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
4571 look into 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;
6181 else if (OP(scan) == OPEN) {
6182 if (stopparen != (I32)ARG(scan))
6185 else if (OP(scan) == CLOSE) {
6186 if (stopparen == (I32)ARG(scan)) {
6189 if ((I32)ARG(scan) == is_par) {
6190 next = regnext(scan);
6192 if ( next && (OP(next) != WHILEM) && next < last)
6193 is_par = 0; /* Disable optimization */
6196 *(data->last_closep) = ARG(scan);
6198 else if (OP(scan) == EVAL) {
6200 data->flags |= SF_HAS_EVAL;
6202 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6203 if (flags & SCF_DO_SUBSTR) {
6204 scan_commit(pRExC_state, data, minlenp, is_inf);
6205 flags &= ~SCF_DO_SUBSTR;
6207 if (data && OP(scan)==ACCEPT) {
6208 data->flags |= SCF_SEEN_ACCEPT;
6213 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6215 if (flags & SCF_DO_SUBSTR) {
6216 scan_commit(pRExC_state, data, minlenp, is_inf);
6217 data->cur_is_floating = 1; /* float */
6219 is_inf = is_inf_internal = 1;
6220 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6221 ssc_anything(data->start_class);
6222 flags &= ~SCF_DO_STCLASS;
6224 else if (OP(scan) == GPOS) {
6225 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6226 !(delta || is_inf || (data && data->pos_delta)))
6228 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6229 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6230 if (RExC_rx->gofs < (STRLEN)min)
6231 RExC_rx->gofs = min;
6233 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6237 #ifdef TRIE_STUDY_OPT
6238 #ifdef FULL_TRIE_STUDY
6239 else if (PL_regkind[OP(scan)] == TRIE) {
6240 /* NOTE - There is similar code to this block above for handling
6241 BRANCH nodes on the initial study. If you change stuff here
6243 regnode *trie_node= scan;
6244 regnode *tail= regnext(scan);
6245 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6246 SSize_t max1 = 0, min1 = SSize_t_MAX;
6249 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6250 /* Cannot merge strings after this. */
6251 scan_commit(pRExC_state, data, minlenp, is_inf);
6253 if (flags & SCF_DO_STCLASS)
6254 ssc_init_zero(pRExC_state, &accum);
6260 const regnode *nextbranch= NULL;
6263 for ( word=1 ; word <= trie->wordcount ; word++)
6265 SSize_t deltanext=0, minnext=0, f = 0, fake;
6266 regnode_ssc this_class;
6268 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6270 data_fake.whilem_c = data->whilem_c;
6271 data_fake.last_closep = data->last_closep;
6274 data_fake.last_closep = &fake;
6275 data_fake.pos_delta = delta;
6276 if (flags & SCF_DO_STCLASS) {
6277 ssc_init(pRExC_state, &this_class);
6278 data_fake.start_class = &this_class;
6279 f = SCF_DO_STCLASS_AND;
6281 if (flags & SCF_WHILEM_VISITED_POS)
6282 f |= SCF_WHILEM_VISITED_POS;
6284 if (trie->jump[word]) {
6286 nextbranch = trie_node + trie->jump[0];
6287 scan= trie_node + trie->jump[word];
6288 /* We go from the jump point to the branch that follows
6289 it. Note this means we need the vestigal unused
6290 branches even though they arent otherwise used. */
6291 /* optimise study_chunk() for TRIE */
6292 minnext = study_chunk(pRExC_state, &scan, minlenp,
6293 &deltanext, (regnode *)nextbranch, &data_fake,
6294 stopparen, recursed_depth, NULL, f, depth+1);
6296 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6297 nextbranch= regnext((regnode*)nextbranch);
6299 if (min1 > (SSize_t)(minnext + trie->minlen))
6300 min1 = minnext + trie->minlen;
6301 if (deltanext == SSize_t_MAX) {
6302 is_inf = is_inf_internal = 1;
6304 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6305 max1 = minnext + deltanext + trie->maxlen;
6307 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6309 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6310 if ( stopmin > min + min1)
6311 stopmin = min + min1;
6312 flags &= ~SCF_DO_SUBSTR;
6314 data->flags |= SCF_SEEN_ACCEPT;
6317 if (data_fake.flags & SF_HAS_EVAL)
6318 data->flags |= SF_HAS_EVAL;
6319 data->whilem_c = data_fake.whilem_c;
6321 if (flags & SCF_DO_STCLASS)
6322 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6325 if (flags & SCF_DO_SUBSTR) {
6326 data->pos_min += min1;
6327 data->pos_delta += max1 - min1;
6328 if (max1 != min1 || is_inf)
6329 data->cur_is_floating = 1; /* float */
6332 if (delta != SSize_t_MAX) {
6333 if (SSize_t_MAX - (max1 - min1) >= delta)
6334 delta += max1 - min1;
6336 delta = SSize_t_MAX;
6338 if (flags & SCF_DO_STCLASS_OR) {
6339 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6341 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6342 flags &= ~SCF_DO_STCLASS;
6345 else if (flags & SCF_DO_STCLASS_AND) {
6347 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6348 flags &= ~SCF_DO_STCLASS;
6351 /* Switch to OR mode: cache the old value of
6352 * data->start_class */
6354 StructCopy(data->start_class, and_withp, regnode_ssc);
6355 flags &= ~SCF_DO_STCLASS_AND;
6356 StructCopy(&accum, data->start_class, regnode_ssc);
6357 flags |= SCF_DO_STCLASS_OR;
6364 else if (PL_regkind[OP(scan)] == TRIE) {
6365 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6368 min += trie->minlen;
6369 delta += (trie->maxlen - trie->minlen);
6370 flags &= ~SCF_DO_STCLASS; /* xxx */
6371 if (flags & SCF_DO_SUBSTR) {
6372 /* Cannot expect anything... */
6373 scan_commit(pRExC_state, data, minlenp, is_inf);
6374 data->pos_min += trie->minlen;
6375 data->pos_delta += (trie->maxlen - trie->minlen);
6376 if (trie->maxlen != trie->minlen)
6377 data->cur_is_floating = 1; /* float */
6379 if (trie->jump) /* no more substrings -- for now /grr*/
6380 flags &= ~SCF_DO_SUBSTR;
6382 #endif /* old or new */
6383 #endif /* TRIE_STUDY_OPT */
6385 /* Else: zero-length, ignore. */
6386 scan = regnext(scan);
6391 /* we need to unwind recursion. */
6394 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6395 DEBUG_PEEP("fend", scan, depth, flags);
6397 /* restore previous context */
6398 last = frame->last_regnode;
6399 scan = frame->next_regnode;
6400 stopparen = frame->stopparen;
6401 recursed_depth = frame->prev_recursed_depth;
6403 RExC_frame_last = frame->prev_frame;
6404 frame = frame->this_prev_frame;
6405 goto fake_study_recurse;
6409 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6412 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6414 if (flags & SCF_DO_SUBSTR && is_inf)
6415 data->pos_delta = SSize_t_MAX - data->pos_min;
6416 if (is_par > (I32)U8_MAX)
6418 if (is_par && pars==1 && data) {
6419 data->flags |= SF_IN_PAR;
6420 data->flags &= ~SF_HAS_PAR;
6422 else if (pars && data) {
6423 data->flags |= SF_HAS_PAR;
6424 data->flags &= ~SF_IN_PAR;
6426 if (flags & SCF_DO_STCLASS_OR)
6427 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6428 if (flags & SCF_TRIE_RESTUDY)
6429 data->flags |= SCF_TRIE_RESTUDY;
6431 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6434 SSize_t final_minlen= min < stopmin ? min : stopmin;
6436 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6437 if (final_minlen > SSize_t_MAX - delta)
6438 RExC_maxlen = SSize_t_MAX;
6439 else if (RExC_maxlen < final_minlen + delta)
6440 RExC_maxlen = final_minlen + delta;
6442 return final_minlen;
6444 NOT_REACHED; /* NOTREACHED */
6448 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6450 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6452 PERL_ARGS_ASSERT_ADD_DATA;
6454 Renewc(RExC_rxi->data,
6455 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6456 char, struct reg_data);
6458 Renew(RExC_rxi->data->what, count + n, U8);
6460 Newx(RExC_rxi->data->what, n, U8);
6461 RExC_rxi->data->count = count + n;
6462 Copy(s, RExC_rxi->data->what + count, n, U8);
6466 /*XXX: todo make this not included in a non debugging perl, but appears to be
6467 * used anyway there, in 'use re' */
6468 #ifndef PERL_IN_XSUB_RE
6470 Perl_reginitcolors(pTHX)
6472 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6474 char *t = savepv(s);
6478 t = strchr(t, '\t');
6484 PL_colors[i] = t = (char *)"";
6489 PL_colors[i++] = (char *)"";
6496 #ifdef TRIE_STUDY_OPT
6497 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6500 (data.flags & SCF_TRIE_RESTUDY) \
6508 #define CHECK_RESTUDY_GOTO_butfirst
6512 * pregcomp - compile a regular expression into internal code
6514 * Decides which engine's compiler to call based on the hint currently in
6518 #ifndef PERL_IN_XSUB_RE
6520 /* return the currently in-scope regex engine (or the default if none) */
6522 regexp_engine const *
6523 Perl_current_re_engine(pTHX)
6525 if (IN_PERL_COMPILETIME) {
6526 HV * const table = GvHV(PL_hintgv);
6529 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6530 return &PL_core_reg_engine;
6531 ptr = hv_fetchs(table, "regcomp", FALSE);
6532 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6533 return &PL_core_reg_engine;
6534 return INT2PTR(regexp_engine*, SvIV(*ptr));
6538 if (!PL_curcop->cop_hints_hash)
6539 return &PL_core_reg_engine;
6540 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6541 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6542 return &PL_core_reg_engine;
6543 return INT2PTR(regexp_engine*, SvIV(ptr));
6549 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6551 regexp_engine const *eng = current_re_engine();
6552 GET_RE_DEBUG_FLAGS_DECL;
6554 PERL_ARGS_ASSERT_PREGCOMP;
6556 /* Dispatch a request to compile a regexp to correct regexp engine. */
6558 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6561 return CALLREGCOMP_ENG(eng, pattern, flags);
6565 /* public(ish) entry point for the perl core's own regex compiling code.
6566 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6567 * pattern rather than a list of OPs, and uses the internal engine rather
6568 * than the current one */
6571 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6573 SV *pat = pattern; /* defeat constness! */
6574 PERL_ARGS_ASSERT_RE_COMPILE;
6575 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6576 #ifdef PERL_IN_XSUB_RE
6579 &PL_core_reg_engine,
6581 NULL, NULL, rx_flags, 0);
6586 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6590 if (--cbs->refcnt > 0)
6592 for (n = 0; n < cbs->count; n++) {
6593 REGEXP *rx = cbs->cb[n].src_regex;
6595 cbs->cb[n].src_regex = NULL;
6596 SvREFCNT_dec_NN(rx);
6604 static struct reg_code_blocks *
6605 S_alloc_code_blocks(pTHX_ int ncode)
6607 struct reg_code_blocks *cbs;
6608 Newx(cbs, 1, struct reg_code_blocks);
6611 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6613 Newx(cbs->cb, ncode, struct reg_code_block);
6620 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6621 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6622 * point to the realloced string and length.
6624 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6628 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6629 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6631 U8 *const src = (U8*)*pat_p;
6636 GET_RE_DEBUG_FLAGS_DECL;
6638 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6639 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6641 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6642 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6645 while (s < *plen_p) {
6646 append_utf8_from_native_byte(src[s], &d);
6648 if (n < num_code_blocks) {
6649 assert(pRExC_state->code_blocks);
6650 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6651 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6652 assert(*(d - 1) == '(');
6655 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6656 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6657 assert(*(d - 1) == ')');
6666 *pat_p = (char*) dst;
6668 RExC_orig_utf8 = RExC_utf8 = 1;
6673 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6674 * while recording any code block indices, and handling overloading,
6675 * nested qr// objects etc. If pat is null, it will allocate a new
6676 * string, or just return the first arg, if there's only one.
6678 * Returns the malloced/updated pat.
6679 * patternp and pat_count is the array of SVs to be concatted;
6680 * oplist is the optional list of ops that generated the SVs;
6681 * recompile_p is a pointer to a boolean that will be set if
6682 * the regex will need to be recompiled.
6683 * delim, if non-null is an SV that will be inserted between each element
6687 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6688 SV *pat, SV ** const patternp, int pat_count,
6689 OP *oplist, bool *recompile_p, SV *delim)
6693 bool use_delim = FALSE;
6694 bool alloced = FALSE;
6696 /* if we know we have at least two args, create an empty string,
6697 * then concatenate args to that. For no args, return an empty string */
6698 if (!pat && pat_count != 1) {
6704 for (svp = patternp; svp < patternp + pat_count; svp++) {
6707 STRLEN orig_patlen = 0;
6709 SV *msv = use_delim ? delim : *svp;
6710 if (!msv) msv = &PL_sv_undef;
6712 /* if we've got a delimiter, we go round the loop twice for each
6713 * svp slot (except the last), using the delimiter the second
6722 if (SvTYPE(msv) == SVt_PVAV) {
6723 /* we've encountered an interpolated array within
6724 * the pattern, e.g. /...@a..../. Expand the list of elements,
6725 * then recursively append elements.
6726 * The code in this block is based on S_pushav() */
6728 AV *const av = (AV*)msv;
6729 const SSize_t maxarg = AvFILL(av) + 1;
6733 assert(oplist->op_type == OP_PADAV
6734 || oplist->op_type == OP_RV2AV);
6735 oplist = OpSIBLING(oplist);
6738 if (SvRMAGICAL(av)) {
6741 Newx(array, maxarg, SV*);
6743 for (i=0; i < maxarg; i++) {
6744 SV ** const svp = av_fetch(av, i, FALSE);
6745 array[i] = svp ? *svp : &PL_sv_undef;
6749 array = AvARRAY(av);
6751 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6752 array, maxarg, NULL, recompile_p,
6754 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6760 /* we make the assumption here that each op in the list of
6761 * op_siblings maps to one SV pushed onto the stack,
6762 * except for code blocks, with have both an OP_NULL and
6764 * This allows us to match up the list of SVs against the
6765 * list of OPs to find the next code block.
6767 * Note that PUSHMARK PADSV PADSV ..
6769 * PADRANGE PADSV PADSV ..
6770 * so the alignment still works. */
6773 if (oplist->op_type == OP_NULL
6774 && (oplist->op_flags & OPf_SPECIAL))
6776 assert(n < pRExC_state->code_blocks->count);
6777 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6778 pRExC_state->code_blocks->cb[n].block = oplist;
6779 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6782 oplist = OpSIBLING(oplist); /* skip CONST */
6785 oplist = OpSIBLING(oplist);;
6788 /* apply magic and QR overloading to arg */
6791 if (SvROK(msv) && SvAMAGIC(msv)) {
6792 SV *sv = AMG_CALLunary(msv, regexp_amg);
6796 if (SvTYPE(sv) != SVt_REGEXP)
6797 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6802 /* try concatenation overload ... */
6803 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6804 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6807 /* overloading involved: all bets are off over literal
6808 * code. Pretend we haven't seen it */
6810 pRExC_state->code_blocks->count -= n;
6814 /* ... or failing that, try "" overload */
6815 while (SvAMAGIC(msv)
6816 && (sv = AMG_CALLunary(msv, string_amg))
6820 && SvRV(msv) == SvRV(sv))
6825 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6829 /* this is a partially unrolled
6830 * sv_catsv_nomg(pat, msv);
6831 * that allows us to adjust code block indices if
6834 char *dst = SvPV_force_nomg(pat, dlen);
6836 if (SvUTF8(msv) && !SvUTF8(pat)) {
6837 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6838 sv_setpvn(pat, dst, dlen);
6841 sv_catsv_nomg(pat, msv);
6845 /* We have only one SV to process, but we need to verify
6846 * it is properly null terminated or we will fail asserts
6847 * later. In theory we probably shouldn't get such SV's,
6848 * but if we do we should handle it gracefully. */
6849 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6850 /* not a string, or a string with a trailing null */
6853 /* a string with no trailing null, we need to copy it
6854 * so it has a trailing null */
6855 pat = sv_2mortal(newSVsv(msv));
6860 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6863 /* extract any code blocks within any embedded qr//'s */
6864 if (rx && SvTYPE(rx) == SVt_REGEXP
6865 && RX_ENGINE((REGEXP*)rx)->op_comp)
6868 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6869 if (ri->code_blocks && ri->code_blocks->count) {
6871 /* the presence of an embedded qr// with code means
6872 * we should always recompile: the text of the
6873 * qr// may not have changed, but it may be a
6874 * different closure than last time */
6876 if (pRExC_state->code_blocks) {
6877 int new_count = pRExC_state->code_blocks->count
6878 + ri->code_blocks->count;
6879 Renew(pRExC_state->code_blocks->cb,
6880 new_count, struct reg_code_block);
6881 pRExC_state->code_blocks->count = new_count;
6884 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6885 ri->code_blocks->count);
6887 for (i=0; i < ri->code_blocks->count; i++) {
6888 struct reg_code_block *src, *dst;
6889 STRLEN offset = orig_patlen
6890 + ReANY((REGEXP *)rx)->pre_prefix;
6891 assert(n < pRExC_state->code_blocks->count);
6892 src = &ri->code_blocks->cb[i];
6893 dst = &pRExC_state->code_blocks->cb[n];
6894 dst->start = src->start + offset;
6895 dst->end = src->end + offset;
6896 dst->block = src->block;
6897 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6906 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6915 /* see if there are any run-time code blocks in the pattern.
