 SSize_t last_end; /* min value, <0 unless valid. */
521 SSize_t last_start_min;
522 SSize_t last_start_max;
523 U8 cur_is_floating; /* whether the last_* values should be set as
524 * the next fixed (0) or floating (1)
527 /* [0] is longest fixed substring so far, [1] is longest float so far */
528 struct scan_data_substrs substrs[2];
530 I32 flags; /* common SF_* and SCF_* flags */
532 SSize_t *last_closep;
533 regnode_ssc *start_class;
537 * Forward declarations for pregcomp()'s friends.
540 static const scan_data_t zero_scan_data = {
541 0, 0, NULL, 0, 0, 0, 0,
543 { NULL, 0, 0, 0, 0, 0 },
544 { NULL, 0, 0, 0, 0, 0 },
551 #define SF_BEFORE_SEOL 0x0001
552 #define SF_BEFORE_MEOL 0x0002
553 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
555 #define SF_IS_INF 0x0040
556 #define SF_HAS_PAR 0x0080
557 #define SF_IN_PAR 0x0100
558 #define SF_HAS_EVAL 0x0200
561 /* SCF_DO_SUBSTR is the flag that tells the regexp analyzer to track the
562 * longest substring in the pattern. When it is not set the optimiser keeps
563 * track of position, but does not keep track of the actual strings seen,
565 * So for instance /foo/ will be parsed with SCF_DO_SUBSTR being true, but
568 * Similarly, /foo.*(blah|erm|huh).*fnorble/ will have "foo" and "fnorble"
569 * parsed with SCF_DO_SUBSTR on, but while processing the (...) it will be
570 * turned off because of the alternation (BRANCH). */
571 #define SCF_DO_SUBSTR 0x0400
573 #define SCF_DO_STCLASS_AND 0x0800
574 #define SCF_DO_STCLASS_OR 0x1000
575 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
576 #define SCF_WHILEM_VISITED_POS 0x2000
578 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
579 #define SCF_SEEN_ACCEPT 0x8000
580 #define SCF_TRIE_DOING_RESTUDY 0x10000
581 #define SCF_IN_DEFINE 0x20000
586 #define UTF cBOOL(RExC_utf8)
588 /* The enums for all these are ordered so things work out correctly */
589 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
590 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
591 == REGEX_DEPENDS_CHARSET)
592 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
593 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
594 >= REGEX_UNICODE_CHARSET)
595 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
596 == REGEX_ASCII_RESTRICTED_CHARSET)
597 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
598 >= REGEX_ASCII_RESTRICTED_CHARSET)
599 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
600 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
602 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
604 /* For programs that want to be strictly Unicode compatible by dying if any
605 * attempt is made to match a non-Unicode code point against a Unicode
607 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
609 #define OOB_NAMEDCLASS -1
611 /* There is no code point that is out-of-bounds, so this is problematic. But
612 * its only current use is to initialize a variable that is always set before
614 #define OOB_UNICODE 0xDEADBEEF
616 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
619 /* length of regex to show in messages that don't mark a position within */
620 #define RegexLengthToShowInErrorMessages 127
623 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
624 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
625 * op/pragma/warn/regcomp.
627 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
628 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
630 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
631 " in m/%" UTF8f MARKER2 "%" UTF8f "/"
633 /* The code in this file in places uses one level of recursion with parsing
634 * rebased to an alternate string constructed by us in memory. This can take
635 * the form of something that is completely different from the input, or
636 * something that uses the input as part of the alternate. In the first case,
637 * there should be no possibility of an error, as we are in complete control of
638 * the alternate string. But in the second case we don't completely control
639 * the input portion, so there may be errors in that. Here's an example:
641 * is handled specially because \x{df} folds to a sequence of more than one
642 * character: 'ss'. What is done is to create and parse an alternate string,
643 * which looks like this:
644 * /(?:\x{DF}|[abc\x{DF}def])/ui
645 * where it uses the input unchanged in the middle of something it constructs,
646 * which is a branch for the DF outside the character class, and clustering
647 * parens around the whole thing. (It knows enough to skip the DF inside the
648 * class while in this substitute parse.) 'abc' and 'def' may have errors that
649 * need to be reported. The general situation looks like this:
651 * |<------- identical ------>|
653 * Input: ---------------------------------------------------------------
654 * Constructed: ---------------------------------------------------
656 * |<------- identical ------>|
658 * sI..eI is the portion of the input pattern we are concerned with here.
659 * sC..EC is the constructed substitute parse string.
660 * sC..tC is constructed by us
661 * tC..eC is an exact duplicate of the portion of the input pattern tI..eI.
662 * In the diagram, these are vertically aligned.
663 * eC..EC is also constructed by us.
664 * xC is the position in the substitute parse string where we found a
666 * xI is the position in the original pattern corresponding to xC.
668 * We want to display a message showing the real input string. Thus we need to
669 * translate from xC to xI. We know that xC >= tC, since the portion of the
670 * string sC..tC has been constructed by us, and so shouldn't have errors. We
672 * xI = tI + (xC - tC)
674 * When the substitute parse is constructed, the code needs to set:
677 * RExC_copy_start_in_input (tI)
678 * RExC_copy_start_in_constructed (tC)
679 * and restore them when done.
681 * During normal processing of the input pattern, both
682 * 'RExC_copy_start_in_input' and 'RExC_copy_start_in_constructed' are set to
683 * sI, so that xC equals xI.
686 #define sI RExC_precomp
687 #define eI RExC_precomp_end
688 #define sC RExC_start
690 #define tI RExC_copy_start_in_input
691 #define tC RExC_copy_start_in_constructed
692 #define xI(xC) (tI + (xC - tC))
693 #define xI_offset(xC) (xI(xC) - sI)
695 #define REPORT_LOCATION_ARGS(xC) \
697 (xI(xC) > eI) /* Don't run off end */ \
698 ? eI - sI /* Length before the <--HERE */ \
699 : ((xI_offset(xC) >= 0) \
701 : (Perl_croak(aTHX_ "panic: %s: %d: negative offset: %" \
702 IVdf " trying to output message for " \
704 __FILE__, __LINE__, (IV) xI_offset(xC), \
705 ((int) (eC - sC)), sC), 0)), \
706 sI), /* The input pattern printed up to the <--HERE */ \
708 (xI(xC) > eI) ? 0 : eI - xI(xC), /* Length after <--HERE */ \
709 (xI(xC) > eI) ? eI : xI(xC)) /* pattern after <--HERE */
711 /* Used to point after bad bytes for an error message, but avoid skipping
712 * past a nul byte. */
713 #define SKIP_IF_CHAR(s, e) (!*(s) ? 0 : UTF ? UTF8_SAFE_SKIP(s, e) : 1)
715 /* Set up to clean up after our imminent demise */
716 #define PREPARE_TO_DIE \
719 SAVEFREESV(RExC_rx_sv); \
720 if (RExC_open_parens) \
721 SAVEFREEPV(RExC_open_parens); \
722 if (RExC_close_parens) \
723 SAVEFREEPV(RExC_close_parens); \
727 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
728 * arg. Show regex, up to a maximum length. If it's too long, chop and add
731 #define _FAIL(code) STMT_START { \
732 const char *ellipses = ""; \
733 IV len = RExC_precomp_end - RExC_precomp; \
736 if (len > RegexLengthToShowInErrorMessages) { \
737 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
738 len = RegexLengthToShowInErrorMessages - 10; \
744 #define FAIL(msg) _FAIL( \
745 Perl_croak(aTHX_ "%s in regex m/%" UTF8f "%s/", \
746 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
748 #define FAIL2(msg,arg) _FAIL( \
749 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
750 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
752 #define FAIL3(msg,arg1,arg2) _FAIL( \
753 Perl_croak(aTHX_ msg " in regex m/%" UTF8f "%s/", \
754 arg1, arg2, UTF8fARG(UTF, len, RExC_precomp), ellipses))
757 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
759 #define Simple_vFAIL(m) STMT_START { \
760 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
761 m, REPORT_LOCATION_ARGS(RExC_parse)); \
765 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
767 #define vFAIL(m) STMT_START { \
773 * Like Simple_vFAIL(), but accepts two arguments.
775 #define Simple_vFAIL2(m,a1) STMT_START { \
776 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
777 REPORT_LOCATION_ARGS(RExC_parse)); \
781 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
783 #define vFAIL2(m,a1) STMT_START { \
785 Simple_vFAIL2(m, a1); \
790 * Like Simple_vFAIL(), but accepts three arguments.
792 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
793 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
794 REPORT_LOCATION_ARGS(RExC_parse)); \
798 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
800 #define vFAIL3(m,a1,a2) STMT_START { \
802 Simple_vFAIL3(m, a1, a2); \
806 * Like Simple_vFAIL(), but accepts four arguments.
808 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
809 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
810 REPORT_LOCATION_ARGS(RExC_parse)); \
813 #define vFAIL4(m,a1,a2,a3) STMT_START { \
815 Simple_vFAIL4(m, a1, a2, a3); \
818 /* A specialized version of vFAIL2 that works with UTF8f */
819 #define vFAIL2utf8f(m, a1) STMT_START { \
821 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
822 REPORT_LOCATION_ARGS(RExC_parse)); \
825 #define vFAIL3utf8f(m, a1, a2) STMT_START { \
827 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
828 REPORT_LOCATION_ARGS(RExC_parse)); \
831 /* Setting this to NULL is a signal to not output warnings */
832 #define TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE \
834 RExC_save_copy_start_in_constructed = RExC_copy_start_in_constructed;\
835 RExC_copy_start_in_constructed = NULL; \
837 #define RESTORE_WARNINGS \
838 RExC_copy_start_in_constructed = RExC_save_copy_start_in_constructed
840 /* Since a warning can be generated multiple times as the input is reparsed, we
841 * output it the first time we come to that point in the parse, but suppress it
842 * otherwise. 'RExC_copy_start_in_constructed' being NULL is a flag to not
843 * generate any warnings */
844 #define TO_OUTPUT_WARNINGS(loc) \
845 ( RExC_copy_start_in_constructed \
846 && ((xI(loc)) - RExC_precomp) > (Ptrdiff_t) RExC_latest_warn_offset)
848 /* After we've emitted a warning, we save the position in the input so we don't
850 #define UPDATE_WARNINGS_LOC(loc) \
852 if (TO_OUTPUT_WARNINGS(loc)) { \
853 RExC_latest_warn_offset = MAX(sI, MIN(eI, xI(loc))) \
858 /* 'warns' is the output of the packWARNx macro used in 'code' */
859 #define _WARN_HELPER(loc, warns, code) \
861 if (! RExC_copy_start_in_constructed) { \
862 Perl_croak( aTHX_ "panic! %s: %d: Tried to warn when none" \
863 " expected at '%s'", \
864 __FILE__, __LINE__, loc); \
866 if (TO_OUTPUT_WARNINGS(loc)) { \
870 UPDATE_WARNINGS_LOC(loc); \
874 /* m is not necessarily a "literal string", in this macro */
875 #define reg_warn_non_literal_string(loc, m) \
876 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
877 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
878 "%s" REPORT_LOCATION, \
879 m, REPORT_LOCATION_ARGS(loc)))
881 #define ckWARNreg(loc,m) \
882 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
883 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
885 REPORT_LOCATION_ARGS(loc)))
887 #define vWARN(loc, m) \
888 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
889 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
891 REPORT_LOCATION_ARGS(loc))) \
893 #define vWARN_dep(loc, m) \
894 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
895 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), \
897 REPORT_LOCATION_ARGS(loc)))
899 #define ckWARNdep(loc,m) \
900 _WARN_HELPER(loc, packWARN(WARN_DEPRECATED), \
901 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
903 REPORT_LOCATION_ARGS(loc)))
905 #define ckWARNregdep(loc,m) \
906 _WARN_HELPER(loc, packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
907 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, \
910 REPORT_LOCATION_ARGS(loc)))
912 #define ckWARN2reg_d(loc,m, a1) \
913 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
914 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
916 a1, REPORT_LOCATION_ARGS(loc)))
918 #define ckWARN2reg(loc, m, a1) \
919 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
920 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
922 a1, REPORT_LOCATION_ARGS(loc)))
924 #define vWARN3(loc, m, a1, a2) \
925 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
926 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
928 a1, a2, REPORT_LOCATION_ARGS(loc)))
930 #define ckWARN3reg(loc, m, a1, a2) \
931 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
932 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
935 REPORT_LOCATION_ARGS(loc)))
937 #define vWARN4(loc, m, a1, a2, a3) \
938 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
939 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
942 REPORT_LOCATION_ARGS(loc)))
944 #define ckWARN4reg(loc, m, a1, a2, a3) \
945 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
946 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), \
949 REPORT_LOCATION_ARGS(loc)))
951 #define vWARN5(loc, m, a1, a2, a3, a4) \
952 _WARN_HELPER(loc, packWARN(WARN_REGEXP), \
953 Perl_warner(aTHX_ packWARN(WARN_REGEXP), \
956 REPORT_LOCATION_ARGS(loc)))
958 #define ckWARNexperimental(loc, class, m) \
959 _WARN_HELPER(loc, packWARN(class), \
960 Perl_ck_warner_d(aTHX_ packWARN(class), \
962 REPORT_LOCATION_ARGS(loc)))
964 /* Convert between a pointer to a node and its offset from the beginning of the
966 #define REGNODE_p(offset) (RExC_emit_start + (offset))
967 #define REGNODE_OFFSET(node) ((node) - RExC_emit_start)
969 /* Macros for recording node offsets. 20001227 mjd@plover.com
970 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
971 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
972 * Element 0 holds the number n.
973 * Position is 1 indexed.
975 #ifndef RE_TRACK_PATTERN_OFFSETS
976 #define Set_Node_Offset_To_R(offset,byte)
977 #define Set_Node_Offset(node,byte)
978 #define Set_Cur_Node_Offset
979 #define Set_Node_Length_To_R(node,len)
980 #define Set_Node_Length(node,len)
981 #define Set_Node_Cur_Length(node,start)
982 #define Node_Offset(n)
983 #define Node_Length(n)
984 #define Set_Node_Offset_Length(node,offset,len)
985 #define ProgLen(ri) ri->u.proglen
986 #define SetProgLen(ri,x) ri->u.proglen = x
987 #define Track_Code(code)
989 #define ProgLen(ri) ri->u.offsets[0]
990 #define SetProgLen(ri,x) ri->u.offsets[0] = x
991 #define Set_Node_Offset_To_R(offset,byte) STMT_START { \
992 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
993 __LINE__, (int)(offset), (int)(byte))); \
995 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
998 RExC_offsets[2*(offset)-1] = (byte); \
1002 #define Set_Node_Offset(node,byte) \
1003 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (byte)-RExC_start)
1004 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
1006 #define Set_Node_Length_To_R(node,len) STMT_START { \
1007 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
1008 __LINE__, (int)(node), (int)(len))); \
1010 Perl_croak(aTHX_ "value of node is %d in Length macro", \
1013 RExC_offsets[2*(node)] = (len); \
1017 #define Set_Node_Length(node,len) \
1018 Set_Node_Length_To_R(REGNODE_OFFSET(node), len)
1019 #define Set_Node_Cur_Length(node, start) \
1020 Set_Node_Length(node, RExC_parse - start)
1022 /* Get offsets and lengths */
1023 #define Node_Offset(n) (RExC_offsets[2*(REGNODE_OFFSET(n))-1])
1024 #define Node_Length(n) (RExC_offsets[2*(REGNODE_OFFSET(n))])
1026 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
1027 Set_Node_Offset_To_R(REGNODE_OFFSET(node), (offset)); \
1028 Set_Node_Length_To_R(REGNODE_OFFSET(node), (len)); \
1031 #define Track_Code(code) STMT_START { code } STMT_END
1034 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
1035 #define EXPERIMENTAL_INPLACESCAN
1036 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
1040 Perl_re_printf(pTHX_ const char *fmt, ...)
1044 PerlIO *f= Perl_debug_log;
1045 PERL_ARGS_ASSERT_RE_PRINTF;
1047 result = PerlIO_vprintf(f, fmt, ap);
1053 Perl_re_indentf(pTHX_ const char *fmt, U32 depth, ...)
1057 PerlIO *f= Perl_debug_log;
1058 PERL_ARGS_ASSERT_RE_INDENTF;
1059 va_start(ap, depth);
1060 PerlIO_printf(f, "%*s", ( (int)depth % 20 ) * 2, "");
1061 result = PerlIO_vprintf(f, fmt, ap);
1065 #endif /* DEBUGGING */
1067 #define DEBUG_RExC_seen() \
1068 DEBUG_OPTIMISE_MORE_r({ \
1069 Perl_re_printf( aTHX_ "RExC_seen: "); \
1071 if (RExC_seen & REG_ZERO_LEN_SEEN) \
1072 Perl_re_printf( aTHX_ "REG_ZERO_LEN_SEEN "); \
1074 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
1075 Perl_re_printf( aTHX_ "REG_LOOKBEHIND_SEEN "); \
1077 if (RExC_seen & REG_GPOS_SEEN) \
1078 Perl_re_printf( aTHX_ "REG_GPOS_SEEN "); \
1080 if (RExC_seen & REG_RECURSE_SEEN) \
1081 Perl_re_printf( aTHX_ "REG_RECURSE_SEEN "); \
1083 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
1084 Perl_re_printf( aTHX_ "REG_TOP_LEVEL_BRANCHES_SEEN "); \
1086 if (RExC_seen & REG_VERBARG_SEEN) \
1087 Perl_re_printf( aTHX_ "REG_VERBARG_SEEN "); \
1089 if (RExC_seen & REG_CUTGROUP_SEEN) \
1090 Perl_re_printf( aTHX_ "REG_CUTGROUP_SEEN "); \
1092 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
1093 Perl_re_printf( aTHX_ "REG_RUN_ON_COMMENT_SEEN "); \
1095 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
1096 Perl_re_printf( aTHX_ "REG_UNFOLDED_MULTI_SEEN "); \
1098 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
1099 Perl_re_printf( aTHX_ "REG_UNBOUNDED_QUANTIFIER_SEEN "); \
1101 Perl_re_printf( aTHX_ "\n"); \
1104 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
1105 if ((flags) & flag) Perl_re_printf( aTHX_ "%s ", #flag)
1110 S_debug_show_study_flags(pTHX_ U32 flags, const char *open_str,
1111 const char *close_str)
1116 Perl_re_printf( aTHX_ "%s", open_str);
1117 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_SEOL);
1118 DEBUG_SHOW_STUDY_FLAG(flags, SF_BEFORE_MEOL);
1119 DEBUG_SHOW_STUDY_FLAG(flags, SF_IS_INF);
1120 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_PAR);
1121 DEBUG_SHOW_STUDY_FLAG(flags, SF_IN_PAR);
1122 DEBUG_SHOW_STUDY_FLAG(flags, SF_HAS_EVAL);
1123 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_SUBSTR);
1124 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_AND);
1125 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS_OR);
1126 DEBUG_SHOW_STUDY_FLAG(flags, SCF_DO_STCLASS);
1127 DEBUG_SHOW_STUDY_FLAG(flags, SCF_WHILEM_VISITED_POS);
1128 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_RESTUDY);
1129 DEBUG_SHOW_STUDY_FLAG(flags, SCF_SEEN_ACCEPT);
1130 DEBUG_SHOW_STUDY_FLAG(flags, SCF_TRIE_DOING_RESTUDY);
1131 DEBUG_SHOW_STUDY_FLAG(flags, SCF_IN_DEFINE);
1132 Perl_re_printf( aTHX_ "%s", close_str);
1137 S_debug_studydata(pTHX_ const char *where, scan_data_t *data,
1138 U32 depth, int is_inf)
1140 GET_RE_DEBUG_FLAGS_DECL;
1142 DEBUG_OPTIMISE_MORE_r({
1145 Perl_re_indentf(aTHX_ "%s: Pos:%" IVdf "/%" IVdf " Flags: 0x%" UVXf,
1149 (IV)data->pos_delta,
1153 S_debug_show_study_flags(aTHX_ data->flags," [","]");
1155 Perl_re_printf( aTHX_
1156 " Whilem_c: %" IVdf " Lcp: %" IVdf " %s",
1158 (IV)(data->last_closep ? *((data)->last_closep) : -1),
1159 is_inf ? "INF " : ""
1162 if (data->last_found) {
1164 Perl_re_printf(aTHX_
1165 "Last:'%s' %" IVdf ":%" IVdf "/%" IVdf,
1166 SvPVX_const(data->last_found),
1168 (IV)data->last_start_min,
1169 (IV)data->last_start_max
1172 for (i = 0; i < 2; i++) {
1173 Perl_re_printf(aTHX_
1174 " %s%s: '%s' @ %" IVdf "/%" IVdf,
1175 data->cur_is_floating == i ? "*" : "",
1176 i ? "Float" : "Fixed",
1177 SvPVX_const(data->substrs[i].str),
1178 (IV)data->substrs[i].min_offset,
1179 (IV)data->substrs[i].max_offset
1181 S_debug_show_study_flags(aTHX_ data->substrs[i].flags," [","]");
1185 Perl_re_printf( aTHX_ "\n");
1191 S_debug_peep(pTHX_ const char *str, const RExC_state_t *pRExC_state,
1192 regnode *scan, U32 depth, U32 flags)
1194 GET_RE_DEBUG_FLAGS_DECL;
1201 Next = regnext(scan);
1202 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
1203 Perl_re_indentf( aTHX_ "%s>%3d: %s (%d)",
1206 REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),
1207 Next ? (REG_NODE_NUM(Next)) : 0 );
1208 S_debug_show_study_flags(aTHX_ flags," [ ","]");
1209 Perl_re_printf( aTHX_ "\n");
1214 # define DEBUG_STUDYDATA(where, data, depth, is_inf) \
1215 S_debug_studydata(aTHX_ where, data, depth, is_inf)
1217 # define DEBUG_PEEP(str, scan, depth, flags) \
1218 S_debug_peep(aTHX_ str, pRExC_state, scan, depth, flags)
1221 # define DEBUG_STUDYDATA(where, data, depth, is_inf) NOOP
1222 # define DEBUG_PEEP(str, scan, depth, flags) NOOP
1226 /* =========================================================
1227 * BEGIN edit_distance stuff.
1229 * This calculates how many single character changes of any type are needed to
1230 * transform a string into another one. It is taken from version 3.1 of
1232 * https://metacpan.org/pod/Text::Levenshtein::Damerau::XS
1235 /* Our unsorted dictionary linked list. */
1236 /* Note we use UVs, not chars. */
1241 struct dictionary* next;
1243 typedef struct dictionary item;
1246 PERL_STATIC_INLINE item*
1247 push(UV key, item* curr)
1250 Newx(head, 1, item);
1258 PERL_STATIC_INLINE item*
1259 find(item* head, UV key)
1261 item* iterator = head;
1263 if (iterator->key == key){
1266 iterator = iterator->next;
1272 PERL_STATIC_INLINE item*
1273 uniquePush(item* head, UV key)
1275 item* iterator = head;
1278 if (iterator->key == key) {
1281 iterator = iterator->next;
1284 return push(key, head);
1287 PERL_STATIC_INLINE void
1288 dict_free(item* head)
1290 item* iterator = head;
1293 item* temp = iterator;
1294 iterator = iterator->next;
1301 /* End of Dictionary Stuff */
1303 /* All calculations/work are done here */
1305 S_edit_distance(const UV* src,
1307 const STRLEN x, /* length of src[] */
1308 const STRLEN y, /* length of tgt[] */
1309 const SSize_t maxDistance
1313 UV swapCount, swapScore, targetCharCount, i, j;
1315 UV score_ceil = x + y;
1317 PERL_ARGS_ASSERT_EDIT_DISTANCE;
1319 /* intialize matrix start values */
1320 Newx(scores, ( (x + 2) * (y + 2)), UV);
1321 scores[0] = score_ceil;
1322 scores[1 * (y + 2) + 0] = score_ceil;
1323 scores[0 * (y + 2) + 1] = score_ceil;
1324 scores[1 * (y + 2) + 1] = 0;
1325 head = uniquePush(uniquePush(head, src[0]), tgt[0]);
1330 for (i=1;i<=x;i++) {
1332 head = uniquePush(head, src[i]);
1333 scores[(i+1) * (y + 2) + 1] = i;
1334 scores[(i+1) * (y + 2) + 0] = score_ceil;
1337 for (j=1;j<=y;j++) {
1340 head = uniquePush(head, tgt[j]);
1341 scores[1 * (y + 2) + (j + 1)] = j;
1342 scores[0 * (y + 2) + (j + 1)] = score_ceil;
1345 targetCharCount = find(head, tgt[j-1])->value;
1346 swapScore = scores[targetCharCount * (y + 2) + swapCount] + i - targetCharCount - 1 + j - swapCount;
1348 if (src[i-1] != tgt[j-1]){
1349 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));
1353 scores[(i+1) * (y + 2) + (j + 1)] = MIN(scores[i * (y + 2) + j], swapScore);
1357 find(head, src[i-1])->value = i;
1361 IV score = scores[(x+1) * (y + 2) + (y + 1)];
1364 return (maxDistance != 0 && maxDistance < score)?(-1):score;
1368 /* END of edit_distance() stuff
1369 * ========================================================= */
1371 /* is c a control character for which we have a mnemonic? */
1372 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
1375 S_cntrl_to_mnemonic(const U8 c)
1377 /* Returns the mnemonic string that represents character 'c', if one
1378 * exists; NULL otherwise. The only ones that exist for the purposes of
1379 * this routine are a few control characters */
1382 case '\a': return "\\a";
1383 case '\b': return "\\b";
1384 case ESC_NATIVE: return "\\e";
1385 case '\f': return "\\f";
1386 case '\n': return "\\n";
1387 case '\r': return "\\r";
1388 case '\t': return "\\t";
1394 /* Mark that we cannot extend a found fixed substring at this point.
1395 Update the longest found anchored substring or the longest found
1396 floating substrings if needed. */
1399 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
1400 SSize_t *minlenp, int is_inf)
1402 const STRLEN l = CHR_SVLEN(data->last_found);
1403 SV * const longest_sv = data->substrs[data->cur_is_floating].str;
1404 const STRLEN old_l = CHR_SVLEN(longest_sv);
1405 GET_RE_DEBUG_FLAGS_DECL;
1407 PERL_ARGS_ASSERT_SCAN_COMMIT;
1409 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
1410 const U8 i = data->cur_is_floating;
1411 SvSetMagicSV(longest_sv, data->last_found);
1412 data->substrs[i].min_offset = l ? data->last_start_min : data->pos_min;
1415 data->substrs[0].max_offset = data->substrs[0].min_offset;
1417 data->substrs[1].max_offset = (l
1418 ? data->last_start_max
1419 : (data->pos_delta > SSize_t_MAX - data->pos_min
1421 : data->pos_min + data->pos_delta));
1423 || (STRLEN)data->substrs[1].max_offset > (STRLEN)SSize_t_MAX)
1424 data->substrs[1].max_offset = SSize_t_MAX;
1427 if (data->flags & SF_BEFORE_EOL)
1428 data->substrs[i].flags |= (data->flags & SF_BEFORE_EOL);
1430 data->substrs[i].flags &= ~SF_BEFORE_EOL;
1431 data->substrs[i].minlenp = minlenp;
1432 data->substrs[i].lookbehind = 0;
1435 SvCUR_set(data->last_found, 0);
1437 SV * const sv = data->last_found;
1438 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1439 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1444 data->last_end = -1;
1445 data->flags &= ~SF_BEFORE_EOL;
1446 DEBUG_STUDYDATA("commit", data, 0, is_inf);
1449 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1450 * list that describes which code points it matches */
1453 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1455 /* Set the SSC 'ssc' to match an empty string or any code point */
1457 PERL_ARGS_ASSERT_SSC_ANYTHING;
1459 assert(is_ANYOF_SYNTHETIC(ssc));
1461 /* mortalize so won't leak */
1462 ssc->invlist = sv_2mortal(_add_range_to_invlist(NULL, 0, UV_MAX));
1463 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1467 S_ssc_is_anything(const regnode_ssc *ssc)
1469 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1470 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1471 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1472 * in any way, so there's no point in using it */
1477 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1479 assert(is_ANYOF_SYNTHETIC(ssc));
1481 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1485 /* See if the list consists solely of the range 0 - Infinity */
1486 invlist_iterinit(ssc->invlist);
1487 ret = invlist_iternext(ssc->invlist, &start, &end)
1491 invlist_iterfinish(ssc->invlist);
1497 /* If e.g., both \w and \W are set, matches everything */
1498 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1500 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1501 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1511 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1513 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1514 * string, any code point, or any posix class under locale */
1516 PERL_ARGS_ASSERT_SSC_INIT;
1518 Zero(ssc, 1, regnode_ssc);
1519 set_ANYOF_SYNTHETIC(ssc);
1520 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1523 /* If any portion of the regex is to operate under locale rules that aren't
1524 * fully known at compile time, initialization includes it. The reason
1525 * this isn't done for all regexes is that the optimizer was written under
1526 * the assumption that locale was all-or-nothing. Given the complexity and
1527 * lack of documentation in the optimizer, and that there are inadequate
1528 * test cases for locale, many parts of it may not work properly, it is
1529 * safest to avoid locale unless necessary. */
1530 if (RExC_contains_locale) {
1531 ANYOF_POSIXL_SETALL(ssc);
1534 ANYOF_POSIXL_ZERO(ssc);
1539 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1540 const regnode_ssc *ssc)
1542 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1543 * to the list of code points matched, and locale posix classes; hence does
1544 * not check its flags) */
1549 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1551 assert(is_ANYOF_SYNTHETIC(ssc));
1553 invlist_iterinit(ssc->invlist);
1554 ret = invlist_iternext(ssc->invlist, &start, &end)
1558 invlist_iterfinish(ssc->invlist);
1564 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1571 #define INVLIST_INDEX 0
1572 #define ONLY_LOCALE_MATCHES_INDEX 1
1573 #define DEFERRED_USER_DEFINED_INDEX 2
1576 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1577 const regnode_charclass* const node)
1579 /* Returns a mortal inversion list defining which code points are matched
1580 * by 'node', which is of type ANYOF. Handles complementing the result if
1581 * appropriate. If some code points aren't knowable at this time, the
1582 * returned list must, and will, contain every code point that is a
1587 SV* only_utf8_locale_invlist = NULL;
1589 const U32 n = ARG(node);
1590 bool new_node_has_latin1 = FALSE;
1591 const U8 flags = (inRANGE(OP(node), ANYOFH, ANYOFHr))
1593 : ANYOF_FLAGS(node);
1595 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1597 /* Look at the data structure created by S_set_ANYOF_arg() */
1598 if (n != ANYOF_ONLY_HAS_BITMAP) {
1599 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1600 AV * const av = MUTABLE_AV(SvRV(rv));
1601 SV **const ary = AvARRAY(av);
1602 assert(RExC_rxi->data->what[n] == 's');
1604 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
1606 /* Here there are things that won't be known until runtime -- we
1607 * have to assume it could be anything */
1608 invlist = sv_2mortal(_new_invlist(1));
1609 return _add_range_to_invlist(invlist, 0, UV_MAX);
1611 else if (ary[INVLIST_INDEX]) {
1613 /* Use the node's inversion list */
1614 invlist = sv_2mortal(invlist_clone(ary[INVLIST_INDEX], NULL));
1617 /* Get the code points valid only under UTF-8 locales */
1618 if ( (flags & ANYOFL_FOLD)
1619 && av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX)
1621 only_utf8_locale_invlist = ary[ONLY_LOCALE_MATCHES_INDEX];
1626 invlist = sv_2mortal(_new_invlist(0));
1629 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1630 * code points, and an inversion list for the others, but if there are code
1631 * points that should match only conditionally on the target string being
1632 * UTF-8, those are placed in the inversion list, and not the bitmap.
1633 * Since there are circumstances under which they could match, they are
1634 * included in the SSC. But if the ANYOF node is to be inverted, we have
1635 * to exclude them here, so that when we invert below, the end result
1636 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1637 * have to do this here before we add the unconditionally matched code
1639 if (flags & ANYOF_INVERT) {
1640 _invlist_intersection_complement_2nd(invlist,
1645 /* Add in the points from the bit map */
1646 if (! inRANGE(OP(node), ANYOFH, ANYOFHr)) {
1647 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1648 if (ANYOF_BITMAP_TEST(node, i)) {
1649 unsigned int start = i++;
1651 for (; i < NUM_ANYOF_CODE_POINTS
1652 && ANYOF_BITMAP_TEST(node, i); ++i)
1656 invlist = _add_range_to_invlist(invlist, start, i-1);
1657 new_node_has_latin1 = TRUE;
1662 /* If this can match all upper Latin1 code points, have to add them
1663 * as well. But don't add them if inverting, as when that gets done below,
1664 * it would exclude all these characters, including the ones it shouldn't
1665 * that were added just above */
1666 if (! (flags & ANYOF_INVERT) && OP(node) == ANYOFD
1667 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1669 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1672 /* Similarly for these */
1673 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1674 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1677 if (flags & ANYOF_INVERT) {
1678 _invlist_invert(invlist);
1680 else if (flags & ANYOFL_FOLD) {
1681 if (new_node_has_latin1) {
1683 /* Under /li, any 0-255 could fold to any other 0-255, depending on
1684 * the locale. We can skip this if there are no 0-255 at all. */
1685 _invlist_union(invlist, PL_Latin1, &invlist);
1687 invlist = add_cp_to_invlist(invlist, LATIN_SMALL_LETTER_DOTLESS_I);
1688 invlist = add_cp_to_invlist(invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
1691 if (_invlist_contains_cp(invlist, LATIN_SMALL_LETTER_DOTLESS_I)) {
1692 invlist = add_cp_to_invlist(invlist, 'I');
1694 if (_invlist_contains_cp(invlist,
1695 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE))
1697 invlist = add_cp_to_invlist(invlist, 'i');
1702 /* Similarly add the UTF-8 locale possible matches. These have to be
1703 * deferred until after the non-UTF-8 locale ones are taken care of just
1704 * above, or it leads to wrong results under ANYOF_INVERT */
1705 if (only_utf8_locale_invlist) {
1706 _invlist_union_maybe_complement_2nd(invlist,
1707 only_utf8_locale_invlist,
1708 flags & ANYOF_INVERT,
1715 /* These two functions currently do the exact same thing */
1716 #define ssc_init_zero ssc_init
1718 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1719 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1721 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1722 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1723 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1726 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1727 const regnode_charclass *and_with)
1729 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1730 * another SSC or a regular ANYOF class. Can create false positives. */
1733 U8 and_with_flags = inRANGE(OP(and_with), ANYOFH, ANYOFHr)
1735 : ANYOF_FLAGS(and_with);
1738 PERL_ARGS_ASSERT_SSC_AND;
1740 assert(is_ANYOF_SYNTHETIC(ssc));
1742 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1743 * the code point inversion list and just the relevant flags */
1744 if (is_ANYOF_SYNTHETIC(and_with)) {
1745 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1746 anded_flags = and_with_flags;
1748 /* XXX This is a kludge around what appears to be deficiencies in the
1749 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1750 * there are paths through the optimizer where it doesn't get weeded
1751 * out when it should. And if we don't make some extra provision for
1752 * it like the code just below, it doesn't get added when it should.
1753 * This solution is to add it only when AND'ing, which is here, and
1754 * only when what is being AND'ed is the pristine, original node
1755 * matching anything. Thus it is like adding it to ssc_anything() but
1756 * only when the result is to be AND'ed. Probably the same solution
1757 * could be adopted for the same problem we have with /l matching,
1758 * which is solved differently in S_ssc_init(), and that would lead to
1759 * fewer false positives than that solution has. But if this solution
1760 * creates bugs, the consequences are only that a warning isn't raised
1761 * that should be; while the consequences for having /l bugs is
1762 * incorrect matches */
1763 if (ssc_is_anything((regnode_ssc *)and_with)) {
1764 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1768 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1769 if (OP(and_with) == ANYOFD) {
1770 anded_flags = and_with_flags & ANYOF_COMMON_FLAGS;
1773 anded_flags = and_with_flags
1774 &( ANYOF_COMMON_FLAGS
1775 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1776 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1777 if (ANYOFL_UTF8_LOCALE_REQD(and_with_flags)) {
1779 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1784 ANYOF_FLAGS(ssc) &= anded_flags;
1786 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1787 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1788 * 'and_with' may be inverted. When not inverted, we have the situation of
1790 * (C1 | P1) & (C2 | P2)
1791 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1792 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1793 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1794 * <= ((C1 & C2) | P1 | P2)
1795 * Alternatively, the last few steps could be:
1796 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1797 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1798 * <= (C1 | C2 | (P1 & P2))
1799 * We favor the second approach if either P1 or P2 is non-empty. This is
1800 * because these components are a barrier to doing optimizations, as what
1801 * they match cannot be known until the moment of matching as they are
1802 * dependent on the current locale, 'AND"ing them likely will reduce or
1804 * But we can do better if we know that C1,P1 are in their initial state (a
1805 * frequent occurrence), each matching everything:
1806 * (<everything>) & (C2 | P2) = C2 | P2
1807 * Similarly, if C2,P2 are in their initial state (again a frequent
1808 * occurrence), the result is a no-op
1809 * (C1 | P1) & (<everything>) = C1 | P1
1812 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1813 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1814 * <= (C1 & ~C2) | (P1 & ~P2)
1817 if ((and_with_flags & ANYOF_INVERT)
1818 && ! is_ANYOF_SYNTHETIC(and_with))
1822 ssc_intersection(ssc,
1824 FALSE /* Has already been inverted */
1827 /* If either P1 or P2 is empty, the intersection will be also; can skip
1829 if (! (and_with_flags & ANYOF_MATCHES_POSIXL)) {
1830 ANYOF_POSIXL_ZERO(ssc);
1832 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1834 /* Note that the Posix class component P from 'and_with' actually
1836 * P = Pa | Pb | ... | Pn
1837 * where each component is one posix class, such as in [\w\s].
1839 * ~P = ~(Pa | Pb | ... | Pn)
1840 * = ~Pa & ~Pb & ... & ~Pn
1841 * <= ~Pa | ~Pb | ... | ~Pn
1842 * The last is something we can easily calculate, but unfortunately
1843 * is likely to have many false positives. We could do better
1844 * in some (but certainly not all) instances if two classes in
1845 * P have known relationships. For example
1846 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1848 * :lower: & :print: = :lower:
1849 * And similarly for classes that must be disjoint. For example,
1850 * since \s and \w can have no elements in common based on rules in
1851 * the POSIX standard,
1852 * \w & ^\S = nothing
1853 * Unfortunately, some vendor locales do not meet the Posix
1854 * standard, in particular almost everything by Microsoft.
1855 * The loop below just changes e.g., \w into \W and vice versa */
1857 regnode_charclass_posixl temp;
1858 int add = 1; /* To calculate the index of the complement */
1860 Zero(&temp, 1, regnode_charclass_posixl);
1861 ANYOF_POSIXL_ZERO(&temp);
1862 for (i = 0; i < ANYOF_MAX; i++) {
1864 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1865 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1867 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1868 ANYOF_POSIXL_SET(&temp, i + add);
1870 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1872 ANYOF_POSIXL_AND(&temp, ssc);
1874 } /* else ssc already has no posixes */
1875 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1876 in its initial state */
1877 else if (! is_ANYOF_SYNTHETIC(and_with)
1878 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1880 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1881 * copy it over 'ssc' */
1882 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1883 if (is_ANYOF_SYNTHETIC(and_with)) {
1884 StructCopy(and_with, ssc, regnode_ssc);
1887 ssc->invlist = anded_cp_list;
1888 ANYOF_POSIXL_ZERO(ssc);
1889 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1890 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1894 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1895 || (and_with_flags & ANYOF_MATCHES_POSIXL))
1897 /* One or the other of P1, P2 is non-empty. */
1898 if (and_with_flags & ANYOF_MATCHES_POSIXL) {
1899 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1901 ssc_union(ssc, anded_cp_list, FALSE);
1903 else { /* P1 = P2 = empty */
1904 ssc_intersection(ssc, anded_cp_list, FALSE);
1910 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1911 const regnode_charclass *or_with)
1913 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1914 * another SSC or a regular ANYOF class. Can create false positives if
1915 * 'or_with' is to be inverted. */
1919 U8 or_with_flags = inRANGE(OP(or_with), ANYOFH, ANYOFHr)
1921 : ANYOF_FLAGS(or_with);
1923 PERL_ARGS_ASSERT_SSC_OR;
1925 assert(is_ANYOF_SYNTHETIC(ssc));
1927 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1928 * the code point inversion list and just the relevant flags */
1929 if (is_ANYOF_SYNTHETIC(or_with)) {
1930 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1931 ored_flags = or_with_flags;
1934 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1935 ored_flags = or_with_flags & ANYOF_COMMON_FLAGS;
1936 if (OP(or_with) != ANYOFD) {
1939 & ( ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
1940 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP);
1941 if (ANYOFL_UTF8_LOCALE_REQD(or_with_flags)) {
1943 ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
1948 ANYOF_FLAGS(ssc) |= ored_flags;
1950 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1951 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1952 * 'or_with' may be inverted. When not inverted, we have the simple
1953 * situation of computing:
1954 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1955 * If P1|P2 yields a situation with both a class and its complement are
1956 * set, like having both \w and \W, this matches all code points, and we
1957 * can delete these from the P component of the ssc going forward. XXX We
1958 * might be able to delete all the P components, but I (khw) am not certain
1959 * about this, and it is better to be safe.
1962 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1963 * <= (C1 | P1) | ~C2
1964 * <= (C1 | ~C2) | P1
1965 * (which results in actually simpler code than the non-inverted case)
1968 if ((or_with_flags & ANYOF_INVERT)
1969 && ! is_ANYOF_SYNTHETIC(or_with))
1971 /* We ignore P2, leaving P1 going forward */
1972 } /* else Not inverted */
1973 else if (or_with_flags & ANYOF_MATCHES_POSIXL) {
1974 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1975 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1977 for (i = 0; i < ANYOF_MAX; i += 2) {
1978 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1980 ssc_match_all_cp(ssc);
1981 ANYOF_POSIXL_CLEAR(ssc, i);
1982 ANYOF_POSIXL_CLEAR(ssc, i+1);
1990 FALSE /* Already has been inverted */
1994 PERL_STATIC_INLINE void
1995 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1997 PERL_ARGS_ASSERT_SSC_UNION;
1999 assert(is_ANYOF_SYNTHETIC(ssc));
2001 _invlist_union_maybe_complement_2nd(ssc->invlist,
2007 PERL_STATIC_INLINE void
2008 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
2010 const bool invert2nd)
2012 PERL_ARGS_ASSERT_SSC_INTERSECTION;
2014 assert(is_ANYOF_SYNTHETIC(ssc));
2016 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
2022 PERL_STATIC_INLINE void
2023 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
2025 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
2027 assert(is_ANYOF_SYNTHETIC(ssc));
2029 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
2032 PERL_STATIC_INLINE void
2033 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
2035 /* AND just the single code point 'cp' into the SSC 'ssc' */
2037 SV* cp_list = _new_invlist(2);
2039 PERL_ARGS_ASSERT_SSC_CP_AND;
2041 assert(is_ANYOF_SYNTHETIC(ssc));
2043 cp_list = add_cp_to_invlist(cp_list, cp);
2044 ssc_intersection(ssc, cp_list,
2045 FALSE /* Not inverted */
2047 SvREFCNT_dec_NN(cp_list);
2050 PERL_STATIC_INLINE void
2051 S_ssc_clear_locale(regnode_ssc *ssc)
2053 /* Set the SSC 'ssc' to not match any locale things */
2054 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
2056 assert(is_ANYOF_SYNTHETIC(ssc));
2058 ANYOF_POSIXL_ZERO(ssc);
2059 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
2062 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
2065 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
2067 /* The synthetic start class is used to hopefully quickly winnow down
2068 * places where a pattern could start a match in the target string. If it
2069 * doesn't really narrow things down that much, there isn't much point to
2070 * having the overhead of using it. This function uses some very crude
2071 * heuristics to decide if to use the ssc or not.
2073 * It returns TRUE if 'ssc' rules out more than half what it considers to
2074 * be the "likely" possible matches, but of course it doesn't know what the
2075 * actual things being matched are going to be; these are only guesses
2077 * For /l matches, it assumes that the only likely matches are going to be
2078 * in the 0-255 range, uniformly distributed, so half of that is 127
2079 * For /a and /d matches, it assumes that the likely matches will be just
2080 * the ASCII range, so half of that is 63
2081 * For /u and there isn't anything matching above the Latin1 range, it
2082 * assumes that that is the only range likely to be matched, and uses
2083 * half that as the cut-off: 127. If anything matches above Latin1,
2084 * it assumes that all of Unicode could match (uniformly), except for
2085 * non-Unicode code points and things in the General Category "Other"
2086 * (unassigned, private use, surrogates, controls and formats). This
2087 * is a much large number. */
2089 U32 count = 0; /* Running total of number of code points matched by
2091 UV start, end; /* Start and end points of current range in inversion
2092 XXX outdated. UTF-8 locales are common, what about invert? list */
2093 const U32 max_code_points = (LOC)
2095 : (( ! UNI_SEMANTICS
2096 || invlist_highest(ssc->invlist) < 256)
2099 const U32 max_match = max_code_points / 2;
2101 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
2103 invlist_iterinit(ssc->invlist);
2104 while (invlist_iternext(ssc->invlist, &start, &end)) {
2105 if (start >= max_code_points) {
2108 end = MIN(end, max_code_points - 1);
2109 count += end - start + 1;
2110 if (count >= max_match) {
2111 invlist_iterfinish(ssc->invlist);
2121 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
2123 /* The inversion list in the SSC is marked mortal; now we need a more
2124 * permanent copy, which is stored the same way that is done in a regular
2125 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
2128 SV* invlist = invlist_clone(ssc->invlist, NULL);
2130 PERL_ARGS_ASSERT_SSC_FINALIZE;
2132 assert(is_ANYOF_SYNTHETIC(ssc));
2134 /* The code in this file assumes that all but these flags aren't relevant
2135 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
2136 * by the time we reach here */
2137 assert(! (ANYOF_FLAGS(ssc)
2138 & ~( ANYOF_COMMON_FLAGS
2139 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER
2140 |ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)));
2142 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
2144 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL);
2146 /* Make sure is clone-safe */
2147 ssc->invlist = NULL;
2149 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2150 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2151 OP(ssc) = ANYOFPOSIXL;
2153 else if (RExC_contains_locale) {
2157 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2160 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2161 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2162 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2163 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2164 ? (TRIE_LIST_CUR( idx ) - 1) \
2170 dump_trie(trie,widecharmap,revcharmap)
2171 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2172 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2174 These routines dump out a trie in a somewhat readable format.
2175 The _interim_ variants are used for debugging the interim
2176 tables that are used to generate the final compressed
2177 representation which is what dump_trie expects.
2179 Part of the reason for their existence is to provide a form
2180 of documentation as to how the different representations function.
2185 Dumps the final compressed table form of the trie to Perl_debug_log.
2186 Used for debugging make_trie().
2190 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2191 AV *revcharmap, U32 depth)
2194 SV *sv=sv_newmortal();
2195 int colwidth= widecharmap ? 6 : 4;
2197 GET_RE_DEBUG_FLAGS_DECL;
2199 PERL_ARGS_ASSERT_DUMP_TRIE;
2201 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2202 depth+1, "Match","Base","Ofs" );
2204 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2205 SV ** const tmp = av_fetch( revcharmap, state, 0);
2207 Perl_re_printf( aTHX_ "%*s",
2209 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2210 PL_colors[0], PL_colors[1],
2211 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2212 PERL_PV_ESCAPE_FIRSTCHAR
2217 Perl_re_printf( aTHX_ "\n");
2218 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2220 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2221 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2222 Perl_re_printf( aTHX_ "\n");
2224 for( state = 1 ; state < trie->statecount ; state++ ) {
2225 const U32 base = trie->states[ state ].trans.base;
2227 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2229 if ( trie->states[ state ].wordnum ) {
2230 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2232 Perl_re_printf( aTHX_ "%6s", "" );
2235 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2240 while( ( base + ofs < trie->uniquecharcount ) ||
2241 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2242 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2246 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2248 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2249 if ( ( base + ofs >= trie->uniquecharcount )
2250 && ( base + ofs - trie->uniquecharcount
2252 && trie->trans[ base + ofs
2253 - trie->uniquecharcount ].check == state )
2255 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2256 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2259 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2263 Perl_re_printf( aTHX_ "]");
2266 Perl_re_printf( aTHX_ "\n" );
2268 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2270 for (word=1; word <= trie->wordcount; word++) {
2271 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2272 (int)word, (int)(trie->wordinfo[word].prev),
2273 (int)(trie->wordinfo[word].len));
2275 Perl_re_printf( aTHX_ "\n" );
2278 Dumps a fully constructed but uncompressed trie in list form.
2279 List tries normally only are used for construction when the number of
2280 possible chars (trie->uniquecharcount) is very high.
2281 Used for debugging make_trie().
2284 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2285 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2289 SV *sv=sv_newmortal();
2290 int colwidth= widecharmap ? 6 : 4;
2291 GET_RE_DEBUG_FLAGS_DECL;
2293 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2295 /* print out the table precompression. */
2296 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2298 Perl_re_indentf( aTHX_ "%s",
2299 depth+1, "------:-----+-----------------\n" );
2301 for( state=1 ; state < next_alloc ; state ++ ) {
2304 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2305 depth+1, (UV)state );
2306 if ( ! trie->states[ state ].wordnum ) {
2307 Perl_re_printf( aTHX_ "%5s| ","");
2309 Perl_re_printf( aTHX_ "W%4x| ",
2310 trie->states[ state ].wordnum
2313 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2314 SV ** const tmp = av_fetch( revcharmap,
2315 TRIE_LIST_ITEM(state, charid).forid, 0);
2317 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2319 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2321 PL_colors[0], PL_colors[1],
2322 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2323 | PERL_PV_ESCAPE_FIRSTCHAR
2325 TRIE_LIST_ITEM(state, charid).forid,
2326 (UV)TRIE_LIST_ITEM(state, charid).newstate
2329 Perl_re_printf( aTHX_ "\n%*s| ",
2330 (int)((depth * 2) + 14), "");
2333 Perl_re_printf( aTHX_ "\n");
2338 Dumps a fully constructed but uncompressed trie in table form.
2339 This is the normal DFA style state transition table, with a few
2340 twists to facilitate compression later.
2341 Used for debugging make_trie().
2344 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2345 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2350 SV *sv=sv_newmortal();
2351 int colwidth= widecharmap ? 6 : 4;
2352 GET_RE_DEBUG_FLAGS_DECL;
2354 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2357 print out the table precompression so that we can do a visual check
2358 that they are identical.
2361 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2363 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2364 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2366 Perl_re_printf( aTHX_ "%*s",
2368 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2369 PL_colors[0], PL_colors[1],
2370 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2371 PERL_PV_ESCAPE_FIRSTCHAR
2377 Perl_re_printf( aTHX_ "\n");
2378 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2380 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2381 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2384 Perl_re_printf( aTHX_ "\n" );
2386 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2388 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2390 (UV)TRIE_NODENUM( state ) );
2392 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2393 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2395 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2397 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2399 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2400 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2401 (UV)trie->trans[ state ].check );
2403 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2404 (UV)trie->trans[ state ].check,
2405 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2413 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2414 startbranch: the first branch in the whole branch sequence
2415 first : start branch of sequence of branch-exact nodes.
2416 May be the same as startbranch
2417 last : Thing following the last branch.
2418 May be the same as tail.
2419 tail : item following the branch sequence
2420 count : words in the sequence
2421 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2422 depth : indent depth
2424 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2426 A trie is an N'ary tree where the branches are determined by digital
2427 decomposition of the key. IE, at the root node you look up the 1st character and
2428 follow that branch repeat until you find the end of the branches. Nodes can be
2429 marked as "accepting" meaning they represent a complete word. Eg:
2433 would convert into the following structure. Numbers represent states, letters
2434 following numbers represent valid transitions on the letter from that state, if
2435 the number is in square brackets it represents an accepting state, otherwise it
2436 will be in parenthesis.
2438 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2442 (1) +-i->(6)-+-s->[7]
2444 +-s->(3)-+-h->(4)-+-e->[5]
2446 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2448 This shows that when matching against the string 'hers' we will begin at state 1
2449 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2450 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2451 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2452 single traverse. We store a mapping from accepting to state to which word was
2453 matched, and then when we have multiple possibilities we try to complete the
2454 rest of the regex in the order in which they occurred in the alternation.
2456 The only prior NFA like behaviour that would be changed by the TRIE support is
2457 the silent ignoring of duplicate alternations which are of the form:
2459 / (DUPE|DUPE) X? (?{ ... }) Y /x
2461 Thus EVAL blocks following a trie may be called a different number of times with
2462 and without the optimisation. With the optimisations dupes will be silently
2463 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2464 the following demonstrates:
2466 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2468 which prints out 'word' three times, but
2470 'words'=~/(word|word|word)(?{ print $1 })S/
2472 which doesnt print it out at all. This is due to other optimisations kicking in.
2474 Example of what happens on a structural level:
2476 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2478 1: CURLYM[1] {1,32767}(18)
2489 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2490 and should turn into:
2492 1: CURLYM[1] {1,32767}(18)
2494 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2502 Cases where tail != last would be like /(?foo|bar)baz/:
2512 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2513 and would end up looking like:
2516 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2523 d = uvchr_to_utf8_flags(d, uv, 0);
2525 is the recommended Unicode-aware way of saying
2530 #define TRIE_STORE_REVCHAR(val) \
2533 SV *zlopp = newSV(UTF8_MAXBYTES); \
2534 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2535 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2537 SvCUR_set(zlopp, kapow - flrbbbbb); \
2540 av_push(revcharmap, zlopp); \
2542 char ooooff = (char)val; \
2543 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2547 /* This gets the next character from the input, folding it if not already
2549 #define TRIE_READ_CHAR STMT_START { \
2552 /* if it is UTF then it is either already folded, or does not need \
2554 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2556 else if (folder == PL_fold_latin1) { \
2557 /* This folder implies Unicode rules, which in the range expressible \
2558 * by not UTF is the lower case, with the two exceptions, one of \
2559 * which should have been taken care of before calling this */ \
2560 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2561 uvc = toLOWER_L1(*uc); \
2562 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2565 /* raw data, will be folded later if needed */ \
2573 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2574 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2575 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2576 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2577 TRIE_LIST_LEN( state ) = ging; \
2579 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2580 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2581 TRIE_LIST_CUR( state )++; \
2584 #define TRIE_LIST_NEW(state) STMT_START { \
2585 Newx( trie->states[ state ].trans.list, \
2586 4, reg_trie_trans_le ); \
2587 TRIE_LIST_CUR( state ) = 1; \
2588 TRIE_LIST_LEN( state ) = 4; \
2591 #define TRIE_HANDLE_WORD(state) STMT_START { \
2592 U16 dupe= trie->states[ state ].wordnum; \
2593 regnode * const noper_next = regnext( noper ); \
2596 /* store the word for dumping */ \
2598 if (OP(noper) != NOTHING) \
2599 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2601 tmp = newSVpvn_utf8( "", 0, UTF ); \
2602 av_push( trie_words, tmp ); \
2606 trie->wordinfo[curword].prev = 0; \
2607 trie->wordinfo[curword].len = wordlen; \
2608 trie->wordinfo[curword].accept = state; \
2610 if ( noper_next < tail ) { \
2612 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2614 trie->jump[curword] = (U16)(noper_next - convert); \
2616 jumper = noper_next; \
2618 nextbranch= regnext(cur); \
2622 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2623 /* chain, so that when the bits of chain are later */\
2624 /* linked together, the dups appear in the chain */\
2625 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2626 trie->wordinfo[dupe].prev = curword; \
2628 /* we haven't inserted this word yet. */ \
2629 trie->states[ state ].wordnum = curword; \
2634 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2635 ( ( base + charid >= ucharcount \
2636 && base + charid < ubound \
2637 && state == trie->trans[ base - ucharcount + charid ].check \
2638 && trie->trans[ base - ucharcount + charid ].next ) \
2639 ? trie->trans[ base - ucharcount + charid ].next \
2640 : ( state==1 ? special : 0 ) \
2643 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2645 TRIE_BITMAP_SET(trie, uvc); \
2646 /* store the folded codepoint */ \
2648 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2651 /* store first byte of utf8 representation of */ \
2652 /* variant codepoints */ \
2653 if (! UVCHR_IS_INVARIANT(uvc)) { \
2654 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2659 #define MADE_JUMP_TRIE 2
2660 #define MADE_EXACT_TRIE 4
2663 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2664 regnode *first, regnode *last, regnode *tail,
2665 U32 word_count, U32 flags, U32 depth)
2667 /* first pass, loop through and scan words */
2668 reg_trie_data *trie;
2669 HV *widecharmap = NULL;
2670 AV *revcharmap = newAV();
2676 regnode *jumper = NULL;
2677 regnode *nextbranch = NULL;
2678 regnode *convert = NULL;
2679 U32 *prev_states; /* temp array mapping each state to previous one */
2680 /* we just use folder as a flag in utf8 */
2681 const U8 * folder = NULL;
2683 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2684 * which stands for one trie structure, one hash, optionally followed
2687 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2688 AV *trie_words = NULL;
2689 /* along with revcharmap, this only used during construction but both are
2690 * useful during debugging so we store them in the struct when debugging.
2693 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2694 STRLEN trie_charcount=0;
2696 SV *re_trie_maxbuff;
2697 GET_RE_DEBUG_FLAGS_DECL;
2699 PERL_ARGS_ASSERT_MAKE_TRIE;
2701 PERL_UNUSED_ARG(depth);
2705 case EXACT: case EXACT_REQ8: case EXACTL: break;
2709 case EXACTFLU8: folder = PL_fold_latin1; break;
2710 case EXACTF: folder = PL_fold; break;
2711 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2714 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2716 trie->startstate = 1;
2717 trie->wordcount = word_count;
2718 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2719 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2720 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2721 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2722 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2723 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2726 trie_words = newAV();
2729 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2730 assert(re_trie_maxbuff);
2731 if (!SvIOK(re_trie_maxbuff)) {
2732 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2734 DEBUG_TRIE_COMPILE_r({
2735 Perl_re_indentf( aTHX_
2736 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2738 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2739 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2742 /* Find the node we are going to overwrite */
2743 if ( first == startbranch && OP( last ) != BRANCH ) {
2744 /* whole branch chain */
2747 /* branch sub-chain */
2748 convert = NEXTOPER( first );
2751 /* -- First loop and Setup --
2753 We first traverse the branches and scan each word to determine if it
2754 contains widechars, and how many unique chars there are, this is
2755 important as we have to build a table with at least as many columns as we
2758 We use an array of integers to represent the character codes 0..255
2759 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2760 the native representation of the character value as the key and IV's for
2763 *TODO* If we keep track of how many times each character is used we can
2764 remap the columns so that the table compression later on is more
2765 efficient in terms of memory by ensuring the most common value is in the
2766 middle and the least common are on the outside. IMO this would be better
2767 than a most to least common mapping as theres a decent chance the most
2768 common letter will share a node with the least common, meaning the node
2769 will not be compressible. With a middle is most common approach the worst
2770 case is when we have the least common nodes twice.
2774 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2775 regnode *noper = NEXTOPER( cur );
2779 U32 wordlen = 0; /* required init */
2780 STRLEN minchars = 0;
2781 STRLEN maxchars = 0;
2782 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2785 if (OP(noper) == NOTHING) {
2786 /* skip past a NOTHING at the start of an alternation
2787 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2789 regnode *noper_next= regnext(noper);
2790 if (noper_next < tail)
2795 && ( OP(noper) == flags
2796 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2797 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2798 || OP(noper) == EXACTFUP))))
2800 uc= (U8*)STRING(noper);
2801 e= uc + STR_LEN(noper);
2808 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2809 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2810 regardless of encoding */
2811 if (OP( noper ) == EXACTFUP) {
2812 /* false positives are ok, so just set this */
2813 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2817 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2819 TRIE_CHARCOUNT(trie)++;
2822 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2823 * is in effect. Under /i, this character can match itself, or
2824 * anything that folds to it. If not under /i, it can match just
2825 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2826 * all fold to k, and all are single characters. But some folds
2827 * expand to more than one character, so for example LATIN SMALL
2828 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2829 * the string beginning at 'uc' is 'ffi', it could be matched by
2830 * three characters, or just by the one ligature character. (It
2831 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2832 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2833 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2834 * match.) The trie needs to know the minimum and maximum number
2835 * of characters that could match so that it can use size alone to
2836 * quickly reject many match attempts. The max is simple: it is
2837 * the number of folded characters in this branch (since a fold is
2838 * never shorter than what folds to it. */
2842 /* And the min is equal to the max if not under /i (indicated by
2843 * 'folder' being NULL), or there are no multi-character folds. If
2844 * there is a multi-character fold, the min is incremented just
2845 * once, for the character that folds to the sequence. Each
2846 * character in the sequence needs to be added to the list below of
2847 * characters in the trie, but we count only the first towards the
2848 * min number of characters needed. This is done through the
2849 * variable 'foldlen', which is returned by the macros that look
2850 * for these sequences as the number of bytes the sequence
2851 * occupies. Each time through the loop, we decrement 'foldlen' by
2852 * how many bytes the current char occupies. Only when it reaches
2853 * 0 do we increment 'minchars' or look for another multi-character
2855 if (folder == NULL) {
2858 else if (foldlen > 0) {
2859 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2864 /* See if *uc is the beginning of a multi-character fold. If
2865 * so, we decrement the length remaining to look at, to account
2866 * for the current character this iteration. (We can use 'uc'
2867 * instead of the fold returned by TRIE_READ_CHAR because for
2868 * non-UTF, the latin1_safe macro is smart enough to account
2869 * for all the unfolded characters, and because for UTF, the
2870 * string will already have been folded earlier in the
2871 * compilation process */
2873 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2874 foldlen -= UTF8SKIP(uc);
2877 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2882 /* The current character (and any potential folds) should be added
2883 * to the possible matching characters for this position in this
2887 U8 folded= folder[ (U8) uvc ];
2888 if ( !trie->charmap[ folded ] ) {
2889 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2890 TRIE_STORE_REVCHAR( folded );
2893 if ( !trie->charmap[ uvc ] ) {
2894 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2895 TRIE_STORE_REVCHAR( uvc );
2898 /* store the codepoint in the bitmap, and its folded
2900 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2901 set_bit = 0; /* We've done our bit :-) */
2905 /* XXX We could come up with the list of code points that fold
2906 * to this using PL_utf8_foldclosures, except not for
2907 * multi-char folds, as there may be multiple combinations
2908 * there that could work, which needs to wait until runtime to
2909 * resolve (The comment about LIGATURE FFI above is such an
2914 widecharmap = newHV();
2916 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2919 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2921 if ( !SvTRUE( *svpp ) ) {
2922 sv_setiv( *svpp, ++trie->uniquecharcount );
2923 TRIE_STORE_REVCHAR(uvc);
2926 } /* end loop through characters in this branch of the trie */
2928 /* We take the min and max for this branch and combine to find the min
2929 * and max for all branches processed so far */
2930 if( cur == first ) {
2931 trie->minlen = minchars;
2932 trie->maxlen = maxchars;
2933 } else if (minchars < trie->minlen) {
2934 trie->minlen = minchars;
2935 } else if (maxchars > trie->maxlen) {
2936 trie->maxlen = maxchars;
2938 } /* end first pass */
2939 DEBUG_TRIE_COMPILE_r(
2940 Perl_re_indentf( aTHX_
2941 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2943 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2944 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2945 (int)trie->minlen, (int)trie->maxlen )
2949 We now know what we are dealing with in terms of unique chars and
2950 string sizes so we can calculate how much memory a naive
2951 representation using a flat table will take. If it's over a reasonable
2952 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2953 conservative but potentially much slower representation using an array
2956 At the end we convert both representations into the same compressed
2957 form that will be used in regexec.c for matching with. The latter
2958 is a form that cannot be used to construct with but has memory
2959 properties similar to the list form and access properties similar
2960 to the table form making it both suitable for fast searches and
2961 small enough that its feasable to store for the duration of a program.
2963 See the comment in the code where the compressed table is produced
2964 inplace from the flat tabe representation for an explanation of how
2965 the compression works.
2970 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2973 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2974 > SvIV(re_trie_maxbuff) )
2977 Second Pass -- Array Of Lists Representation
2979 Each state will be represented by a list of charid:state records
2980 (reg_trie_trans_le) the first such element holds the CUR and LEN
2981 points of the allocated array. (See defines above).
2983 We build the initial structure using the lists, and then convert
2984 it into the compressed table form which allows faster lookups
2985 (but cant be modified once converted).
2988 STRLEN transcount = 1;
2990 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2993 trie->states = (reg_trie_state *)
2994 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2995 sizeof(reg_trie_state) );
2999 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3001 regnode *noper = NEXTOPER( cur );
3002 U32 state = 1; /* required init */
3003 U16 charid = 0; /* sanity init */
3004 U32 wordlen = 0; /* required init */
3006 if (OP(noper) == NOTHING) {
3007 regnode *noper_next= regnext(noper);
3008 if (noper_next < tail)
3013 && ( OP(noper) == flags
3014 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3015 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3016 || OP(noper) == EXACTFUP))))
3018 const U8 *uc= (U8*)STRING(noper);
3019 const U8 *e= uc + STR_LEN(noper);
3021 for ( ; uc < e ; uc += len ) {
3026 charid = trie->charmap[ uvc ];
3028 SV** const svpp = hv_fetch( widecharmap,
3035 charid=(U16)SvIV( *svpp );
3038 /* charid is now 0 if we dont know the char read, or
3039 * nonzero if we do */
3046 if ( !trie->states[ state ].trans.list ) {
3047 TRIE_LIST_NEW( state );
3050 check <= TRIE_LIST_USED( state );
3053 if ( TRIE_LIST_ITEM( state, check ).forid
3056 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3061 newstate = next_alloc++;
3062 prev_states[newstate] = state;
3063 TRIE_LIST_PUSH( state, charid, newstate );
3068 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3072 TRIE_HANDLE_WORD(state);
3074 } /* end second pass */
3076 /* next alloc is the NEXT state to be allocated */
3077 trie->statecount = next_alloc;
3078 trie->states = (reg_trie_state *)
3079 PerlMemShared_realloc( trie->states,
3081 * sizeof(reg_trie_state) );
3083 /* and now dump it out before we compress it */
3084 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3085 revcharmap, next_alloc,
3089 trie->trans = (reg_trie_trans *)
3090 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3097 for( state=1 ; state < next_alloc ; state ++ ) {
3101 DEBUG_TRIE_COMPILE_MORE_r(
3102 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3106 if (trie->states[state].trans.list) {
3107 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3111 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3112 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3113 if ( forid < minid ) {
3115 } else if ( forid > maxid ) {
3119 if ( transcount < tp + maxid - minid + 1) {
3121 trie->trans = (reg_trie_trans *)
3122 PerlMemShared_realloc( trie->trans,
3124 * sizeof(reg_trie_trans) );
3125 Zero( trie->trans + (transcount / 2),
3129 base = trie->uniquecharcount + tp - minid;
3130 if ( maxid == minid ) {
3132 for ( ; zp < tp ; zp++ ) {
3133 if ( ! trie->trans[ zp ].next ) {
3134 base = trie->uniquecharcount + zp - minid;
3135 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3137 trie->trans[ zp ].check = state;
3143 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3145 trie->trans[ tp ].check = state;
3150 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3151 const U32 tid = base
3152 - trie->uniquecharcount
3153 + TRIE_LIST_ITEM( state, idx ).forid;
3154 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3156 trie->trans[ tid ].check = state;
3158 tp += ( maxid - minid + 1 );
3160 Safefree(trie->states[ state ].trans.list);
3163 DEBUG_TRIE_COMPILE_MORE_r(
3164 Perl_re_printf( aTHX_ " base: %d\n",base);
3167 trie->states[ state ].trans.base=base;
3169 trie->lasttrans = tp + 1;
3173 Second Pass -- Flat Table Representation.
3175 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3176 each. We know that we will need Charcount+1 trans at most to store
3177 the data (one row per char at worst case) So we preallocate both
3178 structures assuming worst case.
3180 We then construct the trie using only the .next slots of the entry
3183 We use the .check field of the first entry of the node temporarily
3184 to make compression both faster and easier by keeping track of how
3185 many non zero fields are in the node.
3187 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3190 There are two terms at use here: state as a TRIE_NODEIDX() which is
3191 a number representing the first entry of the node, and state as a
3192 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3193 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3194 if there are 2 entrys per node. eg:
3202 The table is internally in the right hand, idx form. However as we
3203 also have to deal with the states array which is indexed by nodenum
3204 we have to use TRIE_NODENUM() to convert.
3207 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3210 trie->trans = (reg_trie_trans *)
3211 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3212 * trie->uniquecharcount + 1,
3213 sizeof(reg_trie_trans) );
3214 trie->states = (reg_trie_state *)
3215 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3216 sizeof(reg_trie_state) );
3217 next_alloc = trie->uniquecharcount + 1;
3220 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3222 regnode *noper = NEXTOPER( cur );
3224 U32 state = 1; /* required init */
3226 U16 charid = 0; /* sanity init */
3227 U32 accept_state = 0; /* sanity init */
3229 U32 wordlen = 0; /* required init */
3231 if (OP(noper) == NOTHING) {
3232 regnode *noper_next= regnext(noper);
3233 if (noper_next < tail)
3238 && ( OP(noper) == flags
3239 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3240 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3241 || OP(noper) == EXACTFUP))))
3243 const U8 *uc= (U8*)STRING(noper);
3244 const U8 *e= uc + STR_LEN(noper);
3246 for ( ; uc < e ; uc += len ) {
3251 charid = trie->charmap[ uvc ];
3253 SV* const * const svpp = hv_fetch( widecharmap,
3257 charid = svpp ? (U16)SvIV(*svpp) : 0;
3261 if ( !trie->trans[ state + charid ].next ) {
3262 trie->trans[ state + charid ].next = next_alloc;
3263 trie->trans[ state ].check++;
3264 prev_states[TRIE_NODENUM(next_alloc)]
3265 = TRIE_NODENUM(state);
3266 next_alloc += trie->uniquecharcount;
3268 state = trie->trans[ state + charid ].next;
3270 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3272 /* charid is now 0 if we dont know the char read, or
3273 * nonzero if we do */
3276 accept_state = TRIE_NODENUM( state );
3277 TRIE_HANDLE_WORD(accept_state);
3279 } /* end second pass */
3281 /* and now dump it out before we compress it */
3282 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3284 next_alloc, depth+1));
3288 * Inplace compress the table.*
3290 For sparse data sets the table constructed by the trie algorithm will
3291 be mostly 0/FAIL transitions or to put it another way mostly empty.
3292 (Note that leaf nodes will not contain any transitions.)
3294 This algorithm compresses the tables by eliminating most such
3295 transitions, at the cost of a modest bit of extra work during lookup:
3297 - Each states[] entry contains a .base field which indicates the
3298 index in the state[] array wheres its transition data is stored.
3300 - If .base is 0 there are no valid transitions from that node.
3302 - If .base is nonzero then charid is added to it to find an entry in
3305 -If trans[states[state].base+charid].check!=state then the
3306 transition is taken to be a 0/Fail transition. Thus if there are fail
3307 transitions at the front of the node then the .base offset will point
3308 somewhere inside the previous nodes data (or maybe even into a node
3309 even earlier), but the .check field determines if the transition is
3313 The following process inplace converts the table to the compressed
3314 table: We first do not compress the root node 1,and mark all its
3315 .check pointers as 1 and set its .base pointer as 1 as well. This
3316 allows us to do a DFA construction from the compressed table later,
3317 and ensures that any .base pointers we calculate later are greater
3320 - We set 'pos' to indicate the first entry of the second node.
3322 - We then iterate over the columns of the node, finding the first and
3323 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3324 and set the .check pointers accordingly, and advance pos
3325 appropriately and repreat for the next node. Note that when we copy
3326 the next pointers we have to convert them from the original
3327 NODEIDX form to NODENUM form as the former is not valid post
3330 - If a node has no transitions used we mark its base as 0 and do not
3331 advance the pos pointer.
3333 - If a node only has one transition we use a second pointer into the
3334 structure to fill in allocated fail transitions from other states.
3335 This pointer is independent of the main pointer and scans forward
3336 looking for null transitions that are allocated to a state. When it
3337 finds one it writes the single transition into the "hole". If the
3338 pointer doesnt find one the single transition is appended as normal.
3340 - Once compressed we can Renew/realloc the structures to release the
3343 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3344 specifically Fig 3.47 and the associated pseudocode.