6916 * False positives are allowed */
6919 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6920 char *pat, STRLEN plen)
6925 PERL_UNUSED_CONTEXT;
6927 for (s = 0; s < plen; s++) {
6928 if ( pRExC_state->code_blocks
6929 && n < pRExC_state->code_blocks->count
6930 && s == pRExC_state->code_blocks->cb[n].start)
6932 s = pRExC_state->code_blocks->cb[n].end;
6936 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6938 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6940 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6947 /* Handle run-time code blocks. We will already have compiled any direct
6948 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6949 * copy of it, but with any literal code blocks blanked out and
6950 * appropriate chars escaped; then feed it into
6952 * eval "qr'modified_pattern'"
6956 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6960 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6962 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6963 * and merge them with any code blocks of the original regexp.
6965 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6966 * instead, just save the qr and return FALSE; this tells our caller that
6967 * the original pattern needs upgrading to utf8.
6971 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6972 char *pat, STRLEN plen)
6976 GET_RE_DEBUG_FLAGS_DECL;
6978 if (pRExC_state->runtime_code_qr) {
6979 /* this is the second time we've been called; this should
6980 * only happen if the main pattern got upgraded to utf8
6981 * during compilation; re-use the qr we compiled first time
6982 * round (which should be utf8 too)
6984 qr = pRExC_state->runtime_code_qr;
6985 pRExC_state->runtime_code_qr = NULL;
6986 assert(RExC_utf8 && SvUTF8(qr));
6992 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
6996 /* determine how many extra chars we need for ' and \ escaping */
6997 for (s = 0; s < plen; s++) {
6998 if (pat[s] == '\'' || pat[s] == '\\')
7002 Newx(newpat, newlen, char);
7004 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7006 for (s = 0; s < plen; s++) {
7007 if ( pRExC_state->code_blocks
7008 && n < pRExC_state->code_blocks->count
7009 && s == pRExC_state->code_blocks->cb[n].start)
7011 /* blank out literal code block so that they aren't
7012 * recompiled: eg change from/to:
7022 assert(pat[s] == '(');
7023 assert(pat[s+1] == '?');
7027 while (s < pRExC_state->code_blocks->cb[n].end) {
7035 if (pat[s] == '\'' || pat[s] == '\\')
7040 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7042 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7048 Perl_re_printf( aTHX_
7049 "%sre-parsing pattern for runtime code:%s %s\n",
7050 PL_colors[4], PL_colors[5], newpat);
7053 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7059 PUSHSTACKi(PERLSI_REQUIRE);
7060 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7061 * parsing qr''; normally only q'' does this. It also alters
7063 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7064 SvREFCNT_dec_NN(sv);
7069 SV * const errsv = ERRSV;
7070 if (SvTRUE_NN(errsv))
7071 /* use croak_sv ? */
7072 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7074 assert(SvROK(qr_ref));
7076 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7077 /* the leaving below frees the tmp qr_ref.
7078 * Give qr a life of its own */
7086 if (!RExC_utf8 && SvUTF8(qr)) {
7087 /* first time through; the pattern got upgraded; save the
7088 * qr for the next time through */
7089 assert(!pRExC_state->runtime_code_qr);
7090 pRExC_state->runtime_code_qr = qr;
7095 /* extract any code blocks within the returned qr// */
7098 /* merge the main (r1) and run-time (r2) code blocks into one */
7100 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7101 struct reg_code_block *new_block, *dst;
7102 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7106 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7108 SvREFCNT_dec_NN(qr);
7112 if (!r1->code_blocks)
7113 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7115 r1c = r1->code_blocks->count;
7116 r2c = r2->code_blocks->count;
7118 Newx(new_block, r1c + r2c, struct reg_code_block);
7122 while (i1 < r1c || i2 < r2c) {
7123 struct reg_code_block *src;
7127 src = &r2->code_blocks->cb[i2++];
7131 src = &r1->code_blocks->cb[i1++];
7132 else if ( r1->code_blocks->cb[i1].start
7133 < r2->code_blocks->cb[i2].start)
7135 src = &r1->code_blocks->cb[i1++];
7136 assert(src->end < r2->code_blocks->cb[i2].start);
7139 assert( r1->code_blocks->cb[i1].start
7140 > r2->code_blocks->cb[i2].start);
7141 src = &r2->code_blocks->cb[i2++];
7143 assert(src->end < r1->code_blocks->cb[i1].start);
7146 assert(pat[src->start] == '(');
7147 assert(pat[src->end] == ')');
7148 dst->start = src->start;
7149 dst->end = src->end;
7150 dst->block = src->block;
7151 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7155 r1->code_blocks->count += r2c;
7156 Safefree(r1->code_blocks->cb);
7157 r1->code_blocks->cb = new_block;
7160 SvREFCNT_dec_NN(qr);
7166 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7167 struct reg_substr_datum *rsd,
7168 struct scan_data_substrs *sub,
7169 STRLEN longest_length)
7171 /* This is the common code for setting up the floating and fixed length
7172 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7173 * as to whether succeeded or not */
7177 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7178 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7180 if (! (longest_length
7181 || (eol /* Can't have SEOL and MULTI */
7182 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7184 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7185 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7190 /* copy the information about the longest from the reg_scan_data
7191 over to the program. */
7192 if (SvUTF8(sub->str)) {
7194 rsd->utf8_substr = sub->str;
7196 rsd->substr = sub->str;
7197 rsd->utf8_substr = NULL;
7199 /* end_shift is how many chars that must be matched that
7200 follow this item. We calculate it ahead of time as once the
7201 lookbehind offset is added in we lose the ability to correctly
7203 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7204 rsd->end_shift = ml - sub->min_offset
7206 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7208 + (SvTAIL(sub->str) != 0)
7212 t = (eol/* Can't have SEOL and MULTI */
7213 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7214 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7220 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7222 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7223 * properly wrapped with the right modifiers */
7225 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7226 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7227 != REGEX_DEPENDS_CHARSET);
7229 /* The caret is output if there are any defaults: if not all the STD
7230 * flags are set, or if no character set specifier is needed */
7232 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7234 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7235 == REG_RUN_ON_COMMENT_SEEN);
7236 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7237 >> RXf_PMf_STD_PMMOD_SHIFT);
7238 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7240 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7242 /* We output all the necessary flags; we never output a minus, as all
7243 * those are defaults, so are
7244 * covered by the caret */
7245 const STRLEN wraplen = pat_len + has_p + has_runon
7246 + has_default /* If needs a caret */
7247 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7249 /* If needs a character set specifier */
7250 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7251 + (sizeof("(?:)") - 1);
7253 PERL_ARGS_ASSERT_SET_REGEX_PV;
7255 /* make sure PL_bitcount bounds not exceeded */
7256 assert(sizeof(STD_PAT_MODS) <= 8);
7258 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7261 SvFLAGS(Rx) |= SVf_UTF8;
7264 /* If a default, cover it using the caret */
7266 *p++= DEFAULT_PAT_MOD;
7272 name = get_regex_charset_name(RExC_rx->extflags, &len);
7273 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7275 name = UNICODE_PAT_MODS;
7276 len = sizeof(UNICODE_PAT_MODS) - 1;
7278 Copy(name, p, len, char);
7282 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7285 while((ch = *fptr++)) {
7293 Copy(RExC_precomp, p, pat_len, char);
7294 assert ((RX_WRAPPED(Rx) - p) < 16);
7295 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7298 /* Adding a trailing \n causes this to compile properly:
7299 my $R = qr / A B C # D E/x; /($R)/
7300 Otherwise the parens are considered part of the comment */
7305 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7309 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7310 * regular expression into internal code.
7311 * The pattern may be passed either as:
7312 * a list of SVs (patternp plus pat_count)
7313 * a list of OPs (expr)
7314 * If both are passed, the SV list is used, but the OP list indicates
7315 * which SVs are actually pre-compiled code blocks
7317 * The SVs in the list have magic and qr overloading applied to them (and
7318 * the list may be modified in-place with replacement SVs in the latter
7321 * If the pattern hasn't changed from old_re, then old_re will be
7324 * eng is the current engine. If that engine has an op_comp method, then
7325 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7326 * do the initial concatenation of arguments and pass on to the external
7329 * If is_bare_re is not null, set it to a boolean indicating whether the
7330 * arg list reduced (after overloading) to a single bare regex which has
7331 * been returned (i.e. /$qr/).
7333 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7335 * pm_flags contains the PMf_* flags, typically based on those from the
7336 * pm_flags field of the related PMOP. Currently we're only interested in
7337 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
7339 * For many years this code had an initial sizing pass that calculated
7340 * (sometimes incorrectly, leading to security holes) the size needed for the
7341 * compiled pattern. That was changed by commit
7342 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7343 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7344 * references to this sizing pass.
7346 * Now, an initial crude guess as to the size needed is made, based on the
7347 * length of the pattern. Patches welcome to improve that guess. That amount
7348 * of space is malloc'd and then immediately freed, and then clawed back node
7349 * by node. This design is to minimze, to the extent possible, memory churn
7350 * when doing the the reallocs.
7352 * A separate parentheses counting pass may be needed in some cases.
7353 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7356 * The existence of a sizing pass necessitated design decisions that are no
7357 * longer needed. There are potential areas of simplification.
7359 * Beware that the optimization-preparation code in here knows about some
7360 * of the structure of the compiled regexp. [I'll say.]
7364 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7365 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7366 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7369 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7377 SV** new_patternp = patternp;
7379 /* these are all flags - maybe they should be turned
7380 * into a single int with different bit masks */
7381 I32 sawlookahead = 0;
7386 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7388 bool runtime_code = 0;
7390 RExC_state_t RExC_state;
7391 RExC_state_t * const pRExC_state = &RExC_state;
7392 #ifdef TRIE_STUDY_OPT
7394 RExC_state_t copyRExC_state;
7396 GET_RE_DEBUG_FLAGS_DECL;
7398 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7400 DEBUG_r(if (!PL_colorset) reginitcolors());
7402 /* Initialize these here instead of as-needed, as is quick and avoids
7403 * having to test them each time otherwise */
7404 if (! PL_InBitmap) {
7406 char * dump_len_string;
7409 /* This is calculated here, because the Perl program that generates the
7410 * static global ones doesn't currently have access to
7411 * NUM_ANYOF_CODE_POINTS */
7412 PL_InBitmap = _new_invlist(2);
7413 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
7414 NUM_ANYOF_CODE_POINTS - 1);
7416 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
7417 if ( ! dump_len_string
7418 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
7420 PL_dump_re_max_len = 60; /* A reasonable default */
7425 pRExC_state->warn_text = NULL;
7426 pRExC_state->unlexed_names = NULL;
7427 pRExC_state->code_blocks = NULL;
7430 *is_bare_re = FALSE;
7432 if (expr && (expr->op_type == OP_LIST ||
7433 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7434 /* allocate code_blocks if needed */
7438 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7439 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7440 ncode++; /* count of DO blocks */
7443 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7447 /* compile-time pattern with just OP_CONSTs and DO blocks */
7452 /* find how many CONSTs there are */
7455 if (expr->op_type == OP_CONST)
7458 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7459 if (o->op_type == OP_CONST)
7463 /* fake up an SV array */
7465 assert(!new_patternp);
7466 Newx(new_patternp, n, SV*);
7467 SAVEFREEPV(new_patternp);
7471 if (expr->op_type == OP_CONST)
7472 new_patternp[n] = cSVOPx_sv(expr);
7474 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7475 if (o->op_type == OP_CONST)
7476 new_patternp[n++] = cSVOPo_sv;
7481 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7482 "Assembling pattern from %d elements%s\n", pat_count,
7483 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7485 /* set expr to the first arg op */
7487 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7488 && expr->op_type != OP_CONST)
7490 expr = cLISTOPx(expr)->op_first;
7491 assert( expr->op_type == OP_PUSHMARK
7492 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7493 || expr->op_type == OP_PADRANGE);
7494 expr = OpSIBLING(expr);
7497 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7498 expr, &recompile, NULL);
7500 /* handle bare (possibly after overloading) regex: foo =~ $re */
7505 if (SvTYPE(re) == SVt_REGEXP) {
7509 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7510 "Precompiled pattern%s\n",
7511 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7517 exp = SvPV_nomg(pat, plen);
7519 if (!eng->op_comp) {
7520 if ((SvUTF8(pat) && IN_BYTES)
7521 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7523 /* make a temporary copy; either to convert to bytes,
7524 * or to avoid repeating get-magic / overloaded stringify */
7525 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7526 (IN_BYTES ? 0 : SvUTF8(pat)));
7528 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7531 /* ignore the utf8ness if the pattern is 0 length */
7532 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7533 RExC_uni_semantics = 0;
7534 RExC_contains_locale = 0;
7535 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7536 RExC_in_script_run = 0;
7537 RExC_study_started = 0;
7538 pRExC_state->runtime_code_qr = NULL;
7539 RExC_frame_head= NULL;
7540 RExC_frame_last= NULL;
7541 RExC_frame_count= 0;
7542 RExC_latest_warn_offset = 0;
7543 RExC_use_BRANCHJ = 0;
7544 RExC_total_parens = 0;
7545 RExC_open_parens = NULL;
7546 RExC_close_parens = NULL;
7547 RExC_paren_names = NULL;
7549 RExC_seen_d_op = FALSE;
7551 RExC_paren_name_list = NULL;
7555 RExC_mysv1= sv_newmortal();
7556 RExC_mysv2= sv_newmortal();
7560 SV *dsv= sv_newmortal();
7561 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7562 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7563 PL_colors[4], PL_colors[5], s);
7566 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7569 if ((pm_flags & PMf_USE_RE_EVAL)
7570 /* this second condition covers the non-regex literal case,
7571 * i.e. $foo =~ '(?{})'. */
7572 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7574 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7577 /* return old regex if pattern hasn't changed */
7578 /* XXX: note in the below we have to check the flags as well as the
7581 * Things get a touch tricky as we have to compare the utf8 flag
7582 * independently from the compile flags. */
7586 && !!RX_UTF8(old_re) == !!RExC_utf8
7587 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7588 && RX_PRECOMP(old_re)
7589 && RX_PRELEN(old_re) == plen
7590 && memEQ(RX_PRECOMP(old_re), exp, plen)
7591 && !runtime_code /* with runtime code, always recompile */ )
7594 SV *dsv= sv_newmortal();
7595 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7596 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7597 PL_colors[4], PL_colors[5], s);
7602 /* Allocate the pattern's SV */
7603 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7604 RExC_rx = ReANY(Rx);
7605 if ( RExC_rx == NULL )
7606 FAIL("Regexp out of space");
7608 rx_flags = orig_rx_flags;
7610 if ( (UTF || RExC_uni_semantics)
7611 && initial_charset == REGEX_DEPENDS_CHARSET)
7614 /* Set to use unicode semantics if the pattern is in utf8 and has the
7615 * 'depends' charset specified, as it means unicode when utf8 */
7616 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7617 RExC_uni_semantics = 1;
7620 RExC_pm_flags = pm_flags;
7623 assert(TAINTING_get || !TAINT_get);
7625 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7627 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7628 /* whoops, we have a non-utf8 pattern, whilst run-time code
7629 * got compiled as utf8. Try again with a utf8 pattern */
7630 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7631 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7635 assert(!pRExC_state->runtime_code_qr);
7641 RExC_in_lookbehind = 0;
7642 RExC_in_lookahead = 0;
7643 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7644 RExC_recode_x_to_native = 0;
7645 RExC_in_multi_char_class = 0;
7647 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7648 RExC_precomp_end = RExC_end = exp + plen;
7650 RExC_whilem_seen = 0;
7652 RExC_recurse = NULL;
7653 RExC_study_chunk_recursed = NULL;
7654 RExC_study_chunk_recursed_bytes= 0;
7655 RExC_recurse_count = 0;
7656 pRExC_state->code_index = 0;
7658 /* Initialize the string in the compiled pattern. This is so that there is
7659 * something to output if necessary */
7660 set_regex_pv(pRExC_state, Rx);
7663 Perl_re_printf( aTHX_
7664 "Starting parse and generation\n");
7666 RExC_lastparse=NULL;
7669 /* Allocate space and zero-initialize. Note, the two step process
7670 of zeroing when in debug mode, thus anything assigned has to
7671 happen after that */
7674 /* On the first pass of the parse, we guess how big this will be. Then
7675 * we grow in one operation to that amount and then give it back. As
7676 * we go along, we re-allocate what we need.
7678 * XXX Currently the guess is essentially that the pattern will be an
7679 * EXACT node with one byte input, one byte output. This is crude, and
7680 * better heuristics are welcome.
7682 * On any subsequent passes, we guess what we actually computed in the
7683 * latest earlier pass. Such a pass probably didn't complete so is
7684 * missing stuff. We could improve those guesses by knowing where the
7685 * parse stopped, and use the length so far plus apply the above
7686 * assumption to what's left. */
7687 RExC_size = STR_SZ(RExC_end - RExC_start);
7690 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7691 if ( RExC_rxi == NULL )
7692 FAIL("Regexp out of space");
7694 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7695 RXi_SET( RExC_rx, RExC_rxi );
7697 /* We start from 0 (over from 0 in the case this is a reparse. The first
7698 * node parsed will give back any excess memory we have allocated so far).
7702 /* non-zero initialization begins here */
7703 RExC_rx->engine= eng;
7704 RExC_rx->extflags = rx_flags;
7705 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7707 if (pm_flags & PMf_IS_QR) {
7708 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7709 if (RExC_rxi->code_blocks) {
7710 RExC_rxi->code_blocks->refcnt++;
7714 RExC_rx->intflags = 0;
7716 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7719 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7720 * code makes sure the final byte is an uncounted NUL. But should this
7721 * ever not be the case, lots of things could read beyond the end of the
7722 * buffer: loops like
7723 * while(isFOO(*RExC_parse)) RExC_parse++;
7724 * strchr(RExC_parse, "foo");
7725 * etc. So it is worth noting. */
7726 assert(*RExC_end == '\0');
7730 RExC_parens_buf_size = 0;
7731 RExC_emit_start = RExC_rxi->program;
7732 pRExC_state->code_index = 0;
7734 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7738 if (reg(pRExC_state, 0, &flags, 1)) {
7740 /* Success!, But we may need to redo the parse knowing how many parens
7741 * there actually are */
7742 if (IN_PARENS_PASS) {
7743 flags |= RESTART_PARSE;
7746 /* We have that number in RExC_npar */
7747 RExC_total_parens = RExC_npar;
7749 else if (! MUST_RESTART(flags)) {
7751 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7754 /* Here, we either have success, or we have to redo the parse for some reason */
7755 if (MUST_RESTART(flags)) {
7757 /* It's possible to write a regexp in ascii that represents Unicode
7758 codepoints outside of the byte range, such as via \x{100}. If we
7759 detect such a sequence we have to convert the entire pattern to utf8
7760 and then recompile, as our sizing calculation will have been based
7761 on 1 byte == 1 character, but we will need to use utf8 to encode
7762 at least some part of the pattern, and therefore must convert the whole
7765 if (flags & NEED_UTF8) {
7767 /* We have stored the offset of the final warning output so far.
7768 * That must be adjusted. Any variant characters between the start
7769 * of the pattern and this warning count for 2 bytes in the final,
7770 * so just add them again */
7771 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7772 RExC_latest_warn_offset +=
7773 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7774 + RExC_latest_warn_offset);
7776 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7777 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7778 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7781 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7784 if (ALL_PARENS_COUNTED) {
7785 /* Make enough room for all the known parens, and zero it */
7786 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7787 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7788 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7790 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7791 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7793 else { /* Parse did not complete. Reinitialize the parentheses
7795 RExC_total_parens = 0;
7796 if (RExC_open_parens) {
7797 Safefree(RExC_open_parens);
7798 RExC_open_parens = NULL;
7800 if (RExC_close_parens) {
7801 Safefree(RExC_close_parens);
7802 RExC_close_parens = NULL;
7806 /* Clean up what we did in this parse */
7807 SvREFCNT_dec_NN(RExC_rx_sv);
7812 /* Here, we have successfully parsed and generated the pattern's program
7813 * for the regex engine. We are ready to finish things up and look for
7816 /* Update the string to compile, with correct modifiers, etc */
7817 set_regex_pv(pRExC_state, Rx);
7819 RExC_rx->nparens = RExC_total_parens - 1;
7821 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7822 if (RExC_whilem_seen > 15)
7823 RExC_whilem_seen = 15;
7826 Perl_re_printf( aTHX_
7827 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7829 RExC_lastparse=NULL;
7832 #ifdef RE_TRACK_PATTERN_OFFSETS
7833 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7834 "%s %" UVuf " bytes for offset annotations.\n",
7835 RExC_offsets ? "Got" : "Couldn't get",
7836 (UV)((RExC_offsets[0] * 2 + 1))));
7837 DEBUG_OFFSETS_r(if (RExC_offsets) {
7838 const STRLEN len = RExC_offsets[0];
7840 GET_RE_DEBUG_FLAGS_DECL;
7841 Perl_re_printf( aTHX_
7842 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7843 for (i = 1; i <= len; i++) {
7844 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7845 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7846 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7848 Perl_re_printf( aTHX_ "\n");
7852 SetProgLen(RExC_rxi,RExC_size);
7855 DEBUG_DUMP_PRE_OPTIMIZE_r({
7856 SV * const sv = sv_newmortal();
7857 RXi_GET_DECL(RExC_rx, ri);
7859 Perl_re_printf( aTHX_ "Program before optimization:\n");
7861 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
7866 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7869 /* XXXX To minimize changes to RE engine we always allocate
7870 3-units-long substrs field. */
7871 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7872 if (RExC_recurse_count) {
7873 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7874 SAVEFREEPV(RExC_recurse);
7877 if (RExC_seen & REG_RECURSE_SEEN) {
7878 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
7879 * So its 1 if there are no parens. */
7880 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
7881 ((RExC_total_parens & 0x07) != 0);
7882 Newx(RExC_study_chunk_recursed,
7883 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7884 SAVEFREEPV(RExC_study_chunk_recursed);
7888 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7890 RExC_study_chunk_recursed_count= 0;
7892 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
7893 if (RExC_study_chunk_recursed) {
7894 Zero(RExC_study_chunk_recursed,
7895 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7899 #ifdef TRIE_STUDY_OPT
7901 StructCopy(&zero_scan_data, &data, scan_data_t);
7902 copyRExC_state = RExC_state;
7905 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7907 RExC_state = copyRExC_state;
7908 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7909 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7911 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7912 StructCopy(&zero_scan_data, &data, scan_data_t);
7915 StructCopy(&zero_scan_data, &data, scan_data_t);
7918 /* Dig out information for optimizations. */
7919 RExC_rx->extflags = RExC_flags; /* was pm_op */
7920 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7923 SvUTF8_on(Rx); /* Unicode in it? */
7924 RExC_rxi->regstclass = NULL;
7925 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7926 RExC_rx->intflags |= PREGf_NAUGHTY;
7927 scan = RExC_rxi->program + 1; /* First BRANCH. */
7929 /* testing for BRANCH here tells us whether there is "must appear"
7930 data in the pattern. If there is then we can use it for optimisations */
7931 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7934 STRLEN longest_length[2];
7935 regnode_ssc ch_class; /* pointed to by data */
7937 SSize_t last_close = 0; /* pointed to by data */
7938 regnode *first= scan;
7939 regnode *first_next= regnext(first);
7943 * Skip introductions and multiplicators >= 1
7944 * so that we can extract the 'meat' of the pattern that must
7945 * match in the large if() sequence following.
7946 * NOTE that EXACT is NOT covered here, as it is normally
7947 * picked up by the optimiser separately.
7949 * This is unfortunate as the optimiser isnt handling lookahead
7950 * properly currently.
7953 while ((OP(first) == OPEN && (sawopen = 1)) ||
7954 /* An OR of *one* alternative - should not happen now. */
7955 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7956 /* for now we can't handle lookbehind IFMATCH*/
7957 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7958 (OP(first) == PLUS) ||
7959 (OP(first) == MINMOD) ||
7960 /* An {n,m} with n>0 */
7961 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7962 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7965 * the only op that could be a regnode is PLUS, all the rest
7966 * will be regnode_1 or regnode_2.
7968 * (yves doesn't think this is true)
7970 if (OP(first) == PLUS)
7973 if (OP(first) == MINMOD)
7975 first += regarglen[OP(first)];
7977 first = NEXTOPER(first);
7978 first_next= regnext(first);
7981 /* Starting-point info. */
7983 DEBUG_PEEP("first:", first, 0, 0);
7984 /* Ignore EXACT as we deal with it later. */
7985 if (PL_regkind[OP(first)] == EXACT) {
7986 if ( OP(first) == EXACT
7987 || OP(first) == LEXACT
7988 || OP(first) == EXACT_REQ8
7989 || OP(first) == LEXACT_REQ8
7990 || OP(first) == EXACTL)
7992 NOOP; /* Empty, get anchored substr later. */
7995 RExC_rxi->regstclass = first;
7998 else if (PL_regkind[OP(first)] == TRIE &&
7999 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
8001 /* this can happen only on restudy */
8002 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8005 else if (REGNODE_SIMPLE(OP(first)))
8006 RExC_rxi->regstclass = first;
8007 else if (PL_regkind[OP(first)] == BOUND ||
8008 PL_regkind[OP(first)] == NBOUND)
8009 RExC_rxi->regstclass = first;
8010 else if (PL_regkind[OP(first)] == BOL) {
8011 RExC_rx->intflags |= (OP(first) == MBOL
8014 first = NEXTOPER(first);
8017 else if (OP(first) == GPOS) {
8018 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8019 first = NEXTOPER(first);
8022 else if ((!sawopen || !RExC_sawback) &&
8024 (OP(first) == STAR &&
8025 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8026 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8028 /* turn .* into ^.* with an implied $*=1 */
8030 (OP(NEXTOPER(first)) == REG_ANY)
8033 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8034 first = NEXTOPER(first);
8037 if (sawplus && !sawminmod && !sawlookahead
8038 && (!sawopen || !RExC_sawback)
8039 && !pRExC_state->code_blocks) /* May examine pos and $& */
8040 /* x+ must match at the 1st pos of run of x's */
8041 RExC_rx->intflags |= PREGf_SKIP;
8043 /* Scan is after the zeroth branch, first is atomic matcher. */
8044 #ifdef TRIE_STUDY_OPT
8047 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8048 (IV)(first - scan + 1))
8052 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8053 (IV)(first - scan + 1))
8059 * If there's something expensive in the r.e., find the
8060 * longest literal string that must appear and make it the
8061 * regmust. Resolve ties in favor of later strings, since
8062 * the regstart check works with the beginning of the r.e.
8063 * and avoiding duplication strengthens checking. Not a
8064 * strong reason, but sufficient in the absence of others.
8065 * [Now we resolve ties in favor of the earlier string if
8066 * it happens that c_offset_min has been invalidated, since the
8067 * earlier string may buy us something the later one won't.]