3348 const U32 laststate = TRIE_NODENUM( next_alloc );
3351 trie->statecount = laststate;
3353 for ( state = 1 ; state < laststate ; state++ ) {
3355 const U32 stateidx = TRIE_NODEIDX( state );
3356 const U32 o_used = trie->trans[ stateidx ].check;
3357 U32 used = trie->trans[ stateidx ].check;
3358 trie->trans[ stateidx ].check = 0;
3361 used && charid < trie->uniquecharcount;
3364 if ( flag || trie->trans[ stateidx + charid ].next ) {
3365 if ( trie->trans[ stateidx + charid ].next ) {
3367 for ( ; zp < pos ; zp++ ) {
3368 if ( ! trie->trans[ zp ].next ) {
3372 trie->states[ state ].trans.base
3374 + trie->uniquecharcount
3376 trie->trans[ zp ].next
3377 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3379 trie->trans[ zp ].check = state;
3380 if ( ++zp > pos ) pos = zp;
3387 trie->states[ state ].trans.base
3388 = pos + trie->uniquecharcount - charid ;
3390 trie->trans[ pos ].next
3391 = SAFE_TRIE_NODENUM(
3392 trie->trans[ stateidx + charid ].next );
3393 trie->trans[ pos ].check = state;
3398 trie->lasttrans = pos + 1;
3399 trie->states = (reg_trie_state *)
3400 PerlMemShared_realloc( trie->states, laststate
3401 * sizeof(reg_trie_state) );
3402 DEBUG_TRIE_COMPILE_MORE_r(
3403 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3405 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3409 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3412 } /* end table compress */
3414 DEBUG_TRIE_COMPILE_MORE_r(
3415 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3417 (UV)trie->statecount,
3418 (UV)trie->lasttrans)
3420 /* resize the trans array to remove unused space */
3421 trie->trans = (reg_trie_trans *)
3422 PerlMemShared_realloc( trie->trans, trie->lasttrans
3423 * sizeof(reg_trie_trans) );
3425 { /* Modify the program and insert the new TRIE node */
3426 U8 nodetype =(U8)(flags & 0xFF);
3430 regnode *optimize = NULL;
3431 #ifdef RE_TRACK_PATTERN_OFFSETS
3434 U32 mjd_nodelen = 0;
3435 #endif /* RE_TRACK_PATTERN_OFFSETS */
3436 #endif /* DEBUGGING */
3438 This means we convert either the first branch or the first Exact,
3439 depending on whether the thing following (in 'last') is a branch
3440 or not and whther first is the startbranch (ie is it a sub part of
3441 the alternation or is it the whole thing.)
3442 Assuming its a sub part we convert the EXACT otherwise we convert
3443 the whole branch sequence, including the first.
3445 /* Find the node we are going to overwrite */
3446 if ( first != startbranch || OP( last ) == BRANCH ) {
3447 /* branch sub-chain */
3448 NEXT_OFF( first ) = (U16)(last - first);
3449 #ifdef RE_TRACK_PATTERN_OFFSETS
3451 mjd_offset= Node_Offset((convert));
3452 mjd_nodelen= Node_Length((convert));
3455 /* whole branch chain */
3457 #ifdef RE_TRACK_PATTERN_OFFSETS
3460 const regnode *nop = NEXTOPER( convert );
3461 mjd_offset= Node_Offset((nop));
3462 mjd_nodelen= Node_Length((nop));
3466 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3468 (UV)mjd_offset, (UV)mjd_nodelen)
3471 /* But first we check to see if there is a common prefix we can
3472 split out as an EXACT and put in front of the TRIE node. */
3473 trie->startstate= 1;
3474 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3475 /* we want to find the first state that has more than
3476 * one transition, if that state is not the first state
3477 * then we have a common prefix which we can remove.
3480 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3482 I32 first_ofs = -1; /* keeps track of the ofs of the first
3483 transition, -1 means none */
3485 const U32 base = trie->states[ state ].trans.base;
3487 /* does this state terminate an alternation? */
3488 if ( trie->states[state].wordnum )
3491 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3492 if ( ( base + ofs >= trie->uniquecharcount ) &&
3493 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3494 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3496 if ( ++count > 1 ) {
3497 /* we have more than one transition */
3500 /* if this is the first state there is no common prefix
3501 * to extract, so we can exit */
3502 if ( state == 1 ) break;
3503 tmp = av_fetch( revcharmap, ofs, 0);
3504 ch = (U8*)SvPV_nolen_const( *tmp );
3506 /* if we are on count 2 then we need to initialize the
3507 * bitmap, and store the previous char if there was one
3510 /* clear the bitmap */
3511 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3513 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3516 if (first_ofs >= 0) {
3517 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3518 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3520 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3522 Perl_re_printf( aTHX_ "%s", (char*)ch)
3526 /* store the current firstchar in the bitmap */
3527 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3528 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3534 /* This state has only one transition, its transition is part
3535 * of a common prefix - we need to concatenate the char it
3536 * represents to what we have so far. */
3537 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3539 char *ch = SvPV( *tmp, len );
3541 SV *sv=sv_newmortal();
3542 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3544 (UV)state, (UV)first_ofs,
3545 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3546 PL_colors[0], PL_colors[1],
3547 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3548 PERL_PV_ESCAPE_FIRSTCHAR
3553 OP( convert ) = nodetype;
3554 str=STRING(convert);
3555 setSTR_LEN(convert, 0);
3557 setSTR_LEN(convert, STR_LEN(convert) + len);
3563 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3568 trie->prefixlen = (state-1);
3570 regnode *n = convert+NODE_SZ_STR(convert);
3571 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3572 trie->startstate = state;
3573 trie->minlen -= (state - 1);
3574 trie->maxlen -= (state - 1);
3576 /* At least the UNICOS C compiler choked on this
3577 * being argument to DEBUG_r(), so let's just have
3580 #ifdef PERL_EXT_RE_BUILD
3586 regnode *fix = convert;
3587 U32 word = trie->wordcount;
3588 #ifdef RE_TRACK_PATTERN_OFFSETS
3591 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3592 while( ++fix < n ) {
3593 Set_Node_Offset_Length(fix, 0, 0);
3596 SV ** const tmp = av_fetch( trie_words, word, 0 );
3598 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3599 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3601 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3609 NEXT_OFF(convert) = (U16)(tail - convert);
3610 DEBUG_r(optimize= n);
3616 if ( trie->maxlen ) {
3617 NEXT_OFF( convert ) = (U16)(tail - convert);
3618 ARG_SET( convert, data_slot );
3619 /* Store the offset to the first unabsorbed branch in
3620 jump[0], which is otherwise unused by the jump logic.
3621 We use this when dumping a trie and during optimisation. */
3623 trie->jump[0] = (U16)(nextbranch - convert);
3625 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3626 * and there is a bitmap
3627 * and the first "jump target" node we found leaves enough room
3628 * then convert the TRIE node into a TRIEC node, with the bitmap
3629 * embedded inline in the opcode - this is hypothetically faster.
3631 if ( !trie->states[trie->startstate].wordnum
3633 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3635 OP( convert ) = TRIEC;
3636 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3637 PerlMemShared_free(trie->bitmap);
3640 OP( convert ) = TRIE;
3642 /* store the type in the flags */
3643 convert->flags = nodetype;
3647 + regarglen[ OP( convert ) ];
3649 /* XXX We really should free up the resource in trie now,
3650 as we won't use them - (which resources?) dmq */
3652 /* needed for dumping*/
3653 DEBUG_r(if (optimize) {
3654 regnode *opt = convert;
3656 while ( ++opt < optimize) {
3657 Set_Node_Offset_Length(opt, 0, 0);
3660 Try to clean up some of the debris left after the
3663 while( optimize < jumper ) {
3664 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3665 OP( optimize ) = OPTIMIZED;
3666 Set_Node_Offset_Length(optimize, 0, 0);
3669 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3671 } /* end node insert */
3673 /* Finish populating the prev field of the wordinfo array. Walk back
3674 * from each accept state until we find another accept state, and if
3675 * so, point the first word's .prev field at the second word. If the
3676 * second already has a .prev field set, stop now. This will be the
3677 * case either if we've already processed that word's accept state,
3678 * or that state had multiple words, and the overspill words were
3679 * already linked up earlier.
3686 for (word=1; word <= trie->wordcount; word++) {
3688 if (trie->wordinfo[word].prev)
3690 state = trie->wordinfo[word].accept;
3692 state = prev_states[state];
3695 prev = trie->states[state].wordnum;
3699 trie->wordinfo[word].prev = prev;
3701 Safefree(prev_states);
3705 /* and now dump out the compressed format */
3706 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3708 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3710 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3711 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3713 SvREFCNT_dec_NN(revcharmap);
3717 : trie->startstate>1
3723 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3725 /* The Trie is constructed and compressed now so we can build a fail array if
3728 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3730 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3734 We find the fail state for each state in the trie, this state is the longest
3735 proper suffix of the current state's 'word' that is also a proper prefix of
3736 another word in our trie. State 1 represents the word '' and is thus the
3737 default fail state. This allows the DFA not to have to restart after its
3738 tried and failed a word at a given point, it simply continues as though it
3739 had been matching the other word in the first place.
3741 'abcdgu'=~/abcdefg|cdgu/
3742 When we get to 'd' we are still matching the first word, we would encounter
3743 'g' which would fail, which would bring us to the state representing 'd' in
3744 the second word where we would try 'g' and succeed, proceeding to match
3747 /* add a fail transition */
3748 const U32 trie_offset = ARG(source);
3749 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3751 const U32 ucharcount = trie->uniquecharcount;
3752 const U32 numstates = trie->statecount;
3753 const U32 ubound = trie->lasttrans + ucharcount;
3757 U32 base = trie->states[ 1 ].trans.base;
3760 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3762 GET_RE_DEBUG_FLAGS_DECL;
3764 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3765 PERL_UNUSED_CONTEXT;
3767 PERL_UNUSED_ARG(depth);
3770 if ( OP(source) == TRIE ) {
3771 struct regnode_1 *op = (struct regnode_1 *)
3772 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3773 StructCopy(source, op, struct regnode_1);
3774 stclass = (regnode *)op;
3776 struct regnode_charclass *op = (struct regnode_charclass *)
3777 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3778 StructCopy(source, op, struct regnode_charclass);
3779 stclass = (regnode *)op;
3781 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3783 ARG_SET( stclass, data_slot );
3784 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3785 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3786 aho->trie=trie_offset;
3787 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3788 Copy( trie->states, aho->states, numstates, reg_trie_state );
3789 Newx( q, numstates, U32);
3790 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3793 /* initialize fail[0..1] to be 1 so that we always have
3794 a valid final fail state */
3795 fail[ 0 ] = fail[ 1 ] = 1;
3797 for ( charid = 0; charid < ucharcount ; charid++ ) {
3798 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3800 q[ q_write ] = newstate;
3801 /* set to point at the root */
3802 fail[ q[ q_write++ ] ]=1;
3805 while ( q_read < q_write) {
3806 const U32 cur = q[ q_read++ % numstates ];
3807 base = trie->states[ cur ].trans.base;
3809 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3810 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3812 U32 fail_state = cur;
3815 fail_state = fail[ fail_state ];
3816 fail_base = aho->states[ fail_state ].trans.base;
3817 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3819 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3820 fail[ ch_state ] = fail_state;
3821 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3823 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3825 q[ q_write++ % numstates] = ch_state;
3829 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3830 when we fail in state 1, this allows us to use the
3831 charclass scan to find a valid start char. This is based on the principle
3832 that theres a good chance the string being searched contains lots of stuff
3833 that cant be a start char.
3835 fail[ 0 ] = fail[ 1 ] = 0;
3836 DEBUG_TRIE_COMPILE_r({
3837 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3838 depth, (UV)numstates
3840 for( q_read=1; q_read<numstates; q_read++ ) {
3841 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3843 Perl_re_printf( aTHX_ "\n");
3846 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3851 /* The below joins as many adjacent EXACTish nodes as possible into a single
3852 * one. The regop may be changed if the node(s) contain certain sequences that
3853 * require special handling. The joining is only done if:
3854 * 1) there is room in the current conglomerated node to entirely contain the
3856 * 2) they are compatible node types
3858 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3859 * these get optimized out
3861 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3862 * as possible, even if that means splitting an existing node so that its first
3863 * part is moved to the preceeding node. This would maximise the efficiency of
3864 * memEQ during matching.
3866 * If a node is to match under /i (folded), the number of characters it matches
3867 * can be different than its character length if it contains a multi-character
3868 * fold. *min_subtract is set to the total delta number of characters of the
3871 * And *unfolded_multi_char is set to indicate whether or not the node contains
3872 * an unfolded multi-char fold. This happens when it won't be known until
3873 * runtime whether the fold is valid or not; namely
3874 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3875 * target string being matched against turns out to be UTF-8 is that fold
3877 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3879 * (Multi-char folds whose components are all above the Latin1 range are not
3880 * run-time locale dependent, and have already been folded by the time this
3881 * function is called.)
3883 * This is as good a place as any to discuss the design of handling these
3884 * multi-character fold sequences. It's been wrong in Perl for a very long
3885 * time. There are three code points in Unicode whose multi-character folds
3886 * were long ago discovered to mess things up. The previous designs for
3887 * dealing with these involved assigning a special node for them. This
3888 * approach doesn't always work, as evidenced by this example:
3889 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3890 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3891 * would match just the \xDF, it won't be able to handle the case where a
3892 * successful match would have to cross the node's boundary. The new approach
3893 * that hopefully generally solves the problem generates an EXACTFUP node
3894 * that is "sss" in this case.
3896 * It turns out that there are problems with all multi-character folds, and not
3897 * just these three. Now the code is general, for all such cases. The
3898 * approach taken is:
3899 * 1) This routine examines each EXACTFish node that could contain multi-
3900 * character folded sequences. Since a single character can fold into
3901 * such a sequence, the minimum match length for this node is less than
3902 * the number of characters in the node. This routine returns in
3903 * *min_subtract how many characters to subtract from the the actual
3904 * length of the string to get a real minimum match length; it is 0 if
3905 * there are no multi-char foldeds. This delta is used by the caller to
3906 * adjust the min length of the match, and the delta between min and max,
3907 * so that the optimizer doesn't reject these possibilities based on size
3910 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3911 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3912 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3913 * EXACTFU nodes. The node type of such nodes is then changed to
3914 * EXACTFUP, indicating it is problematic, and needs careful handling.
3915 * (The procedures in step 1) above are sufficient to handle this case in
3916 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3917 * the only case where there is a possible fold length change in non-UTF-8
3918 * patterns. By reserving a special node type for problematic cases, the
3919 * far more common regular EXACTFU nodes can be processed faster.
3920 * regexec.c takes advantage of this.
3922 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3923 * problematic cases. These all only occur when the pattern is not
3924 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3925 * length change, it handles the situation where the string cannot be
3926 * entirely folded. The strings in an EXACTFish node are folded as much
3927 * as possible during compilation in regcomp.c. This saves effort in
3928 * regex matching. By using an EXACTFUP node when it is not possible to
3929 * fully fold at compile time, regexec.c can know that everything in an
3930 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3931 * case where folding in EXACTFU nodes can't be done at compile time is
3932 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3933 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3934 * handle two very different cases. Alternatively, there could have been
3935 * a node type where there are length changes, one for unfolded, and one
3936 * for both. If yet another special case needed to be created, the number
3937 * of required node types would have to go to 7. khw figures that even
3938 * though there are plenty of node types to spare, that the maintenance
3939 * cost wasn't worth the small speedup of doing it that way, especially
3940 * since he thinks the MICRO SIGN is rarely encountered in practice.
3942 * There are other cases where folding isn't done at compile time, but
3943 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3944 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3945 * changes. Some folds in EXACTF depend on if the runtime target string
3946 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3947 * when no fold in it depends on the UTF-8ness of the target string.)
3949 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3950 * validity of the fold won't be known until runtime, and so must remain
3951 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3952 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3953 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3954 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3955 * The reason this is a problem is that the optimizer part of regexec.c
3956 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3957 * that a character in the pattern corresponds to at most a single
3958 * character in the target string. (And I do mean character, and not byte
3959 * here, unlike other parts of the documentation that have never been
3960 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3961 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3962 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3963 * EXACTFL nodes, violate the assumption, and they are the only instances
3964 * where it is violated. I'm reluctant to try to change the assumption,
3965 * as the code involved is impenetrable to me (khw), so instead the code
3966 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3967 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3968 * boolean indicating whether or not the node contains such a fold. When
3969 * it is true, the caller sets a flag that later causes the optimizer in
3970 * this file to not set values for the floating and fixed string lengths,
3971 * and thus avoids the optimizer code in regexec.c that makes the invalid
3972 * assumption. Thus, there is no optimization based on string lengths for
3973 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3974 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3975 * assumption is wrong only in these cases is that all other non-UTF-8
3976 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3977 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3978 * EXACTF nodes because we don't know at compile time if it actually
3979 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3980 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3981 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3982 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3983 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3984 * string would require the pattern to be forced into UTF-8, the overhead
3985 * of which we want to avoid. Similarly the unfolded multi-char folds in
3986 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3989 * Similarly, the code that generates tries doesn't currently handle
3990 * not-already-folded multi-char folds, and it looks like a pain to change
3991 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3992 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3993 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3994 * using /iaa matching will be doing so almost entirely with ASCII
3995 * strings, so this should rarely be encountered in practice */
3997 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3998 if (PL_regkind[OP(scan)] == EXACT && OP(scan) != LEXACT \
3999 && OP(scan) != LEXACT_REQ8) \
4000 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
4003 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4004 UV *min_subtract, bool *unfolded_multi_char,
4005 U32 flags, regnode *val, U32 depth)
4007 /* Merge several consecutive EXACTish nodes into one. */
4009 regnode *n = regnext(scan);
4011 regnode *next = scan + NODE_SZ_STR(scan);
4015 regnode *stop = scan;
4016 GET_RE_DEBUG_FLAGS_DECL;
4018 PERL_UNUSED_ARG(depth);
4021 PERL_ARGS_ASSERT_JOIN_EXACT;
4022 #ifndef EXPERIMENTAL_INPLACESCAN
4023 PERL_UNUSED_ARG(flags);
4024 PERL_UNUSED_ARG(val);
4026 DEBUG_PEEP("join", scan, depth, 0);
4028 assert(PL_regkind[OP(scan)] == EXACT);
4030 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4031 * EXACT ones that are mergeable to the current one. */
4033 && ( PL_regkind[OP(n)] == NOTHING
4034 || (stringok && PL_regkind[OP(n)] == EXACT))
4036 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4039 if (OP(n) == TAIL || n > next)
4041 if (PL_regkind[OP(n)] == NOTHING) {
4042 DEBUG_PEEP("skip:", n, depth, 0);
4043 NEXT_OFF(scan) += NEXT_OFF(n);
4044 next = n + NODE_STEP_REGNODE;
4051 else if (stringok) {
4052 const unsigned int oldl = STR_LEN(scan);
4053 regnode * const nnext = regnext(n);
4055 /* XXX I (khw) kind of doubt that this works on platforms (should
4056 * Perl ever run on one) where U8_MAX is above 255 because of lots
4057 * of other assumptions */
4058 /* Don't join if the sum can't fit into a single node */
4059 if (oldl + STR_LEN(n) > U8_MAX)
4062 /* Joining something that requires UTF-8 with something that
4063 * doesn't, means the result requires UTF-8. */
4064 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4065 OP(scan) = EXACT_REQ8;
4067 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4068 ; /* join is compatible, no need to change OP */
4070 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4071 OP(scan) = EXACTFU_REQ8;
4073 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4074 ; /* join is compatible, no need to change OP */
4076 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4077 ; /* join is compatible, no need to change OP */
4079 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4081 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4082 * which can join with EXACTFU ones. We check for this case
4083 * here. These need to be resolved to either EXACTFU or
4084 * EXACTF at joining time. They have nothing in them that
4085 * would forbid them from being the more desirable EXACTFU
4086 * nodes except that they begin and/or end with a single [Ss].
4087 * The reason this is problematic is because they could be
4088 * joined in this loop with an adjacent node that ends and/or
4089 * begins with [Ss] which would then form the sequence 'ss',
4090 * which matches differently under /di than /ui, in which case
4091 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4092 * formed, the nodes get absorbed into any adjacent EXACTFU
4093 * node. And if the only adjacent node is EXACTF, they get
4094 * absorbed into that, under the theory that a longer node is
4095 * better than two shorter ones, even if one is EXACTFU. Note
4096 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4097 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4099 if (STRING(n)[STR_LEN(n)-1] == 's') {
4101 /* Here the joined node would end with 's'. If the node
4102 * following the combination is an EXACTF one, it's better to
4103 * join this trailing edge 's' node with that one, leaving the
4104 * current one in 'scan' be the more desirable EXACTFU */
4105 if (OP(nnext) == EXACTF) {
4109 OP(scan) = EXACTFU_S_EDGE;
4111 } /* Otherwise, the beginning 's' of the 2nd node just
4112 becomes an interior 's' in 'scan' */
4114 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4115 ; /* join is compatible, no need to change OP */
4117 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4119 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4120 * nodes. But the latter nodes can be also joined with EXACTFU
4121 * ones, and that is a better outcome, so if the node following
4122 * 'n' is EXACTFU, quit now so that those two can be joined
4124 if (OP(nnext) == EXACTFU) {
4128 /* The join is compatible, and the combined node will be
4129 * EXACTF. (These don't care if they begin or end with 's' */
4131 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4132 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4133 && STRING(n)[0] == 's')
4135 /* When combined, we have the sequence 'ss', which means we
4136 * have to remain /di */
4140 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4141 if (STRING(n)[0] == 's') {
4142 ; /* Here the join is compatible and the combined node
4143 starts with 's', no need to change OP */
4145 else { /* Now the trailing 's' is in the interior */
4149 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4151 /* The join is compatible, and the combined node will be
4152 * EXACTF. (These don't care if they begin or end with 's' */
4155 else if (OP(scan) != OP(n)) {
4157 /* The only other compatible joinings are the same node type */
4161 DEBUG_PEEP("merg", n, depth, 0);
4164 NEXT_OFF(scan) += NEXT_OFF(n);
4165 setSTR_LEN(scan, STR_LEN(scan) + STR_LEN(n));
4166 next = n + NODE_SZ_STR(n);
4167 /* Now we can overwrite *n : */
4168 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4176 #ifdef EXPERIMENTAL_INPLACESCAN
4177 if (flags && !NEXT_OFF(n)) {
4178 DEBUG_PEEP("atch", val, depth, 0);
4179 if (reg_off_by_arg[OP(n)]) {
4180 ARG_SET(n, val - n);
4183 NEXT_OFF(n) = val - n;
4190 /* This temporary node can now be turned into EXACTFU, and must, as
4191 * regexec.c doesn't handle it */
4192 if (OP(scan) == EXACTFU_S_EDGE) {
4197 *unfolded_multi_char = FALSE;
4199 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4200 * can now analyze for sequences of problematic code points. (Prior to
4201 * this final joining, sequences could have been split over boundaries, and
4202 * hence missed). The sequences only happen in folding, hence for any
4203 * non-EXACT EXACTish node */
4204 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4205 U8* s0 = (U8*) STRING(scan);
4207 U8* s_end = s0 + STR_LEN(scan);
4209 int total_count_delta = 0; /* Total delta number of characters that
4210 multi-char folds expand to */
4212 /* One pass is made over the node's string looking for all the
4213 * possibilities. To avoid some tests in the loop, there are two main
4214 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4219 if (OP(scan) == EXACTFL) {
4222 /* An EXACTFL node would already have been changed to another
4223 * node type unless there is at least one character in it that
4224 * is problematic; likely a character whose fold definition
4225 * won't be known until runtime, and so has yet to be folded.
4226 * For all but the UTF-8 locale, folds are 1-1 in length, but
4227 * to handle the UTF-8 case, we need to create a temporary
4228 * folded copy using UTF-8 locale rules in order to analyze it.
4229 * This is because our macros that look to see if a sequence is
4230 * a multi-char fold assume everything is folded (otherwise the
4231 * tests in those macros would be too complicated and slow).
4232 * Note that here, the non-problematic folds will have already
4233 * been done, so we can just copy such characters. We actually
4234 * don't completely fold the EXACTFL string. We skip the
4235 * unfolded multi-char folds, as that would just create work
4236 * below to figure out the size they already are */
4238 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4241 STRLEN s_len = UTF8SKIP(s);
4242 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4243 Copy(s, d, s_len, U8);
4246 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4247 *unfolded_multi_char = TRUE;
4248 Copy(s, d, s_len, U8);
4251 else if (isASCII(*s)) {
4252 *(d++) = toFOLD(*s);
4256 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4262 /* Point the remainder of the routine to look at our temporary
4266 } /* End of creating folded copy of EXACTFL string */
4268 /* Examine the string for a multi-character fold sequence. UTF-8
4269 * patterns have all characters pre-folded by the time this code is
4271 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4272 length sequence we are looking for is 2 */
4274 int count = 0; /* How many characters in a multi-char fold */
4275 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4276 if (! len) { /* Not a multi-char fold: get next char */
4281 { /* Here is a generic multi-char fold. */
4282 U8* multi_end = s + len;
4284 /* Count how many characters are in it. In the case of
4285 * /aa, no folds which contain ASCII code points are
4286 * allowed, so check for those, and skip if found. */
4287 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4288 count = utf8_length(s, multi_end);
4292 while (s < multi_end) {
4295 goto next_iteration;
4305 /* The delta is how long the sequence is minus 1 (1 is how long
4306 * the character that folds to the sequence is) */
4307 total_count_delta += count - 1;
4311 /* We created a temporary folded copy of the string in EXACTFL
4312 * nodes. Therefore we need to be sure it doesn't go below zero,
4313 * as the real string could be shorter */
4314 if (OP(scan) == EXACTFL) {
4315 int total_chars = utf8_length((U8*) STRING(scan),
4316 (U8*) STRING(scan) + STR_LEN(scan));
4317 if (total_count_delta > total_chars) {
4318 total_count_delta = total_chars;
4322 *min_subtract += total_count_delta;
4325 else if (OP(scan) == EXACTFAA) {
4327 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4328 * fold to the ASCII range (and there are no existing ones in the
4329 * upper latin1 range). But, as outlined in the comments preceding
4330 * this function, we need to flag any occurrences of the sharp s.
4331 * This character forbids trie formation (because of added
4333 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4334 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4335 || UNICODE_DOT_DOT_VERSION > 0)
4337 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4338 OP(scan) = EXACTFAA_NO_TRIE;
4339 *unfolded_multi_char = TRUE;
4347 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4348 * folds that are all Latin1. As explained in the comments
4349 * preceding this function, we look also for the sharp s in EXACTF
4350 * and EXACTFL nodes; it can be in the final position. Otherwise
4351 * we can stop looking 1 byte earlier because have to find at least
4352 * two characters for a multi-fold */
4353 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4358 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4359 if (! len) { /* Not a multi-char fold. */
4360 if (*s == LATIN_SMALL_LETTER_SHARP_S
4361 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4363 *unfolded_multi_char = TRUE;
4370 && isALPHA_FOLD_EQ(*s, 's')
4371 && isALPHA_FOLD_EQ(*(s+1), 's'))
4374 /* EXACTF nodes need to know that the minimum length
4375 * changed so that a sharp s in the string can match this
4376 * ss in the pattern, but they remain EXACTF nodes, as they
4377 * won't match this unless the target string is is UTF-8,
4378 * which we don't know until runtime. EXACTFL nodes can't
4379 * transform into EXACTFU nodes */
4380 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4381 OP(scan) = EXACTFUP;
4385 *min_subtract += len - 1;
4391 if ( STR_LEN(scan) == 1
4392 && isALPHA_A(* STRING(scan))
4393 && ( OP(scan) == EXACTFAA
4394 || ( OP(scan) == EXACTFU
4395 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(scan)))))
4397 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
4399 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
4400 * with the mask set to the complement of the bit that differs
4401 * between upper and lower case, and the lowest code point of the
4402 * pair (which the '&' forces) */
4404 ARG_SET(scan, *STRING(scan) & mask);
4410 /* Allow dumping but overwriting the collection of skipped
4411 * ops and/or strings with fake optimized ops */
4412 n = scan + NODE_SZ_STR(scan);
4420 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4424 /* REx optimizer. Converts nodes into quicker variants "in place".
4425 Finds fixed substrings. */
4427 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4428 to the position after last scanned or to NULL. */
4430 #define INIT_AND_WITHP \
4431 assert(!and_withp); \
4432 Newx(and_withp, 1, regnode_ssc); \
4433 SAVEFREEPV(and_withp)
4437 S_unwind_scan_frames(pTHX_ const void *p)
4439 scan_frame *f= (scan_frame *)p;
4441 scan_frame *n= f->next_frame;
4447 /* the return from this sub is the minimum length that could possibly match */
4449 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4450 SSize_t *minlenp, SSize_t *deltap,
4455 regnode_ssc *and_withp,
4456 U32 flags, U32 depth)
4457 /* scanp: Start here (read-write). */
4458 /* deltap: Write maxlen-minlen here. */
4459 /* last: Stop before this one. */
4460 /* data: string data about the pattern */
4461 /* stopparen: treat close N as END */
4462 /* recursed: which subroutines have we recursed into */
4463 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4466 /* There must be at least this number of characters to match */
4469 regnode *scan = *scanp, *next;
4471 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4472 int is_inf_internal = 0; /* The studied chunk is infinite */
4473 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4474 scan_data_t data_fake;
4475 SV *re_trie_maxbuff = NULL;
4476 regnode *first_non_open = scan;
4477 SSize_t stopmin = SSize_t_MAX;
4478 scan_frame *frame = NULL;
4479 GET_RE_DEBUG_FLAGS_DECL;
4481 PERL_ARGS_ASSERT_STUDY_CHUNK;
4482 RExC_study_started= 1;
4484 Zero(&data_fake, 1, scan_data_t);
4487 while (first_non_open && OP(first_non_open) == OPEN)
4488 first_non_open=regnext(first_non_open);
4494 RExC_study_chunk_recursed_count++;
4496 DEBUG_OPTIMISE_MORE_r(
4498 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4499 depth, (long)stopparen,
4500 (unsigned long)RExC_study_chunk_recursed_count,
4501 (unsigned long)depth, (unsigned long)recursed_depth,
4504 if (recursed_depth) {
4507 for ( j = 0 ; j < recursed_depth ; j++ ) {
4508 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4510 PAREN_TEST(RExC_study_chunk_recursed +
4511 ( j * RExC_study_chunk_recursed_bytes), i )
4514 !PAREN_TEST(RExC_study_chunk_recursed +
4515 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4518 Perl_re_printf( aTHX_ " %d",(int)i);
4522 if ( j + 1 < recursed_depth ) {
4523 Perl_re_printf( aTHX_ ",");
4527 Perl_re_printf( aTHX_ "\n");
4530 while ( scan && OP(scan) != END && scan < last ){
4531 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4532 node length to get a real minimum (because
4533 the folded version may be shorter) */
4534 bool unfolded_multi_char = FALSE;
4535 /* Peephole optimizer: */
4536 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4537 DEBUG_PEEP("Peep", scan, depth, flags);
4540 /* The reason we do this here is that we need to deal with things like
4541 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4542 * parsing code, as each (?:..) is handled by a different invocation of
4545 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4547 /* Follow the next-chain of the current node and optimize
4548 away all the NOTHINGs from it. */
4549 if (OP(scan) != CURLYX) {
4550 const int max = (reg_off_by_arg[OP(scan)]
4552 /* I32 may be smaller than U16 on CRAYs! */
4553 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4554 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4558 /* Skip NOTHING and LONGJMP. */
4559 while ( (n = regnext(n))
4560 && ( (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4561 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4562 && off + noff < max)
4564 if (reg_off_by_arg[OP(scan)])
4567 NEXT_OFF(scan) = off;
4570 /* The principal pseudo-switch. Cannot be a switch, since we look into
4571 * several different things. */
4572 if ( OP(scan) == DEFINEP ) {
4574 SSize_t deltanext = 0;
4575 SSize_t fake_last_close = 0;
4576 I32 f = SCF_IN_DEFINE;
4578 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4579 scan = regnext(scan);
4580 assert( OP(scan) == IFTHEN );
4581 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4583 data_fake.last_closep= &fake_last_close;
4585 next = regnext(scan);
4586 scan = NEXTOPER(NEXTOPER(scan));
4587 DEBUG_PEEP("scan", scan, depth, flags);
4588 DEBUG_PEEP("next", next, depth, flags);
4590 /* we suppose the run is continuous, last=next...
4591 * NOTE we dont use the return here! */
4592 /* DEFINEP study_chunk() recursion */
4593 (void)study_chunk(pRExC_state, &scan, &minlen,
4594 &deltanext, next, &data_fake, stopparen,
4595 recursed_depth, NULL, f, depth+1);
4600 OP(scan) == BRANCH ||
4601 OP(scan) == BRANCHJ ||
4604 next = regnext(scan);
4607 /* The op(next)==code check below is to see if we
4608 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4609 * IFTHEN is special as it might not appear in pairs.
4610 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4611 * we dont handle it cleanly. */
4612 if (OP(next) == code || code == IFTHEN) {
4613 /* NOTE - There is similar code to this block below for
4614 * handling TRIE nodes on a re-study. If you change stuff here
4615 * check there too. */
4616 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4618 regnode * const startbranch=scan;
4620 if (flags & SCF_DO_SUBSTR) {
4621 /* Cannot merge strings after this. */
4622 scan_commit(pRExC_state, data, minlenp, is_inf);
4625 if (flags & SCF_DO_STCLASS)
4626 ssc_init_zero(pRExC_state, &accum);
4628 while (OP(scan) == code) {
4629 SSize_t deltanext, minnext, fake;
4631 regnode_ssc this_class;
4633 DEBUG_PEEP("Branch", scan, depth, flags);
4636 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4638 data_fake.whilem_c = data->whilem_c;
4639 data_fake.last_closep = data->last_closep;
4642 data_fake.last_closep = &fake;
4644 data_fake.pos_delta = delta;
4645 next = regnext(scan);
4647 scan = NEXTOPER(scan); /* everything */
4648 if (code != BRANCH) /* everything but BRANCH */
4649 scan = NEXTOPER(scan);
4651 if (flags & SCF_DO_STCLASS) {
4652 ssc_init(pRExC_state, &this_class);
4653 data_fake.start_class = &this_class;
4654 f = SCF_DO_STCLASS_AND;
4656 if (flags & SCF_WHILEM_VISITED_POS)
4657 f |= SCF_WHILEM_VISITED_POS;
4659 /* we suppose the run is continuous, last=next...*/
4660 /* recurse study_chunk() for each BRANCH in an alternation */
4661 minnext = study_chunk(pRExC_state, &scan, minlenp,
4662 &deltanext, next, &data_fake, stopparen,
4663 recursed_depth, NULL, f, depth+1);
4667 if (deltanext == SSize_t_MAX) {
4668 is_inf = is_inf_internal = 1;
4670 } else if (max1 < minnext + deltanext)
4671 max1 = minnext + deltanext;
4673 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4675 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4676 if ( stopmin > minnext)
4677 stopmin = min + min1;
4678 flags &= ~SCF_DO_SUBSTR;
4680 data->flags |= SCF_SEEN_ACCEPT;
4683 if (data_fake.flags & SF_HAS_EVAL)
4684 data->flags |= SF_HAS_EVAL;
4685 data->whilem_c = data_fake.whilem_c;
4687 if (flags & SCF_DO_STCLASS)
4688 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4690 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4692 if (flags & SCF_DO_SUBSTR) {
4693 data->pos_min += min1;
4694 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4695 data->pos_delta = SSize_t_MAX;
4697 data->pos_delta += max1 - min1;
4698 if (max1 != min1 || is_inf)
4699 data->cur_is_floating = 1;
4702 if (delta == SSize_t_MAX
4703 || SSize_t_MAX - delta - (max1 - min1) < 0)
4704 delta = SSize_t_MAX;
4706 delta += max1 - min1;
4707 if (flags & SCF_DO_STCLASS_OR) {
4708 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4710 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4711 flags &= ~SCF_DO_STCLASS;
4714 else if (flags & SCF_DO_STCLASS_AND) {
4716 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4717 flags &= ~SCF_DO_STCLASS;
4720 /* Switch to OR mode: cache the old value of
4721 * data->start_class */
4723 StructCopy(data->start_class, and_withp, regnode_ssc);
4724 flags &= ~SCF_DO_STCLASS_AND;
4725 StructCopy(&accum, data->start_class, regnode_ssc);
4726 flags |= SCF_DO_STCLASS_OR;
4730 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4731 OP( startbranch ) == BRANCH )
4735 Assuming this was/is a branch we are dealing with: 'scan'
4736 now points at the item that follows the branch sequence,
4737 whatever it is. We now start at the beginning of the
4738 sequence and look for subsequences of
4744 which would be constructed from a pattern like
4747 If we can find such a subsequence we need to turn the first
4748 element into a trie and then add the subsequent branch exact
4749 strings to the trie.
4753 1. patterns where the whole set of branches can be
4756 2. patterns where only a subset can be converted.
4758 In case 1 we can replace the whole set with a single regop
4759 for the trie. In case 2 we need to keep the start and end
4762 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4763 becomes BRANCH TRIE; BRANCH X;
4765 There is an additional case, that being where there is a
4766 common prefix, which gets split out into an EXACT like node
4767 preceding the TRIE node.
4769 If x(1..n)==tail then we can do a simple trie, if not we make
4770 a "jump" trie, such that when we match the appropriate word
4771 we "jump" to the appropriate tail node. Essentially we turn
4772 a nested if into a case structure of sorts.
4777 if (!re_trie_maxbuff) {
4778 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4779 if (!SvIOK(re_trie_maxbuff))
4780 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4782 if ( SvIV(re_trie_maxbuff)>=0 ) {
4784 regnode *first = (regnode *)NULL;
4785 regnode *last = (regnode *)NULL;
4786 regnode *tail = scan;
4790 /* var tail is used because there may be a TAIL
4791 regop in the way. Ie, the exacts will point to the
4792 thing following the TAIL, but the last branch will
4793 point at the TAIL. So we advance tail. If we
4794 have nested (?:) we may have to move through several
4798 while ( OP( tail ) == TAIL ) {
4799 /* this is the TAIL generated by (?:) */
4800 tail = regnext( tail );
4804 DEBUG_TRIE_COMPILE_r({
4805 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4806 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4808 "Looking for TRIE'able sequences. Tail node is ",
4809 (UV) REGNODE_OFFSET(tail),
4810 SvPV_nolen_const( RExC_mysv )
4816 Step through the branches
4817 cur represents each branch,
4818 noper is the first thing to be matched as part
4820 noper_next is the regnext() of that node.
4822 We normally handle a case like this
4823 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4824 support building with NOJUMPTRIE, which restricts
4825 the trie logic to structures like /FOO|BAR/.
4827 If noper is a trieable nodetype then the branch is
4828 a possible optimization target. If we are building
4829 under NOJUMPTRIE then we require that noper_next is
4830 the same as scan (our current position in the regex
4833 Once we have two or more consecutive such branches
4834 we can create a trie of the EXACT's contents and
4835 stitch it in place into the program.
4837 If the sequence represents all of the branches in
4838 the alternation we replace the entire thing with a
4841 Otherwise when it is a subsequence we need to
4842 stitch it in place and replace only the relevant
4843 branches. This means the first branch has to remain
4844 as it is used by the alternation logic, and its
4845 next pointer, and needs to be repointed at the item
4846 on the branch chain following the last branch we
4847 have optimized away.
4849 This could be either a BRANCH, in which case the
4850 subsequence is internal, or it could be the item
4851 following the branch sequence in which case the
4852 subsequence is at the end (which does not
4853 necessarily mean the first node is the start of the
4856 TRIE_TYPE(X) is a define which maps the optype to a
4860 ----------------+-----------
4865 EXACTFU_REQ8 | EXACTFU
4869 EXACTFLU8 | EXACTFLU8
4873 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4875 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4877 : ( EXACTFU == (X) \
4878 || EXACTFU_REQ8 == (X) \
4879 || EXACTFUP == (X) ) \
4881 : ( EXACTFAA == (X) ) \
4883 : ( EXACTL == (X) ) \
4885 : ( EXACTFLU8 == (X) ) \
4889 /* dont use tail as the end marker for this traverse */
4890 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4891 regnode * const noper = NEXTOPER( cur );
4892 U8 noper_type = OP( noper );
4893 U8 noper_trietype = TRIE_TYPE( noper_type );
4894 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4895 regnode * const noper_next = regnext( noper );
4896 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4897 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4900 DEBUG_TRIE_COMPILE_r({
4901 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4902 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4904 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4906 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4907 Perl_re_printf( aTHX_ " -> %d:%s",
4908 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4911 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4912 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4913 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4915 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4916 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4917 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4921 /* Is noper a trieable nodetype that can be merged
4922 * with the current trie (if there is one)? */
4926 ( noper_trietype == NOTHING )
4927 || ( trietype == NOTHING )
4928 || ( trietype == noper_trietype )
4931 && noper_next >= tail
4935 /* Handle mergable triable node Either we are
4936 * the first node in a new trieable sequence,
4937 * in which case we do some bookkeeping,
4938 * otherwise we update the end pointer. */
4941 if ( noper_trietype == NOTHING ) {
4942 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4943 regnode * const noper_next = regnext( noper );
4944 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4945 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4948 if ( noper_next_trietype ) {
4949 trietype = noper_next_trietype;
4950 } else if (noper_next_type) {
4951 /* a NOTHING regop is 1 regop wide.
4952 * We need at least two for a trie
4953 * so we can't merge this in */
4957 trietype = noper_trietype;
4960 if ( trietype == NOTHING )
4961 trietype = noper_trietype;
4966 } /* end handle mergable triable node */
4968 /* handle unmergable node -
4969 * noper may either be a triable node which can
4970 * not be tried together with the current trie,
4971 * or a non triable node */
4973 /* If last is set and trietype is not
4974 * NOTHING then we have found at least two
4975 * triable branch sequences in a row of a
4976 * similar trietype so we can turn them
4977 * into a trie. If/when we allow NOTHING to
4978 * start a trie sequence this condition
4979 * will be required, and it isn't expensive
4980 * so we leave it in for now. */
4981 if ( trietype && trietype != NOTHING )
4982 make_trie( pRExC_state,
4983 startbranch, first, cur, tail,
4984 count, trietype, depth+1 );
4985 last = NULL; /* note: we clear/update
4986 first, trietype etc below,
4987 so we dont do it here */
4991 && noper_next >= tail
4994 /* noper is triable, so we can start a new
4998 trietype = noper_trietype;
5000 /* if we already saw a first but the
5001 * current node is not triable then we have
5002 * to reset the first information. */
5007 } /* end handle unmergable node */
5008 } /* loop over branches */
5009 DEBUG_TRIE_COMPILE_r({
5010 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5011 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5012 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5013 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5014 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
5015 PL_reg_name[trietype]
5019 if ( last && trietype ) {
5020 if ( trietype != NOTHING ) {
5021 /* the last branch of the sequence was part of
5022 * a trie, so we have to construct it here
5023 * outside of the loop */
5024 made= make_trie( pRExC_state, startbranch,
5025 first, scan, tail, count,
5026 trietype, depth+1 );
5027 #ifdef TRIE_STUDY_OPT
5028 if ( ((made == MADE_EXACT_TRIE &&
5029 startbranch == first)
5030 || ( first_non_open == first )) &&
5032 flags |= SCF_TRIE_RESTUDY;
5033 if ( startbranch == first
5036 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5041 /* at this point we know whatever we have is a
5042 * NOTHING sequence/branch AND if 'startbranch'
5043 * is 'first' then we can turn the whole thing
5046 if ( startbranch == first ) {
5048 /* the entire thing is a NOTHING sequence,
5049 * something like this: (?:|) So we can
5050 * turn it into a plain NOTHING op. */
5051 DEBUG_TRIE_COMPILE_r({
5052 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5053 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5055 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5058 OP(startbranch)= NOTHING;
5059 NEXT_OFF(startbranch)= tail - startbranch;
5060 for ( opt= startbranch + 1; opt < tail ; opt++ )
5064 } /* end if ( last) */
5065 } /* TRIE_MAXBUF is non zero */
5070 else if ( code == BRANCHJ ) { /* single branch is optimized. */
5071 scan = NEXTOPER(NEXTOPER(scan));
5072 } else /* single branch is optimized. */
5073 scan = NEXTOPER(scan);
5075 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB) {
5077 regnode *start = NULL;
5078 regnode *end = NULL;
5079 U32 my_recursed_depth= recursed_depth;
5081 if (OP(scan) != SUSPEND) { /* GOSUB */
5082 /* Do setup, note this code has side effects beyond
5083 * the rest of this block. Specifically setting
5084 * RExC_recurse[] must happen at least once during
5087 RExC_recurse[ARG2L(scan)] = scan;
5088 start = REGNODE_p(RExC_open_parens[paren]);
5089 end = REGNODE_p(RExC_close_parens[paren]);
5091 /* NOTE we MUST always execute the above code, even
5092 * if we do nothing with a GOSUB */
5094 ( flags & SCF_IN_DEFINE )
5097 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
5099 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
5102 /* no need to do anything here if we are in a define. */
5103 /* or we are after some kind of infinite construct
5104 * so we can skip recursing into this item.
5105 * Since it is infinite we will not change the maxlen
5106 * or delta, and if we miss something that might raise
5107 * the minlen it will merely pessimise a little.
5109 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
5110 * might result in a minlen of 1 and not of 4,
5111 * but this doesn't make us mismatch, just try a bit
5112 * harder than we should.
5114 scan= regnext(scan);
5121 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
5123 /* it is quite possible that there are more efficient ways
5124 * to do this. We maintain a bitmap per level of recursion
5125 * of which patterns we have entered so we can detect if a
5126 * pattern creates a possible infinite loop. When we
5127 * recurse down a level we copy the previous levels bitmap
5128 * down. When we are at recursion level 0 we zero the top
5129 * level bitmap. It would be nice to implement a different
5130 * more efficient way of doing this. In particular the top
5131 * level bitmap may be unnecessary.
5133 if (!recursed_depth) {
5134 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
5136 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
5137 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
5138 RExC_study_chunk_recursed_bytes, U8);
5140 /* we havent recursed into this paren yet, so recurse into it */
5141 DEBUG_STUDYDATA("gosub-set", data, depth, is_inf);
5142 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
5143 my_recursed_depth= recursed_depth + 1;
5145 DEBUG_STUDYDATA("gosub-inf", data, depth, is_inf);
5146 /* some form of infinite recursion, assume infinite length
5148 if (flags & SCF_DO_SUBSTR) {
5149 scan_commit(pRExC_state, data, minlenp, is_inf);
5150 data->cur_is_floating = 1;
5152 is_inf = is_inf_internal = 1;
5153 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5154 ssc_anything(data->start_class);
5155 flags &= ~SCF_DO_STCLASS;
5157 start= NULL; /* reset start so we dont recurse later on. */
5162 end = regnext(scan);
5165 scan_frame *newframe;
5167 if (!RExC_frame_last) {
5168 Newxz(newframe, 1, scan_frame);
5169 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
5170 RExC_frame_head= newframe;
5172 } else if (!RExC_frame_last->next_frame) {
5173 Newxz(newframe, 1, scan_frame);
5174 RExC_frame_last->next_frame= newframe;
5175 newframe->prev_frame= RExC_frame_last;
5178 newframe= RExC_frame_last->next_frame;
5180 RExC_frame_last= newframe;
5182 newframe->next_regnode = regnext(scan);
5183 newframe->last_regnode = last;
5184 newframe->stopparen = stopparen;
5185 newframe->prev_recursed_depth = recursed_depth;
5186 newframe->this_prev_frame= frame;
5188 DEBUG_STUDYDATA("frame-new", data, depth, is_inf);
5189 DEBUG_PEEP("fnew", scan, depth, flags);
5196 recursed_depth= my_recursed_depth;
5201 else if ( OP(scan) == EXACT
5202 || OP(scan) == LEXACT
5203 || OP(scan) == EXACT_REQ8
5204 || OP(scan) == LEXACT_REQ8
5205 || OP(scan) == EXACTL)
5207 SSize_t l = STR_LEN(scan);
5211 const U8 * const s = (U8*)STRING(scan);
5212 uc = utf8_to_uvchr_buf(s, s + l, NULL);
5213 l = utf8_length(s, s + l);
5215 uc = *((U8*)STRING(scan));
5218 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
5219 /* The code below prefers earlier match for fixed
5220 offset, later match for variable offset. */
5221 if (data->last_end == -1) { /* Update the start info. */
5222 data->last_start_min = data->pos_min;
5223 data->last_start_max = is_inf
5224 ? SSize_t_MAX : data->pos_min + data->pos_delta;
5226 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
5228 SvUTF8_on(data->last_found);
5230 SV * const sv = data->last_found;
5231 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5232 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5233 if (mg && mg->mg_len >= 0)
5234 mg->mg_len += utf8_length((U8*)STRING(scan),
5235 (U8*)STRING(scan)+STR_LEN(scan));
5237 data->last_end = data->pos_min + l;
5238 data->pos_min += l; /* As in the first entry. */
5239 data->flags &= ~SF_BEFORE_EOL;
5242 /* ANDing the code point leaves at most it, and not in locale, and
5243 * can't match null string */
5244 if (flags & SCF_DO_STCLASS_AND) {
5245 ssc_cp_and(data->start_class, uc);
5246 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5247 ssc_clear_locale(data->start_class);
5249 else if (flags & SCF_DO_STCLASS_OR) {
5250 ssc_add_cp(data->start_class, uc);
5251 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5253 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5254 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5256 flags &= ~SCF_DO_STCLASS;
5258 else if (PL_regkind[OP(scan)] == EXACT) {
5259 /* But OP != EXACT!, so is EXACTFish */
5260 SSize_t l = STR_LEN(scan);
5261 const U8 * s = (U8*)STRING(scan);
5263 /* Search for fixed substrings supports EXACT only. */
5264 if (flags & SCF_DO_SUBSTR) {
5266 scan_commit(pRExC_state, data, minlenp, is_inf);
5269 l = utf8_length(s, s + l);
5271 if (unfolded_multi_char) {
5272 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
5274 min += l - min_subtract;
5276 delta += min_subtract;
5277 if (flags & SCF_DO_SUBSTR) {
5278 data->pos_min += l - min_subtract;
5279 if (data->pos_min < 0) {
5282 data->pos_delta += min_subtract;
5284 data->cur_is_floating = 1; /* float */
5288 if (flags & SCF_DO_STCLASS) {
5289 SV* EXACTF_invlist = make_exactf_invlist(pRExC_state, scan);
5291 assert(EXACTF_invlist);
5292 if (flags & SCF_DO_STCLASS_AND) {
5293 if (OP(scan) != EXACTFL)
5294 ssc_clear_locale(data->start_class);
5295 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5296 ANYOF_POSIXL_ZERO(data->start_class);
5297 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
5299 else { /* SCF_DO_STCLASS_OR */
5300 ssc_union(data->start_class, EXACTF_invlist, FALSE);
5301 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5303 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5304 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5306 flags &= ~SCF_DO_STCLASS;
5307 SvREFCNT_dec(EXACTF_invlist);
5310 else if (REGNODE_VARIES(OP(scan))) {
5311 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
5312 I32 fl = 0, f = flags;
5313 regnode * const oscan = scan;
5314 regnode_ssc this_class;
5315 regnode_ssc *oclass = NULL;
5316 I32 next_is_eval = 0;
5318 switch (PL_regkind[OP(scan)]) {
5319 case WHILEM: /* End of (?:...)* . */
5320 scan = NEXTOPER(scan);
5323 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
5324 next = NEXTOPER(scan);
5325 if ( OP(next) == EXACT
5326 || OP(next) == LEXACT
5327 || OP(next) == EXACT_REQ8
5328 || OP(next) == LEXACT_REQ8
5329 || OP(next) == EXACTL
5330 || (flags & SCF_DO_STCLASS))
5333 maxcount = REG_INFTY;
5334 next = regnext(scan);
5335 scan = NEXTOPER(scan);
5339 if (flags & SCF_DO_SUBSTR)
5344 next = NEXTOPER(scan);
5346 /* This temporary node can now be turned into EXACTFU, and
5347 * must, as regexec.c doesn't handle it */
5348 if (OP(next) == EXACTFU_S_EDGE) {
5352 if ( STR_LEN(next) == 1
5353 && isALPHA_A(* STRING(next))
5354 && ( OP(next) == EXACTFAA
5355 || ( OP(next) == EXACTFU
5356 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(next)))))
5358 /* These differ in just one bit */
5359 U8 mask = ~ ('A' ^ 'a');
5361 assert(isALPHA_A(* STRING(next)));
5363 /* Then replace it by an ANYOFM node, with
5364 * the mask set to the complement of the
5365 * bit that differs between upper and lower
5366 * case, and the lowest code point of the
5367 * pair (which the '&' forces) */
5369 ARG_SET(next, *STRING(next) & mask);
5373 if (flags & SCF_DO_STCLASS) {
5375 maxcount = REG_INFTY;
5376 next = regnext(scan);
5377 scan = NEXTOPER(scan);
5380 if (flags & SCF_DO_SUBSTR) {
5381 scan_commit(pRExC_state, data, minlenp, is_inf);
5382 /* Cannot extend fixed substrings */
5383 data->cur_is_floating = 1; /* float */
5385 is_inf = is_inf_internal = 1;
5386 scan = regnext(scan);
5387 goto optimize_curly_tail;
5389 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
5390 && (scan->flags == stopparen))
5395 mincount = ARG1(scan);
5396 maxcount = ARG2(scan);
5398 next = regnext(scan);
5399 if (OP(scan) == CURLYX) {
5400 I32 lp = (data ? *(data->last_closep) : 0);
5401 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
5403 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
5404 next_is_eval = (OP(scan) == EVAL);
5406 if (flags & SCF_DO_SUBSTR) {
5408 scan_commit(pRExC_state, data, minlenp, is_inf);
5409 /* Cannot extend fixed substrings */
5410 pos_before = data->pos_min;
5414 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
5416 data->flags |= SF_IS_INF;
5418 if (flags & SCF_DO_STCLASS) {
5419 ssc_init(pRExC_state, &this_class);
5420 oclass = data->start_class;
5421 data->start_class = &this_class;
5422 f |= SCF_DO_STCLASS_AND;
5423 f &= ~SCF_DO_STCLASS_OR;
5425 /* Exclude from super-linear cache processing any {n,m}
5426 regops for which the combination of input pos and regex
5427 pos is not enough information to determine if a match
5430 For example, in the regex /foo(bar\s*){4,8}baz/ with the
5431 regex pos at the \s*, the prospects for a match depend not
5432 only on the input position but also on how many (bar\s*)
5433 repeats into the {4,8} we are. */
5434 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
5435 f &= ~SCF_WHILEM_VISITED_POS;
5437 /* This will finish on WHILEM, setting scan, or on NULL: */
5438 /* recurse study_chunk() on loop bodies */
5439 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
5440 last, data, stopparen, recursed_depth, NULL,
5442 ? (f & ~SCF_DO_SUBSTR)
5446 if (flags & SCF_DO_STCLASS)
5447 data->start_class = oclass;
5448 if (mincount == 0 || minnext == 0) {
5449 if (flags & SCF_DO_STCLASS_OR) {
5450 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5452 else if (flags & SCF_DO_STCLASS_AND) {
5453 /* Switch to OR mode: cache the old value of
5454 * data->start_class */
5456 StructCopy(data->start_class, and_withp, regnode_ssc);
5457 flags &= ~SCF_DO_STCLASS_AND;
5458 StructCopy(&this_class, data->start_class, regnode_ssc);
5459 flags |= SCF_DO_STCLASS_OR;
5460 ANYOF_FLAGS(data->start_class)
5461 |= SSC_MATCHES_EMPTY_STRING;
5463 } else { /* Non-zero len */
5464 if (flags & SCF_DO_STCLASS_OR) {
5465 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5466 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5468 else if (flags & SCF_DO_STCLASS_AND)
5469 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
5470 flags &= ~SCF_DO_STCLASS;
5472 if (!scan) /* It was not CURLYX, but CURLY. */
5474 if (((flags & (SCF_TRIE_DOING_RESTUDY|SCF_DO_SUBSTR))==SCF_DO_SUBSTR)
5475 /* ? quantifier ok, except for (?{ ... }) */
5476 && (next_is_eval || !(mincount == 0 && maxcount == 1))
5477 && (minnext == 0) && (deltanext == 0)
5478 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
5479 && maxcount <= REG_INFTY/3) /* Complement check for big
5482 _WARN_HELPER(RExC_precomp_end, packWARN(WARN_REGEXP),
5483 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
5484 "Quantifier unexpected on zero-length expression "
5485 "in regex m/%" UTF8f "/",
5486 UTF8fARG(UTF, RExC_precomp_end - RExC_precomp,
5490 min += minnext * mincount;
5491 is_inf_internal |= deltanext == SSize_t_MAX
5492 || (maxcount == REG_INFTY && minnext + deltanext > 0);
5493 is_inf |= is_inf_internal;
5495 delta = SSize_t_MAX;
5497 delta += (minnext + deltanext) * maxcount
5498 - minnext * mincount;
5500 /* Try powerful optimization CURLYX => CURLYN. */
5501 if ( OP(oscan) == CURLYX && data
5502 && data->flags & SF_IN_PAR
5503 && !(data->flags & SF_HAS_EVAL)
5504 && !deltanext && minnext == 1 ) {
5505 /* Try to optimize to CURLYN. */
5506 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
5507 regnode * const nxt1 = nxt;
5514 if (!REGNODE_SIMPLE(OP(nxt))
5515 && !(PL_regkind[OP(nxt)] == EXACT
5516 && STR_LEN(nxt) == 1))
5522 if (OP(nxt) != CLOSE)
5524 if (RExC_open_parens) {
5527 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5530 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt) + 2;
5532 /* Now we know that nxt2 is the only contents: */
5533 oscan->flags = (U8)ARG(nxt);
5535 OP(nxt1) = NOTHING; /* was OPEN. */
5538 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5539 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
5540 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
5541 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5542 OP(nxt + 1) = OPTIMIZED; /* was count. */
5543 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
5548 /* Try optimization CURLYX => CURLYM. */
5549 if ( OP(oscan) == CURLYX && data
5550 && !(data->flags & SF_HAS_PAR)
5551 && !(data->flags & SF_HAS_EVAL)
5552 && !deltanext /* atom is fixed width */
5553 && minnext != 0 /* CURLYM can't handle zero width */
5555 /* Nor characters whose fold at run-time may be
5556 * multi-character */
5557 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
5559 /* XXXX How to optimize if data == 0? */
5560 /* Optimize to a simpler form. */
5561 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
5565 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
5566 && (OP(nxt2) != WHILEM))
5568 OP(nxt2) = SUCCEED; /* Whas WHILEM */
5569 /* Need to optimize away parenths. */
5570 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
5571 /* Set the parenth number. */
5572 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
5574 oscan->flags = (U8)ARG(nxt);
5575 if (RExC_open_parens) {
5577 RExC_open_parens[ARG(nxt1)] = REGNODE_OFFSET(oscan);
5580 RExC_close_parens[ARG(nxt1)] = REGNODE_OFFSET(nxt2)
5583 OP(nxt1) = OPTIMIZED; /* was OPEN. */
5584 OP(nxt) = OPTIMIZED; /* was CLOSE. */
5587 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
5588 OP(nxt + 1) = OPTIMIZED; /* was count. */
5589 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
5590 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
5593 while ( nxt1 && (OP(nxt1) != WHILEM)) {
5594 regnode *nnxt = regnext(nxt1);
5596 if (reg_off_by_arg[OP(nxt1)])
5597 ARG_SET(nxt1, nxt2 - nxt1);
5598 else if (nxt2 - nxt1 < U16_MAX)
5599 NEXT_OFF(nxt1) = nxt2 - nxt1;
5601 OP(nxt) = NOTHING; /* Cannot beautify */
5606 /* Optimize again: */
5607 /* recurse study_chunk() on optimised CURLYX => CURLYM */
5608 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
5609 NULL, stopparen, recursed_depth, NULL, 0,
5615 else if ((OP(oscan) == CURLYX)
5616 && (flags & SCF_WHILEM_VISITED_POS)
5617 /* See the comment on a similar expression above.
5618 However, this time it's not a subexpression
5619 we care about, but the expression itself. */
5620 && (maxcount == REG_INFTY)
5622 /* This stays as CURLYX, we can put the count/of pair. */
5623 /* Find WHILEM (as in regexec.c) */
5624 regnode *nxt = oscan + NEXT_OFF(oscan);
5626 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
5628 nxt = PREVOPER(nxt);
5629 if (nxt->flags & 0xf) {
5630 /* we've already set whilem count on this node */
5631 } else if (++data->whilem_c < 16) {
5632 assert(data->whilem_c <= RExC_whilem_seen);
5633 nxt->flags = (U8)(data->whilem_c
5634 | (RExC_whilem_seen << 4)); /* On WHILEM */
5637 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
5639 if (flags & SCF_DO_SUBSTR) {
5640 SV *last_str = NULL;
5641 STRLEN last_chrs = 0;
5642 int counted = mincount != 0;
5644 if (data->last_end > 0 && mincount != 0) { /* Ends with a
5646 SSize_t b = pos_before >= data->last_start_min
5647 ? pos_before : data->last_start_min;
5649 const char * const s = SvPV_const(data->last_found, l);
5650 SSize_t old = b - data->last_start_min;
5654 old = utf8_hop_forward((U8*)s, old,
5655 (U8 *) SvEND(data->last_found))
5658 /* Get the added string: */
5659 last_str = newSVpvn_utf8(s + old, l, UTF);
5660 last_chrs = UTF ? utf8_length((U8*)(s + old),
5661 (U8*)(s + old + l)) : l;
5662 if (deltanext == 0 && pos_before == b) {
5663 /* What was added is a constant string */
5666 SvGROW(last_str, (mincount * l) + 1);
5667 repeatcpy(SvPVX(last_str) + l,
5668 SvPVX_const(last_str), l,
5670 SvCUR_set(last_str, SvCUR(last_str) * mincount);
5671 /* Add additional parts. */
5672 SvCUR_set(data->last_found,
5673 SvCUR(data->last_found) - l);
5674 sv_catsv(data->last_found, last_str);
5676 SV * sv = data->last_found;
5678 SvUTF8(sv) && SvMAGICAL(sv) ?
5679 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5680 if (mg && mg->mg_len >= 0)
5681 mg->mg_len += last_chrs * (mincount-1);
5683 last_chrs *= mincount;
5684 data->last_end += l * (mincount - 1);
5687 /* start offset must point into the last copy */
5688 data->last_start_min += minnext * (mincount - 1);
5689 data->last_start_max =
5692 : data->last_start_max +
5693 (maxcount - 1) * (minnext + data->pos_delta);
5696 /* It is counted once already... */
5697 data->pos_min += minnext * (mincount - counted);
5699 Perl_re_printf( aTHX_ "counted=%" UVuf " deltanext=%" UVuf
5700 " SSize_t_MAX=%" UVuf " minnext=%" UVuf
5701 " maxcount=%" UVuf " mincount=%" UVuf "\n",
5702 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
5704 if (deltanext != SSize_t_MAX)
5705 Perl_re_printf( aTHX_ "LHS=%" UVuf " RHS=%" UVuf "\n",
5706 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5707 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5709 if (deltanext == SSize_t_MAX
5710 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5711 data->pos_delta = SSize_t_MAX;
5713 data->pos_delta += - counted * deltanext +
5714 (minnext + deltanext) * maxcount - minnext * mincount;
5715 if (mincount != maxcount) {
5716 /* Cannot extend fixed substrings found inside
5718 scan_commit(pRExC_state, data, minlenp, is_inf);
5719 if (mincount && last_str) {
5720 SV * const sv = data->last_found;
5721 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5722 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5726 sv_setsv(sv, last_str);
5727 data->last_end = data->pos_min;
5728 data->last_start_min = data->pos_min - last_chrs;
5729 data->last_start_max = is_inf
5731 : data->pos_min + data->pos_delta - last_chrs;
5733 data->cur_is_floating = 1; /* float */
5735 SvREFCNT_dec(last_str);
5737 if (data && (fl & SF_HAS_EVAL))
5738 data->flags |= SF_HAS_EVAL;
5739 optimize_curly_tail:
5740 if (OP(oscan) != CURLYX) {
5741 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5743 NEXT_OFF(oscan) += NEXT_OFF(next);
5749 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5754 if (flags & SCF_DO_SUBSTR) {
5755 /* Cannot expect anything... */
5756 scan_commit(pRExC_state, data, minlenp, is_inf);
5757 data->cur_is_floating = 1; /* float */
5759 is_inf = is_inf_internal = 1;
5760 if (flags & SCF_DO_STCLASS_OR) {
5761 if (OP(scan) == CLUMP) {
5762 /* Actually is any start char, but very few code points
5763 * aren't start characters */
5764 ssc_match_all_cp(data->start_class);
5767 ssc_anything(data->start_class);
5770 flags &= ~SCF_DO_STCLASS;
5774 else if (OP(scan) == LNBREAK) {
5775 if (flags & SCF_DO_STCLASS) {
5776 if (flags & SCF_DO_STCLASS_AND) {
5777 ssc_intersection(data->start_class,
5778 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5779 ssc_clear_locale(data->start_class);
5780 ANYOF_FLAGS(data->start_class)
5781 &= ~SSC_MATCHES_EMPTY_STRING;
5783 else if (flags & SCF_DO_STCLASS_OR) {
5784 ssc_union(data->start_class,
5785 PL_XPosix_ptrs[_CC_VERTSPACE],
5787 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5789 /* See commit msg for
5790 * 749e076fceedeb708a624933726e7989f2302f6a */
5791 ANYOF_FLAGS(data->start_class)
5792 &= ~SSC_MATCHES_EMPTY_STRING;
5794 flags &= ~SCF_DO_STCLASS;
5797 if (delta != SSize_t_MAX)
5798 delta++; /* Because of the 2 char string cr-lf */
5799 if (flags & SCF_DO_SUBSTR) {
5800 /* Cannot expect anything... */
5801 scan_commit(pRExC_state, data, minlenp, is_inf);
5803 if (data->pos_delta != SSize_t_MAX) {
5804 data->pos_delta += 1;
5806 data->cur_is_floating = 1; /* float */
5809 else if (REGNODE_SIMPLE(OP(scan))) {
5811 if (flags & SCF_DO_SUBSTR) {
5812 scan_commit(pRExC_state, data, minlenp, is_inf);
5816 if (flags & SCF_DO_STCLASS) {
5818 SV* my_invlist = NULL;
5821 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5822 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5824 /* Some of the logic below assumes that switching
5825 locale on will only add false positives. */
5830 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5834 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5835 ssc_match_all_cp(data->start_class);
5840 SV* REG_ANY_invlist = _new_invlist(2);
5841 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5843 if (flags & SCF_DO_STCLASS_OR) {
5844 ssc_union(data->start_class,
5846 TRUE /* TRUE => invert, hence all but \n
5850 else if (flags & SCF_DO_STCLASS_AND) {
5851 ssc_intersection(data->start_class,
5853 TRUE /* TRUE => invert */
5855 ssc_clear_locale(data->start_class);
5857 SvREFCNT_dec_NN(REG_ANY_invlist);
5868 if (flags & SCF_DO_STCLASS_AND)
5869 ssc_and(pRExC_state, data->start_class,
5870 (regnode_charclass *) scan);
5872 ssc_or(pRExC_state, data->start_class,
5873 (regnode_charclass *) scan);
5879 SV* cp_list = get_ANYOFM_contents(scan);
5881 if (flags & SCF_DO_STCLASS_OR) {
5882 ssc_union(data->start_class, cp_list, invert);
5884 else if (flags & SCF_DO_STCLASS_AND) {
5885 ssc_intersection(data->start_class, cp_list, invert);
5888 SvREFCNT_dec_NN(cp_list);
5897 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5898 if (flags & SCF_DO_STCLASS_AND) {
5899 bool was_there = cBOOL(
5900 ANYOF_POSIXL_TEST(data->start_class,
5902 ANYOF_POSIXL_ZERO(data->start_class);
5903 if (was_there) { /* Do an AND */
5904 ANYOF_POSIXL_SET(data->start_class, namedclass);
5906 /* No individual code points can now match */
5907 data->start_class->invlist
5908 = sv_2mortal(_new_invlist(0));
5911 int complement = namedclass + ((invert) ? -1 : 1);
5913 assert(flags & SCF_DO_STCLASS_OR);
5915 /* If the complement of this class was already there,
5916 * the result is that they match all code points,
5917 * (\d + \D == everything). Remove the classes from
5918 * future consideration. Locale is not relevant in
5920 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5921 ssc_match_all_cp(data->start_class);
5922 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5923 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5925 else { /* The usual case; just add this class to the
5927 ANYOF_POSIXL_SET(data->start_class, namedclass);
5932 case NPOSIXA: /* For these, we always know the exact set of
5937 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
5938 goto join_posix_and_ascii;
5946 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
5948 /* NPOSIXD matches all upper Latin1 code points unless the
5949 * target string being matched is UTF-8, which is
5950 * unknowable until match time. Since we are going to
5951 * invert, we want to get rid of all of them so that the
5952 * inversion will match all */
5953 if (OP(scan) == NPOSIXD) {
5954 _invlist_subtract(my_invlist, PL_UpperLatin1,
5958 join_posix_and_ascii:
5960 if (flags & SCF_DO_STCLASS_AND) {
5961 ssc_intersection(data->start_class, my_invlist, invert);
5962 ssc_clear_locale(data->start_class);
5965 assert(flags & SCF_DO_STCLASS_OR);
5966 ssc_union(data->start_class, my_invlist, invert);
5968 SvREFCNT_dec(my_invlist);
5970 if (flags & SCF_DO_STCLASS_OR)
5971 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5972 flags &= ~SCF_DO_STCLASS;
5975 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5976 data->flags |= (OP(scan) == MEOL
5979 scan_commit(pRExC_state, data, minlenp, is_inf);
5982 else if ( PL_regkind[OP(scan)] == BRANCHJ
5983 /* Lookbehind, or need to calculate parens/evals/stclass: */
5984 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5985 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5987 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5988 || OP(scan) == UNLESSM )
5990 /* Negative Lookahead/lookbehind
5991 In this case we can't do fixed string optimisation.
5994 SSize_t deltanext, minnext, fake = 0;
5999 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6001 data_fake.whilem_c = data->whilem_c;
6002 data_fake.last_closep = data->last_closep;
6005 data_fake.last_closep = &fake;
6006 data_fake.pos_delta = delta;
6007 if ( flags & SCF_DO_STCLASS && !scan->flags
6008 && OP(scan) == IFMATCH ) { /* Lookahead */
6009 ssc_init(pRExC_state, &intrnl);
6010 data_fake.start_class = &intrnl;
6011 f |= SCF_DO_STCLASS_AND;
6013 if (flags & SCF_WHILEM_VISITED_POS)
6014 f |= SCF_WHILEM_VISITED_POS;
6015 next = regnext(scan);
6016 nscan = NEXTOPER(NEXTOPER(scan));
6018 /* recurse study_chunk() for lookahead body */
6019 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6020 last, &data_fake, stopparen,
6021 recursed_depth, NULL, f, depth+1);
6024 || deltanext > (I32) U8_MAX
6025 || minnext > (I32)U8_MAX
6026 || minnext + deltanext > (I32)U8_MAX)
6028 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6032 /* The 'next_off' field has been repurposed to count the
6033 * additional starting positions to try beyond the initial
6034 * one. (This leaves it at 0 for non-variable length
6035 * matches to avoid breakage for those not using this
6038 scan->next_off = deltanext;
6039 ckWARNexperimental(RExC_parse,
6040 WARN_EXPERIMENTAL__VLB,
6041 "Variable length lookbehind is experimental");
6043 scan->flags = (U8)minnext + deltanext;
6046 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6048 if (data_fake.flags & SF_HAS_EVAL)
6049 data->flags |= SF_HAS_EVAL;
6050 data->whilem_c = data_fake.whilem_c;
6052 if (f & SCF_DO_STCLASS_AND) {
6053 if (flags & SCF_DO_STCLASS_OR) {
6054 /* OR before, AND after: ideally we would recurse with
6055 * data_fake to get the AND applied by study of the
6056 * remainder of the pattern, and then derecurse;
6057 * *** HACK *** for now just treat as "no information".
6058 * See [perl #56690].
6060 ssc_init(pRExC_state, data->start_class);
6062 /* AND before and after: combine and continue. These
6063 * assertions are zero-length, so can match an EMPTY
6065 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6066 ANYOF_FLAGS(data->start_class)
6067 |= SSC_MATCHES_EMPTY_STRING;
6071 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6073 /* Positive Lookahead/lookbehind
6074 In this case we can do fixed string optimisation,
6075 but we must be careful about it. Note in the case of
6076 lookbehind the positions will be offset by the minimum
6077 length of the pattern, something we won't know about
6078 until after the recurse.
6080 SSize_t deltanext, fake = 0;
6084 /* We use SAVEFREEPV so that when the full compile
6085 is finished perl will clean up the allocated
6086 minlens when it's all done. This way we don't
6087 have to worry about freeing them when we know
6088 they wont be used, which would be a pain.
6091 Newx( minnextp, 1, SSize_t );
6092 SAVEFREEPV(minnextp);
6095 StructCopy(data, &data_fake, scan_data_t);
6096 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6099 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6100 data_fake.last_found=newSVsv(data->last_found);
6104 data_fake.last_closep = &fake;
6105 data_fake.flags = 0;
6106 data_fake.substrs[0].flags = 0;
6107 data_fake.substrs[1].flags = 0;
6108 data_fake.pos_delta = delta;
6110 data_fake.flags |= SF_IS_INF;
6111 if ( flags & SCF_DO_STCLASS && !scan->flags
6112 && OP(scan) == IFMATCH ) { /* Lookahead */
6113 ssc_init(pRExC_state, &intrnl);
6114 data_fake.start_class = &intrnl;
6115 f |= SCF_DO_STCLASS_AND;
6117 if (flags & SCF_WHILEM_VISITED_POS)
6118 f |= SCF_WHILEM_VISITED_POS;
6119 next = regnext(scan);
6120 nscan = NEXTOPER(NEXTOPER(scan));
6122 /* positive lookahead study_chunk() recursion */
6123 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6124 &deltanext, last, &data_fake,
6125 stopparen, recursed_depth, NULL,
6128 assert(0); /* This code has never been tested since this
6129 is normally not compiled */
6131 || deltanext > (I32) U8_MAX
6132 || *minnextp > (I32)U8_MAX
6133 || *minnextp + deltanext > (I32)U8_MAX)
6135 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6140 scan->next_off = deltanext;
6142 scan->flags = (U8)*minnextp + deltanext;
6147 if (f & SCF_DO_STCLASS_AND) {
6148 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6149 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6152 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6154 if (data_fake.flags & SF_HAS_EVAL)
6155 data->flags |= SF_HAS_EVAL;
6156 data->whilem_c = data_fake.whilem_c;
6157 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6159 if (RExC_rx->minlen<*minnextp)
6160 RExC_rx->minlen=*minnextp;
6161 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6162 SvREFCNT_dec_NN(data_fake.last_found);
6164 for (i = 0; i < 2; i++) {
6165 if (data_fake.substrs[i].minlenp != minlenp) {
6166 data->substrs[i].min_offset =
6167 data_fake.substrs[i].min_offset;
6168 data->substrs[i].max_offset =
6169 data_fake.substrs[i].max_offset;
6170 data->substrs[i].minlenp =
6171 data_fake.substrs[i].minlenp;
6172 data->substrs[i].lookbehind += scan->flags;
6180 else if (OP(scan) == OPEN) {
6181 if (stopparen != (I32)ARG(scan))
6184 else if (OP(scan) == CLOSE) {
6185 if (stopparen == (I32)ARG(scan)) {
6188 if ((I32)ARG(scan) == is_par) {
6189 next = regnext(scan);
6191 if ( next && (OP(next) != WHILEM) && next < last)
6192 is_par = 0; /* Disable optimization */
6195 *(data->last_closep) = ARG(scan);
6197 else if (OP(scan) == EVAL) {
6199 data->flags |= SF_HAS_EVAL;
6201 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6202 if (flags & SCF_DO_SUBSTR) {
6203 scan_commit(pRExC_state, data, minlenp, is_inf);
6204 flags &= ~SCF_DO_SUBSTR;
6206 if (data && OP(scan)==ACCEPT) {
6207 data->flags |= SCF_SEEN_ACCEPT;
6212 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6214 if (flags & SCF_DO_SUBSTR) {
6215 scan_commit(pRExC_state, data, minlenp, is_inf);
6216 data->cur_is_floating = 1; /* float */
6218 is_inf = is_inf_internal = 1;
6219 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6220 ssc_anything(data->start_class);
6221 flags &= ~SCF_DO_STCLASS;
6223 else if (OP(scan) == GPOS) {
6224 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6225 !(delta || is_inf || (data && data->pos_delta)))
6227 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6228 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6229 if (RExC_rx->gofs < (STRLEN)min)
6230 RExC_rx->gofs = min;
6232 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6236 #ifdef TRIE_STUDY_OPT
6237 #ifdef FULL_TRIE_STUDY
6238 else if (PL_regkind[OP(scan)] == TRIE) {
6239 /* NOTE - There is similar code to this block above for handling
6240 BRANCH nodes on the initial study. If you change stuff here
6242 regnode *trie_node= scan;
6243 regnode *tail= regnext(scan);
6244 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6245 SSize_t max1 = 0, min1 = SSize_t_MAX;
6248 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6249 /* Cannot merge strings after this. */
6250 scan_commit(pRExC_state, data, minlenp, is_inf);
6252 if (flags & SCF_DO_STCLASS)
6253 ssc_init_zero(pRExC_state, &accum);
6259 const regnode *nextbranch= NULL;
6262 for ( word=1 ; word <= trie->wordcount ; word++)
6264 SSize_t deltanext=0, minnext=0, f = 0, fake;
6265 regnode_ssc this_class;
6267 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6269 data_fake.whilem_c = data->whilem_c;
6270 data_fake.last_closep = data->last_closep;
6273 data_fake.last_closep = &fake;
6274 data_fake.pos_delta = delta;
6275 if (flags & SCF_DO_STCLASS) {
6276 ssc_init(pRExC_state, &this_class);
6277 data_fake.start_class = &this_class;
6278 f = SCF_DO_STCLASS_AND;
6280 if (flags & SCF_WHILEM_VISITED_POS)
6281 f |= SCF_WHILEM_VISITED_POS;
6283 if (trie->jump[word]) {
6285 nextbranch = trie_node + trie->jump[0];
6286 scan= trie_node + trie->jump[word];
6287 /* We go from the jump point to the branch that follows
6288 it. Note this means we need the vestigal unused
6289 branches even though they arent otherwise used. */
6290 /* optimise study_chunk() for TRIE */
6291 minnext = study_chunk(pRExC_state, &scan, minlenp,
6292 &deltanext, (regnode *)nextbranch, &data_fake,
6293 stopparen, recursed_depth, NULL, f, depth+1);
6295 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6296 nextbranch= regnext((regnode*)nextbranch);
6298 if (min1 > (SSize_t)(minnext + trie->minlen))
6299 min1 = minnext + trie->minlen;
6300 if (deltanext == SSize_t_MAX) {
6301 is_inf = is_inf_internal = 1;
6303 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6304 max1 = minnext + deltanext + trie->maxlen;
6306 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6308 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6309 if ( stopmin > min + min1)
6310 stopmin = min + min1;
6311 flags &= ~SCF_DO_SUBSTR;
6313 data->flags |= SCF_SEEN_ACCEPT;
6316 if (data_fake.flags & SF_HAS_EVAL)
6317 data->flags |= SF_HAS_EVAL;
6318 data->whilem_c = data_fake.whilem_c;
6320 if (flags & SCF_DO_STCLASS)
6321 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6324 if (flags & SCF_DO_SUBSTR) {
6325 data->pos_min += min1;
6326 data->pos_delta += max1 - min1;
6327 if (max1 != min1 || is_inf)
6328 data->cur_is_floating = 1; /* float */
6331 if (delta != SSize_t_MAX) {
6332 if (SSize_t_MAX - (max1 - min1) >= delta)
6333 delta += max1 - min1;
6335 delta = SSize_t_MAX;
6337 if (flags & SCF_DO_STCLASS_OR) {
6338 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6340 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6341 flags &= ~SCF_DO_STCLASS;
6344 else if (flags & SCF_DO_STCLASS_AND) {
6346 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6347 flags &= ~SCF_DO_STCLASS;
6350 /* Switch to OR mode: cache the old value of
6351 * data->start_class */
6353 StructCopy(data->start_class, and_withp, regnode_ssc);
6354 flags &= ~SCF_DO_STCLASS_AND;
6355 StructCopy(&accum, data->start_class, regnode_ssc);
6356 flags |= SCF_DO_STCLASS_OR;
6363 else if (PL_regkind[OP(scan)] == TRIE) {
6364 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6367 min += trie->minlen;
6368 delta += (trie->maxlen - trie->minlen);
6369 flags &= ~SCF_DO_STCLASS; /* xxx */
6370 if (flags & SCF_DO_SUBSTR) {
6371 /* Cannot expect anything... */
6372 scan_commit(pRExC_state, data, minlenp, is_inf);
6373 data->pos_min += trie->minlen;
6374 data->pos_delta += (trie->maxlen - trie->minlen);
6375 if (trie->maxlen != trie->minlen)
6376 data->cur_is_floating = 1; /* float */
6378 if (trie->jump) /* no more substrings -- for now /grr*/
6379 flags &= ~SCF_DO_SUBSTR;
6381 #endif /* old or new */
6382 #endif /* TRIE_STUDY_OPT */
6384 /* Else: zero-length, ignore. */
6385 scan = regnext(scan);
6390 /* we need to unwind recursion. */
6393 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6394 DEBUG_PEEP("fend", scan, depth, flags);
6396 /* restore previous context */
6397 last = frame->last_regnode;
6398 scan = frame->next_regnode;
6399 stopparen = frame->stopparen;
6400 recursed_depth = frame->prev_recursed_depth;
6402 RExC_frame_last = frame->prev_frame;
6403 frame = frame->this_prev_frame;
6404 goto fake_study_recurse;
6408 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6411 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6413 if (flags & SCF_DO_SUBSTR && is_inf)
6414 data->pos_delta = SSize_t_MAX - data->pos_min;
6415 if (is_par > (I32)U8_MAX)
6417 if (is_par && pars==1 && data) {
6418 data->flags |= SF_IN_PAR;
6419 data->flags &= ~SF_HAS_PAR;
6421 else if (pars && data) {
6422 data->flags |= SF_HAS_PAR;
6423 data->flags &= ~SF_IN_PAR;
6425 if (flags & SCF_DO_STCLASS_OR)
6426 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6427 if (flags & SCF_TRIE_RESTUDY)
6428 data->flags |= SCF_TRIE_RESTUDY;
6430 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6433 SSize_t final_minlen= min < stopmin ? min : stopmin;
6435 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6436 if (final_minlen > SSize_t_MAX - delta)
6437 RExC_maxlen = SSize_t_MAX;
6438 else if (RExC_maxlen < final_minlen + delta)
6439 RExC_maxlen = final_minlen + delta;
6441 return final_minlen;
6443 NOT_REACHED; /* NOTREACHED */
6447 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6449 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6451 PERL_ARGS_ASSERT_ADD_DATA;
6453 Renewc(RExC_rxi->data,
6454 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6455 char, struct reg_data);
6457 Renew(RExC_rxi->data->what, count + n, U8);
6459 Newx(RExC_rxi->data->what, n, U8);
6460 RExC_rxi->data->count = count + n;
6461 Copy(s, RExC_rxi->data->what + count, n, U8);
6465 /*XXX: todo make this not included in a non debugging perl, but appears to be
6466 * used anyway there, in 'use re' */
6467 #ifndef PERL_IN_XSUB_RE
6469 Perl_reginitcolors(pTHX)
6471 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6473 char *t = savepv(s);
6477 t = strchr(t, '\t');
6483 PL_colors[i] = t = (char *)"";
6488 PL_colors[i++] = (char *)"";
6495 #ifdef TRIE_STUDY_OPT
6496 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6499 (data.flags & SCF_TRIE_RESTUDY) \
6507 #define CHECK_RESTUDY_GOTO_butfirst
6511 * pregcomp - compile a regular expression into internal code
6513 * Decides which engine's compiler to call based on the hint currently in
6517 #ifndef PERL_IN_XSUB_RE
6519 /* return the currently in-scope regex engine (or the default if none) */
6521 regexp_engine const *
6522 Perl_current_re_engine(pTHX)
6524 if (IN_PERL_COMPILETIME) {
6525 HV * const table = GvHV(PL_hintgv);
6528 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6529 return &PL_core_reg_engine;
6530 ptr = hv_fetchs(table, "regcomp", FALSE);
6531 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6532 return &PL_core_reg_engine;
6533 return INT2PTR(regexp_engine*, SvIV(*ptr));
6537 if (!PL_curcop->cop_hints_hash)
6538 return &PL_core_reg_engine;
6539 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6540 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6541 return &PL_core_reg_engine;
6542 return INT2PTR(regexp_engine*, SvIV(ptr));
6548 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6550 regexp_engine const *eng = current_re_engine();
6551 GET_RE_DEBUG_FLAGS_DECL;
6553 PERL_ARGS_ASSERT_PREGCOMP;
6555 /* Dispatch a request to compile a regexp to correct regexp engine. */
6557 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6560 return CALLREGCOMP_ENG(eng, pattern, flags);
6564 /* public(ish) entry point for the perl core's own regex compiling code.
6565 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6566 * pattern rather than a list of OPs, and uses the internal engine rather
6567 * than the current one */
6570 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6572 SV *pat = pattern; /* defeat constness! */
6573 PERL_ARGS_ASSERT_RE_COMPILE;
6574 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6575 #ifdef PERL_IN_XSUB_RE
6578 &PL_core_reg_engine,
6580 NULL, NULL, rx_flags, 0);
6585 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6589 if (--cbs->refcnt > 0)
6591 for (n = 0; n < cbs->count; n++) {
6592 REGEXP *rx = cbs->cb[n].src_regex;
6594 cbs->cb[n].src_regex = NULL;
6595 SvREFCNT_dec_NN(rx);
6603 static struct reg_code_blocks *
6604 S_alloc_code_blocks(pTHX_ int ncode)
6606 struct reg_code_blocks *cbs;
6607 Newx(cbs, 1, struct reg_code_blocks);
6610 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6612 Newx(cbs->cb, ncode, struct reg_code_block);
6619 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6620 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6621 * point to the realloced string and length.
6623 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6627 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6628 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6630 U8 *const src = (U8*)*pat_p;
6635 GET_RE_DEBUG_FLAGS_DECL;
6637 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6638 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6640 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6641 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6644 while (s < *plen_p) {
6645 append_utf8_from_native_byte(src[s], &d);
6647 if (n < num_code_blocks) {
6648 assert(pRExC_state->code_blocks);
6649 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6650 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6651 assert(*(d - 1) == '(');
6654 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6655 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6656 assert(*(d - 1) == ')');
6665 *pat_p = (char*) dst;
6667 RExC_orig_utf8 = RExC_utf8 = 1;
6672 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6673 * while recording any code block indices, and handling overloading,
6674 * nested qr// objects etc. If pat is null, it will allocate a new
6675 * string, or just return the first arg, if there's only one.
6677 * Returns the malloced/updated pat.
6678 * patternp and pat_count is the array of SVs to be concatted;
6679 * oplist is the optional list of ops that generated the SVs;
6680 * recompile_p is a pointer to a boolean that will be set if
6681 * the regex will need to be recompiled.
6682 * delim, if non-null is an SV that will be inserted between each element
6686 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6687 SV *pat, SV ** const patternp, int pat_count,
6688 OP *oplist, bool *recompile_p, SV *delim)
6692 bool use_delim = FALSE;
6693 bool alloced = FALSE;
6695 /* if we know we have at least two args, create an empty string,
6696 * then concatenate args to that. For no args, return an empty string */
6697 if (!pat && pat_count != 1) {
6703 for (svp = patternp; svp < patternp + pat_count; svp++) {
6706 STRLEN orig_patlen = 0;
6708 SV *msv = use_delim ? delim : *svp;
6709 if (!msv) msv = &PL_sv_undef;
6711 /* if we've got a delimiter, we go round the loop twice for each
6712 * svp slot (except the last), using the delimiter the second
6721 if (SvTYPE(msv) == SVt_PVAV) {
6722 /* we've encountered an interpolated array within
6723 * the pattern, e.g. /...@a..../. Expand the list of elements,
6724 * then recursively append elements.
6725 * The code in this block is based on S_pushav() */
6727 AV *const av = (AV*)msv;
6728 const SSize_t maxarg = AvFILL(av) + 1;
6732 assert(oplist->op_type == OP_PADAV
6733 || oplist->op_type == OP_RV2AV);
6734 oplist = OpSIBLING(oplist);
6737 if (SvRMAGICAL(av)) {
6740 Newx(array, maxarg, SV*);
6742 for (i=0; i < maxarg; i++) {
6743 SV ** const svp = av_fetch(av, i, FALSE);
6744 array[i] = svp ? *svp : &PL_sv_undef;
6748 array = AvARRAY(av);
6750 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6751 array, maxarg, NULL, recompile_p,
6753 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6759 /* we make the assumption here that each op in the list of
6760 * op_siblings maps to one SV pushed onto the stack,
6761 * except for code blocks, with have both an OP_NULL and
6763 * This allows us to match up the list of SVs against the
6764 * list of OPs to find the next code block.
6766 * Note that PUSHMARK PADSV PADSV ..
6768 * PADRANGE PADSV PADSV ..
6769 * so the alignment still works. */
6772 if (oplist->op_type == OP_NULL
6773 && (oplist->op_flags & OPf_SPECIAL))
6775 assert(n < pRExC_state->code_blocks->count);
6776 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6777 pRExC_state->code_blocks->cb[n].block = oplist;
6778 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6781 oplist = OpSIBLING(oplist); /* skip CONST */
6784 oplist = OpSIBLING(oplist);;
6787 /* apply magic and QR overloading to arg */
6790 if (SvROK(msv) && SvAMAGIC(msv)) {
6791 SV *sv = AMG_CALLunary(msv, regexp_amg);
6795 if (SvTYPE(sv) != SVt_REGEXP)
6796 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6801 /* try concatenation overload ... */
6802 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6803 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6806 /* overloading involved: all bets are off over literal
6807 * code. Pretend we haven't seen it */
6809 pRExC_state->code_blocks->count -= n;
6813 /* ... or failing that, try "" overload */
6814 while (SvAMAGIC(msv)
6815 && (sv = AMG_CALLunary(msv, string_amg))
6819 && SvRV(msv) == SvRV(sv))
6824 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6828 /* this is a partially unrolled
6829 * sv_catsv_nomg(pat, msv);
6830 * that allows us to adjust code block indices if
6833 char *dst = SvPV_force_nomg(pat, dlen);
6835 if (SvUTF8(msv) && !SvUTF8(pat)) {
6836 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6837 sv_setpvn(pat, dst, dlen);
6840 sv_catsv_nomg(pat, msv);
6844 /* We have only one SV to process, but we need to verify
6845 * it is properly null terminated or we will fail asserts
6846 * later. In theory we probably shouldn't get such SV's,
6847 * but if we do we should handle it gracefully. */
6848 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6849 /* not a string, or a string with a trailing null */
6852 /* a string with no trailing null, we need to copy it
6853 * so it has a trailing null */
6854 pat = sv_2mortal(newSVsv(msv));
6859 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6862 /* extract any code blocks within any embedded qr//'s */
6863 if (rx && SvTYPE(rx) == SVt_REGEXP
6864 && RX_ENGINE((REGEXP*)rx)->op_comp)
6867 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6868 if (ri->code_blocks && ri->code_blocks->count) {
6870 /* the presence of an embedded qr// with code means
6871 * we should always recompile: the text of the
6872 * qr// may not have changed, but it may be a
6873 * different closure than last time */
6875 if (pRExC_state->code_blocks) {
6876 int new_count = pRExC_state->code_blocks->count
6877 + ri->code_blocks->count;
6878 Renew(pRExC_state->code_blocks->cb,
6879 new_count, struct reg_code_block);
6880 pRExC_state->code_blocks->count = new_count;
6883 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6884 ri->code_blocks->count);
6886 for (i=0; i < ri->code_blocks->count; i++) {
6887 struct reg_code_block *src, *dst;
6888 STRLEN offset = orig_patlen
6889 + ReANY((REGEXP *)rx)->pre_prefix;
6890 assert(n < pRExC_state->code_blocks->count);
6891 src = &ri->code_blocks->cb[i];
6892 dst = &pRExC_state->code_blocks->cb[n];
6893 dst->start = src->start + offset;
6894 dst->end = src->end + offset;
6895 dst->block = src->block;
6896 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6905 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6914 /* see if there are any run-time code blocks in the pattern.
6915 * False positives are allowed */
6918 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6919 char *pat, STRLEN plen)
6924 PERL_UNUSED_CONTEXT;
6926 for (s = 0; s < plen; s++) {
6927 if ( pRExC_state->code_blocks
6928 && n < pRExC_state->code_blocks->count
6929 && s == pRExC_state->code_blocks->cb[n].start)
6931 s = pRExC_state->code_blocks->cb[n].end;
6935 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6937 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6939 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6946 /* Handle run-time code blocks. We will already have compiled any direct
6947 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6948 * copy of it, but with any literal code blocks blanked out and
6949 * appropriate chars escaped; then feed it into
6951 * eval "qr'modified_pattern'"
6955 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6959 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6961 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6962 * and merge them with any code blocks of the original regexp.
6964 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6965 * instead, just save the qr and return FALSE; this tells our caller that
6966 * the original pattern needs upgrading to utf8.
6970 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6971 char *pat, STRLEN plen)
6975 GET_RE_DEBUG_FLAGS_DECL;
6977 if (pRExC_state->runtime_code_qr) {
6978 /* this is the second time we've been called; this should
6979 * only happen if the main pattern got upgraded to utf8
6980 * during compilation; re-use the qr we compiled first time
6981 * round (which should be utf8 too)
6983 qr = pRExC_state->runtime_code_qr;
6984 pRExC_state->runtime_code_qr = NULL;
6985 assert(RExC_utf8 && SvUTF8(qr));
6991 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6995 /* determine how many extra chars we need for ' and \ escaping */
6996 for (s = 0; s < plen; s++) {
6997 if (pat[s] == '\'' || pat[s] == '\\')
7001 Newx(newpat, newlen, char);
7003 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7005 for (s = 0; s < plen; s++) {
7006 if ( pRExC_state->code_blocks
7007 && n < pRExC_state->code_blocks->count
7008 && s == pRExC_state->code_blocks->cb[n].start)
7010 /* blank out literal code block so that they aren't
7011 * recompiled: eg change from/to:
7021 assert(pat[s] == '(');
7022 assert(pat[s+1] == '?');
7026 while (s < pRExC_state->code_blocks->cb[n].end) {
7034 if (pat[s] == '\'' || pat[s] == '\\')
7039 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7041 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7047 Perl_re_printf( aTHX_
7048 "%sre-parsing pattern for runtime code:%s %s\n",
7049 PL_colors[4], PL_colors[5], newpat);
7052 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7058 PUSHSTACKi(PERLSI_REQUIRE);
7059 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7060 * parsing qr''; normally only q'' does this. It also alters
7062 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7063 SvREFCNT_dec_NN(sv);
7068 SV * const errsv = ERRSV;
7069 if (SvTRUE_NN(errsv))
7070 /* use croak_sv ? */
7071 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7073 assert(SvROK(qr_ref));
7075 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7076 /* the leaving below frees the tmp qr_ref.
7077 * Give qr a life of its own */
7085 if (!RExC_utf8 && SvUTF8(qr)) {
7086 /* first time through; the pattern got upgraded; save the
7087 * qr for the next time through */
7088 assert(!pRExC_state->runtime_code_qr);
7089 pRExC_state->runtime_code_qr = qr;
7094 /* extract any code blocks within the returned qr// */
7097 /* merge the main (r1) and run-time (r2) code blocks into one */
7099 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7100 struct reg_code_block *new_block, *dst;
7101 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7105 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7107 SvREFCNT_dec_NN(qr);
7111 if (!r1->code_blocks)
7112 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7114 r1c = r1->code_blocks->count;
7115 r2c = r2->code_blocks->count;
7117 Newx(new_block, r1c + r2c, struct reg_code_block);
7121 while (i1 < r1c || i2 < r2c) {
7122 struct reg_code_block *src;
7126 src = &r2->code_blocks->cb[i2++];
7130 src = &r1->code_blocks->cb[i1++];
7131 else if ( r1->code_blocks->cb[i1].start
7132 < r2->code_blocks->cb[i2].start)
7134 src = &r1->code_blocks->cb[i1++];
7135 assert(src->end < r2->code_blocks->cb[i2].start);
7138 assert( r1->code_blocks->cb[i1].start
7139 > r2->code_blocks->cb[i2].start);
7140 src = &r2->code_blocks->cb[i2++];
7142 assert(src->end < r1->code_blocks->cb[i1].start);
7145 assert(pat[src->start] == '(');
7146 assert(pat[src->end] == ')');
7147 dst->start = src->start;
7148 dst->end = src->end;
7149 dst->block = src->block;
7150 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7154 r1->code_blocks->count += r2c;
7155 Safefree(r1->code_blocks->cb);
7156 r1->code_blocks->cb = new_block;
7159 SvREFCNT_dec_NN(qr);
7165 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7166 struct reg_substr_datum *rsd,
7167 struct scan_data_substrs *sub,
7168 STRLEN longest_length)
7170 /* This is the common code for setting up the floating and fixed length
7171 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7172 * as to whether succeeded or not */
7176 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7177 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7179 if (! (longest_length
7180 || (eol /* Can't have SEOL and MULTI */
7181 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7183 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7184 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7189 /* copy the information about the longest from the reg_scan_data
7190 over to the program. */
7191 if (SvUTF8(sub->str)) {
7193 rsd->utf8_substr = sub->str;
7195 rsd->substr = sub->str;
7196 rsd->utf8_substr = NULL;
7198 /* end_shift is how many chars that must be matched that
7199 follow this item. We calculate it ahead of time as once the
7200 lookbehind offset is added in we lose the ability to correctly
7202 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7203 rsd->end_shift = ml - sub->min_offset
7205 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7207 + (SvTAIL(sub->str) != 0)
7211 t = (eol/* Can't have SEOL and MULTI */
7212 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7213 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7219 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7221 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7222 * properly wrapped with the right modifiers */
7224 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7225 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7226 != REGEX_DEPENDS_CHARSET);
7228 /* The caret is output if there are any defaults: if not all the STD
7229 * flags are set, or if no character set specifier is needed */
7231 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7233 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7234 == REG_RUN_ON_COMMENT_SEEN);
7235 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7236 >> RXf_PMf_STD_PMMOD_SHIFT);
7237 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7239 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7241 /* We output all the necessary flags; we never output a minus, as all
7242 * those are defaults, so are
7243 * covered by the caret */
7244 const STRLEN wraplen = pat_len + has_p + has_runon
7245 + has_default /* If needs a caret */
7246 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7248 /* If needs a character set specifier */
7249 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7250 + (sizeof("(?:)") - 1);
7252 PERL_ARGS_ASSERT_SET_REGEX_PV;
7254 /* make sure PL_bitcount bounds not exceeded */
7255 assert(sizeof(STD_PAT_MODS) <= 8);
7257 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7260 SvFLAGS(Rx) |= SVf_UTF8;
7263 /* If a default, cover it using the caret */
7265 *p++= DEFAULT_PAT_MOD;
7271 name = get_regex_charset_name(RExC_rx->extflags, &len);
7272 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7274 name = UNICODE_PAT_MODS;
7275 len = sizeof(UNICODE_PAT_MODS) - 1;
7277 Copy(name, p, len, char);
7281 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7284 while((ch = *fptr++)) {
7292 Copy(RExC_precomp, p, pat_len, char);
7293 assert ((RX_WRAPPED(Rx) - p) < 16);
7294 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7297 /* Adding a trailing \n causes this to compile properly:
7298 my $R = qr / A B C # D E/x; /($R)/
7299 Otherwise the parens are considered part of the comment */
7304 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7308 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7309 * regular expression into internal code.
7310 * The pattern may be passed either as:
7311 * a list of SVs (patternp plus pat_count)
7312 * a list of OPs (expr)
7313 * If both are passed, the SV list is used, but the OP list indicates
7314 * which SVs are actually pre-compiled code blocks
7316 * The SVs in the list have magic and qr overloading applied to them (and
7317 * the list may be modified in-place with replacement SVs in the latter
7320 * If the pattern hasn't changed from old_re, then old_re will be
7323 * eng is the current engine. If that engine has an op_comp method, then
7324 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7325 * do the initial concatenation of arguments and pass on to the external
7328 * If is_bare_re is not null, set it to a boolean indicating whether the
7329 * arg list reduced (after overloading) to a single bare regex which has
7330 * been returned (i.e. /$qr/).
7332 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7334 * pm_flags contains the PMf_* flags, typically based on those from the
7335 * pm_flags field of the related PMOP. Currently we're only interested in
7336 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
7338 * For many years this code had an initial sizing pass that calculated
7339 * (sometimes incorrectly, leading to security holes) the size needed for the
7340 * compiled pattern. That was changed by commit
7341 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7342 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7343 * references to this sizing pass.
7345 * Now, an initial crude guess as to the size needed is made, based on the
7346 * length of the pattern. Patches welcome to improve that guess. That amount
7347 * of space is malloc'd and then immediately freed, and then clawed back node
7348 * by node. This design is to minimze, to the extent possible, memory churn
7349 * when doing the the reallocs.
7351 * A separate parentheses counting pass may be needed in some cases.
7352 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7355 * The existence of a sizing pass necessitated design decisions that are no
7356 * longer needed. There are potential areas of simplification.
7358 * Beware that the optimization-preparation code in here knows about some
7359 * of the structure of the compiled regexp. [I'll say.]
7363 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7364 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7365 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7368 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7376 SV** new_patternp = patternp;
7378 /* these are all flags - maybe they should be turned
7379 * into a single int with different bit masks */
7380 I32 sawlookahead = 0;
7385 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7387 bool runtime_code = 0;
7389 RExC_state_t RExC_state;
7390 RExC_state_t * const pRExC_state = &RExC_state;
7391 #ifdef TRIE_STUDY_OPT
7393 RExC_state_t copyRExC_state;
7395 GET_RE_DEBUG_FLAGS_DECL;
7397 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7399 DEBUG_r(if (!PL_colorset) reginitcolors());
7401 /* Initialize these here instead of as-needed, as is quick and avoids
7402 * having to test them each time otherwise */
7403 if (! PL_InBitmap) {
7405 char * dump_len_string;
7408 /* This is calculated here, because the Perl program that generates the
7409 * static global ones doesn't currently have access to
7410 * NUM_ANYOF_CODE_POINTS */
7411 PL_InBitmap = _new_invlist(2);
7412 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
7413 NUM_ANYOF_CODE_POINTS - 1);
7415 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
7416 if ( ! dump_len_string
7417 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
7419 PL_dump_re_max_len = 60; /* A reasonable default */
7424 pRExC_state->warn_text = NULL;
7425 pRExC_state->unlexed_names = NULL;
7426 pRExC_state->code_blocks = NULL;
7429 *is_bare_re = FALSE;
7431 if (expr && (expr->op_type == OP_LIST ||
7432 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7433 /* allocate code_blocks if needed */
7437 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7438 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7439 ncode++; /* count of DO blocks */
7442 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7446 /* compile-time pattern with just OP_CONSTs and DO blocks */
7451 /* find how many CONSTs there are */
7454 if (expr->op_type == OP_CONST)
7457 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7458 if (o->op_type == OP_CONST)
7462 /* fake up an SV array */
7464 assert(!new_patternp);
7465 Newx(new_patternp, n, SV*);
7466 SAVEFREEPV(new_patternp);
7470 if (expr->op_type == OP_CONST)
7471 new_patternp[n] = cSVOPx_sv(expr);
7473 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7474 if (o->op_type == OP_CONST)
7475 new_patternp[n++] = cSVOPo_sv;
7480 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7481 "Assembling pattern from %d elements%s\n", pat_count,
7482 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7484 /* set expr to the first arg op */
7486 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7487 && expr->op_type != OP_CONST)
7489 expr = cLISTOPx(expr)->op_first;
7490 assert( expr->op_type == OP_PUSHMARK
7491 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7492 || expr->op_type == OP_PADRANGE);
7493 expr = OpSIBLING(expr);
7496 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7497 expr, &recompile, NULL);
7499 /* handle bare (possibly after overloading) regex: foo =~ $re */
7504 if (SvTYPE(re) == SVt_REGEXP) {
7508 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7509 "Precompiled pattern%s\n",
7510 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7516 exp = SvPV_nomg(pat, plen);
7518 if (!eng->op_comp) {
7519 if ((SvUTF8(pat) && IN_BYTES)
7520 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7522 /* make a temporary copy; either to convert to bytes,
7523 * or to avoid repeating get-magic / overloaded stringify */
7524 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7525 (IN_BYTES ? 0 : SvUTF8(pat)));
7527 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7530 /* ignore the utf8ness if the pattern is 0 length */
7531 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7532 RExC_uni_semantics = 0;
7533 RExC_contains_locale = 0;
7534 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7535 RExC_in_script_run = 0;
7536 RExC_study_started = 0;
7537 pRExC_state->runtime_code_qr = NULL;
7538 RExC_frame_head= NULL;
7539 RExC_frame_last= NULL;
7540 RExC_frame_count= 0;
7541 RExC_latest_warn_offset = 0;
7542 RExC_use_BRANCHJ = 0;
7543 RExC_total_parens = 0;
7544 RExC_open_parens = NULL;
7545 RExC_close_parens = NULL;
7546 RExC_paren_names = NULL;
7548 RExC_seen_d_op = FALSE;
7550 RExC_paren_name_list = NULL;
7554 RExC_mysv1= sv_newmortal();
7555 RExC_mysv2= sv_newmortal();
7559 SV *dsv= sv_newmortal();
7560 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7561 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7562 PL_colors[4], PL_colors[5], s);
7565 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7568 if ((pm_flags & PMf_USE_RE_EVAL)
7569 /* this second condition covers the non-regex literal case,
7570 * i.e. $foo =~ '(?{})'. */
7571 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7573 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7576 /* return old regex if pattern hasn't changed */
7577 /* XXX: note in the below we have to check the flags as well as the
7580 * Things get a touch tricky as we have to compare the utf8 flag
7581 * independently from the compile flags. */
7585 && !!RX_UTF8(old_re) == !!RExC_utf8
7586 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7587 && RX_PRECOMP(old_re)
7588 && RX_PRELEN(old_re) == plen
7589 && memEQ(RX_PRECOMP(old_re), exp, plen)
7590 && !runtime_code /* with runtime code, always recompile */ )
7593 SV *dsv= sv_newmortal();
7594 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7595 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7596 PL_colors[4], PL_colors[5], s);
7601 /* Allocate the pattern's SV */
7602 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7603 RExC_rx = ReANY(Rx);
7604 if ( RExC_rx == NULL )
7605 FAIL("Regexp out of space");
7607 rx_flags = orig_rx_flags;
7609 if ( (UTF || RExC_uni_semantics)
7610 && initial_charset == REGEX_DEPENDS_CHARSET)
7613 /* Set to use unicode semantics if the pattern is in utf8 and has the
7614 * 'depends' charset specified, as it means unicode when utf8 */
7615 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7616 RExC_uni_semantics = 1;
7619 RExC_pm_flags = pm_flags;
7622 assert(TAINTING_get || !TAINT_get);
7624 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7626 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7627 /* whoops, we have a non-utf8 pattern, whilst run-time code
7628 * got compiled as utf8. Try again with a utf8 pattern */
7629 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7630 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7634 assert(!pRExC_state->runtime_code_qr);
7640 RExC_in_lookbehind = 0;
7641 RExC_in_lookahead = 0;
7642 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7643 RExC_recode_x_to_native = 0;
7644 RExC_in_multi_char_class = 0;
7646 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7647 RExC_precomp_end = RExC_end = exp + plen;
7649 RExC_whilem_seen = 0;
7651 RExC_recurse = NULL;
7652 RExC_study_chunk_recursed = NULL;
7653 RExC_study_chunk_recursed_bytes= 0;
7654 RExC_recurse_count = 0;
7655 pRExC_state->code_index = 0;
7657 /* Initialize the string in the compiled pattern. This is so that there is
7658 * something to output if necessary */
7659 set_regex_pv(pRExC_state, Rx);
7662 Perl_re_printf( aTHX_
7663 "Starting parse and generation\n");
7665 RExC_lastparse=NULL;
7668 /* Allocate space and zero-initialize. Note, the two step process
7669 of zeroing when in debug mode, thus anything assigned has to
7670 happen after that */
7673 /* On the first pass of the parse, we guess how big this will be. Then
7674 * we grow in one operation to that amount and then give it back. As
7675 * we go along, we re-allocate what we need.
7677 * XXX Currently the guess is essentially that the pattern will be an
7678 * EXACT node with one byte input, one byte output. This is crude, and
7679 * better heuristics are welcome.
7681 * On any subsequent passes, we guess what we actually computed in the
7682 * latest earlier pass. Such a pass probably didn't complete so is
7683 * missing stuff. We could improve those guesses by knowing where the
7684 * parse stopped, and use the length so far plus apply the above
7685 * assumption to what's left. */
7686 RExC_size = STR_SZ(RExC_end - RExC_start);
7689 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7690 if ( RExC_rxi == NULL )
7691 FAIL("Regexp out of space");
7693 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7694 RXi_SET( RExC_rx, RExC_rxi );
7696 /* We start from 0 (over from 0 in the case this is a reparse. The first
7697 * node parsed will give back any excess memory we have allocated so far).
7701 /* non-zero initialization begins here */
7702 RExC_rx->engine= eng;
7703 RExC_rx->extflags = rx_flags;
7704 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7706 if (pm_flags & PMf_IS_QR) {
7707 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7708 if (RExC_rxi->code_blocks) {
7709 RExC_rxi->code_blocks->refcnt++;
7713 RExC_rx->intflags = 0;
7715 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7718 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7719 * code makes sure the final byte is an uncounted NUL. But should this
7720 * ever not be the case, lots of things could read beyond the end of the
7721 * buffer: loops like
7722 * while(isFOO(*RExC_parse)) RExC_parse++;
7723 * strchr(RExC_parse, "foo");
7724 * etc. So it is worth noting. */
7725 assert(*RExC_end == '\0');
7729 RExC_parens_buf_size = 0;
7730 RExC_emit_start = RExC_rxi->program;
7731 pRExC_state->code_index = 0;
7733 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7737 if (reg(pRExC_state, 0, &flags, 1)) {
7739 /* Success!, But we may need to redo the parse knowing how many parens
7740 * there actually are */
7741 if (IN_PARENS_PASS) {
7742 flags |= RESTART_PARSE;
7745 /* We have that number in RExC_npar */
7746 RExC_total_parens = RExC_npar;
7748 else if (! MUST_RESTART(flags)) {
7750 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7753 /* Here, we either have success, or we have to redo the parse for some reason */
7754 if (MUST_RESTART(flags)) {
7756 /* It's possible to write a regexp in ascii that represents Unicode
7757 codepoints outside of the byte range, such as via \x{100}. If we
7758 detect such a sequence we have to convert the entire pattern to utf8
7759 and then recompile, as our sizing calculation will have been based
7760 on 1 byte == 1 character, but we will need to use utf8 to encode
7761 at least some part of the pattern, and therefore must convert the whole
7764 if (flags & NEED_UTF8) {
7766 /* We have stored the offset of the final warning output so far.
7767 * That must be adjusted. Any variant characters between the start
7768 * of the pattern and this warning count for 2 bytes in the final,
7769 * so just add them again */
7770 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7771 RExC_latest_warn_offset +=
7772 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7773 + RExC_latest_warn_offset);
7775 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7776 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7777 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7780 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7783 if (ALL_PARENS_COUNTED) {
7784 /* Make enough room for all the known parens, and zero it */
7785 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7786 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7787 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7789 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7790 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7792 else { /* Parse did not complete. Reinitialize the parentheses
7794 RExC_total_parens = 0;
7795 if (RExC_open_parens) {
7796 Safefree(RExC_open_parens);
7797 RExC_open_parens = NULL;
7799 if (RExC_close_parens) {
7800 Safefree(RExC_close_parens);
7801 RExC_close_parens = NULL;
7805 /* Clean up what we did in this parse */
7806 SvREFCNT_dec_NN(RExC_rx_sv);
7811 /* Here, we have successfully parsed and generated the pattern's program
7812 * for the regex engine. We are ready to finish things up and look for
7815 /* Update the string to compile, with correct modifiers, etc */
7816 set_regex_pv(pRExC_state, Rx);
7818 RExC_rx->nparens = RExC_total_parens - 1;
7820 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7821 if (RExC_whilem_seen > 15)
7822 RExC_whilem_seen = 15;
7825 Perl_re_printf( aTHX_
7826 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7828 RExC_lastparse=NULL;
7831 #ifdef RE_TRACK_PATTERN_OFFSETS
7832 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7833 "%s %" UVuf " bytes for offset annotations.\n",
7834 RExC_offsets ? "Got" : "Couldn't get",
7835 (UV)((RExC_offsets[0] * 2 + 1))));
7836 DEBUG_OFFSETS_r(if (RExC_offsets) {
7837 const STRLEN len = RExC_offsets[0];
7839 GET_RE_DEBUG_FLAGS_DECL;
7840 Perl_re_printf( aTHX_
7841 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7842 for (i = 1; i <= len; i++) {
7843 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7844 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7845 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7847 Perl_re_printf( aTHX_ "\n");
7851 SetProgLen(RExC_rxi,RExC_size);
7854 DEBUG_DUMP_PRE_OPTIMIZE_r({
7855 SV * const sv = sv_newmortal();
7856 RXi_GET_DECL(RExC_rx, ri);
7858 Perl_re_printf( aTHX_ "Program before optimization:\n");
7860 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
7865 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7868 /* XXXX To minimize changes to RE engine we always allocate
7869 3-units-long substrs field. */
7870 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7871 if (RExC_recurse_count) {
7872 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7873 SAVEFREEPV(RExC_recurse);
7876 if (RExC_seen & REG_RECURSE_SEEN) {
7877 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
7878 * So its 1 if there are no parens. */
7879 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
7880 ((RExC_total_parens & 0x07) != 0);
7881 Newx(RExC_study_chunk_recursed,
7882 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7883 SAVEFREEPV(RExC_study_chunk_recursed);
7887 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7889 RExC_study_chunk_recursed_count= 0;
7891 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
7892 if (RExC_study_chunk_recursed) {
7893 Zero(RExC_study_chunk_recursed,
7894 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7898 #ifdef TRIE_STUDY_OPT
7900 StructCopy(&zero_scan_data, &data, scan_data_t);
7901 copyRExC_state = RExC_state;
7904 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7906 RExC_state = copyRExC_state;
7907 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7908 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7910 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7911 StructCopy(&zero_scan_data, &data, scan_data_t);
7914 StructCopy(&zero_scan_data, &data, scan_data_t);
7917 /* Dig out information for optimizations. */
7918 RExC_rx->extflags = RExC_flags; /* was pm_op */
7919 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7922 SvUTF8_on(Rx); /* Unicode in it? */
7923 RExC_rxi->regstclass = NULL;
7924 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7925 RExC_rx->intflags |= PREGf_NAUGHTY;
7926 scan = RExC_rxi->program + 1; /* First BRANCH. */
7928 /* testing for BRANCH here tells us whether there is "must appear"
7929 data in the pattern. If there is then we can use it for optimisations */
7930 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7933 STRLEN longest_length[2];
7934 regnode_ssc ch_class; /* pointed to by data */
7936 SSize_t last_close = 0; /* pointed to by data */
7937 regnode *first= scan;
7938 regnode *first_next= regnext(first);
7942 * Skip introductions and multiplicators >= 1
7943 * so that we can extract the 'meat' of the pattern that must
7944 * match in the large if() sequence following.
7945 * NOTE that EXACT is NOT covered here, as it is normally
7946 * picked up by the optimiser separately.
7948 * This is unfortunate as the optimiser isnt handling lookahead
7949 * properly currently.
7952 while ((OP(first) == OPEN && (sawopen = 1)) ||
7953 /* An OR of *one* alternative - should not happen now. */
7954 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7955 /* for now we can't handle lookbehind IFMATCH*/
7956 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7957 (OP(first) == PLUS) ||
7958 (OP(first) == MINMOD) ||
7959 /* An {n,m} with n>0 */
7960 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7961 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7964 * the only op that could be a regnode is PLUS, all the rest
7965 * will be regnode_1 or regnode_2.
7967 * (yves doesn't think this is true)
7969 if (OP(first) == PLUS)
7972 if (OP(first) == MINMOD)
7974 first += regarglen[OP(first)];
7976 first = NEXTOPER(first);
7977 first_next= regnext(first);
7980 /* Starting-point info. */
7982 DEBUG_PEEP("first:", first, 0, 0);
7983 /* Ignore EXACT as we deal with it later. */
7984 if (PL_regkind[OP(first)] == EXACT) {
7985 if ( OP(first) == EXACT
7986 || OP(first) == LEXACT
7987 || OP(first) == EXACT_REQ8
7988 || OP(first) == LEXACT_REQ8
7989 || OP(first) == EXACTL)
7991 NOOP; /* Empty, get anchored substr later. */
7994 RExC_rxi->regstclass = first;
7997 else if (PL_regkind[OP(first)] == TRIE &&
7998 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8000 /* this can happen only on restudy */
8001 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8004 else if (REGNODE_SIMPLE(OP(first)))
8005 RExC_rxi->regstclass = first;
8006 else if (PL_regkind[OP(first)] == BOUND ||
8007 PL_regkind[OP(first)] == NBOUND)
8008 RExC_rxi->regstclass = first;
8009 else if (PL_regkind[OP(first)] == BOL) {
8010 RExC_rx->intflags |= (OP(first) == MBOL
8013 first = NEXTOPER(first);
8016 else if (OP(first) == GPOS) {
8017 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8018 first = NEXTOPER(first);
8021 else if ((!sawopen || !RExC_sawback) &&
8023 (OP(first) == STAR &&
8024 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8025 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8027 /* turn .* into ^.* with an implied $*=1 */
8029 (OP(NEXTOPER(first)) == REG_ANY)
8032 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8033 first = NEXTOPER(first);
8036 if (sawplus && !sawminmod && !sawlookahead
8037 && (!sawopen || !RExC_sawback)
8038 && !pRExC_state->code_blocks) /* May examine pos and $& */
8039 /* x+ must match at the 1st pos of run of x's */
8040 RExC_rx->intflags |= PREGf_SKIP;
8042 /* Scan is after the zeroth branch, first is atomic matcher. */
8043 #ifdef TRIE_STUDY_OPT
8046 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8047 (IV)(first - scan + 1))
8051 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8052 (IV)(first - scan + 1))
8058 * If there's something expensive in the r.e., find the
8059 * longest literal string that must appear and make it the
8060 * regmust. Resolve ties in favor of later strings, since
8061 * the regstart check works with the beginning of the r.e.
8062 * and avoiding duplication strengthens checking. Not a
8063 * strong reason, but sufficient in the absence of others.
8064 * [Now we resolve ties in favor of the earlier string if
8065 * it happens that c_offset_min has been invalidated, since the
8066 * earlier string may buy us something the later one won't.]
8069 data.substrs[0].str = newSVpvs("");
8070 data.substrs[1].str = newSVpvs("");
8071 data.last_found = newSVpvs("");
8072 data.cur_is_floating = 0; /* initially any found substring is fixed */
8073 ENTER_with_name("study_chunk");
8074 SAVEFREESV(data.substrs[0].str);
8075 SAVEFREESV(data.substrs[1].str);
8076 SAVEFREESV(data.last_found);
8078 if (!RExC_rxi->regstclass) {
8079 ssc_init(pRExC_state, &ch_class);
8080 data.start_class = &ch_class;
8081 stclass_flag = SCF_DO_STCLASS_AND;
8082 } else /* XXXX Check for BOUND? */
8084 data.last_closep = &last_close;
8088 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8089 * (NO top level branches)
8091 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8092 scan + RExC_size, /* Up to end */
8094 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8095 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8099 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8102 if ( RExC_total_parens == 1 && !data.cur_is_floating
8103 && data.last_start_min == 0 && data.last_end > 0
8104 && !RExC_seen_zerolen
8105 && !(RExC_seen & REG_VERBARG_SEEN)
8106 && !(RExC_seen & REG_GPOS_SEEN)
8108 RExC_rx->extflags |= RXf_CHECK_ALL;
8110 scan_commit(pRExC_state, &data,&minlen, 0);
8113 /* XXX this is done in reverse order because that's the way the
8114 * code was before it was parameterised. Don't know whether it
8115 * actually needs doing in reverse order. DAPM */
8116 for (i = 1; i >= 0; i--) {
8117 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8120 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8121 && data.substrs[0].min_offset
8122 == data.substrs[1].min_offset
8123 && SvCUR(data.substrs[0].str)
8124 == SvCUR(data.substrs[1].str)
8126 && S_setup_longest (aTHX_ pRExC_state,
8127 &(RExC_rx->substrs->data[i]),
8131 RExC_rx->substrs->data[i].min_offset =
8132 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8134 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8135 /* Don't offset infinity */
8136 if (data.substrs[i].max_offset < SSize_t_MAX)
8137 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8138 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8141 RExC_rx->substrs->data[i].substr = NULL;
8142 RExC_rx->substrs->data[i].utf8_substr = NULL;
8143 longest_length[i] = 0;
8147 LEAVE_with_name("study_chunk");
8149 if (RExC_rxi->regstclass
8150 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8151 RExC_rxi->regstclass = NULL;
8153 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8154 || RExC_rx->substrs->data[0].min_offset)
8156 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8157 && is_ssc_worth_it(pRExC_state, data.start_class))
8159 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8161 ssc_finalize(pRExC_state, data.start_class);
8163 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8164 StructCopy(data.start_class,
8165 (regnode_ssc*)RExC_rxi->data->data[n],
8167 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8168 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8169 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8170 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8171 Perl_re_printf( aTHX_
8172 "synthetic stclass \"%s\".\n",
8173 SvPVX_const(sv));});
8174 data.start_class = NULL;
8177 /* A temporary algorithm prefers floated substr to fixed one of
8178 * same length to dig more info. */
8179 i = (longest_length[0] <= longest_length[1]);
8180 RExC_rx->substrs->check_ix = i;
8181 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8182 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8183 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8184 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8185 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8186 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8187 RExC_rx->intflags |= PREGf_NOSCAN;
8189 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8190 RExC_rx->extflags |= RXf_USE_INTUIT;
8191 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8192 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8195 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8196 if ( (STRLEN)minlen < longest_length[1] )
8197 minlen= longest_length[1];
8198 if ( (STRLEN)minlen < longest_length[0] )
8199 minlen= longest_length[0];
8203 /* Several toplevels. Best we can is to set minlen. */
8205 regnode_ssc ch_class;
8206 SSize_t last_close = 0;
8208 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8210 scan = RExC_rxi->program + 1;
8211 ssc_init(pRExC_state, &ch_class);
8212 data.start_class = &ch_class;
8213 data.last_closep = &last_close;
8217 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8218 * (patterns WITH top level branches)
8220 minlen = study_chunk(pRExC_state,
8221 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8222 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8223 ? SCF_TRIE_DOING_RESTUDY
8227 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8229 RExC_rx->check_substr = NULL;
8230 RExC_rx->check_utf8 = NULL;
8231 RExC_rx->substrs->data[0].substr = NULL;
8232 RExC_rx->substrs->data[0].utf8_substr = NULL;
8233 RExC_rx->substrs->data[1].substr = NULL;
8234 RExC_rx->substrs->data[1].utf8_substr = NULL;
8236 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8237 && is_ssc_worth_it(pRExC_state, data.start_class))
8239 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8241 ssc_finalize(pRExC_state, data.start_class);
8243 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8244 StructCopy(data.start_class,
8245 (regnode_ssc*)RExC_rxi->data->data[n],
8247 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8248 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8249 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8250 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8251 Perl_re_printf( aTHX_
8252 "synthetic stclass \"%s\".\n",
8253 SvPVX_const(sv));});
8254 data.start_class = NULL;
8258 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8259 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8260 RExC_rx->maxlen = REG_INFTY;
8263 RExC_rx->maxlen = RExC_maxlen;
8266 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8267 the "real" pattern. */
8269 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8270 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8272 RExC_rx->minlenret = minlen;
8273 if (RExC_rx->minlen < minlen)
8274 RExC_rx->minlen = minlen;
8276 if (RExC_seen & REG_RECURSE_SEEN ) {
8277 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8278 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8280 if (RExC_seen & REG_GPOS_SEEN)
8281 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8282 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8283 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8285 if (pRExC_state->code_blocks)
8286 RExC_rx->extflags |= RXf_EVAL_SEEN;
8287 if (RExC_seen & REG_VERBARG_SEEN)
8289 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8290 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8292 if (RExC_seen & REG_CUTGROUP_SEEN)
8293 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8294 if (pm_flags & PMf_USE_RE_EVAL)
8295 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8296 if (RExC_paren_names)
8297 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8299 RXp_PAREN_NAMES(RExC_rx) = NULL;
8301 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8302 * so it can be used in pp.c */
8303 if (RExC_rx->intflags & PREGf_ANCH)
8304 RExC_rx->extflags |= RXf_IS_ANCHORED;
8308 /* this is used to identify "special" patterns that might result
8309 * in Perl NOT calling the regex engine and instead doing the match "itself",
8310 * particularly special cases in split//. By having the regex compiler
8311 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8312 * we avoid weird issues with equivalent patterns resulting in different behavior,
8313 * AND we allow non Perl engines to get the same optimizations by the setting the
8314 * flags appropriately - Yves */
8315 regnode *first = RExC_rxi->program + 1;
8317 regnode *next = regnext(first);
8320 if (PL_regkind[fop] == NOTHING && nop == END)
8321 RExC_rx->extflags |= RXf_NULL;
8322 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8323 /* when fop is SBOL first->flags will be true only when it was
8324 * produced by parsing /\A/, and not when parsing /^/. This is
8325 * very important for the split code as there we want to
8326 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8327 * See rt #122761 for more details. -- Yves */
8328 RExC_rx->extflags |= RXf_START_ONLY;
8329 else if (fop == PLUS
8330 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8332 RExC_rx->extflags |= RXf_WHITE;
8333 else if ( RExC_rx->extflags & RXf_SPLIT
8334 && ( fop == EXACT || fop == LEXACT
8335 || fop == EXACT_REQ8 || fop == LEXACT_REQ8
8337 && STR_LEN(first) == 1
8338 && *(STRING(first)) == ' '
8340 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8344 if (RExC_contains_locale) {
8345 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8349 if (RExC_paren_names) {
8350 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8351 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8352 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8355 RExC_rxi->name_list_idx = 0;
8357 while ( RExC_recurse_count > 0 ) {
8358 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8360 * This data structure is set up in study_chunk() and is used
8361 * to calculate the distance between a GOSUB regopcode and
8362 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8365 * If for some reason someone writes code that optimises
8366 * away a GOSUB opcode then the assert should be changed to
8367 * an if(scan) to guard the ARG2L_SET() - Yves
8370 assert(scan && OP(scan) == GOSUB);
8371 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8374 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8375 /* assume we don't need to swap parens around before we match */
8377 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8378 (unsigned long)RExC_study_chunk_recursed_count);
8382 Perl_re_printf( aTHX_ "Final program:\n");
8386 if (RExC_open_parens) {
8387 Safefree(RExC_open_parens);
8388 RExC_open_parens = NULL;
8390 if (RExC_close_parens) {
8391 Safefree(RExC_close_parens);
8392 RExC_close_parens = NULL;
8396 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8397 * by setting the regexp SV to readonly-only instead. If the
8398 * pattern's been recompiled, the USEDness should remain. */
8399 if (old_re && SvREADONLY(old_re))
8407 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8410 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8412 PERL_UNUSED_ARG(value);
8414 if (flags & RXapif_FETCH) {
8415 return reg_named_buff_fetch(rx, key, flags);
8416 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8417 Perl_croak_no_modify();
8419 } else if (flags & RXapif_EXISTS) {
8420 return reg_named_buff_exists(rx, key, flags)
8423 } else if (flags & RXapif_REGNAMES) {
8424 return reg_named_buff_all(rx, flags);
8425 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8426 return reg_named_buff_scalar(rx, flags);
8428 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8434 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8437 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8438 PERL_UNUSED_ARG(lastkey);
8440 if (flags & RXapif_FIRSTKEY)
8441 return reg_named_buff_firstkey(rx, flags);
8442 else if (flags & RXapif_NEXTKEY)
8443 return reg_named_buff_nextkey(rx, flags);
8445 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8452 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8456 struct regexp *const rx = ReANY(r);
8458 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8460 if (rx && RXp_PAREN_NAMES(rx)) {
8461 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8464 SV* sv_dat=HeVAL(he_str);
8465 I32 *nums=(I32*)SvPVX(sv_dat);
8466 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8467 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8468 if ((I32)(rx->nparens) >= nums[i]
8469 && rx->offs[nums[i]].start != -1
8470 && rx->offs[nums[i]].end != -1)
8473 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8478 ret = newSVsv(&PL_sv_undef);
8481 av_push(retarray, ret);
8484 return newRV_noinc(MUTABLE_SV(retarray));
8491 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8494 struct regexp *const rx = ReANY(r);
8496 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8498 if (rx && RXp_PAREN_NAMES(rx)) {
8499 if (flags & RXapif_ALL) {
8500 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8502 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8504 SvREFCNT_dec_NN(sv);
8516 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8518 struct regexp *const rx = ReANY(r);
8520 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8522 if ( rx && RXp_PAREN_NAMES(rx) ) {
8523 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8525 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8532 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8534 struct regexp *const rx = ReANY(r);
8535 GET_RE_DEBUG_FLAGS_DECL;
8537 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8539 if (rx && RXp_PAREN_NAMES(rx)) {
8540 HV *hv = RXp_PAREN_NAMES(rx);
8542 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8545 SV* sv_dat = HeVAL(temphe);
8546 I32 *nums = (I32*)SvPVX(sv_dat);
8547 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8548 if ((I32)(rx->lastparen) >= nums[i] &&
8549 rx->offs[nums[i]].start != -1 &&
8550 rx->offs[nums[i]].end != -1)
8556 if (parno || flags & RXapif_ALL) {
8557 return newSVhek(HeKEY_hek(temphe));
8565 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8570 struct regexp *const rx = ReANY(r);
8572 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8574 if (rx && RXp_PAREN_NAMES(rx)) {
8575 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8576 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8577 } else if (flags & RXapif_ONE) {
8578 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8579 av = MUTABLE_AV(SvRV(ret));
8580 length = av_tindex(av);
8581 SvREFCNT_dec_NN(ret);
8582 return newSViv(length + 1);
8584 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8589 return &PL_sv_undef;
8593 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8595 struct regexp *const rx = ReANY(r);
8598 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8600 if (rx && RXp_PAREN_NAMES(rx)) {
8601 HV *hv= RXp_PAREN_NAMES(rx);
8603 (void)hv_iterinit(hv);
8604 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8607 SV* sv_dat = HeVAL(temphe);
8608 I32 *nums = (I32*)SvPVX(sv_dat);
8609 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8610 if ((I32)(rx->lastparen) >= nums[i] &&
8611 rx->offs[nums[i]].start != -1 &&
8612 rx->offs[nums[i]].end != -1)
8618 if (parno || flags & RXapif_ALL) {
8619 av_push(av, newSVhek(HeKEY_hek(temphe)));
8624 return newRV_noinc(MUTABLE_SV(av));
8628 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8631 struct regexp *const rx = ReANY(r);
8637 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8639 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8640 || n == RX_BUFF_IDX_CARET_FULLMATCH
8641 || n == RX_BUFF_IDX_CARET_POSTMATCH
8644 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8646 /* on something like
8649 * the KEEPCOPY is set on the PMOP rather than the regex */
8650 if (PL_curpm && r == PM_GETRE(PL_curpm))
8651 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8660 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8661 /* no need to distinguish between them any more */
8662 n = RX_BUFF_IDX_FULLMATCH;
8664 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8665 && rx->offs[0].start != -1)
8667 /* $`, ${^PREMATCH} */
8668 i = rx->offs[0].start;
8672 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8673 && rx->offs[0].end != -1)
8675 /* $', ${^POSTMATCH} */
8676 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8677 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8680 if ( 0 <= n && n <= (I32)rx->nparens &&
8681 (s1 = rx->offs[n].start) != -1 &&
8682 (t1 = rx->offs[n].end) != -1)
8684 /* $&, ${^MATCH}, $1 ... */
8686 s = rx->subbeg + s1 - rx->suboffset;
8691 assert(s >= rx->subbeg);
8692 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8694 #ifdef NO_TAINT_SUPPORT
8695 sv_setpvn(sv, s, i);
8697 const int oldtainted = TAINT_get;
8699 sv_setpvn(sv, s, i);
8700 TAINT_set(oldtainted);
8702 if (RXp_MATCH_UTF8(rx))
8707 if (RXp_MATCH_TAINTED(rx)) {
8708 if (SvTYPE(sv) >= SVt_PVMG) {
8709 MAGIC* const mg = SvMAGIC(sv);
8712 SvMAGIC_set(sv, mg->mg_moremagic);
8714 if ((mgt = SvMAGIC(sv))) {
8715 mg->mg_moremagic = mgt;
8716 SvMAGIC_set(sv, mg);
8733 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8734 SV const * const value)
8736 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8738 PERL_UNUSED_ARG(rx);
8739 PERL_UNUSED_ARG(paren);
8740 PERL_UNUSED_ARG(value);
8743 Perl_croak_no_modify();
8747 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8750 struct regexp *const rx = ReANY(r);
8754 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8756 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8757 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8758 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8761 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8763 /* on something like
8766 * the KEEPCOPY is set on the PMOP rather than the regex */
8767 if (PL_curpm && r == PM_GETRE(PL_curpm))
8768 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8774 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8776 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8777 case RX_BUFF_IDX_PREMATCH: /* $` */
8778 if (rx->offs[0].start != -1) {
8779 i = rx->offs[0].start;
8788 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8789 case RX_BUFF_IDX_POSTMATCH: /* $' */
8790 if (rx->offs[0].end != -1) {
8791 i = rx->sublen - rx->offs[0].end;
8793 s1 = rx->offs[0].end;
8800 default: /* $& / ${^MATCH}, $1, $2, ... */
8801 if (paren <= (I32)rx->nparens &&
8802 (s1 = rx->offs[paren].start) != -1 &&
8803 (t1 = rx->offs[paren].end) != -1)
8809 if (ckWARN(WARN_UNINITIALIZED))
8810 report_uninit((const SV *)sv);
8815 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8816 const char * const s = rx->subbeg - rx->suboffset + s1;
8821 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8828 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8830 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8831 PERL_UNUSED_ARG(rx);
8835 return newSVpvs("Regexp");
8838 /* Scans the name of a named buffer from the pattern.
8839 * If flags is REG_RSN_RETURN_NULL returns null.
8840 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8841 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8842 * to the parsed name as looked up in the RExC_paren_names hash.
8843 * If there is an error throws a vFAIL().. type exception.
8846 #define REG_RSN_RETURN_NULL 0
8847 #define REG_RSN_RETURN_NAME 1
8848 #define REG_RSN_RETURN_DATA 2
8851 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8853 char *name_start = RExC_parse;
8856 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8858 assert (RExC_parse <= RExC_end);
8859 if (RExC_parse == RExC_end) NOOP;
8860 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8861 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8862 * using do...while */
8865 RExC_parse += UTF8SKIP(RExC_parse);
8866 } while ( RExC_parse < RExC_end
8867 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8871 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8873 RExC_parse++; /* so the <- from the vFAIL is after the offending
8875 vFAIL("Group name must start with a non-digit word character");
8877 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8878 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8879 if ( flags == REG_RSN_RETURN_NAME)
8881 else if (flags==REG_RSN_RETURN_DATA) {
8884 if ( ! sv_name ) /* should not happen*/
8885 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8886 if (RExC_paren_names)
8887 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8889 sv_dat = HeVAL(he_str);
8890 if ( ! sv_dat ) { /* Didn't find group */
8892 /* It might be a forward reference; we can't fail until we
8893 * know, by completing the parse to get all the groups, and
8895 if (ALL_PARENS_COUNTED) {
8896 vFAIL("Reference to nonexistent named group");
8899 REQUIRE_PARENS_PASS;
8905 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8906 (unsigned long) flags);
8909 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8910 if (RExC_lastparse!=RExC_parse) { \
8911 Perl_re_printf( aTHX_ "%s", \
8912 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8913 RExC_end - RExC_parse, 16, \
8915 PERL_PV_ESCAPE_UNI_DETECT | \
8916 PERL_PV_PRETTY_ELLIPSES | \
8917 PERL_PV_PRETTY_LTGT | \
8918 PERL_PV_ESCAPE_RE | \
8919 PERL_PV_PRETTY_EXACTSIZE \
8923 Perl_re_printf( aTHX_ "%16s",""); \
8925 if (RExC_lastnum!=RExC_emit) \
8926 Perl_re_printf( aTHX_ "|%4d", RExC_emit); \
8928 Perl_re_printf( aTHX_ "|%4s",""); \
8929 Perl_re_printf( aTHX_ "|%*s%-4s", \
8930 (int)((depth*2)), "", \
8933 RExC_lastnum=RExC_emit; \
8934 RExC_lastparse=RExC_parse; \
8939 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8940 DEBUG_PARSE_MSG((funcname)); \
8941 Perl_re_printf( aTHX_ "%4s","\n"); \
8943 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8944 DEBUG_PARSE_MSG((funcname)); \
8945 Perl_re_printf( aTHX_ fmt "\n",args); \
8948 /* This section of code defines the inversion list object and its methods. The
8949 * interfaces are highly subject to change, so as much as possible is static to
8950 * this file. An inversion list is here implemented as a malloc'd C UV array
8951 * as an SVt_INVLIST scalar.
8953 * An inversion list for Unicode is an array of code points, sorted by ordinal
8954 * number. Each element gives the code point that begins a range that extends
8955 * up-to but not including the code point given by the next element. The final
8956 * element gives the first code point of a range that extends to the platform's
8957 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8958 * ...) give ranges whose code points are all in the inversion list. We say
8959 * that those ranges are in the set. The odd-numbered elements give ranges
8960 * whose code points are not in the inversion list, and hence not in the set.
8961 * Thus, element [0] is the first code point in the list. Element [1]
8962 * is the first code point beyond that not in the list; and element [2] is the
8963 * first code point beyond that that is in the list. In other words, the first
8964 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8965 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8966 * all code points in that range are not in the inversion list. The third
8967 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8968 * list, and so forth. Thus every element whose index is divisible by two
8969 * gives the beginning of a range that is in the list, and every element whose
8970 * index is not divisible by two gives the beginning of a range not in the
8971 * list. If the final element's index is divisible by two, the inversion list
8972 * extends to the platform's infinity; otherwise the highest code point in the
8973 * inversion list is the contents of that element minus 1.
8975 * A range that contains just a single code point N will look like
8977 * invlist[i+1] == N+1
8979 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8980 * impossible to represent, so element [i+1] is omitted. The single element
8982 * invlist[0] == UV_MAX
8983 * contains just UV_MAX, but is interpreted as matching to infinity.
8985 * Taking the complement (inverting) an inversion list is quite simple, if the
8986 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8987 * This implementation reserves an element at the beginning of each inversion
8988 * list to always contain 0; there is an additional flag in the header which
8989 * indicates if the list begins at the 0, or is offset to begin at the next
8990 * element. This means that the inversion list can be inverted without any
8991 * copying; just flip the flag.
8993 * More about inversion lists can be found in "Unicode Demystified"
8994 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8996 * The inversion list data structure is currently implemented as an SV pointing
8997 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8998 * array of UV whose memory management is automatically handled by the existing
8999 * facilities for SV's.
9001 * Some of the methods should always be private to the implementation, and some
9002 * should eventually be made public */
9004 /* The header definitions are in F<invlist_inline.h> */
9006 #ifndef PERL_IN_XSUB_RE
9008 PERL_STATIC_INLINE UV*
9009 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9011 /* Returns a pointer to the first element in the inversion list's array.
9012 * This is called upon initialization of an inversion list. Where the
9013 * array begins depends on whether the list has the code point U+0000 in it
9014 * or not. The other parameter tells it whether the code that follows this
9015 * call is about to put a 0 in the inversion list or not. The first
9016 * element is either the element reserved for 0, if TRUE, or the element
9017 * after it, if FALSE */
9019 bool* offset = get_invlist_offset_addr(invlist);
9020 UV* zero_addr = (UV *) SvPVX(invlist);
9022 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9025 assert(! _invlist_len(invlist));
9029 /* 1^1 = 0; 1^0 = 1 */
9030 *offset = 1 ^ will_have_0;
9031 return zero_addr + *offset;
9035 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9037 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9038 * steals the list from 'src', so 'src' is made to have a NULL list. This
9039 * is similar to what SvSetMagicSV() would do, if it were implemented on
9040 * inversion lists, though this routine avoids a copy */
9042 const UV src_len = _invlist_len(src);
9043 const bool src_offset = *get_invlist_offset_addr(src);
9044 const STRLEN src_byte_len = SvLEN(src);
9045 char * array = SvPVX(src);
9047 const int oldtainted = TAINT_get;
9049 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9051 assert(is_invlist(src));
9052 assert(is_invlist(dest));
9053 assert(! invlist_is_iterating(src));
9054 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9056 /* Make sure it ends in the right place with a NUL, as our inversion list
9057 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9059 array[src_byte_len - 1] = '\0';
9061 TAINT_NOT; /* Otherwise it breaks */
9062 sv_usepvn_flags(dest,
9066 /* This flag is documented to cause a copy to be avoided */
9067 SV_HAS_TRAILING_NUL);
9068 TAINT_set(oldtainted);
9073 /* Finish up copying over the other fields in an inversion list */
9074 *get_invlist_offset_addr(dest) = src_offset;
9075 invlist_set_len(dest, src_len, src_offset);
9076 *get_invlist_previous_index_addr(dest) = 0;
9077 invlist_iterfinish(dest);
9080 PERL_STATIC_INLINE IV*
9081 S_get_invlist_previous_index_addr(SV* invlist)
9083 /* Return the address of the IV that is reserved to hold the cached index
9085 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9087 assert(is_invlist(invlist));
9089 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9092 PERL_STATIC_INLINE IV
9093 S_invlist_previous_index(SV* const invlist)
9095 /* Returns cached index of previous search */
9097 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9099 return *get_invlist_previous_index_addr(invlist);
9102 PERL_STATIC_INLINE void
9103 S_invlist_set_previous_index(SV* const invlist, const IV index)
9105 /* Caches <index> for later retrieval */
9107 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9109 assert(index == 0 || index < (int) _invlist_len(invlist));
9111 *get_invlist_previous_index_addr(invlist) = index;
9114 PERL_STATIC_INLINE void
9115 S_invlist_trim(SV* invlist)
9117 /* Free the not currently-being-used space in an inversion list */
9119 /* But don't free up the space needed for the 0 UV that is always at the
9120 * beginning of the list, nor the trailing NUL */
9121 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9123 PERL_ARGS_ASSERT_INVLIST_TRIM;
9125 assert(is_invlist(invlist));
9127 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9130 PERL_STATIC_INLINE void
9131 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9133 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9135 assert(is_invlist(invlist));
9137 invlist_set_len(invlist, 0, 0);
9138 invlist_trim(invlist);
9141 #endif /* ifndef PERL_IN_XSUB_RE */
9143 PERL_STATIC_INLINE bool
9144 S_invlist_is_iterating(SV* const invlist)
9146 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9148 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9151 #ifndef PERL_IN_XSUB_RE
9153 PERL_STATIC_INLINE UV
9154 S_invlist_max(SV* const invlist)
9156 /* Returns the maximum number of elements storable in the inversion list's
9157 * array, without having to realloc() */
9159 PERL_ARGS_ASSERT_INVLIST_MAX;
9161 assert(is_invlist(invlist));
9163 /* Assumes worst case, in which the 0 element is not counted in the
9164 * inversion list, so subtracts 1 for that */
9165 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9166 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9167 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9171 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9173 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9175 /* First 1 is in case the zero element isn't in the list; second 1 is for
9177 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9178 invlist_set_len(invlist, 0, 0);
9180 /* Force iterinit() to be used to get iteration to work */
9181 invlist_iterfinish(invlist);
9183 *get_invlist_previous_index_addr(invlist) = 0;
9184 SvPOK_on(invlist); /* This allows B to extract the PV */
9188 Perl__new_invlist(pTHX_ IV initial_size)
9191 /* Return a pointer to a newly constructed inversion list, with enough
9192 * space to store 'initial_size' elements. If that number is negative, a
9193 * system default is used instead */
9197 if (initial_size < 0) {
9201 new_list = newSV_type(SVt_INVLIST);
9202 initialize_invlist_guts(new_list, initial_size);
9208 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9210 /* Return a pointer to a newly constructed inversion list, initialized to
9211 * point to <list>, which has to be in the exact correct inversion list
9212 * form, including internal fields. Thus this is a dangerous routine that
9213 * should not be used in the wrong hands. The passed in 'list' contains
9214 * several header fields at the beginning that are not part of the
9215 * inversion list body proper */
9217 const STRLEN length = (STRLEN) list[0];
9218 const UV version_id = list[1];
9219 const bool offset = cBOOL(list[2]);
9220 #define HEADER_LENGTH 3
9221 /* If any of the above changes in any way, you must change HEADER_LENGTH
9222 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9223 * perl -E 'say int(rand 2**31-1)'
9225 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9226 data structure type, so that one being
9227 passed in can be validated to be an
9228 inversion list of the correct vintage.
9231 SV* invlist = newSV_type(SVt_INVLIST);
9233 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9235 if (version_id != INVLIST_VERSION_ID) {
9236 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9239 /* The generated array passed in includes header elements that aren't part
9240 * of the list proper, so start it just after them */
9241 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9243 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9244 shouldn't touch it */
9246 *(get_invlist_offset_addr(invlist)) = offset;
9248 /* The 'length' passed to us is the physical number of elements in the
9249 * inversion list. But if there is an offset the logical number is one
9251 invlist_set_len(invlist, length - offset, offset);
9253 invlist_set_previous_index(invlist, 0);
9255 /* Initialize the iteration pointer. */
9256 invlist_iterfinish(invlist);
9258 SvREADONLY_on(invlist);
9265 S__append_range_to_invlist(pTHX_ SV* const invlist,
9266 const UV start, const UV end)
9268 /* Subject to change or removal. Append the range from 'start' to 'end' at
9269 * the end of the inversion list. The range must be above any existing
9273 UV max = invlist_max(invlist);
9274 UV len = _invlist_len(invlist);
9277 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9279 if (len == 0) { /* Empty lists must be initialized */
9280 offset = start != 0;
9281 array = _invlist_array_init(invlist, ! offset);
9284 /* Here, the existing list is non-empty. The current max entry in the
9285 * list is generally the first value not in the set, except when the
9286 * set extends to the end of permissible values, in which case it is
9287 * the first entry in that final set, and so this call is an attempt to
9288 * append out-of-order */
9290 UV final_element = len - 1;
9291 array = invlist_array(invlist);
9292 if ( array[final_element] > start
9293 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9295 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",
9296 array[final_element], start,
9297 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9300 /* Here, it is a legal append. If the new range begins 1 above the end
9301 * of the range below it, it is extending the range below it, so the
9302 * new first value not in the set is one greater than the newly
9303 * extended range. */
9304 offset = *get_invlist_offset_addr(invlist);
9305 if (array[final_element] == start) {
9306 if (end != UV_MAX) {
9307 array[final_element] = end + 1;
9310 /* But if the end is the maximum representable on the machine,
9311 * assume that infinity was actually what was meant. Just let
9312 * the range that this would extend to have no end */
9313 invlist_set_len(invlist, len - 1, offset);
9319 /* Here the new range doesn't extend any existing set. Add it */
9321 len += 2; /* Includes an element each for the start and end of range */
9323 /* If wll overflow the existing space, extend, which may cause the array to
9326 invlist_extend(invlist, len);
9328 /* Have to set len here to avoid assert failure in invlist_array() */
9329 invlist_set_len(invlist, len, offset);
9331 array = invlist_array(invlist);
9334 invlist_set_len(invlist, len, offset);
9337 /* The next item on the list starts the range, the one after that is
9338 * one past the new range. */
9339 array[len - 2] = start;
9340 if (end != UV_MAX) {
9341 array[len - 1] = end + 1;
9344 /* But if the end is the maximum representable on the machine, just let
9345 * the range have no end */
9346 invlist_set_len(invlist, len - 1, offset);
9351 Perl__invlist_search(SV* const invlist, const UV cp)
9353 /* Searches the inversion list for the entry that contains the input code
9354 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9355 * return value is the index into the list's array of the range that
9356 * contains <cp>, that is, 'i' such that
9357 * array[i] <= cp < array[i+1]
9362 IV high = _invlist_len(invlist);
9363 const IV highest_element = high - 1;
9366 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9368 /* If list is empty, return failure. */
9373 /* (We can't get the array unless we know the list is non-empty) */
9374 array = invlist_array(invlist);
9376 mid = invlist_previous_index(invlist);
9378 if (mid > highest_element) {
9379 mid = highest_element;
9382 /* <mid> contains the cache of the result of the previous call to this
9383 * function (0 the first time). See if this call is for the same result,
9384 * or if it is for mid-1. This is under the theory that calls to this
9385 * function will often be for related code points that are near each other.