8070 data.substrs[0].str = newSVpvs("");
8071 data.substrs[1].str = newSVpvs("");
8072 data.last_found = newSVpvs("");
8073 data.cur_is_floating = 0; /* initially any found substring is fixed */
8074 ENTER_with_name("study_chunk");
8075 SAVEFREESV(data.substrs[0].str);
8076 SAVEFREESV(data.substrs[1].str);
8077 SAVEFREESV(data.last_found);
8079 if (!RExC_rxi->regstclass) {
8080 ssc_init(pRExC_state, &ch_class);
8081 data.start_class = &ch_class;
8082 stclass_flag = SCF_DO_STCLASS_AND;
8083 } else /* XXXX Check for BOUND? */
8085 data.last_closep = &last_close;
8089 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8090 * (NO top level branches)
8092 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8093 scan + RExC_size, /* Up to end */
8095 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8096 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8100 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8103 if ( RExC_total_parens == 1 && !data.cur_is_floating
8104 && data.last_start_min == 0 && data.last_end > 0
8105 && !RExC_seen_zerolen
8106 && !(RExC_seen & REG_VERBARG_SEEN)
8107 && !(RExC_seen & REG_GPOS_SEEN)
8109 RExC_rx->extflags |= RXf_CHECK_ALL;
8111 scan_commit(pRExC_state, &data,&minlen, 0);
8114 /* XXX this is done in reverse order because that's the way the
8115 * code was before it was parameterised. Don't know whether it
8116 * actually needs doing in reverse order. DAPM */
8117 for (i = 1; i >= 0; i--) {
8118 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8121 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8122 && data.substrs[0].min_offset
8123 == data.substrs[1].min_offset
8124 && SvCUR(data.substrs[0].str)
8125 == SvCUR(data.substrs[1].str)
8127 && S_setup_longest (aTHX_ pRExC_state,
8128 &(RExC_rx->substrs->data[i]),
8132 RExC_rx->substrs->data[i].min_offset =
8133 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8135 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8136 /* Don't offset infinity */
8137 if (data.substrs[i].max_offset < SSize_t_MAX)
8138 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8139 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8142 RExC_rx->substrs->data[i].substr = NULL;
8143 RExC_rx->substrs->data[i].utf8_substr = NULL;
8144 longest_length[i] = 0;
8148 LEAVE_with_name("study_chunk");
8150 if (RExC_rxi->regstclass
8151 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8152 RExC_rxi->regstclass = NULL;
8154 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8155 || RExC_rx->substrs->data[0].min_offset)
8157 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8158 && is_ssc_worth_it(pRExC_state, data.start_class))
8160 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8162 ssc_finalize(pRExC_state, data.start_class);
8164 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8165 StructCopy(data.start_class,
8166 (regnode_ssc*)RExC_rxi->data->data[n],
8168 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8169 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8170 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8171 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8172 Perl_re_printf( aTHX_
8173 "synthetic stclass \"%s\".\n",
8174 SvPVX_const(sv));});
8175 data.start_class = NULL;
8178 /* A temporary algorithm prefers floated substr to fixed one of
8179 * same length to dig more info. */
8180 i = (longest_length[0] <= longest_length[1]);
8181 RExC_rx->substrs->check_ix = i;
8182 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8183 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8184 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8185 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8186 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8187 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8188 RExC_rx->intflags |= PREGf_NOSCAN;
8190 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8191 RExC_rx->extflags |= RXf_USE_INTUIT;
8192 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8193 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8196 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8197 if ( (STRLEN)minlen < longest_length[1] )
8198 minlen= longest_length[1];
8199 if ( (STRLEN)minlen < longest_length[0] )
8200 minlen= longest_length[0];
8204 /* Several toplevels. Best we can is to set minlen. */
8206 regnode_ssc ch_class;
8207 SSize_t last_close = 0;
8209 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8211 scan = RExC_rxi->program + 1;
8212 ssc_init(pRExC_state, &ch_class);
8213 data.start_class = &ch_class;
8214 data.last_closep = &last_close;
8218 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8219 * (patterns WITH top level branches)
8221 minlen = study_chunk(pRExC_state,
8222 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8223 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8224 ? SCF_TRIE_DOING_RESTUDY
8228 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8230 RExC_rx->check_substr = NULL;
8231 RExC_rx->check_utf8 = NULL;
8232 RExC_rx->substrs->data[0].substr = NULL;
8233 RExC_rx->substrs->data[0].utf8_substr = NULL;
8234 RExC_rx->substrs->data[1].substr = NULL;
8235 RExC_rx->substrs->data[1].utf8_substr = NULL;
8237 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8238 && is_ssc_worth_it(pRExC_state, data.start_class))
8240 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8242 ssc_finalize(pRExC_state, data.start_class);
8244 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8245 StructCopy(data.start_class,
8246 (regnode_ssc*)RExC_rxi->data->data[n],
8248 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8249 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8250 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8251 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8252 Perl_re_printf( aTHX_
8253 "synthetic stclass \"%s\".\n",
8254 SvPVX_const(sv));});
8255 data.start_class = NULL;
8259 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8260 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8261 RExC_rx->maxlen = REG_INFTY;
8264 RExC_rx->maxlen = RExC_maxlen;
8267 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8268 the "real" pattern. */
8270 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8271 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8273 RExC_rx->minlenret = minlen;
8274 if (RExC_rx->minlen < minlen)
8275 RExC_rx->minlen = minlen;
8277 if (RExC_seen & REG_RECURSE_SEEN ) {
8278 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8279 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8281 if (RExC_seen & REG_GPOS_SEEN)
8282 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8283 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8284 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8286 if (pRExC_state->code_blocks)
8287 RExC_rx->extflags |= RXf_EVAL_SEEN;
8288 if (RExC_seen & REG_VERBARG_SEEN)
8290 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8291 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8293 if (RExC_seen & REG_CUTGROUP_SEEN)
8294 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8295 if (pm_flags & PMf_USE_RE_EVAL)
8296 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8297 if (RExC_paren_names)
8298 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8300 RXp_PAREN_NAMES(RExC_rx) = NULL;
8302 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8303 * so it can be used in pp.c */
8304 if (RExC_rx->intflags & PREGf_ANCH)
8305 RExC_rx->extflags |= RXf_IS_ANCHORED;
8309 /* this is used to identify "special" patterns that might result
8310 * in Perl NOT calling the regex engine and instead doing the match "itself",
8311 * particularly special cases in split//. By having the regex compiler
8312 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8313 * we avoid weird issues with equivalent patterns resulting in different behavior,
8314 * AND we allow non Perl engines to get the same optimizations by the setting the
8315 * flags appropriately - Yves */
8316 regnode *first = RExC_rxi->program + 1;
8318 regnode *next = regnext(first);
8321 if (PL_regkind[fop] == NOTHING && nop == END)
8322 RExC_rx->extflags |= RXf_NULL;
8323 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8324 /* when fop is SBOL first->flags will be true only when it was
8325 * produced by parsing /\A/, and not when parsing /^/. This is
8326 * very important for the split code as there we want to
8327 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8328 * See rt #122761 for more details. -- Yves */
8329 RExC_rx->extflags |= RXf_START_ONLY;
8330 else if (fop == PLUS
8331 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8333 RExC_rx->extflags |= RXf_WHITE;
8334 else if ( RExC_rx->extflags & RXf_SPLIT
8335 && ( fop == EXACT || fop == LEXACT
8336 || fop == EXACT_REQ8 || fop == LEXACT_REQ8
8338 && STR_LEN(first) == 1
8339 && *(STRING(first)) == ' '
8341 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8345 if (RExC_contains_locale) {
8346 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8350 if (RExC_paren_names) {
8351 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8352 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8353 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8356 RExC_rxi->name_list_idx = 0;
8358 while ( RExC_recurse_count > 0 ) {
8359 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8361 * This data structure is set up in study_chunk() and is used
8362 * to calculate the distance between a GOSUB regopcode and
8363 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8366 * If for some reason someone writes code that optimises
8367 * away a GOSUB opcode then the assert should be changed to
8368 * an if(scan) to guard the ARG2L_SET() - Yves
8371 assert(scan && OP(scan) == GOSUB);
8372 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8375 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8376 /* assume we don't need to swap parens around before we match */
8378 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8379 (unsigned long)RExC_study_chunk_recursed_count);
8383 Perl_re_printf( aTHX_ "Final program:\n");
8387 if (RExC_open_parens) {
8388 Safefree(RExC_open_parens);
8389 RExC_open_parens = NULL;
8391 if (RExC_close_parens) {
8392 Safefree(RExC_close_parens);
8393 RExC_close_parens = NULL;
8397 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8398 * by setting the regexp SV to readonly-only instead. If the
8399 * pattern's been recompiled, the USEDness should remain. */
8400 if (old_re && SvREADONLY(old_re))
8408 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8411 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8413 PERL_UNUSED_ARG(value);
8415 if (flags & RXapif_FETCH) {
8416 return reg_named_buff_fetch(rx, key, flags);
8417 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8418 Perl_croak_no_modify();
8420 } else if (flags & RXapif_EXISTS) {
8421 return reg_named_buff_exists(rx, key, flags)
8424 } else if (flags & RXapif_REGNAMES) {
8425 return reg_named_buff_all(rx, flags);
8426 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8427 return reg_named_buff_scalar(rx, flags);
8429 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8435 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8438 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8439 PERL_UNUSED_ARG(lastkey);
8441 if (flags & RXapif_FIRSTKEY)
8442 return reg_named_buff_firstkey(rx, flags);
8443 else if (flags & RXapif_NEXTKEY)
8444 return reg_named_buff_nextkey(rx, flags);
8446 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8453 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8457 struct regexp *const rx = ReANY(r);
8459 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8461 if (rx && RXp_PAREN_NAMES(rx)) {
8462 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8465 SV* sv_dat=HeVAL(he_str);
8466 I32 *nums=(I32*)SvPVX(sv_dat);
8467 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8468 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8469 if ((I32)(rx->nparens) >= nums[i]
8470 && rx->offs[nums[i]].start != -1
8471 && rx->offs[nums[i]].end != -1)
8474 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8479 ret = newSVsv(&PL_sv_undef);
8482 av_push(retarray, ret);
8485 return newRV_noinc(MUTABLE_SV(retarray));
8492 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8495 struct regexp *const rx = ReANY(r);
8497 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8499 if (rx && RXp_PAREN_NAMES(rx)) {
8500 if (flags & RXapif_ALL) {
8501 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8503 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8505 SvREFCNT_dec_NN(sv);
8517 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8519 struct regexp *const rx = ReANY(r);
8521 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8523 if ( rx && RXp_PAREN_NAMES(rx) ) {
8524 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8526 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8533 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8535 struct regexp *const rx = ReANY(r);
8536 GET_RE_DEBUG_FLAGS_DECL;
8538 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8540 if (rx && RXp_PAREN_NAMES(rx)) {
8541 HV *hv = RXp_PAREN_NAMES(rx);
8543 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8546 SV* sv_dat = HeVAL(temphe);
8547 I32 *nums = (I32*)SvPVX(sv_dat);
8548 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8549 if ((I32)(rx->lastparen) >= nums[i] &&
8550 rx->offs[nums[i]].start != -1 &&
8551 rx->offs[nums[i]].end != -1)
8557 if (parno || flags & RXapif_ALL) {
8558 return newSVhek(HeKEY_hek(temphe));
8566 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8571 struct regexp *const rx = ReANY(r);
8573 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8575 if (rx && RXp_PAREN_NAMES(rx)) {
8576 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8577 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8578 } else if (flags & RXapif_ONE) {
8579 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8580 av = MUTABLE_AV(SvRV(ret));
8581 length = av_tindex(av);
8582 SvREFCNT_dec_NN(ret);
8583 return newSViv(length + 1);
8585 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8590 return &PL_sv_undef;
8594 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8596 struct regexp *const rx = ReANY(r);
8599 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8601 if (rx && RXp_PAREN_NAMES(rx)) {
8602 HV *hv= RXp_PAREN_NAMES(rx);
8604 (void)hv_iterinit(hv);
8605 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8608 SV* sv_dat = HeVAL(temphe);
8609 I32 *nums = (I32*)SvPVX(sv_dat);
8610 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8611 if ((I32)(rx->lastparen) >= nums[i] &&
8612 rx->offs[nums[i]].start != -1 &&
8613 rx->offs[nums[i]].end != -1)
8619 if (parno || flags & RXapif_ALL) {
8620 av_push(av, newSVhek(HeKEY_hek(temphe)));
8625 return newRV_noinc(MUTABLE_SV(av));
8629 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8632 struct regexp *const rx = ReANY(r);
8638 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8640 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8641 || n == RX_BUFF_IDX_CARET_FULLMATCH
8642 || n == RX_BUFF_IDX_CARET_POSTMATCH
8645 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8647 /* on something like
8650 * the KEEPCOPY is set on the PMOP rather than the regex */
8651 if (PL_curpm && r == PM_GETRE(PL_curpm))
8652 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8661 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8662 /* no need to distinguish between them any more */
8663 n = RX_BUFF_IDX_FULLMATCH;
8665 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8666 && rx->offs[0].start != -1)
8668 /* $`, ${^PREMATCH} */
8669 i = rx->offs[0].start;
8673 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8674 && rx->offs[0].end != -1)
8676 /* $', ${^POSTMATCH} */
8677 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8678 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8681 if ( 0 <= n && n <= (I32)rx->nparens &&
8682 (s1 = rx->offs[n].start) != -1 &&
8683 (t1 = rx->offs[n].end) != -1)
8685 /* $&, ${^MATCH}, $1 ... */
8687 s = rx->subbeg + s1 - rx->suboffset;
8692 assert(s >= rx->subbeg);
8693 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8695 #ifdef NO_TAINT_SUPPORT
8696 sv_setpvn(sv, s, i);
8698 const int oldtainted = TAINT_get;
8700 sv_setpvn(sv, s, i);
8701 TAINT_set(oldtainted);
8703 if (RXp_MATCH_UTF8(rx))
8708 if (RXp_MATCH_TAINTED(rx)) {
8709 if (SvTYPE(sv) >= SVt_PVMG) {
8710 MAGIC* const mg = SvMAGIC(sv);
8713 SvMAGIC_set(sv, mg->mg_moremagic);
8715 if ((mgt = SvMAGIC(sv))) {
8716 mg->mg_moremagic = mgt;
8717 SvMAGIC_set(sv, mg);
8734 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8735 SV const * const value)
8737 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8739 PERL_UNUSED_ARG(rx);
8740 PERL_UNUSED_ARG(paren);
8741 PERL_UNUSED_ARG(value);
8744 Perl_croak_no_modify();
8748 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8751 struct regexp *const rx = ReANY(r);
8755 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8757 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8758 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8759 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8762 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8764 /* on something like
8767 * the KEEPCOPY is set on the PMOP rather than the regex */
8768 if (PL_curpm && r == PM_GETRE(PL_curpm))
8769 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8775 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8777 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8778 case RX_BUFF_IDX_PREMATCH: /* $` */
8779 if (rx->offs[0].start != -1) {
8780 i = rx->offs[0].start;
8789 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8790 case RX_BUFF_IDX_POSTMATCH: /* $' */
8791 if (rx->offs[0].end != -1) {
8792 i = rx->sublen - rx->offs[0].end;
8794 s1 = rx->offs[0].end;
8801 default: /* $& / ${^MATCH}, $1, $2, ... */
8802 if (paren <= (I32)rx->nparens &&
8803 (s1 = rx->offs[paren].start) != -1 &&
8804 (t1 = rx->offs[paren].end) != -1)
8810 if (ckWARN(WARN_UNINITIALIZED))
8811 report_uninit((const SV *)sv);
8816 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8817 const char * const s = rx->subbeg - rx->suboffset + s1;
8822 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8829 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8831 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8832 PERL_UNUSED_ARG(rx);
8836 return newSVpvs("Regexp");
8839 /* Scans the name of a named buffer from the pattern.
8840 * If flags is REG_RSN_RETURN_NULL returns null.
8841 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8842 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8843 * to the parsed name as looked up in the RExC_paren_names hash.
8844 * If there is an error throws a vFAIL().. type exception.
8847 #define REG_RSN_RETURN_NULL 0
8848 #define REG_RSN_RETURN_NAME 1
8849 #define REG_RSN_RETURN_DATA 2
8852 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8854 char *name_start = RExC_parse;
8857 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8859 assert (RExC_parse <= RExC_end);
8860 if (RExC_parse == RExC_end) NOOP;
8861 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8862 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8863 * using do...while */
8866 RExC_parse += UTF8SKIP(RExC_parse);
8867 } while ( RExC_parse < RExC_end
8868 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8872 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8874 RExC_parse++; /* so the <- from the vFAIL is after the offending
8876 vFAIL("Group name must start with a non-digit word character");
8878 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8879 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8880 if ( flags == REG_RSN_RETURN_NAME)
8882 else if (flags==REG_RSN_RETURN_DATA) {
8885 if ( ! sv_name ) /* should not happen*/
8886 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8887 if (RExC_paren_names)
8888 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8890 sv_dat = HeVAL(he_str);
8891 if ( ! sv_dat ) { /* Didn't find group */
8893 /* It might be a forward reference; we can't fail until we
8894 * know, by completing the parse to get all the groups, and
8896 if (ALL_PARENS_COUNTED) {
8897 vFAIL("Reference to nonexistent named group");
8900 REQUIRE_PARENS_PASS;
8906 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8907 (unsigned long) flags);
8910 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8911 if (RExC_lastparse!=RExC_parse) { \
8912 Perl_re_printf( aTHX_ "%s", \
8913 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8914 RExC_end - RExC_parse, 16, \
8916 PERL_PV_ESCAPE_UNI_DETECT | \
8917 PERL_PV_PRETTY_ELLIPSES | \
8918 PERL_PV_PRETTY_LTGT | \
8919 PERL_PV_ESCAPE_RE | \
8920 PERL_PV_PRETTY_EXACTSIZE \
8924 Perl_re_printf( aTHX_ "%16s",""); \
8926 if (RExC_lastnum!=RExC_emit) \
8927 Perl_re_printf( aTHX_ "|%4d", RExC_emit); \
8929 Perl_re_printf( aTHX_ "|%4s",""); \
8930 Perl_re_printf( aTHX_ "|%*s%-4s", \
8931 (int)((depth*2)), "", \
8934 RExC_lastnum=RExC_emit; \
8935 RExC_lastparse=RExC_parse; \
8940 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8941 DEBUG_PARSE_MSG((funcname)); \
8942 Perl_re_printf( aTHX_ "%4s","\n"); \
8944 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8945 DEBUG_PARSE_MSG((funcname)); \
8946 Perl_re_printf( aTHX_ fmt "\n",args); \
8949 /* This section of code defines the inversion list object and its methods. The
8950 * interfaces are highly subject to change, so as much as possible is static to
8951 * this file. An inversion list is here implemented as a malloc'd C UV array
8952 * as an SVt_INVLIST scalar.
8954 * An inversion list for Unicode is an array of code points, sorted by ordinal
8955 * number. Each element gives the code point that begins a range that extends
8956 * up-to but not including the code point given by the next element. The final
8957 * element gives the first code point of a range that extends to the platform's
8958 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8959 * ...) give ranges whose code points are all in the inversion list. We say
8960 * that those ranges are in the set. The odd-numbered elements give ranges
8961 * whose code points are not in the inversion list, and hence not in the set.
8962 * Thus, element [0] is the first code point in the list. Element [1]
8963 * is the first code point beyond that not in the list; and element [2] is the
8964 * first code point beyond that that is in the list. In other words, the first
8965 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8966 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8967 * all code points in that range are not in the inversion list. The third
8968 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8969 * list, and so forth. Thus every element whose index is divisible by two
8970 * gives the beginning of a range that is in the list, and every element whose
8971 * index is not divisible by two gives the beginning of a range not in the
8972 * list. If the final element's index is divisible by two, the inversion list
8973 * extends to the platform's infinity; otherwise the highest code point in the
8974 * inversion list is the contents of that element minus 1.
8976 * A range that contains just a single code point N will look like
8978 * invlist[i+1] == N+1
8980 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8981 * impossible to represent, so element [i+1] is omitted. The single element
8983 * invlist[0] == UV_MAX
8984 * contains just UV_MAX, but is interpreted as matching to infinity.
8986 * Taking the complement (inverting) an inversion list is quite simple, if the
8987 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8988 * This implementation reserves an element at the beginning of each inversion
8989 * list to always contain 0; there is an additional flag in the header which
8990 * indicates if the list begins at the 0, or is offset to begin at the next
8991 * element. This means that the inversion list can be inverted without any
8992 * copying; just flip the flag.
8994 * More about inversion lists can be found in "Unicode Demystified"
8995 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8997 * The inversion list data structure is currently implemented as an SV pointing
8998 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8999 * array of UV whose memory management is automatically handled by the existing
9000 * facilities for SV's.
9002 * Some of the methods should always be private to the implementation, and some
9003 * should eventually be made public */
9005 /* The header definitions are in F<invlist_inline.h> */
9007 #ifndef PERL_IN_XSUB_RE
9009 PERL_STATIC_INLINE UV*
9010 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9012 /* Returns a pointer to the first element in the inversion list's array.
9013 * This is called upon initialization of an inversion list. Where the
9014 * array begins depends on whether the list has the code point U+0000 in it
9015 * or not. The other parameter tells it whether the code that follows this
9016 * call is about to put a 0 in the inversion list or not. The first
9017 * element is either the element reserved for 0, if TRUE, or the element
9018 * after it, if FALSE */
9020 bool* offset = get_invlist_offset_addr(invlist);
9021 UV* zero_addr = (UV *) SvPVX(invlist);
9023 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9026 assert(! _invlist_len(invlist));
9030 /* 1^1 = 0; 1^0 = 1 */
9031 *offset = 1 ^ will_have_0;
9032 return zero_addr + *offset;
9036 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9038 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9039 * steals the list from 'src', so 'src' is made to have a NULL list. This
9040 * is similar to what SvSetMagicSV() would do, if it were implemented on
9041 * inversion lists, though this routine avoids a copy */
9043 const UV src_len = _invlist_len(src);
9044 const bool src_offset = *get_invlist_offset_addr(src);
9045 const STRLEN src_byte_len = SvLEN(src);
9046 char * array = SvPVX(src);
9048 const int oldtainted = TAINT_get;
9050 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9052 assert(is_invlist(src));
9053 assert(is_invlist(dest));
9054 assert(! invlist_is_iterating(src));
9055 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9057 /* Make sure it ends in the right place with a NUL, as our inversion list
9058 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9060 array[src_byte_len - 1] = '\0';
9062 TAINT_NOT; /* Otherwise it breaks */
9063 sv_usepvn_flags(dest,
9067 /* This flag is documented to cause a copy to be avoided */
9068 SV_HAS_TRAILING_NUL);
9069 TAINT_set(oldtainted);
9074 /* Finish up copying over the other fields in an inversion list */
9075 *get_invlist_offset_addr(dest) = src_offset;
9076 invlist_set_len(dest, src_len, src_offset);
9077 *get_invlist_previous_index_addr(dest) = 0;
9078 invlist_iterfinish(dest);
9081 PERL_STATIC_INLINE IV*
9082 S_get_invlist_previous_index_addr(SV* invlist)
9084 /* Return the address of the IV that is reserved to hold the cached index
9086 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9088 assert(is_invlist(invlist));
9090 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9093 PERL_STATIC_INLINE IV
9094 S_invlist_previous_index(SV* const invlist)
9096 /* Returns cached index of previous search */
9098 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9100 return *get_invlist_previous_index_addr(invlist);
9103 PERL_STATIC_INLINE void
9104 S_invlist_set_previous_index(SV* const invlist, const IV index)
9106 /* Caches <index> for later retrieval */
9108 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9110 assert(index == 0 || index < (int) _invlist_len(invlist));
9112 *get_invlist_previous_index_addr(invlist) = index;
9115 PERL_STATIC_INLINE void
9116 S_invlist_trim(SV* invlist)
9118 /* Free the not currently-being-used space in an inversion list */
9120 /* But don't free up the space needed for the 0 UV that is always at the
9121 * beginning of the list, nor the trailing NUL */
9122 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9124 PERL_ARGS_ASSERT_INVLIST_TRIM;
9126 assert(is_invlist(invlist));
9128 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9131 PERL_STATIC_INLINE void
9132 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9134 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9136 assert(is_invlist(invlist));
9138 invlist_set_len(invlist, 0, 0);
9139 invlist_trim(invlist);
9142 #endif /* ifndef PERL_IN_XSUB_RE */
9144 PERL_STATIC_INLINE bool
9145 S_invlist_is_iterating(SV* const invlist)
9147 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9149 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9152 #ifndef PERL_IN_XSUB_RE
9154 PERL_STATIC_INLINE UV
9155 S_invlist_max(SV* const invlist)
9157 /* Returns the maximum number of elements storable in the inversion list's
9158 * array, without having to realloc() */
9160 PERL_ARGS_ASSERT_INVLIST_MAX;
9162 assert(is_invlist(invlist));
9164 /* Assumes worst case, in which the 0 element is not counted in the
9165 * inversion list, so subtracts 1 for that */
9166 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9167 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9168 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9172 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9174 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9176 /* First 1 is in case the zero element isn't in the list; second 1 is for
9178 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9179 invlist_set_len(invlist, 0, 0);
9181 /* Force iterinit() to be used to get iteration to work */
9182 invlist_iterfinish(invlist);
9184 *get_invlist_previous_index_addr(invlist) = 0;
9185 SvPOK_on(invlist); /* This allows B to extract the PV */
9189 Perl__new_invlist(pTHX_ IV initial_size)
9192 /* Return a pointer to a newly constructed inversion list, with enough
9193 * space to store 'initial_size' elements. If that number is negative, a
9194 * system default is used instead */
9198 if (initial_size < 0) {
9202 new_list = newSV_type(SVt_INVLIST);
9203 initialize_invlist_guts(new_list, initial_size);
9209 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9211 /* Return a pointer to a newly constructed inversion list, initialized to
9212 * point to <list>, which has to be in the exact correct inversion list
9213 * form, including internal fields. Thus this is a dangerous routine that
9214 * should not be used in the wrong hands. The passed in 'list' contains
9215 * several header fields at the beginning that are not part of the
9216 * inversion list body proper */
9218 const STRLEN length = (STRLEN) list[0];
9219 const UV version_id = list[1];
9220 const bool offset = cBOOL(list[2]);
9221 #define HEADER_LENGTH 3
9222 /* If any of the above changes in any way, you must change HEADER_LENGTH
9223 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9224 * perl -E 'say int(rand 2**31-1)'
9226 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9227 data structure type, so that one being
9228 passed in can be validated to be an
9229 inversion list of the correct vintage.
9232 SV* invlist = newSV_type(SVt_INVLIST);
9234 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9236 if (version_id != INVLIST_VERSION_ID) {
9237 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9240 /* The generated array passed in includes header elements that aren't part
9241 * of the list proper, so start it just after them */
9242 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9244 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9245 shouldn't touch it */
9247 *(get_invlist_offset_addr(invlist)) = offset;
9249 /* The 'length' passed to us is the physical number of elements in the
9250 * inversion list. But if there is an offset the logical number is one
9252 invlist_set_len(invlist, length - offset, offset);
9254 invlist_set_previous_index(invlist, 0);
9256 /* Initialize the iteration pointer. */
9257 invlist_iterfinish(invlist);
9259 SvREADONLY_on(invlist);
9266 S__append_range_to_invlist(pTHX_ SV* const invlist,
9267 const UV start, const UV end)
9269 /* Subject to change or removal. Append the range from 'start' to 'end' at
9270 * the end of the inversion list. The range must be above any existing
9274 UV max = invlist_max(invlist);
9275 UV len = _invlist_len(invlist);
9278 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9280 if (len == 0) { /* Empty lists must be initialized */
9281 offset = start != 0;
9282 array = _invlist_array_init(invlist, ! offset);
9285 /* Here, the existing list is non-empty. The current max entry in the
9286 * list is generally the first value not in the set, except when the
9287 * set extends to the end of permissible values, in which case it is
9288 * the first entry in that final set, and so this call is an attempt to
9289 * append out-of-order */
9291 UV final_element = len - 1;
9292 array = invlist_array(invlist);
9293 if ( array[final_element] > start
9294 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9296 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",
9297 array[final_element], start,
9298 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9301 /* Here, it is a legal append. If the new range begins 1 above the end
9302 * of the range below it, it is extending the range below it, so the
9303 * new first value not in the set is one greater than the newly
9304 * extended range. */
9305 offset = *get_invlist_offset_addr(invlist);
9306 if (array[final_element] == start) {
9307 if (end != UV_MAX) {
9308 array[final_element] = end + 1;
9311 /* But if the end is the maximum representable on the machine,
9312 * assume that infinity was actually what was meant. Just let
9313 * the range that this would extend to have no end */
9314 invlist_set_len(invlist, len - 1, offset);
9320 /* Here the new range doesn't extend any existing set. Add it */
9322 len += 2; /* Includes an element each for the start and end of range */
9324 /* If wll overflow the existing space, extend, which may cause the array to
9327 invlist_extend(invlist, len);
9329 /* Have to set len here to avoid assert failure in invlist_array() */
9330 invlist_set_len(invlist, len, offset);
9332 array = invlist_array(invlist);
9335 invlist_set_len(invlist, len, offset);
9338 /* The next item on the list starts the range, the one after that is
9339 * one past the new range. */
9340 array[len - 2] = start;
9341 if (end != UV_MAX) {
9342 array[len - 1] = end + 1;
9345 /* But if the end is the maximum representable on the machine, just let
9346 * the range have no end */
9347 invlist_set_len(invlist, len - 1, offset);
9352 Perl__invlist_search(SV* const invlist, const UV cp)
9354 /* Searches the inversion list for the entry that contains the input code
9355 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9356 * return value is the index into the list's array of the range that
9357 * contains <cp>, that is, 'i' such that
9358 * array[i] <= cp < array[i+1]
9363 IV high = _invlist_len(invlist);
9364 const IV highest_element = high - 1;
9367 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9369 /* If list is empty, return failure. */
9374 /* (We can't get the array unless we know the list is non-empty) */
9375 array = invlist_array(invlist);
9377 mid = invlist_previous_index(invlist);
9379 if (mid > highest_element) {
9380 mid = highest_element;
9383 /* <mid> contains the cache of the result of the previous call to this
9384 * function (0 the first time). See if this call is for the same result,
9385 * or if it is for mid-1. This is under the theory that calls to this
9386 * function will often be for related code points that are near each other.
9387 * And benchmarks show that caching gives better results. We also test
9388 * here if the code point is within the bounds of the list. These tests
9389 * replace others that would have had to be made anyway to make sure that
9390 * the array bounds were not exceeded, and these give us extra information
9391 * at the same time */
9392 if (cp >= array[mid]) {
9393 if (cp >= array[highest_element]) {
9394 return highest_element;
9397 /* Here, array[mid] <= cp < array[highest_element]. This means that
9398 * the final element is not the answer, so can exclude it; it also
9399 * means that <mid> is not the final element, so can refer to 'mid + 1'
9401 if (cp < array[mid + 1]) {
9407 else { /* cp < aray[mid] */
9408 if (cp < array[0]) { /* Fail if outside the array */
9412 if (cp >= array[mid - 1]) {
9417 /* Binary search. What we are looking for is <i> such that
9418 * array[i] <= cp < array[i+1]
9419 * The loop below converges on the i+1. Note that there may not be an
9420 * (i+1)th element in the array, and things work nonetheless */
9421 while (low < high) {
9422 mid = (low + high) / 2;
9423 assert(mid <= highest_element);
9424 if (array[mid] <= cp) { /* cp >= array[mid] */
9427 /* We could do this extra test to exit the loop early.
9428 if (cp < array[low]) {
9433 else { /* cp < array[mid] */
9440 invlist_set_previous_index(invlist, high);
9445 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9446 const bool complement_b, SV** output)
9448 /* Take the union of two inversion lists and point '*output' to it. On
9449 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9450 * even 'a' or 'b'). If to an inversion list, the contents of the original
9451 * list will be replaced by the union. The first list, 'a', may be
9452 * NULL, in which case a copy of the second list is placed in '*output'.
9453 * If 'complement_b' is TRUE, the union is taken of the complement
9454 * (inversion) of 'b' instead of b itself.
9456 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9457 * Richard Gillam, published by Addison-Wesley, and explained at some
9458 * length there. The preface says to incorporate its examples into your
9459 * code at your own risk.
9461 * The algorithm is like a merge sort. */
9463 const UV* array_a; /* a's array */
9465 UV len_a; /* length of a's array */
9468 SV* u; /* the resulting union */
9472 UV i_a = 0; /* current index into a's array */
9476 /* running count, as explained in the algorithm source book; items are
9477 * stopped accumulating and are output when the count changes to/from 0.
9478 * The count is incremented when we start a range that's in an input's set,
9479 * and decremented when we start a range that's not in a set. So this
9480 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9481 * and hence nothing goes into the union; 1, just one of the inputs is in
9482 * its set (and its current range gets added to the union); and 2 when both
9483 * inputs are in their sets. */
9486 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9488 assert(*output == NULL || is_invlist(*output));
9490 len_b = _invlist_len(b);
9493 /* Here, 'b' is empty, hence it's complement is all possible code
9494 * points. So if the union includes the complement of 'b', it includes
9495 * everything, and we need not even look at 'a'. It's easiest to
9496 * create a new inversion list that matches everything. */
9498 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9500 if (*output == NULL) { /* If the output didn't exist, just point it
9502 *output = everything;
9504 else { /* Otherwise, replace its contents with the new list */
9505 invlist_replace_list_destroys_src(*output, everything);
9506 SvREFCNT_dec_NN(everything);
9512 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9513 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9514 * output will be empty */
9516 if (a == NULL || _invlist_len(a) == 0) {
9517 if (*output == NULL) {
9518 *output = _new_invlist(0);
9521 invlist_clear(*output);
9526 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9527 * union. We can just return a copy of 'a' if '*output' doesn't point
9528 * to an existing list */
9529 if (*output == NULL) {
9530 *output = invlist_clone(a, NULL);
9534 /* If the output is to overwrite 'a', we have a no-op, as it's
9540 /* Here, '*output' is to be overwritten by 'a' */
9541 u = invlist_clone(a, NULL);
9542 invlist_replace_list_destroys_src(*output, u);
9548 /* Here 'b' is not empty. See about 'a' */
9550 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9552 /* Here, 'a' is empty (and b is not). That means the union will come
9553 * entirely from 'b'. If '*output' is NULL, we can directly return a
9554 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9557 SV ** dest = (*output == NULL) ? output : &u;
9558 *dest = invlist_clone(b, NULL);
9560 _invlist_invert(*dest);
9564 invlist_replace_list_destroys_src(*output, u);
9571 /* Here both lists exist and are non-empty */
9572 array_a = invlist_array(a);
9573 array_b = invlist_array(b);
9575 /* If are to take the union of 'a' with the complement of b, set it
9576 * up so are looking at b's complement. */
9579 /* To complement, we invert: if the first element is 0, remove it. To
9580 * do this, we just pretend the array starts one later */
9581 if (array_b[0] == 0) {
9587 /* But if the first element is not zero, we pretend the list starts
9588 * at the 0 that is always stored immediately before the array. */
9594 /* Size the union for the worst case: that the sets are completely
9596 u = _new_invlist(len_a + len_b);
9598 /* Will contain U+0000 if either component does */
9599 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9600 || (len_b > 0 && array_b[0] == 0));
9602 /* Go through each input list item by item, stopping when have exhausted
9604 while (i_a < len_a && i_b < len_b) {
9605 UV cp; /* The element to potentially add to the union's array */
9606 bool cp_in_set; /* is it in the the input list's set or not */
9608 /* We need to take one or the other of the two inputs for the union.
9609 * Since we are merging two sorted lists, we take the smaller of the
9610 * next items. In case of a tie, we take first the one that is in its
9611 * set. If we first took the one not in its set, it would decrement
9612 * the count, possibly to 0 which would cause it to be output as ending
9613 * the range, and the next time through we would take the same number,
9614 * and output it again as beginning the next range. By doing it the
9615 * opposite way, there is no possibility that the count will be
9616 * momentarily decremented to 0, and thus the two adjoining ranges will
9617 * be seamlessly merged. (In a tie and both are in the set or both not
9618 * in the set, it doesn't matter which we take first.) */
9619 if ( array_a[i_a] < array_b[i_b]
9620 || ( array_a[i_a] == array_b[i_b]
9621 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9623 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9624 cp = array_a[i_a++];
9627 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9628 cp = array_b[i_b++];
9631 /* Here, have chosen which of the two inputs to look at. Only output
9632 * if the running count changes to/from 0, which marks the
9633 * beginning/end of a range that's in the set */
9636 array_u[i_u++] = cp;
9643 array_u[i_u++] = cp;
9649 /* The loop above increments the index into exactly one of the input lists
9650 * each iteration, and ends when either index gets to its list end. That
9651 * means the other index is lower than its end, and so something is
9652 * remaining in that one. We decrement 'count', as explained below, if
9653 * that list is in its set. (i_a and i_b each currently index the element
9654 * beyond the one we care about.) */
9655 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9656 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9661 /* Above we decremented 'count' if the list that had unexamined elements in
9662 * it was in its set. This has made it so that 'count' being non-zero
9663 * means there isn't anything left to output; and 'count' equal to 0 means
9664 * that what is left to output is precisely that which is left in the
9665 * non-exhausted input list.