9386 * And benchmarks show that caching gives better results. We also test
9387 * here if the code point is within the bounds of the list. These tests
9388 * replace others that would have had to be made anyway to make sure that
9389 * the array bounds were not exceeded, and these give us extra information
9390 * at the same time */
9391 if (cp >= array[mid]) {
9392 if (cp >= array[highest_element]) {
9393 return highest_element;
9396 /* Here, array[mid] <= cp < array[highest_element]. This means that
9397 * the final element is not the answer, so can exclude it; it also
9398 * means that <mid> is not the final element, so can refer to 'mid + 1'
9400 if (cp < array[mid + 1]) {
9406 else { /* cp < aray[mid] */
9407 if (cp < array[0]) { /* Fail if outside the array */
9411 if (cp >= array[mid - 1]) {
9416 /* Binary search. What we are looking for is <i> such that
9417 * array[i] <= cp < array[i+1]
9418 * The loop below converges on the i+1. Note that there may not be an
9419 * (i+1)th element in the array, and things work nonetheless */
9420 while (low < high) {
9421 mid = (low + high) / 2;
9422 assert(mid <= highest_element);
9423 if (array[mid] <= cp) { /* cp >= array[mid] */
9426 /* We could do this extra test to exit the loop early.
9427 if (cp < array[low]) {
9432 else { /* cp < array[mid] */
9439 invlist_set_previous_index(invlist, high);
9444 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9445 const bool complement_b, SV** output)
9447 /* Take the union of two inversion lists and point '*output' to it. On
9448 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9449 * even 'a' or 'b'). If to an inversion list, the contents of the original
9450 * list will be replaced by the union. The first list, 'a', may be
9451 * NULL, in which case a copy of the second list is placed in '*output'.
9452 * If 'complement_b' is TRUE, the union is taken of the complement
9453 * (inversion) of 'b' instead of b itself.
9455 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9456 * Richard Gillam, published by Addison-Wesley, and explained at some
9457 * length there. The preface says to incorporate its examples into your
9458 * code at your own risk.
9460 * The algorithm is like a merge sort. */
9462 const UV* array_a; /* a's array */
9464 UV len_a; /* length of a's array */
9467 SV* u; /* the resulting union */
9471 UV i_a = 0; /* current index into a's array */
9475 /* running count, as explained in the algorithm source book; items are
9476 * stopped accumulating and are output when the count changes to/from 0.
9477 * The count is incremented when we start a range that's in an input's set,
9478 * and decremented when we start a range that's not in a set. So this
9479 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9480 * and hence nothing goes into the union; 1, just one of the inputs is in
9481 * its set (and its current range gets added to the union); and 2 when both
9482 * inputs are in their sets. */
9485 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9487 assert(*output == NULL || is_invlist(*output));
9489 len_b = _invlist_len(b);
9492 /* Here, 'b' is empty, hence it's complement is all possible code
9493 * points. So if the union includes the complement of 'b', it includes
9494 * everything, and we need not even look at 'a'. It's easiest to
9495 * create a new inversion list that matches everything. */
9497 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9499 if (*output == NULL) { /* If the output didn't exist, just point it
9501 *output = everything;
9503 else { /* Otherwise, replace its contents with the new list */
9504 invlist_replace_list_destroys_src(*output, everything);
9505 SvREFCNT_dec_NN(everything);
9511 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9512 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9513 * output will be empty */
9515 if (a == NULL || _invlist_len(a) == 0) {
9516 if (*output == NULL) {
9517 *output = _new_invlist(0);
9520 invlist_clear(*output);
9525 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9526 * union. We can just return a copy of 'a' if '*output' doesn't point
9527 * to an existing list */
9528 if (*output == NULL) {
9529 *output = invlist_clone(a, NULL);
9533 /* If the output is to overwrite 'a', we have a no-op, as it's
9539 /* Here, '*output' is to be overwritten by 'a' */
9540 u = invlist_clone(a, NULL);
9541 invlist_replace_list_destroys_src(*output, u);
9547 /* Here 'b' is not empty. See about 'a' */
9549 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9551 /* Here, 'a' is empty (and b is not). That means the union will come
9552 * entirely from 'b'. If '*output' is NULL, we can directly return a
9553 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9556 SV ** dest = (*output == NULL) ? output : &u;
9557 *dest = invlist_clone(b, NULL);
9559 _invlist_invert(*dest);
9563 invlist_replace_list_destroys_src(*output, u);
9570 /* Here both lists exist and are non-empty */
9571 array_a = invlist_array(a);
9572 array_b = invlist_array(b);
9574 /* If are to take the union of 'a' with the complement of b, set it
9575 * up so are looking at b's complement. */
9578 /* To complement, we invert: if the first element is 0, remove it. To
9579 * do this, we just pretend the array starts one later */
9580 if (array_b[0] == 0) {
9586 /* But if the first element is not zero, we pretend the list starts
9587 * at the 0 that is always stored immediately before the array. */
9593 /* Size the union for the worst case: that the sets are completely
9595 u = _new_invlist(len_a + len_b);
9597 /* Will contain U+0000 if either component does */
9598 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9599 || (len_b > 0 && array_b[0] == 0));
9601 /* Go through each input list item by item, stopping when have exhausted
9603 while (i_a < len_a && i_b < len_b) {
9604 UV cp; /* The element to potentially add to the union's array */
9605 bool cp_in_set; /* is it in the the input list's set or not */
9607 /* We need to take one or the other of the two inputs for the union.
9608 * Since we are merging two sorted lists, we take the smaller of the
9609 * next items. In case of a tie, we take first the one that is in its
9610 * set. If we first took the one not in its set, it would decrement
9611 * the count, possibly to 0 which would cause it to be output as ending
9612 * the range, and the next time through we would take the same number,
9613 * and output it again as beginning the next range. By doing it the
9614 * opposite way, there is no possibility that the count will be
9615 * momentarily decremented to 0, and thus the two adjoining ranges will
9616 * be seamlessly merged. (In a tie and both are in the set or both not
9617 * in the set, it doesn't matter which we take first.) */
9618 if ( array_a[i_a] < array_b[i_b]
9619 || ( array_a[i_a] == array_b[i_b]
9620 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9622 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9623 cp = array_a[i_a++];
9626 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9627 cp = array_b[i_b++];
9630 /* Here, have chosen which of the two inputs to look at. Only output
9631 * if the running count changes to/from 0, which marks the
9632 * beginning/end of a range that's in the set */
9635 array_u[i_u++] = cp;
9642 array_u[i_u++] = cp;
9648 /* The loop above increments the index into exactly one of the input lists
9649 * each iteration, and ends when either index gets to its list end. That
9650 * means the other index is lower than its end, and so something is
9651 * remaining in that one. We decrement 'count', as explained below, if
9652 * that list is in its set. (i_a and i_b each currently index the element
9653 * beyond the one we care about.) */
9654 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9655 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9660 /* Above we decremented 'count' if the list that had unexamined elements in
9661 * it was in its set. This has made it so that 'count' being non-zero
9662 * means there isn't anything left to output; and 'count' equal to 0 means
9663 * that what is left to output is precisely that which is left in the
9664 * non-exhausted input list.
9666 * To see why, note first that the exhausted input obviously has nothing
9667 * left to add to the union. If it was in its set at its end, that means
9668 * the set extends from here to the platform's infinity, and hence so does
9669 * the union and the non-exhausted set is irrelevant. The exhausted set
9670 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9671 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9672 * 'count' remains at 1. This is consistent with the decremented 'count'
9673 * != 0 meaning there's nothing left to add to the union.
9675 * But if the exhausted input wasn't in its set, it contributed 0 to
9676 * 'count', and the rest of the union will be whatever the other input is.
9677 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9678 * otherwise it gets decremented to 0. This is consistent with 'count'
9679 * == 0 meaning the remainder of the union is whatever is left in the
9680 * non-exhausted list. */
9685 IV copy_count = len_a - i_a;
9686 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9687 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9689 else { /* The non-exhausted input is b */
9690 copy_count = len_b - i_b;
9691 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9693 len_u = i_u + copy_count;
9696 /* Set the result to the final length, which can change the pointer to
9697 * array_u, so re-find it. (Note that it is unlikely that this will
9698 * change, as we are shrinking the space, not enlarging it) */
9699 if (len_u != _invlist_len(u)) {
9700 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9702 array_u = invlist_array(u);
9705 if (*output == NULL) { /* Simply return the new inversion list */
9709 /* Otherwise, overwrite the inversion list that was in '*output'. We
9710 * could instead free '*output', and then set it to 'u', but experience
9711 * has shown [perl #127392] that if the input is a mortal, we can get a
9712 * huge build-up of these during regex compilation before they get
9714 invlist_replace_list_destroys_src(*output, u);
9722 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9723 const bool complement_b, SV** i)
9725 /* Take the intersection of two inversion lists and point '*i' to it. On
9726 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9727 * even 'a' or 'b'). If to an inversion list, the contents of the original
9728 * list will be replaced by the intersection. The first list, 'a', may be
9729 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9730 * TRUE, the result will be the intersection of 'a' and the complement (or
9731 * inversion) of 'b' instead of 'b' directly.
9733 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9734 * Richard Gillam, published by Addison-Wesley, and explained at some
9735 * length there. The preface says to incorporate its examples into your
9736 * code at your own risk. In fact, it had bugs
9738 * The algorithm is like a merge sort, and is essentially the same as the
9742 const UV* array_a; /* a's array */
9744 UV len_a; /* length of a's array */
9747 SV* r; /* the resulting intersection */
9751 UV i_a = 0; /* current index into a's array */
9755 /* running count of how many of the two inputs are postitioned at ranges
9756 * that are in their sets. As explained in the algorithm source book,
9757 * items are stopped accumulating and are output when the count changes
9758 * to/from 2. The count is incremented when we start a range that's in an
9759 * input's set, and decremented when we start a range that's not in a set.
9760 * Only when it is 2 are we in the intersection. */
9763 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9765 assert(*i == NULL || is_invlist(*i));
9767 /* Special case if either one is empty */
9768 len_a = (a == NULL) ? 0 : _invlist_len(a);
9769 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9770 if (len_a != 0 && complement_b) {
9772 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9773 * must be empty. Here, also we are using 'b's complement, which
9774 * hence must be every possible code point. Thus the intersection
9777 if (*i == a) { /* No-op */
9782 *i = invlist_clone(a, NULL);
9786 r = invlist_clone(a, NULL);
9787 invlist_replace_list_destroys_src(*i, r);
9792 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9793 * intersection must be empty */
9795 *i = _new_invlist(0);
9803 /* Here both lists exist and are non-empty */
9804 array_a = invlist_array(a);
9805 array_b = invlist_array(b);
9807 /* If are to take the intersection of 'a' with the complement of b, set it
9808 * up so are looking at b's complement. */
9811 /* To complement, we invert: if the first element is 0, remove it. To
9812 * do this, we just pretend the array starts one later */
9813 if (array_b[0] == 0) {
9819 /* But if the first element is not zero, we pretend the list starts
9820 * at the 0 that is always stored immediately before the array. */
9826 /* Size the intersection for the worst case: that the intersection ends up
9827 * fragmenting everything to be completely disjoint */
9828 r= _new_invlist(len_a + len_b);
9830 /* Will contain U+0000 iff both components do */
9831 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9832 && len_b > 0 && array_b[0] == 0);
9834 /* Go through each list item by item, stopping when have exhausted one of
9836 while (i_a < len_a && i_b < len_b) {
9837 UV cp; /* The element to potentially add to the intersection's
9839 bool cp_in_set; /* Is it in the input list's set or not */
9841 /* We need to take one or the other of the two inputs for the
9842 * intersection. Since we are merging two sorted lists, we take the
9843 * smaller of the next items. In case of a tie, we take first the one
9844 * that is not in its set (a difference from the union algorithm). If
9845 * we first took the one in its set, it would increment the count,
9846 * possibly to 2 which would cause it to be output as starting a range
9847 * in the intersection, and the next time through we would take that
9848 * same number, and output it again as ending the set. By doing the
9849 * opposite of this, there is no possibility that the count will be
9850 * momentarily incremented to 2. (In a tie and both are in the set or
9851 * both not in the set, it doesn't matter which we take first.) */
9852 if ( array_a[i_a] < array_b[i_b]
9853 || ( array_a[i_a] == array_b[i_b]
9854 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9856 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9857 cp = array_a[i_a++];
9860 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9864 /* Here, have chosen which of the two inputs to look at. Only output
9865 * if the running count changes to/from 2, which marks the
9866 * beginning/end of a range that's in the intersection */
9870 array_r[i_r++] = cp;
9875 array_r[i_r++] = cp;
9882 /* The loop above increments the index into exactly one of the input lists
9883 * each iteration, and ends when either index gets to its list end. That
9884 * means the other index is lower than its end, and so something is
9885 * remaining in that one. We increment 'count', as explained below, if the
9886 * exhausted list was in its set. (i_a and i_b each currently index the
9887 * element beyond the one we care about.) */
9888 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9889 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9894 /* Above we incremented 'count' if the exhausted list was in its set. This
9895 * has made it so that 'count' being below 2 means there is nothing left to
9896 * output; otheriwse what's left to add to the intersection is precisely
9897 * that which is left in the non-exhausted input list.
9899 * To see why, note first that the exhausted input obviously has nothing
9900 * left to affect the intersection. If it was in its set at its end, that
9901 * means the set extends from here to the platform's infinity, and hence
9902 * anything in the non-exhausted's list will be in the intersection, and
9903 * anything not in it won't be. Hence, the rest of the intersection is
9904 * precisely what's in the non-exhausted list The exhausted set also
9905 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9906 * it means 'count' is now at least 2. This is consistent with the
9907 * incremented 'count' being >= 2 means to add the non-exhausted list to
9910 * But if the exhausted input wasn't in its set, it contributed 0 to
9911 * 'count', and the intersection can't include anything further; the
9912 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9913 * incremented. This is consistent with 'count' being < 2 meaning nothing
9914 * further to add to the intersection. */
9915 if (count < 2) { /* Nothing left to put in the intersection. */
9918 else { /* copy the non-exhausted list, unchanged. */
9919 IV copy_count = len_a - i_a;
9920 if (copy_count > 0) { /* a is the one with stuff left */
9921 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9923 else { /* b is the one with stuff left */
9924 copy_count = len_b - i_b;
9925 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9927 len_r = i_r + copy_count;
9930 /* Set the result to the final length, which can change the pointer to
9931 * array_r, so re-find it. (Note that it is unlikely that this will
9932 * change, as we are shrinking the space, not enlarging it) */
9933 if (len_r != _invlist_len(r)) {
9934 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9936 array_r = invlist_array(r);
9939 if (*i == NULL) { /* Simply return the calculated intersection */
9942 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9943 instead free '*i', and then set it to 'r', but experience has
9944 shown [perl #127392] that if the input is a mortal, we can get a
9945 huge build-up of these during regex compilation before they get
9948 invlist_replace_list_destroys_src(*i, r);
9960 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9962 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9963 * set. A pointer to the inversion list is returned. This may actually be
9964 * a new list, in which case the passed in one has been destroyed. The
9965 * passed-in inversion list can be NULL, in which case a new one is created
9966 * with just the one range in it. The new list is not necessarily
9967 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9968 * result of this function. The gain would not be large, and in many
9969 * cases, this is called multiple times on a single inversion list, so
9970 * anything freed may almost immediately be needed again.
9972 * This used to mostly call the 'union' routine, but that is much more
9973 * heavyweight than really needed for a single range addition */
9975 UV* array; /* The array implementing the inversion list */
9976 UV len; /* How many elements in 'array' */
9977 SSize_t i_s; /* index into the invlist array where 'start'
9979 SSize_t i_e = 0; /* And the index where 'end' should go */
9980 UV cur_highest; /* The highest code point in the inversion list
9981 upon entry to this function */
9983 /* This range becomes the whole inversion list if none already existed */
9984 if (invlist == NULL) {
9985 invlist = _new_invlist(2);
9986 _append_range_to_invlist(invlist, start, end);
9990 /* Likewise, if the inversion list is currently empty */
9991 len = _invlist_len(invlist);
9993 _append_range_to_invlist(invlist, start, end);
9997 /* Starting here, we have to know the internals of the list */
9998 array = invlist_array(invlist);
10000 /* If the new range ends higher than the current highest ... */
10001 cur_highest = invlist_highest(invlist);
10002 if (end > cur_highest) {
10004 /* If the whole range is higher, we can just append it */
10005 if (start > cur_highest) {
10006 _append_range_to_invlist(invlist, start, end);
10010 /* Otherwise, add the portion that is higher ... */
10011 _append_range_to_invlist(invlist, cur_highest + 1, end);
10013 /* ... and continue on below to handle the rest. As a result of the
10014 * above append, we know that the index of the end of the range is the
10015 * final even numbered one of the array. Recall that the final element
10016 * always starts a range that extends to infinity. If that range is in
10017 * the set (meaning the set goes from here to infinity), it will be an
10018 * even index, but if it isn't in the set, it's odd, and the final
10019 * range in the set is one less, which is even. */
10020 if (end == UV_MAX) {
10028 /* We have dealt with appending, now see about prepending. If the new
10029 * range starts lower than the current lowest ... */
10030 if (start < array[0]) {
10032 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10033 * Let the union code handle it, rather than having to know the
10034 * trickiness in two code places. */
10035 if (UNLIKELY(start == 0)) {
10038 range_invlist = _new_invlist(2);
10039 _append_range_to_invlist(range_invlist, start, end);
10041 _invlist_union(invlist, range_invlist, &invlist);
10043 SvREFCNT_dec_NN(range_invlist);
10048 /* If the whole new range comes before the first entry, and doesn't
10049 * extend it, we have to insert it as an additional range */
10050 if (end < array[0] - 1) {
10052 goto splice_in_new_range;
10055 /* Here the new range adjoins the existing first range, extending it
10059 /* And continue on below to handle the rest. We know that the index of
10060 * the beginning of the range is the first one of the array */
10063 else { /* Not prepending any part of the new range to the existing list.
10064 * Find where in the list it should go. This finds i_s, such that:
10065 * invlist[i_s] <= start < array[i_s+1]
10067 i_s = _invlist_search(invlist, start);
10070 /* At this point, any extending before the beginning of the inversion list
10071 * and/or after the end has been done. This has made it so that, in the
10072 * code below, each endpoint of the new range is either in a range that is
10073 * in the set, or is in a gap between two ranges that are. This means we
10074 * don't have to worry about exceeding the array bounds.
10076 * Find where in the list the new range ends (but we can skip this if we
10077 * have already determined what it is, or if it will be the same as i_s,
10078 * which we already have computed) */
10080 i_e = (start == end)
10082 : _invlist_search(invlist, end);
10085 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10086 * is a range that goes to infinity there is no element at invlist[i_e+1],
10087 * so only the first relation holds. */
10089 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10091 /* Here, the ranges on either side of the beginning of the new range
10092 * are in the set, and this range starts in the gap between them.
10094 * The new range extends the range above it downwards if the new range
10095 * ends at or above that range's start */
10096 const bool extends_the_range_above = ( end == UV_MAX
10097 || end + 1 >= array[i_s+1]);
10099 /* The new range extends the range below it upwards if it begins just
10100 * after where that range ends */
10101 if (start == array[i_s]) {
10103 /* If the new range fills the entire gap between the other ranges,
10104 * they will get merged together. Other ranges may also get
10105 * merged, depending on how many of them the new range spans. In
10106 * the general case, we do the merge later, just once, after we
10107 * figure out how many to merge. But in the case where the new
10108 * range exactly spans just this one gap (possibly extending into
10109 * the one above), we do the merge here, and an early exit. This
10110 * is done here to avoid having to special case later. */
10111 if (i_e - i_s <= 1) {
10113 /* If i_e - i_s == 1, it means that the new range terminates
10114 * within the range above, and hence 'extends_the_range_above'
10115 * must be true. (If the range above it extends to infinity,
10116 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10117 * will be 0, so no harm done.) */
10118 if (extends_the_range_above) {
10119 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10120 invlist_set_len(invlist,
10122 *(get_invlist_offset_addr(invlist)));
10126 /* Here, i_e must == i_s. We keep them in sync, as they apply
10127 * to the same range, and below we are about to decrement i_s
10132 /* Here, the new range is adjacent to the one below. (It may also
10133 * span beyond the range above, but that will get resolved later.)
10134 * Extend the range below to include this one. */
10135 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10137 start = array[i_s];
10139 else if (extends_the_range_above) {
10141 /* Here the new range only extends the range above it, but not the
10142 * one below. It merges with the one above. Again, we keep i_e
10143 * and i_s in sync if they point to the same range */
10148 array[i_s] = start;
10152 /* Here, we've dealt with the new range start extending any adjoining
10155 * If the new range extends to infinity, it is now the final one,
10156 * regardless of what was there before */
10157 if (UNLIKELY(end == UV_MAX)) {
10158 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10162 /* If i_e started as == i_s, it has also been dealt with,
10163 * and been updated to the new i_s, which will fail the following if */
10164 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10166 /* Here, the ranges on either side of the end of the new range are in
10167 * the set, and this range ends in the gap between them.
10169 * If this range is adjacent to (hence extends) the range above it, it
10170 * becomes part of that range; likewise if it extends the range below,
10171 * it becomes part of that range */
10172 if (end + 1 == array[i_e+1]) {
10174 array[i_e] = start;
10176 else if (start <= array[i_e]) {
10177 array[i_e] = end + 1;
10184 /* If the range fits entirely in an existing range (as possibly already
10185 * extended above), it doesn't add anything new */
10186 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10190 /* Here, no part of the range is in the list. Must add it. It will
10191 * occupy 2 more slots */
10192 splice_in_new_range:
10194 invlist_extend(invlist, len + 2);
10195 array = invlist_array(invlist);
10196 /* Move the rest of the array down two slots. Don't include any
10198 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10200 /* Do the actual splice */
10201 array[i_e+1] = start;
10202 array[i_e+2] = end + 1;
10203 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10207 /* Here the new range crossed the boundaries of a pre-existing range. The
10208 * code above has adjusted things so that both ends are in ranges that are
10209 * in the set. This means everything in between must also be in the set.
10210 * Just squash things together */
10211 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10212 invlist_set_len(invlist,
10214 *(get_invlist_offset_addr(invlist)));
10220 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10221 UV** other_elements_ptr)
10223 /* Create and return an inversion list whose contents are to be populated
10224 * by the caller. The caller gives the number of elements (in 'size') and
10225 * the very first element ('element0'). This function will set
10226 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10227 * are to be placed.
10229 * Obviously there is some trust involved that the caller will properly
10230 * fill in the other elements of the array.
10232 * (The first element needs to be passed in, as the underlying code does
10233 * things differently depending on whether it is zero or non-zero) */
10235 SV* invlist = _new_invlist(size);
10238 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10240 invlist = add_cp_to_invlist(invlist, element0);
10241 offset = *get_invlist_offset_addr(invlist);
10243 invlist_set_len(invlist, size, offset);
10244 *other_elements_ptr = invlist_array(invlist) + 1;
10250 #ifndef PERL_IN_XSUB_RE
10252 Perl__invlist_invert(pTHX_ SV* const invlist)
10254 /* Complement the input inversion list. This adds a 0 if the list didn't
10255 * have a zero; removes it otherwise. As described above, the data
10256 * structure is set up so that this is very efficient */
10258 PERL_ARGS_ASSERT__INVLIST_INVERT;
10260 assert(! invlist_is_iterating(invlist));
10262 /* The inverse of matching nothing is matching everything */
10263 if (_invlist_len(invlist) == 0) {
10264 _append_range_to_invlist(invlist, 0, UV_MAX);
10268 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10272 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10274 /* Return a new inversion list that is a copy of the input one, which is
10275 * unchanged. The new list will not be mortal even if the old one was. */
10277 const STRLEN nominal_length = _invlist_len(invlist);
10278 const STRLEN physical_length = SvCUR(invlist);
10279 const bool offset = *(get_invlist_offset_addr(invlist));
10281 PERL_ARGS_ASSERT_INVLIST_CLONE;
10283 if (new_invlist == NULL) {
10284 new_invlist = _new_invlist(nominal_length);
10287 sv_upgrade(new_invlist, SVt_INVLIST);
10288 initialize_invlist_guts(new_invlist, nominal_length);
10291 *(get_invlist_offset_addr(new_invlist)) = offset;
10292 invlist_set_len(new_invlist, nominal_length, offset);
10293 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10295 return new_invlist;
10301 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10303 /* Get the contents of an inversion list into a string SV so that they can
10304 * be printed out. If 'traditional_style' is TRUE, it uses the format
10305 * traditionally done for debug tracing; otherwise it uses a format
10306 * suitable for just copying to the output, with blanks between ranges and
10307 * a dash between range components */
10311 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10312 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10314 if (traditional_style) {
10315 output = newSVpvs("\n");
10318 output = newSVpvs("");
10321 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10323 assert(! invlist_is_iterating(invlist));
10325 invlist_iterinit(invlist);
10326 while (invlist_iternext(invlist, &start, &end)) {
10327 if (end == UV_MAX) {
10328 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10329 start, intra_range_delimiter,
10330 inter_range_delimiter);
10332 else if (end != start) {
10333 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10335 intra_range_delimiter,
10336 end, inter_range_delimiter);
10339 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10340 start, inter_range_delimiter);
10344 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10345 SvCUR_set(output, SvCUR(output) - 1);
10351 #ifndef PERL_IN_XSUB_RE
10353 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10354 const char * const indent, SV* const invlist)
10356 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10357 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10358 * the string 'indent'. The output looks like this:
10359 [0] 0x000A .. 0x000D
10361 [4] 0x2028 .. 0x2029
10362 [6] 0x3104 .. INFTY
10363 * This means that the first range of code points matched by the list are
10364 * 0xA through 0xD; the second range contains only the single code point
10365 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10366 * are used to define each range (except if the final range extends to
10367 * infinity, only a single element is needed). The array index of the
10368 * first element for the corresponding range is given in brackets. */
10373 PERL_ARGS_ASSERT__INVLIST_DUMP;
10375 if (invlist_is_iterating(invlist)) {
10376 Perl_dump_indent(aTHX_ level, file,
10377 "%sCan't dump inversion list because is in middle of iterating\n",
10382 invlist_iterinit(invlist);
10383 while (invlist_iternext(invlist, &start, &end)) {
10384 if (end == UV_MAX) {
10385 Perl_dump_indent(aTHX_ level, file,
10386 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10387 indent, (UV)count, start);
10389 else if (end != start) {
10390 Perl_dump_indent(aTHX_ level, file,
10391 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10392 indent, (UV)count, start, end);
10395 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10396 indent, (UV)count, start);
10404 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10406 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10408 /* Return a boolean as to if the two passed in inversion lists are
10409 * identical. The final argument, if TRUE, says to take the complement of
10410 * the second inversion list before doing the comparison */
10412 const UV len_a = _invlist_len(a);
10413 UV len_b = _invlist_len(b);
10415 const UV* array_a = NULL;
10416 const UV* array_b = NULL;
10418 PERL_ARGS_ASSERT__INVLISTEQ;
10420 /* This code avoids accessing the arrays unless it knows the length is
10425 return ! complement_b;
10429 array_a = invlist_array(a);
10433 array_b = invlist_array(b);
10436 /* If are to compare 'a' with the complement of b, set it
10437 * up so are looking at b's complement. */
10438 if (complement_b) {
10440 /* The complement of nothing is everything, so <a> would have to have
10441 * just one element, starting at zero (ending at infinity) */
10443 return (len_a == 1 && array_a[0] == 0);
10445 if (array_b[0] == 0) {
10447 /* Otherwise, to complement, we invert. Here, the first element is
10448 * 0, just remove it. To do this, we just pretend the array starts
10456 /* But if the first element is not zero, we pretend the list starts
10457 * at the 0 that is always stored immediately before the array. */
10463 return len_a == len_b
10464 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10470 * As best we can, determine the characters that can match the start of
10471 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10472 * can be false positive matches
10474 * Returns the invlist as a new SV*; it is the caller's responsibility to
10475 * call SvREFCNT_dec() when done with it.
10478 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10481 const U8 * s = (U8*)STRING(node);
10482 SSize_t bytelen = STR_LEN(node);
10484 /* Start out big enough for 2 separate code points */
10485 SV* invlist = _new_invlist(4);
10487 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10492 /* We punt and assume can match anything if the node begins
10493 * with a multi-character fold. Things are complicated. For
10494 * example, /ffi/i could match any of:
10495 * "\N{LATIN SMALL LIGATURE FFI}"
10496 * "\N{LATIN SMALL LIGATURE FF}I"
10497 * "F\N{LATIN SMALL LIGATURE FI}"
10498 * plus several other things; and making sure we have all the
10499 * possibilities is hard. */
10500 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10501 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10504 /* Any Latin1 range character can potentially match any
10505 * other depending on the locale, and in Turkic locales, U+130 and
10507 if (OP(node) == EXACTFL) {
10508 _invlist_union(invlist, PL_Latin1, &invlist);
10509 invlist = add_cp_to_invlist(invlist,
10510 LATIN_SMALL_LETTER_DOTLESS_I);
10511 invlist = add_cp_to_invlist(invlist,
10512 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10515 /* But otherwise, it matches at least itself. We can
10516 * quickly tell if it has a distinct fold, and if so,
10517 * it matches that as well */
10518 invlist = add_cp_to_invlist(invlist, uc);
10519 if (IS_IN_SOME_FOLD_L1(uc))
10520 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10523 /* Some characters match above-Latin1 ones under /i. This
10524 * is true of EXACTFL ones when the locale is UTF-8 */
10525 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10526 && (! isASCII(uc) || (OP(node) != EXACTFAA
10527 && OP(node) != EXACTFAA_NO_TRIE)))
10529 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10533 else { /* Pattern is UTF-8 */
10534 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10535 const U8* e = s + bytelen;
10538 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10540 /* The only code points that aren't folded in a UTF EXACTFish
10541 * node are are the problematic ones in EXACTFL nodes */
10542 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10543 /* We need to check for the possibility that this EXACTFL
10544 * node begins with a multi-char fold. Therefore we fold
10545 * the first few characters of it so that we can make that
10551 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10553 *(d++) = (U8) toFOLD(*s);
10554 if (fc < 0) { /* Save the first fold */
10561 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10562 if (fc < 0) { /* Save the first fold */
10570 /* And set up so the code below that looks in this folded
10571 * buffer instead of the node's string */
10576 /* When we reach here 's' points to the fold of the first
10577 * character(s) of the node; and 'e' points to far enough along
10578 * the folded string to be just past any possible multi-char
10581 * Unlike the non-UTF-8 case, the macro for determining if a
10582 * string is a multi-char fold requires all the characters to
10583 * already be folded. This is because of all the complications
10584 * if not. Note that they are folded anyway, except in EXACTFL
10585 * nodes. Like the non-UTF case above, we punt if the node
10586 * begins with a multi-char fold */
10588 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10589 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10591 else { /* Single char fold */
10593 unsigned int first_fold;
10594 const unsigned int * remaining_folds;
10595 Size_t folds_count;
10597 /* It matches itself */
10598 invlist = add_cp_to_invlist(invlist, fc);
10600 /* ... plus all the things that fold to it, which are found in
10601 * PL_utf8_foldclosures */
10602 folds_count = _inverse_folds(fc, &first_fold,
10604 for (k = 0; k < folds_count; k++) {
10605 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10607 /* /aa doesn't allow folds between ASCII and non- */
10608 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10609 && isASCII(c) != isASCII(fc))
10614 invlist = add_cp_to_invlist(invlist, c);
10617 if (OP(node) == EXACTFL) {
10619 /* If either [iI] are present in an EXACTFL node the above code
10620 * should have added its normal case pair, but under a Turkish
10621 * locale they could match instead the case pairs from it. Add
10622 * those as potential matches as well */
10623 if (isALPHA_FOLD_EQ(fc, 'I')) {
10624 invlist = add_cp_to_invlist(invlist,
10625 LATIN_SMALL_LETTER_DOTLESS_I);
10626 invlist = add_cp_to_invlist(invlist,
10627 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10629 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10630 invlist = add_cp_to_invlist(invlist, 'I');
10632 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10633 invlist = add_cp_to_invlist(invlist, 'i');
10642 #undef HEADER_LENGTH
10643 #undef TO_INTERNAL_SIZE
10644 #undef FROM_INTERNAL_SIZE
10645 #undef INVLIST_VERSION_ID
10647 /* End of inversion list object */
10650 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10652 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10653 * constructs, and updates RExC_flags with them. On input, RExC_parse
10654 * should point to the first flag; it is updated on output to point to the
10655 * final ')' or ':'. There needs to be at least one flag, or this will
10658 /* for (?g), (?gc), and (?o) warnings; warning
10659 about (?c) will warn about (?g) -- japhy */
10661 #define WASTED_O 0x01
10662 #define WASTED_G 0x02
10663 #define WASTED_C 0x04
10664 #define WASTED_GC (WASTED_G|WASTED_C)
10665 I32 wastedflags = 0x00;
10666 U32 posflags = 0, negflags = 0;
10667 U32 *flagsp = &posflags;
10668 char has_charset_modifier = '\0';
10670 bool has_use_defaults = FALSE;
10671 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10672 int x_mod_count = 0;
10674 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10676 /* '^' as an initial flag sets certain defaults */
10677 if (UCHARAT(RExC_parse) == '^') {
10679 has_use_defaults = TRUE;
10680 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10681 cs = (RExC_uni_semantics)
10682 ? REGEX_UNICODE_CHARSET
10683 : REGEX_DEPENDS_CHARSET;
10684 set_regex_charset(&RExC_flags, cs);
10687 cs = get_regex_charset(RExC_flags);
10688 if ( cs == REGEX_DEPENDS_CHARSET
10689 && RExC_uni_semantics)
10691 cs = REGEX_UNICODE_CHARSET;
10695 while (RExC_parse < RExC_end) {
10696 /* && strchr("iogcmsx", *RExC_parse) */
10697 /* (?g), (?gc) and (?o) are useless here
10698 and must be globally applied -- japhy */
10699 switch (*RExC_parse) {
10701 /* Code for the imsxn flags */
10702 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10704 case LOCALE_PAT_MOD:
10705 if (has_charset_modifier) {
10706 goto excess_modifier;
10708 else if (flagsp == &negflags) {
10711 cs = REGEX_LOCALE_CHARSET;
10712 has_charset_modifier = LOCALE_PAT_MOD;
10714 case UNICODE_PAT_MOD:
10715 if (has_charset_modifier) {
10716 goto excess_modifier;
10718 else if (flagsp == &negflags) {
10721 cs = REGEX_UNICODE_CHARSET;
10722 has_charset_modifier = UNICODE_PAT_MOD;
10724 case ASCII_RESTRICT_PAT_MOD:
10725 if (flagsp == &negflags) {
10728 if (has_charset_modifier) {
10729 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10730 goto excess_modifier;
10732 /* Doubled modifier implies more restricted */
10733 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10736 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10738 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10740 case DEPENDS_PAT_MOD:
10741 if (has_use_defaults) {
10742 goto fail_modifiers;
10744 else if (flagsp == &negflags) {
10747 else if (has_charset_modifier) {
10748 goto excess_modifier;
10751 /* The dual charset means unicode semantics if the
10752 * pattern (or target, not known until runtime) are
10753 * utf8, or something in the pattern indicates unicode
10755 cs = (RExC_uni_semantics)
10756 ? REGEX_UNICODE_CHARSET
10757 : REGEX_DEPENDS_CHARSET;
10758 has_charset_modifier = DEPENDS_PAT_MOD;
10762 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10763 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10765 else if (has_charset_modifier == *(RExC_parse - 1)) {
10766 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10767 *(RExC_parse - 1));
10770 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10772 NOT_REACHED; /*NOTREACHED*/
10775 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10776 *(RExC_parse - 1));
10777 NOT_REACHED; /*NOTREACHED*/
10778 case ONCE_PAT_MOD: /* 'o' */
10779 case GLOBAL_PAT_MOD: /* 'g' */
10780 if (ckWARN(WARN_REGEXP)) {
10781 const I32 wflagbit = *RExC_parse == 'o'
10784 if (! (wastedflags & wflagbit) ) {
10785 wastedflags |= wflagbit;
10786 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10789 "Useless (%s%c) - %suse /%c modifier",
10790 flagsp == &negflags ? "?-" : "?",
10792 flagsp == &negflags ? "don't " : "",
10799 case CONTINUE_PAT_MOD: /* 'c' */
10800 if (ckWARN(WARN_REGEXP)) {
10801 if (! (wastedflags & WASTED_C) ) {
10802 wastedflags |= WASTED_GC;
10803 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10806 "Useless (%sc) - %suse /gc modifier",
10807 flagsp == &negflags ? "?-" : "?",
10808 flagsp == &negflags ? "don't " : ""
10813 case KEEPCOPY_PAT_MOD: /* 'p' */
10814 if (flagsp == &negflags) {
10815 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10817 *flagsp |= RXf_PMf_KEEPCOPY;
10821 /* A flag is a default iff it is following a minus, so
10822 * if there is a minus, it means will be trying to
10823 * re-specify a default which is an error */
10824 if (has_use_defaults || flagsp == &negflags) {
10825 goto fail_modifiers;
10827 flagsp = &negflags;
10828 wastedflags = 0; /* reset so (?g-c) warns twice */
10834 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10835 negflags |= RXf_PMf_EXTENDED_MORE;
10837 RExC_flags |= posflags;
10839 if (negflags & RXf_PMf_EXTENDED) {
10840 negflags |= RXf_PMf_EXTENDED_MORE;
10842 RExC_flags &= ~negflags;
10843 set_regex_charset(&RExC_flags, cs);
10848 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
10849 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10850 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10851 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10852 NOT_REACHED; /*NOTREACHED*/
10855 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10858 vFAIL("Sequence (?... not terminated");
10862 - reg - regular expression, i.e. main body or parenthesized thing
10864 * Caller must absorb opening parenthesis.
10866 * Combining parenthesis handling with the base level of regular expression
10867 * is a trifle forced, but the need to tie the tails of the branches to what
10868 * follows makes it hard to avoid.
10870 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10872 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10874 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10877 PERL_STATIC_INLINE regnode_offset
10878 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10880 char * parse_start,
10884 regnode_offset ret;
10885 char* name_start = RExC_parse;
10887 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
10888 GET_RE_DEBUG_FLAGS_DECL;
10890 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10892 if (RExC_parse == name_start || *RExC_parse != ch) {
10893 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10894 vFAIL2("Sequence %.3s... not terminated", parse_start);
10898 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10899 RExC_rxi->data->data[num]=(void*)sv_dat;
10900 SvREFCNT_inc_simple_void_NN(sv_dat);
10903 ret = reganode(pRExC_state,
10906 : (ASCII_FOLD_RESTRICTED)
10908 : (AT_LEAST_UNI_SEMANTICS)
10914 *flagp |= HASWIDTH;
10916 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
10917 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
10919 nextchar(pRExC_state);
10923 /* On success, returns the offset at which any next node should be placed into
10924 * the regex engine program being compiled.
10926 * Returns 0 otherwise, with *flagp set to indicate why:
10927 * TRYAGAIN at the end of (?) that only sets flags.
10928 * RESTART_PARSE if the parse needs to be restarted, or'd with
10929 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
10930 * Otherwise would only return 0 if regbranch() returns 0, which cannot
10932 STATIC regnode_offset
10933 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
10934 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10935 * 2 is like 1, but indicates that nextchar() has been called to advance
10936 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10937 * this flag alerts us to the need to check for that */
10939 regnode_offset ret = 0; /* Will be the head of the group. */
10941 regnode_offset lastbr;
10942 regnode_offset ender = 0;
10945 U32 oregflags = RExC_flags;
10946 bool have_branch = 0;
10948 I32 freeze_paren = 0;
10949 I32 after_freeze = 0;
10950 I32 num; /* numeric backreferences */
10951 SV * max_open; /* Max number of unclosed parens */
10953 char * parse_start = RExC_parse; /* MJD */
10954 char * const oregcomp_parse = RExC_parse;
10956 GET_RE_DEBUG_FLAGS_DECL;
10958 PERL_ARGS_ASSERT_REG;
10959 DEBUG_PARSE("reg ");
10961 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
10963 if (!SvIOK(max_open)) {
10964 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
10966 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
10968 vFAIL("Too many nested open parens");
10971 *flagp = 0; /* Tentatively. */
10973 if (RExC_in_lookbehind) {
10974 RExC_in_lookbehind++;
10976 if (RExC_in_lookahead) {
10977 RExC_in_lookahead++;
10980 /* Having this true makes it feasible to have a lot fewer tests for the
10981 * parse pointer being in scope. For example, we can write
10982 * while(isFOO(*RExC_parse)) RExC_parse++;
10984 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10986 assert(*RExC_end == '\0');
10988 /* Make an OPEN node, if parenthesized. */
10991 /* Under /x, space and comments can be gobbled up between the '(' and
10992 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10993 * intervening space, as the sequence is a token, and a token should be
10995 bool has_intervening_patws = (paren == 2)
10996 && *(RExC_parse - 1) != '(';
10998 if (RExC_parse >= RExC_end) {
10999 vFAIL("Unmatched (");
11002 if (paren == 'r') { /* Atomic script run */
11006 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11007 char *start_verb = RExC_parse + 1;
11009 char *start_arg = NULL;
11010 unsigned char op = 0;
11011 int arg_required = 0;
11012 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11013 bool has_upper = FALSE;
11015 if (has_intervening_patws) {
11016 RExC_parse++; /* past the '*' */
11018 /* For strict backwards compatibility, don't change the message
11019 * now that we also have lowercase operands */
11020 if (isUPPER(*RExC_parse)) {
11021 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11024 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11027 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11028 if ( *RExC_parse == ':' ) {
11029 start_arg = RExC_parse + 1;
11033 if (isUPPER(*RExC_parse)) {
11039 RExC_parse += UTF8SKIP(RExC_parse);
11042 verb_len = RExC_parse - start_verb;
11044 if (RExC_parse >= RExC_end) {
11045 goto unterminated_verb_pattern;
11048 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11049 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11050 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11052 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11053 unterminated_verb_pattern:
11055 vFAIL("Unterminated verb pattern argument");
11058 vFAIL("Unterminated '(*...' argument");
11062 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11064 vFAIL("Unterminated verb pattern");
11067 vFAIL("Unterminated '(*...' construct");
11072 /* Here, we know that RExC_parse < RExC_end */
11074 switch ( *start_verb ) {
11075 case 'A': /* (*ACCEPT) */
11076 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11078 internal_argval = RExC_nestroot;
11081 case 'C': /* (*COMMIT) */
11082 if ( memEQs(start_verb, verb_len,"COMMIT") )
11085 case 'F': /* (*FAIL) */
11086 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11090 case ':': /* (*:NAME) */
11091 case 'M': /* (*MARK:NAME) */
11092 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11097 case 'P': /* (*PRUNE) */
11098 if ( memEQs(start_verb, verb_len,"PRUNE") )
11101 case 'S': /* (*SKIP) */
11102 if ( memEQs(start_verb, verb_len,"SKIP") )
11105 case 'T': /* (*THEN) */
11106 /* [19:06] <TimToady> :: is then */
11107 if ( memEQs(start_verb, verb_len,"THEN") ) {
11109 RExC_seen |= REG_CUTGROUP_SEEN;
11113 if ( memEQs(start_verb, verb_len, "asr")
11114 || memEQs(start_verb, verb_len, "atomic_script_run"))
11116 paren = 'r'; /* Mnemonic: recursed run */
11119 else if (memEQs(start_verb, verb_len, "atomic")) {
11120 paren = 't'; /* AtOMIC */
11121 goto alpha_assertions;
11125 if ( memEQs(start_verb, verb_len, "plb")
11126 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11129 goto lookbehind_alpha_assertions;
11131 else if ( memEQs(start_verb, verb_len, "pla")
11132 || memEQs(start_verb, verb_len, "positive_lookahead"))
11135 goto alpha_assertions;
11139 if ( memEQs(start_verb, verb_len, "nlb")
11140 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11143 goto lookbehind_alpha_assertions;
11145 else if ( memEQs(start_verb, verb_len, "nla")
11146 || memEQs(start_verb, verb_len, "negative_lookahead"))
11149 goto alpha_assertions;
11153 if ( memEQs(start_verb, verb_len, "sr")
11154 || memEQs(start_verb, verb_len, "script_run"))
11156 regnode_offset atomic;
11162 /* This indicates Unicode rules. */
11163 REQUIRE_UNI_RULES(flagp, 0);
11169 RExC_parse = start_arg;
11171 if (RExC_in_script_run) {
11173 /* Nested script runs are treated as no-ops, because
11174 * if the nested one fails, the outer one must as
11175 * well. It could fail sooner, and avoid (??{} with
11176 * side effects, but that is explicitly documented as
11177 * undefined behavior. */
11181 if (paren == 's') {
11186 /* But, the atomic part of a nested atomic script run
11187 * isn't a no-op, but can be treated just like a '(?>'
11193 if (paren == 's') {
11194 /* Here, we're starting a new regular script run */
11195 ret = reg_node(pRExC_state, SROPEN);
11196 RExC_in_script_run = 1;
11201 /* Here, we are starting an atomic script run. This is
11202 * handled by recursing to deal with the atomic portion
11203 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11205 ret = reg_node(pRExC_state, SROPEN);
11207 RExC_in_script_run = 1;
11209 atomic = reg(pRExC_state, 'r', &flags, depth);
11210 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11211 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11215 if (! REGTAIL(pRExC_state, ret, atomic)) {
11216 REQUIRE_BRANCHJ(flagp, 0);
11219 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11222 REQUIRE_BRANCHJ(flagp, 0);
11225 RExC_in_script_run = 0;
11231 lookbehind_alpha_assertions:
11232 RExC_seen |= REG_LOOKBEHIND_SEEN;
11233 RExC_in_lookbehind++;
11238 RExC_seen_zerolen++;
11244 /* An empty negative lookahead assertion simply is failure */
11245 if (paren == 'A' && RExC_parse == start_arg) {
11246 ret=reganode(pRExC_state, OPFAIL, 0);
11247 nextchar(pRExC_state);
11251 RExC_parse = start_arg;
11256 "'(*%" UTF8f "' requires a terminating ':'",
11257 UTF8fARG(UTF, verb_len, start_verb));
11258 NOT_REACHED; /*NOTREACHED*/
11260 } /* End of switch */
11263 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11265 if (has_upper || verb_len == 0) {
11267 "Unknown verb pattern '%" UTF8f "'",
11268 UTF8fARG(UTF, verb_len, start_verb));
11272 "Unknown '(*...)' construct '%" UTF8f "'",
11273 UTF8fARG(UTF, verb_len, start_verb));
11276 if ( RExC_parse == start_arg ) {
11279 if ( arg_required && !start_arg ) {
11280 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11281 verb_len, start_verb);
11283 if (internal_argval == -1) {
11284 ret = reganode(pRExC_state, op, 0);
11286 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11288 RExC_seen |= REG_VERBARG_SEEN;
11290 SV *sv = newSVpvn( start_arg,
11291 RExC_parse - start_arg);
11292 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11293 STR_WITH_LEN("S"));
11294 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11295 FLAGS(REGNODE_p(ret)) = 1;
11297 FLAGS(REGNODE_p(ret)) = 0;
11299 if ( internal_argval != -1 )
11300 ARG2L_SET(REGNODE_p(ret), internal_argval);
11301 nextchar(pRExC_state);
11304 else if (*RExC_parse == '?') { /* (?...) */
11305 bool is_logical = 0;
11306 const char * const seqstart = RExC_parse;
11307 const char * endptr;
11308 if (has_intervening_patws) {
11310 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11313 RExC_parse++; /* past the '?' */
11314 paren = *RExC_parse; /* might be a trailing NUL, if not
11316 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11317 if (RExC_parse > RExC_end) {
11320 ret = 0; /* For look-ahead/behind. */
11323 case 'P': /* (?P...) variants for those used to PCRE/Python */
11324 paren = *RExC_parse;
11325 if ( paren == '<') { /* (?P<...>) named capture */
11327 if (RExC_parse >= RExC_end) {
11328 vFAIL("Sequence (?P<... not terminated");
11330 goto named_capture;
11332 else if (paren == '>') { /* (?P>name) named recursion */
11334 if (RExC_parse >= RExC_end) {
11335 vFAIL("Sequence (?P>... not terminated");
11337 goto named_recursion;
11339 else if (paren == '=') { /* (?P=...) named backref */
11341 return handle_named_backref(pRExC_state, flagp,
11344 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11345 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11346 vFAIL3("Sequence (%.*s...) not recognized",
11347 RExC_parse-seqstart, seqstart);
11348 NOT_REACHED; /*NOTREACHED*/
11349 case '<': /* (?<...) */
11350 if (*RExC_parse == '!')
11352 else if (*RExC_parse != '=')
11359 case '\'': /* (?'...') */
11360 name_start = RExC_parse;
11361 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11362 if ( RExC_parse == name_start
11363 || RExC_parse >= RExC_end
11364 || *RExC_parse != paren)
11366 vFAIL2("Sequence (?%c... not terminated",
11367 paren=='>' ? '<' : paren);
11372 if (!svname) /* shouldn't happen */
11374 "panic: reg_scan_name returned NULL");
11375 if (!RExC_paren_names) {
11376 RExC_paren_names= newHV();
11377 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11379 RExC_paren_name_list= newAV();
11380 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11383 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11385 sv_dat = HeVAL(he_str);
11387 /* croak baby croak */
11389 "panic: paren_name hash element allocation failed");
11390 } else if ( SvPOK(sv_dat) ) {
11391 /* (?|...) can mean we have dupes so scan to check
11392 its already been stored. Maybe a flag indicating
11393 we are inside such a construct would be useful,
11394 but the arrays are likely to be quite small, so
11395 for now we punt -- dmq */
11396 IV count = SvIV(sv_dat);
11397 I32 *pv = (I32*)SvPVX(sv_dat);
11399 for ( i = 0 ; i < count ; i++ ) {
11400 if ( pv[i] == RExC_npar ) {
11406 pv = (I32*)SvGROW(sv_dat,
11407 SvCUR(sv_dat) + sizeof(I32)+1);
11408 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11409 pv[count] = RExC_npar;
11410 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11413 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11414 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11417 SvIV_set(sv_dat, 1);
11420 /* Yes this does cause a memory leak in debugging Perls
11422 if (!av_store(RExC_paren_name_list,
11423 RExC_npar, SvREFCNT_inc_NN(svname)))
11424 SvREFCNT_dec_NN(svname);
11427 /*sv_dump(sv_dat);*/
11429 nextchar(pRExC_state);
11431 goto capturing_parens;
11434 RExC_seen |= REG_LOOKBEHIND_SEEN;
11435 RExC_in_lookbehind++;
11437 if (RExC_parse >= RExC_end) {
11438 vFAIL("Sequence (?... not terminated");
11440 RExC_seen_zerolen++;
11442 case '=': /* (?=...) */
11443 RExC_seen_zerolen++;
11444 RExC_in_lookahead++;
11446 case '!': /* (?!...) */
11447 RExC_seen_zerolen++;
11448 /* check if we're really just a "FAIL" assertion */
11449 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11450 FALSE /* Don't force to /x */ );
11451 if (*RExC_parse == ')') {
11452 ret=reganode(pRExC_state, OPFAIL, 0);
11453 nextchar(pRExC_state);
11457 case '|': /* (?|...) */
11458 /* branch reset, behave like a (?:...) except that
11459 buffers in alternations share the same numbers */
11461 after_freeze = freeze_paren = RExC_npar;
11463 /* XXX This construct currently requires an extra pass.
11464 * Investigation would be required to see if that could be
11466 REQUIRE_PARENS_PASS;
11468 case ':': /* (?:...) */
11469 case '>': /* (?>...) */
11471 case '$': /* (?$...) */
11472 case '@': /* (?@...) */
11473 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11475 case '0' : /* (?0) */
11476 case 'R' : /* (?R) */
11477 if (RExC_parse == RExC_end || *RExC_parse != ')')
11478 FAIL("Sequence (?R) not terminated");
11480 RExC_seen |= REG_RECURSE_SEEN;
11482 /* XXX These constructs currently require an extra pass.
11483 * It probably could be changed */
11484 REQUIRE_PARENS_PASS;
11486 *flagp |= POSTPONED;
11487 goto gen_recurse_regop;
11489 /* named and numeric backreferences */
11490 case '&': /* (?&NAME) */
11491 parse_start = RExC_parse - 1;
11494 SV *sv_dat = reg_scan_name(pRExC_state,
11495 REG_RSN_RETURN_DATA);
11496 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11498 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11499 vFAIL("Sequence (?&... not terminated");
11500 goto gen_recurse_regop;
11503 if (! inRANGE(RExC_parse[0], '1', '9')) {
11505 vFAIL("Illegal pattern");
11507 goto parse_recursion;
11509 case '-': /* (?-1) */
11510 if (! inRANGE(RExC_parse[0], '1', '9')) {
11511 RExC_parse--; /* rewind to let it be handled later */
11515 case '1': case '2': case '3': case '4': /* (?1) */
11516 case '5': case '6': case '7': case '8': case '9':
11517 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11520 bool is_neg = FALSE;
11522 parse_start = RExC_parse - 1; /* MJD */
11523 if (*RExC_parse == '-') {
11528 if (grok_atoUV(RExC_parse, &unum, &endptr)
11532 RExC_parse = (char*)endptr;
11536 /* Some limit for num? */
11540 if (*RExC_parse!=')')
11541 vFAIL("Expecting close bracket");
11544 if ( paren == '-' ) {
11546 Diagram of capture buffer numbering.
11547 Top line is the normal capture buffer numbers
11548 Bottom line is the negative indexing as from
11552 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11556 num = RExC_npar + num;
11559 /* It might be a forward reference; we can't fail until
11560 * we know, by completing the parse to get all the
11561 * groups, and then reparsing */
11562 if (ALL_PARENS_COUNTED) {
11564 vFAIL("Reference to nonexistent group");
11567 REQUIRE_PARENS_PASS;
11570 } else if ( paren == '+' ) {
11571 num = RExC_npar + num - 1;
11573 /* We keep track how many GOSUB items we have produced.
11574 To start off the ARG2L() of the GOSUB holds its "id",
11575 which is used later in conjunction with RExC_recurse
11576 to calculate the offset we need to jump for the GOSUB,
11577 which it will store in the final representation.
11578 We have to defer the actual calculation until much later
11579 as the regop may move.
11582 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11583 if (num >= RExC_npar) {
11585 /* It might be a forward reference; we can't fail until we
11586 * know, by completing the parse to get all the groups, and
11587 * then reparsing */
11588 if (ALL_PARENS_COUNTED) {
11589 if (num >= RExC_total_parens) {
11591 vFAIL("Reference to nonexistent group");
11595 REQUIRE_PARENS_PASS;
11598 RExC_recurse_count++;
11599 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11600 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11601 22, "| |", (int)(depth * 2 + 1), "",
11602 (UV)ARG(REGNODE_p(ret)),
11603 (IV)ARG2L(REGNODE_p(ret))));
11604 RExC_seen |= REG_RECURSE_SEEN;
11606 Set_Node_Length(REGNODE_p(ret),
11607 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11608 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11610 *flagp |= POSTPONED;
11611 assert(*RExC_parse == ')');
11612 nextchar(pRExC_state);
11617 case '?': /* (??...) */
11619 if (*RExC_parse != '{') {
11620 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11621 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11623 "Sequence (%" UTF8f "...) not recognized",
11624 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11625 NOT_REACHED; /*NOTREACHED*/
11627 *flagp |= POSTPONED;
11631 case '{': /* (?{...}) */
11634 struct reg_code_block *cb;
11637 RExC_seen_zerolen++;
11639 if ( !pRExC_state->code_blocks
11640 || pRExC_state->code_index
11641 >= pRExC_state->code_blocks->count
11642 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11643 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11646 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11647 FAIL("panic: Sequence (?{...}): no code block found\n");
11648 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11650 /* this is a pre-compiled code block (?{...}) */
11651 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11652 RExC_parse = RExC_start + cb->end;
11654 if (cb->src_regex) {
11655 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11656 RExC_rxi->data->data[n] =
11657 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11658 RExC_rxi->data->data[n+1] = (void*)o;
11661 n = add_data(pRExC_state,
11662 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11663 RExC_rxi->data->data[n] = (void*)o;
11665 pRExC_state->code_index++;
11666 nextchar(pRExC_state);
11669 regnode_offset eval;
11670 ret = reg_node(pRExC_state, LOGICAL);
11672 eval = reg2Lanode(pRExC_state, EVAL,
11675 /* for later propagation into (??{})
11677 RExC_flags & RXf_PMf_COMPILETIME
11679 FLAGS(REGNODE_p(ret)) = 2;
11680 if (! REGTAIL(pRExC_state, ret, eval)) {
11681 REQUIRE_BRANCHJ(flagp, 0);
11683 /* deal with the length of this later - MJD */
11686 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11687 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11688 Set_Node_Offset(REGNODE_p(ret), parse_start);
11691 case '(': /* (?(?{...})...) and (?(?=...)...) */
11694 const int DEFINE_len = sizeof("DEFINE") - 1;
11695 if ( RExC_parse < RExC_end - 1
11696 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11697 && ( RExC_parse[1] == '='
11698 || RExC_parse[1] == '!'
11699 || RExC_parse[1] == '<'
11700 || RExC_parse[1] == '{'))
11701 || ( RExC_parse[0] == '*' /* (?(*...)) */
11702 && ( memBEGINs(RExC_parse + 1,
11703 (Size_t) (RExC_end - (RExC_parse + 1)),
11705 || memBEGINs(RExC_parse + 1,
11706 (Size_t) (RExC_end - (RExC_parse + 1)),
11708 || memBEGINs(RExC_parse + 1,
11709 (Size_t) (RExC_end - (RExC_parse + 1)),
11711 || memBEGINs(RExC_parse + 1,
11712 (Size_t) (RExC_end - (RExC_parse + 1)),
11714 || memBEGINs(RExC_parse + 1,
11715 (Size_t) (RExC_end - (RExC_parse + 1)),
11716 "positive_lookahead:")
11717 || memBEGINs(RExC_parse + 1,
11718 (Size_t) (RExC_end - (RExC_parse + 1)),
11719 "positive_lookbehind:")
11720 || memBEGINs(RExC_parse + 1,
11721 (Size_t) (RExC_end - (RExC_parse + 1)),
11722 "negative_lookahead:")
11723 || memBEGINs(RExC_parse + 1,
11724 (Size_t) (RExC_end - (RExC_parse + 1)),
11725 "negative_lookbehind:"))))
11726 ) { /* Lookahead or eval. */
11728 regnode_offset tail;
11730 ret = reg_node(pRExC_state, LOGICAL);
11731 FLAGS(REGNODE_p(ret)) = 1;
11733 tail = reg(pRExC_state, 1, &flag, depth+1);
11734 RETURN_FAIL_ON_RESTART(flag, flagp);
11735 if (! REGTAIL(pRExC_state, ret, tail)) {
11736 REQUIRE_BRANCHJ(flagp, 0);
11740 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11741 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11743 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11744 char *name_start= RExC_parse++;
11746 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11747 if ( RExC_parse == name_start
11748 || RExC_parse >= RExC_end
11749 || *RExC_parse != ch)
11751 vFAIL2("Sequence (?(%c... not terminated",
11752 (ch == '>' ? '<' : ch));
11756 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11757 RExC_rxi->data->data[num]=(void*)sv_dat;
11758 SvREFCNT_inc_simple_void_NN(sv_dat);
11760 ret = reganode(pRExC_state, GROUPPN, num);
11761 goto insert_if_check_paren;
11763 else if (memBEGINs(RExC_parse,
11764 (STRLEN) (RExC_end - RExC_parse),
11767 ret = reganode(pRExC_state, DEFINEP, 0);
11768 RExC_parse += DEFINE_len;
11770 goto insert_if_check_paren;
11772 else if (RExC_parse[0] == 'R') {
11774 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11775 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11776 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11779 if (RExC_parse[0] == '0') {
11783 else if (inRANGE(RExC_parse[0], '1', '9')) {
11786 if (grok_atoUV(RExC_parse, &uv, &endptr)
11789 parno = (I32)uv + 1;
11790 RExC_parse = (char*)endptr;
11792 /* else "Switch condition not recognized" below */
11793 } else if (RExC_parse[0] == '&') {
11796 sv_dat = reg_scan_name(pRExC_state,
11797 REG_RSN_RETURN_DATA);
11799 parno = 1 + *((I32 *)SvPVX(sv_dat));
11801 ret = reganode(pRExC_state, INSUBP, parno);
11802 goto insert_if_check_paren;
11804 else if (inRANGE(RExC_parse[0], '1', '9')) {
11809 if (grok_atoUV(RExC_parse, &uv, &endptr)
11813 RExC_parse = (char*)endptr;
11816 vFAIL("panic: grok_atoUV returned FALSE");
11818 ret = reganode(pRExC_state, GROUPP, parno);
11820 insert_if_check_paren:
11821 if (UCHARAT(RExC_parse) != ')') {
11823 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11825 vFAIL("Switch condition not recognized");
11827 nextchar(pRExC_state);
11829 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
11832 REQUIRE_BRANCHJ(flagp, 0);
11834 br = regbranch(pRExC_state, &flags, 1, depth+1);
11836 RETURN_FAIL_ON_RESTART(flags,flagp);
11837 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11840 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
11843 REQUIRE_BRANCHJ(flagp, 0);
11845 c = UCHARAT(RExC_parse);
11846 nextchar(pRExC_state);
11847 if (flags&HASWIDTH)
11848 *flagp |= HASWIDTH;
11851 vFAIL("(?(DEFINE)....) does not allow branches");
11853 /* Fake one for optimizer. */
11854 lastbr = reganode(pRExC_state, IFTHEN, 0);
11856 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
11857 RETURN_FAIL_ON_RESTART(flags, flagp);
11858 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11861 if (! REGTAIL(pRExC_state, ret, lastbr)) {
11862 REQUIRE_BRANCHJ(flagp, 0);
11864 if (flags&HASWIDTH)
11865 *flagp |= HASWIDTH;
11866 c = UCHARAT(RExC_parse);
11867 nextchar(pRExC_state);
11872 if (RExC_parse >= RExC_end)
11873 vFAIL("Switch (?(condition)... not terminated");
11875 vFAIL("Switch (?(condition)... contains too many branches");
11877 ender = reg_node(pRExC_state, TAIL);
11878 if (! REGTAIL(pRExC_state, br, ender)) {
11879 REQUIRE_BRANCHJ(flagp, 0);
11882 if (! REGTAIL(pRExC_state, lastbr, ender)) {
11883 REQUIRE_BRANCHJ(flagp, 0);
11885 if (! REGTAIL(pRExC_state,
11888 NEXTOPER(REGNODE_p(lastbr)))),
11891 REQUIRE_BRANCHJ(flagp, 0);
11895 if (! REGTAIL(pRExC_state, ret, ender)) {
11896 REQUIRE_BRANCHJ(flagp, 0);
11898 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
11899 RExC_size++; /* XXX WHY do we need this?!!
11900 For large programs it seems to be required
11901 but I can't figure out why. -- dmq*/
11906 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11908 vFAIL("Unknown switch condition (?(...))");
11910 case '[': /* (?[ ... ]) */
11911 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
11913 case 0: /* A NUL */
11914 RExC_parse--; /* for vFAIL to print correctly */
11915 vFAIL("Sequence (? incomplete");
11919 if (RExC_strict) { /* [perl #132851] */
11920 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
11923 default: /* e.g., (?i) */
11924 RExC_parse = (char *) seqstart + 1;
11926 parse_lparen_question_flags(pRExC_state);
11927 if (UCHARAT(RExC_parse) != ':') {
11928 if (RExC_parse < RExC_end)
11929 nextchar(pRExC_state);
11934 nextchar(pRExC_state);
11939 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11943 if (! ALL_PARENS_COUNTED) {
11944 /* If we are in our first pass through (and maybe only pass),
11945 * we need to allocate memory for the capturing parentheses
11949 if (!RExC_parens_buf_size) {
11950 /* first guess at number of parens we might encounter */
11951 RExC_parens_buf_size = 10;
11953 /* setup RExC_open_parens, which holds the address of each
11954 * OPEN tag, and to make things simpler for the 0 index the
11955 * start of the program - this is used later for offsets */
11956 Newxz(RExC_open_parens, RExC_parens_buf_size,
11958 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
11960 /* setup RExC_close_parens, which holds the address of each
11961 * CLOSE tag, and to make things simpler for the 0 index
11962 * the end of the program - this is used later for offsets
11964 Newxz(RExC_close_parens, RExC_parens_buf_size,
11966 /* we dont know where end op starts yet, so we dont need to
11967 * set RExC_close_parens[0] like we do RExC_open_parens[0]
11970 else if (RExC_npar > RExC_parens_buf_size) {
11971 I32 old_size = RExC_parens_buf_size;
11973 RExC_parens_buf_size *= 2;
11975 Renew(RExC_open_parens, RExC_parens_buf_size,
11977 Zero(RExC_open_parens + old_size,
11978 RExC_parens_buf_size - old_size, regnode_offset);
11980 Renew(RExC_close_parens, RExC_parens_buf_size,
11982 Zero(RExC_close_parens + old_size,
11983 RExC_parens_buf_size - old_size, regnode_offset);
11987 ret = reganode(pRExC_state, OPEN, parno);
11988 if (!RExC_nestroot)
11989 RExC_nestroot = parno;
11990 if (RExC_open_parens && !RExC_open_parens[parno])
11992 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11993 "%*s%*s Setting open paren #%" IVdf " to %d\n",
11994 22, "| |", (int)(depth * 2 + 1), "",
11996 RExC_open_parens[parno]= ret;
11999 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12000 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12003 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12012 /* Pick up the branches, linking them together. */
12013 parse_start = RExC_parse; /* MJD */
12014 br = regbranch(pRExC_state, &flags, 1, depth+1);
12016 /* branch_len = (paren != 0); */
12019 RETURN_FAIL_ON_RESTART(flags, flagp);
12020 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12022 if (*RExC_parse == '|') {
12023 if (RExC_use_BRANCHJ) {
12024 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12027 reginsert(pRExC_state, BRANCH, br, depth+1);
12028 Set_Node_Length(REGNODE_p(br), paren != 0);
12029 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12033 else if (paren == ':') {
12034 *flagp |= flags&SIMPLE;
12036 if (is_open) { /* Starts with OPEN. */
12037 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12038 REQUIRE_BRANCHJ(flagp, 0);
12041 else if (paren != '?') /* Not Conditional */
12043 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12045 while (*RExC_parse == '|') {
12046 if (RExC_use_BRANCHJ) {
12049 ender = reganode(pRExC_state, LONGJMP, 0);
12051 /* Append to the previous. */
12052 shut_gcc_up = REGTAIL(pRExC_state,
12053 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12055 PERL_UNUSED_VAR(shut_gcc_up);
12057 nextchar(pRExC_state);
12058 if (freeze_paren) {
12059 if (RExC_npar > after_freeze)
12060 after_freeze = RExC_npar;
12061 RExC_npar = freeze_paren;
12063 br = regbranch(pRExC_state, &flags, 0, depth+1);
12066 RETURN_FAIL_ON_RESTART(flags, flagp);
12067 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12069 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12070 REQUIRE_BRANCHJ(flagp, 0);
12073 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12076 if (have_branch || paren != ':') {
12079 /* Make a closing node, and hook it on the end. */
12082 ender = reg_node(pRExC_state, TAIL);
12085 ender = reganode(pRExC_state, CLOSE, parno);
12086 if ( RExC_close_parens ) {
12087 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12088 "%*s%*s Setting close paren #%" IVdf " to %d\n",
12089 22, "| |", (int)(depth * 2 + 1), "",
12090 (IV)parno, ender));
12091 RExC_close_parens[parno]= ender;
12092 if (RExC_nestroot == parno)
12095 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12096 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12099 ender = reg_node(pRExC_state, SRCLOSE);
12100 RExC_in_script_run = 0;
12110 *flagp &= ~HASWIDTH;
12112 case 't': /* aTomic */
12114 ender = reg_node(pRExC_state, SUCCEED);
12117 ender = reg_node(pRExC_state, END);
12118 assert(!RExC_end_op); /* there can only be one! */
12119 RExC_end_op = REGNODE_p(ender);
12120 if (RExC_close_parens) {
12121 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12122 "%*s%*s Setting close paren #0 (END) to %d\n",
12123 22, "| |", (int)(depth * 2 + 1), "",
12126 RExC_close_parens[0]= ender;
12131 DEBUG_PARSE_MSG("lsbr");
12132 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12133 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12134 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12135 SvPV_nolen_const(RExC_mysv1),
12137 SvPV_nolen_const(RExC_mysv2),
12139 (IV)(ender - lastbr)
12142 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12143 REQUIRE_BRANCHJ(flagp, 0);
12147 char is_nothing= 1;
12149 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12151 /* Hook the tails of the branches to the closing node. */
12152 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12153 const U8 op = PL_regkind[OP(br)];
12154 if (op == BRANCH) {
12155 if (! REGTAIL_STUDY(pRExC_state,
12156 REGNODE_OFFSET(NEXTOPER(br)),
12159 REQUIRE_BRANCHJ(flagp, 0);
12161 if ( OP(NEXTOPER(br)) != NOTHING
12162 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12165 else if (op == BRANCHJ) {
12166 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12167 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12169 PERL_UNUSED_VAR(shut_gcc_up);
12170 /* for now we always disable this optimisation * /
12171 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12172 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12178 regnode * ret_as_regnode = REGNODE_p(ret);
12179 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12180 ? regnext(ret_as_regnode)
12183 DEBUG_PARSE_MSG("NADA");
12184 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12185 NULL, pRExC_state);
12186 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12187 NULL, pRExC_state);
12188 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12189 SvPV_nolen_const(RExC_mysv1),
12190 (IV)REG_NODE_NUM(ret_as_regnode),
12191 SvPV_nolen_const(RExC_mysv2),
12197 if (OP(REGNODE_p(ender)) == TAIL) {
12199 RExC_emit= REGNODE_OFFSET(br) + 1;
12202 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12203 OP(opt)= OPTIMIZED;
12204 NEXT_OFF(br)= REGNODE_p(ender) - br;
12212 /* Even/odd or x=don't care: 010101x10x */
12213 static const char parens[] = "=!aA<,>Bbt";
12214 /* flag below is set to 0 up through 'A'; 1 for larger */
12216 if (paren && (p = strchr(parens, paren))) {
12217 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12218 int flag = (p - parens) > 3;
12220 if (paren == '>' || paren == 't') {
12221 node = SUSPEND, flag = 0;
12224 reginsert(pRExC_state, node, ret, depth+1);
12225 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12226 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12227 FLAGS(REGNODE_p(ret)) = flag;
12228 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12230 REQUIRE_BRANCHJ(flagp, 0);
12235 /* Check for proper termination. */
12237 /* restore original flags, but keep (?p) and, if we've encountered
12238 * something in the parse that changes /d rules into /u, keep the /u */
12239 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12240 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
12241 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12243 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12244 RExC_parse = oregcomp_parse;
12245 vFAIL("Unmatched (");
12247 nextchar(pRExC_state);
12249 else if (!paren && RExC_parse < RExC_end) {
12250 if (*RExC_parse == ')') {
12252 vFAIL("Unmatched )");
12255 FAIL("Junk on end of regexp"); /* "Can't happen". */
12256 NOT_REACHED; /* NOTREACHED */
12259 if (RExC_in_lookbehind) {
12260 RExC_in_lookbehind--;
12262 if (RExC_in_lookahead) {
12263 RExC_in_lookahead--;
12265 if (after_freeze > RExC_npar)
12266 RExC_npar = after_freeze;
12271 - regbranch - one alternative of an | operator
12273 * Implements the concatenation operator.
12275 * On success, returns the offset at which any next node should be placed into
12276 * the regex engine program being compiled.