9667 * To see why, note first that the exhausted input obviously has nothing
9668 * left to add to the union. If it was in its set at its end, that means
9669 * the set extends from here to the platform's infinity, and hence so does
9670 * the union and the non-exhausted set is irrelevant. The exhausted set
9671 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9672 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9673 * 'count' remains at 1. This is consistent with the decremented 'count'
9674 * != 0 meaning there's nothing left to add to the union.
9676 * But if the exhausted input wasn't in its set, it contributed 0 to
9677 * 'count', and the rest of the union will be whatever the other input is.
9678 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9679 * otherwise it gets decremented to 0. This is consistent with 'count'
9680 * == 0 meaning the remainder of the union is whatever is left in the
9681 * non-exhausted list. */
9686 IV copy_count = len_a - i_a;
9687 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9688 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9690 else { /* The non-exhausted input is b */
9691 copy_count = len_b - i_b;
9692 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9694 len_u = i_u + copy_count;
9697 /* Set the result to the final length, which can change the pointer to
9698 * array_u, so re-find it. (Note that it is unlikely that this will
9699 * change, as we are shrinking the space, not enlarging it) */
9700 if (len_u != _invlist_len(u)) {
9701 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9703 array_u = invlist_array(u);
9706 if (*output == NULL) { /* Simply return the new inversion list */
9710 /* Otherwise, overwrite the inversion list that was in '*output'. We
9711 * could instead free '*output', and then set it to 'u', but experience
9712 * has shown [perl #127392] that if the input is a mortal, we can get a
9713 * huge build-up of these during regex compilation before they get
9715 invlist_replace_list_destroys_src(*output, u);
9723 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9724 const bool complement_b, SV** i)
9726 /* Take the intersection of two inversion lists and point '*i' to it. On
9727 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9728 * even 'a' or 'b'). If to an inversion list, the contents of the original
9729 * list will be replaced by the intersection. The first list, 'a', may be
9730 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9731 * TRUE, the result will be the intersection of 'a' and the complement (or
9732 * inversion) of 'b' instead of 'b' directly.
9734 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9735 * Richard Gillam, published by Addison-Wesley, and explained at some
9736 * length there. The preface says to incorporate its examples into your
9737 * code at your own risk. In fact, it had bugs
9739 * The algorithm is like a merge sort, and is essentially the same as the
9743 const UV* array_a; /* a's array */
9745 UV len_a; /* length of a's array */
9748 SV* r; /* the resulting intersection */
9752 UV i_a = 0; /* current index into a's array */
9756 /* running count of how many of the two inputs are postitioned at ranges
9757 * that are in their sets. As explained in the algorithm source book,
9758 * items are stopped accumulating and are output when the count changes
9759 * to/from 2. The count is incremented when we start a range that's in an
9760 * input's set, and decremented when we start a range that's not in a set.
9761 * Only when it is 2 are we in the intersection. */
9764 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9766 assert(*i == NULL || is_invlist(*i));
9768 /* Special case if either one is empty */
9769 len_a = (a == NULL) ? 0 : _invlist_len(a);
9770 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9771 if (len_a != 0 && complement_b) {
9773 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9774 * must be empty. Here, also we are using 'b's complement, which
9775 * hence must be every possible code point. Thus the intersection
9778 if (*i == a) { /* No-op */
9783 *i = invlist_clone(a, NULL);
9787 r = invlist_clone(a, NULL);
9788 invlist_replace_list_destroys_src(*i, r);
9793 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9794 * intersection must be empty */
9796 *i = _new_invlist(0);
9804 /* Here both lists exist and are non-empty */
9805 array_a = invlist_array(a);
9806 array_b = invlist_array(b);
9808 /* If are to take the intersection of 'a' with the complement of b, set it
9809 * up so are looking at b's complement. */
9812 /* To complement, we invert: if the first element is 0, remove it. To
9813 * do this, we just pretend the array starts one later */
9814 if (array_b[0] == 0) {
9820 /* But if the first element is not zero, we pretend the list starts
9821 * at the 0 that is always stored immediately before the array. */
9827 /* Size the intersection for the worst case: that the intersection ends up
9828 * fragmenting everything to be completely disjoint */
9829 r= _new_invlist(len_a + len_b);
9831 /* Will contain U+0000 iff both components do */
9832 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9833 && len_b > 0 && array_b[0] == 0);
9835 /* Go through each list item by item, stopping when have exhausted one of
9837 while (i_a < len_a && i_b < len_b) {
9838 UV cp; /* The element to potentially add to the intersection's
9840 bool cp_in_set; /* Is it in the input list's set or not */
9842 /* We need to take one or the other of the two inputs for the
9843 * intersection. Since we are merging two sorted lists, we take the
9844 * smaller of the next items. In case of a tie, we take first the one
9845 * that is not in its set (a difference from the union algorithm). If
9846 * we first took the one in its set, it would increment the count,
9847 * possibly to 2 which would cause it to be output as starting a range
9848 * in the intersection, and the next time through we would take that
9849 * same number, and output it again as ending the set. By doing the
9850 * opposite of this, there is no possibility that the count will be
9851 * momentarily incremented to 2. (In a tie and both are in the set or
9852 * both not in the set, it doesn't matter which we take first.) */
9853 if ( array_a[i_a] < array_b[i_b]
9854 || ( array_a[i_a] == array_b[i_b]
9855 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9857 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9858 cp = array_a[i_a++];
9861 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9865 /* Here, have chosen which of the two inputs to look at. Only output
9866 * if the running count changes to/from 2, which marks the
9867 * beginning/end of a range that's in the intersection */
9871 array_r[i_r++] = cp;
9876 array_r[i_r++] = cp;
9883 /* The loop above increments the index into exactly one of the input lists
9884 * each iteration, and ends when either index gets to its list end. That
9885 * means the other index is lower than its end, and so something is
9886 * remaining in that one. We increment 'count', as explained below, if the
9887 * exhausted list was in its set. (i_a and i_b each currently index the
9888 * element beyond the one we care about.) */
9889 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9890 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9895 /* Above we incremented 'count' if the exhausted list was in its set. This
9896 * has made it so that 'count' being below 2 means there is nothing left to
9897 * output; otheriwse what's left to add to the intersection is precisely
9898 * that which is left in the non-exhausted input list.
9900 * To see why, note first that the exhausted input obviously has nothing
9901 * left to affect the intersection. If it was in its set at its end, that
9902 * means the set extends from here to the platform's infinity, and hence
9903 * anything in the non-exhausted's list will be in the intersection, and
9904 * anything not in it won't be. Hence, the rest of the intersection is
9905 * precisely what's in the non-exhausted list The exhausted set also
9906 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9907 * it means 'count' is now at least 2. This is consistent with the
9908 * incremented 'count' being >= 2 means to add the non-exhausted list to
9911 * But if the exhausted input wasn't in its set, it contributed 0 to
9912 * 'count', and the intersection can't include anything further; the
9913 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9914 * incremented. This is consistent with 'count' being < 2 meaning nothing
9915 * further to add to the intersection. */
9916 if (count < 2) { /* Nothing left to put in the intersection. */
9919 else { /* copy the non-exhausted list, unchanged. */
9920 IV copy_count = len_a - i_a;
9921 if (copy_count > 0) { /* a is the one with stuff left */
9922 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9924 else { /* b is the one with stuff left */
9925 copy_count = len_b - i_b;
9926 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9928 len_r = i_r + copy_count;
9931 /* Set the result to the final length, which can change the pointer to
9932 * array_r, so re-find it. (Note that it is unlikely that this will
9933 * change, as we are shrinking the space, not enlarging it) */
9934 if (len_r != _invlist_len(r)) {
9935 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9937 array_r = invlist_array(r);
9940 if (*i == NULL) { /* Simply return the calculated intersection */
9943 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9944 instead free '*i', and then set it to 'r', but experience has
9945 shown [perl #127392] that if the input is a mortal, we can get a
9946 huge build-up of these during regex compilation before they get
9949 invlist_replace_list_destroys_src(*i, r);
9961 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9963 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9964 * set. A pointer to the inversion list is returned. This may actually be
9965 * a new list, in which case the passed in one has been destroyed. The
9966 * passed-in inversion list can be NULL, in which case a new one is created
9967 * with just the one range in it. The new list is not necessarily
9968 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9969 * result of this function. The gain would not be large, and in many
9970 * cases, this is called multiple times on a single inversion list, so
9971 * anything freed may almost immediately be needed again.
9973 * This used to mostly call the 'union' routine, but that is much more
9974 * heavyweight than really needed for a single range addition */
9976 UV* array; /* The array implementing the inversion list */
9977 UV len; /* How many elements in 'array' */
9978 SSize_t i_s; /* index into the invlist array where 'start'
9980 SSize_t i_e = 0; /* And the index where 'end' should go */
9981 UV cur_highest; /* The highest code point in the inversion list
9982 upon entry to this function */
9984 /* This range becomes the whole inversion list if none already existed */
9985 if (invlist == NULL) {
9986 invlist = _new_invlist(2);
9987 _append_range_to_invlist(invlist, start, end);
9991 /* Likewise, if the inversion list is currently empty */
9992 len = _invlist_len(invlist);
9994 _append_range_to_invlist(invlist, start, end);
9998 /* Starting here, we have to know the internals of the list */
9999 array = invlist_array(invlist);
10001 /* If the new range ends higher than the current highest ... */
10002 cur_highest = invlist_highest(invlist);
10003 if (end > cur_highest) {
10005 /* If the whole range is higher, we can just append it */
10006 if (start > cur_highest) {
10007 _append_range_to_invlist(invlist, start, end);
10011 /* Otherwise, add the portion that is higher ... */
10012 _append_range_to_invlist(invlist, cur_highest + 1, end);
10014 /* ... and continue on below to handle the rest. As a result of the
10015 * above append, we know that the index of the end of the range is the
10016 * final even numbered one of the array. Recall that the final element
10017 * always starts a range that extends to infinity. If that range is in
10018 * the set (meaning the set goes from here to infinity), it will be an
10019 * even index, but if it isn't in the set, it's odd, and the final
10020 * range in the set is one less, which is even. */
10021 if (end == UV_MAX) {
10029 /* We have dealt with appending, now see about prepending. If the new
10030 * range starts lower than the current lowest ... */
10031 if (start < array[0]) {
10033 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10034 * Let the union code handle it, rather than having to know the
10035 * trickiness in two code places. */
10036 if (UNLIKELY(start == 0)) {
10039 range_invlist = _new_invlist(2);
10040 _append_range_to_invlist(range_invlist, start, end);
10042 _invlist_union(invlist, range_invlist, &invlist);
10044 SvREFCNT_dec_NN(range_invlist);
10049 /* If the whole new range comes before the first entry, and doesn't
10050 * extend it, we have to insert it as an additional range */
10051 if (end < array[0] - 1) {
10053 goto splice_in_new_range;
10056 /* Here the new range adjoins the existing first range, extending it
10060 /* And continue on below to handle the rest. We know that the index of
10061 * the beginning of the range is the first one of the array */
10064 else { /* Not prepending any part of the new range to the existing list.
10065 * Find where in the list it should go. This finds i_s, such that:
10066 * invlist[i_s] <= start < array[i_s+1]
10068 i_s = _invlist_search(invlist, start);
10071 /* At this point, any extending before the beginning of the inversion list
10072 * and/or after the end has been done. This has made it so that, in the
10073 * code below, each endpoint of the new range is either in a range that is
10074 * in the set, or is in a gap between two ranges that are. This means we
10075 * don't have to worry about exceeding the array bounds.
10077 * Find where in the list the new range ends (but we can skip this if we
10078 * have already determined what it is, or if it will be the same as i_s,
10079 * which we already have computed) */
10081 i_e = (start == end)
10083 : _invlist_search(invlist, end);
10086 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10087 * is a range that goes to infinity there is no element at invlist[i_e+1],
10088 * so only the first relation holds. */
10090 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10092 /* Here, the ranges on either side of the beginning of the new range
10093 * are in the set, and this range starts in the gap between them.
10095 * The new range extends the range above it downwards if the new range
10096 * ends at or above that range's start */
10097 const bool extends_the_range_above = ( end == UV_MAX
10098 || end + 1 >= array[i_s+1]);
10100 /* The new range extends the range below it upwards if it begins just
10101 * after where that range ends */
10102 if (start == array[i_s]) {
10104 /* If the new range fills the entire gap between the other ranges,
10105 * they will get merged together. Other ranges may also get
10106 * merged, depending on how many of them the new range spans. In
10107 * the general case, we do the merge later, just once, after we
10108 * figure out how many to merge. But in the case where the new
10109 * range exactly spans just this one gap (possibly extending into
10110 * the one above), we do the merge here, and an early exit. This
10111 * is done here to avoid having to special case later. */
10112 if (i_e - i_s <= 1) {
10114 /* If i_e - i_s == 1, it means that the new range terminates
10115 * within the range above, and hence 'extends_the_range_above'
10116 * must be true. (If the range above it extends to infinity,
10117 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10118 * will be 0, so no harm done.) */
10119 if (extends_the_range_above) {
10120 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10121 invlist_set_len(invlist,
10123 *(get_invlist_offset_addr(invlist)));
10127 /* Here, i_e must == i_s. We keep them in sync, as they apply
10128 * to the same range, and below we are about to decrement i_s
10133 /* Here, the new range is adjacent to the one below. (It may also
10134 * span beyond the range above, but that will get resolved later.)
10135 * Extend the range below to include this one. */
10136 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10138 start = array[i_s];
10140 else if (extends_the_range_above) {
10142 /* Here the new range only extends the range above it, but not the
10143 * one below. It merges with the one above. Again, we keep i_e
10144 * and i_s in sync if they point to the same range */
10149 array[i_s] = start;
10153 /* Here, we've dealt with the new range start extending any adjoining
10156 * If the new range extends to infinity, it is now the final one,
10157 * regardless of what was there before */
10158 if (UNLIKELY(end == UV_MAX)) {
10159 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10163 /* If i_e started as == i_s, it has also been dealt with,
10164 * and been updated to the new i_s, which will fail the following if */
10165 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10167 /* Here, the ranges on either side of the end of the new range are in
10168 * the set, and this range ends in the gap between them.
10170 * If this range is adjacent to (hence extends) the range above it, it
10171 * becomes part of that range; likewise if it extends the range below,
10172 * it becomes part of that range */
10173 if (end + 1 == array[i_e+1]) {
10175 array[i_e] = start;
10177 else if (start <= array[i_e]) {
10178 array[i_e] = end + 1;
10185 /* If the range fits entirely in an existing range (as possibly already
10186 * extended above), it doesn't add anything new */
10187 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10191 /* Here, no part of the range is in the list. Must add it. It will
10192 * occupy 2 more slots */
10193 splice_in_new_range:
10195 invlist_extend(invlist, len + 2);
10196 array = invlist_array(invlist);
10197 /* Move the rest of the array down two slots. Don't include any
10199 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10201 /* Do the actual splice */
10202 array[i_e+1] = start;
10203 array[i_e+2] = end + 1;
10204 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10208 /* Here the new range crossed the boundaries of a pre-existing range. The
10209 * code above has adjusted things so that both ends are in ranges that are
10210 * in the set. This means everything in between must also be in the set.
10211 * Just squash things together */
10212 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10213 invlist_set_len(invlist,
10215 *(get_invlist_offset_addr(invlist)));
10221 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10222 UV** other_elements_ptr)
10224 /* Create and return an inversion list whose contents are to be populated
10225 * by the caller. The caller gives the number of elements (in 'size') and
10226 * the very first element ('element0'). This function will set
10227 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10228 * are to be placed.
10230 * Obviously there is some trust involved that the caller will properly
10231 * fill in the other elements of the array.