12278 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12279 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12282 STATIC regnode_offset
12283 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12285 regnode_offset ret;
12286 regnode_offset chain = 0;
12287 regnode_offset latest;
12288 I32 flags = 0, c = 0;
12289 GET_RE_DEBUG_FLAGS_DECL;
12291 PERL_ARGS_ASSERT_REGBRANCH;
12293 DEBUG_PARSE("brnc");
12298 if (RExC_use_BRANCHJ)
12299 ret = reganode(pRExC_state, BRANCHJ, 0);
12301 ret = reg_node(pRExC_state, BRANCH);
12302 Set_Node_Length(REGNODE_p(ret), 1);
12306 *flagp = WORST; /* Tentatively. */
12308 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12309 FALSE /* Don't force to /x */ );
12310 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12311 flags &= ~TRYAGAIN;
12312 latest = regpiece(pRExC_state, &flags, depth+1);
12314 if (flags & TRYAGAIN)
12316 RETURN_FAIL_ON_RESTART(flags, flagp);
12317 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12321 *flagp |= flags&(HASWIDTH|POSTPONED);
12322 if (chain == 0) /* First piece. */
12323 *flagp |= flags&SPSTART;
12325 /* FIXME adding one for every branch after the first is probably
12326 * excessive now we have TRIE support. (hv) */
12328 if (! REGTAIL(pRExC_state, chain, latest)) {
12329 /* XXX We could just redo this branch, but figuring out what
12330 * bookkeeping needs to be reset is a pain, and it's likely
12331 * that other branches that goto END will also be too large */
12332 REQUIRE_BRANCHJ(flagp, 0);
12338 if (chain == 0) { /* Loop ran zero times. */
12339 chain = reg_node(pRExC_state, NOTHING);
12344 *flagp |= flags&SIMPLE;
12351 - regpiece - something followed by possible quantifier * + ? {n,m}
12353 * Note that the branching code sequences used for ? and the general cases
12354 * of * and + are somewhat optimized: they use the same NOTHING node as
12355 * both the endmarker for their branch list and the body of the last branch.
12356 * It might seem that this node could be dispensed with entirely, but the
12357 * endmarker role is not redundant.
12359 * On success, returns the offset at which any next node should be placed into
12360 * the regex engine program being compiled.
12362 * Returns 0 otherwise, with *flagp set to indicate why:
12363 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12364 * RESTART_PARSE if the parse needs to be restarted, or'd with
12365 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12367 STATIC regnode_offset
12368 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12370 regnode_offset ret;
12374 const char * const origparse = RExC_parse;
12376 I32 max = REG_INFTY;
12377 #ifdef RE_TRACK_PATTERN_OFFSETS
12380 const char *maxpos = NULL;
12383 /* Save the original in case we change the emitted regop to a FAIL. */
12384 const regnode_offset orig_emit = RExC_emit;
12386 GET_RE_DEBUG_FLAGS_DECL;
12388 PERL_ARGS_ASSERT_REGPIECE;
12390 DEBUG_PARSE("piec");
12392 ret = regatom(pRExC_state, &flags, depth+1);
12394 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12395 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12400 if (op == '{' && regcurly(RExC_parse)) {
12402 #ifdef RE_TRACK_PATTERN_OFFSETS
12403 parse_start = RExC_parse; /* MJD */
12405 next = RExC_parse + 1;
12406 while (isDIGIT(*next) || *next == ',') {
12407 if (*next == ',') {
12415 if (*next == '}') { /* got one */
12416 const char* endptr;
12420 if (isDIGIT(*RExC_parse)) {
12422 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12423 vFAIL("Invalid quantifier in {,}");
12424 if (uv >= REG_INFTY)
12425 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12430 if (*maxpos == ',')
12433 maxpos = RExC_parse;
12434 if (isDIGIT(*maxpos)) {
12436 if (!grok_atoUV(maxpos, &uv, &endptr))
12437 vFAIL("Invalid quantifier in {,}");
12438 if (uv >= REG_INFTY)
12439 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12442 max = REG_INFTY; /* meaning "infinity" */
12445 nextchar(pRExC_state);
12446 if (max < min) { /* If can't match, warn and optimize to fail
12448 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12449 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12450 NEXT_OFF(REGNODE_p(orig_emit)) =
12451 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12454 else if (min == max && *RExC_parse == '?')
12456 ckWARN2reg(RExC_parse + 1,
12457 "Useless use of greediness modifier '%c'",
12462 if ((flags&SIMPLE)) {
12463 if (min == 0 && max == REG_INFTY) {
12464 reginsert(pRExC_state, STAR, ret, depth+1);
12466 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12469 if (min == 1 && max == REG_INFTY) {
12470 reginsert(pRExC_state, PLUS, ret, depth+1);
12472 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12475 MARK_NAUGHTY_EXP(2, 2);
12476 reginsert(pRExC_state, CURLY, ret, depth+1);
12477 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12478 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12481 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12483 FLAGS(REGNODE_p(w)) = 0;
12484 if (! REGTAIL(pRExC_state, ret, w)) {
12485 REQUIRE_BRANCHJ(flagp, 0);
12487 if (RExC_use_BRANCHJ) {
12488 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12489 reginsert(pRExC_state, NOTHING, ret, depth+1);
12490 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12492 reginsert(pRExC_state, CURLYX, ret, depth+1);
12494 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12495 Set_Node_Length(REGNODE_p(ret),
12496 op == '{' ? (RExC_parse - parse_start) : 1);
12498 if (RExC_use_BRANCHJ)
12499 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12501 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12504 REQUIRE_BRANCHJ(flagp, 0);
12506 RExC_whilem_seen++;
12507 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12509 FLAGS(REGNODE_p(ret)) = 0;
12514 *flagp |= HASWIDTH;
12515 ARG1_SET(REGNODE_p(ret), (U16)min);
12516 ARG2_SET(REGNODE_p(ret), (U16)max);
12517 if (max == REG_INFTY)
12518 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12524 if (!ISMULT1(op)) {
12529 #if 0 /* Now runtime fix should be reliable. */
12531 /* if this is reinstated, don't forget to put this back into perldiag:
12533 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12535 (F) The part of the regexp subject to either the * or + quantifier
12536 could match an empty string. The {#} shows in the regular
12537 expression about where the problem was discovered.
12541 if (!(flags&HASWIDTH) && op != '?')
12542 vFAIL("Regexp *+ operand could be empty");
12545 #ifdef RE_TRACK_PATTERN_OFFSETS
12546 parse_start = RExC_parse;
12548 nextchar(pRExC_state);
12550 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12556 else if (op == '+') {
12560 else if (op == '?') {
12565 if (!(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12566 ckWARN2reg(RExC_parse,
12567 "%" UTF8f " matches null string many times",
12568 UTF8fARG(UTF, (RExC_parse >= origparse
12569 ? RExC_parse - origparse
12574 if (*RExC_parse == '?') {
12575 nextchar(pRExC_state);
12576 reginsert(pRExC_state, MINMOD, ret, depth+1);
12577 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12578 REQUIRE_BRANCHJ(flagp, 0);
12581 else if (*RExC_parse == '+') {
12582 regnode_offset ender;
12583 nextchar(pRExC_state);
12584 ender = reg_node(pRExC_state, SUCCEED);
12585 if (! REGTAIL(pRExC_state, ret, ender)) {
12586 REQUIRE_BRANCHJ(flagp, 0);
12588 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12589 ender = reg_node(pRExC_state, TAIL);
12590 if (! REGTAIL(pRExC_state, ret, ender)) {
12591 REQUIRE_BRANCHJ(flagp, 0);
12595 if (ISMULT2(RExC_parse)) {
12597 vFAIL("Nested quantifiers");
12604 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12605 regnode_offset * node_p,
12613 /* This routine teases apart the various meanings of \N and returns
12614 * accordingly. The input parameters constrain which meaning(s) is/are valid
12615 * in the current context.
12617 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12619 * If <code_point_p> is not NULL, the context is expecting the result to be a
12620 * single code point. If this \N instance turns out to a single code point,
12621 * the function returns TRUE and sets *code_point_p to that code point.
12623 * If <node_p> is not NULL, the context is expecting the result to be one of
12624 * the things representable by a regnode. If this \N instance turns out to be
12625 * one such, the function generates the regnode, returns TRUE and sets *node_p
12626 * to point to the offset of that regnode into the regex engine program being
12629 * If this instance of \N isn't legal in any context, this function will
12630 * generate a fatal error and not return.
12632 * On input, RExC_parse should point to the first char following the \N at the
12633 * time of the call. On successful return, RExC_parse will have been updated
12634 * to point to just after the sequence identified by this routine. Also
12635 * *flagp has been updated as needed.
12637 * When there is some problem with the current context and this \N instance,
12638 * the function returns FALSE, without advancing RExC_parse, nor setting
12639 * *node_p, nor *code_point_p, nor *flagp.
12641 * If <cp_count> is not NULL, the caller wants to know the length (in code
12642 * points) that this \N sequence matches. This is set, and the input is
12643 * parsed for errors, even if the function returns FALSE, as detailed below.
12645 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12647 * Probably the most common case is for the \N to specify a single code point.
12648 * *cp_count will be set to 1, and *code_point_p will be set to that code
12651 * Another possibility is for the input to be an empty \N{}. This is no
12652 * longer accepted, and will generate a fatal error.
12654 * Another possibility is for a custom charnames handler to be in effect which
12655 * translates the input name to an empty string. *cp_count will be set to 0.
12656 * *node_p will be set to a generated NOTHING node.
12658 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12659 * set to 0. *node_p will be set to a generated REG_ANY node.
12661 * The fifth possibility is that \N resolves to a sequence of more than one
12662 * code points. *cp_count will be set to the number of code points in the
12663 * sequence. *node_p will be set to a generated node returned by this
12664 * function calling S_reg().
12666 * The final possibility is that it is premature to be calling this function;
12667 * the parse needs to be restarted. This can happen when this changes from
12668 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12669 * latter occurs only when the fifth possibility would otherwise be in
12670 * effect, and is because one of those code points requires the pattern to be
12671 * recompiled as UTF-8. The function returns FALSE, and sets the
12672 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12673 * happens, the caller needs to desist from continuing parsing, and return
12674 * this information to its caller. This is not set for when there is only one
12675 * code point, as this can be called as part of an ANYOF node, and they can
12676 * store above-Latin1 code points without the pattern having to be in UTF-8.
12678 * For non-single-quoted regexes, the tokenizer has resolved character and
12679 * sequence names inside \N{...} into their Unicode values, normalizing the
12680 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12681 * hex-represented code points in the sequence. This is done there because
12682 * the names can vary based on what charnames pragma is in scope at the time,
12683 * so we need a way to take a snapshot of what they resolve to at the time of
12684 * the original parse. [perl #56444].
12686 * That parsing is skipped for single-quoted regexes, so here we may get
12687 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12688 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12689 * the native character set for non-ASCII platforms. The other possibilities
12690 * are already native, so no translation is done. */
12692 char * endbrace; /* points to '}' following the name */
12693 char* p = RExC_parse; /* Temporary */
12695 SV * substitute_parse = NULL;
12700 GET_RE_DEBUG_FLAGS_DECL;
12702 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12704 GET_RE_DEBUG_FLAGS;
12706 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12707 assert(! (node_p && cp_count)); /* At most 1 should be set */
12709 if (cp_count) { /* Initialize return for the most common case */
12713 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12714 * modifier. The other meanings do not, so use a temporary until we find
12715 * out which we are being called with */
12716 skip_to_be_ignored_text(pRExC_state, &p,
12717 FALSE /* Don't force to /x */ );
12719 /* Disambiguate between \N meaning a named character versus \N meaning
12720 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12721 * quantifier, or if there is no '{' at all */
12722 if (*p != '{' || regcurly(p)) {
12732 *node_p = reg_node(pRExC_state, REG_ANY);
12733 *flagp |= HASWIDTH|SIMPLE;
12735 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
12739 /* The test above made sure that the next real character is a '{', but
12740 * under the /x modifier, it could be separated by space (or a comment and
12741 * \n) and this is not allowed (for consistency with \x{...} and the
12742 * tokenizer handling of \N{NAME}). */
12743 if (*RExC_parse != '{') {
12744 vFAIL("Missing braces on \\N{}");
12747 RExC_parse++; /* Skip past the '{' */
12749 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12750 if (! endbrace) { /* no trailing brace */
12751 vFAIL2("Missing right brace on \\%c{}", 'N');
12754 /* Here, we have decided it should be a named character or sequence. These
12755 * imply Unicode semantics */
12756 REQUIRE_UNI_RULES(flagp, FALSE);
12758 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
12759 * nothing at all (not allowed under strict) */
12760 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
12761 RExC_parse = endbrace;
12763 RExC_parse++; /* Position after the "}" */
12764 vFAIL("Zero length \\N{}");
12770 nextchar(pRExC_state);
12775 *node_p = reg_node(pRExC_state, NOTHING);
12779 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
12781 /* Here, the name isn't of the form U+.... This can happen if the
12782 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
12783 * is the time to find out what the name means */
12785 const STRLEN name_len = endbrace - RExC_parse;
12786 SV * value_sv; /* What does this name evaluate to */
12788 const U8 * value; /* string of name's value */
12789 STRLEN value_len; /* and its length */
12791 /* RExC_unlexed_names is a hash of names that weren't evaluated by
12792 * toke.c, and their values. Make sure is initialized */
12793 if (! RExC_unlexed_names) {
12794 RExC_unlexed_names = newHV();
12797 /* If we have already seen this name in this pattern, use that. This
12798 * allows us to only call the charnames handler once per name per
12799 * pattern. A broken or malicious handler could return something
12800 * different each time, which could cause the results to vary depending
12801 * on if something gets added or subtracted from the pattern that
12802 * causes the number of passes to change, for example */
12803 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
12806 value_sv = *value_svp;
12808 else { /* Otherwise we have to go out and get the name */
12809 const char * error_msg = NULL;
12810 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
12814 RExC_parse = endbrace;
12818 /* If no error message, should have gotten a valid return */
12821 /* Save the name's meaning for later use */
12822 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
12825 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
12829 /* Here, we have the value the name evaluates to in 'value_sv' */
12830 value = (U8 *) SvPV(value_sv, value_len);
12832 /* See if the result is one code point vs 0 or multiple */
12833 if (value_len > 0 && value_len <= (UV) ((SvUTF8(value_sv))
12837 /* Here, exactly one code point. If that isn't what is wanted,
12839 if (! code_point_p) {
12844 /* Convert from string to numeric code point */
12845 *code_point_p = (SvUTF8(value_sv))
12846 ? valid_utf8_to_uvchr(value, NULL)
12849 /* Have parsed this entire single code point \N{...}. *cp_count
12850 * has already been set to 1, so don't do it again. */
12851 RExC_parse = endbrace;
12852 nextchar(pRExC_state);
12854 } /* End of is a single code point */
12856 /* Count the code points, if caller desires. The API says to do this
12857 * even if we will later return FALSE */
12861 *cp_count = (SvUTF8(value_sv))
12862 ? utf8_length(value, value + value_len)
12866 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12867 * But don't back the pointer up if the caller wants to know how many
12868 * code points there are (they need to handle it themselves in this
12877 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
12878 * reg recursively to parse it. That way, it retains its atomicness,
12879 * while not having to worry about any special handling that some code
12880 * points may have. */
12882 substitute_parse = newSVpvs("?:");
12883 sv_catsv(substitute_parse, value_sv);
12884 sv_catpv(substitute_parse, ")");
12886 /* The value should already be native, so no need to convert on EBCDIC
12888 assert(! RExC_recode_x_to_native);
12891 else { /* \N{U+...} */
12892 Size_t count = 0; /* code point count kept internally */
12894 /* We can get to here when the input is \N{U+...} or when toke.c has
12895 * converted a name to the \N{U+...} form. This include changing a
12896 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
12898 RExC_parse += 2; /* Skip past the 'U+' */
12900 /* Code points are separated by dots. The '}' terminates the whole
12903 do { /* Loop until the ending brace */
12905 char * start_digit; /* The first of the current code point */
12906 if (! isXDIGIT(*RExC_parse)) {
12908 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12911 start_digit = RExC_parse;
12914 /* Loop through the hex digits of the current code point */
12916 /* Adding this digit will shift the result 4 bits. If that
12917 * result would be above the legal max, it's overflow */
12918 if (cp > MAX_LEGAL_CP >> 4) {
12920 /* Find the end of the code point */
12923 } while (isXDIGIT(*RExC_parse) || *RExC_parse == '_');
12925 /* Be sure to synchronize this message with the similar one
12927 vFAIL4("Use of code point 0x%.*s is not allowed; the"
12928 " permissible max is 0x%" UVxf,
12929 (int) (RExC_parse - start_digit), start_digit,
12933 /* Accumulate this (valid) digit into the running total */
12934 cp = (cp << 4) + READ_XDIGIT(RExC_parse);
12936 /* READ_XDIGIT advanced the input pointer. Ignore a single
12937 * underscore separator */
12938 if (*RExC_parse == '_' && isXDIGIT(RExC_parse[1])) {
12941 } while (isXDIGIT(*RExC_parse));
12943 /* Here, have accumulated the next code point */
12944 if (RExC_parse >= endbrace) { /* If done ... */
12949 /* Here, is a single code point; fail if doesn't want that */
12950 if (! code_point_p) {
12955 /* A single code point is easy to handle; just return it */
12956 *code_point_p = UNI_TO_NATIVE(cp);
12957 RExC_parse = endbrace;
12958 nextchar(pRExC_state);
12962 /* Here, the only legal thing would be a multiple character
12963 * sequence (of the form "\N{U+c1.c2. ... }". So the next
12964 * character must be a dot (and the one after that can't be the
12965 * endbrace, or we'd have something like \N{U+100.} ) */
12966 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
12967 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12968 ? UTF8SKIP(RExC_parse)
12970 if (RExC_parse >= endbrace) { /* Guard against malformed utf8 */
12971 RExC_parse = endbrace;
12973 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12976 /* Here, looks like its really a multiple character sequence. Fail
12977 * if that's not what the caller wants. But continue with counting
12978 * and error checking if they still want a count */
12979 if (! node_p && ! cp_count) {
12983 /* What is done here is to convert this to a sub-pattern of the
12984 * form \x{char1}\x{char2}... and then call reg recursively to
12985 * parse it (enclosing in "(?: ... )" ). That way, it retains its
12986 * atomicness, while not having to worry about special handling
12987 * that some code points may have. We don't create a subpattern,
12988 * but go through the motions of code point counting and error
12989 * checking, if the caller doesn't want a node returned. */
12991 if (node_p && count == 1) {
12992 substitute_parse = newSVpvs("?:");
12998 /* Convert to notation the rest of the code understands */
12999 sv_catpvs(substitute_parse, "\\x{");
13000 sv_catpvn(substitute_parse, start_digit,
13001 RExC_parse - start_digit);
13002 sv_catpvs(substitute_parse, "}");
13005 /* Move to after the dot (or ending brace the final time through.)
13010 } while (RExC_parse < endbrace);
13012 if (! node_p) { /* Doesn't want the node */
13019 sv_catpvs(substitute_parse, ")");
13021 /* The values are Unicode, and therefore have to be converted to native
13022 * on a non-Unicode (meaning non-ASCII) platform. */
13023 SET_recode_x_to_native(1);
13026 /* Here, we have the string the name evaluates to, ready to be parsed,
13027 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13028 * constructs. This can be called from within a substitute parse already.
13029 * The error reporting mechanism doesn't work for 2 levels of this, but the
13030 * code above has validated this new construct, so there should be no
13031 * errors generated by the below. And this isn' an exact copy, so the
13032 * mechanism to seamlessly deal with this won't work, so turn off warnings
13034 save_start = RExC_start;
13035 orig_end = RExC_end;
13037 RExC_parse = RExC_start = SvPVX(substitute_parse);
13038 RExC_end = RExC_parse + SvCUR(substitute_parse);
13039 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13041 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13043 /* Restore the saved values */
13045 RExC_start = save_start;
13046 RExC_parse = endbrace;
13047 RExC_end = orig_end;
13048 SET_recode_x_to_native(0);
13050 SvREFCNT_dec_NN(substitute_parse);
13053 RETURN_FAIL_ON_RESTART(flags, flagp);
13054 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13057 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13059 nextchar(pRExC_state);
13065 PERL_STATIC_INLINE U8
13066 S_compute_EXACTish(RExC_state_t *pRExC_state)
13070 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13078 op = get_regex_charset(RExC_flags);
13079 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13080 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13081 been, so there is no hole */
13084 return op + EXACTF;
13088 S_new_regcurly(const char *s, const char *e)
13090 /* This is a temporary function designed to match the most lenient form of
13091 * a {m,n} quantifier we ever envision, with either number omitted, and
13092 * spaces anywhere between/before/after them.
13094 * If this function fails, then the string it matches is very unlikely to
13095 * ever be considered a valid quantifier, so we can allow the '{' that
13096 * begins it to be considered as a literal */
13098 bool has_min = FALSE;
13099 bool has_max = FALSE;
13101 PERL_ARGS_ASSERT_NEW_REGCURLY;
13103 if (s >= e || *s++ != '{')
13106 while (s < e && isSPACE(*s)) {
13109 while (s < e && isDIGIT(*s)) {
13113 while (s < e && isSPACE(*s)) {
13119 while (s < e && isSPACE(*s)) {
13122 while (s < e && isDIGIT(*s)) {
13126 while (s < e && isSPACE(*s)) {
13131 return s < e && *s == '}' && (has_min || has_max);
13134 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13135 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13138 S_backref_value(char *p, char *e)
13140 const char* endptr = e;
13142 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13149 - regatom - the lowest level
13151 Try to identify anything special at the start of the current parse position.
13152 If there is, then handle it as required. This may involve generating a
13153 single regop, such as for an assertion; or it may involve recursing, such as
13154 to handle a () structure.
13156 If the string doesn't start with something special then we gobble up
13157 as much literal text as we can. If we encounter a quantifier, we have to
13158 back off the final literal character, as that quantifier applies to just it
13159 and not to the whole string of literals.
13161 Once we have been able to handle whatever type of thing started the
13162 sequence, we return the offset into the regex engine program being compiled
13163 at which any next regnode should be placed.
13165 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13166 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13167 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13168 Otherwise does not return 0.
13170 Note: we have to be careful with escapes, as they can be both literal
13171 and special, and in the case of \10 and friends, context determines which.
13173 A summary of the code structure is:
13175 switch (first_byte) {
13176 cases for each special:
13177 handle this special;
13180 switch (2nd byte) {
13181 cases for each unambiguous special:
13182 handle this special;
13184 cases for each ambigous special/literal:
13186 if (special) handle here
13188 default: // unambiguously literal:
13191 default: // is a literal char
13194 create EXACTish node for literal;
13195 while (more input and node isn't full) {
13196 switch (input_byte) {
13197 cases for each special;
13198 make sure parse pointer is set so that the next call to
13199 regatom will see this special first
13200 goto loopdone; // EXACTish node terminated by prev. char
13202 append char to EXACTISH node;
13204 get next input byte;
13208 return the generated node;
13210 Specifically there are two separate switches for handling
13211 escape sequences, with the one for handling literal escapes requiring
13212 a dummy entry for all of the special escapes that are actually handled
13217 STATIC regnode_offset
13218 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13221 regnode_offset ret = 0;
13227 GET_RE_DEBUG_FLAGS_DECL;
13229 *flagp = WORST; /* Tentatively. */
13231 DEBUG_PARSE("atom");
13233 PERL_ARGS_ASSERT_REGATOM;
13236 parse_start = RExC_parse;
13237 assert(RExC_parse < RExC_end);
13238 switch ((U8)*RExC_parse) {
13240 RExC_seen_zerolen++;
13241 nextchar(pRExC_state);
13242 if (RExC_flags & RXf_PMf_MULTILINE)
13243 ret = reg_node(pRExC_state, MBOL);
13245 ret = reg_node(pRExC_state, SBOL);
13246 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13249 nextchar(pRExC_state);
13251 RExC_seen_zerolen++;
13252 if (RExC_flags & RXf_PMf_MULTILINE)
13253 ret = reg_node(pRExC_state, MEOL);
13255 ret = reg_node(pRExC_state, SEOL);
13256 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13259 nextchar(pRExC_state);
13260 if (RExC_flags & RXf_PMf_SINGLELINE)
13261 ret = reg_node(pRExC_state, SANY);
13263 ret = reg_node(pRExC_state, REG_ANY);
13264 *flagp |= HASWIDTH|SIMPLE;
13266 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13270 char * const oregcomp_parse = ++RExC_parse;
13271 ret = regclass(pRExC_state, flagp, depth+1,
13272 FALSE, /* means parse the whole char class */
13273 TRUE, /* allow multi-char folds */
13274 FALSE, /* don't silence non-portable warnings. */
13275 (bool) RExC_strict,
13276 TRUE, /* Allow an optimized regnode result */
13279 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13280 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13283 if (*RExC_parse != ']') {
13284 RExC_parse = oregcomp_parse;
13285 vFAIL("Unmatched [");
13287 nextchar(pRExC_state);
13288 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13292 nextchar(pRExC_state);
13293 ret = reg(pRExC_state, 2, &flags, depth+1);
13295 if (flags & TRYAGAIN) {
13296 if (RExC_parse >= RExC_end) {
13297 /* Make parent create an empty node if needed. */
13298 *flagp |= TRYAGAIN;
13303 RETURN_FAIL_ON_RESTART(flags, flagp);
13304 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13307 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13311 if (flags & TRYAGAIN) {
13312 *flagp |= TRYAGAIN;
13315 vFAIL("Internal urp");
13316 /* Supposed to be caught earlier. */
13322 vFAIL("Quantifier follows nothing");
13327 This switch handles escape sequences that resolve to some kind
13328 of special regop and not to literal text. Escape sequences that
13329 resolve to literal text are handled below in the switch marked
13332 Every entry in this switch *must* have a corresponding entry
13333 in the literal escape switch. However, the opposite is not
13334 required, as the default for this switch is to jump to the
13335 literal text handling code.
13338 switch ((U8)*RExC_parse) {
13339 /* Special Escapes */
13341 RExC_seen_zerolen++;
13342 ret = reg_node(pRExC_state, SBOL);
13343 /* SBOL is shared with /^/ so we set the flags so we can tell
13344 * /\A/ from /^/ in split. */
13345 FLAGS(REGNODE_p(ret)) = 1;
13347 goto finish_meta_pat;
13349 ret = reg_node(pRExC_state, GPOS);
13350 RExC_seen |= REG_GPOS_SEEN;
13352 goto finish_meta_pat;
13354 if (!RExC_in_lookbehind && !RExC_in_lookahead) {
13355 RExC_seen_zerolen++;
13356 ret = reg_node(pRExC_state, KEEPS);
13358 /* XXX:dmq : disabling in-place substitution seems to
13359 * be necessary here to avoid cases of memory corruption, as
13360 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13362 RExC_seen |= REG_LOOKBEHIND_SEEN;
13363 goto finish_meta_pat;
13366 ++RExC_parse; /* advance past the 'K' */
13367 vFAIL("\\K not permitted in lookahead/lookbehind");
13370 ret = reg_node(pRExC_state, SEOL);
13372 RExC_seen_zerolen++; /* Do not optimize RE away */
13373 goto finish_meta_pat;
13375 ret = reg_node(pRExC_state, EOS);
13377 RExC_seen_zerolen++; /* Do not optimize RE away */
13378 goto finish_meta_pat;
13380 vFAIL("\\C no longer supported");
13382 ret = reg_node(pRExC_state, CLUMP);
13383 *flagp |= HASWIDTH;
13384 goto finish_meta_pat;
13392 regex_charset charset = get_regex_charset(RExC_flags);
13394 RExC_seen_zerolen++;
13395 RExC_seen |= REG_LOOKBEHIND_SEEN;
13396 op = BOUND + charset;
13398 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13399 flags = TRADITIONAL_BOUND;
13400 if (op > BOUNDA) { /* /aa is same as /a */
13406 char name = *RExC_parse;
13407 char * endbrace = NULL;
13409 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13412 vFAIL2("Missing right brace on \\%c{}", name);
13414 /* XXX Need to decide whether to take spaces or not. Should be
13415 * consistent with \p{}, but that currently is SPACE, which
13416 * means vertical too, which seems wrong
13417 * while (isBLANK(*RExC_parse)) {
13420 if (endbrace == RExC_parse) {
13421 RExC_parse++; /* After the '}' */
13422 vFAIL2("Empty \\%c{}", name);
13424 length = endbrace - RExC_parse;
13425 /*while (isBLANK(*(RExC_parse + length - 1))) {
13428 switch (*RExC_parse) {
13431 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13433 goto bad_bound_type;
13438 if (length != 2 || *(RExC_parse + 1) != 'b') {
13439 goto bad_bound_type;
13444 if (length != 2 || *(RExC_parse + 1) != 'b') {
13445 goto bad_bound_type;
13450 if (length != 2 || *(RExC_parse + 1) != 'b') {
13451 goto bad_bound_type;
13457 RExC_parse = endbrace;
13459 "'%" UTF8f "' is an unknown bound type",
13460 UTF8fARG(UTF, length, endbrace - length));
13461 NOT_REACHED; /*NOTREACHED*/
13463 RExC_parse = endbrace;
13464 REQUIRE_UNI_RULES(flagp, 0);
13469 else if (op >= BOUNDA) { /* /aa is same as /a */
13473 /* Don't have to worry about UTF-8, in this message because
13474 * to get here the contents of the \b must be ASCII */
13475 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13476 "Using /u for '%.*s' instead of /%s",
13478 endbrace - length + 1,
13479 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13480 ? ASCII_RESTRICT_PAT_MODS
13481 : ASCII_MORE_RESTRICT_PAT_MODS);
13486 RExC_seen_d_op = TRUE;
13488 else if (op == BOUNDL) {
13489 RExC_contains_locale = 1;
13493 op += NBOUND - BOUND;
13496 ret = reg_node(pRExC_state, op);
13497 FLAGS(REGNODE_p(ret)) = flags;
13501 goto finish_meta_pat;
13505 ret = reg_node(pRExC_state, LNBREAK);
13506 *flagp |= HASWIDTH|SIMPLE;
13507 goto finish_meta_pat;
13521 /* These all have the same meaning inside [brackets], and it knows
13522 * how to do the best optimizations for them. So, pretend we found
13523 * these within brackets, and let it do the work */
13526 ret = regclass(pRExC_state, flagp, depth+1,
13527 TRUE, /* means just parse this element */
13528 FALSE, /* don't allow multi-char folds */
13529 FALSE, /* don't silence non-portable warnings. It
13530 would be a bug if these returned
13532 (bool) RExC_strict,
13533 TRUE, /* Allow an optimized regnode result */
13535 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13536 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13537 * multi-char folds are allowed. */
13539 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13542 RExC_parse--; /* regclass() leaves this one too far ahead */
13545 /* The escapes above that don't take a parameter can't be
13546 * followed by a '{'. But 'pX', 'p{foo}' and
13547 * correspondingly 'P' can be */
13548 if ( RExC_parse - parse_start == 1
13549 && UCHARAT(RExC_parse + 1) == '{'
13550 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13553 vFAIL("Unescaped left brace in regex is illegal here");
13555 Set_Node_Offset(REGNODE_p(ret), parse_start);
13556 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13557 nextchar(pRExC_state);
13560 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13561 * \N{...} evaluates to a sequence of more than one code points).
13562 * The function call below returns a regnode, which is our result.
13563 * The parameters cause it to fail if the \N{} evaluates to a
13564 * single code point; we handle those like any other literal. The
13565 * reason that the multicharacter case is handled here and not as
13566 * part of the EXACtish code is because of quantifiers. In
13567 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13568 * this way makes that Just Happen. dmq.
13569 * join_exact() will join this up with adjacent EXACTish nodes
13570 * later on, if appropriate. */
13572 if (grok_bslash_N(pRExC_state,
13573 &ret, /* Want a regnode returned */
13574 NULL, /* Fail if evaluates to a single code
13576 NULL, /* Don't need a count of how many code
13585 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13587 /* Here, evaluates to a single code point. Go get that */
13588 RExC_parse = parse_start;
13591 case 'k': /* Handle \k<NAME> and \k'NAME' */
13595 if ( RExC_parse >= RExC_end - 1
13596 || (( ch = RExC_parse[1]) != '<'
13601 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13602 vFAIL2("Sequence %.2s... not terminated", parse_start);
13605 ret = handle_named_backref(pRExC_state,
13617 case '1': case '2': case '3': case '4':
13618 case '5': case '6': case '7': case '8': case '9':
13623 if (*RExC_parse == 'g') {
13627 if (*RExC_parse == '{') {
13631 if (*RExC_parse == '-') {
13635 if (hasbrace && !isDIGIT(*RExC_parse)) {
13636 if (isrel) RExC_parse--;
13638 goto parse_named_seq;
13641 if (RExC_parse >= RExC_end) {
13642 goto unterminated_g;
13644 num = S_backref_value(RExC_parse, RExC_end);
13646 vFAIL("Reference to invalid group 0");
13647 else if (num == I32_MAX) {
13648 if (isDIGIT(*RExC_parse))
13649 vFAIL("Reference to nonexistent group");
13652 vFAIL("Unterminated \\g... pattern");
13656 num = RExC_npar - num;
13658 vFAIL("Reference to nonexistent or unclosed group");
13662 num = S_backref_value(RExC_parse, RExC_end);
13663 /* bare \NNN might be backref or octal - if it is larger
13664 * than or equal RExC_npar then it is assumed to be an
13665 * octal escape. Note RExC_npar is +1 from the actual
13666 * number of parens. */
13667 /* Note we do NOT check if num == I32_MAX here, as that is
13668 * handled by the RExC_npar check */
13671 /* any numeric escape < 10 is always a backref */
13673 /* any numeric escape < RExC_npar is a backref */
13674 && num >= RExC_npar
13675 /* cannot be an octal escape if it starts with 8 */
13676 && *RExC_parse != '8'
13677 /* cannot be an octal escape if it starts with 9 */
13678 && *RExC_parse != '9'
13680 /* Probably not meant to be a backref, instead likely
13681 * to be an octal character escape, e.g. \35 or \777.
13682 * The above logic should make it obvious why using
13683 * octal escapes in patterns is problematic. - Yves */
13684 RExC_parse = parse_start;
13689 /* At this point RExC_parse points at a numeric escape like
13690 * \12 or \88 or something similar, which we should NOT treat
13691 * as an octal escape. It may or may not be a valid backref
13692 * escape. For instance \88888888 is unlikely to be a valid
13694 while (isDIGIT(*RExC_parse))
13697 if (*RExC_parse != '}')
13698 vFAIL("Unterminated \\g{...} pattern");
13701 if (num >= (I32)RExC_npar) {
13703 /* It might be a forward reference; we can't fail until we
13704 * know, by completing the parse to get all the groups, and
13705 * then reparsing */
13706 if (ALL_PARENS_COUNTED) {
13707 if (num >= RExC_total_parens) {
13708 vFAIL("Reference to nonexistent group");
13712 REQUIRE_PARENS_PASS;
13716 ret = reganode(pRExC_state,
13719 : (ASCII_FOLD_RESTRICTED)
13721 : (AT_LEAST_UNI_SEMANTICS)
13727 if (OP(REGNODE_p(ret)) == REFF) {
13728 RExC_seen_d_op = TRUE;
13730 *flagp |= HASWIDTH;
13732 /* override incorrect value set in reganode MJD */
13733 Set_Node_Offset(REGNODE_p(ret), parse_start);
13734 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
13735 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13736 FALSE /* Don't force to /x */ );
13740 if (RExC_parse >= RExC_end)
13741 FAIL("Trailing \\");
13744 /* Do not generate "unrecognized" warnings here, we fall
13745 back into the quick-grab loop below */
13746 RExC_parse = parse_start;
13748 } /* end of switch on a \foo sequence */
13753 /* '#' comments should have been spaced over before this function was
13755 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13757 if (RExC_flags & RXf_PMf_EXTENDED) {
13758 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13759 if (RExC_parse < RExC_end)
13769 /* Here, we have determined that the next thing is probably a
13770 * literal character. RExC_parse points to the first byte of its
13771 * definition. (It still may be an escape sequence that evaluates
13772 * to a single character) */
13779 U32 max_string_len = 255;
13781 /* We may have to reparse the node, artificially stopping filling
13782 * it early, based on info gleaned in the first parse. This
13783 * variable gives where we stop. Make it above the normal stopping
13784 * place first time through; otherwise it would stop too early */
13785 U32 upper_fill = max_string_len + 1;
13787 /* We start out as an EXACT node, even if under /i, until we find a
13788 * character which is in a fold. The algorithm now segregates into
13789 * separate nodes, characters that fold from those that don't under
13790 * /i. (This hopefully will create nodes that are fixed strings
13791 * even under /i, giving the optimizer something to grab on to.)
13792 * So, if a node has something in it and the next character is in
13793 * the opposite category, that node is closed up, and the function
13794 * returns. Then regatom is called again, and a new node is
13795 * created for the new category. */
13796 U8 node_type = EXACT;
13798 /* Assume the node will be fully used; the excess is given back at
13799 * the end. Under /i, leave enough extra room so that we won't
13800 * overflow the buffer when we fold a character which would end up
13801 * overflowing the node. We can't make any other length
13802 * assumptions, as a byte input sequence could shrink down. */
13803 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
13807 ? UTF8_MAXBYTES_CASE
13808 /* Max non-UTF-8 expansion is 2 */ : 2)));
13810 bool next_is_quantifier;
13811 char * oldp = NULL;
13812 char * old_oldp = NULL;
13814 /* We can convert EXACTF nodes to EXACTFU if they contain only
13815 * characters that match identically regardless of the target
13816 * string's UTF8ness. The reason to do this is that EXACTF is not
13817 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
13820 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13821 * contain only above-Latin1 characters (hence must be in UTF8),
13822 * which don't participate in folds with Latin1-range characters,
13823 * as the latter's folds aren't known until runtime. */
13824 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13826 /* Single-character EXACTish nodes are almost always SIMPLE. This
13827 * allows us to override this as encountered */
13828 U8 maybe_SIMPLE = SIMPLE;
13830 /* Does this node contain something that can't match unless the
13831 * target string is (also) in UTF-8 */
13832 bool requires_utf8_target = FALSE;
13834 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
13835 bool has_ss = FALSE;
13837 /* So is the MICRO SIGN */
13838 bool has_micro_sign = FALSE;
13840 /* Set when we fill up the current node and there is still more
13841 * text to process */
13844 /* Allocate an EXACT node. The node_type may change below to
13845 * another EXACTish node, but since the size of the node doesn't
13846 * change, it works */
13847 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
13849 FILL_NODE(ret, node_type);
13852 s = STRING(REGNODE_p(ret));
13863 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13864 maybe_SIMPLE = SIMPLE;
13865 requires_utf8_target = FALSE;
13867 has_micro_sign = FALSE;
13871 /* This breaks under rare circumstances. If folding, we do not
13872 * want to split a node at a character that is a non-final in a
13873 * multi-char fold, as an input string could just happen to want to
13874 * match across the node boundary. The code at the end of the loop
13875 * looks for this, and backs off until it finds not such a
13876 * character, but it is possible (though extremely, extremely
13877 * unlikely) for all characters in the node to be non-final fold
13878 * ones, in which case we just leave the node fully filled, and
13879 * hope that it doesn't match the string in just the wrong place */
13881 assert( ! UTF /* Is at the beginning of a character */
13882 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13883 || UTF8_IS_START(UCHARAT(RExC_parse)));
13885 overflowed = FALSE;
13887 /* Here, we have a literal character. Find the maximal string of
13888 * them in the input that we can fit into a single EXACTish node.
13889 * We quit at the first non-literal or when the node gets full, or
13890 * under /i the categorization of folding/non-folding character
13892 while (p < RExC_end && len < upper_fill) {
13894 /* In most cases each iteration adds one byte to the output.
13895 * The exceptions override this */
13896 Size_t added_len = 1;
13901 /* White space has already been ignored */
13902 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13903 || ! is_PATWS_safe((p), RExC_end, UTF));
13915 /* Literal Escapes Switch
13917 This switch is meant to handle escape sequences that
13918 resolve to a literal character.
13920 Every escape sequence that represents something
13921 else, like an assertion or a char class, is handled
13922 in the switch marked 'Special Escapes' above in this
13923 routine, but also has an entry here as anything that
13924 isn't explicitly mentioned here will be treated as
13925 an unescaped equivalent literal.
13928 switch ((U8)*++p) {
13930 /* These are all the special escapes. */
13931 case 'A': /* Start assertion */
13932 case 'b': case 'B': /* Word-boundary assertion*/
13933 case 'C': /* Single char !DANGEROUS! */
13934 case 'd': case 'D': /* digit class */
13935 case 'g': case 'G': /* generic-backref, pos assertion */
13936 case 'h': case 'H': /* HORIZWS */
13937 case 'k': case 'K': /* named backref, keep marker */
13938 case 'p': case 'P': /* Unicode property */
13939 case 'R': /* LNBREAK */
13940 case 's': case 'S': /* space class */
13941 case 'v': case 'V': /* VERTWS */
13942 case 'w': case 'W': /* word class */
13943 case 'X': /* eXtended Unicode "combining
13944 character sequence" */
13945 case 'z': case 'Z': /* End of line/string assertion */
13949 /* Anything after here is an escape that resolves to a
13950 literal. (Except digits, which may or may not)
13956 case 'N': /* Handle a single-code point named character. */
13957 RExC_parse = p + 1;
13958 if (! grok_bslash_N(pRExC_state,
13959 NULL, /* Fail if evaluates to
13960 anything other than a
13961 single code point */
13962 &ender, /* The returned single code
13964 NULL, /* Don't need a count of
13965 how many code points */
13970 if (*flagp & NEED_UTF8)
13971 FAIL("panic: grok_bslash_N set NEED_UTF8");
13972 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13974 /* Here, it wasn't a single code point. Go close
13975 * up this EXACTish node. The switch() prior to
13976 * this switch handles the other cases */
13977 RExC_parse = p = oldp;
13981 RExC_parse = parse_start;
13983 /* The \N{} means the pattern, if previously /d,
13984 * becomes /u. That means it can't be an EXACTF node,
13985 * but an EXACTFU */
13986 if (node_type == EXACTF) {
13987 node_type = EXACTFU;
13989 /* If the node already contains something that
13990 * differs between EXACTF and EXACTFU, reparse it
13992 if (! maybe_exactfu) {
14013 ender = ESC_NATIVE;
14023 const char* error_msg;
14025 bool valid = grok_bslash_o(&p,
14029 TO_OUTPUT_WARNINGS(p),
14030 (bool) RExC_strict,
14031 TRUE, /* Output warnings
14036 RExC_parse = p; /* going to die anyway; point
14037 to exact spot of failure */
14040 UPDATE_WARNINGS_LOC(p - 1);
14046 UV result = UV_MAX; /* initialize to erroneous
14048 const char* error_msg;
14050 bool valid = grok_bslash_x(&p,
14054 TO_OUTPUT_WARNINGS(p),
14055 (bool) RExC_strict,
14056 TRUE, /* Silence warnings
14061 RExC_parse = p; /* going to die anyway; point
14062 to exact spot of failure */
14065 UPDATE_WARNINGS_LOC(p - 1);
14069 if (ender < 0x100) {
14070 if (RExC_recode_x_to_native) {
14071 ender = LATIN1_TO_NATIVE(ender);
14079 ender = grok_bslash_c(*p, TO_OUTPUT_WARNINGS(p));
14080 UPDATE_WARNINGS_LOC(p);
14083 case '8': case '9': /* must be a backreference */
14085 /* we have an escape like \8 which cannot be an octal escape
14086 * so we exit the loop, and let the outer loop handle this
14087 * escape which may or may not be a legitimate backref. */
14089 case '1': case '2': case '3':case '4':
14090 case '5': case '6': case '7':
14091 /* When we parse backslash escapes there is ambiguity
14092 * between backreferences and octal escapes. Any escape
14093 * from \1 - \9 is a backreference, any multi-digit
14094 * escape which does not start with 0 and which when
14095 * evaluated as decimal could refer to an already
14096 * parsed capture buffer is a back reference. Anything
14099 * Note this implies that \118 could be interpreted as
14100 * 118 OR as "\11" . "8" depending on whether there
14101 * were 118 capture buffers defined already in the
14104 /* NOTE, RExC_npar is 1 more than the actual number of
14105 * parens we have seen so far, hence the "<" as opposed
14107 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14108 { /* Not to be treated as an octal constant, go
14116 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14118 ender = grok_oct(p, &numlen, &flags, NULL);
14120 if ( isDIGIT(*p) /* like \08, \178 */
14121 && ckWARN(WARN_REGEXP)
14124 reg_warn_non_literal_string(
14126 form_short_octal_warning(p, numlen));
14132 FAIL("Trailing \\");
14135 if (isALPHANUMERIC(*p)) {
14136 /* An alpha followed by '{' is going to fail next
14137 * iteration, so don't output this warning in that
14139 if (! isALPHA(*p) || *(p + 1) != '{') {
14140 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14141 " passed through", p);
14144 goto normal_default;
14145 } /* End of switch on '\' */
14148 /* Trying to gain new uses for '{' without breaking too
14149 * much existing code is hard. The solution currently
14151 * 1) If there is no ambiguity that a '{' should always
14152 * be taken literally, at the start of a construct, we
14154 * 2) If the literal '{' conflicts with our desired use
14155 * of it as a metacharacter, we die. The deprecation
14156 * cycles for this have come and gone.
14157 * 3) If there is ambiguity, we raise a simple warning.
14158 * This could happen, for example, if the user
14159 * intended it to introduce a quantifier, but slightly
14160 * misspelled the quantifier. Without this warning,
14161 * the quantifier would silently be taken as a literal
14162 * string of characters instead of a meta construct */
14163 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14165 || ( p > parse_start + 1
14166 && isALPHA_A(*(p - 1))
14167 && *(p - 2) == '\\')
14168 || new_regcurly(p, RExC_end))
14170 RExC_parse = p + 1;
14171 vFAIL("Unescaped left brace in regex is "
14174 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14175 " passed through");
14177 goto normal_default;
14180 if (p > RExC_parse && RExC_strict) {
14181 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14184 default: /* A literal character */
14186 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14188 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14189 &numlen, UTF8_ALLOW_DEFAULT);
14195 } /* End of switch on the literal */
14197 /* Here, have looked at the literal character, and <ender>
14198 * contains its ordinal; <p> points to the character after it.
14202 REQUIRE_UTF8(flagp);
14205 /* We need to check if the next non-ignored thing is a
14206 * quantifier. Move <p> to after anything that should be
14207 * ignored, which, as a side effect, positions <p> for the next
14208 * loop iteration */
14209 skip_to_be_ignored_text(pRExC_state, &p,
14210 FALSE /* Don't force to /x */ );
14212 /* If the next thing is a quantifier, it applies to this
14213 * character only, which means that this character has to be in
14214 * its own node and can't just be appended to the string in an
14215 * existing node, so if there are already other characters in
14216 * the node, close the node with just them, and set up to do
14217 * this character again next time through, when it will be the
14218 * only thing in its new node */
14220 next_is_quantifier = LIKELY(p < RExC_end)
14221 && UNLIKELY(ISMULT2(p));
14223 if (next_is_quantifier && LIKELY(len)) {
14228 /* Ready to add 'ender' to the node */
14230 if (! FOLD) { /* The simple case, just append the literal */
14233 /* Don't output if it would overflow */
14234 if (UNLIKELY(len > max_string_len - ((UTF)
14235 ? UVCHR_SKIP(ender)
14242 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14243 *(s++) = (char) ender;
14246 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14247 added_len = (char *) new_s - s;
14248 s = (char *) new_s;
14251 requires_utf8_target = TRUE;
14255 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14257 /* Here are folding under /l, and the code point is
14258 * problematic. If this is the first character in the
14259 * node, change the node type to folding. Otherwise, if
14260 * this is the first problematic character, close up the
14261 * existing node, so can start a new node with this one */
14263 node_type = EXACTFL;
14264 RExC_contains_locale = 1;
14266 else if (node_type == EXACT) {
14271 /* This problematic code point means we can't simplify
14273 maybe_exactfu = FALSE;
14275 /* Here, we are adding a problematic fold character.
14276 * "Problematic" in this context means that its fold isn't
14277 * known until runtime. (The non-problematic code points
14278 * are the above-Latin1 ones that fold to also all
14279 * above-Latin1. Their folds don't vary no matter what the
14280 * locale is.) But here we have characters whose fold
14281 * depends on the locale. We just add in the unfolded
14282 * character, and wait until runtime to fold it */
14283 goto not_fold_common;
14285 else /* regular fold; see if actually is in a fold */
14286 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14288 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14290 /* Here, folding, but the character isn't in a fold.
14292 * Start a new node if previous characters in the node were
14294 if (len && node_type != EXACT) {
14299 /* Here, continuing a node with non-folded characters. Add
14301 goto not_fold_common;
14303 else { /* Here, does participate in some fold */
14305 /* If this is the first character in the node, change its
14306 * type to folding. Otherwise, if this is the first
14307 * folding character in the node, close up the existing
14308 * node, so can start a new node with this one. */
14310 node_type = compute_EXACTish(pRExC_state);
14312 else if (node_type == EXACT) {
14317 if (UTF) { /* Alway use the folded value for UTF-8
14319 if (UVCHR_IS_INVARIANT(ender)) {
14320 if (UNLIKELY(len + 1 > max_string_len)) {
14325 *(s)++ = (U8) toFOLD(ender);
14328 UV folded = _to_uni_fold_flags(
14330 (U8 *) s, /* We have allocated extra space
14331 in 's' so can't run off the
14334 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14335 ? FOLD_FLAGS_NOMIX_ASCII
14337 if (UNLIKELY(len + added_len > max_string_len)) {
14345 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14347 /* U+B5 folds to the MU, so its possible for a
14348 * non-UTF-8 target to match it */
14349 requires_utf8_target = TRUE;
14353 else { /* Here is non-UTF8. */
14355 /* The fold will be one or (rarely) two characters.
14356 * Check that there's room for at least a single one
14357 * before setting any flags, etc. Because otherwise an
14358 * overflowing character could cause a flag to be set
14359 * even though it doesn't end up in this node. (For
14360 * the two character fold, we check again, before
14361 * setting any flags) */
14362 if (UNLIKELY(len + 1 > max_string_len)) {
14367 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14368 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14369 || UNICODE_DOT_DOT_VERSION > 0)
14371 /* On non-ancient Unicodes, check for the only possible
14372 * multi-char fold */
14373 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14375 /* This potential multi-char fold means the node
14376 * can't be simple (because it could match more
14377 * than a single char). And in some cases it will
14378 * match 'ss', so set that flag */
14382 /* It can't change to be an EXACTFU (unless already
14383 * is one). We fold it iff under /u rules. */
14384 if (node_type != EXACTFU) {
14385 maybe_exactfu = FALSE;
14388 if (UNLIKELY(len + 2 > max_string_len)) {
14397 goto done_with_this_char;
14400 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14402 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14404 /* Also, the sequence 'ss' is special when not
14405 * under /u. If the target string is UTF-8, it
14406 * should match SHARP S; otherwise it won't. So,
14407 * here we have to exclude the possibility of this
14408 * node moving to /u.*/
14410 maybe_exactfu = FALSE;
14413 /* Here, the fold will be a single character */
14415 if (UNLIKELY(ender == MICRO_SIGN)) {
14416 has_micro_sign = TRUE;
14418 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14420 /* If the character's fold differs between /d and
14421 * /u, this can't change to be an EXACTFU node */
14422 maybe_exactfu = FALSE;
14425 *(s++) = (DEPENDS_SEMANTICS)
14426 ? (char) toFOLD(ender)
14428 /* Under /u, the fold of any character in
14429 * the 0-255 range happens to be its
14430 * lowercase equivalent, except for LATIN
14431 * SMALL LETTER SHARP S, which was handled
14432 * above, and the MICRO SIGN, whose fold
14433 * requires UTF-8 to represent. */
14434 : (char) toLOWER_L1(ender);
14436 } /* End of adding current character to the node */
14438 done_with_this_char:
14442 if (next_is_quantifier) {
14444 /* Here, the next input is a quantifier, and to get here,
14445 * the current character is the only one in the node. */
14449 } /* End of loop through literal characters */
14451 /* Here we have either exhausted the input or run out of room in
14452 * the node. If the former, we are done. (If we encountered a
14453 * character that can't be in the node, transfer is made directly
14454 * to <loopdone>, and so we wouldn't have fallen off the end of the
14456 if (LIKELY(! overflowed)) {
14460 /* Here we have run out of room. We can grow plain EXACT and
14461 * LEXACT nodes. If the pattern is gigantic enough, though,
14462 * eventually we'll have to artificially chunk the pattern into
14463 * multiple nodes. */
14464 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14465 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14466 Size_t overhead_expansion = 0;
14468 Size_t max_nodes_for_string;
14472 /* Here we couldn't fit the final character in the current
14473 * node, so it will have to be reparsed, no matter what else we
14477 /* If would have overflowed a regular EXACT node, switch
14478 * instead to an LEXACT. The code below is structured so that
14479 * the actual growing code is common to changing from an EXACT
14480 * or just increasing the LEXACT size. This means that we have
14481 * to save the string in the EXACT case before growing, and
14482 * then copy it afterwards to its new location */
14483 if (node_type == EXACT) {
14484 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14485 RExC_emit += overhead_expansion;
14486 Copy(s0, temp, len, char);
14489 /* Ready to grow. If it was a plain EXACT, the string was
14490 * saved, and the first few bytes of it overwritten by adding
14491 * an argument field. We assume, as we do elsewhere in this
14492 * file, that one byte of remaining input will translate into
14493 * one byte of output, and if that's too small, we grow again,
14494 * if too large the excess memory is freed at the end */
14496 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14497 achievable = MIN(max_nodes_for_string,
14498 current_string_nodes + STR_SZ(RExC_end - p));
14499 delta = achievable - current_string_nodes;
14501 /* If there is just no more room, go finish up this chunk of
14507 change_engine_size(pRExC_state, delta + overhead_expansion);
14508 current_string_nodes += delta;
14510 = sizeof(struct regnode) * current_string_nodes;
14511 upper_fill = max_string_len + 1;
14513 /* If the length was small, we know this was originally an
14514 * EXACT node now converted to LEXACT, and the string has to be
14515 * restored. Otherwise the string was untouched. 260 is just
14516 * a number safely above 255 so don't have to worry about
14517 * getting it precise */
14519 node_type = LEXACT;
14520 FILL_NODE(ret, node_type);
14521 s0 = STRING(REGNODE_p(ret));
14522 Copy(temp, s0, len, char);
14526 goto continue_parse;
14528 else if (! LOC) { /* XXX shouldn't /l assume could be a UTF-8
14529 locale, and prepare for that? */
14531 /* Here is /i. Running out of room creates a problem if we are
14532 * folding, and the split happens in the middle of a
14533 * multi-character fold, as a match that should have occurred,
14534 * won't, due to the way nodes are matched, and our artificial
14535 * boundary. So back off until we aren't splitting such a
14536 * fold. If there is no such place to back off to, we end up
14537 * taking the entire node as-is. This can happen if the node
14538 * consists entirely of 'f' or entirely of 's' characters (or
14539 * things that fold to them) as 'ff' and 'ss' are
14540 * multi-character folds.
14543 * old_oldp points to the beginning in the input of the
14544 * penultimate character in the node.
14545 * oldp points to the beginning in the input of the
14546 * final character in the node.
14547 * p points to the beginning in the input of the
14548 * next character in the input, the one that won't
14551 * We aren't in the middle of a multi-char fold unless the
14552 * final character in the node can appear in a non-final
14553 * position in such a fold. Very few characters actually
14554 * participate in multi-character folds, and fewer still can be
14555 * in the non-final position. But it's complicated to know
14556 * here if that final character is folded or not, so skip this
14559 /* Make sure enough space for final char of node,
14560 * first char of following node, and the fold of the
14561 * following char (so we don't have to worry about
14562 * that fold running off the end */
14563 U8 foldbuf[UTF8_MAXBYTES_CASE * 5 + 1];
14566 char * const sav_oldp = oldp;
14570 /* The Unicode standard says that multi character folds consist
14571 * of either two or three characters. So we create a buffer
14572 * containing a window of three. The first is the final
14573 * character in the node (folded), and then the two that begin
14574 * the following node. But if the first character of the
14575 * following node can't be in a non-final fold position, there
14576 * is no need to look at its successor character. The macros
14577 * used below to check for multi character folds require folded
14578 * inputs, so we have to fold these. (The fold of p was likely
14579 * calculated in the loop above, but it hasn't beeen saved, and
14580 * khw thinks it would be too entangled to change to do so) */
14582 if (UTF || LIKELY(UCHARAT(p) != MICRO_SIGN)) {
14583 folded = _to_uni_fold_flags(ender,
14589 foldbuf[0] = folded = MICRO_SIGN;
14593 /* Here, foldbuf contains the fold of the first character in
14594 * the next node. We may also need the next one (if there is
14595 * one) to get our third, but if the first character folded to
14596 * more than one, those extra one(s) will serve as the third.
14597 * Also, we don't need a third unless the previous one can
14598 * appear in a non-final position in a fold */
14599 if ( ((RExC_end - p) > ((UTF) ? UVCHR_SKIP(ender) : 1))
14600 && (fold_len == 1 || ( UTF
14601 && UVCHR_SKIP(folded) == fold_len))
14602 && UNLIKELY(_invlist_contains_cp(PL_NonFinalFold, folded)))
14605 STRLEN next_fold_len;
14607 toFOLD_utf8_safe((U8*) p + UTF8SKIP(p),
14608 (U8*) RExC_end, foldbuf + fold_len,
14610 fold_len += next_fold_len;
14613 if (UNLIKELY(p[1] == LATIN_SMALL_LETTER_SHARP_S)) {
14614 foldbuf[fold_len] = 's';
14617 foldbuf[fold_len] = toLOWER_L1(p[1]);
14623 /* Here foldbuf contains the the fold of p, and if appropriate
14624 * that of the character following p in the input. */
14626 /* Search backwards until find a place that doesn't split a
14627 * multi-char fold */
14630 char s_fold_buf[UTF8_MAXBYTES_CASE];
14631 char * s_fold = s_fold_buf;
14635 /* There's no safe place in the node to split. Quit so
14636 * will take the whole node */
14641 /* Backup 1 character. The first time through this moves s
14642 * to point to the final character in the node */
14644 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s0);
14650 /* 's' may or may not be folded; so make sure it is, and
14651 * use just the final character in its fold (should there
14652 * be more than one */
14654 toFOLD_utf8_safe((U8*) s,
14655 (U8*) s + UTF8SKIP(s),
14656 (U8 *) s_fold_buf, &s_len);
14657 while (s_fold + UTF8SKIP(s_fold) < s_fold_buf + s_len)
14659 s_fold += UTF8SKIP(s_fold);
14661 s_len = UTF8SKIP(s_fold);
14664 if (UNLIKELY(UCHARAT(s) == LATIN_SMALL_LETTER_SHARP_S))
14666 s_fold_buf[0] = 's';
14668 else { /* This works for all other non-UTF-8 folds
14670 s_fold_buf[0] = toLOWER_L1(UCHARAT(s));
14675 /* Unshift this character to the beginning of the buffer,
14676 * No longer needed trailing characters are overwritten.
14678 Move(foldbuf, foldbuf + s_len, sizeof(foldbuf) - s_len, U8);
14679 Copy(s_fold, foldbuf, s_len, U8);
14681 /* If this isn't a multi-character fold, we have found a
14682 * splittable place. If this is the final character in the
14683 * node, that means the node is valid as-is, and can quit.
14684 * Otherwise, we note how much we can fill the node before
14685 * coming to a non-splittable position, and go parse it
14686 * again, stopping there. This is done because we know
14687 * where in the output to stop, but we don't have a map to
14688 * where that is in the input. One could be created, but
14689 * it seems like overkill for such a rare event as we are
14690 * dealing with here */
14692 if (! is_MULTI_CHAR_FOLD_utf8_safe(foldbuf,
14693 foldbuf + UTF8_MAXBYTES_CASE))
14695 upper_fill = s + UTF8SKIP(s) - s0;
14696 if (LIKELY(oldp)) {
14702 else if (! is_MULTI_CHAR_FOLD_latin1_safe(foldbuf,
14703 foldbuf + UTF8_MAXBYTES_CASE))
14705 upper_fill = s + 1 - s0;
14706 if (LIKELY(oldp)) {
14715 } /* End of loop backing up through the node */
14716 /* Here the node consists entirely of non-final multi-char
14717 * folds. (Likely it is all 'f's or all 's's.) There's no
14718 * decent place to split it, so give up and just take the
14721 } /* End of verifying node ends with an appropriate char */
14725 loopdone: /* Jumped to when encounters something that shouldn't be
14728 /* Free up any over-allocated space; cast is to silence bogus
14729 * warning in MS VC */
14730 change_engine_size(pRExC_state,
14731 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
14733 /* I (khw) don't know if you can get here with zero length, but the
14734 * old code handled this situation by creating a zero-length EXACT
14735 * node. Might as well be NOTHING instead */
14737 OP(REGNODE_p(ret)) = NOTHING;
14741 /* If the node type is EXACT here, check to see if it
14742 * should be EXACTL, or EXACT_REQ8. */
14743 if (node_type == EXACT) {
14745 node_type = EXACTL;
14747 else if (requires_utf8_target) {
14748 node_type = EXACT_REQ8;
14751 else if (node_type == LEXACT) {
14752 if (requires_utf8_target) {
14753 node_type = LEXACT_REQ8;
14757 if ( UNLIKELY(has_micro_sign || has_ss)
14758 && (node_type == EXACTFU || ( node_type == EXACTF
14759 && maybe_exactfu)))
14760 { /* These two conditions are problematic in non-UTF-8
14763 node_type = EXACTFUP;
14765 else if (node_type == EXACTFL) {
14767 /* 'maybe_exactfu' is deliberately set above to
14768 * indicate this node type, where all code points in it
14770 if (maybe_exactfu) {
14771 node_type = EXACTFLU8;
14774 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
14776 /* A character that folds to more than one will
14777 * match multiple characters, so can't be SIMPLE.
14778 * We don't have to worry about this with EXACTFLU8
14779 * nodes just above, as they have already been
14780 * folded (since the fold doesn't vary at run
14781 * time). Here, if the final character in the node
14782 * folds to multiple, it can't be simple. (This
14783 * only has an effect if the node has only a single
14784 * character, hence the final one, as elsewhere we
14785 * turn off simple for nodes whose length > 1 */
14789 else if (node_type == EXACTF) { /* Means is /di */
14791 /* This intermediate variable is needed solely because
14792 * the asserts in the macro where used exceed Win32's
14793 * literal string capacity */
14794 char first_char = * STRING(REGNODE_p(ret));
14796 /* If 'maybe_exactfu' is clear, then we need to stay
14797 * /di. If it is set, it means there are no code
14798 * points that match differently depending on UTF8ness
14799 * of the target string, so it can become an EXACTFU
14801 if (! maybe_exactfu) {
14802 RExC_seen_d_op = TRUE;
14804 else if ( isALPHA_FOLD_EQ(first_char, 's')
14805 || isALPHA_FOLD_EQ(ender, 's'))
14807 /* But, if the node begins or ends in an 's' we
14808 * have to defer changing it into an EXACTFU, as
14809 * the node could later get joined with another one
14810 * that ends or begins with 's' creating an 'ss'
14811 * sequence which would then wrongly match the
14812 * sharp s without the target being UTF-8. We
14813 * create a special node that we resolve later when
14814 * we join nodes together */
14816 node_type = EXACTFU_S_EDGE;
14819 node_type = EXACTFU;
14823 if (requires_utf8_target && node_type == EXACTFU) {
14824 node_type = EXACTFU_REQ8;
14828 OP(REGNODE_p(ret)) = node_type;
14829 setSTR_LEN(REGNODE_p(ret), len);
14830 RExC_emit += STR_SZ(len);
14832 /* If the node isn't a single character, it can't be SIMPLE */
14833 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
14837 *flagp |= HASWIDTH | maybe_SIMPLE;
14840 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
14844 /* len is STRLEN which is unsigned, need to copy to signed */
14847 vFAIL("Internal disaster");
14850 } /* End of label 'defchar:' */
14852 } /* End of giant switch on input character */
14854 /* Position parse to next real character */
14855 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14856 FALSE /* Don't force to /x */ );
14857 if ( *RExC_parse == '{'
14858 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
14860 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
14862 vFAIL("Unescaped left brace in regex is illegal here");
14864 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
14865 " passed through");
14873 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
14875 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
14876 * sets up the bitmap and any flags, removing those code points from the
14877 * inversion list, setting it to NULL should it become completely empty */
14881 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
14882 assert(PL_regkind[OP(node)] == ANYOF);
14884 /* There is no bitmap for this node type */
14885 if (inRANGE(OP(node), ANYOFH, ANYOFHr)) {
14889 ANYOF_BITMAP_ZERO(node);
14890 if (*invlist_ptr) {
14892 /* This gets set if we actually need to modify things */
14893 bool change_invlist = FALSE;
14897 /* Start looking through *invlist_ptr */
14898 invlist_iterinit(*invlist_ptr);
14899 while (invlist_iternext(*invlist_ptr, &start, &end)) {
14903 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
14904 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
14907 /* Quit if are above what we should change */
14908 if (start >= NUM_ANYOF_CODE_POINTS) {
14912 change_invlist = TRUE;
14914 /* Set all the bits in the range, up to the max that we are doing */
14915 high = (end < NUM_ANYOF_CODE_POINTS - 1)
14917 : NUM_ANYOF_CODE_POINTS - 1;
14918 for (i = start; i <= (int) high; i++) {
14919 if (! ANYOF_BITMAP_TEST(node, i)) {
14920 ANYOF_BITMAP_SET(node, i);
14924 invlist_iterfinish(*invlist_ptr);
14926 /* Done with loop; remove any code points that are in the bitmap from
14927 * *invlist_ptr; similarly for code points above the bitmap if we have
14928 * a flag to match all of them anyways */
14929 if (change_invlist) {
14930 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
14932 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14933 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14936 /* If have completely emptied it, remove it completely */
14937 if (_invlist_len(*invlist_ptr) == 0) {
14938 SvREFCNT_dec_NN(*invlist_ptr);
14939 *invlist_ptr = NULL;
14944 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14945 Character classes ([:foo:]) can also be negated ([:^foo:]).
14946 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14947 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14948 but trigger failures because they are currently unimplemented. */
14950 #define POSIXCC_DONE(c) ((c) == ':')
14951 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14952 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14953 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14955 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14956 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14957 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14959 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14961 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14963 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14964 if (posix_warnings) { \
14965 if (! RExC_warn_text ) RExC_warn_text = \
14966 (AV *) sv_2mortal((SV *) newAV()); \
14967 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14971 REPORT_LOCATION_ARGS(p))); \
14974 #define CLEAR_POSIX_WARNINGS() \
14976 if (posix_warnings && RExC_warn_text) \
14977 av_clear(RExC_warn_text); \
14980 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14982 CLEAR_POSIX_WARNINGS(); \
14987 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14989 const char * const s, /* Where the putative posix class begins.
14990 Normally, this is one past the '['. This
14991 parameter exists so it can be somewhere
14992 besides RExC_parse. */
14993 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14995 AV ** posix_warnings, /* Where to place any generated warnings, or
14997 const bool check_only /* Don't die if error */
15000 /* This parses what the caller thinks may be one of the three POSIX
15002 * 1) a character class, like [:blank:]
15003 * 2) a collating symbol, like [. .]
15004 * 3) an equivalence class, like [= =]
15005 * In the latter two cases, it croaks if it finds a syntactically legal
15006 * one, as these are not handled by Perl.
15008 * The main purpose is to look for a POSIX character class. It returns:
15009 * a) the class number
15010 * if it is a completely syntactically and semantically legal class.
15011 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15012 * closing ']' of the class
15013 * b) OOB_NAMEDCLASS
15014 * if it appears that one of the three POSIX constructs was meant, but
15015 * its specification was somehow defective. 'updated_parse_ptr', if
15016 * not NULL, is set to point to the character just after the end
15017 * character of the class. See below for handling of warnings.
15018 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15019 * if it doesn't appear that a POSIX construct was intended.
15020 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15023 * In b) there may be errors or warnings generated. If 'check_only' is
15024 * TRUE, then any errors are discarded. Warnings are returned to the
15025 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15026 * instead it is NULL, warnings are suppressed.
15028 * The reason for this function, and its complexity is that a bracketed
15029 * character class can contain just about anything. But it's easy to
15030 * mistype the very specific posix class syntax but yielding a valid
15031 * regular bracketed class, so it silently gets compiled into something
15032 * quite unintended.
15034 * The solution adopted here maintains backward compatibility except that
15035 * it adds a warning if it looks like a posix class was intended but
15036 * improperly specified. The warning is not raised unless what is input
15037 * very closely resembles one of the 14 legal posix classes. To do this,
15038 * it uses fuzzy parsing. It calculates how many single-character edits it
15039 * would take to transform what was input into a legal posix class. Only
15040 * if that number is quite small does it think that the intention was a
15041 * posix class. Obviously these are heuristics, and there will be cases
15042 * where it errs on one side or another, and they can be tweaked as
15043 * experience informs.
15045 * The syntax for a legal posix class is:
15047 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15049 * What this routine considers syntactically to be an intended posix class
15050 * is this (the comments indicate some restrictions that the pattern
15053 * qr/(?x: \[? # The left bracket, possibly
15055 * \h* # possibly followed by blanks
15056 * (?: \^ \h* )? # possibly a misplaced caret
15057 * [:;]? # The opening class character,
15058 * # possibly omitted. A typo
15059 * # semi-colon can also be used.
15061 * \^? # possibly a correctly placed
15062 * # caret, but not if there was also
15063 * # a misplaced one
15065 * .{3,15} # The class name. If there are
15066 * # deviations from the legal syntax,
15067 * # its edit distance must be close
15068 * # to a real class name in order
15069 * # for it to be considered to be
15070 * # an intended posix class.
15072 * [[:punct:]]? # The closing class character,
15073 * # possibly omitted. If not a colon
15074 * # nor semi colon, the class name
15075 * # must be even closer to a valid
15078 * \]? # The right bracket, possibly
15082 * In the above, \h must be ASCII-only.
15084 * These are heuristics, and can be tweaked as field experience dictates.
15085 * There will be cases when someone didn't intend to specify a posix class
15086 * that this warns as being so. The goal is to minimize these, while
15087 * maximizing the catching of things intended to be a posix class that
15088 * aren't parsed as such.
15092 const char * const e = RExC_end;
15093 unsigned complement = 0; /* If to complement the class */
15094 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15095 bool has_opening_bracket = FALSE;
15096 bool has_opening_colon = FALSE;
15097 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15099 const char * possible_end = NULL; /* used for a 2nd parse pass */
15100 const char* name_start; /* ptr to class name first char */
15102 /* If the number of single-character typos the input name is away from a
15103 * legal name is no more than this number, it is considered to have meant
15104 * the legal name */
15105 int max_distance = 2;
15107 /* to store the name. The size determines the maximum length before we
15108 * decide that no posix class was intended. Should be at least
15109 * sizeof("alphanumeric") */
15111 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15113 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15115 CLEAR_POSIX_WARNINGS();
15118 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15121 if (*(p - 1) != '[') {
15122 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15123 found_problem = TRUE;
15126 has_opening_bracket = TRUE;
15129 /* They could be confused and think you can put spaces between the
15132 found_problem = TRUE;
15136 } while (p < e && isBLANK(*p));
15138 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15141 /* For [. .] and [= =]. These are quite different internally from [: :],
15142 * so they are handled separately. */
15143 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15144 and 1 for at least one char in it
15147 const char open_char = *p;
15148 const char * temp_ptr = p + 1;
15150 /* These two constructs are not handled by perl, and if we find a
15151 * syntactically valid one, we croak. khw, who wrote this code, finds
15152 * this explanation of them very unclear:
15153 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15154 * And searching the rest of the internet wasn't very helpful either.
15155 * It looks like just about any byte can be in these constructs,
15156 * depending on the locale. But unless the pattern is being compiled
15157 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15158 * In that case, it looks like [= =] isn't allowed at all, and that
15159 * [. .] could be any single code point, but for longer strings the
15160 * constituent characters would have to be the ASCII alphabetics plus
15161 * the minus-hyphen. Any sensible locale definition would limit itself
15162 * to these. And any portable one definitely should. Trying to parse
15163 * the general case is a nightmare (see [perl #127604]). So, this code
15164 * looks only for interiors of these constructs that match:
15166 * Using \w relaxes the apparent rules a little, without adding much
15167 * danger of mistaking something else for one of these constructs.
15169 * [. .] in some implementations described on the internet is usable to
15170 * escape a character that otherwise is special in bracketed character
15171 * classes. For example [.].] means a literal right bracket instead of
15172 * the ending of the class
15174 * [= =] can legitimately contain a [. .] construct, but we don't
15175 * handle this case, as that [. .] construct will later get parsed
15176 * itself and croak then. And [= =] is checked for even when not under
15177 * /l, as Perl has long done so.
15179 * The code below relies on there being a trailing NUL, so it doesn't
15180 * have to keep checking if the parse ptr < e.
15182 if (temp_ptr[1] == open_char) {
15185 else while ( temp_ptr < e
15186 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15191 if (*temp_ptr == open_char) {
15193 if (*temp_ptr == ']') {
15195 if (! found_problem && ! check_only) {
15196 RExC_parse = (char *) temp_ptr;
15197 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15198 "extensions", open_char, open_char);
15201 /* Here, the syntax wasn't completely valid, or else the call
15202 * is to check-only */
15203 if (updated_parse_ptr) {
15204 *updated_parse_ptr = (char *) temp_ptr;
15207 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15211 /* If we find something that started out to look like one of these
15212 * constructs, but isn't, we continue below so that it can be checked
15213 * for being a class name with a typo of '.' or '=' instead of a colon.