10233 * (The first element needs to be passed in, as the underlying code does
10234 * things differently depending on whether it is zero or non-zero) */
10236 SV* invlist = _new_invlist(size);
10239 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10241 invlist = add_cp_to_invlist(invlist, element0);
10242 offset = *get_invlist_offset_addr(invlist);
10244 invlist_set_len(invlist, size, offset);
10245 *other_elements_ptr = invlist_array(invlist) + 1;
10251 #ifndef PERL_IN_XSUB_RE
10253 Perl__invlist_invert(pTHX_ SV* const invlist)
10255 /* Complement the input inversion list. This adds a 0 if the list didn't
10256 * have a zero; removes it otherwise. As described above, the data
10257 * structure is set up so that this is very efficient */
10259 PERL_ARGS_ASSERT__INVLIST_INVERT;
10261 assert(! invlist_is_iterating(invlist));
10263 /* The inverse of matching nothing is matching everything */
10264 if (_invlist_len(invlist) == 0) {
10265 _append_range_to_invlist(invlist, 0, UV_MAX);
10269 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10273 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10275 /* Return a new inversion list that is a copy of the input one, which is
10276 * unchanged. The new list will not be mortal even if the old one was. */
10278 const STRLEN nominal_length = _invlist_len(invlist);
10279 const STRLEN physical_length = SvCUR(invlist);
10280 const bool offset = *(get_invlist_offset_addr(invlist));
10282 PERL_ARGS_ASSERT_INVLIST_CLONE;
10284 if (new_invlist == NULL) {
10285 new_invlist = _new_invlist(nominal_length);
10288 sv_upgrade(new_invlist, SVt_INVLIST);
10289 initialize_invlist_guts(new_invlist, nominal_length);
10292 *(get_invlist_offset_addr(new_invlist)) = offset;
10293 invlist_set_len(new_invlist, nominal_length, offset);
10294 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10296 return new_invlist;
10302 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10304 /* Get the contents of an inversion list into a string SV so that they can
10305 * be printed out. If 'traditional_style' is TRUE, it uses the format
10306 * traditionally done for debug tracing; otherwise it uses a format
10307 * suitable for just copying to the output, with blanks between ranges and
10308 * a dash between range components */
10312 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10313 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10315 if (traditional_style) {
10316 output = newSVpvs("\n");
10319 output = newSVpvs("");
10322 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10324 assert(! invlist_is_iterating(invlist));
10326 invlist_iterinit(invlist);
10327 while (invlist_iternext(invlist, &start, &end)) {
10328 if (end == UV_MAX) {
10329 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10330 start, intra_range_delimiter,
10331 inter_range_delimiter);
10333 else if (end != start) {
10334 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10336 intra_range_delimiter,
10337 end, inter_range_delimiter);
10340 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10341 start, inter_range_delimiter);
10345 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10346 SvCUR_set(output, SvCUR(output) - 1);
10352 #ifndef PERL_IN_XSUB_RE
10354 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10355 const char * const indent, SV* const invlist)
10357 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10358 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10359 * the string 'indent'. The output looks like this:
10360 [0] 0x000A .. 0x000D
10362 [4] 0x2028 .. 0x2029
10363 [6] 0x3104 .. INFTY
10364 * This means that the first range of code points matched by the list are
10365 * 0xA through 0xD; the second range contains only the single code point
10366 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10367 * are used to define each range (except if the final range extends to
10368 * infinity, only a single element is needed). The array index of the
10369 * first element for the corresponding range is given in brackets. */
10374 PERL_ARGS_ASSERT__INVLIST_DUMP;
10376 if (invlist_is_iterating(invlist)) {
10377 Perl_dump_indent(aTHX_ level, file,
10378 "%sCan't dump inversion list because is in middle of iterating\n",
10383 invlist_iterinit(invlist);
10384 while (invlist_iternext(invlist, &start, &end)) {
10385 if (end == UV_MAX) {
10386 Perl_dump_indent(aTHX_ level, file,
10387 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10388 indent, (UV)count, start);
10390 else if (end != start) {
10391 Perl_dump_indent(aTHX_ level, file,
10392 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10393 indent, (UV)count, start, end);
10396 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10397 indent, (UV)count, start);
10405 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10407 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10409 /* Return a boolean as to if the two passed in inversion lists are
10410 * identical. The final argument, if TRUE, says to take the complement of
10411 * the second inversion list before doing the comparison */
10413 const UV len_a = _invlist_len(a);
10414 UV len_b = _invlist_len(b);
10416 const UV* array_a = NULL;
10417 const UV* array_b = NULL;
10419 PERL_ARGS_ASSERT__INVLISTEQ;
10421 /* This code avoids accessing the arrays unless it knows the length is
10426 return ! complement_b;
10430 array_a = invlist_array(a);
10434 array_b = invlist_array(b);
10437 /* If are to compare 'a' with the complement of b, set it
10438 * up so are looking at b's complement. */
10439 if (complement_b) {
10441 /* The complement of nothing is everything, so <a> would have to have
10442 * just one element, starting at zero (ending at infinity) */
10444 return (len_a == 1 && array_a[0] == 0);
10446 if (array_b[0] == 0) {
10448 /* Otherwise, to complement, we invert. Here, the first element is
10449 * 0, just remove it. To do this, we just pretend the array starts
10457 /* But if the first element is not zero, we pretend the list starts
10458 * at the 0 that is always stored immediately before the array. */
10464 return len_a == len_b
10465 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10471 * As best we can, determine the characters that can match the start of
10472 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10473 * can be false positive matches
10475 * Returns the invlist as a new SV*; it is the caller's responsibility to
10476 * call SvREFCNT_dec() when done with it.
10479 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10482 const U8 * s = (U8*)STRING(node);
10483 SSize_t bytelen = STR_LEN(node);
10485 /* Start out big enough for 2 separate code points */
10486 SV* invlist = _new_invlist(4);
10488 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10493 /* We punt and assume can match anything if the node begins
10494 * with a multi-character fold. Things are complicated. For
10495 * example, /ffi/i could match any of:
10496 * "\N{LATIN SMALL LIGATURE FFI}"
10497 * "\N{LATIN SMALL LIGATURE FF}I"
10498 * "F\N{LATIN SMALL LIGATURE FI}"
10499 * plus several other things; and making sure we have all the
10500 * possibilities is hard. */
10501 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10502 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10505 /* Any Latin1 range character can potentially match any
10506 * other depending on the locale, and in Turkic locales, U+130 and
10508 if (OP(node) == EXACTFL) {
10509 _invlist_union(invlist, PL_Latin1, &invlist);
10510 invlist = add_cp_to_invlist(invlist,
10511 LATIN_SMALL_LETTER_DOTLESS_I);
10512 invlist = add_cp_to_invlist(invlist,
10513 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10516 /* But otherwise, it matches at least itself. We can
10517 * quickly tell if it has a distinct fold, and if so,
10518 * it matches that as well */
10519 invlist = add_cp_to_invlist(invlist, uc);
10520 if (IS_IN_SOME_FOLD_L1(uc))
10521 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10524 /* Some characters match above-Latin1 ones under /i. This
10525 * is true of EXACTFL ones when the locale is UTF-8 */
10526 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10527 && (! isASCII(uc) || (OP(node) != EXACTFAA
10528 && OP(node) != EXACTFAA_NO_TRIE)))
10530 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10534 else { /* Pattern is UTF-8 */
10535 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10536 const U8* e = s + bytelen;
10539 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10541 /* The only code points that aren't folded in a UTF EXACTFish
10542 * node are are the problematic ones in EXACTFL nodes */
10543 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10544 /* We need to check for the possibility that this EXACTFL
10545 * node begins with a multi-char fold. Therefore we fold
10546 * the first few characters of it so that we can make that
10552 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10554 *(d++) = (U8) toFOLD(*s);
10555 if (fc < 0) { /* Save the first fold */
10562 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10563 if (fc < 0) { /* Save the first fold */
10571 /* And set up so the code below that looks in this folded
10572 * buffer instead of the node's string */
10577 /* When we reach here 's' points to the fold of the first
10578 * character(s) of the node; and 'e' points to far enough along
10579 * the folded string to be just past any possible multi-char
10582 * Unlike the non-UTF-8 case, the macro for determining if a
10583 * string is a multi-char fold requires all the characters to
10584 * already be folded. This is because of all the complications
10585 * if not. Note that they are folded anyway, except in EXACTFL
10586 * nodes. Like the non-UTF case above, we punt if the node
10587 * begins with a multi-char fold */
10589 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10590 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10592 else { /* Single char fold */
10594 unsigned int first_fold;
10595 const unsigned int * remaining_folds;
10596 Size_t folds_count;
10598 /* It matches itself */
10599 invlist = add_cp_to_invlist(invlist, fc);
10601 /* ... plus all the things that fold to it, which are found in
10602 * PL_utf8_foldclosures */
10603 folds_count = _inverse_folds(fc, &first_fold,
10605 for (k = 0; k < folds_count; k++) {
10606 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10608 /* /aa doesn't allow folds between ASCII and non- */
10609 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10610 && isASCII(c) != isASCII(fc))
10615 invlist = add_cp_to_invlist(invlist, c);
10618 if (OP(node) == EXACTFL) {
10620 /* If either [iI] are present in an EXACTFL node the above code
10621 * should have added its normal case pair, but under a Turkish
10622 * locale they could match instead the case pairs from it. Add
10623 * those as potential matches as well */
10624 if (isALPHA_FOLD_EQ(fc, 'I')) {
10625 invlist = add_cp_to_invlist(invlist,
10626 LATIN_SMALL_LETTER_DOTLESS_I);
10627 invlist = add_cp_to_invlist(invlist,
10628 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10630 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10631 invlist = add_cp_to_invlist(invlist, 'I');
10633 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10634 invlist = add_cp_to_invlist(invlist, 'i');
10643 #undef HEADER_LENGTH
10644 #undef TO_INTERNAL_SIZE
10645 #undef FROM_INTERNAL_SIZE
10646 #undef INVLIST_VERSION_ID
10648 /* End of inversion list object */
10651 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10653 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10654 * constructs, and updates RExC_flags with them. On input, RExC_parse
10655 * should point to the first flag; it is updated on output to point to the
10656 * final ')' or ':'. There needs to be at least one flag, or this will
10659 /* for (?g), (?gc), and (?o) warnings; warning
10660 about (?c) will warn about (?g) -- japhy */
10662 #define WASTED_O 0x01
10663 #define WASTED_G 0x02
10664 #define WASTED_C 0x04
10665 #define WASTED_GC (WASTED_G|WASTED_C)
10666 I32 wastedflags = 0x00;
10667 U32 posflags = 0, negflags = 0;
10668 U32 *flagsp = &posflags;
10669 char has_charset_modifier = '\0';
10671 bool has_use_defaults = FALSE;
10672 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10673 int x_mod_count = 0;
10675 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10677 /* '^' as an initial flag sets certain defaults */
10678 if (UCHARAT(RExC_parse) == '^') {
10680 has_use_defaults = TRUE;
10681 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10682 cs = (RExC_uni_semantics)
10683 ? REGEX_UNICODE_CHARSET
10684 : REGEX_DEPENDS_CHARSET;
10685 set_regex_charset(&RExC_flags, cs);
10688 cs = get_regex_charset(RExC_flags);
10689 if ( cs == REGEX_DEPENDS_CHARSET
10690 && RExC_uni_semantics)
10692 cs = REGEX_UNICODE_CHARSET;
10696 while (RExC_parse < RExC_end) {
10697 /* && strchr("iogcmsx", *RExC_parse) */
10698 /* (?g), (?gc) and (?o) are useless here
10699 and must be globally applied -- japhy */
10700 switch (*RExC_parse) {
10702 /* Code for the imsxn flags */
10703 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10705 case LOCALE_PAT_MOD:
10706 if (has_charset_modifier) {
10707 goto excess_modifier;
10709 else if (flagsp == &negflags) {
10712 cs = REGEX_LOCALE_CHARSET;
10713 has_charset_modifier = LOCALE_PAT_MOD;
10715 case UNICODE_PAT_MOD:
10716 if (has_charset_modifier) {
10717 goto excess_modifier;
10719 else if (flagsp == &negflags) {
10722 cs = REGEX_UNICODE_CHARSET;
10723 has_charset_modifier = UNICODE_PAT_MOD;
10725 case ASCII_RESTRICT_PAT_MOD:
10726 if (flagsp == &negflags) {
10729 if (has_charset_modifier) {
10730 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10731 goto excess_modifier;
10733 /* Doubled modifier implies more restricted */
10734 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10737 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10739 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10741 case DEPENDS_PAT_MOD:
10742 if (has_use_defaults) {
10743 goto fail_modifiers;
10745 else if (flagsp == &negflags) {
10748 else if (has_charset_modifier) {
10749 goto excess_modifier;
10752 /* The dual charset means unicode semantics if the
10753 * pattern (or target, not known until runtime) are
10754 * utf8, or something in the pattern indicates unicode
10756 cs = (RExC_uni_semantics)
10757 ? REGEX_UNICODE_CHARSET
10758 : REGEX_DEPENDS_CHARSET;
10759 has_charset_modifier = DEPENDS_PAT_MOD;
10763 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10764 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10766 else if (has_charset_modifier == *(RExC_parse - 1)) {
10767 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10768 *(RExC_parse - 1));
10771 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10773 NOT_REACHED; /*NOTREACHED*/
10776 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10777 *(RExC_parse - 1));
10778 NOT_REACHED; /*NOTREACHED*/
10779 case ONCE_PAT_MOD: /* 'o' */
10780 case GLOBAL_PAT_MOD: /* 'g' */
10781 if (ckWARN(WARN_REGEXP)) {
10782 const I32 wflagbit = *RExC_parse == 'o'
10785 if (! (wastedflags & wflagbit) ) {
10786 wastedflags |= wflagbit;
10787 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10790 "Useless (%s%c) - %suse /%c modifier",
10791 flagsp == &negflags ? "?-" : "?",
10793 flagsp == &negflags ? "don't " : "",
10800 case CONTINUE_PAT_MOD: /* 'c' */
10801 if (ckWARN(WARN_REGEXP)) {
10802 if (! (wastedflags & WASTED_C) ) {
10803 wastedflags |= WASTED_GC;
10804 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10807 "Useless (%sc) - %suse /gc modifier",
10808 flagsp == &negflags ? "?-" : "?",
10809 flagsp == &negflags ? "don't " : ""
10814 case KEEPCOPY_PAT_MOD: /* 'p' */
10815 if (flagsp == &negflags) {
10816 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10818 *flagsp |= RXf_PMf_KEEPCOPY;
10822 /* A flag is a default iff it is following a minus, so
10823 * if there is a minus, it means will be trying to
10824 * re-specify a default which is an error */
10825 if (has_use_defaults || flagsp == &negflags) {
10826 goto fail_modifiers;
10828 flagsp = &negflags;
10829 wastedflags = 0; /* reset so (?g-c) warns twice */
10835 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10836 negflags |= RXf_PMf_EXTENDED_MORE;
10838 RExC_flags |= posflags;
10840 if (negflags & RXf_PMf_EXTENDED) {
10841 negflags |= RXf_PMf_EXTENDED_MORE;
10843 RExC_flags &= ~negflags;
10844 set_regex_charset(&RExC_flags, cs);
10849 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
10850 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10851 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10852 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10853 NOT_REACHED; /*NOTREACHED*/
10856 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10859 vFAIL("Sequence (?... not terminated");
10863 - reg - regular expression, i.e. main body or parenthesized thing
10865 * Caller must absorb opening parenthesis.
10867 * Combining parenthesis handling with the base level of regular expression
10868 * is a trifle forced, but the need to tie the tails of the branches to what
10869 * follows makes it hard to avoid.
10871 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10873 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10875 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10878 PERL_STATIC_INLINE regnode_offset
10879 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10881 char * parse_start,
10885 regnode_offset ret;
10886 char* name_start = RExC_parse;
10888 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
10889 GET_RE_DEBUG_FLAGS_DECL;
10891 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10893 if (RExC_parse == name_start || *RExC_parse != ch) {
10894 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10895 vFAIL2("Sequence %.3s... not terminated", parse_start);
10899 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10900 RExC_rxi->data->data[num]=(void*)sv_dat;
10901 SvREFCNT_inc_simple_void_NN(sv_dat);
10904 ret = reganode(pRExC_state,
10907 : (ASCII_FOLD_RESTRICTED)
10909 : (AT_LEAST_UNI_SEMANTICS)
10915 *flagp |= HASWIDTH;
10917 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
10918 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
10920 nextchar(pRExC_state);
10924 /* On success, returns the offset at which any next node should be placed into
10925 * the regex engine program being compiled.