15217 /* Here, we think there is a possibility that a [: :] class was meant, and
15218 * we have the first real character. It could be they think the '^' comes
15221 found_problem = TRUE;
15222 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15227 found_problem = TRUE;
15231 } while (p < e && isBLANK(*p));
15233 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15237 /* But the first character should be a colon, which they could have easily
15238 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15239 * distinguish from a colon, so treat that as a colon). */
15242 has_opening_colon = TRUE;
15244 else if (*p == ';') {
15245 found_problem = TRUE;
15247 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15248 has_opening_colon = TRUE;
15251 found_problem = TRUE;
15252 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15254 /* Consider an initial punctuation (not one of the recognized ones) to
15255 * be a left terminator */
15256 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15261 /* They may think that you can put spaces between the components */
15263 found_problem = TRUE;
15267 } while (p < e && isBLANK(*p));
15269 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15274 /* We consider something like [^:^alnum:]] to not have been intended to
15275 * be a posix class, but XXX maybe we should */
15277 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15284 /* Again, they may think that you can put spaces between the components */
15286 found_problem = TRUE;
15290 } while (p < e && isBLANK(*p));
15292 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15297 /* XXX This ']' may be a typo, and something else was meant. But
15298 * treating it as such creates enough complications, that that
15299 * possibility isn't currently considered here. So we assume that the
15300 * ']' is what is intended, and if we've already found an initial '[',
15301 * this leaves this construct looking like [:] or [:^], which almost
15302 * certainly weren't intended to be posix classes */
15303 if (has_opening_bracket) {
15304 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15307 /* But this function can be called when we parse the colon for
15308 * something like qr/[alpha:]]/, so we back up to look for the
15313 found_problem = TRUE;
15314 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15316 else if (*p != ':') {
15318 /* XXX We are currently very restrictive here, so this code doesn't
15319 * consider the possibility that, say, /[alpha.]]/ was intended to
15320 * be a posix class. */
15321 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15324 /* Here we have something like 'foo:]'. There was no initial colon,
15325 * and we back up over 'foo. XXX Unlike the going forward case, we
15326 * don't handle typos of non-word chars in the middle */
15327 has_opening_colon = FALSE;
15330 while (p > RExC_start && isWORDCHAR(*p)) {
15335 /* Here, we have positioned ourselves to where we think the first
15336 * character in the potential class is */
15339 /* Now the interior really starts. There are certain key characters that
15340 * can end the interior, or these could just be typos. To catch both
15341 * cases, we may have to do two passes. In the first pass, we keep on
15342 * going unless we come to a sequence that matches
15343 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15344 * This means it takes a sequence to end the pass, so two typos in a row if
15345 * that wasn't what was intended. If the class is perfectly formed, just
15346 * this one pass is needed. We also stop if there are too many characters
15347 * being accumulated, but this number is deliberately set higher than any
15348 * real class. It is set high enough so that someone who thinks that
15349 * 'alphanumeric' is a correct name would get warned that it wasn't.
15350 * While doing the pass, we keep track of where the key characters were in
15351 * it. If we don't find an end to the class, and one of the key characters
15352 * was found, we redo the pass, but stop when we get to that character.
15353 * Thus the key character was considered a typo in the first pass, but a
15354 * terminator in the second. If two key characters are found, we stop at
15355 * the second one in the first pass. Again this can miss two typos, but
15356 * catches a single one
15358 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15359 * point to the first key character. For the second pass, it starts as -1.
15365 bool has_blank = FALSE;
15366 bool has_upper = FALSE;
15367 bool has_terminating_colon = FALSE;
15368 bool has_terminating_bracket = FALSE;
15369 bool has_semi_colon = FALSE;
15370 unsigned int name_len = 0;
15371 int punct_count = 0;
15375 /* Squeeze out blanks when looking up the class name below */
15376 if (isBLANK(*p) ) {
15378 found_problem = TRUE;
15383 /* The name will end with a punctuation */
15385 const char * peek = p + 1;
15387 /* Treat any non-']' punctuation followed by a ']' (possibly
15388 * with intervening blanks) as trying to terminate the class.
15389 * ']]' is very likely to mean a class was intended (but
15390 * missing the colon), but the warning message that gets
15391 * generated shows the error position better if we exit the
15392 * loop at the bottom (eventually), so skip it here. */
15394 if (peek < e && isBLANK(*peek)) {
15396 found_problem = TRUE;
15399 } while (peek < e && isBLANK(*peek));
15402 if (peek < e && *peek == ']') {
15403 has_terminating_bracket = TRUE;
15405 has_terminating_colon = TRUE;
15407 else if (*p == ';') {
15408 has_semi_colon = TRUE;
15409 has_terminating_colon = TRUE;
15412 found_problem = TRUE;
15419 /* Here we have punctuation we thought didn't end the class.
15420 * Keep track of the position of the key characters that are
15421 * more likely to have been class-enders */
15422 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15424 /* Allow just one such possible class-ender not actually
15425 * ending the class. */
15426 if (possible_end) {
15432 /* If we have too many punctuation characters, no use in
15434 if (++punct_count > max_distance) {
15438 /* Treat the punctuation as a typo. */
15439 input_text[name_len++] = *p;
15442 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15443 input_text[name_len++] = toLOWER(*p);
15445 found_problem = TRUE;
15447 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15448 input_text[name_len++] = *p;
15452 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15456 /* The declaration of 'input_text' is how long we allow a potential
15457 * class name to be, before saying they didn't mean a class name at
15459 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15464 /* We get to here when the possible class name hasn't been properly
15465 * terminated before:
15466 * 1) we ran off the end of the pattern; or
15467 * 2) found two characters, each of which might have been intended to
15468 * be the name's terminator
15469 * 3) found so many punctuation characters in the purported name,
15470 * that the edit distance to a valid one is exceeded
15471 * 4) we decided it was more characters than anyone could have
15472 * intended to be one. */
15474 found_problem = TRUE;
15476 /* In the final two cases, we know that looking up what we've
15477 * accumulated won't lead to a match, even a fuzzy one. */
15478 if ( name_len >= C_ARRAY_LENGTH(input_text)
15479 || punct_count > max_distance)
15481 /* If there was an intermediate key character that could have been
15482 * an intended end, redo the parse, but stop there */
15483 if (possible_end && possible_end != (char *) -1) {
15484 possible_end = (char *) -1; /* Special signal value to say
15485 we've done a first pass */
15490 /* Otherwise, it can't have meant to have been a class */
15491 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15494 /* If we ran off the end, and the final character was a punctuation
15495 * one, back up one, to look at that final one just below. Later, we
15496 * will restore the parse pointer if appropriate */
15497 if (name_len && p == e && isPUNCT(*(p-1))) {
15502 if (p < e && isPUNCT(*p)) {
15504 has_terminating_bracket = TRUE;
15506 /* If this is a 2nd ']', and the first one is just below this
15507 * one, consider that to be the real terminator. This gives a
15508 * uniform and better positioning for the warning message */
15510 && possible_end != (char *) -1
15511 && *possible_end == ']'
15512 && name_len && input_text[name_len - 1] == ']')
15517 /* And this is actually equivalent to having done the 2nd
15518 * pass now, so set it to not try again */
15519 possible_end = (char *) -1;
15524 has_terminating_colon = TRUE;
15526 else if (*p == ';') {
15527 has_semi_colon = TRUE;
15528 has_terminating_colon = TRUE;
15536 /* Here, we have a class name to look up. We can short circuit the
15537 * stuff below for short names that can't possibly be meant to be a
15538 * class name. (We can do this on the first pass, as any second pass
15539 * will yield an even shorter name) */
15540 if (name_len < 3) {
15541 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15544 /* Find which class it is. Initially switch on the length of the name.
15546 switch (name_len) {
15548 if (memEQs(name_start, 4, "word")) {
15549 /* this is not POSIX, this is the Perl \w */
15550 class_number = ANYOF_WORDCHAR;
15554 /* Names all of length 5: alnum alpha ascii blank cntrl digit
15555 * graph lower print punct space upper
15556 * Offset 4 gives the best switch position. */
15557 switch (name_start[4]) {
15559 if (memBEGINs(name_start, 5, "alph")) /* alpha */
15560 class_number = ANYOF_ALPHA;
15563 if (memBEGINs(name_start, 5, "spac")) /* space */
15564 class_number = ANYOF_SPACE;
15567 if (memBEGINs(name_start, 5, "grap")) /* graph */
15568 class_number = ANYOF_GRAPH;
15571 if (memBEGINs(name_start, 5, "asci")) /* ascii */
15572 class_number = ANYOF_ASCII;
15575 if (memBEGINs(name_start, 5, "blan")) /* blank */
15576 class_number = ANYOF_BLANK;
15579 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
15580 class_number = ANYOF_CNTRL;
15583 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
15584 class_number = ANYOF_ALPHANUMERIC;
15587 if (memBEGINs(name_start, 5, "lowe")) /* lower */
15588 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
15589 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
15590 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
15593 if (memBEGINs(name_start, 5, "digi")) /* digit */
15594 class_number = ANYOF_DIGIT;
15595 else if (memBEGINs(name_start, 5, "prin")) /* print */
15596 class_number = ANYOF_PRINT;
15597 else if (memBEGINs(name_start, 5, "punc")) /* punct */
15598 class_number = ANYOF_PUNCT;
15603 if (memEQs(name_start, 6, "xdigit"))
15604 class_number = ANYOF_XDIGIT;
15608 /* If the name exactly matches a posix class name the class number will
15609 * here be set to it, and the input almost certainly was meant to be a
15610 * posix class, so we can skip further checking. If instead the syntax
15611 * is exactly correct, but the name isn't one of the legal ones, we
15612 * will return that as an error below. But if neither of these apply,
15613 * it could be that no posix class was intended at all, or that one
15614 * was, but there was a typo. We tease these apart by doing fuzzy
15615 * matching on the name */
15616 if (class_number == OOB_NAMEDCLASS && found_problem) {
15617 const UV posix_names[][6] = {
15618 { 'a', 'l', 'n', 'u', 'm' },
15619 { 'a', 'l', 'p', 'h', 'a' },
15620 { 'a', 's', 'c', 'i', 'i' },
15621 { 'b', 'l', 'a', 'n', 'k' },
15622 { 'c', 'n', 't', 'r', 'l' },
15623 { 'd', 'i', 'g', 'i', 't' },
15624 { 'g', 'r', 'a', 'p', 'h' },
15625 { 'l', 'o', 'w', 'e', 'r' },
15626 { 'p', 'r', 'i', 'n', 't' },
15627 { 'p', 'u', 'n', 'c', 't' },
15628 { 's', 'p', 'a', 'c', 'e' },
15629 { 'u', 'p', 'p', 'e', 'r' },
15630 { 'w', 'o', 'r', 'd' },
15631 { 'x', 'd', 'i', 'g', 'i', 't' }
15633 /* The names of the above all have added NULs to make them the same
15634 * size, so we need to also have the real lengths */
15635 const UV posix_name_lengths[] = {
15636 sizeof("alnum") - 1,
15637 sizeof("alpha") - 1,
15638 sizeof("ascii") - 1,
15639 sizeof("blank") - 1,
15640 sizeof("cntrl") - 1,
15641 sizeof("digit") - 1,
15642 sizeof("graph") - 1,
15643 sizeof("lower") - 1,
15644 sizeof("print") - 1,
15645 sizeof("punct") - 1,
15646 sizeof("space") - 1,
15647 sizeof("upper") - 1,
15648 sizeof("word") - 1,
15649 sizeof("xdigit")- 1
15652 int temp_max = max_distance; /* Use a temporary, so if we
15653 reparse, we haven't changed the
15656 /* Use a smaller max edit distance if we are missing one of the
15658 if ( has_opening_bracket + has_opening_colon < 2
15659 || has_terminating_bracket + has_terminating_colon < 2)
15664 /* See if the input name is close to a legal one */
15665 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
15667 /* Short circuit call if the lengths are too far apart to be
15669 if (abs( (int) (name_len - posix_name_lengths[i]))
15675 if (edit_distance(input_text,
15678 posix_name_lengths[i],
15682 { /* If it is close, it probably was intended to be a class */
15683 goto probably_meant_to_be;
15687 /* Here the input name is not close enough to a valid class name
15688 * for us to consider it to be intended to be a posix class. If
15689 * we haven't already done so, and the parse found a character that
15690 * could have been terminators for the name, but which we absorbed
15691 * as typos during the first pass, repeat the parse, signalling it
15692 * to stop at that character */
15693 if (possible_end && possible_end != (char *) -1) {
15694 possible_end = (char *) -1;
15699 /* Here neither pass found a close-enough class name */
15700 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15703 probably_meant_to_be:
15705 /* Here we think that a posix specification was intended. Update any
15707 if (updated_parse_ptr) {
15708 *updated_parse_ptr = (char *) p;
15711 /* If a posix class name was intended but incorrectly specified, we
15712 * output or return the warnings */
15713 if (found_problem) {
15715 /* We set flags for these issues in the parse loop above instead of
15716 * adding them to the list of warnings, because we can parse it
15717 * twice, and we only want one warning instance */
15719 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
15722 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15724 if (has_semi_colon) {
15725 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15727 else if (! has_terminating_colon) {
15728 ADD_POSIX_WARNING(p, "there is no terminating ':'");
15730 if (! has_terminating_bracket) {
15731 ADD_POSIX_WARNING(p, "there is no terminating ']'");
15734 if ( posix_warnings
15736 && av_top_index(RExC_warn_text) > -1)
15738 *posix_warnings = RExC_warn_text;
15741 else if (class_number != OOB_NAMEDCLASS) {
15742 /* If it is a known class, return the class. The class number
15743 * #defines are structured so each complement is +1 to the normal
15745 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
15747 else if (! check_only) {
15749 /* Here, it is an unrecognized class. This is an error (unless the
15750 * call is to check only, which we've already handled above) */
15751 const char * const complement_string = (complement)
15754 RExC_parse = (char *) p;
15755 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
15757 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
15761 return OOB_NAMEDCLASS;
15763 #undef ADD_POSIX_WARNING
15765 STATIC unsigned int
15766 S_regex_set_precedence(const U8 my_operator) {
15768 /* Returns the precedence in the (?[...]) construct of the input operator,
15769 * specified by its character representation. The precedence follows
15770 * general Perl rules, but it extends this so that ')' and ']' have (low)
15771 * precedence even though they aren't really operators */
15773 switch (my_operator) {
15789 NOT_REACHED; /* NOTREACHED */
15790 return 0; /* Silence compiler warning */
15793 STATIC regnode_offset
15794 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
15795 I32 *flagp, U32 depth,
15796 char * const oregcomp_parse)
15798 /* Handle the (?[...]) construct to do set operations */
15800 U8 curchar; /* Current character being parsed */
15801 UV start, end; /* End points of code point ranges */
15802 SV* final = NULL; /* The end result inversion list */
15803 SV* result_string; /* 'final' stringified */
15804 AV* stack; /* stack of operators and operands not yet
15806 AV* fence_stack = NULL; /* A stack containing the positions in
15807 'stack' of where the undealt-with left
15808 parens would be if they were actually
15810 /* The 'volatile' is a workaround for an optimiser bug
15811 * in Solaris Studio 12.3. See RT #127455 */
15812 volatile IV fence = 0; /* Position of where most recent undealt-
15813 with left paren in stack is; -1 if none.
15815 STRLEN len; /* Temporary */
15816 regnode_offset node; /* Temporary, and final regnode returned by
15818 const bool save_fold = FOLD; /* Temporary */
15819 char *save_end, *save_parse; /* Temporaries */
15820 const bool in_locale = LOC; /* we turn off /l during processing */
15822 GET_RE_DEBUG_FLAGS_DECL;
15824 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
15826 DEBUG_PARSE("xcls");
15829 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
15832 /* The use of this operator implies /u. This is required so that the
15833 * compile time values are valid in all runtime cases */
15834 REQUIRE_UNI_RULES(flagp, 0);
15836 ckWARNexperimental(RExC_parse,
15837 WARN_EXPERIMENTAL__REGEX_SETS,
15838 "The regex_sets feature is experimental");
15840 /* Everything in this construct is a metacharacter. Operands begin with
15841 * either a '\' (for an escape sequence), or a '[' for a bracketed
15842 * character class. Any other character should be an operator, or
15843 * parenthesis for grouping. Both types of operands are handled by calling
15844 * regclass() to parse them. It is called with a parameter to indicate to
15845 * return the computed inversion list. The parsing here is implemented via
15846 * a stack. Each entry on the stack is a single character representing one
15847 * of the operators; or else a pointer to an operand inversion list. */
15849 #define IS_OPERATOR(a) SvIOK(a)
15850 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15852 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15853 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15854 * with pronouncing it called it Reverse Polish instead, but now that YOU
15855 * know how to pronounce it you can use the correct term, thus giving due
15856 * credit to the person who invented it, and impressing your geek friends.
15857 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15858 * it is now more like an English initial W (as in wonk) than an L.)
15860 * This means that, for example, 'a | b & c' is stored on the stack as
15868 * where the numbers in brackets give the stack [array] element number.
15869 * In this implementation, parentheses are not stored on the stack.
15870 * Instead a '(' creates a "fence" so that the part of the stack below the
15871 * fence is invisible except to the corresponding ')' (this allows us to
15872 * replace testing for parens, by using instead subtraction of the fence
15873 * position). As new operands are processed they are pushed onto the stack
15874 * (except as noted in the next paragraph). New operators of higher
15875 * precedence than the current final one are inserted on the stack before
15876 * the lhs operand (so that when the rhs is pushed next, everything will be
15877 * in the correct positions shown above. When an operator of equal or
15878 * lower precedence is encountered in parsing, all the stacked operations
15879 * of equal or higher precedence are evaluated, leaving the result as the
15880 * top entry on the stack. This makes higher precedence operations
15881 * evaluate before lower precedence ones, and causes operations of equal
15882 * precedence to left associate.
15884 * The only unary operator '!' is immediately pushed onto the stack when
15885 * encountered. When an operand is encountered, if the top of the stack is
15886 * a '!", the complement is immediately performed, and the '!' popped. The
15887 * resulting value is treated as a new operand, and the logic in the
15888 * previous paragraph is executed. Thus in the expression
15890 * the stack looks like
15896 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15903 * A ')' is treated as an operator with lower precedence than all the
15904 * aforementioned ones, which causes all operations on the stack above the
15905 * corresponding '(' to be evaluated down to a single resultant operand.
15906 * Then the fence for the '(' is removed, and the operand goes through the
15907 * algorithm above, without the fence.
15909 * A separate stack is kept of the fence positions, so that the position of
15910 * the latest so-far unbalanced '(' is at the top of it.
15912 * The ']' ending the construct is treated as the lowest operator of all,
15913 * so that everything gets evaluated down to a single operand, which is the
15916 sv_2mortal((SV *)(stack = newAV()));
15917 sv_2mortal((SV *)(fence_stack = newAV()));
15919 while (RExC_parse < RExC_end) {
15920 I32 top_index; /* Index of top-most element in 'stack' */
15921 SV** top_ptr; /* Pointer to top 'stack' element */
15922 SV* current = NULL; /* To contain the current inversion list
15924 SV* only_to_avoid_leaks;
15926 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15927 TRUE /* Force /x */ );
15928 if (RExC_parse >= RExC_end) { /* Fail */
15932 curchar = UCHARAT(RExC_parse);
15936 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15937 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15938 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15939 stack, fence, fence_stack));
15942 top_index = av_tindex_skip_len_mg(stack);
15945 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15946 char stacked_operator; /* The topmost operator on the 'stack'. */
15947 SV* lhs; /* Operand to the left of the operator */
15948 SV* rhs; /* Operand to the right of the operator */
15949 SV* fence_ptr; /* Pointer to top element of the fence
15954 if ( RExC_parse < RExC_end - 2
15955 && UCHARAT(RExC_parse + 1) == '?'
15956 && UCHARAT(RExC_parse + 2) == '^')
15958 /* If is a '(?', could be an embedded '(?^flags:(?[...])'.
15959 * This happens when we have some thing like
15961 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15963 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15965 * Here we would be handling the interpolated
15966 * '$thai_or_lao'. We handle this by a recursive call to
15967 * ourselves which returns the inversion list the
15968 * interpolated expression evaluates to. We use the flags
15969 * from the interpolated pattern. */
15970 U32 save_flags = RExC_flags;
15971 const char * save_parse;
15973 RExC_parse += 2; /* Skip past the '(?' */
15974 save_parse = RExC_parse;
15976 /* Parse the flags for the '(?'. We already know the first
15977 * flag to parse is a '^' */
15978 parse_lparen_question_flags(pRExC_state);
15980 if ( RExC_parse >= RExC_end - 4
15981 || UCHARAT(RExC_parse) != ':'
15982 || UCHARAT(++RExC_parse) != '('
15983 || UCHARAT(++RExC_parse) != '?'
15984 || UCHARAT(++RExC_parse) != '[')
15987 /* In combination with the above, this moves the
15988 * pointer to the point just after the first erroneous
15990 if (RExC_parse >= RExC_end - 4) {
15991 RExC_parse = RExC_end;
15993 else if (RExC_parse != save_parse) {
15994 RExC_parse += (UTF)
15995 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
15998 vFAIL("Expecting '(?flags:(?[...'");
16001 /* Recurse, with the meat of the embedded expression */
16003 if (! handle_regex_sets(pRExC_state, ¤t, flagp,
16004 depth+1, oregcomp_parse))
16006 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16009 /* Here, 'current' contains the embedded expression's
16010 * inversion list, and RExC_parse points to the trailing
16011 * ']'; the next character should be the ')' */
16013 if (UCHARAT(RExC_parse) != ')')
16014 vFAIL("Expecting close paren for nested extended charclass");
16016 /* Then the ')' matching the original '(' handled by this
16017 * case: statement */
16019 if (UCHARAT(RExC_parse) != ')')
16020 vFAIL("Expecting close paren for wrapper for nested extended charclass");
16022 RExC_flags = save_flags;
16023 goto handle_operand;
16026 /* A regular '('. Look behind for illegal syntax */
16027 if (top_index - fence >= 0) {
16028 /* If the top entry on the stack is an operator, it had
16029 * better be a '!', otherwise the entry below the top
16030 * operand should be an operator */
16031 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16032 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16033 || ( IS_OPERAND(*top_ptr)
16034 && ( top_index - fence < 1
16035 || ! (stacked_ptr = av_fetch(stack,
16038 || ! IS_OPERATOR(*stacked_ptr))))
16041 vFAIL("Unexpected '(' with no preceding operator");
16045 /* Stack the position of this undealt-with left paren */
16046 av_push(fence_stack, newSViv(fence));
16047 fence = top_index + 1;
16051 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16052 * multi-char folds are allowed. */
16053 if (!regclass(pRExC_state, flagp, depth+1,
16054 TRUE, /* means parse just the next thing */
16055 FALSE, /* don't allow multi-char folds */
16056 FALSE, /* don't silence non-portable warnings. */
16058 FALSE, /* Require return to be an ANYOF */
16061 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16062 goto regclass_failed;
16065 /* regclass() will return with parsing just the \ sequence,
16066 * leaving the parse pointer at the next thing to parse */
16068 goto handle_operand;
16070 case '[': /* Is a bracketed character class */
16072 /* See if this is a [:posix:] class. */
16073 bool is_posix_class = (OOB_NAMEDCLASS
16074 < handle_possible_posix(pRExC_state,
16078 TRUE /* checking only */));
16079 /* If it is a posix class, leave the parse pointer at the '['
16080 * to fool regclass() into thinking it is part of a
16081 * '[[:posix:]]'. */
16082 if (! is_posix_class) {
16086 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16087 * multi-char folds are allowed. */
16088 if (!regclass(pRExC_state, flagp, depth+1,
16089 is_posix_class, /* parse the whole char
16090 class only if not a
16092 FALSE, /* don't allow multi-char folds */
16093 TRUE, /* silence non-portable warnings. */
16095 FALSE, /* Require return to be an ANYOF */
16098 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16099 goto regclass_failed;
16106 /* function call leaves parse pointing to the ']', except if we
16108 if (is_posix_class) {
16112 goto handle_operand;
16116 if (top_index >= 1) {
16117 goto join_operators;
16120 /* Only a single operand on the stack: are done */
16124 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16125 if (UCHARAT(RExC_parse - 1) == ']') {
16129 vFAIL("Unexpected ')'");
16132 /* If nothing after the fence, is missing an operand */
16133 if (top_index - fence < 0) {
16137 /* If at least two things on the stack, treat this as an
16139 if (top_index - fence >= 1) {
16140 goto join_operators;
16143 /* Here only a single thing on the fenced stack, and there is a
16144 * fence. Get rid of it */
16145 fence_ptr = av_pop(fence_stack);
16147 fence = SvIV(fence_ptr);
16148 SvREFCNT_dec_NN(fence_ptr);
16155 /* Having gotten rid of the fence, we pop the operand at the
16156 * stack top and process it as a newly encountered operand */
16157 current = av_pop(stack);
16158 if (IS_OPERAND(current)) {
16159 goto handle_operand;
16171 /* These binary operators should have a left operand already
16173 if ( top_index - fence < 0
16174 || top_index - fence == 1
16175 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16176 || ! IS_OPERAND(*top_ptr))
16178 goto unexpected_binary;
16181 /* If only the one operand is on the part of the stack visible
16182 * to us, we just place this operator in the proper position */
16183 if (top_index - fence < 2) {
16185 /* Place the operator before the operand */
16187 SV* lhs = av_pop(stack);
16188 av_push(stack, newSVuv(curchar));
16189 av_push(stack, lhs);
16193 /* But if there is something else on the stack, we need to
16194 * process it before this new operator if and only if the
16195 * stacked operation has equal or higher precedence than the
16200 /* The operator on the stack is supposed to be below both its
16202 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16203 || IS_OPERAND(*stacked_ptr))
16205 /* But if not, it's legal and indicates we are completely
16206 * done if and only if we're currently processing a ']',
16207 * which should be the final thing in the expression */
16208 if (curchar == ']') {
16214 vFAIL2("Unexpected binary operator '%c' with no "
16215 "preceding operand", curchar);
16217 stacked_operator = (char) SvUV(*stacked_ptr);
16219 if (regex_set_precedence(curchar)
16220 > regex_set_precedence(stacked_operator))
16222 /* Here, the new operator has higher precedence than the
16223 * stacked one. This means we need to add the new one to
16224 * the stack to await its rhs operand (and maybe more
16225 * stuff). We put it before the lhs operand, leaving
16226 * untouched the stacked operator and everything below it
16228 lhs = av_pop(stack);
16229 assert(IS_OPERAND(lhs));
16231 av_push(stack, newSVuv(curchar));
16232 av_push(stack, lhs);
16236 /* Here, the new operator has equal or lower precedence than
16237 * what's already there. This means the operation already
16238 * there should be performed now, before the new one. */
16240 rhs = av_pop(stack);
16241 if (! IS_OPERAND(rhs)) {
16243 /* This can happen when a ! is not followed by an operand,
16244 * like in /(?[\t &!])/ */
16248 lhs = av_pop(stack);
16250 if (! IS_OPERAND(lhs)) {
16252 /* This can happen when there is an empty (), like in
16253 * /(?[[0]+()+])/ */
16257 switch (stacked_operator) {
16259 _invlist_intersection(lhs, rhs, &rhs);
16264 _invlist_union(lhs, rhs, &rhs);
16268 _invlist_subtract(lhs, rhs, &rhs);
16271 case '^': /* The union minus the intersection */
16276 _invlist_union(lhs, rhs, &u);
16277 _invlist_intersection(lhs, rhs, &i);
16278 _invlist_subtract(u, i, &rhs);
16279 SvREFCNT_dec_NN(i);
16280 SvREFCNT_dec_NN(u);
16286 /* Here, the higher precedence operation has been done, and the
16287 * result is in 'rhs'. We overwrite the stacked operator with
16288 * the result. Then we redo this code to either push the new
16289 * operator onto the stack or perform any higher precedence
16290 * stacked operation */
16291 only_to_avoid_leaks = av_pop(stack);
16292 SvREFCNT_dec(only_to_avoid_leaks);
16293 av_push(stack, rhs);
16296 case '!': /* Highest priority, right associative */
16298 /* If what's already at the top of the stack is another '!",
16299 * they just cancel each other out */
16300 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16301 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16303 only_to_avoid_leaks = av_pop(stack);
16304 SvREFCNT_dec(only_to_avoid_leaks);
16306 else { /* Otherwise, since it's right associative, just push
16308 av_push(stack, newSVuv(curchar));
16313 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16314 if (RExC_parse >= RExC_end) {
16317 vFAIL("Unexpected character");
16321 /* Here 'current' is the operand. If something is already on the
16322 * stack, we have to check if it is a !. But first, the code above
16323 * may have altered the stack in the time since we earlier set
16326 top_index = av_tindex_skip_len_mg(stack);
16327 if (top_index - fence >= 0) {
16328 /* If the top entry on the stack is an operator, it had better
16329 * be a '!', otherwise the entry below the top operand should
16330 * be an operator */
16331 top_ptr = av_fetch(stack, top_index, FALSE);
16333 if (IS_OPERATOR(*top_ptr)) {
16335 /* The only permissible operator at the top of the stack is
16336 * '!', which is applied immediately to this operand. */
16337 curchar = (char) SvUV(*top_ptr);
16338 if (curchar != '!') {
16339 SvREFCNT_dec(current);
16340 vFAIL2("Unexpected binary operator '%c' with no "
16341 "preceding operand", curchar);
16344 _invlist_invert(current);
16346 only_to_avoid_leaks = av_pop(stack);
16347 SvREFCNT_dec(only_to_avoid_leaks);
16349 /* And we redo with the inverted operand. This allows
16350 * handling multiple ! in a row */
16351 goto handle_operand;
16353 /* Single operand is ok only for the non-binary ')'
16355 else if ((top_index - fence == 0 && curchar != ')')
16356 || (top_index - fence > 0
16357 && (! (stacked_ptr = av_fetch(stack,
16360 || IS_OPERAND(*stacked_ptr))))
16362 SvREFCNT_dec(current);
16363 vFAIL("Operand with no preceding operator");
16367 /* Here there was nothing on the stack or the top element was
16368 * another operand. Just add this new one */
16369 av_push(stack, current);
16371 } /* End of switch on next parse token */
16373 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16374 } /* End of loop parsing through the construct */
16376 vFAIL("Syntax error in (?[...])");
16380 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16381 if (RExC_parse < RExC_end) {
16385 vFAIL("Unexpected ']' with no following ')' in (?[...");
16388 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16389 vFAIL("Unmatched (");
16392 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16393 || ((final = av_pop(stack)) == NULL)
16394 || ! IS_OPERAND(final)
16395 || ! is_invlist(final)
16396 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16399 SvREFCNT_dec(final);
16400 vFAIL("Incomplete expression within '(?[ ])'");
16403 /* Here, 'final' is the resultant inversion list from evaluating the
16404 * expression. Return it if so requested */
16405 if (return_invlist) {
16406 *return_invlist = final;
16410 /* Otherwise generate a resultant node, based on 'final'. regclass() is
16411 * expecting a string of ranges and individual code points */
16412 invlist_iterinit(final);
16413 result_string = newSVpvs("");
16414 while (invlist_iternext(final, &start, &end)) {
16415 if (start == end) {
16416 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16419 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
16424 /* About to generate an ANYOF (or similar) node from the inversion list we
16425 * have calculated */
16426 save_parse = RExC_parse;
16427 RExC_parse = SvPV(result_string, len);
16428 save_end = RExC_end;
16429 RExC_end = RExC_parse + len;
16430 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16432 /* We turn off folding around the call, as the class we have constructed
16433 * already has all folding taken into consideration, and we don't want
16434 * regclass() to add to that */
16435 RExC_flags &= ~RXf_PMf_FOLD;
16436 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16437 * folds are allowed. */
16438 node = regclass(pRExC_state, flagp, depth+1,
16439 FALSE, /* means parse the whole char class */
16440 FALSE, /* don't allow multi-char folds */
16441 TRUE, /* silence non-portable warnings. The above may very
16442 well have generated non-portable code points, but
16443 they're valid on this machine */
16444 FALSE, /* similarly, no need for strict */
16445 FALSE, /* Require return to be an ANYOF */
16450 RExC_parse = save_parse + 1;
16451 RExC_end = save_end;
16452 SvREFCNT_dec_NN(final);
16453 SvREFCNT_dec_NN(result_string);
16456 RExC_flags |= RXf_PMf_FOLD;
16460 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16461 goto regclass_failed;
16464 /* Fix up the node type if we are in locale. (We have pretended we are
16465 * under /u for the purposes of regclass(), as this construct will only
16466 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
16467 * as to cause any warnings about bad locales to be output in regexec.c),
16468 * and add the flag that indicates to check if not in a UTF-8 locale. The
16469 * reason we above forbid optimization into something other than an ANYOF
16470 * node is simply to minimize the number of code changes in regexec.c.
16471 * Otherwise we would have to create new EXACTish node types and deal with
16472 * them. This decision could be revisited should this construct become
16475 * (One might think we could look at the resulting ANYOF node and suppress
16476 * the flag if everything is above 255, as those would be UTF-8 only,
16477 * but this isn't true, as the components that led to that result could
16478 * have been locale-affected, and just happen to cancel each other out
16479 * under UTF-8 locales.) */
16481 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16483 assert(OP(REGNODE_p(node)) == ANYOF);
16485 OP(REGNODE_p(node)) = ANYOFL;
16486 ANYOF_FLAGS(REGNODE_p(node))
16487 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16490 nextchar(pRExC_state);
16491 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
16495 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
16499 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16502 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16503 AV * stack, const IV fence, AV * fence_stack)
16504 { /* Dumps the stacks in handle_regex_sets() */
16506 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16507 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16510 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16512 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16514 if (stack_top < 0) {
16515 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16518 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16519 for (i = stack_top; i >= 0; i--) {
16520 SV ** element_ptr = av_fetch(stack, i, FALSE);
16521 if (! element_ptr) {
16524 if (IS_OPERATOR(*element_ptr)) {
16525 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16526 (int) i, (int) SvIV(*element_ptr));
16529 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16530 sv_dump(*element_ptr);
16535 if (fence_stack_top < 0) {
16536 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16539 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16540 for (i = fence_stack_top; i >= 0; i--) {
16541 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16542 if (! element_ptr) {
16545 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16546 (int) i, (int) SvIV(*element_ptr));
16557 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16559 /* This adds the Latin1/above-Latin1 folding rules.
16561 * This should be called only for a Latin1-range code points, cp, which is
16562 * known to be involved in a simple fold with other code points above
16563 * Latin1. It would give false results if /aa has been specified.
16564 * Multi-char folds are outside the scope of this, and must be handled
16567 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
16569 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
16571 /* The rules that are valid for all Unicode versions are hard-coded in */
16576 add_cp_to_invlist(*invlist, KELVIN_SIGN);
16580 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
16583 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
16584 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
16586 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
16587 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
16588 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
16590 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
16591 *invlist = add_cp_to_invlist(*invlist,
16592 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
16595 default: /* Other code points are checked against the data for the
16596 current Unicode version */
16598 Size_t folds_count;
16599 unsigned int first_fold;
16600 const unsigned int * remaining_folds;
16604 folded_cp = toFOLD(cp);
16607 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
16609 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
16612 if (folded_cp > 255) {
16613 *invlist = add_cp_to_invlist(*invlist, folded_cp);
16616 folds_count = _inverse_folds(folded_cp, &first_fold,
16618 if (folds_count == 0) {
16620 /* Use deprecated warning to increase the chances of this being
16622 ckWARN2reg_d(RExC_parse,
16623 "Perl folding rules are not up-to-date for 0x%02X;"
16624 " please use the perlbug utility to report;", cp);
16629 if (first_fold > 255) {
16630 *invlist = add_cp_to_invlist(*invlist, first_fold);
16632 for (i = 0; i < folds_count - 1; i++) {
16633 if (remaining_folds[i] > 255) {
16634 *invlist = add_cp_to_invlist(*invlist,
16635 remaining_folds[i]);
16645 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
16647 /* Output the elements of the array given by '*posix_warnings' as REGEXP
16651 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
16653 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
16655 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
16659 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
16660 if (first_is_fatal) { /* Avoid leaking this */
16661 av_undef(posix_warnings); /* This isn't necessary if the
16662 array is mortal, but is a
16664 (void) sv_2mortal(msg);
16667 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
16668 SvREFCNT_dec_NN(msg);
16671 UPDATE_WARNINGS_LOC(RExC_parse);
16675 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
16677 /* This adds the string scalar <multi_string> to the array
16678 * <multi_char_matches>. <multi_string> is known to have exactly
16679 * <cp_count> code points in it. This is used when constructing a
16680 * bracketed character class and we find something that needs to match more
16681 * than a single character.
16683 * <multi_char_matches> is actually an array of arrays. Each top-level
16684 * element is an array that contains all the strings known so far that are
16685 * the same length. And that length (in number of code points) is the same
16686 * as the index of the top-level array. Hence, the [2] element is an
16687 * array, each element thereof is a string containing TWO code points;
16688 * while element [3] is for strings of THREE characters, and so on. Since
16689 * this is for multi-char strings there can never be a [0] nor [1] element.
16691 * When we rewrite the character class below, we will do so such that the
16692 * longest strings are written first, so that it prefers the longest
16693 * matching strings first. This is done even if it turns out that any
16694 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
16695 * Christiansen has agreed that this is ok. This makes the test for the
16696 * ligature 'ffi' come before the test for 'ff', for example */
16699 AV** this_array_ptr;
16701 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
16703 if (! multi_char_matches) {
16704 multi_char_matches = newAV();
16707 if (av_exists(multi_char_matches, cp_count)) {
16708 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
16709 this_array = *this_array_ptr;
16712 this_array = newAV();
16713 av_store(multi_char_matches, cp_count,
16716 av_push(this_array, multi_string);
16718 return multi_char_matches;
16721 /* The names of properties whose definitions are not known at compile time are
16722 * stored in this SV, after a constant heading. So if the length has been
16723 * changed since initialization, then there is a run-time definition. */
16724 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
16725 (SvCUR(listsv) != initial_listsv_len)
16727 /* There is a restricted set of white space characters that are legal when
16728 * ignoring white space in a bracketed character class. This generates the
16729 * code to skip them.
16731 * There is a line below that uses the same white space criteria but is outside
16732 * this macro. Both here and there must use the same definition */
16733 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
16736 while (isBLANK_A(UCHARAT(p))) \
16743 STATIC regnode_offset
16744 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
16745 const bool stop_at_1, /* Just parse the next thing, don't
16746 look for a full character class */
16747 bool allow_mutiple_chars,
16748 const bool silence_non_portable, /* Don't output warnings
16752 bool optimizable, /* ? Allow a non-ANYOF return
16754 SV** ret_invlist /* Return an inversion list, not a node */
16757 /* parse a bracketed class specification. Most of these will produce an
16758 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
16759 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
16760 * under /i with multi-character folds: it will be rewritten following the
16761 * paradigm of this example, where the <multi-fold>s are characters which
16762 * fold to multiple character sequences:
16763 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
16764 * gets effectively rewritten as:
16765 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
16766 * reg() gets called (recursively) on the rewritten version, and this
16767 * function will return what it constructs. (Actually the <multi-fold>s
16768 * aren't physically removed from the [abcdefghi], it's just that they are
16769 * ignored in the recursion by means of a flag:
16770 * <RExC_in_multi_char_class>.)
16772 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
16773 * characters, with the corresponding bit set if that character is in the
16774 * list. For characters above this, an inversion list is used. There
16775 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
16776 * determinable at compile time
16778 * On success, returns the offset at which any next node should be placed
16779 * into the regex engine program being compiled.
16781 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
16782 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
16787 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
16789 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
16790 regnode_offset ret = -1; /* Initialized to an illegal value */
16792 int namedclass = OOB_NAMEDCLASS;
16793 char *rangebegin = NULL;
16794 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
16795 aren't available at the time this was called */
16796 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
16797 than just initialized. */
16798 SV* properties = NULL; /* Code points that match \p{} \P{} */
16799 SV* posixes = NULL; /* Code points that match classes like [:word:],
16800 extended beyond the Latin1 range. These have to
16801 be kept separate from other code points for much
16802 of this function because their handling is
16803 different under /i, and for most classes under
16805 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
16806 separate for a while from the non-complemented
16807 versions because of complications with /d
16809 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
16810 treated more simply than the general case,
16811 leading to less compilation and execution
16813 UV element_count = 0; /* Number of distinct elements in the class.
16814 Optimizations may be possible if this is tiny */
16815 AV * multi_char_matches = NULL; /* Code points that fold to more than one
16816 character; used under /i */
16818 char * stop_ptr = RExC_end; /* where to stop parsing */
16820 /* ignore unescaped whitespace? */
16821 const bool skip_white = cBOOL( ret_invlist
16822 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
16824 /* inversion list of code points this node matches only when the target
16825 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
16827 SV* upper_latin1_only_utf8_matches = NULL;
16829 /* Inversion list of code points this node matches regardless of things
16830 * like locale, folding, utf8ness of the target string */
16831 SV* cp_list = NULL;
16833 /* Like cp_list, but code points on this list need to be checked for things
16834 * that fold to/from them under /i */
16835 SV* cp_foldable_list = NULL;
16837 /* Like cp_list, but code points on this list are valid only when the
16838 * runtime locale is UTF-8 */
16839 SV* only_utf8_locale_list = NULL;
16841 /* In a range, if one of the endpoints is non-character-set portable,
16842 * meaning that it hard-codes a code point that may mean a different
16843 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16844 * mnemonic '\t' which each mean the same character no matter which
16845 * character set the platform is on. */
16846 unsigned int non_portable_endpoint = 0;
16848 /* Is the range unicode? which means on a platform that isn't 1-1 native
16849 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16850 * to be a Unicode value. */
16851 bool unicode_range = FALSE;
16852 bool invert = FALSE; /* Is this class to be complemented */
16854 bool warn_super = ALWAYS_WARN_SUPER;
16856 const char * orig_parse = RExC_parse;
16858 /* This variable is used to mark where the end in the input is of something
16859 * that looks like a POSIX construct but isn't. During the parse, when
16860 * something looks like it could be such a construct is encountered, it is
16861 * checked for being one, but not if we've already checked this area of the
16862 * input. Only after this position is reached do we check again */
16863 char *not_posix_region_end = RExC_parse - 1;
16865 AV* posix_warnings = NULL;
16866 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
16867 U8 op = END; /* The returned node-type, initialized to an impossible
16869 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
16870 U32 posixl = 0; /* bit field of posix classes matched under /l */
16873 /* Flags as to what things aren't knowable until runtime. (Note that these are
16874 * mutually exclusive.) */
16875 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
16876 haven't been defined as of yet */
16877 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
16879 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
16880 what gets folded */
16881 U32 has_runtime_dependency = 0; /* OR of the above flags */
16883 GET_RE_DEBUG_FLAGS_DECL;
16885 PERL_ARGS_ASSERT_REGCLASS;
16887 PERL_UNUSED_ARG(depth);
16891 /* If wants an inversion list returned, we can't optimize to something
16894 optimizable = FALSE;
16897 DEBUG_PARSE("clas");
16899 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16900 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16901 && UNICODE_DOT_DOT_VERSION == 0)
16902 allow_mutiple_chars = FALSE;
16905 /* We include the /i status at the beginning of this so that we can
16906 * know it at runtime */
16907 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
16908 initial_listsv_len = SvCUR(listsv);
16909 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16911 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16913 assert(RExC_parse <= RExC_end);
16915 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16918 allow_mutiple_chars = FALSE;
16920 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16923 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16924 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16925 int maybe_class = handle_possible_posix(pRExC_state,
16927 ¬_posix_region_end,
16929 TRUE /* checking only */);
16930 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16931 ckWARN4reg(not_posix_region_end,
16932 "POSIX syntax [%c %c] belongs inside character classes%s",
16933 *RExC_parse, *RExC_parse,
16934 (maybe_class == OOB_NAMEDCLASS)
16935 ? ((POSIXCC_NOTYET(*RExC_parse))
16936 ? " (but this one isn't implemented)"
16937 : " (but this one isn't fully valid)")
16943 /* If the caller wants us to just parse a single element, accomplish this
16944 * by faking the loop ending condition */
16945 if (stop_at_1 && RExC_end > RExC_parse) {
16946 stop_ptr = RExC_parse + 1;
16949 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16950 if (UCHARAT(RExC_parse) == ']')
16951 goto charclassloop;
16955 if ( posix_warnings
16956 && av_tindex_skip_len_mg(posix_warnings) >= 0
16957 && RExC_parse > not_posix_region_end)
16959 /* Warnings about posix class issues are considered tentative until
16960 * we are far enough along in the parse that we can no longer
16961 * change our mind, at which point we output them. This is done
16962 * each time through the loop so that a later class won't zap them
16963 * before they have been dealt with. */
16964 output_posix_warnings(pRExC_state, posix_warnings);
16967 if (RExC_parse >= stop_ptr) {
16971 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16973 if (UCHARAT(RExC_parse) == ']') {
16979 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16980 save_value = value;
16981 save_prevvalue = prevvalue;
16984 rangebegin = RExC_parse;
16986 non_portable_endpoint = 0;
16988 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16989 value = utf8n_to_uvchr((U8*)RExC_parse,
16990 RExC_end - RExC_parse,
16991 &numlen, UTF8_ALLOW_DEFAULT);
16992 RExC_parse += numlen;
16995 value = UCHARAT(RExC_parse++);
16997 if (value == '[') {
16998 char * posix_class_end;
16999 namedclass = handle_possible_posix(pRExC_state,
17002 do_posix_warnings ? &posix_warnings : NULL,
17003 FALSE /* die if error */);
17004 if (namedclass > OOB_NAMEDCLASS) {
17006 /* If there was an earlier attempt to parse this particular
17007 * posix class, and it failed, it was a false alarm, as this
17008 * successful one proves */
17009 if ( posix_warnings
17010 && av_tindex_skip_len_mg(posix_warnings) >= 0
17011 && not_posix_region_end >= RExC_parse
17012 && not_posix_region_end <= posix_class_end)
17014 av_undef(posix_warnings);
17017 RExC_parse = posix_class_end;
17019 else if (namedclass == OOB_NAMEDCLASS) {
17020 not_posix_region_end = posix_class_end;
17023 namedclass = OOB_NAMEDCLASS;
17026 else if ( RExC_parse - 1 > not_posix_region_end
17027 && MAYBE_POSIXCC(value))
17029 (void) handle_possible_posix(
17031 RExC_parse - 1, /* -1 because parse has already been
17033 ¬_posix_region_end,
17034 do_posix_warnings ? &posix_warnings : NULL,
17035 TRUE /* checking only */);
17037 else if ( strict && ! skip_white
17038 && ( _generic_isCC(value, _CC_VERTSPACE)
17039 || is_VERTWS_cp_high(value)))
17041 vFAIL("Literal vertical space in [] is illegal except under /x");
17043 else if (value == '\\') {
17044 /* Is a backslash; get the code point of the char after it */
17046 if (RExC_parse >= RExC_end) {
17047 vFAIL("Unmatched [");
17050 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17051 value = utf8n_to_uvchr((U8*)RExC_parse,
17052 RExC_end - RExC_parse,
17053 &numlen, UTF8_ALLOW_DEFAULT);
17054 RExC_parse += numlen;
17057 value = UCHARAT(RExC_parse++);
17059 /* Some compilers cannot handle switching on 64-bit integer
17060 * values, therefore value cannot be an UV. Yes, this will
17061 * be a problem later if we want switch on Unicode.
17062 * A similar issue a little bit later when switching on
17063 * namedclass. --jhi */
17065 /* If the \ is escaping white space when white space is being
17066 * skipped, it means that that white space is wanted literally, and
17067 * is already in 'value'. Otherwise, need to translate the escape
17068 * into what it signifies. */
17069 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17071 case 'w': namedclass = ANYOF_WORDCHAR; break;
17072 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17073 case 's': namedclass = ANYOF_SPACE; break;
17074 case 'S': namedclass = ANYOF_NSPACE; break;
17075 case 'd': namedclass = ANYOF_DIGIT; break;
17076 case 'D': namedclass = ANYOF_NDIGIT; break;
17077 case 'v': namedclass = ANYOF_VERTWS; break;
17078 case 'V': namedclass = ANYOF_NVERTWS; break;
17079 case 'h': namedclass = ANYOF_HORIZWS; break;
17080 case 'H': namedclass = ANYOF_NHORIZWS; break;
17081 case 'N': /* Handle \N{NAME} in class */
17083 const char * const backslash_N_beg = RExC_parse - 2;
17086 if (! grok_bslash_N(pRExC_state,
17087 NULL, /* No regnode */
17088 &value, /* Yes single value */
17089 &cp_count, /* Multiple code pt count */
17095 if (*flagp & NEED_UTF8)
17096 FAIL("panic: grok_bslash_N set NEED_UTF8");
17098 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17100 if (cp_count < 0) {
17101 vFAIL("\\N in a character class must be a named character: \\N{...}");
17103 else if (cp_count == 0) {
17104 ckWARNreg(RExC_parse,
17105 "Ignoring zero length \\N{} in character class");
17107 else { /* cp_count > 1 */
17108 assert(cp_count > 1);
17109 if (! RExC_in_multi_char_class) {
17110 if ( ! allow_mutiple_chars
17113 || *RExC_parse == '-')
17117 vFAIL("\\N{} here is restricted to one character");
17119 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17120 break; /* <value> contains the first code
17121 point. Drop out of the switch to
17125 SV * multi_char_N = newSVpvn(backslash_N_beg,
17126 RExC_parse - backslash_N_beg);
17128 = add_multi_match(multi_char_matches,
17133 } /* End of cp_count != 1 */
17135 /* This element should not be processed further in this
17138 value = save_value;
17139 prevvalue = save_prevvalue;
17140 continue; /* Back to top of loop to get next char */
17143 /* Here, is a single code point, and <value> contains it */
17144 unicode_range = TRUE; /* \N{} are Unicode */
17152 /* \p means they want Unicode semantics */
17153 REQUIRE_UNI_RULES(flagp, 0);
17155 if (RExC_parse >= RExC_end)
17156 vFAIL2("Empty \\%c", (U8)value);
17157 if (*RExC_parse == '{') {
17158 const U8 c = (U8)value;
17159 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17162 vFAIL2("Missing right brace on \\%c{}", c);
17167 /* White space is allowed adjacent to the braces and after
17168 * any '^', even when not under /x */
17169 while (isSPACE(*RExC_parse)) {
17173 if (UCHARAT(RExC_parse) == '^') {
17175 /* toggle. (The rhs xor gets the single bit that
17176 * differs between P and p; the other xor inverts just
17178 value ^= 'P' ^ 'p';
17181 while (isSPACE(*RExC_parse)) {
17186 if (e == RExC_parse)
17187 vFAIL2("Empty \\%c{}", c);
17189 n = e - RExC_parse;
17190 while (isSPACE(*(RExC_parse + n - 1)))
17193 } /* The \p isn't immediately followed by a '{' */
17194 else if (! isALPHA(*RExC_parse)) {
17195 RExC_parse += (UTF)
17196 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17198 vFAIL2("Character following \\%c must be '{' or a "
17199 "single-character Unicode property name",
17207 char* name = RExC_parse;
17209 /* Any message returned about expanding the definition */
17210 SV* msg = newSVpvs_flags("", SVs_TEMP);
17212 /* If set TRUE, the property is user-defined as opposed to
17213 * official Unicode */
17214 bool user_defined = FALSE;
17216 SV * prop_definition = parse_uniprop_string(
17217 name, n, UTF, FOLD,
17218 FALSE, /* This is compile-time */
17220 /* We can't defer this defn when
17221 * the full result is required in
17223 ! cBOOL(ret_invlist),
17229 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17230 assert(prop_definition == NULL);
17231 RExC_parse = e + 1;
17232 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17233 thing so, or else the display is
17237 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17238 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17239 SvCUR(msg), SvPVX(msg)));
17242 if (! is_invlist(prop_definition)) {
17244 /* Here, the definition isn't known, so we have gotten
17245 * returned a string that will be evaluated if and when
17246 * encountered at runtime. We add it to the list of
17247 * such properties, along with whether it should be
17248 * complemented or not */
17249 if (value == 'P') {
17250 sv_catpvs(listsv, "!");
17253 sv_catpvs(listsv, "+");
17255 sv_catsv(listsv, prop_definition);
17257 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17259 /* We don't know yet what this matches, so have to flag
17261 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17264 assert (prop_definition && is_invlist(prop_definition));
17266 /* Here we do have the complete property definition
17268 * Temporary workaround for [perl #133136]. For this
17269 * precise input that is in the .t that is failing,
17270 * load utf8.pm, which is what the test wants, so that
17271 * that .t passes */
17272 if ( memEQs(RExC_start, e + 1 - RExC_start,
17274 && ! hv_common(GvHVn(PL_incgv),
17276 "utf8.pm", sizeof("utf8.pm") - 1,
17277 0, HV_FETCH_ISEXISTS, NULL, 0))
17279 require_pv("utf8.pm");
17282 if (! user_defined &&
17283 /* We warn on matching an above-Unicode code point
17284 * if the match would return true, except don't
17285 * warn for \p{All}, which has exactly one element
17287 (_invlist_contains_cp(prop_definition, 0x110000)
17288 && (! (_invlist_len(prop_definition) == 1
17289 && *invlist_array(prop_definition) == 0))))
17294 /* Invert if asking for the complement */
17295 if (value == 'P') {
17296 _invlist_union_complement_2nd(properties,
17301 _invlist_union(properties, prop_definition, &properties);
17306 RExC_parse = e + 1;
17307 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17311 case 'n': value = '\n'; break;
17312 case 'r': value = '\r'; break;
17313 case 't': value = '\t'; break;
17314 case 'f': value = '\f'; break;
17315 case 'b': value = '\b'; break;
17316 case 'e': value = ESC_NATIVE; break;
17317 case 'a': value = '\a'; break;
17319 RExC_parse--; /* function expects to be pointed at the 'o' */
17321 const char* error_msg;
17322 bool valid = grok_bslash_o(&RExC_parse,
17326 TO_OUTPUT_WARNINGS(RExC_parse),
17328 silence_non_portable,
17333 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17335 non_portable_endpoint++;
17338 RExC_parse--; /* function expects to be pointed at the 'x' */
17340 const char* error_msg;
17341 bool valid = grok_bslash_x(&RExC_parse,
17345 TO_OUTPUT_WARNINGS(RExC_parse),
17347 silence_non_portable,
17352 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17354 non_portable_endpoint++;
17357 value = grok_bslash_c(*RExC_parse, TO_OUTPUT_WARNINGS(RExC_parse));
17358 UPDATE_WARNINGS_LOC(RExC_parse);
17360 non_portable_endpoint++;
17362 case '0': case '1': case '2': case '3': case '4':
17363 case '5': case '6': case '7':
17365 /* Take 1-3 octal digits */
17366 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
17367 numlen = (strict) ? 4 : 3;
17368 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17369 RExC_parse += numlen;
17372 RExC_parse += (UTF)
17373 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17375 vFAIL("Need exactly 3 octal digits");
17377 else if ( numlen < 3 /* like \08, \178 */
17378 && RExC_parse < RExC_end
17379 && isDIGIT(*RExC_parse)
17380 && ckWARN(WARN_REGEXP))
17382 reg_warn_non_literal_string(
17384 form_short_octal_warning(RExC_parse, numlen));
17387 non_portable_endpoint++;
17391 /* Allow \_ to not give an error */
17392 if (isWORDCHAR(value) && value != '_') {
17394 vFAIL2("Unrecognized escape \\%c in character class",
17398 ckWARN2reg(RExC_parse,
17399 "Unrecognized escape \\%c in character class passed through",
17404 } /* End of switch on char following backslash */
17405 } /* end of handling backslash escape sequences */
17407 /* Here, we have the current token in 'value' */
17409 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17412 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17413 * literal, as is the character that began the false range, i.e.
17414 * the 'a' in the examples */
17416 const int w = (RExC_parse >= rangebegin)
17417 ? RExC_parse - rangebegin
17421 "False [] range \"%" UTF8f "\"",
17422 UTF8fARG(UTF, w, rangebegin));
17425 ckWARN2reg(RExC_parse,
17426 "False [] range \"%" UTF8f "\"",
17427 UTF8fARG(UTF, w, rangebegin));
17428 cp_list = add_cp_to_invlist(cp_list, '-');
17429 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17433 range = 0; /* this was not a true range */
17434 element_count += 2; /* So counts for three values */
17437 classnum = namedclass_to_classnum(namedclass);
17439 if (LOC && namedclass < ANYOF_POSIXL_MAX
17440 #ifndef HAS_ISASCII
17441 && classnum != _CC_ASCII
17444 SV* scratch_list = NULL;
17446 /* What the Posix classes (like \w, [:space:]) match isn't
17447 * generally knowable under locale until actual match time. A
17448 * special node is used for these which has extra space for a
17449 * bitmap, with a bit reserved for each named class that is to
17450 * be matched against. (This isn't needed for \p{} and
17451 * pseudo-classes, as they are not affected by locale, and
17452 * hence are dealt with separately.) However, if a named class
17453 * and its complement are both present, then it matches
17454 * everything, and there is no runtime dependency. Odd numbers
17455 * are the complements of the next lower number, so xor works.
17456 * (Note that something like [\w\D] should match everything,
17457 * because \d should be a proper subset of \w. But rather than
17458 * trust that the locale is well behaved, we leave this to
17459 * runtime to sort out) */
17460 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
17461 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
17462 POSIXL_ZERO(posixl);
17463 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
17464 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
17465 continue; /* We could ignore the rest of the class, but
17466 best to parse it for any errors */
17468 else { /* Here, isn't the complement of any already parsed
17470 POSIXL_SET(posixl, namedclass);
17471 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
17472 anyof_flags |= ANYOF_MATCHES_POSIXL;
17474 /* The above-Latin1 characters are not subject to locale
17475 * rules. Just add them to the unconditionally-matched
17478 /* Get the list of the above-Latin1 code points this
17480 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17481 PL_XPosix_ptrs[classnum],
17483 /* Odd numbers are complements,
17484 * like NDIGIT, NASCII, ... */
17485 namedclass % 2 != 0,
17487 /* Checking if 'cp_list' is NULL first saves an extra
17488 * clone. Its reference count will be decremented at the
17489 * next union, etc, or if this is the only instance, at the
17490 * end of the routine */
17492 cp_list = scratch_list;
17495 _invlist_union(cp_list, scratch_list, &cp_list);
17496 SvREFCNT_dec_NN(scratch_list);
17498 continue; /* Go get next character */
17503 /* Here, is not /l, or is a POSIX class for which /l doesn't
17504 * matter (or is a Unicode property, which is skipped here). */
17505 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17506 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17508 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17509 * nor /l make a difference in what these match,
17510 * therefore we just add what they match to cp_list. */
17511 if (classnum != _CC_VERTSPACE) {
17512 assert( namedclass == ANYOF_HORIZWS
17513 || namedclass == ANYOF_NHORIZWS);
17515 /* It turns out that \h is just a synonym for
17517 classnum = _CC_BLANK;
17520 _invlist_union_maybe_complement_2nd(
17522 PL_XPosix_ptrs[classnum],
17523 namedclass % 2 != 0, /* Complement if odd
17524 (NHORIZWS, NVERTWS)
17529 else if ( AT_LEAST_UNI_SEMANTICS
17530 || classnum == _CC_ASCII
17531 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
17532 || classnum == _CC_XDIGIT)))
17534 /* We usually have to worry about /d affecting what POSIX
17535 * classes match, with special code needed because we won't
17536 * know until runtime what all matches. But there is no
17537 * extra work needed under /u and /a; and [:ascii:] is
17538 * unaffected by /d; and :digit: and :xdigit: don't have
17539 * runtime differences under /d. So we can special case
17540 * these, and avoid some extra work below, and at runtime.
17542 _invlist_union_maybe_complement_2nd(
17544 ((AT_LEAST_ASCII_RESTRICTED)
17545 ? PL_Posix_ptrs[classnum]
17546 : PL_XPosix_ptrs[classnum]),
17547 namedclass % 2 != 0,
17550 else { /* Garden variety class. If is NUPPER, NALPHA, ...
17551 complement and use nposixes */
17552 SV** posixes_ptr = namedclass % 2 == 0
17555 _invlist_union_maybe_complement_2nd(
17557 PL_XPosix_ptrs[classnum],
17558 namedclass % 2 != 0,
17562 } /* end of namedclass \blah */
17564 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17566 /* If 'range' is set, 'value' is the ending of a range--check its
17567 * validity. (If value isn't a single code point in the case of a
17568 * range, we should have figured that out above in the code that
17569 * catches false ranges). Later, we will handle each individual code
17570 * point in the range. If 'range' isn't set, this could be the
17571 * beginning of a range, so check for that by looking ahead to see if
17572 * the next real character to be processed is the range indicator--the
17577 /* For unicode ranges, we have to test that the Unicode as opposed
17578 * to the native values are not decreasing. (Above 255, there is
17579 * no difference between native and Unicode) */
17580 if (unicode_range && prevvalue < 255 && value < 255) {
17581 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
17582 goto backwards_range;
17587 if (prevvalue > value) /* b-a */ {
17592 w = RExC_parse - rangebegin;
17594 "Invalid [] range \"%" UTF8f "\"",
17595 UTF8fARG(UTF, w, rangebegin));
17596 NOT_REACHED; /* NOTREACHED */
17600 prevvalue = value; /* save the beginning of the potential range */
17601 if (! stop_at_1 /* Can't be a range if parsing just one thing */
17602 && *RExC_parse == '-')
17604 char* next_char_ptr = RExC_parse + 1;
17606 /* Get the next real char after the '-' */
17607 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
17609 /* If the '-' is at the end of the class (just before the ']',
17610 * it is a literal minus; otherwise it is a range */
17611 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
17612 RExC_parse = next_char_ptr;
17614 /* a bad range like \w-, [:word:]- ? */
17615 if (namedclass > OOB_NAMEDCLASS) {
17616 if (strict || ckWARN(WARN_REGEXP)) {
17617 const int w = RExC_parse >= rangebegin
17618 ? RExC_parse - rangebegin
17621 vFAIL4("False [] range \"%*.*s\"",
17626 "False [] range \"%*.*s\"",
17630 cp_list = add_cp_to_invlist(cp_list, '-');
17633 range = 1; /* yeah, it's a range! */
17634 continue; /* but do it the next time */
17639 if (namedclass > OOB_NAMEDCLASS) {
17643 /* Here, we have a single value this time through the loop, and
17644 * <prevvalue> is the beginning of the range, if any; or <value> if
17647 /* non-Latin1 code point implies unicode semantics. */
17649 REQUIRE_UNI_RULES(flagp, 0);
17652 /* Ready to process either the single value, or the completed range.
17653 * For single-valued non-inverted ranges, we consider the possibility
17654 * of multi-char folds. (We made a conscious decision to not do this
17655 * for the other cases because it can often lead to non-intuitive
17656 * results. For example, you have the peculiar case that:
17657 * "s s" =~ /^[^\xDF]+$/i => Y
17658 * "ss" =~ /^[^\xDF]+$/i => N
17660 * See [perl #89750] */
17661 if (FOLD && allow_mutiple_chars && value == prevvalue) {
17662 if ( value == LATIN_SMALL_LETTER_SHARP_S
17663 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
17666 /* Here <value> is indeed a multi-char fold. Get what it is */
17668 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17671 UV folded = _to_uni_fold_flags(
17675 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
17676 ? FOLD_FLAGS_NOMIX_ASCII
17680 /* Here, <folded> should be the first character of the
17681 * multi-char fold of <value>, with <foldbuf> containing the
17682 * whole thing. But, if this fold is not allowed (because of
17683 * the flags), <fold> will be the same as <value>, and should
17684 * be processed like any other character, so skip the special
17686 if (folded != value) {
17688 /* Skip if we are recursed, currently parsing the class
17689 * again. Otherwise add this character to the list of
17690 * multi-char folds. */
17691 if (! RExC_in_multi_char_class) {
17692 STRLEN cp_count = utf8_length(foldbuf,
17693 foldbuf + foldlen);
17694 SV* multi_fold = sv_2mortal(newSVpvs(""));
17696 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
17699 = add_multi_match(multi_char_matches,
17705 /* This element should not be processed further in this
17708 value = save_value;
17709 prevvalue = save_prevvalue;
17715 if (strict && ckWARN(WARN_REGEXP)) {
17718 /* If the range starts above 255, everything is portable and
17719 * likely to be so for any forseeable character set, so don't
17721 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
17722 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
17724 else if (prevvalue != value) {
17726 /* Under strict, ranges that stop and/or end in an ASCII
17727 * printable should have each end point be a portable value
17728 * for it (preferably like 'A', but we don't warn if it is
17729 * a (portable) Unicode name or code point), and the range
17730 * must be be all digits or all letters of the same case.
17731 * Otherwise, the range is non-portable and unclear as to
17732 * what it contains */
17733 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
17734 && ( non_portable_endpoint
17735 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
17736 || (isLOWER_A(prevvalue) && isLOWER_A(value))
17737 || (isUPPER_A(prevvalue) && isUPPER_A(value))
17739 vWARN(RExC_parse, "Ranges of ASCII printables should"
17740 " be some subset of \"0-9\","
17741 " \"A-Z\", or \"a-z\"");
17743 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
17744 SSize_t index_start;
17745 SSize_t index_final;
17747 /* But the nature of Unicode and languages mean we
17748 * can't do the same checks for above-ASCII ranges,
17749 * except in the case of digit ones. These should
17750 * contain only digits from the same group of 10. The
17751 * ASCII case is handled just above. Hence here, the
17752 * range could be a range of digits. First some
17753 * unlikely special cases. Grandfather in that a range
17754 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
17755 * if its starting value is one of the 10 digits prior
17756 * to it. This is because it is an alternate way of
17757 * writing 19D1, and some people may expect it to be in
17758 * that group. But it is bad, because it won't give
17759 * the expected results. In Unicode 5.2 it was
17760 * considered to be in that group (of 11, hence), but
17761 * this was fixed in the next version */
17763 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
17764 goto warn_bad_digit_range;
17766 else if (UNLIKELY( prevvalue >= 0x1D7CE
17767 && value <= 0x1D7FF))
17769 /* This is the only other case currently in Unicode
17770 * where the algorithm below fails. The code
17771 * points just above are the end points of a single
17772 * range containing only decimal digits. It is 5
17773 * different series of 0-9. All other ranges of
17774 * digits currently in Unicode are just a single
17775 * series. (And mktables will notify us if a later
17776 * Unicode version breaks this.)
17778 * If the range being checked is at most 9 long,
17779 * and the digit values represented are in
17780 * numerical order, they are from the same series.
17782 if ( value - prevvalue > 9
17783 || ((( value - 0x1D7CE) % 10)
17784 <= (prevvalue - 0x1D7CE) % 10))
17786 goto warn_bad_digit_range;
17791 /* For all other ranges of digits in Unicode, the
17792 * algorithm is just to check if both end points
17793 * are in the same series, which is the same range.
17795 index_start = _invlist_search(
17796 PL_XPosix_ptrs[_CC_DIGIT],
17799 /* Warn if the range starts and ends with a digit,
17800 * and they are not in the same group of 10. */
17801 if ( index_start >= 0
17802 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17804 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17805 value)) != index_start
17806 && index_final >= 0
17807 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17809 warn_bad_digit_range:
17810 vWARN(RExC_parse, "Ranges of digits should be"
17811 " from the same group of"
17818 if ((! range || prevvalue == value) && non_portable_endpoint) {
17819 if (isPRINT_A(value)) {
17822 if (isBACKSLASHED_PUNCT(value)) {
17823 literal[d++] = '\\';
17825 literal[d++] = (char) value;
17826 literal[d++] = '\0';
17829 "\"%.*s\" is more clearly written simply as \"%s\"",
17830 (int) (RExC_parse - rangebegin),
17835 else if (isMNEMONIC_CNTRL(value)) {
17837 "\"%.*s\" is more clearly written simply as \"%s\"",
17838 (int) (RExC_parse - rangebegin),
17840 cntrl_to_mnemonic((U8) value)
17846 /* Deal with this element of the class */
17849 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17852 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
17853 * that don't require special handling, we can just add the range like
17854 * we do for ASCII platforms */
17855 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17856 || ! (prevvalue < 256
17858 || (! non_portable_endpoint
17859 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17860 || (isUPPER_A(prevvalue)
17861 && isUPPER_A(value)))))))
17863 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17867 /* Here, requires special handling. This can be because it is a
17868 * range whose code points are considered to be Unicode, and so
17869 * must be individually translated into native, or because its a
17870 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
17871 * EBCDIC, but we have defined them to include only the "expected"
17872 * upper or lower case ASCII alphabetics. Subranges above 255 are
17873 * the same in native and Unicode, so can be added as a range */
17874 U8 start = NATIVE_TO_LATIN1(prevvalue);
17876 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17877 for (j = start; j <= end; j++) {
17878 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17881 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17887 range = 0; /* this range (if it was one) is done now */
17888 } /* End of loop through all the text within the brackets */
17890 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17891 output_posix_warnings(pRExC_state, posix_warnings);
17894 /* If anything in the class expands to more than one character, we have to
17895 * deal with them by building up a substitute parse string, and recursively
17896 * calling reg() on it, instead of proceeding */
17897 if (multi_char_matches) {
17898 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17901 char *save_end = RExC_end;
17902 char *save_parse = RExC_parse;
17903 char *save_start = RExC_start;
17904 Size_t constructed_prefix_len = 0; /* This gives the length of the
17905 constructed portion of the
17906 substitute parse. */
17907 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17912 /* Only one level of recursion allowed */
17913 assert(RExC_copy_start_in_constructed == RExC_precomp);
17915 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17916 because too confusing */
17918 sv_catpvs(substitute_parse, "(?:");
17922 /* Look at the longest folds first */
17923 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17928 if (av_exists(multi_char_matches, cp_count)) {
17929 AV** this_array_ptr;
17932 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17934 while ((this_sequence = av_pop(*this_array_ptr)) !=
17937 if (! first_time) {
17938 sv_catpvs(substitute_parse, "|");
17940 first_time = FALSE;
17942 sv_catpv(substitute_parse, SvPVX(this_sequence));
17947 /* If the character class contains anything else besides these
17948 * multi-character folds, have to include it in recursive parsing */
17949 if (element_count) {
17950 sv_catpvs(substitute_parse, "|[");
17951 constructed_prefix_len = SvCUR(substitute_parse);
17952 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17954 /* Put in a closing ']' only if not going off the end, as otherwise
17955 * we are adding something that really isn't there */
17956 if (RExC_parse < RExC_end) {
17957 sv_catpvs(substitute_parse, "]");
17961 sv_catpvs(substitute_parse, ")");
17964 /* This is a way to get the parse to skip forward a whole named
17965 * sequence instead of matching the 2nd character when it fails the
17967 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17971 /* Set up the data structure so that any errors will be properly
17972 * reported. See the comments at the definition of
17973 * REPORT_LOCATION_ARGS for details */
17974 RExC_copy_start_in_input = (char *) orig_parse;
17975 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17976 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
17977 RExC_end = RExC_parse + len;
17978 RExC_in_multi_char_class = 1;
17980 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17982 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PARSE|NEED_UTF8);
17984 /* And restore so can parse the rest of the pattern */
17985 RExC_parse = save_parse;
17986 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
17987 RExC_end = save_end;
17988 RExC_in_multi_char_class = 0;
17989 SvREFCNT_dec_NN(multi_char_matches);
17993 /* If folding, we calculate all characters that could fold to or from the
17994 * ones already on the list */
17995 if (cp_foldable_list) {
17997 UV start, end; /* End points of code point ranges */
17999 SV* fold_intersection = NULL;
18002 /* Our calculated list will be for Unicode rules. For locale
18003 * matching, we have to keep a separate list that is consulted at
18004 * runtime only when the locale indicates Unicode rules (and we
18005 * don't include potential matches in the ASCII/Latin1 range, as
18006 * any code point could fold to any other, based on the run-time
18007 * locale). For non-locale, we just use the general list */
18009 use_list = &only_utf8_locale_list;
18012 use_list = &cp_list;
18015 /* Only the characters in this class that participate in folds need
18016 * be checked. Get the intersection of this class and all the
18017 * possible characters that are foldable. This can quickly narrow
18018 * down a large class */
18019 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18020 &fold_intersection);
18022 /* Now look at the foldable characters in this class individually */
18023 invlist_iterinit(fold_intersection);
18024 while (invlist_iternext(fold_intersection, &start, &end)) {
18028 /* Look at every character in the range */
18029 for (j = start; j <= end; j++) {
18030 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18033 Size_t folds_count;
18034 unsigned int first_fold;
18035 const unsigned int * remaining_folds;
18039 /* Under /l, we don't know what code points below 256
18040 * fold to, except we do know the MICRO SIGN folds to
18041 * an above-255 character if the locale is UTF-8, so we
18042 * add it to the special list (in *use_list) Otherwise
18043 * we know now what things can match, though some folds
18044 * are valid under /d only if the target is UTF-8.
18045 * Those go in a separate list */
18046 if ( IS_IN_SOME_FOLD_L1(j)
18047 && ! (LOC && j != MICRO_SIGN))
18050 /* ASCII is always matched; non-ASCII is matched
18051 * only under Unicode rules (which could happen
18052 * under /l if the locale is a UTF-8 one */
18053 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18054 *use_list = add_cp_to_invlist(*use_list,
18055 PL_fold_latin1[j]);
18057 else if (j != PL_fold_latin1[j]) {
18058 upper_latin1_only_utf8_matches
18059 = add_cp_to_invlist(
18060 upper_latin1_only_utf8_matches,
18061 PL_fold_latin1[j]);
18065 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18066 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18068 add_above_Latin1_folds(pRExC_state,
18075 /* Here is an above Latin1 character. We don't have the
18076 * rules hard-coded for it. First, get its fold. This is
18077 * the simple fold, as the multi-character folds have been
18078 * handled earlier and separated out */
18079 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18080 (ASCII_FOLD_RESTRICTED)
18081 ? FOLD_FLAGS_NOMIX_ASCII
18084 /* Single character fold of above Latin1. Add everything
18085 * in its fold closure to the list that this node should
18087 folds_count = _inverse_folds(folded, &first_fold,
18089 for (k = 0; k <= folds_count; k++) {
18090 UV c = (k == 0) /* First time through use itself */
18092 : (k == 1) /* 2nd time use, the first fold */
18095 /* Then the remaining ones */
18096 : remaining_folds[k-2];
18098 /* /aa doesn't allow folds between ASCII and non- */
18099 if (( ASCII_FOLD_RESTRICTED
18100 && (isASCII(c) != isASCII(j))))
18105 /* Folds under /l which cross the 255/256 boundary are
18106 * added to a separate list. (These are valid only
18107 * when the locale is UTF-8.) */
18108 if (c < 256 && LOC) {
18109 *use_list = add_cp_to_invlist(*use_list, c);
18113 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18115 cp_list = add_cp_to_invlist(cp_list, c);
18118 /* Similarly folds involving non-ascii Latin1
18119 * characters under /d are added to their list */
18120 upper_latin1_only_utf8_matches
18121 = add_cp_to_invlist(
18122 upper_latin1_only_utf8_matches,
18128 SvREFCNT_dec_NN(fold_intersection);
18131 /* Now that we have finished adding all the folds, there is no reason
18132 * to keep the foldable list separate */
18133 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18134 SvREFCNT_dec_NN(cp_foldable_list);
18137 /* And combine the result (if any) with any inversion lists from posix
18138 * classes. The lists are kept separate up to now because we don't want to
18139 * fold the classes */
18140 if (simple_posixes) { /* These are the classes known to be unaffected by
18143 _invlist_union(cp_list, simple_posixes, &cp_list);
18144 SvREFCNT_dec_NN(simple_posixes);
18147 cp_list = simple_posixes;
18150 if (posixes || nposixes) {
18151 if (! DEPENDS_SEMANTICS) {
18153 /* For everything but /d, we can just add the current 'posixes' and
18154 * 'nposixes' to the main list */
18157 _invlist_union(cp_list, posixes, &cp_list);
18158 SvREFCNT_dec_NN(posixes);
18166 _invlist_union(cp_list, nposixes, &cp_list);
18167 SvREFCNT_dec_NN(nposixes);
18170 cp_list = nposixes;
18175 /* Under /d, things like \w match upper Latin1 characters only if
18176 * the target string is in UTF-8. But things like \W match all the
18177 * upper Latin1 characters if the target string is not in UTF-8.
18179 * Handle the case with something like \W separately */
18181 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18183 /* A complemented posix class matches all upper Latin1
18184 * characters if not in UTF-8. And it matches just certain
18185 * ones when in UTF-8. That means those certain ones are
18186 * matched regardless, so can just be added to the
18187 * unconditional list */
18189 _invlist_union(cp_list, nposixes, &cp_list);
18190 SvREFCNT_dec_NN(nposixes);
18194 cp_list = nposixes;
18197 /* Likewise for 'posixes' */
18198 _invlist_union(posixes, cp_list, &cp_list);
18199 SvREFCNT_dec(posixes);
18201 /* Likewise for anything else in the range that matched only
18203 if (upper_latin1_only_utf8_matches) {
18204 _invlist_union(cp_list,
18205 upper_latin1_only_utf8_matches,
18207 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18208 upper_latin1_only_utf8_matches = NULL;
18211 /* If we don't match all the upper Latin1 characters regardless
18212 * of UTF-8ness, we have to set a flag to match the rest when
18214 _invlist_subtract(only_non_utf8_list, cp_list,
18215 &only_non_utf8_list);
18216 if (_invlist_len(only_non_utf8_list) != 0) {
18217 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18219 SvREFCNT_dec_NN(only_non_utf8_list);
18222 /* Here there were no complemented posix classes. That means
18223 * the upper Latin1 characters in 'posixes' match only when the
18224 * target string is in UTF-8. So we have to add them to the
18225 * list of those types of code points, while adding the
18226 * remainder to the unconditional list.
18228 * First calculate what they are */
18229 SV* nonascii_but_latin1_properties = NULL;
18230 _invlist_intersection(posixes, PL_UpperLatin1,
18231 &nonascii_but_latin1_properties);
18233 /* And add them to the final list of such characters. */
18234 _invlist_union(upper_latin1_only_utf8_matches,
18235 nonascii_but_latin1_properties,
18236 &upper_latin1_only_utf8_matches);
18238 /* Remove them from what now becomes the unconditional list */
18239 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18242 /* And add those unconditional ones to the final list */
18244 _invlist_union(cp_list, posixes, &cp_list);
18245 SvREFCNT_dec_NN(posixes);
18252 SvREFCNT_dec(nonascii_but_latin1_properties);
18254 /* Get rid of any characters from the conditional list that we
18255 * now know are matched unconditionally, which may make that
18257 _invlist_subtract(upper_latin1_only_utf8_matches,
18259 &upper_latin1_only_utf8_matches);
18260 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18261 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18262 upper_latin1_only_utf8_matches = NULL;
18268 /* And combine the result (if any) with any inversion list from properties.
18269 * The lists are kept separate up to now so that we can distinguish the two
18270 * in regards to matching above-Unicode. A run-time warning is generated
18271 * if a Unicode property is matched against a non-Unicode code point. But,
18272 * we allow user-defined properties to match anything, without any warning,
18273 * and we also suppress the warning if there is a portion of the character
18274 * class that isn't a Unicode property, and which matches above Unicode, \W
18275 * or [\x{110000}] for example.
18276 * (Note that in this case, unlike the Posix one above, there is no
18277 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18278 * forces Unicode semantics */
18282 /* If it matters to the final outcome, see if a non-property
18283 * component of the class matches above Unicode. If so, the
18284 * warning gets suppressed. This is true even if just a single
18285 * such code point is specified, as, though not strictly correct if
18286 * another such code point is matched against, the fact that they
18287 * are using above-Unicode code points indicates they should know
18288 * the issues involved */
18290 warn_super = ! (invert
18291 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18294 _invlist_union(properties, cp_list, &cp_list);
18295 SvREFCNT_dec_NN(properties);
18298 cp_list = properties;
18303 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18305 /* Because an ANYOF node is the only one that warns, this node
18306 * can't be optimized into something else */
18307 optimizable = FALSE;
18311 /* Here, we have calculated what code points should be in the character
18314 * Now we can see about various optimizations. Fold calculation (which we
18315 * did above) needs to take place before inversion. Otherwise /[^k]/i
18316 * would invert to include K, which under /i would match k, which it
18317 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18318 * folded until runtime */
18320 /* If we didn't do folding, it's because some information isn't available
18321 * until runtime; set the run-time fold flag for these We know to set the
18322 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18323 * at least one 0-255 range code point */
18326 /* Some things on the list might be unconditionally included because of
18327 * other components. Remove them, and clean up the list if it goes to
18329 if (only_utf8_locale_list && cp_list) {
18330 _invlist_subtract(only_utf8_locale_list, cp_list,
18331 &only_utf8_locale_list);
18333 if (_invlist_len(only_utf8_locale_list) == 0) {
18334 SvREFCNT_dec_NN(only_utf8_locale_list);
18335 only_utf8_locale_list = NULL;
18338 if ( only_utf8_locale_list
18339 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18340 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18342 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18345 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18347 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
18349 invlist_iterinit(cp_list);
18350 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
18351 anyof_flags |= ANYOFL_FOLD;
18352 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18354 invlist_iterfinish(cp_list);
18357 else if ( DEPENDS_SEMANTICS
18358 && ( upper_latin1_only_utf8_matches
18359 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18361 RExC_seen_d_op = TRUE;
18362 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
18365 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
18369 && ! has_runtime_dependency)
18371 _invlist_invert(cp_list);
18373 /* Clear the invert flag since have just done it here */
18378 *ret_invlist = cp_list;
18383 /* All possible optimizations below still have these characteristics.
18384 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
18386 *flagp |= HASWIDTH|SIMPLE;
18388 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
18389 RExC_contains_locale = 1;
18392 /* Some character classes are equivalent to other nodes. Such nodes take
18393 * up less room, and some nodes require fewer operations to execute, than
18394 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
18395 * improve efficiency. */
18398 PERL_UINT_FAST8_T i;
18399 Size_t partial_cp_count = 0;
18400 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
18401 UV end[MAX_FOLD_FROMS+1] = { 0 };
18403 if (cp_list) { /* Count the code points in enough ranges that we would
18404 see all the ones possible in any fold in this version
18407 invlist_iterinit(cp_list);
18408 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
18409 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
18412 partial_cp_count += end[i] - start[i] + 1;
18415 invlist_iterfinish(cp_list);
18418 /* If we know at compile time that this matches every possible code
18419 * point, any run-time dependencies don't matter */
18420 if (start[0] == 0 && end[0] == UV_MAX) {
18422 ret = reganode(pRExC_state, OPFAIL, 0);
18425 ret = reg_node(pRExC_state, SANY);
18431 /* Similarly, for /l posix classes, if both a class and its
18432 * complement match, any run-time dependencies don't matter */
18434 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
18437 if ( POSIXL_TEST(posixl, namedclass) /* class */
18438 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
18441 ret = reganode(pRExC_state, OPFAIL, 0);
18444 ret = reg_node(pRExC_state, SANY);
18451 /* For well-behaved locales, some classes are subsets of others,
18452 * so complementing the subset and including the non-complemented
18453 * superset should match everything, like [\D[:alnum:]], and
18454 * [[:^alpha:][:alnum:]], but some implementations of locales are
18455 * buggy, and khw thinks its a bad idea to have optimization change
18456 * behavior, even if it avoids an OS bug in a given case */
18458 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
18460 /* If is a single posix /l class, can optimize to just that op.
18461 * Such a node will not match anything in the Latin1 range, as that
18462 * is not determinable until runtime, but will match whatever the
18463 * class does outside that range. (Note that some classes won't
18464 * match anything outside the range, like [:ascii:]) */
18465 if ( isSINGLE_BIT_SET(posixl)
18466 && (partial_cp_count == 0 || start[0] > 255))
18469 SV * class_above_latin1 = NULL;
18470 bool already_inverted;
18471 bool are_equivalent;
18473 /* Compute which bit is set, which is the same thing as, e.g.,
18474 * ANYOF_CNTRL. From
18475 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
18477 static const int MultiplyDeBruijnBitPosition2[32] =
18479 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
18480 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
18483 namedclass = MultiplyDeBruijnBitPosition2[(posixl
18484 * 0x077CB531U) >> 27];
18485 classnum = namedclass_to_classnum(namedclass);
18487 /* The named classes are such that the inverted number is one
18488 * larger than the non-inverted one */
18489 already_inverted = namedclass
18490 - classnum_to_namedclass(classnum);
18492 /* Create an inversion list of the official property, inverted
18493 * if the constructed node list is inverted, and restricted to
18494 * only the above latin1 code points, which are the only ones
18495 * known at compile time */
18496 _invlist_intersection_maybe_complement_2nd(
18498 PL_XPosix_ptrs[classnum],
18500 &class_above_latin1);
18501 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
18503 SvREFCNT_dec_NN(class_above_latin1);
18505 if (are_equivalent) {
18507 /* Resolve the run-time inversion flag with this possibly
18508 * inverted class */
18509 invert = invert ^ already_inverted;
18511 ret = reg_node(pRExC_state,
18512 POSIXL + invert * (NPOSIXL - POSIXL));
18513 FLAGS(REGNODE_p(ret)) = classnum;
18519 /* khw can't think of any other possible transformation involving
18521 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
18525 if (! has_runtime_dependency) {
18527 /* If the list is empty, nothing matches. This happens, for
18528 * example, when a Unicode property that doesn't match anything is
18529 * the only element in the character class (perluniprops.pod notes
18530 * such properties). */
18531 if (partial_cp_count == 0) {
18533 ret = reg_node(pRExC_state, SANY);
18536 ret = reganode(pRExC_state, OPFAIL, 0);
18542 /* If matches everything but \n */
18543 if ( start[0] == 0 && end[0] == '\n' - 1
18544 && start[1] == '\n' + 1 && end[1] == UV_MAX)
18547 ret = reg_node(pRExC_state, REG_ANY);
18553 /* Next see if can optimize classes that contain just a few code points
18554 * into an EXACTish node. The reason to do this is to let the
18555 * optimizer join this node with adjacent EXACTish ones, and ANYOF
18556 * nodes require conversion to code point from UTF-8.
18558 * An EXACTFish node can be generated even if not under /i, and vice
18559 * versa. But care must be taken. An EXACTFish node has to be such
18560 * that it only matches precisely the code points in the class, but we
18561 * want to generate the least restrictive one that does that, to
18562 * increase the odds of being able to join with an adjacent node. For
18563 * example, if the class contains [kK], we have to make it an EXACTFAA
18564 * node to prevent the KELVIN SIGN from matching. Whether we are under
18565 * /i or not is irrelevant in this case. Less obvious is the pattern
18566 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
18567 * supposed to match the single character U+0149 LATIN SMALL LETTER N
18568 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
18569 * that includes \X{02BC}, there is a multi-char fold that does, and so
18570 * the node generated for it must be an EXACTFish one. On the other
18571 * hand qr/:/i should generate a plain EXACT node since the colon
18572 * participates in no fold whatsoever, and having it EXACT tells the
18573 * optimizer the target string cannot match unless it has a colon in
18576 * We don't typically generate an EXACTish node if doing so would
18577 * require changing the pattern to UTF-8, as that affects /d and
18578 * otherwise is slower. However, under /i, not changing to UTF-8 can
18579 * miss some potential multi-character folds. We calculate the
18580 * EXACTish node, and then decide if something would be missed if we
18585 /* Only try if there are no more code points in the class than
18586 * in the max possible fold */
18587 && partial_cp_count > 0 && partial_cp_count <= MAX_FOLD_FROMS + 1
18589 && (start[0] < 256 || UTF || FOLD))
18591 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
18593 /* We can always make a single code point class into an
18594 * EXACTish node. */
18598 /* Here is /l: Use EXACTL, except /li indicates EXACTFL,
18599 * as that means there is a fold not known until runtime so
18600 * shows as only a single code point here. */
18601 op = (FOLD) ? EXACTFL : EXACTL;
18603 else if (! FOLD) { /* Not /l and not /i */
18604 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
18606 else if (start[0] < 256) { /* /i, not /l, and the code point is
18609 /* Under /i, it gets a little tricky. A code point that
18610 * doesn't participate in a fold should be an EXACT node.
18611 * We know this one isn't the result of a simple fold, or
18612 * there'd be more than one code point in the list, but it
18613 * could be part of a multi- character fold. In that case
18614 * we better not create an EXACT node, as we would wrongly
18615 * be telling the optimizer that this code point must be in
18616 * the target string, and that is wrong. This is because
18617 * if the sequence around this code point forms a
18618 * multi-char fold, what needs to be in the string could be
18619 * the code point that folds to the sequence.
18621 * This handles the case of below-255 code points, as we
18622 * have an easy look up for those. The next clause handles
18623 * the above-256 one */
18624 op = IS_IN_SOME_FOLD_L1(start[0])
18628 else { /* /i, larger code point. Since we are under /i, and
18629 have just this code point, we know that it can't
18630 fold to something else, so PL_InMultiCharFold
18632 op = _invlist_contains_cp(PL_InMultiCharFold,
18640 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
18641 && _invlist_contains_cp(PL_in_some_fold, start[0]))
18643 /* Here, the only runtime dependency, if any, is from /d, and
18644 * the class matches more than one code point, and the lowest
18645 * code point participates in some fold. It might be that the
18646 * other code points are /i equivalent to this one, and hence
18647 * they would representable by an EXACTFish node. Above, we
18648 * eliminated classes that contain too many code points to be
18649 * EXACTFish, with the test for MAX_FOLD_FROMS
18651 * First, special case the ASCII fold pairs, like 'B' and 'b'.
18652 * We do this because we have EXACTFAA at our disposal for the
18654 if (partial_cp_count == 2 && isASCII(start[0])) {
18656 /* The only ASCII characters that participate in folds are
18658 assert(isALPHA(start[0]));
18659 if ( end[0] == start[0] /* First range is a single
18660 character, so 2nd exists */
18661 && isALPHA_FOLD_EQ(start[0], start[1]))
18664 /* Here, is part of an ASCII fold pair */
18666 if ( ASCII_FOLD_RESTRICTED
18667 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
18669 /* If the second clause just above was true, it
18670 * means we can't be under /i, or else the list
18671 * would have included more than this fold pair.
18672 * Therefore we have to exclude the possibility of
18673 * whatever else it is that folds to these, by
18674 * using EXACTFAA */
18677 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
18679 /* Here, there's no simple fold that start[0] is part
18680 * of, but there is a multi-character one. If we
18681 * are not under /i, we want to exclude that
18682 * possibility; if under /i, we want to include it
18684 op = (FOLD) ? EXACTFU : EXACTFAA;
18688 /* Here, the only possible fold start[0] particpates in
18689 * is with start[1]. /i or not isn't relevant */
18693 value = toFOLD(start[0]);
18696 else if ( ! upper_latin1_only_utf8_matches
18697 || ( _invlist_len(upper_latin1_only_utf8_matches)
18700 invlist_highest(upper_latin1_only_utf8_matches)]
18703 /* Here, the smallest character is non-ascii or there are
18704 * more than 2 code points matched by this node. Also, we
18705 * either don't have /d UTF-8 dependent matches, or if we
18706 * do, they look like they could be a single character that
18707 * is the fold of the lowest one in the always-match list.
18708 * This test quickly excludes most of the false positives
18709 * when there are /d UTF-8 depdendent matches. These are
18710 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
18711 * SMALL LETTER A WITH GRAVE iff the target string is
18712 * UTF-8. (We don't have to worry above about exceeding
18713 * the array bounds of PL_fold_latin1[] because any code
18714 * point in 'upper_latin1_only_utf8_matches' is below 256.)
18716 * EXACTFAA would apply only to pairs (hence exactly 2 code
18717 * points) in the ASCII range, so we can't use it here to
18718 * artificially restrict the fold domain, so we check if
18719 * the class does or does not match some EXACTFish node.
18720 * Further, if we aren't under /i, and and the folded-to
18721 * character is part of a multi-character fold, we can't do
18722 * this optimization, as the sequence around it could be
18723 * that multi-character fold, and we don't here know the
18724 * context, so we have to assume it is that multi-char
18725 * fold, to prevent potential bugs.
18727 * To do the general case, we first find the fold of the
18728 * lowest code point (which may be higher than the lowest
18729 * one), then find everything that folds to it. (The data
18730 * structure we have only maps from the folded code points,
18731 * so we have to do the earlier step.) */
18734 U8 foldbuf[UTF8_MAXBYTES_CASE];
18735 UV folded = _to_uni_fold_flags(start[0],
18736 foldbuf, &foldlen, 0);
18737 unsigned int first_fold;
18738 const unsigned int * remaining_folds;
18739 Size_t folds_to_this_cp_count = _inverse_folds(
18743 Size_t folds_count = folds_to_this_cp_count + 1;
18744 SV * fold_list = _new_invlist(folds_count);
18747 /* If there are UTF-8 dependent matches, create a temporary
18748 * list of what this node matches, including them. */
18749 SV * all_cp_list = NULL;
18750 SV ** use_this_list = &cp_list;
18752 if (upper_latin1_only_utf8_matches) {
18753 all_cp_list = _new_invlist(0);
18754 use_this_list = &all_cp_list;
18755 _invlist_union(cp_list,
18756 upper_latin1_only_utf8_matches,
18760 /* Having gotten everything that participates in the fold
18761 * containing the lowest code point, we turn that into an
18762 * inversion list, making sure everything is included. */
18763 fold_list = add_cp_to_invlist(fold_list, start[0]);
18764 fold_list = add_cp_to_invlist(fold_list, folded);
18765 if (folds_to_this_cp_count > 0) {
18766 fold_list = add_cp_to_invlist(fold_list, first_fold);
18767 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
18768 fold_list = add_cp_to_invlist(fold_list,
18769 remaining_folds[i]);
18773 /* If the fold list is identical to what's in this ANYOF
18774 * node, the node can be represented by an EXACTFish one
18776 if (_invlistEQ(*use_this_list, fold_list,
18777 0 /* Don't complement */ )
18780 /* But, we have to be careful, as mentioned above.
18781 * Just the right sequence of characters could match
18782 * this if it is part of a multi-character fold. That
18783 * IS what we want if we are under /i. But it ISN'T
18784 * what we want if not under /i, as it could match when
18785 * it shouldn't. So, when we aren't under /i and this
18786 * character participates in a multi-char fold, we
18787 * don't optimize into an EXACTFish node. So, for each
18788 * case below we have to check if we are folding
18789 * and if not, if it is not part of a multi-char fold.
18791 if (start[0] > 255) { /* Highish code point */
18792 if (FOLD || ! _invlist_contains_cp(
18793 PL_InMultiCharFold, folded))
18797 : (ASCII_FOLD_RESTRICTED)
18802 } /* Below, the lowest code point < 256 */
18805 && DEPENDS_SEMANTICS)
18806 { /* An EXACTF node containing a single character
18807 's', can be an EXACTFU if it doesn't get
18808 joined with an adjacent 's' */
18809 op = EXACTFU_S_EDGE;
18813 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
18815 if (upper_latin1_only_utf8_matches) {
18818 /* We can't use the fold, as that only matches
18822 else if ( UNLIKELY(start[0] == MICRO_SIGN)
18824 { /* EXACTFUP is a special node for this
18826 op = (ASCII_FOLD_RESTRICTED)
18829 value = MICRO_SIGN;
18831 else if ( ASCII_FOLD_RESTRICTED
18832 && ! isASCII(start[0]))
18833 { /* For ASCII under /iaa, we can use EXACTFU
18845 SvREFCNT_dec_NN(fold_list);
18846 SvREFCNT_dec(all_cp_list);
18852 /* Here, we have calculated what EXACTish node we would use.
18853 * But we don't use it if it would require converting the
18854 * pattern to UTF-8, unless not using it could cause us to miss
18855 * some folds (hence be buggy) */
18857 if (! UTF && value > 255) {
18858 SV * in_multis = NULL;
18862 /* If there is no code point that is part of a multi-char
18863 * fold, then there aren't any matches, so we don't do this
18864 * optimization. Otherwise, it could match depending on
18865 * the context around us, so we do upgrade */
18866 _invlist_intersection(PL_InMultiCharFold, cp_list, &in_multis);
18867 if (UNLIKELY(_invlist_len(in_multis) != 0)) {
18868 REQUIRE_UTF8(flagp);
18876 U8 len = (UTF) ? UVCHR_SKIP(value) : 1;
18878 ret = regnode_guts(pRExC_state, op, len, "exact");
18879 FILL_NODE(ret, op);
18880 RExC_emit += 1 + STR_SZ(len);
18881 setSTR_LEN(REGNODE_p(ret), len);
18883 *STRING(REGNODE_p(ret)) = (U8) value;
18886 uvchr_to_utf8((U8 *) STRING(REGNODE_p(ret)), value);
18893 if (! has_runtime_dependency) {
18895 /* See if this can be turned into an ANYOFM node. Think about the
18896 * bit patterns in two different bytes. In some positions, the
18897 * bits in each will be 1; and in other positions both will be 0;
18898 * and in some positions the bit will be 1 in one byte, and 0 in
18899 * the other. Let 'n' be the number of positions where the bits
18900 * differ. We create a mask which has exactly 'n' 0 bits, each in
18901 * a position where the two bytes differ. Now take the set of all
18902 * bytes that when ANDed with the mask yield the same result. That
18903 * set has 2**n elements, and is representable by just two 8 bit
18904 * numbers: the result and the mask. Importantly, matching the set
18905 * can be vectorized by creating a word full of the result bytes,
18906 * and a word full of the mask bytes, yielding a significant speed
18907 * up. Here, see if this node matches such a set. As a concrete
18908 * example consider [01], and the byte representing '0' which is
18909 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
18910 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
18911 * 0x30. Any other bytes ANDed yield something else. So [01],
18912 * which is a common usage, is optimizable into ANYOFM, and can
18913 * benefit from the speed up. We can only do this on UTF-8
18914 * invariant bytes, because they have the same bit patterns under
18916 PERL_UINT_FAST8_T inverted = 0;
18918 const PERL_UINT_FAST8_T max_permissible = 0xFF;
18920 const PERL_UINT_FAST8_T max_permissible = 0x7F;
18922 /* If doesn't fit the criteria for ANYOFM, invert and try again.
18923 * If that works we will instead later generate an NANYOFM, and
18924 * invert back when through */
18925 if (invlist_highest(cp_list) > max_permissible) {
18926 _invlist_invert(cp_list);
18930 if (invlist_highest(cp_list) <= max_permissible) {
18931 UV this_start, this_end;
18932 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
18933 U8 bits_differing = 0;
18934 Size_t full_cp_count = 0;
18935 bool first_time = TRUE;
18937 /* Go through the bytes and find the bit positions that differ
18939 invlist_iterinit(cp_list);
18940 while (invlist_iternext(cp_list, &this_start, &this_end)) {
18941 unsigned int i = this_start;
18944 if (! UVCHR_IS_INVARIANT(i)) {
18948 first_time = FALSE;
18949 lowest_cp = this_start;
18951 /* We have set up the code point to compare with.
18952 * Don't compare it with itself */
18956 /* Find the bit positions that differ from the lowest code
18957 * point in the node. Keep track of all such positions by
18959 for (; i <= this_end; i++) {
18960 if (! UVCHR_IS_INVARIANT(i)) {
18964 bits_differing |= i ^ lowest_cp;
18967 full_cp_count += this_end - this_start + 1;
18970 /* At the end of the loop, we count how many bits differ from
18971 * the bits in lowest code point, call the count 'd'. If the
18972 * set we found contains 2**d elements, it is the closure of
18973 * all code points that differ only in those bit positions. To
18974 * convince yourself of that, first note that the number in the
18975 * closure must be a power of 2, which we test for. The only
18976 * way we could have that count and it be some differing set,
18977 * is if we got some code points that don't differ from the
18978 * lowest code point in any position, but do differ from each
18979 * other in some other position. That means one code point has
18980 * a 1 in that position, and another has a 0. But that would
18981 * mean that one of them differs from the lowest code point in
18982 * that position, which possibility we've already excluded. */
18983 if ( (inverted || full_cp_count > 1)
18984 && full_cp_count == 1U << PL_bitcount[bits_differing])
18988 op = ANYOFM + inverted;;
18990 /* We need to make the bits that differ be 0's */
18991 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
18993 /* The argument is the lowest code point */
18994 ret = reganode(pRExC_state, op, lowest_cp);
18995 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
18999 invlist_iterfinish(cp_list);
19003 _invlist_invert(cp_list);
19010 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19011 * all were invariants, it wasn't inverted, and there is a single
19012 * range. This would be faster than some of the posix nodes we
19013 * create below like /\d/a, but would be twice the size. Without
19014 * having actually measured the gain, khw doesn't think the
19015 * tradeoff is really worth it */
19018 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
19019 PERL_UINT_FAST8_T type;
19020 SV * intersection = NULL;
19021 SV* d_invlist = NULL;
19023 /* See if this matches any of the POSIX classes. The POSIXA and
19024 * POSIXD ones are about the same speed as ANYOF ops, but take less
19025 * room; the ones that have above-Latin1 code point matches are
19026 * somewhat faster than ANYOF. */
19028 for (type = POSIXA; type >= POSIXD; type--) {
19031 if (type == POSIXL) { /* But not /l posix classes */
19035 for (posix_class = 0;
19036 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19039 SV** our_code_points = &cp_list;
19040 SV** official_code_points;
19043 if (type == POSIXA) {
19044 official_code_points = &PL_Posix_ptrs[posix_class];
19047 official_code_points = &PL_XPosix_ptrs[posix_class];
19050 /* Skip non-existent classes of this type. e.g. \v only
19051 * has an entry in PL_XPosix_ptrs */
19052 if (! *official_code_points) {
19056 /* Try both the regular class, and its inversion */
19057 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19058 bool this_inverted = invert ^ try_inverted;
19060 if (type != POSIXD) {
19062 /* This class that isn't /d can't match if we have
19063 * /d dependencies */
19064 if (has_runtime_dependency
19065 & HAS_D_RUNTIME_DEPENDENCY)
19070 else /* is /d */ if (! this_inverted) {
19072 /* /d classes don't match anything non-ASCII below
19073 * 256 unconditionally (which cp_list contains) */
19074 _invlist_intersection(cp_list, PL_UpperLatin1,
19076 if (_invlist_len(intersection) != 0) {
19080 SvREFCNT_dec(d_invlist);
19081 d_invlist = invlist_clone(cp_list, NULL);
19083 /* But under UTF-8 it turns into using /u rules.
19084 * Add the things it matches under these conditions
19085 * so that we check below that these are identical
19086 * to what the tested class should match */
19087 if (upper_latin1_only_utf8_matches) {
19090 upper_latin1_only_utf8_matches,
19093 our_code_points = &d_invlist;
19095 else { /* POSIXD, inverted. If this doesn't have this
19096 flag set, it isn't /d. */
19097 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
19101 our_code_points = &cp_list;
19104 /* Here, have weeded out some things. We want to see
19105 * if the list of characters this node contains
19106 * ('*our_code_points') precisely matches those of the
19107 * class we are currently checking against
19108 * ('*official_code_points'). */
19109 if (_invlistEQ(*our_code_points,
19110 *official_code_points,
19113 /* Here, they precisely match. Optimize this ANYOF
19114 * node into its equivalent POSIX one of the
19115 * correct type, possibly inverted */
19116 ret = reg_node(pRExC_state, (try_inverted)
19120 FLAGS(REGNODE_p(ret)) = posix_class;
19121 SvREFCNT_dec(d_invlist);
19122 SvREFCNT_dec(intersection);
19128 SvREFCNT_dec(d_invlist);
19129 SvREFCNT_dec(intersection);
19132 /* If didn't find an optimization and there is no need for a bitmap,
19133 * optimize to indicate that */
19134 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19136 && ! upper_latin1_only_utf8_matches
19137 && anyof_flags == 0)
19139 U8 low_utf8[UTF8_MAXBYTES+1];
19140 UV highest_cp = invlist_highest(cp_list);
19144 /* Currently the maximum allowed code point by the system is
19145 * IV_MAX. Higher ones are reserved for future internal use. This
19146 * particular regnode can be used for higher ones, but we can't
19147 * calculate the code point of those. IV_MAX suffices though, as
19148 * it will be a large first byte */
19149 (void) uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX));
19151 /* We store the lowest possible first byte of the UTF-8
19152 * representation, using the flags field. This allows for quick
19153 * ruling out of some inputs without having to convert from UTF-8
19154 * to code point. For EBCDIC, we use I8, as not doing that
19155 * transformation would not rule out nearly so many things */
19156 anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
19158 /* If the first UTF-8 start byte for the highest code point in the
19159 * range is suitably small, we may be able to get an upper bound as
19161 if (highest_cp <= IV_MAX) {
19162 U8 high_utf8[UTF8_MAXBYTES+1];
19164 (void) uvchr_to_utf8(high_utf8, highest_cp);
19166 /* If the lowest and highest are the same, we can get an exact
19167 * first byte instead of a just minimum. We signal this with a
19168 * different regnode */
19169 if (low_utf8[0] == high_utf8[0]) {
19171 /* No need to convert to I8 for EBCDIC as this is an exact
19173 anyof_flags = low_utf8[0];
19176 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
19179 /* Here, the high byte is not the same as the low, but is
19180 * small enough that its reasonable to have a loose upper
19181 * bound, which is packed in with the strict lower bound.
19182 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
19183 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
19184 * is the same thing as UTF-8 */
19187 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - anyof_flags;
19188 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
19191 if (range_diff <= max_range_diff / 8) {
19194 else if (range_diff <= max_range_diff / 4) {
19197 else if (range_diff <= max_range_diff / 2) {
19200 anyof_flags = (anyof_flags - 0xC0) << 2 | bits;
19205 goto done_finding_op;
19207 } /* End of seeing if can optimize it into a different node */
19209 is_anyof: /* It's going to be an ANYOF node. */
19210 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19220 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19221 FILL_NODE(ret, op); /* We set the argument later */
19222 RExC_emit += 1 + regarglen[op];
19223 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19225 /* Here, <cp_list> contains all the code points we can determine at
19226 * compile time that match under all conditions. Go through it, and
19227 * for things that belong in the bitmap, put them there, and delete from
19228 * <cp_list>. While we are at it, see if everything above 255 is in the
19229 * list, and if so, set a flag to speed up execution */
19231 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19234 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19238 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19241 /* Here, the bitmap has been populated with all the Latin1 code points that
19242 * always match. Can now add to the overall list those that match only
19243 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19245 if (upper_latin1_only_utf8_matches) {
19247 _invlist_union(cp_list,
19248 upper_latin1_only_utf8_matches,
19250 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19253 cp_list = upper_latin1_only_utf8_matches;
19255 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19258 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19259 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19261 only_utf8_locale_list);
19266 /* Here, the node is getting optimized into something that's not an ANYOF
19267 * one. Finish up. */
19269 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19270 RExC_parse - orig_parse);;
19271 SvREFCNT_dec(cp_list);;
19275 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19278 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19279 regnode* const node,
19281 SV* const runtime_defns,
19282 SV* const only_utf8_locale_list)
19284 /* Sets the arg field of an ANYOF-type node 'node', using information about
19285 * the node passed-in. If there is nothing outside the node's bitmap, the
19286 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
19287 * the count returned by add_data(), having allocated and stored an array,
19290 * av[0] stores the inversion list defining this class as far as known at
19291 * this time, or PL_sv_undef if nothing definite is now known.
19292 * av[1] stores the inversion list of code points that match only if the
19293 * current locale is UTF-8, or if none, PL_sv_undef if there is an
19294 * av[2], or no entry otherwise.
19295 * av[2] stores the list of user-defined properties whose subroutine
19296 * definitions aren't known at this time, or no entry if none. */
19300 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
19302 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
19303 assert(! (ANYOF_FLAGS(node)
19304 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
19305 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
19308 AV * const av = newAV();
19312 av_store(av, INVLIST_INDEX, cp_list);
19315 if (only_utf8_locale_list) {
19316 av_store(av, ONLY_LOCALE_MATCHES_INDEX, only_utf8_locale_list);
19319 if (runtime_defns) {
19320 av_store(av, DEFERRED_USER_DEFINED_INDEX, SvREFCNT_inc(runtime_defns));
19323 rv = newRV_noinc(MUTABLE_SV(av));
19324 n = add_data(pRExC_state, STR_WITH_LEN("s"));
19325 RExC_rxi->data->data[n] = (void*)rv;
19330 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
19332 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
19333 const regnode* node,
19336 SV** only_utf8_locale_ptr,
19337 SV** output_invlist)
19340 /* For internal core use only.
19341 * Returns the inversion list for the input 'node' in the regex 'prog'.
19342 * If <doinit> is 'true', will attempt to create the inversion list if not
19344 * If <listsvp> is non-null, will return the printable contents of the
19345 * property definition. This can be used to get debugging information
19346 * even before the inversion list exists, by calling this function with
19347 * 'doinit' set to false, in which case the components that will be used
19348 * to eventually create the inversion list are returned (in a printable
19350 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
19351 * store an inversion list of code points that should match only if the
19352 * execution-time locale is a UTF-8 one.
19353 * If <output_invlist> is not NULL, it is where this routine is to store an
19354 * inversion list of the code points that would be instead returned in
19355 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
19356 * when this parameter is used, is just the non-code point data that
19357 * will go into creating the inversion list. This currently should be just
19358 * user-defined properties whose definitions were not known at compile
19359 * time. Using this parameter allows for easier manipulation of the
19360 * inversion list's data by the caller. It is illegal to call this
19361 * function with this parameter set, but not <listsvp>
19363 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
19364 * that, in spite of this function's name, the inversion list it returns
19365 * may include the bitmap data as well */
19367 SV *si = NULL; /* Input initialization string */
19368 SV* invlist = NULL;
19370 RXi_GET_DECL(prog, progi);
19371 const struct reg_data * const data = prog ? progi->data : NULL;
19373 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
19374 assert(! output_invlist || listsvp);
19376 if (data && data->count) {
19377 const U32 n = ARG(node);
19379 if (data->what[n] == 's') {
19380 SV * const rv = MUTABLE_SV(data->data[n]);
19381 AV * const av = MUTABLE_AV(SvRV(rv));
19382 SV **const ary = AvARRAY(av);
19384 invlist = ary[INVLIST_INDEX];
19386 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
19387 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
19390 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
19391 si = ary[DEFERRED_USER_DEFINED_INDEX];
19394 if (doinit && (si || invlist)) {
19397 SV * msg = newSVpvs_flags("", SVs_TEMP);
19399 SV * prop_definition = handle_user_defined_property(
19400 "", 0, FALSE, /* There is no \p{}, \P{} */
19401 SvPVX_const(si)[1] - '0', /* /i or not has been
19402 stored here for just
19404 TRUE, /* run time */
19405 FALSE, /* This call must find the defn */
19406 si, /* The property definition */
19409 0 /* base level call */
19413 assert(prop_definition == NULL);
19415 Perl_croak(aTHX_ "%" UTF8f,
19416 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
19420 _invlist_union(invlist, prop_definition, &invlist);
19421 SvREFCNT_dec_NN(prop_definition);
19424 invlist = prop_definition;
19427 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
19428 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
19430 av_store(av, INVLIST_INDEX, invlist);
19431 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
19432 ? ONLY_LOCALE_MATCHES_INDEX:
19440 /* If requested, return a printable version of what this ANYOF node matches
19443 SV* matches_string = NULL;
19445 /* This function can be called at compile-time, before everything gets
19446 * resolved, in which case we return the currently best available
19447 * information, which is the string that will eventually be used to do
19448 * that resolving, 'si' */
19450 /* Here, we only have 'si' (and possibly some passed-in data in
19451 * 'invlist', which is handled below) If the caller only wants
19452 * 'si', use that. */
19453 if (! output_invlist) {
19454 matches_string = newSVsv(si);
19457 /* But if the caller wants an inversion list of the node, we
19458 * need to parse 'si' and place as much as possible in the
19459 * desired output inversion list, making 'matches_string' only
19460 * contain the currently unresolvable things */
19461 const char *si_string = SvPVX(si);
19462 STRLEN remaining = SvCUR(si);
19466 /* Ignore everything before the first new-line */
19467 while (*si_string != '\n' && remaining > 0) {
19471 assert(remaining > 0);
19476 while (remaining > 0) {
19478 /* The data consists of just strings defining user-defined
19479 * property names, but in prior incarnations, and perhaps
19480 * somehow from pluggable regex engines, it could still
19481 * hold hex code point definitions. Each component of a
19482 * range would be separated by a tab, and each range by a
19483 * new-line. If these are found, instead add them to the
19484 * inversion list */
19485 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
19486 |PERL_SCAN_SILENT_NON_PORTABLE;
19487 STRLEN len = remaining;
19488 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
19490 /* If the hex decode routine found something, it should go
19491 * up to the next \n */
19492 if ( *(si_string + len) == '\n') {
19493 if (count) { /* 2nd code point on line */
19494 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
19497 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
19500 goto prepare_for_next_iteration;
19503 /* If the hex decode was instead for the lower range limit,
19504 * save it, and go parse the upper range limit */
19505 if (*(si_string + len) == '\t') {
19506 assert(count == 0);
19510 prepare_for_next_iteration:
19511 si_string += len + 1;
19512 remaining -= len + 1;
19516 /* Here, didn't find a legal hex number. Just add it from
19517 * here to the next \n */
19520 while (*(si_string + len) != '\n' && remaining > 0) {
19524 if (*(si_string + len) == '\n') {
19528 if (matches_string) {
19529 sv_catpvn(matches_string, si_string, len - 1);
19532 matches_string = newSVpvn(si_string, len - 1);
19535 sv_catpvs(matches_string, " ");
19536 } /* end of loop through the text */
19538 assert(matches_string);
19539 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
19540 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
19542 } /* end of has an 'si' */
19545 /* Add the stuff that's already known */
19548 /* Again, if the caller doesn't want the output inversion list, put
19549 * everything in 'matches-string' */
19550 if (! output_invlist) {
19551 if ( ! matches_string) {
19552 matches_string = newSVpvs("\n");
19554 sv_catsv(matches_string, invlist_contents(invlist,
19555 TRUE /* traditional style */
19558 else if (! *output_invlist) {
19559 *output_invlist = invlist_clone(invlist, NULL);
19562 _invlist_union(*output_invlist, invlist, output_invlist);
19566 *listsvp = matches_string;
19571 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
19573 /* reg_skipcomment()
19575 Absorbs an /x style # comment from the input stream,
19576 returning a pointer to the first character beyond the comment, or if the
19577 comment terminates the pattern without anything following it, this returns
19578 one past the final character of the pattern (in other words, RExC_end) and
19579 sets the REG_RUN_ON_COMMENT_SEEN flag.
19581 Note it's the callers responsibility to ensure that we are
19582 actually in /x mode
19586 PERL_STATIC_INLINE char*
19587 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
19589 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
19593 while (p < RExC_end) {
19594 if (*(++p) == '\n') {
19599 /* we ran off the end of the pattern without ending the comment, so we have
19600 * to add an \n when wrapping */
19601 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
19606 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
19608 const bool force_to_xmod
19611 /* If the text at the current parse position '*p' is a '(?#...)' comment,
19612 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
19613 * is /x whitespace, advance '*p' so that on exit it points to the first
19614 * byte past all such white space and comments */
19616 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
19618 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
19620 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
19623 if (RExC_end - (*p) >= 3
19625 && *(*p + 1) == '?'
19626 && *(*p + 2) == '#')
19628 while (*(*p) != ')') {
19629 if ((*p) == RExC_end)
19630 FAIL("Sequence (?#... not terminated");
19638 const char * save_p = *p;
19639 while ((*p) < RExC_end) {
19641 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
19644 else if (*(*p) == '#') {
19645 (*p) = reg_skipcomment(pRExC_state, (*p));
19651 if (*p != save_p) {
19664 Advances the parse position by one byte, unless that byte is the beginning
19665 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
19666 those two cases, the parse position is advanced beyond all such comments and
19669 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
19673 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
19675 PERL_ARGS_ASSERT_NEXTCHAR;
19677 if (RExC_parse < RExC_end) {
19679 || UTF8_IS_INVARIANT(*RExC_parse)
19680 || UTF8_IS_START(*RExC_parse));
19682 RExC_parse += (UTF)
19683 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
19686 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
19687 FALSE /* Don't force /x */ );
19692 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
19694 /* 'size' is the delta number of smallest regnode equivalents to add or
19695 * subtract from the current memory allocated to the regex engine being
19698 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
19703 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
19704 /* +1 for REG_MAGIC */
19707 if ( RExC_rxi == NULL )
19708 FAIL("Regexp out of space");
19709 RXi_SET(RExC_rx, RExC_rxi);
19711 RExC_emit_start = RExC_rxi->program;
19713 Zero(REGNODE_p(RExC_emit), size, regnode);
19716 #ifdef RE_TRACK_PATTERN_OFFSETS
19717 Renew(RExC_offsets, 2*RExC_size+1, U32);
19719 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
19721 RExC_offsets[0] = RExC_size;
19725 STATIC regnode_offset
19726 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
19728 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
19729 * equivalents space. It aligns and increments RExC_size and RExC_emit
19731 * It returns the regnode's offset into the regex engine program */
19733 const regnode_offset ret = RExC_emit;
19735 GET_RE_DEBUG_FLAGS_DECL;
19737 PERL_ARGS_ASSERT_REGNODE_GUTS;
19739 SIZE_ALIGN(RExC_size);
19740 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
19741 NODE_ALIGN_FILL(REGNODE_p(ret));
19742 #ifndef RE_TRACK_PATTERN_OFFSETS
19743 PERL_UNUSED_ARG(name);
19744 PERL_UNUSED_ARG(op);
19746 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
19748 if (RExC_offsets) { /* MJD */
19750 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
19753 (UV)(RExC_emit) > RExC_offsets[0]
19754 ? "Overwriting end of array!\n" : "OK",
19756 (UV)(RExC_parse - RExC_start),
19757 (UV)RExC_offsets[0]));
19758 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
19765 - reg_node - emit a node
19767 STATIC regnode_offset /* Location. */
19768 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
19770 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
19771 regnode_offset ptr = ret;
19773 PERL_ARGS_ASSERT_REG_NODE;
19775 assert(regarglen[op] == 0);
19777 FILL_ADVANCE_NODE(ptr, op);
19783 - reganode - emit a node with an argument
19785 STATIC regnode_offset /* Location. */
19786 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
19788 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
19789 regnode_offset ptr = ret;
19791 PERL_ARGS_ASSERT_REGANODE;
19793 /* ANYOF are special cased to allow non-length 1 args */
19794 assert(regarglen[op] == 1);
19796 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
19801 STATIC regnode_offset
19802 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
19804 /* emit a node with U32 and I32 arguments */
19806 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
19807 regnode_offset ptr = ret;
19809 PERL_ARGS_ASSERT_REG2LANODE;
19811 assert(regarglen[op] == 2);
19813 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
19819 - reginsert - insert an operator in front of already-emitted operand
19821 * That means that on exit 'operand' is the offset of the newly inserted
19822 * operator, and the original operand has been relocated.
19824 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
19825 * set up NEXT_OFF() of the inserted node if needed. Something like this:
19827 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
19828 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
19830 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
19833 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
19834 const regnode_offset operand, const U32 depth)
19839 const int offset = regarglen[(U8)op];
19840 const int size = NODE_STEP_REGNODE + offset;
19841 GET_RE_DEBUG_FLAGS_DECL;
19843 PERL_ARGS_ASSERT_REGINSERT;
19844 PERL_UNUSED_CONTEXT;
19845 PERL_UNUSED_ARG(depth);
19846 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
19847 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
19848 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
19849 studying. If this is wrong then we need to adjust RExC_recurse
19850 below like we do with RExC_open_parens/RExC_close_parens. */
19851 change_engine_size(pRExC_state, (Ptrdiff_t) size);
19852 src = REGNODE_p(RExC_emit);
19854 dst = REGNODE_p(RExC_emit);
19856 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
19857 * and [perl #133871] shows this can lead to problems, so skip this
19858 * realignment of parens until a later pass when they are reliable */
19859 if (! IN_PARENS_PASS && RExC_open_parens) {
19861 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
19862 /* remember that RExC_npar is rex->nparens + 1,
19863 * iow it is 1 more than the number of parens seen in
19864 * the pattern so far. */
19865 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
19866 /* note, RExC_open_parens[0] is the start of the
19867 * regex, it can't move. RExC_close_parens[0] is the end
19868 * of the regex, it *can* move. */
19869 if ( paren && RExC_open_parens[paren] >= operand ) {
19870 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
19871 RExC_open_parens[paren] += size;
19873 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
19875 if ( RExC_close_parens[paren] >= operand ) {
19876 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
19877 RExC_close_parens[paren] += size;
19879 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
19884 RExC_end_op += size;
19886 while (src > REGNODE_p(operand)) {
19887 StructCopy(--src, --dst, regnode);
19888 #ifdef RE_TRACK_PATTERN_OFFSETS
19889 if (RExC_offsets) { /* MJD 20010112 */
19891 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
19895 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
19896 ? "Overwriting end of array!\n" : "OK",
19897 (UV)REGNODE_OFFSET(src),
19898 (UV)REGNODE_OFFSET(dst),
19899 (UV)RExC_offsets[0]));
19900 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
19901 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
19906 place = REGNODE_p(operand); /* Op node, where operand used to be. */
19907 #ifdef RE_TRACK_PATTERN_OFFSETS
19908 if (RExC_offsets) { /* MJD */
19910 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
19914 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
19915 ? "Overwriting end of array!\n" : "OK",
19916 (UV)REGNODE_OFFSET(place),
19917 (UV)(RExC_parse - RExC_start),
19918 (UV)RExC_offsets[0]));
19919 Set_Node_Offset(place, RExC_parse);
19920 Set_Node_Length(place, 1);
19923 src = NEXTOPER(place);
19925 FILL_NODE(operand, op);
19927 /* Zero out any arguments in the new node */
19928 Zero(src, offset, regnode);
19932 - regtail - set the next-pointer at the end of a node chain of p to val. If
19933 that value won't fit in the space available, instead returns FALSE.
19934 (Except asserts if we can't fit in the largest space the regex
19935 engine is designed for.)
19936 - SEE ALSO: regtail_study
19939 S_regtail(pTHX_ RExC_state_t * pRExC_state,
19940 const regnode_offset p,
19941 const regnode_offset val,
19944 regnode_offset scan;
19945 GET_RE_DEBUG_FLAGS_DECL;
19947 PERL_ARGS_ASSERT_REGTAIL;
19949 PERL_UNUSED_ARG(depth);
19952 /* Find last node. */
19953 scan = (regnode_offset) p;
19955 regnode * const temp = regnext(REGNODE_p(scan));
19957 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
19958 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
19959 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
19960 SvPV_nolen_const(RExC_mysv), scan,
19961 (temp == NULL ? "->" : ""),
19962 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
19967 scan = REGNODE_OFFSET(temp);
19970 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
19971 assert((UV) (val - scan) <= U32_MAX);
19972 ARG_SET(REGNODE_p(scan), val - scan);
19975 if (val - scan > U16_MAX) {
19976 /* Populate this with something that won't loop and will likely
19977 * lead to a crash if the caller ignores the failure return, and
19978 * execution continues */
19979 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
19982 NEXT_OFF(REGNODE_p(scan)) = val - scan;
19990 - regtail_study - set the next-pointer at the end of a node chain of p to val.
19991 - Look for optimizable sequences at the same time.
19992 - currently only looks for EXACT chains.
19994 This is experimental code. The idea is to use this routine to perform
19995 in place optimizations on branches and groups as they are constructed,
19996 with the long term intention of removing optimization from study_chunk so
19997 that it is purely analytical.
19999 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20000 to control which is which.
20002 This used to return a value that was ignored. It was a problem that it is
20003 #ifdef'd to be another function that didn't return a value. khw has changed it
20004 so both currently return a pass/fail return.
20007 /* TODO: All four parms should be const */
20010 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20011 const regnode_offset val, U32 depth)
20013 regnode_offset scan;
20015 #ifdef EXPERIMENTAL_INPLACESCAN
20018 GET_RE_DEBUG_FLAGS_DECL;
20020 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20023 /* Find last node. */
20027 regnode * const temp = regnext(REGNODE_p(scan));
20028 #ifdef EXPERIMENTAL_INPLACESCAN
20029 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20030 bool unfolded_multi_char; /* Unexamined in this routine */
20031 if (join_exact(pRExC_state, scan, &min,
20032 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20033 return TRUE; /* Was return EXACT */
20037 switch (OP(REGNODE_p(scan))) {
20044 case EXACTFU_S_EDGE:
20045 case EXACTFAA_NO_TRIE:
20052 if( exact == PSEUDO )
20053 exact= OP(REGNODE_p(scan));
20054 else if ( exact != OP(REGNODE_p(scan)) )
20063 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20064 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20065 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
20066 SvPV_nolen_const(RExC_mysv),
20068 PL_reg_name[exact]);
20072 scan = REGNODE_OFFSET(temp);
20075 DEBUG_PARSE_MSG("");
20076 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20077 Perl_re_printf( aTHX_
20078 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20079 SvPV_nolen_const(RExC_mysv),
20084 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20085 assert((UV) (val - scan) <= U32_MAX);
20086 ARG_SET(REGNODE_p(scan), val - scan);
20089 if (val - scan > U16_MAX) {
20090 /* Populate this with something that won't loop and will likely
20091 * lead to a crash if the caller ignores the failure return, and
20092 * execution continues */
20093 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20096 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20099 return TRUE; /* Was 'return exact' */
20104 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
20106 /* Returns an inversion list of all the code points matched by the
20107 * ANYOFM/NANYOFM node 'n' */
20109 SV * cp_list = _new_invlist(-1);
20110 const U8 lowest = (U8) ARG(n);
20113 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
20115 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
20117 /* Starting with the lowest code point, any code point that ANDed with the
20118 * mask yields the lowest code point is in the set */
20119 for (i = lowest; i <= 0xFF; i++) {
20120 if ((i & FLAGS(n)) == ARG(n)) {
20121 cp_list = add_cp_to_invlist(cp_list, i);
20124 /* We know how many code points (a power of two) that are in the
20125 * set. No use looking once we've got that number */
20126 if (count >= needed) break;
20130 if (OP(n) == NANYOFM) {
20131 _invlist_invert(cp_list);
20137 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
20142 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
20147 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20149 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
20150 if (flags & (1<<bit)) {
20151 if (!set++ && lead)
20152 Perl_re_printf( aTHX_ "%s", lead);
20153 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
20158 Perl_re_printf( aTHX_ "\n");
20160 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20165 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20171 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20173 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20174 if (flags & (1<<bit)) {
20175 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20178 if (!set++ && lead)
20179 Perl_re_printf( aTHX_ "%s", lead);
20180 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20183 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20184 if (!set++ && lead) {
20185 Perl_re_printf( aTHX_ "%s", lead);
20188 case REGEX_UNICODE_CHARSET:
20189 Perl_re_printf( aTHX_ "UNICODE");
20191 case REGEX_LOCALE_CHARSET:
20192 Perl_re_printf( aTHX_ "LOCALE");
20194 case REGEX_ASCII_RESTRICTED_CHARSET:
20195 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20197 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20198 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20201 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20207 Perl_re_printf( aTHX_ "\n");
20209 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20215 Perl_regdump(pTHX_ const regexp *r)
20219 SV * const sv = sv_newmortal();
20220 SV *dsv= sv_newmortal();
20221 RXi_GET_DECL(r, ri);
20222 GET_RE_DEBUG_FLAGS_DECL;
20224 PERL_ARGS_ASSERT_REGDUMP;
20226 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20228 /* Header fields of interest. */
20229 for (i = 0; i < 2; i++) {
20230 if (r->substrs->data[i].substr) {
20231 RE_PV_QUOTED_DECL(s, 0, dsv,
20232 SvPVX_const(r->substrs->data[i].substr),
20233 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20234 PL_dump_re_max_len);
20235 Perl_re_printf( aTHX_
20236 "%s %s%s at %" IVdf "..%" UVuf " ",
20237 i ? "floating" : "anchored",
20239 RE_SV_TAIL(r->substrs->data[i].substr),
20240 (IV)r->substrs->data[i].min_offset,
20241 (UV)r->substrs->data[i].max_offset);
20243 else if (r->substrs->data[i].utf8_substr) {
20244 RE_PV_QUOTED_DECL(s, 1, dsv,
20245 SvPVX_const(r->substrs->data[i].utf8_substr),
20246 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20248 Perl_re_printf( aTHX_
20249 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20250 i ? "floating" : "anchored",
20252 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20253 (IV)r->substrs->data[i].min_offset,
20254 (UV)r->substrs->data[i].max_offset);
20258 if (r->check_substr || r->check_utf8)
20259 Perl_re_printf( aTHX_
20261 ( r->check_substr == r->substrs->data[1].substr
20262 && r->check_utf8 == r->substrs->data[1].utf8_substr
20263 ? "(checking floating" : "(checking anchored"));
20264 if (r->intflags & PREGf_NOSCAN)
20265 Perl_re_printf( aTHX_ " noscan");
20266 if (r->extflags & RXf_CHECK_ALL)
20267 Perl_re_printf( aTHX_ " isall");
20268 if (r->check_substr || r->check_utf8)
20269 Perl_re_printf( aTHX_ ") ");
20271 if (ri->regstclass) {
20272 regprop(r, sv, ri->regstclass, NULL, NULL);
20273 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
20275 if (r->intflags & PREGf_ANCH) {
20276 Perl_re_printf( aTHX_ "anchored");
20277 if (r->intflags & PREGf_ANCH_MBOL)
20278 Perl_re_printf( aTHX_ "(MBOL)");
20279 if (r->intflags & PREGf_ANCH_SBOL)
20280 Perl_re_printf( aTHX_ "(SBOL)");
20281 if (r->intflags & PREGf_ANCH_GPOS)
20282 Perl_re_printf( aTHX_ "(GPOS)");
20283 Perl_re_printf( aTHX_ " ");
20285 if (r->intflags & PREGf_GPOS_SEEN)
20286 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
20287 if (r->intflags & PREGf_SKIP)
20288 Perl_re_printf( aTHX_ "plus ");
20289 if (r->intflags & PREGf_IMPLICIT)
20290 Perl_re_printf( aTHX_ "implicit ");
20291 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
20292 if (r->extflags & RXf_EVAL_SEEN)
20293 Perl_re_printf( aTHX_ "with eval ");
20294 Perl_re_printf( aTHX_ "\n");
20296 regdump_extflags("r->extflags: ", r->extflags);
20297 regdump_intflags("r->intflags: ", r->intflags);
20300 PERL_ARGS_ASSERT_REGDUMP;
20301 PERL_UNUSED_CONTEXT;
20302 PERL_UNUSED_ARG(r);
20303 #endif /* DEBUGGING */
20306 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
20309 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
20310 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
20311 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
20312 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
20313 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
20314 || _CC_VERTSPACE != 15
20315 # error Need to adjust order of anyofs[]
20317 static const char * const anyofs[] = {
20354 - regprop - printable representation of opcode, with run time support
20358 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
20363 RXi_GET_DECL(prog, progi);
20364 GET_RE_DEBUG_FLAGS_DECL;
20366 PERL_ARGS_ASSERT_REGPROP;
20370 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
20371 if (pRExC_state) { /* This gives more info, if we have it */
20372 FAIL3("panic: corrupted regexp opcode %d > %d",
20373 (int)OP(o), (int)REGNODE_MAX);
20376 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
20377 (int)OP(o), (int)REGNODE_MAX);
20380 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
20382 k = PL_regkind[OP(o)];
20385 sv_catpvs(sv, " ");
20386 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
20387 * is a crude hack but it may be the best for now since
20388 * we have no flag "this EXACTish node was UTF-8"
20390 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
20391 PL_colors[0], PL_colors[1],
20392 PERL_PV_ESCAPE_UNI_DETECT |
20393 PERL_PV_ESCAPE_NONASCII |
20394 PERL_PV_PRETTY_ELLIPSES |
20395 PERL_PV_PRETTY_LTGT |
20396 PERL_PV_PRETTY_NOCLEAR
20398 } else if (k == TRIE) {
20399 /* print the details of the trie in dumpuntil instead, as
20400 * progi->data isn't available here */
20401 const char op = OP(o);
20402 const U32 n = ARG(o);
20403 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
20404 (reg_ac_data *)progi->data->data[n] :
20406 const reg_trie_data * const trie
20407 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
20409 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
20410 DEBUG_TRIE_COMPILE_r({
20412 sv_catpvs(sv, "(JUMP)");
20413 Perl_sv_catpvf(aTHX_ sv,
20414 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
20415 (UV)trie->startstate,
20416 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
20417 (UV)trie->wordcount,
20420 (UV)TRIE_CHARCOUNT(trie),
20421 (UV)trie->uniquecharcount
20424 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
20425 sv_catpvs(sv, "[");
20426 (void) put_charclass_bitmap_innards(sv,
20427 ((IS_ANYOF_TRIE(op))
20429 : TRIE_BITMAP(trie)),
20435 sv_catpvs(sv, "]");
20437 } else if (k == CURLY) {
20438 U32 lo = ARG1(o), hi = ARG2(o);
20439 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
20440 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
20441 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
20442 if (hi == REG_INFTY)
20443 sv_catpvs(sv, "INFTY");
20445 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
20446 sv_catpvs(sv, "}");
20448 else if (k == WHILEM && o->flags) /* Ordinal/of */
20449 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
20450 else if (k == REF || k == OPEN || k == CLOSE
20451 || k == GROUPP || OP(o)==ACCEPT)
20453 AV *name_list= NULL;
20454 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
20455 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
20456 if ( RXp_PAREN_NAMES(prog) ) {
20457 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20458 } else if ( pRExC_state ) {
20459 name_list= RExC_paren_name_list;
20462 if ( k != REF || (OP(o) < REFN)) {
20463 SV **name= av_fetch(name_list, parno, 0 );
20465 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20468 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
20469 I32 *nums=(I32*)SvPVX(sv_dat);
20470 SV **name= av_fetch(name_list, nums[0], 0 );
20473 for ( n=0; n<SvIVX(sv_dat); n++ ) {
20474 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
20475 (n ? "," : ""), (IV)nums[n]);
20477 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20481 if ( k == REF && reginfo) {
20482 U32 n = ARG(o); /* which paren pair */
20483 I32 ln = prog->offs[n].start;
20484 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
20485 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
20486 else if (ln == prog->offs[n].end)
20487 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
20489 const char *s = reginfo->strbeg + ln;
20490 Perl_sv_catpvf(aTHX_ sv, ": ");
20491 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
20492 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
20495 } else if (k == GOSUB) {
20496 AV *name_list= NULL;
20497 if ( RXp_PAREN_NAMES(prog) ) {
20498 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20499 } else if ( pRExC_state ) {
20500 name_list= RExC_paren_name_list;
20503 /* Paren and offset */
20504 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
20505 (int)((o + (int)ARG2L(o)) - progi->program) );
20507 SV **name= av_fetch(name_list, ARG(o), 0 );
20509 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20512 else if (k == LOGICAL)
20513 /* 2: embedded, otherwise 1 */
20514 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
20515 else if (k == ANYOF) {
20516 const U8 flags = inRANGE(OP(o), ANYOFH, ANYOFHr)
20519 bool do_sep = FALSE; /* Do we need to separate various components of
20521 /* Set if there is still an unresolved user-defined property */
20522 SV *unresolved = NULL;
20524 /* Things that are ignored except when the runtime locale is UTF-8 */
20525 SV *only_utf8_locale_invlist = NULL;
20527 /* Code points that don't fit in the bitmap */
20528 SV *nonbitmap_invlist = NULL;
20530 /* And things that aren't in the bitmap, but are small enough to be */
20531 SV* bitmap_range_not_in_bitmap = NULL;
20533 const bool inverted = flags & ANYOF_INVERT;
20535 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
20536 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
20537 sv_catpvs(sv, "{utf8-locale-reqd}");
20539 if (flags & ANYOFL_FOLD) {
20540 sv_catpvs(sv, "{i}");
20544 /* If there is stuff outside the bitmap, get it */
20545 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
20546 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
20548 &only_utf8_locale_invlist,
20549 &nonbitmap_invlist);
20550 /* The non-bitmap data may contain stuff that could fit in the
20551 * bitmap. This could come from a user-defined property being
20552 * finally resolved when this call was done; or much more likely
20553 * because there are matches that require UTF-8 to be valid, and so
20554 * aren't in the bitmap. This is teased apart later */
20555 _invlist_intersection(nonbitmap_invlist,
20557 &bitmap_range_not_in_bitmap);
20558 /* Leave just the things that don't fit into the bitmap */
20559 _invlist_subtract(nonbitmap_invlist,
20561 &nonbitmap_invlist);
20564 /* Obey this flag to add all above-the-bitmap code points */
20565 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
20566 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
20567 NUM_ANYOF_CODE_POINTS,
20571 /* Ready to start outputting. First, the initial left bracket */
20572 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20574 if (! inRANGE(OP(o), ANYOFH, ANYOFHr)) {
20575 /* Then all the things that could fit in the bitmap */
20576 do_sep = put_charclass_bitmap_innards(sv,
20578 bitmap_range_not_in_bitmap,
20579 only_utf8_locale_invlist,
20582 /* Can't try inverting for a
20583 * better display if there
20584 * are things that haven't
20586 unresolved != NULL);
20587 SvREFCNT_dec(bitmap_range_not_in_bitmap);
20589 /* If there are user-defined properties which haven't been defined
20590 * yet, output them. If the result is not to be inverted, it is
20591 * clearest to output them in a separate [] from the bitmap range
20592 * stuff. If the result is to be complemented, we have to show
20593 * everything in one [], as the inversion applies to the whole
20594 * thing. Use {braces} to separate them from anything in the
20595 * bitmap and anything above the bitmap. */
20598 if (! do_sep) { /* If didn't output anything in the bitmap
20600 sv_catpvs(sv, "^");
20602 sv_catpvs(sv, "{");
20605 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
20608 sv_catsv(sv, unresolved);
20610 sv_catpvs(sv, "}");
20612 do_sep = ! inverted;
20616 /* And, finally, add the above-the-bitmap stuff */
20617 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
20620 /* See if truncation size is overridden */
20621 const STRLEN dump_len = (PL_dump_re_max_len > 256)
20622 ? PL_dump_re_max_len
20625 /* This is output in a separate [] */
20627 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
20630 /* And, for easy of understanding, it is shown in the
20631 * uncomplemented form if possible. The one exception being if
20632 * there are unresolved items, where the inversion has to be
20633 * delayed until runtime */
20634 if (inverted && ! unresolved) {
20635 _invlist_invert(nonbitmap_invlist);
20636 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
20639 contents = invlist_contents(nonbitmap_invlist,
20640 FALSE /* output suitable for catsv */
20643 /* If the output is shorter than the permissible maximum, just do it. */
20644 if (SvCUR(contents) <= dump_len) {
20645 sv_catsv(sv, contents);
20648 const char * contents_string = SvPVX(contents);
20649 STRLEN i = dump_len;
20651 /* Otherwise, start at the permissible max and work back to the
20652 * first break possibility */
20653 while (i > 0 && contents_string[i] != ' ') {
20656 if (i == 0) { /* Fail-safe. Use the max if we couldn't
20657 find a legal break */
20661 sv_catpvn(sv, contents_string, i);
20662 sv_catpvs(sv, "...");
20665 SvREFCNT_dec_NN(contents);
20666 SvREFCNT_dec_NN(nonbitmap_invlist);
20669 /* And finally the matching, closing ']' */
20670 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20672 if (inRANGE(OP(o), ANYOFH, ANYOFHr)) {
20673 U8 lowest = (OP(o) != ANYOFHr)
20675 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
20676 U8 highest = (OP(o) == ANYOFHb)
20680 : HIGHEST_ANYOF_HRx_BYTE(FLAGS(o));
20681 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
20682 if (lowest != highest) {
20683 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
20685 Perl_sv_catpvf(aTHX_ sv, ")");
20688 SvREFCNT_dec(unresolved);
20690 else if (k == ANYOFM) {
20691 SV * cp_list = get_ANYOFM_contents(o);
20693 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20694 if (OP(o) == NANYOFM) {
20695 _invlist_invert(cp_list);
20698 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, TRUE);
20699 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20701 SvREFCNT_dec(cp_list);
20703 else if (k == POSIXD || k == NPOSIXD) {
20704 U8 index = FLAGS(o) * 2;
20705 if (index < C_ARRAY_LENGTH(anyofs)) {
20706 if (*anyofs[index] != '[') {
20707 sv_catpvs(sv, "[");
20709 sv_catpv(sv, anyofs[index]);
20710 if (*anyofs[index] != '[') {
20711 sv_catpvs(sv, "]");
20715 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
20718 else if (k == BOUND || k == NBOUND) {
20719 /* Must be synced with order of 'bound_type' in regcomp.h */
20720 const char * const bounds[] = {
20721 "", /* Traditional */
20727 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
20728 sv_catpv(sv, bounds[FLAGS(o)]);
20730 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
20731 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
20733 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
20735 Perl_sv_catpvf(aTHX_ sv, "]");
20737 else if (OP(o) == SBOL)
20738 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
20740 /* add on the verb argument if there is one */
20741 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
20743 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
20744 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
20746 sv_catpvs(sv, ":NULL");
20749 PERL_UNUSED_CONTEXT;
20750 PERL_UNUSED_ARG(sv);
20751 PERL_UNUSED_ARG(o);
20752 PERL_UNUSED_ARG(prog);
20753 PERL_UNUSED_ARG(reginfo);
20754 PERL_UNUSED_ARG(pRExC_state);
20755 #endif /* DEBUGGING */
20761 Perl_re_intuit_string(pTHX_ REGEXP * const r)
20762 { /* Assume that RE_INTUIT is set */
20763 struct regexp *const prog = ReANY(r);
20764 GET_RE_DEBUG_FLAGS_DECL;
20766 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
20767 PERL_UNUSED_CONTEXT;
20771 const char * const s = SvPV_nolen_const(RX_UTF8(r)
20772 ? prog->check_utf8 : prog->check_substr);
20774 if (!PL_colorset) reginitcolors();
20775 Perl_re_printf( aTHX_
20776 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
20778 RX_UTF8(r) ? "utf8 " : "",
20779 PL_colors[5], PL_colors[0],
20782 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
20785 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
20786 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
20792 handles refcounting and freeing the perl core regexp structure. When
20793 it is necessary to actually free the structure the first thing it
20794 does is call the 'free' method of the regexp_engine associated to
20795 the regexp, allowing the handling of the void *pprivate; member
20796 first. (This routine is not overridable by extensions, which is why
20797 the extensions free is called first.)
20799 See regdupe and regdupe_internal if you change anything here.
20801 #ifndef PERL_IN_XSUB_RE
20803 Perl_pregfree(pTHX_ REGEXP *r)
20809 Perl_pregfree2(pTHX_ REGEXP *rx)
20811 struct regexp *const r = ReANY(rx);
20812 GET_RE_DEBUG_FLAGS_DECL;
20814 PERL_ARGS_ASSERT_PREGFREE2;
20819 if (r->mother_re) {
20820 ReREFCNT_dec(r->mother_re);
20822 CALLREGFREE_PVT(rx); /* free the private data */
20823 SvREFCNT_dec(RXp_PAREN_NAMES(r));
20827 for (i = 0; i < 2; i++) {
20828 SvREFCNT_dec(r->substrs->data[i].substr);
20829 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
20831 Safefree(r->substrs);
20833 RX_MATCH_COPY_FREE(rx);
20834 #ifdef PERL_ANY_COW
20835 SvREFCNT_dec(r->saved_copy);
20838 SvREFCNT_dec(r->qr_anoncv);
20839 if (r->recurse_locinput)
20840 Safefree(r->recurse_locinput);
20846 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
20847 except that dsv will be created if NULL.
20849 This function is used in two main ways. First to implement
20850 $r = qr/....; $s = $$r;
20852 Secondly, it is used as a hacky workaround to the structural issue of
20854 being stored in the regexp structure which is in turn stored in
20855 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
20856 could be PL_curpm in multiple contexts, and could require multiple
20857 result sets being associated with the pattern simultaneously, such
20858 as when doing a recursive match with (??{$qr})
20860 The solution is to make a lightweight copy of the regexp structure
20861 when a qr// is returned from the code executed by (??{$qr}) this
20862 lightweight copy doesn't actually own any of its data except for
20863 the starp/end and the actual regexp structure itself.
20869 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
20871 struct regexp *drx;
20872 struct regexp *const srx = ReANY(ssv);
20873 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
20875 PERL_ARGS_ASSERT_REG_TEMP_COPY;
20878 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
20880 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
20882 /* our only valid caller, sv_setsv_flags(), should have done
20883 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
20884 assert(!SvOOK(dsv));
20885 assert(!SvIsCOW(dsv));
20886 assert(!SvROK(dsv));
20888 if (SvPVX_const(dsv)) {
20890 Safefree(SvPVX(dsv));
20895 SvOK_off((SV *)dsv);
20898 /* For PVLVs, the head (sv_any) points to an XPVLV, while
20899 * the LV's xpvlenu_rx will point to a regexp body, which
20900 * we allocate here */
20901 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
20902 assert(!SvPVX(dsv));
20903 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
20904 temp->sv_any = NULL;
20905 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
20906 SvREFCNT_dec_NN(temp);
20907 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
20908 ing below will not set it. */
20909 SvCUR_set(dsv, SvCUR(ssv));
20912 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
20913 sv_force_normal(sv) is called. */
20917 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
20918 SvPV_set(dsv, RX_WRAPPED(ssv));
20919 /* We share the same string buffer as the original regexp, on which we
20920 hold a reference count, incremented when mother_re is set below.
20921 The string pointer is copied here, being part of the regexp struct.
20923 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
20924 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
20928 const I32 npar = srx->nparens+1;
20929 Newx(drx->offs, npar, regexp_paren_pair);
20930 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
20932 if (srx->substrs) {
20934 Newx(drx->substrs, 1, struct reg_substr_data);
20935 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
20937 for (i = 0; i < 2; i++) {
20938 SvREFCNT_inc_void(drx->substrs->data[i].substr);
20939 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
20942 /* check_substr and check_utf8, if non-NULL, point to either their
20943 anchored or float namesakes, and don't hold a second reference. */
20945 RX_MATCH_COPIED_off(dsv);
20946 #ifdef PERL_ANY_COW
20947 drx->saved_copy = NULL;
20949 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
20950 SvREFCNT_inc_void(drx->qr_anoncv);
20951 if (srx->recurse_locinput)
20952 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
20959 /* regfree_internal()
20961 Free the private data in a regexp. This is overloadable by
20962 extensions. Perl takes care of the regexp structure in pregfree(),
20963 this covers the *pprivate pointer which technically perl doesn't
20964 know about, however of course we have to handle the
20965 regexp_internal structure when no extension is in use.