10927 * Returns 0 otherwise, with *flagp set to indicate why:
10928 * TRYAGAIN at the end of (?) that only sets flags.
10929 * RESTART_PARSE if the parse needs to be restarted, or'd with
10930 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
10931 * Otherwise would only return 0 if regbranch() returns 0, which cannot
10933 STATIC regnode_offset
10934 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
10935 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10936 * 2 is like 1, but indicates that nextchar() has been called to advance
10937 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10938 * this flag alerts us to the need to check for that */
10940 regnode_offset ret = 0; /* Will be the head of the group. */
10942 regnode_offset lastbr;
10943 regnode_offset ender = 0;
10946 U32 oregflags = RExC_flags;
10947 bool have_branch = 0;
10949 I32 freeze_paren = 0;
10950 I32 after_freeze = 0;
10951 I32 num; /* numeric backreferences */
10952 SV * max_open; /* Max number of unclosed parens */
10954 char * parse_start = RExC_parse; /* MJD */
10955 char * const oregcomp_parse = RExC_parse;
10957 GET_RE_DEBUG_FLAGS_DECL;
10959 PERL_ARGS_ASSERT_REG;
10960 DEBUG_PARSE("reg ");
10963 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
10965 if (!SvIOK(max_open)) {
10966 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
10968 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
10970 vFAIL("Too many nested open parens");
10973 *flagp = 0; /* Tentatively. */
10975 if (RExC_in_lookbehind) {
10976 RExC_in_lookbehind++;
10978 if (RExC_in_lookahead) {
10979 RExC_in_lookahead++;
10982 /* Having this true makes it feasible to have a lot fewer tests for the
10983 * parse pointer being in scope. For example, we can write
10984 * while(isFOO(*RExC_parse)) RExC_parse++;
10986 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
10988 assert(*RExC_end == '\0');
10990 /* Make an OPEN node, if parenthesized. */
10993 /* Under /x, space and comments can be gobbled up between the '(' and
10994 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
10995 * intervening space, as the sequence is a token, and a token should be
10997 bool has_intervening_patws = (paren == 2)
10998 && *(RExC_parse - 1) != '(';
11000 if (RExC_parse >= RExC_end) {
11001 vFAIL("Unmatched (");
11004 if (paren == 'r') { /* Atomic script run */
11008 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11009 char *start_verb = RExC_parse + 1;
11011 char *start_arg = NULL;
11012 unsigned char op = 0;
11013 int arg_required = 0;
11014 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11015 bool has_upper = FALSE;
11017 if (has_intervening_patws) {
11018 RExC_parse++; /* past the '*' */
11020 /* For strict backwards compatibility, don't change the message
11021 * now that we also have lowercase operands */
11022 if (isUPPER(*RExC_parse)) {
11023 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11026 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11029 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11030 if ( *RExC_parse == ':' ) {
11031 start_arg = RExC_parse + 1;
11035 if (isUPPER(*RExC_parse)) {
11041 RExC_parse += UTF8SKIP(RExC_parse);
11044 verb_len = RExC_parse - start_verb;
11046 if (RExC_parse >= RExC_end) {
11047 goto unterminated_verb_pattern;
11050 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11051 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11052 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11054 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11055 unterminated_verb_pattern:
11057 vFAIL("Unterminated verb pattern argument");
11060 vFAIL("Unterminated '(*...' argument");
11064 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11066 vFAIL("Unterminated verb pattern");
11069 vFAIL("Unterminated '(*...' construct");
11074 /* Here, we know that RExC_parse < RExC_end */
11076 switch ( *start_verb ) {
11077 case 'A': /* (*ACCEPT) */
11078 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11080 internal_argval = RExC_nestroot;
11083 case 'C': /* (*COMMIT) */
11084 if ( memEQs(start_verb, verb_len,"COMMIT") )
11087 case 'F': /* (*FAIL) */
11088 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11092 case ':': /* (*:NAME) */
11093 case 'M': /* (*MARK:NAME) */
11094 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11099 case 'P': /* (*PRUNE) */
11100 if ( memEQs(start_verb, verb_len,"PRUNE") )
11103 case 'S': /* (*SKIP) */
11104 if ( memEQs(start_verb, verb_len,"SKIP") )
11107 case 'T': /* (*THEN) */
11108 /* [19:06] <TimToady> :: is then */
11109 if ( memEQs(start_verb, verb_len,"THEN") ) {
11111 RExC_seen |= REG_CUTGROUP_SEEN;
11115 if ( memEQs(start_verb, verb_len, "asr")
11116 || memEQs(start_verb, verb_len, "atomic_script_run"))
11118 paren = 'r'; /* Mnemonic: recursed run */
11121 else if (memEQs(start_verb, verb_len, "atomic")) {
11122 paren = 't'; /* AtOMIC */
11123 goto alpha_assertions;
11127 if ( memEQs(start_verb, verb_len, "plb")
11128 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11131 goto lookbehind_alpha_assertions;
11133 else if ( memEQs(start_verb, verb_len, "pla")
11134 || memEQs(start_verb, verb_len, "positive_lookahead"))
11137 goto alpha_assertions;
11141 if ( memEQs(start_verb, verb_len, "nlb")
11142 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11145 goto lookbehind_alpha_assertions;
11147 else if ( memEQs(start_verb, verb_len, "nla")
11148 || memEQs(start_verb, verb_len, "negative_lookahead"))
11151 goto alpha_assertions;
11155 if ( memEQs(start_verb, verb_len, "sr")
11156 || memEQs(start_verb, verb_len, "script_run"))
11158 regnode_offset atomic;
11164 /* This indicates Unicode rules. */
11165 REQUIRE_UNI_RULES(flagp, 0);
11171 RExC_parse = start_arg;
11173 if (RExC_in_script_run) {
11175 /* Nested script runs are treated as no-ops, because
11176 * if the nested one fails, the outer one must as
11177 * well. It could fail sooner, and avoid (??{} with
11178 * side effects, but that is explicitly documented as
11179 * undefined behavior. */
11183 if (paren == 's') {
11188 /* But, the atomic part of a nested atomic script run
11189 * isn't a no-op, but can be treated just like a '(?>'
11195 if (paren == 's') {
11196 /* Here, we're starting a new regular script run */
11197 ret = reg_node(pRExC_state, SROPEN);
11198 RExC_in_script_run = 1;
11203 /* Here, we are starting an atomic script run. This is
11204 * handled by recursing to deal with the atomic portion
11205 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11207 ret = reg_node(pRExC_state, SROPEN);
11209 RExC_in_script_run = 1;
11211 atomic = reg(pRExC_state, 'r', &flags, depth);
11212 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11213 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11217 if (! REGTAIL(pRExC_state, ret, atomic)) {
11218 REQUIRE_BRANCHJ(flagp, 0);
11221 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11224 REQUIRE_BRANCHJ(flagp, 0);
11227 RExC_in_script_run = 0;
11233 lookbehind_alpha_assertions:
11234 RExC_seen |= REG_LOOKBEHIND_SEEN;
11235 RExC_in_lookbehind++;
11240 RExC_seen_zerolen++;
11246 /* An empty negative lookahead assertion simply is failure */
11247 if (paren == 'A' && RExC_parse == start_arg) {
11248 ret=reganode(pRExC_state, OPFAIL, 0);
11249 nextchar(pRExC_state);
11253 RExC_parse = start_arg;
11258 "'(*%" UTF8f "' requires a terminating ':'",
11259 UTF8fARG(UTF, verb_len, start_verb));
11260 NOT_REACHED; /*NOTREACHED*/
11262 } /* End of switch */
11265 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11267 if (has_upper || verb_len == 0) {
11269 "Unknown verb pattern '%" UTF8f "'",
11270 UTF8fARG(UTF, verb_len, start_verb));
11274 "Unknown '(*...)' construct '%" UTF8f "'",
11275 UTF8fARG(UTF, verb_len, start_verb));
11278 if ( RExC_parse == start_arg ) {
11281 if ( arg_required && !start_arg ) {
11282 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11283 verb_len, start_verb);
11285 if (internal_argval == -1) {
11286 ret = reganode(pRExC_state, op, 0);
11288 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11290 RExC_seen |= REG_VERBARG_SEEN;
11292 SV *sv = newSVpvn( start_arg,
11293 RExC_parse - start_arg);
11294 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11295 STR_WITH_LEN("S"));
11296 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11297 FLAGS(REGNODE_p(ret)) = 1;
11299 FLAGS(REGNODE_p(ret)) = 0;
11301 if ( internal_argval != -1 )
11302 ARG2L_SET(REGNODE_p(ret), internal_argval);
11303 nextchar(pRExC_state);
11306 else if (*RExC_parse == '?') { /* (?...) */
11307 bool is_logical = 0;
11308 const char * const seqstart = RExC_parse;
11309 const char * endptr;
11310 if (has_intervening_patws) {
11312 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11315 RExC_parse++; /* past the '?' */
11316 paren = *RExC_parse; /* might be a trailing NUL, if not
11318 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11319 if (RExC_parse > RExC_end) {
11322 ret = 0; /* For look-ahead/behind. */
11325 case 'P': /* (?P...) variants for those used to PCRE/Python */
11326 paren = *RExC_parse;
11327 if ( paren == '<') { /* (?P<...>) named capture */
11329 if (RExC_parse >= RExC_end) {
11330 vFAIL("Sequence (?P<... not terminated");
11332 goto named_capture;
11334 else if (paren == '>') { /* (?P>name) named recursion */
11336 if (RExC_parse >= RExC_end) {
11337 vFAIL("Sequence (?P>... not terminated");
11339 goto named_recursion;
11341 else if (paren == '=') { /* (?P=...) named backref */
11343 return handle_named_backref(pRExC_state, flagp,
11346 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11347 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11348 vFAIL3("Sequence (%.*s...) not recognized",
11349 RExC_parse-seqstart, seqstart);
11350 NOT_REACHED; /*NOTREACHED*/
11351 case '<': /* (?<...) */
11352 if (*RExC_parse == '!')
11354 else if (*RExC_parse != '=')
11361 case '\'': /* (?'...') */
11362 name_start = RExC_parse;
11363 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11364 if ( RExC_parse == name_start
11365 || RExC_parse >= RExC_end
11366 || *RExC_parse != paren)
11368 vFAIL2("Sequence (?%c... not terminated",
11369 paren=='>' ? '<' : paren);
11374 if (!svname) /* shouldn't happen */
11376 "panic: reg_scan_name returned NULL");
11377 if (!RExC_paren_names) {
11378 RExC_paren_names= newHV();
11379 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11381 RExC_paren_name_list= newAV();
11382 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11385 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11387 sv_dat = HeVAL(he_str);
11389 /* croak baby croak */
11391 "panic: paren_name hash element allocation failed");
11392 } else if ( SvPOK(sv_dat) ) {
11393 /* (?|...) can mean we have dupes so scan to check
11394 its already been stored. Maybe a flag indicating
11395 we are inside such a construct would be useful,
11396 but the arrays are likely to be quite small, so
11397 for now we punt -- dmq */
11398 IV count = SvIV(sv_dat);
11399 I32 *pv = (I32*)SvPVX(sv_dat);
11401 for ( i = 0 ; i < count ; i++ ) {
11402 if ( pv[i] == RExC_npar ) {
11408 pv = (I32*)SvGROW(sv_dat,
11409 SvCUR(sv_dat) + sizeof(I32)+1);
11410 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11411 pv[count] = RExC_npar;
11412 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11415 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11416 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11419 SvIV_set(sv_dat, 1);
11422 /* Yes this does cause a memory leak in debugging Perls
11424 if (!av_store(RExC_paren_name_list,
11425 RExC_npar, SvREFCNT_inc_NN(svname)))
11426 SvREFCNT_dec_NN(svname);
11429 /*sv_dump(sv_dat);*/
11431 nextchar(pRExC_state);
11433 goto capturing_parens;
11436 RExC_seen |= REG_LOOKBEHIND_SEEN;
11437 RExC_in_lookbehind++;
11439 if (RExC_parse >= RExC_end) {
11440 vFAIL("Sequence (?... not terminated");
11442 RExC_seen_zerolen++;
11444 case '=': /* (?=...) */
11445 RExC_seen_zerolen++;
11446 RExC_in_lookahead++;
11448 case '!': /* (?!...) */
11449 RExC_seen_zerolen++;
11450 /* check if we're really just a "FAIL" assertion */
11451 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11452 FALSE /* Don't force to /x */ );
11453 if (*RExC_parse == ')') {
11454 ret=reganode(pRExC_state, OPFAIL, 0);
11455 nextchar(pRExC_state);
11459 case '|': /* (?|...) */
11460 /* branch reset, behave like a (?:...) except that
11461 buffers in alternations share the same numbers */
11463 after_freeze = freeze_paren = RExC_npar;
11465 /* XXX This construct currently requires an extra pass.
11466 * Investigation would be required to see if that could be
11468 REQUIRE_PARENS_PASS;
11470 case ':': /* (?:...) */
11471 case '>': /* (?>...) */
11473 case '$': /* (?$...) */
11474 case '@': /* (?@...) */
11475 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11477 case '0' : /* (?0) */
11478 case 'R' : /* (?R) */
11479 if (RExC_parse == RExC_end || *RExC_parse != ')')
11480 FAIL("Sequence (?R) not terminated");
11482 RExC_seen |= REG_RECURSE_SEEN;
11484 /* XXX These constructs currently require an extra pass.
11485 * It probably could be changed */
11486 REQUIRE_PARENS_PASS;
11488 *flagp |= POSTPONED;
11489 goto gen_recurse_regop;
11491 /* named and numeric backreferences */
11492 case '&': /* (?&NAME) */
11493 parse_start = RExC_parse - 1;
11496 SV *sv_dat = reg_scan_name(pRExC_state,
11497 REG_RSN_RETURN_DATA);
11498 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11500 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11501 vFAIL("Sequence (?&... not terminated");
11502 goto gen_recurse_regop;
11505 if (! inRANGE(RExC_parse[0], '1', '9')) {
11507 vFAIL("Illegal pattern");
11509 goto parse_recursion;
11511 case '-': /* (?-1) */
11512 if (! inRANGE(RExC_parse[0], '1', '9')) {
11513 RExC_parse--; /* rewind to let it be handled later */
11517 case '1': case '2': case '3': case '4': /* (?1) */
11518 case '5': case '6': case '7': case '8': case '9':
11519 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11522 bool is_neg = FALSE;
11524 parse_start = RExC_parse - 1; /* MJD */
11525 if (*RExC_parse == '-') {
11530 if (grok_atoUV(RExC_parse, &unum, &endptr)
11534 RExC_parse = (char*)endptr;
11538 /* Some limit for num? */
11542 if (*RExC_parse!=')')
11543 vFAIL("Expecting close bracket");
11546 if ( paren == '-' ) {
11548 Diagram of capture buffer numbering.
11549 Top line is the normal capture buffer numbers
11550 Bottom line is the negative indexing as from
11554 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11558 num = RExC_npar + num;
11561 /* It might be a forward reference; we can't fail until
11562 * we know, by completing the parse to get all the
11563 * groups, and then reparsing */
11564 if (ALL_PARENS_COUNTED) {
11566 vFAIL("Reference to nonexistent group");
11569 REQUIRE_PARENS_PASS;
11572 } else if ( paren == '+' ) {
11573 num = RExC_npar + num - 1;
11575 /* We keep track how many GOSUB items we have produced.
11576 To start off the ARG2L() of the GOSUB holds its "id",
11577 which is used later in conjunction with RExC_recurse
11578 to calculate the offset we need to jump for the GOSUB,
11579 which it will store in the final representation.
11580 We have to defer the actual calculation until much later
11581 as the regop may move.
11584 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11585 if (num >= RExC_npar) {
11587 /* It might be a forward reference; we can't fail until we
11588 * know, by completing the parse to get all the groups, and
11589 * then reparsing */
11590 if (ALL_PARENS_COUNTED) {
11591 if (num >= RExC_total_parens) {
11593 vFAIL("Reference to nonexistent group");
11597 REQUIRE_PARENS_PASS;
11600 RExC_recurse_count++;
11601 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11602 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11603 22, "| |", (int)(depth * 2 + 1), "",
11604 (UV)ARG(REGNODE_p(ret)),
11605 (IV)ARG2L(REGNODE_p(ret))));
11606 RExC_seen |= REG_RECURSE_SEEN;