20967 Note this is called before freeing anything in the regexp
20972 Perl_regfree_internal(pTHX_ REGEXP * const rx)
20974 struct regexp *const r = ReANY(rx);
20975 RXi_GET_DECL(r, ri);
20976 GET_RE_DEBUG_FLAGS_DECL;
20978 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
20988 SV *dsv= sv_newmortal();
20989 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
20990 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
20991 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
20992 PL_colors[4], PL_colors[5], s);
20996 #ifdef RE_TRACK_PATTERN_OFFSETS
20998 Safefree(ri->u.offsets); /* 20010421 MJD */
21000 if (ri->code_blocks)
21001 S_free_codeblocks(aTHX_ ri->code_blocks);
21004 int n = ri->data->count;
21007 /* If you add a ->what type here, update the comment in regcomp.h */
21008 switch (ri->data->what[n]) {
21014 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21017 Safefree(ri->data->data[n]);
21023 { /* Aho Corasick add-on structure for a trie node.
21024 Used in stclass optimization only */
21026 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21027 #ifdef USE_ITHREADS
21031 refcount = --aho->refcount;
21034 PerlMemShared_free(aho->states);
21035 PerlMemShared_free(aho->fail);
21036 /* do this last!!!! */
21037 PerlMemShared_free(ri->data->data[n]);
21038 /* we should only ever get called once, so
21039 * assert as much, and also guard the free
21040 * which /might/ happen twice. At the least
21041 * it will make code anlyzers happy and it
21042 * doesn't cost much. - Yves */
21043 assert(ri->regstclass);
21044 if (ri->regstclass) {
21045 PerlMemShared_free(ri->regstclass);
21046 ri->regstclass = 0;
21053 /* trie structure. */
21055 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
21056 #ifdef USE_ITHREADS
21060 refcount = --trie->refcount;
21063 PerlMemShared_free(trie->charmap);
21064 PerlMemShared_free(trie->states);
21065 PerlMemShared_free(trie->trans);
21067 PerlMemShared_free(trie->bitmap);
21069 PerlMemShared_free(trie->jump);
21070 PerlMemShared_free(trie->wordinfo);
21071 /* do this last!!!! */
21072 PerlMemShared_free(ri->data->data[n]);
21077 Perl_croak(aTHX_ "panic: regfree data code '%c'",
21078 ri->data->what[n]);
21081 Safefree(ri->data->what);
21082 Safefree(ri->data);
21088 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
21089 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
21090 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
21093 re_dup_guts - duplicate a regexp.
21095 This routine is expected to clone a given regexp structure. It is only
21096 compiled under USE_ITHREADS.
21098 After all of the core data stored in struct regexp is duplicated
21099 the regexp_engine.dupe method is used to copy any private data
21100 stored in the *pprivate pointer. This allows extensions to handle
21101 any duplication it needs to do.
21103 See pregfree() and regfree_internal() if you change anything here.
21105 #if defined(USE_ITHREADS)
21106 #ifndef PERL_IN_XSUB_RE
21108 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
21112 const struct regexp *r = ReANY(sstr);
21113 struct regexp *ret = ReANY(dstr);
21115 PERL_ARGS_ASSERT_RE_DUP_GUTS;
21117 npar = r->nparens+1;
21118 Newx(ret->offs, npar, regexp_paren_pair);
21119 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
21121 if (ret->substrs) {
21122 /* Do it this way to avoid reading from *r after the StructCopy().
21123 That way, if any of the sv_dup_inc()s dislodge *r from the L1
21124 cache, it doesn't matter. */
21126 const bool anchored = r->check_substr
21127 ? r->check_substr == r->substrs->data[0].substr
21128 : r->check_utf8 == r->substrs->data[0].utf8_substr;
21129 Newx(ret->substrs, 1, struct reg_substr_data);
21130 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
21132 for (i = 0; i < 2; i++) {
21133 ret->substrs->data[i].substr =
21134 sv_dup_inc(ret->substrs->data[i].substr, param);
21135 ret->substrs->data[i].utf8_substr =
21136 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
21139 /* check_substr and check_utf8, if non-NULL, point to either their
21140 anchored or float namesakes, and don't hold a second reference. */
21142 if (ret->check_substr) {
21144 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
21146 ret->check_substr = ret->substrs->data[0].substr;
21147 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21149 assert(r->check_substr == r->substrs->data[1].substr);
21150 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
21152 ret->check_substr = ret->substrs->data[1].substr;
21153 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21155 } else if (ret->check_utf8) {
21157 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21159 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21164 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
21165 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
21166 if (r->recurse_locinput)
21167 Newx(ret->recurse_locinput, r->nparens + 1, char *);
21170 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
21172 if (RX_MATCH_COPIED(dstr))
21173 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
21175 ret->subbeg = NULL;
21176 #ifdef PERL_ANY_COW
21177 ret->saved_copy = NULL;
21180 /* Whether mother_re be set or no, we need to copy the string. We
21181 cannot refrain from copying it when the storage points directly to
21182 our mother regexp, because that's
21183 1: a buffer in a different thread
21184 2: something we no longer hold a reference on
21185 so we need to copy it locally. */
21186 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21187 /* set malloced length to a non-zero value so it will be freed
21188 * (otherwise in combination with SVf_FAKE it looks like an alien
21189 * buffer). It doesn't have to be the actual malloced size, since it
21190 * should never be grown */
21191 SvLEN_set(dstr, SvCUR(sstr)+1);
21192 ret->mother_re = NULL;
21194 #endif /* PERL_IN_XSUB_RE */
21199 This is the internal complement to regdupe() which is used to copy
21200 the structure pointed to by the *pprivate pointer in the regexp.
21201 This is the core version of the extension overridable cloning hook.
21202 The regexp structure being duplicated will be copied by perl prior
21203 to this and will be provided as the regexp *r argument, however
21204 with the /old/ structures pprivate pointer value. Thus this routine
21205 may override any copying normally done by perl.
21207 It returns a pointer to the new regexp_internal structure.
21211 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21214 struct regexp *const r = ReANY(rx);
21215 regexp_internal *reti;
21217 RXi_GET_DECL(r, ri);
21219 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21223 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
21224 char, regexp_internal);
21225 Copy(ri->program, reti->program, len+1, regnode);
21228 if (ri->code_blocks) {
21230 Newx(reti->code_blocks, 1, struct reg_code_blocks);
21231 Newx(reti->code_blocks->cb, ri->code_blocks->count,
21232 struct reg_code_block);
21233 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
21234 ri->code_blocks->count, struct reg_code_block);
21235 for (n = 0; n < ri->code_blocks->count; n++)
21236 reti->code_blocks->cb[n].src_regex = (REGEXP*)
21237 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
21238 reti->code_blocks->count = ri->code_blocks->count;
21239 reti->code_blocks->refcnt = 1;
21242 reti->code_blocks = NULL;
21244 reti->regstclass = NULL;
21247 struct reg_data *d;
21248 const int count = ri->data->count;
21251 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
21252 char, struct reg_data);
21253 Newx(d->what, count, U8);
21256 for (i = 0; i < count; i++) {
21257 d->what[i] = ri->data->what[i];
21258 switch (d->what[i]) {
21259 /* see also regcomp.h and regfree_internal() */
21260 case 'a': /* actually an AV, but the dup function is identical.
21261 values seem to be "plain sv's" generally. */
21262 case 'r': /* a compiled regex (but still just another SV) */
21263 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
21264 this use case should go away, the code could have used
21265 'a' instead - see S_set_ANYOF_arg() for array contents. */
21266 case 'S': /* actually an SV, but the dup function is identical. */
21267 case 'u': /* actually an HV, but the dup function is identical.
21268 values are "plain sv's" */
21269 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
21272 /* Synthetic Start Class - "Fake" charclass we generate to optimize
21273 * patterns which could start with several different things. Pre-TRIE
21274 * this was more important than it is now, however this still helps
21275 * in some places, for instance /x?a+/ might produce a SSC equivalent
21276 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
21279 /* This is cheating. */
21280 Newx(d->data[i], 1, regnode_ssc);
21281 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
21282 reti->regstclass = (regnode*)d->data[i];
21285 /* AHO-CORASICK fail table */
21286 /* Trie stclasses are readonly and can thus be shared
21287 * without duplication. We free the stclass in pregfree
21288 * when the corresponding reg_ac_data struct is freed.
21290 reti->regstclass= ri->regstclass;
21293 /* TRIE transition table */
21295 ((reg_trie_data*)ri->data->data[i])->refcount++;
21298 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
21299 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
21300 is not from another regexp */
21301 d->data[i] = ri->data->data[i];
21304 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
21305 ri->data->what[i]);
21314 reti->name_list_idx = ri->name_list_idx;
21316 #ifdef RE_TRACK_PATTERN_OFFSETS
21317 if (ri->u.offsets) {
21318 Newx(reti->u.offsets, 2*len+1, U32);
21319 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
21322 SetProgLen(reti, len);
21325 return (void*)reti;
21328 #endif /* USE_ITHREADS */
21330 #ifndef PERL_IN_XSUB_RE
21333 - regnext - dig the "next" pointer out of a node
21336 Perl_regnext(pTHX_ regnode *p)
21343 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
21344 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
21345 (int)OP(p), (int)REGNODE_MAX);
21348 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
21358 S_re_croak2(pTHX_ bool utf8, const char* pat1, const char* pat2,...)
21361 STRLEN l1 = strlen(pat1);
21362 STRLEN l2 = strlen(pat2);
21365 const char *message;
21367 PERL_ARGS_ASSERT_RE_CROAK2;
21373 Copy(pat1, buf, l1 , char);
21374 Copy(pat2, buf + l1, l2 , char);
21375 buf[l1 + l2] = '\n';
21376 buf[l1 + l2 + 1] = '\0';
21377 va_start(args, pat2);
21378 msv = vmess(buf, &args);
21380 message = SvPV_const(msv, l1);
21383 Copy(message, buf, l1 , char);
21384 /* l1-1 to avoid \n */
21385 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
21388 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
21390 #ifndef PERL_IN_XSUB_RE
21392 Perl_save_re_context(pTHX)
21397 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
21400 const REGEXP * const rx = PM_GETRE(PL_curpm);
21402 nparens = RX_NPARENS(rx);
21405 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
21406 * that PL_curpm will be null, but that utf8.pm and the modules it
21407 * loads will only use $1..$3.
21408 * The t/porting/re_context.t test file checks this assumption.
21413 for (i = 1; i <= nparens; i++) {
21414 char digits[TYPE_CHARS(long)];
21415 const STRLEN len = my_snprintf(digits, sizeof(digits),
21417 GV *const *const gvp
21418 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
21421 GV * const gv = *gvp;
21422 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
21432 S_put_code_point(pTHX_ SV *sv, UV c)
21434 PERL_ARGS_ASSERT_PUT_CODE_POINT;
21437 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
21439 else if (isPRINT(c)) {
21440 const char string = (char) c;
21442 /* We use {phrase} as metanotation in the class, so also escape literal
21444 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
21445 sv_catpvs(sv, "\\");
21446 sv_catpvn(sv, &string, 1);
21448 else if (isMNEMONIC_CNTRL(c)) {
21449 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
21452 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
21456 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
21459 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
21461 /* Appends to 'sv' a displayable version of the range of code points from
21462 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
21463 * that have them, when they occur at the beginning or end of the range.
21464 * It uses hex to output the remaining code points, unless 'allow_literals'
21465 * is true, in which case the printable ASCII ones are output as-is (though
21466 * some of these will be escaped by put_code_point()).
21468 * NOTE: This is designed only for printing ranges of code points that fit
21469 * inside an ANYOF bitmap. Higher code points are simply suppressed
21472 const unsigned int min_range_count = 3;
21474 assert(start <= end);
21476 PERL_ARGS_ASSERT_PUT_RANGE;
21478 while (start <= end) {
21480 const char * format;
21482 if (end - start < min_range_count) {
21484 /* Output chars individually when they occur in short ranges */
21485 for (; start <= end; start++) {
21486 put_code_point(sv, start);
21491 /* If permitted by the input options, and there is a possibility that
21492 * this range contains a printable literal, look to see if there is
21494 if (allow_literals && start <= MAX_PRINT_A) {
21496 /* If the character at the beginning of the range isn't an ASCII
21497 * printable, effectively split the range into two parts:
21498 * 1) the portion before the first such printable,
21500 * and output them separately. */
21501 if (! isPRINT_A(start)) {
21502 UV temp_end = start + 1;
21504 /* There is no point looking beyond the final possible
21505 * printable, in MAX_PRINT_A */
21506 UV max = MIN(end, MAX_PRINT_A);
21508 while (temp_end <= max && ! isPRINT_A(temp_end)) {
21512 /* Here, temp_end points to one beyond the first printable if
21513 * found, or to one beyond 'max' if not. If none found, make
21514 * sure that we use the entire range */
21515 if (temp_end > MAX_PRINT_A) {
21516 temp_end = end + 1;
21519 /* Output the first part of the split range: the part that
21520 * doesn't have printables, with the parameter set to not look
21521 * for literals (otherwise we would infinitely recurse) */
21522 put_range(sv, start, temp_end - 1, FALSE);
21524 /* The 2nd part of the range (if any) starts here. */
21527 /* We do a continue, instead of dropping down, because even if
21528 * the 2nd part is non-empty, it could be so short that we want
21529 * to output it as individual characters, as tested for at the
21530 * top of this loop. */
21534 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
21535 * output a sub-range of just the digits or letters, then process
21536 * the remaining portion as usual. */
21537 if (isALPHANUMERIC_A(start)) {
21538 UV mask = (isDIGIT_A(start))
21543 UV temp_end = start + 1;
21545 /* Find the end of the sub-range that includes just the
21546 * characters in the same class as the first character in it */
21547 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
21552 /* For short ranges, don't duplicate the code above to output
21553 * them; just call recursively */
21554 if (temp_end - start < min_range_count) {
21555 put_range(sv, start, temp_end, FALSE);
21557 else { /* Output as a range */
21558 put_code_point(sv, start);
21559 sv_catpvs(sv, "-");
21560 put_code_point(sv, temp_end);
21562 start = temp_end + 1;
21566 /* We output any other printables as individual characters */
21567 if (isPUNCT_A(start) || isSPACE_A(start)) {
21568 while (start <= end && (isPUNCT_A(start)
21569 || isSPACE_A(start)))
21571 put_code_point(sv, start);
21576 } /* End of looking for literals */
21578 /* Here is not to output as a literal. Some control characters have
21579 * mnemonic names. Split off any of those at the beginning and end of
21580 * the range to print mnemonically. It isn't possible for many of
21581 * these to be in a row, so this won't overwhelm with output */
21583 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
21585 while (isMNEMONIC_CNTRL(start) && start <= end) {
21586 put_code_point(sv, start);
21590 /* If this didn't take care of the whole range ... */
21591 if (start <= end) {
21593 /* Look backwards from the end to find the final non-mnemonic
21596 while (isMNEMONIC_CNTRL(temp_end)) {
21600 /* And separately output the interior range that doesn't start
21601 * or end with mnemonics */
21602 put_range(sv, start, temp_end, FALSE);
21604 /* Then output the mnemonic trailing controls */
21605 start = temp_end + 1;
21606 while (start <= end) {
21607 put_code_point(sv, start);
21614 /* As a final resort, output the range or subrange as hex. */
21616 if (start >= NUM_ANYOF_CODE_POINTS) {
21619 else { /* Have to split range at the bitmap boundary */
21620 this_end = (end < NUM_ANYOF_CODE_POINTS)
21622 : NUM_ANYOF_CODE_POINTS - 1;
21624 #if NUM_ANYOF_CODE_POINTS > 256
21625 format = (this_end < 256)
21626 ? "\\x%02" UVXf "-\\x%02" UVXf
21627 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
21629 format = "\\x%02" UVXf "-\\x%02" UVXf;
21631 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
21632 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
21633 GCC_DIAG_RESTORE_STMT;
21639 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
21641 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
21645 bool allow_literals = TRUE;
21647 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
21649 /* Generally, it is more readable if printable characters are output as
21650 * literals, but if a range (nearly) spans all of them, it's best to output
21651 * it as a single range. This code will use a single range if all but 2
21652 * ASCII printables are in it */
21653 invlist_iterinit(invlist);
21654 while (invlist_iternext(invlist, &start, &end)) {
21656 /* If the range starts beyond the final printable, it doesn't have any
21658 if (start > MAX_PRINT_A) {
21662 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
21663 * all but two, the range must start and end no later than 2 from
21665 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
21666 if (end > MAX_PRINT_A) {
21672 if (end - start >= MAX_PRINT_A - ' ' - 2) {
21673 allow_literals = FALSE;
21678 invlist_iterfinish(invlist);
21680 /* Here we have figured things out. Output each range */
21681 invlist_iterinit(invlist);
21682 while (invlist_iternext(invlist, &start, &end)) {
21683 if (start >= NUM_ANYOF_CODE_POINTS) {
21686 put_range(sv, start, end, allow_literals);
21688 invlist_iterfinish(invlist);
21694 S_put_charclass_bitmap_innards_common(pTHX_
21695 SV* invlist, /* The bitmap */
21696 SV* posixes, /* Under /l, things like [:word:], \S */
21697 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
21698 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
21699 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
21700 const bool invert /* Is the result to be inverted? */
21703 /* Create and return an SV containing a displayable version of the bitmap
21704 * and associated information determined by the input parameters. If the
21705 * output would have been only the inversion indicator '^', NULL is instead
21711 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
21714 output = newSVpvs("^");
21717 output = newSVpvs("");
21720 /* First, the code points in the bitmap that are unconditionally there */
21721 put_charclass_bitmap_innards_invlist(output, invlist);
21723 /* Traditionally, these have been placed after the main code points */
21725 sv_catsv(output, posixes);
21728 if (only_utf8 && _invlist_len(only_utf8)) {
21729 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
21730 put_charclass_bitmap_innards_invlist(output, only_utf8);
21733 if (not_utf8 && _invlist_len(not_utf8)) {
21734 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
21735 put_charclass_bitmap_innards_invlist(output, not_utf8);
21738 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
21739 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
21740 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
21742 /* This is the only list in this routine that can legally contain code
21743 * points outside the bitmap range. The call just above to
21744 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
21745 * output them here. There's about a half-dozen possible, and none in
21746 * contiguous ranges longer than 2 */
21747 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21749 SV* above_bitmap = NULL;
21751 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
21753 invlist_iterinit(above_bitmap);
21754 while (invlist_iternext(above_bitmap, &start, &end)) {
21757 for (i = start; i <= end; i++) {
21758 put_code_point(output, i);
21761 invlist_iterfinish(above_bitmap);
21762 SvREFCNT_dec_NN(above_bitmap);
21766 if (invert && SvCUR(output) == 1) {
21774 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
21776 SV *nonbitmap_invlist,
21777 SV *only_utf8_locale_invlist,
21778 const regnode * const node,
21779 const bool force_as_is_display)
21781 /* Appends to 'sv' a displayable version of the innards of the bracketed
21782 * character class defined by the other arguments:
21783 * 'bitmap' points to the bitmap, or NULL if to ignore that.
21784 * 'nonbitmap_invlist' is an inversion list of the code points that are in
21785 * the bitmap range, but for some reason aren't in the bitmap; NULL if
21786 * none. The reasons for this could be that they require some
21787 * condition such as the target string being or not being in UTF-8
21788 * (under /d), or because they came from a user-defined property that
21789 * was not resolved at the time of the regex compilation (under /u)
21790 * 'only_utf8_locale_invlist' is an inversion list of the code points that
21791 * are valid only if the runtime locale is a UTF-8 one; NULL if none
21792 * 'node' is the regex pattern ANYOF node. It is needed only when the
21793 * above two parameters are not null, and is passed so that this
21794 * routine can tease apart the various reasons for them.
21795 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
21796 * to invert things to see if that leads to a cleaner display. If
21797 * FALSE, this routine is free to use its judgment about doing this.
21799 * It returns TRUE if there was actually something output. (It may be that
21800 * the bitmap, etc is empty.)
21802 * When called for outputting the bitmap of a non-ANYOF node, just pass the
21803 * bitmap, with the succeeding parameters set to NULL, and the final one to
21807 /* In general, it tries to display the 'cleanest' representation of the
21808 * innards, choosing whether to display them inverted or not, regardless of
21809 * whether the class itself is to be inverted. However, there are some
21810 * cases where it can't try inverting, as what actually matches isn't known
21811 * until runtime, and hence the inversion isn't either. */
21814 bool inverting_allowed = ! force_as_is_display;
21817 STRLEN orig_sv_cur = SvCUR(sv);
21819 SV* invlist; /* Inversion list we accumulate of code points that
21820 are unconditionally matched */
21821 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
21823 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
21825 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
21826 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
21829 SV* as_is_display; /* The output string when we take the inputs
21831 SV* inverted_display; /* The output string when we invert the inputs */
21833 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
21835 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
21837 /* We are biased in favor of displaying things without them being inverted,
21838 * as that is generally easier to understand */
21839 const int bias = 5;
21841 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
21843 /* Start off with whatever code points are passed in. (We clone, so we
21844 * don't change the caller's list) */
21845 if (nonbitmap_invlist) {
21846 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
21847 invlist = invlist_clone(nonbitmap_invlist, NULL);
21849 else { /* Worst case size is every other code point is matched */
21850 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
21854 if (OP(node) == ANYOFD) {
21856 /* This flag indicates that the code points below 0x100 in the
21857 * nonbitmap list are precisely the ones that match only when the
21858 * target is UTF-8 (they should all be non-ASCII). */
21859 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
21861 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
21862 _invlist_subtract(invlist, only_utf8, &invlist);
21865 /* And this flag for matching all non-ASCII 0xFF and below */
21866 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
21868 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
21871 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
21873 /* If either of these flags are set, what matches isn't
21874 * determinable except during execution, so don't know enough here
21876 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
21877 inverting_allowed = FALSE;
21880 /* What the posix classes match also varies at runtime, so these
21881 * will be output symbolically. */
21882 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
21885 posixes = newSVpvs("");
21886 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
21887 if (ANYOF_POSIXL_TEST(node, i)) {
21888 sv_catpv(posixes, anyofs[i]);
21895 /* Accumulate the bit map into the unconditional match list */
21897 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
21898 if (BITMAP_TEST(bitmap, i)) {
21901 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
21904 invlist = _add_range_to_invlist(invlist, start, i-1);
21909 /* Make sure that the conditional match lists don't have anything in them
21910 * that match unconditionally; otherwise the output is quite confusing.
21911 * This could happen if the code that populates these misses some
21914 _invlist_subtract(only_utf8, invlist, &only_utf8);
21917 _invlist_subtract(not_utf8, invlist, ¬_utf8);
21920 if (only_utf8_locale_invlist) {
21922 /* Since this list is passed in, we have to make a copy before
21924 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
21926 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
21928 /* And, it can get really weird for us to try outputting an inverted
21929 * form of this list when it has things above the bitmap, so don't even
21931 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21932 inverting_allowed = FALSE;
21936 /* Calculate what the output would be if we take the input as-is */
21937 as_is_display = put_charclass_bitmap_innards_common(invlist,
21944 /* If have to take the output as-is, just do that */
21945 if (! inverting_allowed) {
21946 if (as_is_display) {
21947 sv_catsv(sv, as_is_display);
21948 SvREFCNT_dec_NN(as_is_display);
21951 else { /* But otherwise, create the output again on the inverted input, and
21952 use whichever version is shorter */
21954 int inverted_bias, as_is_bias;
21956 /* We will apply our bias to whichever of the the results doesn't have
21966 inverted_bias = bias;
21969 /* Now invert each of the lists that contribute to the output,
21970 * excluding from the result things outside the possible range */
21972 /* For the unconditional inversion list, we have to add in all the
21973 * conditional code points, so that when inverted, they will be gone
21975 _invlist_union(only_utf8, invlist, &invlist);
21976 _invlist_union(not_utf8, invlist, &invlist);
21977 _invlist_union(only_utf8_locale, invlist, &invlist);
21978 _invlist_invert(invlist);
21979 _invlist_intersection(invlist, PL_InBitmap, &invlist);
21982 _invlist_invert(only_utf8);
21983 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
21985 else if (not_utf8) {
21987 /* If a code point matches iff the target string is not in UTF-8,
21988 * then complementing the result has it not match iff not in UTF-8,
21989 * which is the same thing as matching iff it is UTF-8. */
21990 only_utf8 = not_utf8;
21994 if (only_utf8_locale) {
21995 _invlist_invert(only_utf8_locale);
21996 _invlist_intersection(only_utf8_locale,
21998 &only_utf8_locale);
22001 inverted_display = put_charclass_bitmap_innards_common(
22006 only_utf8_locale, invert);
22008 /* Use the shortest representation, taking into account our bias
22009 * against showing it inverted */
22010 if ( inverted_display
22011 && ( ! as_is_display
22012 || ( SvCUR(inverted_display) + inverted_bias
22013 < SvCUR(as_is_display) + as_is_bias)))
22015 sv_catsv(sv, inverted_display);
22017 else if (as_is_display) {
22018 sv_catsv(sv, as_is_display);
22021 SvREFCNT_dec(as_is_display);
22022 SvREFCNT_dec(inverted_display);
22025 SvREFCNT_dec_NN(invlist);
22026 SvREFCNT_dec(only_utf8);
22027 SvREFCNT_dec(not_utf8);
22028 SvREFCNT_dec(posixes);
22029 SvREFCNT_dec(only_utf8_locale);
22031 return SvCUR(sv) > orig_sv_cur;
22034 #define CLEAR_OPTSTART \
22035 if (optstart) STMT_START { \
22036 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22037 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22041 #define DUMPUNTIL(b,e) \
22043 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
22045 STATIC const regnode *
22046 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
22047 const regnode *last, const regnode *plast,
22048 SV* sv, I32 indent, U32 depth)
22050 U8 op = PSEUDO; /* Arbitrary non-END op. */
22051 const regnode *next;
22052 const regnode *optstart= NULL;
22054 RXi_GET_DECL(r, ri);
22055 GET_RE_DEBUG_FLAGS_DECL;
22057 PERL_ARGS_ASSERT_DUMPUNTIL;
22059 #ifdef DEBUG_DUMPUNTIL
22060 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
22061 last ? last-start : 0, plast ? plast-start : 0);
22064 if (plast && plast < last)
22067 while (PL_regkind[op] != END && (!last || node < last)) {
22069 /* While that wasn't END last time... */
22072 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
22074 next = regnext((regnode *)node);
22077 if (OP(node) == OPTIMIZED) {
22078 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
22085 regprop(r, sv, node, NULL, NULL);
22086 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
22087 (int)(2*indent + 1), "", SvPVX_const(sv));
22089 if (OP(node) != OPTIMIZED) {
22090 if (next == NULL) /* Next ptr. */
22091 Perl_re_printf( aTHX_ " (0)");
22092 else if (PL_regkind[(U8)op] == BRANCH
22093 && PL_regkind[OP(next)] != BRANCH )
22094 Perl_re_printf( aTHX_ " (FAIL)");
22096 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
22097 Perl_re_printf( aTHX_ "\n");
22101 if (PL_regkind[(U8)op] == BRANCHJ) {
22104 const regnode *nnode = (OP(next) == LONGJMP
22105 ? regnext((regnode *)next)
22107 if (last && nnode > last)
22109 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
22112 else if (PL_regkind[(U8)op] == BRANCH) {
22114 DUMPUNTIL(NEXTOPER(node), next);
22116 else if ( PL_regkind[(U8)op] == TRIE ) {
22117 const regnode *this_trie = node;
22118 const char op = OP(node);
22119 const U32 n = ARG(node);
22120 const reg_ac_data * const ac = op>=AHOCORASICK ?
22121 (reg_ac_data *)ri->data->data[n] :
22123 const reg_trie_data * const trie =
22124 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
22126 AV *const trie_words
22127 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
22129 const regnode *nextbranch= NULL;
22132 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
22133 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
22135 Perl_re_indentf( aTHX_ "%s ",
22138 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
22139 SvCUR(*elem_ptr), PL_dump_re_max_len,
22140 PL_colors[0], PL_colors[1],
22142 ? PERL_PV_ESCAPE_UNI
22144 | PERL_PV_PRETTY_ELLIPSES
22145 | PERL_PV_PRETTY_LTGT
22150 U16 dist= trie->jump[word_idx+1];
22151 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
22152 (UV)((dist ? this_trie + dist : next) - start));
22155 nextbranch= this_trie + trie->jump[0];
22156 DUMPUNTIL(this_trie + dist, nextbranch);
22158 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
22159 nextbranch= regnext((regnode *)nextbranch);
22161 Perl_re_printf( aTHX_ "\n");
22164 if (last && next > last)
22169 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
22170 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
22171 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
22173 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
22175 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
22177 else if ( op == PLUS || op == STAR) {
22178 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
22180 else if (PL_regkind[(U8)op] == EXACT) {
22181 /* Literal string, where present. */
22182 node += NODE_SZ_STR(node) - 1;
22183 node = NEXTOPER(node);
22186 node = NEXTOPER(node);
22187 node += regarglen[(U8)op];
22189 if (op == CURLYX || op == OPEN || op == SROPEN)
22193 #ifdef DEBUG_DUMPUNTIL
22194 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22199 #endif /* DEBUGGING */
22201 #ifndef PERL_IN_XSUB_RE
22203 #include "uni_keywords.h"
22206 Perl_init_uniprops(pTHX)
22210 PL_user_def_props = newHV();
22212 #ifdef USE_ITHREADS
22214 HvSHAREKEYS_off(PL_user_def_props);
22215 PL_user_def_props_aTHX = aTHX;
22219 /* Set up the inversion list global variables */
22221 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22222 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
22223 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
22224 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
22225 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
22226 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
22227 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
22228 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
22229 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
22230 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
22231 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
22232 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
22233 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
22234 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
22235 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
22236 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
22238 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22239 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
22240 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
22241 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
22242 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
22243 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
22244 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
22245 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
22246 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
22247 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
22248 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
22249 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
22250 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
22251 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
22252 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
22253 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
22255 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
22256 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
22257 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
22258 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
22259 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
22261 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
22262 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
22263 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
22265 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
22267 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
22268 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
22270 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
22271 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
22273 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
22274 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22275 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
22276 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22277 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
22278 PL_NonFinalFold = _new_invlist_C_array(uni_prop_ptrs[
22279 UNI__PERL_NON_FINAL_FOLDS]);
22281 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
22282 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
22283 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
22284 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
22285 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
22286 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
22287 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
22288 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
22289 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
22292 /* The below are used only by deprecated functions. They could be removed */
22293 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
22294 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
22295 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
22301 This code was mainly added for backcompat to give a warning for non-portable
22302 code points in user-defined properties. But experiments showed that the
22303 warning in earlier perls were only omitted on overflow, which should be an
22304 error, so there really isnt a backcompat issue, and actually adding the
22305 warning when none was present before might cause breakage, for little gain. So
22306 khw left this code in, but not enabled. Tests were never added.
22309 Ei |const char *|get_extended_utf8_msg|const UV cp
22311 PERL_STATIC_INLINE const char *
22312 S_get_extended_utf8_msg(pTHX_ const UV cp)
22314 U8 dummy[UTF8_MAXBYTES + 1];
22318 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
22321 msg = hv_fetchs(msgs, "text", 0);
22324 (void) sv_2mortal((SV *) msgs);
22326 return SvPVX(*msg);
22332 Perl_handle_user_defined_property(pTHX_
22334 /* Parses the contents of a user-defined property definition; returning the
22335 * expanded definition if possible. If so, the return is an inversion
22338 * If there are subroutines that are part of the expansion and which aren't
22339 * known at the time of the call to this function, this returns what
22340 * parse_uniprop_string() returned for the first one encountered.
22342 * If an error was found, NULL is returned, and 'msg' gets a suitable
22343 * message appended to it. (Appending allows the back trace of how we got
22344 * to the faulty definition to be displayed through nested calls of
22345 * user-defined subs.)
22347 * The caller IS responsible for freeing any returned SV.
22349 * The syntax of the contents is pretty much described in perlunicode.pod,
22350 * but we also allow comments on each line */
22352 const char * name, /* Name of property */
22353 const STRLEN name_len, /* The name's length in bytes */
22354 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22355 const bool to_fold, /* ? Is this under /i */
22356 const bool runtime, /* ? Are we in compile- or run-time */
22357 const bool deferrable, /* Is it ok for this property's full definition
22358 to be deferred until later? */
22359 SV* contents, /* The property's definition */
22360 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
22361 getting called unless this is thought to be
22362 a user-defined property */
22363 SV * msg, /* Any error or warning msg(s) are appended to
22365 const STRLEN level) /* Recursion level of this call */
22368 const char * string = SvPV_const(contents, len);
22369 const char * const e = string + len;
22370 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
22371 const STRLEN msgs_length_on_entry = SvCUR(msg);
22373 const char * s0 = string; /* Points to first byte in the current line
22374 being parsed in 'string' */
22375 const char overflow_msg[] = "Code point too large in \"";
22376 SV* running_definition = NULL;
22378 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
22380 *user_defined_ptr = TRUE;
22382 /* Look at each line */
22384 const char * s; /* Current byte */
22385 char op = '+'; /* Default operation is 'union' */
22386 IV min = 0; /* range begin code point */
22387 IV max = -1; /* and range end */
22388 SV* this_definition;
22390 /* Skip comment lines */
22392 s0 = strchr(s0, '\n');
22400 /* For backcompat, allow an empty first line */
22406 /* First character in the line may optionally be the operation */
22415 /* If the line is one or two hex digits separated by blank space, its
22416 * a range; otherwise it is either another user-defined property or an
22421 if (! isXDIGIT(*s)) {
22422 goto check_if_property;
22425 do { /* Each new hex digit will add 4 bits. */
22426 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
22427 s = strchr(s, '\n');
22431 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22432 sv_catpv(msg, overflow_msg);
22433 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22434 UTF8fARG(is_contents_utf8, s - s0, s0));
22435 sv_catpvs(msg, "\"");
22436 goto return_failure;
22439 /* Accumulate this digit into the value */
22440 min = (min << 4) + READ_XDIGIT(s);
22441 } while (isXDIGIT(*s));
22443 while (isBLANK(*s)) { s++; }
22445 /* We allow comments at the end of the line */
22447 s = strchr(s, '\n');
22453 else if (s < e && *s != '\n') {
22454 if (! isXDIGIT(*s)) {
22455 goto check_if_property;
22458 /* Look for the high point of the range */
22461 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
22462 s = strchr(s, '\n');
22466 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22467 sv_catpv(msg, overflow_msg);
22468 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22469 UTF8fARG(is_contents_utf8, s - s0, s0));
22470 sv_catpvs(msg, "\"");
22471 goto return_failure;
22474 max = (max << 4) + READ_XDIGIT(s);
22475 } while (isXDIGIT(*s));
22477 while (isBLANK(*s)) { s++; }
22480 s = strchr(s, '\n');
22485 else if (s < e && *s != '\n') {
22486 goto check_if_property;
22490 if (max == -1) { /* The line only had one entry */
22493 else if (max < min) {
22494 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22495 sv_catpvs(msg, "Illegal range in \"");
22496 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22497 UTF8fARG(is_contents_utf8, s - s0, s0));
22498 sv_catpvs(msg, "\"");
22499 goto return_failure;
22502 #if 0 /* See explanation at definition above of get_extended_utf8_msg() */
22504 if ( UNICODE_IS_PERL_EXTENDED(min)
22505 || UNICODE_IS_PERL_EXTENDED(max))
22507 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22509 /* If both code points are non-portable, warn only on the lower
22511 sv_catpv(msg, get_extended_utf8_msg(
22512 (UNICODE_IS_PERL_EXTENDED(min))
22514 sv_catpvs(msg, " in \"");
22515 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22516 UTF8fARG(is_contents_utf8, s - s0, s0));
22517 sv_catpvs(msg, "\"");
22522 /* Here, this line contains a legal range */
22523 this_definition = sv_2mortal(_new_invlist(2));
22524 this_definition = _add_range_to_invlist(this_definition, min, max);
22529 /* Here it isn't a legal range line. See if it is a legal property
22530 * line. First find the end of the meat of the line */
22531 s = strpbrk(s, "#\n");
22536 /* Ignore trailing blanks in keeping with the requirements of
22537 * parse_uniprop_string() */
22539 while (s > s0 && isBLANK_A(*s)) {
22544 this_definition = parse_uniprop_string(s0, s - s0,
22545 is_utf8, to_fold, runtime,
22547 user_defined_ptr, msg,
22549 ? level /* Don't increase level
22550 if input is empty */
22553 if (this_definition == NULL) {
22554 goto return_failure; /* 'msg' should have had the reason
22555 appended to it by the above call */
22558 if (! is_invlist(this_definition)) { /* Unknown at this time */
22559 return newSVsv(this_definition);
22563 s = strchr(s, '\n');
22573 _invlist_union(running_definition, this_definition,
22574 &running_definition);
22577 _invlist_subtract(running_definition, this_definition,
22578 &running_definition);
22581 _invlist_intersection(running_definition, this_definition,
22582 &running_definition);
22585 _invlist_union_complement_2nd(running_definition,
22586 this_definition, &running_definition);
22589 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
22590 __FILE__, __LINE__, op);
22594 /* Position past the '\n' */
22596 } /* End of loop through the lines of 'contents' */
22598 /* Here, we processed all the lines in 'contents' without error. If we
22599 * didn't add any warnings, simply return success */
22600 if (msgs_length_on_entry == SvCUR(msg)) {
22602 /* If the expansion was empty, the answer isn't nothing: its an empty
22603 * inversion list */
22604 if (running_definition == NULL) {
22605 running_definition = _new_invlist(1);
22608 return running_definition;
22611 /* Otherwise, add some explanatory text, but we will return success */
22615 running_definition = NULL;
22619 if (name_len > 0) {
22620 sv_catpvs(msg, " in expansion of ");
22621 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
22624 return running_definition;
22627 /* As explained below, certain operations need to take place in the first
22628 * thread created. These macros switch contexts */
22629 #ifdef USE_ITHREADS
22630 # define DECLARATION_FOR_GLOBAL_CONTEXT \
22631 PerlInterpreter * save_aTHX = aTHX;
22632 # define SWITCH_TO_GLOBAL_CONTEXT \
22633 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
22634 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
22635 # define CUR_CONTEXT aTHX
22636 # define ORIGINAL_CONTEXT save_aTHX
22638 # define DECLARATION_FOR_GLOBAL_CONTEXT
22639 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
22640 # define RESTORE_CONTEXT NOOP
22641 # define CUR_CONTEXT NULL
22642 # define ORIGINAL_CONTEXT NULL
22646 S_delete_recursion_entry(pTHX_ void *key)
22648 /* Deletes the entry used to detect recursion when expanding user-defined
22649 * properties. This is a function so it can be set up to be called even if
22650 * the program unexpectedly quits */
22653 SV ** current_entry;
22654 const STRLEN key_len = strlen((const char *) key);
22655 DECLARATION_FOR_GLOBAL_CONTEXT;
22657 SWITCH_TO_GLOBAL_CONTEXT;
22659 /* If the entry is one of these types, it is a permanent entry, and not the
22660 * one used to detect recursions. This function should delete only the
22661 * recursion entry */
22662 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
22664 && ! is_invlist(*current_entry)
22665 && ! SvPOK(*current_entry))
22667 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
22675 S_get_fq_name(pTHX_
22676 const char * const name, /* The first non-blank in the \p{}, \P{} */
22677 const Size_t name_len, /* Its length in bytes, not including any trailing space */
22678 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22679 const bool has_colon_colon
22682 /* Returns a mortal SV containing the fully qualified version of the input
22687 fq_name = newSVpvs_flags("", SVs_TEMP);
22689 /* Use the current package if it wasn't included in our input */
22690 if (! has_colon_colon) {
22691 const HV * pkg = (IN_PERL_COMPILETIME)
22693 : CopSTASH(PL_curcop);
22694 const char* pkgname = HvNAME(pkg);
22696 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22697 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
22698 sv_catpvs(fq_name, "::");
22701 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22702 UTF8fARG(is_utf8, name_len, name));
22707 Perl_parse_uniprop_string(pTHX_
22709 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
22710 * now. If so, the return is an inversion list.
22712 * If the property is user-defined, it is a subroutine, which in turn
22713 * may call other subroutines. This function will call the whole nest of
22714 * them to get the definition they return; if some aren't known at the time
22715 * of the call to this function, the fully qualified name of the highest
22716 * level sub is returned. It is an error to call this function at runtime
22717 * without every sub defined.
22719 * If an error was found, NULL is returned, and 'msg' gets a suitable
22720 * message appended to it. (Appending allows the back trace of how we got
22721 * to the faulty definition to be displayed through nested calls of
22722 * user-defined subs.)
22724 * The caller should NOT try to free any returned inversion list.
22726 * Other parameters will be set on return as described below */
22728 const char * const name, /* The first non-blank in the \p{}, \P{} */
22729 const Size_t name_len, /* Its length in bytes, not including any
22731 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22732 const bool to_fold, /* ? Is this under /i */
22733 const bool runtime, /* TRUE if this is being called at run time */
22734 const bool deferrable, /* TRUE if it's ok for the definition to not be
22735 known at this call */
22736 bool *user_defined_ptr, /* Upon return from this function it will be
22737 set to TRUE if any component is a
22738 user-defined property */
22739 SV * msg, /* Any error or warning msg(s) are appended to
22741 const STRLEN level) /* Recursion level of this call */
22744 char* lookup_name; /* normalized name for lookup in our tables */
22745 unsigned lookup_len; /* Its length */
22746 bool stricter = FALSE; /* Some properties have stricter name
22747 normalization rules, which we decide upon
22748 based on parsing */
22750 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
22751 * (though it requires extra effort to download them from Unicode and
22752 * compile perl to know about them) */
22753 bool is_nv_type = FALSE;
22755 unsigned int i, j = 0;
22756 int equals_pos = -1; /* Where the '=' is found, or negative if none */
22757 int slash_pos = -1; /* Where the '/' is found, or negative if none */
22758 int table_index = 0; /* The entry number for this property in the table
22759 of all Unicode property names */
22760 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
22761 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
22762 the normalized name in certain situations */
22763 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
22764 part of a package name */
22765 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
22766 property rather than a Unicode
22768 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
22769 if an error. If it is an inversion list,
22770 it is the definition. Otherwise it is a
22771 string containing the fully qualified sub
22773 SV * fq_name = NULL; /* For user-defined properties, the fully
22775 bool invert_return = FALSE; /* ? Do we need to complement the result before
22778 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
22780 /* The input will be normalized into 'lookup_name' */
22781 Newx(lookup_name, name_len, char);
22782 SAVEFREEPV(lookup_name);
22784 /* Parse the input. */
22785 for (i = 0; i < name_len; i++) {
22786 char cur = name[i];
22788 /* Most of the characters in the input will be of this ilk, being parts
22790 if (isIDCONT_A(cur)) {
22792 /* Case differences are ignored. Our lookup routine assumes
22793 * everything is lowercase, so normalize to that */
22794 if (isUPPER_A(cur)) {
22795 lookup_name[j++] = toLOWER_A(cur);
22799 if (cur == '_') { /* Don't include these in the normalized name */
22803 lookup_name[j++] = cur;
22805 /* The first character in a user-defined name must be of this type.
22807 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
22808 could_be_user_defined = FALSE;
22814 /* Here, the character is not something typically in a name, But these
22815 * two types of characters (and the '_' above) can be freely ignored in
22816 * most situations. Later it may turn out we shouldn't have ignored
22817 * them, and we have to reparse, but we don't have enough information
22818 * yet to make that decision */
22819 if (cur == '-' || isSPACE_A(cur)) {
22820 could_be_user_defined = FALSE;
22824 /* An equals sign or single colon mark the end of the first part of
22825 * the property name */
22827 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
22829 lookup_name[j++] = '='; /* Treat the colon as an '=' */
22830 equals_pos = j; /* Note where it occurred in the input */
22831 could_be_user_defined = FALSE;
22835 /* Otherwise, this character is part of the name. */
22836 lookup_name[j++] = cur;
22838 /* Here it isn't a single colon, so if it is a colon, it must be a
22842 /* A double colon should be a package qualifier. We note its
22843 * position and continue. Note that one could have
22844 * pkg1::pkg2::...::foo
22845 * so that the position at the end of the loop will be just after
22846 * the final qualifier */
22849 non_pkg_begin = i + 1;
22850 lookup_name[j++] = ':';
22852 else { /* Only word chars (and '::') can be in a user-defined name */
22853 could_be_user_defined = FALSE;
22855 } /* End of parsing through the lhs of the property name (or all of it if
22858 #define STRLENs(s) (sizeof("" s "") - 1)
22860 /* If there is a single package name 'utf8::', it is ambiguous. It could
22861 * be for a user-defined property, or it could be a Unicode property, as
22862 * all of them are considered to be for that package. For the purposes of
22863 * parsing the rest of the property, strip it off */
22864 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
22865 lookup_name += STRLENs("utf8::");
22866 j -= STRLENs("utf8::");
22867 equals_pos -= STRLENs("utf8::");
22870 /* Here, we are either done with the whole property name, if it was simple;
22871 * or are positioned just after the '=' if it is compound. */
22873 if (equals_pos >= 0) {
22874 assert(! stricter); /* We shouldn't have set this yet */
22876 /* Space immediately after the '=' is ignored */
22878 for (; i < name_len; i++) {
22879 if (! isSPACE_A(name[i])) {
22884 /* Most punctuation after the equals indicates a subpattern, like
22886 if ( isPUNCT_A(name[i])
22892 /* Find the property. The table includes the equals sign, so we
22894 table_index = match_uniprop((U8 *) lookup_name, j);
22896 const char * const * prop_values
22897 = UNI_prop_value_ptrs[table_index];
22899 Size_t subpattern_len;
22900 REGEXP * subpattern_re;
22901 char open = name[i++];
22903 const char * pos_in_brackets;
22906 /* A backslash means the real delimitter is the next character.
22908 if (open == '\\') {
22913 /* This data structure is constructed so that the matching
22914 * closing bracket is 3 past its matching opening. The second
22915 * set of closing is so that if the opening is something like
22916 * ']', the closing will be that as well. Something similar is
22917 * done in toke.c */
22918 pos_in_brackets = strchr("([<)]>)]>", open);
22919 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
22922 || name[name_len-1] != close
22923 || (escaped && name[name_len-2] != '\\'))
22925 sv_catpvs(msg, "Unicode property wildcard not terminated");
22926 goto append_name_to_msg;
22929 Perl_ck_warner_d(aTHX_
22930 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
22931 "The Unicode property wildcards feature is experimental");
22933 /* Now create and compile the wildcard subpattern. Use /iaa
22934 * because nothing outside of ASCII will match, and it the
22935 * property values should all match /i. Note that when the
22936 * pattern fails to compile, our added text to the user's
22937 * pattern will be displayed to the user, which is not so
22939 subpattern_len = name_len - i - 1 - escaped;
22940 subpattern = Perl_newSVpvf(aTHX_ "(?iaa:%.*s)",
22941 (unsigned) subpattern_len,
22943 subpattern = sv_2mortal(subpattern);
22944 subpattern_re = re_compile(subpattern, 0);
22945 assert(subpattern_re); /* Should have died if didn't compile
22948 /* For each legal property value, see if the supplied pattern
22950 while (*prop_values) {
22951 const char * const entry = *prop_values;
22952 const Size_t len = strlen(entry);
22953 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
22955 if (pregexec(subpattern_re,
22957 (char *) entry + len,
22961 { /* Here, matched. Add to the returned list */
22962 Size_t total_len = j + len;
22963 SV * sub_invlist = NULL;
22964 char * this_string;
22966 /* We know this is a legal \p{property=value}. Call
22967 * the function to return the list of code points that
22969 Newxz(this_string, total_len + 1, char);
22970 Copy(lookup_name, this_string, j, char);
22971 my_strlcat(this_string, entry, total_len + 1);
22972 SAVEFREEPV(this_string);
22973 sub_invlist = parse_uniprop_string(this_string,
22982 _invlist_union(prop_definition, sub_invlist,
22986 prop_values++; /* Next iteration, look at next propvalue */
22987 } /* End of looking through property values; (the data
22988 structure is terminated by a NULL ptr) */
22990 SvREFCNT_dec_NN(subpattern_re);
22992 if (prop_definition) {
22993 return prop_definition;
22996 sv_catpvs(msg, "No Unicode property value wildcard matches:");
22997 goto append_name_to_msg;
23000 /* Here's how khw thinks we should proceed to handle the properties
23001 * not yet done: Bidi Mirroring Glyph
23002 Bidi Paired Bracket
23003 Case Folding (both full and simple)
23004 Decomposition Mapping
23005 Equivalent Unified Ideograph
23008 Lowercase Mapping (both full and simple)
23010 Titlecase Mapping (both full and simple)
23011 Uppercase Mapping (both full and simple)
23012 * Move the part that looks at the property values into a perl
23013 * script, like utf8_heavy.pl was done. This makes things somewhat
23014 * easier, but most importantly, it avoids always adding all these
23015 * strings to the memory usage when the feature is little-used.
23017 * The property values would all be concatenated into a single
23018 * string per property with each value on a separate line, and the
23019 * code point it's for on alternating lines. Then we match the
23020 * user's input pattern m//mg, without having to worry about their
23021 * uses of '^' and '$'. Only the values that aren't the default
23022 * would be in the strings. Code points would be in UTF-8. The
23023 * search pattern that we would construct would look like
23024 * (?: \n (code-point_re) \n (?aam: user-re ) \n )
23025 * And so $1 would contain the code point that matched the user-re.
23026 * For properties where the default is the code point itself, such
23027 * as any of the case changing mappings, the string would otherwise
23028 * consist of all Unicode code points in UTF-8 strung together.
23029 * This would be impractical. So instead, examine their compiled
23030 * pattern, looking at the ssc. If none, reject the pattern as an
23031 * error. Otherwise run the pattern against every code point in
23032 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
23033 * And it might be good to create an API to return the ssc.
23035 * For the name properties, a new function could be created in
23036 * charnames which essentially does the same thing as above,
23037 * sharing Name.pl with the other charname functions. Don't know
23038 * about loose name matching, or algorithmically determined names.
23039 * Decomposition.pl similarly.
23041 * It might be that a new pattern modifier would have to be
23042 * created, like /t for resTricTed, which changed the behavior of
23043 * some constructs in their subpattern, like \A. */
23044 } /* End of is a wildcard subppattern */
23047 /* Certain properties whose values are numeric need special handling.
23048 * They may optionally be prefixed by 'is'. Ignore that prefix for the
23049 * purposes of checking if this is one of those properties */
23050 if (memBEGINPs(lookup_name, j, "is")) {
23054 /* Then check if it is one of these specially-handled properties. The
23055 * possibilities are hard-coded because easier this way, and the list
23056 * is unlikely to change.
23058 * All numeric value type properties are of this ilk, and are also
23059 * special in a different way later on. So find those first. There
23060 * are several numeric value type properties in the Unihan DB (which is
23061 * unlikely to be compiled with perl, but we handle it here in case it
23062 * does get compiled). They all end with 'numeric'. The interiors
23063 * aren't checked for the precise property. This would stop working if
23064 * a cjk property were to be created that ended with 'numeric' and
23065 * wasn't a numeric type */
23066 is_nv_type = memEQs(lookup_name + lookup_offset,
23067 j - 1 - lookup_offset, "numericvalue")
23068 || memEQs(lookup_name + lookup_offset,
23069 j - 1 - lookup_offset, "nv")
23070 || ( memENDPs(lookup_name + lookup_offset,
23071 j - 1 - lookup_offset, "numeric")
23072 && ( memBEGINPs(lookup_name + lookup_offset,
23073 j - 1 - lookup_offset, "cjk")
23074 || memBEGINPs(lookup_name + lookup_offset,
23075 j - 1 - lookup_offset, "k")));
23077 || memEQs(lookup_name + lookup_offset,
23078 j - 1 - lookup_offset, "canonicalcombiningclass")
23079 || memEQs(lookup_name + lookup_offset,
23080 j - 1 - lookup_offset, "ccc")
23081 || memEQs(lookup_name + lookup_offset,
23082 j - 1 - lookup_offset, "age")
23083 || memEQs(lookup_name + lookup_offset,
23084 j - 1 - lookup_offset, "in")
23085 || memEQs(lookup_name + lookup_offset,
23086 j - 1 - lookup_offset, "presentin"))
23090 /* Since the stuff after the '=' is a number, we can't throw away
23091 * '-' willy-nilly, as those could be a minus sign. Other stricter
23092 * rules also apply. However, these properties all can have the
23093 * rhs not be a number, in which case they contain at least one
23094 * alphabetic. In those cases, the stricter rules don't apply.
23095 * But the numeric type properties can have the alphas [Ee] to
23096 * signify an exponent, and it is still a number with stricter
23097 * rules. So look for an alpha that signifies not-strict */
23099 for (k = i; k < name_len; k++) {
23100 if ( isALPHA_A(name[k])
23101 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
23111 /* A number may have a leading '+' or '-'. The latter is retained
23113 if (name[i] == '+') {
23116 else if (name[i] == '-') {
23117 lookup_name[j++] = '-';
23121 /* Skip leading zeros including single underscores separating the
23122 * zeros, or between the final leading zero and the first other
23124 for (; i < name_len - 1; i++) {
23125 if ( name[i] != '0'
23126 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23133 else { /* No '=' */
23135 /* Only a few properties without an '=' should be parsed with stricter
23136 * rules. The list is unlikely to change. */
23137 if ( memBEGINPs(lookup_name, j, "perl")
23138 && memNEs(lookup_name + 4, j - 4, "space")
23139 && memNEs(lookup_name + 4, j - 4, "word"))
23143 /* We set the inputs back to 0 and the code below will reparse,
23149 /* Here, we have either finished the property, or are positioned to parse
23150 * the remainder, and we know if stricter rules apply. Finish out, if not
23152 for (; i < name_len; i++) {
23153 char cur = name[i];
23155 /* In all instances, case differences are ignored, and we normalize to
23157 if (isUPPER_A(cur)) {
23158 lookup_name[j++] = toLOWER(cur);
23162 /* An underscore is skipped, but not under strict rules unless it
23163 * separates two digits */
23166 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
23167 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
23169 lookup_name[j++] = '_';
23174 /* Hyphens are skipped except under strict */
23175 if (cur == '-' && ! stricter) {
23179 /* XXX Bug in documentation. It says white space skipped adjacent to
23180 * non-word char. Maybe we should, but shouldn't skip it next to a dot
23182 if (isSPACE_A(cur) && ! stricter) {
23186 lookup_name[j++] = cur;
23188 /* Unless this is a non-trailing slash, we are done with it */
23189 if (i >= name_len - 1 || cur != '/') {
23195 /* A slash in the 'numeric value' property indicates that what follows
23196 * is a denominator. It can have a leading '+' and '0's that should be
23197 * skipped. But we have never allowed a negative denominator, so treat
23198 * a minus like every other character. (No need to rule out a second
23199 * '/', as that won't match anything anyway */
23202 if (i < name_len && name[i] == '+') {
23206 /* Skip leading zeros including underscores separating digits */
23207 for (; i < name_len - 1; i++) {
23208 if ( name[i] != '0'
23209 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23215 /* Store the first real character in the denominator */
23216 if (i < name_len) {
23217 lookup_name[j++] = name[i];
23222 /* Here are completely done parsing the input 'name', and 'lookup_name'
23223 * contains a copy, normalized.
23225 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
23226 * different from without the underscores. */
23227 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
23228 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
23229 && UNLIKELY(name[name_len-1] == '_'))
23231 lookup_name[j++] = '&';
23234 /* If the original input began with 'In' or 'Is', it could be a subroutine
23235 * call to a user-defined property instead of a Unicode property name. */
23236 if ( name_len - non_pkg_begin > 2
23237 && name[non_pkg_begin+0] == 'I'
23238 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
23240 /* Names that start with In have different characterstics than those
23241 * that start with Is */
23242 if (name[non_pkg_begin+1] == 's') {
23243 starts_with_Is = TRUE;
23247 could_be_user_defined = FALSE;
23250 if (could_be_user_defined) {
23253 /* If the user defined property returns the empty string, it could
23254 * easily be because the pattern is being compiled before the data it
23255 * actually needs to compile is available. This could be argued to be
23256 * a bug in the perl code, but this is a change of behavior for Perl,
23257 * so we handle it. This means that intentionally returning nothing
23258 * will not be resolved until runtime */
23259 bool empty_return = FALSE;
23261 /* Here, the name could be for a user defined property, which are
23262 * implemented as subs. */
23263 user_sub = get_cvn_flags(name, name_len, 0);
23265 const char insecure[] = "Insecure user-defined property";
23267 /* Here, there is a sub by the correct name. Normally we call it
23268 * to get the property definition */
23270 SV * user_sub_sv = MUTABLE_SV(user_sub);
23271 SV * error; /* Any error returned by calling 'user_sub' */
23272 SV * key; /* The key into the hash of user defined sub names
23275 SV ** saved_user_prop_ptr; /* Hash entry for this property */
23277 /* How many times to retry when another thread is in the middle of
23278 * expanding the same definition we want */
23279 PERL_INT_FAST8_T retry_countdown = 10;
23281 DECLARATION_FOR_GLOBAL_CONTEXT;
23283 /* If we get here, we know this property is user-defined */
23284 *user_defined_ptr = TRUE;
23286 /* We refuse to call a potentially tainted subroutine; returning an
23289 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23290 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23291 goto append_name_to_msg;
23294 /* In principal, we only call each subroutine property definition
23295 * once during the life of the program. This guarantees that the
23296 * property definition never changes. The results of the single
23297 * sub call are stored in a hash, which is used instead for future
23298 * references to this property. The property definition is thus
23299 * immutable. But, to allow the user to have a /i-dependent
23300 * definition, we call the sub once for non-/i, and once for /i,
23301 * should the need arise, passing the /i status as a parameter.
23303 * We start by constructing the hash key name, consisting of the
23304 * fully qualified subroutine name, preceded by the /i status, so
23305 * that there is a key for /i and a different key for non-/i */
23306 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
23307 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23308 non_pkg_begin != 0);
23309 sv_catsv(key, fq_name);
23312 /* We only call the sub once throughout the life of the program
23313 * (with the /i, non-/i exception noted above). That means the
23314 * hash must be global and accessible to all threads. It is
23315 * created at program start-up, before any threads are created, so
23316 * is accessible to all children. But this creates some
23319 * 1) The keys can't be shared, or else problems arise; sharing is
23320 * turned off at hash creation time
23321 * 2) All SVs in it are there for the remainder of the life of the
23322 * program, and must be created in the same interpreter context
23323 * as the hash, or else they will be freed from the wrong pool
23324 * at global destruction time. This is handled by switching to
23325 * the hash's context to create each SV going into it, and then
23326 * immediately switching back
23327 * 3) All accesses to the hash must be controlled by a mutex, to
23328 * prevent two threads from getting an unstable state should
23329 * they simultaneously be accessing it. The code below is
23330 * crafted so that the mutex is locked whenever there is an
23331 * access and unlocked only when the next stable state is
23334 * The hash stores either the definition of the property if it was
23335 * valid, or, if invalid, the error message that was raised. We
23336 * use the type of SV to distinguish.
23338 * There's also the need to guard against the definition expansion
23339 * from infinitely recursing. This is handled by storing the aTHX
23340 * of the expanding thread during the expansion. Again the SV type
23341 * is used to distinguish this from the other two cases. If we
23342 * come to here and the hash entry for this property is our aTHX,
23343 * it means we have recursed, and the code assumes that we would
23344 * infinitely recurse, so instead stops and raises an error.
23345 * (Any recursion has always been treated as infinite recursion in
23348 * If instead, the entry is for a different aTHX, it means that
23349 * that thread has gotten here first, and hasn't finished expanding
23350 * the definition yet. We just have to wait until it is done. We
23351 * sleep and retry a few times, returning an error if the other
23352 * thread doesn't complete. */
23355 USER_PROP_MUTEX_LOCK;
23357 /* If we have an entry for this key, the subroutine has already
23358 * been called once with this /i status. */
23359 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
23360 SvPVX(key), SvCUR(key), 0);
23361 if (saved_user_prop_ptr) {
23363 /* If the saved result is an inversion list, it is the valid
23364 * definition of this property */
23365 if (is_invlist(*saved_user_prop_ptr)) {
23366 prop_definition = *saved_user_prop_ptr;
23368 /* The SV in the hash won't be removed until global
23369 * destruction, so it is stable and we can unlock */
23370 USER_PROP_MUTEX_UNLOCK;
23372 /* The caller shouldn't try to free this SV */
23373 return prop_definition;
23376 /* Otherwise, if it is a string, it is the error message
23377 * that was returned when we first tried to evaluate this
23378 * property. Fail, and append the message */
23379 if (SvPOK(*saved_user_prop_ptr)) {
23380 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23381 sv_catsv(msg, *saved_user_prop_ptr);
23383 /* The SV in the hash won't be removed until global
23384 * destruction, so it is stable and we can unlock */
23385 USER_PROP_MUTEX_UNLOCK;
23390 assert(SvIOK(*saved_user_prop_ptr));
23392 /* Here, we have an unstable entry in the hash. Either another
23393 * thread is in the middle of expanding the property's
23394 * definition, or we are ourselves recursing. We use the aTHX
23395 * in it to distinguish */
23396 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
23398 /* Here, it's another thread doing the expanding. We've
23399 * looked as much as we are going to at the contents of the
23400 * hash entry. It's safe to unlock. */
23401 USER_PROP_MUTEX_UNLOCK;
23403 /* Retry a few times */
23404 if (retry_countdown-- > 0) {
23409 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23410 sv_catpvs(msg, "Timeout waiting for another thread to "
23412 goto append_name_to_msg;
23415 /* Here, we are recursing; don't dig any deeper */
23416 USER_PROP_MUTEX_UNLOCK;
23418 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23420 "Infinite recursion in user-defined property");
23421 goto append_name_to_msg;
23424 /* Here, this thread has exclusive control, and there is no entry
23425 * for this property in the hash. So we have the go ahead to
23426 * expand the definition ourselves. */
23428 PUSHSTACKi(PERLSI_MAGIC);
23431 /* Create a temporary placeholder in the hash to detect recursion
23433 SWITCH_TO_GLOBAL_CONTEXT;
23434 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
23435 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
23438 /* Now that we have a placeholder, we can let other threads
23440 USER_PROP_MUTEX_UNLOCK;
23442 /* Make sure the placeholder always gets destroyed */
23443 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
23448 /* Call the user's function, with the /i status as a parameter.
23449 * Note that we have gone to a lot of trouble to keep this call
23450 * from being within the locked mutex region. */
23451 XPUSHs(boolSV(to_fold));
23454 /* The following block was taken from swash_init(). Presumably
23455 * they apply to here as well, though we no longer use a swash --
23459 /* We might get here via a subroutine signature which uses a utf8
23460 * parameter name, at which point PL_subname will have been set
23461 * but not yet used. */
23462 save_item(PL_subname);
23464 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
23469 if (TAINT_get || SvTRUE(error)) {
23470 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23471 if (SvTRUE(error)) {
23472 sv_catpvs(msg, "Error \"");
23473 sv_catsv(msg, error);
23474 sv_catpvs(msg, "\"");
23477 if (SvTRUE(error)) sv_catpvs(msg, "; ");
23478 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23481 if (name_len > 0) {
23482 sv_catpvs(msg, " in expansion of ");
23483 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
23489 prop_definition = NULL;
23491 else { /* G_SCALAR guarantees a single return value */
23492 SV * contents = POPs;
23494 /* The contents is supposed to be the expansion of the property
23495 * definition. If the definition is deferrable, and we got an
23496 * empty string back, set a flag to later defer it (after clean
23499 && (! SvPOK(contents) || SvCUR(contents) == 0))
23501 empty_return = TRUE;
23503 else { /* Otherwise, call a function to check for valid syntax,
23506 prop_definition = handle_user_defined_property(
23508 is_utf8, to_fold, runtime,
23510 contents, user_defined_ptr,
23516 /* Here, we have the results of the expansion. Delete the
23517 * placeholder, and if the definition is now known, replace it with
23518 * that definition. We need exclusive access to the hash, and we
23519 * can't let anyone else in, between when we delete the placeholder
23520 * and add the permanent entry */
23521 USER_PROP_MUTEX_LOCK;
23523 S_delete_recursion_entry(aTHX_ SvPVX(key));
23525 if ( ! empty_return
23526 && (! prop_definition || is_invlist(prop_definition)))
23528 /* If we got success we use the inversion list defining the
23529 * property; otherwise use the error message */
23530 SWITCH_TO_GLOBAL_CONTEXT;
23531 (void) hv_store_ent(PL_user_def_props,
23534 ? newSVsv(prop_definition)
23540 /* All done, and the hash now has a permanent entry for this
23541 * property. Give up exclusive control */
23542 USER_PROP_MUTEX_UNLOCK;
23548 if (empty_return) {
23549 goto definition_deferred;
23552 if (prop_definition) {
23554 /* If the definition is for something not known at this time,
23555 * we toss it, and go return the main property name, as that's
23556 * the one the user will be aware of */
23557 if (! is_invlist(prop_definition)) {
23558 SvREFCNT_dec_NN(prop_definition);
23559 goto definition_deferred;
23562 sv_2mortal(prop_definition);
23566 return prop_definition;
23568 } /* End of calling the subroutine for the user-defined property */
23569 } /* End of it could be a user-defined property */
23571 /* Here it wasn't a user-defined property that is known at this time. See
23572 * if it is a Unicode property */
23574 lookup_len = j; /* This is a more mnemonic name than 'j' */
23576 /* Get the index into our pointer table of the inversion list corresponding
23577 * to the property */
23578 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23580 /* If it didn't find the property ... */
23581 if (table_index == 0) {
23583 /* Try again stripping off any initial 'Is'. This is because we
23584 * promise that an initial Is is optional. The same isn't true of
23585 * names that start with 'In'. Those can match only blocks, and the
23586 * lookup table already has those accounted for. */
23587 if (starts_with_Is) {
23593 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23596 if (table_index == 0) {
23599 /* Here, we didn't find it. If not a numeric type property, and
23600 * can't be a user-defined one, it isn't a legal property */
23601 if (! is_nv_type) {
23602 if (! could_be_user_defined) {
23606 /* Here, the property name is legal as a user-defined one. At
23607 * compile time, it might just be that the subroutine for that
23608 * property hasn't been encountered yet, but at runtime, it's
23609 * an error to try to use an undefined one */
23610 if (! deferrable) {
23611 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23612 sv_catpvs(msg, "Unknown user-defined property name");
23613 goto append_name_to_msg;
23616 goto definition_deferred;
23617 } /* End of isn't a numeric type property */
23619 /* The numeric type properties need more work to decide. What we
23620 * do is make sure we have the number in canonical form and look
23623 if (slash_pos < 0) { /* No slash */
23625 /* When it isn't a rational, take the input, convert it to a
23626 * NV, then create a canonical string representation of that
23630 SSize_t value_len = lookup_len - equals_pos;
23632 /* Get the value */
23633 if ( value_len <= 0
23634 || my_atof3(lookup_name + equals_pos, &value,
23636 != lookup_name + lookup_len)
23641 /* If the value is an integer, the canonical value is integral
23643 if (Perl_ceil(value) == value) {
23644 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
23645 equals_pos, lookup_name, value);
23647 else { /* Otherwise, it is %e with a known precision */
23650 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
23651 equals_pos, lookup_name,
23652 PL_E_FORMAT_PRECISION, value);
23654 /* The exponent generated is expecting two digits, whereas
23655 * %e on some systems will generate three. Remove leading
23656 * zeros in excess of 2 from the exponent. We start
23657 * looking for them after the '=' */
23658 exp_ptr = strchr(canonical + equals_pos, 'e');
23660 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
23661 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
23663 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
23665 if (excess_exponent_len > 0) {
23666 SSize_t leading_zeros = strspn(cur_ptr, "0");
23667 SSize_t excess_leading_zeros
23668 = MIN(leading_zeros, excess_exponent_len);
23669 if (excess_leading_zeros > 0) {
23670 Move(cur_ptr + excess_leading_zeros,
23672 strlen(cur_ptr) - excess_leading_zeros
23673 + 1, /* Copy the NUL as well */
23680 else { /* Has a slash. Create a rational in canonical form */
23681 UV numerator, denominator, gcd, trial;
23682 const char * end_ptr;
23683 const char * sign = "";
23685 /* We can't just find the numerator, denominator, and do the
23686 * division, then use the method above, because that is
23687 * inexact. And the input could be a rational that is within
23688 * epsilon (given our precision) of a valid rational, and would
23689 * then incorrectly compare valid.
23691 * We're only interested in the part after the '=' */
23692 const char * this_lookup_name = lookup_name + equals_pos;
23693 lookup_len -= equals_pos;
23694 slash_pos -= equals_pos;
23696 /* Handle any leading minus */
23697 if (this_lookup_name[0] == '-') {
23699 this_lookup_name++;
23704 /* Convert the numerator to numeric */
23705 end_ptr = this_lookup_name + slash_pos;
23706 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
23710 /* It better have included all characters before the slash */
23711 if (*end_ptr != '/') {
23715 /* Set to look at just the denominator */
23716 this_lookup_name += slash_pos;
23717 lookup_len -= slash_pos;
23718 end_ptr = this_lookup_name + lookup_len;
23720 /* Convert the denominator to numeric */
23721 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
23725 /* It better be the rest of the characters, and don't divide by
23727 if ( end_ptr != this_lookup_name + lookup_len
23728 || denominator == 0)
23733 /* Get the greatest common denominator using
23734 http://en.wikipedia.org/wiki/Euclidean_algorithm */
23736 trial = denominator;
23737 while (trial != 0) {
23739 trial = gcd % trial;
23743 /* If already in lowest possible terms, we have already tried
23744 * looking this up */
23749 /* Reduce the rational, which should put it in canonical form
23752 denominator /= gcd;
23754 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
23755 equals_pos, lookup_name, sign, numerator, denominator);
23758 /* Here, we have the number in canonical form. Try that */
23759 table_index = match_uniprop((U8 *) canonical, strlen(canonical));
23760 if (table_index == 0) {
23763 } /* End of still didn't find the property in our table */
23764 } /* End of didn't find the property in our table */
23766 /* Here, we have a non-zero return, which is an index into a table of ptrs.
23767 * A negative return signifies that the real index is the absolute value,
23768 * but the result needs to be inverted */
23769 if (table_index < 0) {
23770 invert_return = TRUE;
23771 table_index = -table_index;
23774 /* Out-of band indices indicate a deprecated property. The proper index is
23775 * modulo it with the table size. And dividing by the table size yields
23776 * an offset into a table constructed by regen/mk_invlists.pl to contain
23777 * the corresponding warning message */
23778 if (table_index > MAX_UNI_KEYWORD_INDEX) {
23779 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
23780 table_index %= MAX_UNI_KEYWORD_INDEX;
23781 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
23782 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
23783 (int) name_len, name, deprecated_property_msgs[warning_offset]);
23786 /* In a few properties, a different property is used under /i. These are
23787 * unlikely to change, so are hard-coded here. */
23789 if ( table_index == UNI_XPOSIXUPPER
23790 || table_index == UNI_XPOSIXLOWER
23791 || table_index == UNI_TITLE)
23793 table_index = UNI_CASED;
23795 else if ( table_index == UNI_UPPERCASELETTER
23796 || table_index == UNI_LOWERCASELETTER
23797 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
23798 || table_index == UNI_TITLECASELETTER
23801 table_index = UNI_CASEDLETTER;
23803 else if ( table_index == UNI_POSIXUPPER
23804 || table_index == UNI_POSIXLOWER)
23806 table_index = UNI_POSIXALPHA;
23810 /* Create and return the inversion list */
23811 prop_definition =_new_invlist_C_array(uni_prop_ptrs[table_index]);
23812 sv_2mortal(prop_definition);
23815 /* See if there is a private use override to add to this definition */
23817 COPHH * hinthash = (IN_PERL_COMPILETIME)
23818 ? CopHINTHASH_get(&PL_compiling)
23819 : CopHINTHASH_get(PL_curcop);
23820 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
23822 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
23824 /* See if there is an element in the hints hash for this table */
23825 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
23826 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
23830 SV * pu_definition;
23832 SV * expanded_prop_definition =
23833 sv_2mortal(invlist_clone(prop_definition, NULL));
23835 /* If so, it's definition is the string from here to the next
23836 * \a character. And its format is the same as a user-defined
23838 pos += SvCUR(pu_lookup);
23839 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
23840 pu_invlist = handle_user_defined_property(lookup_name,
23843 0, /* Not folded */
23851 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23852 sv_catpvs(msg, "Insecure private-use override");
23853 goto append_name_to_msg;
23856 /* For now, as a safety measure, make sure that it doesn't
23857 * override non-private use code points */
23858 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
23860 /* Add it to the list to be returned */
23861 _invlist_union(prop_definition, pu_invlist,
23862 &expanded_prop_definition);
23863 prop_definition = expanded_prop_definition;
23864 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
23869 if (invert_return) {
23870 _invlist_invert(prop_definition);
23872 return prop_definition;
23876 if (non_pkg_begin != 0) {
23877 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23878 sv_catpvs(msg, "Illegal user-defined property name");
23881 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23882 sv_catpvs(msg, "Can't find Unicode property definition");
23886 append_name_to_msg:
23888 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
23889 const char * suffix = (runtime && level == 0) ? "}" : "\"";
23891 sv_catpv(msg, prefix);
23892 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23893 sv_catpv(msg, suffix);
23898 definition_deferred:
23900 /* Here it could yet to be defined, so defer evaluation of this
23901 * until its needed at runtime. We need the fully qualified property name
23902 * to avoid ambiguity, and a trailing newline */
23904 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23905 non_pkg_begin != 0 /* If has "::" */
23908 sv_catpvs(fq_name, "\n");
23910 *user_defined_ptr = TRUE;
23917 * ex: set ts=8 sts=4 sw=4 et: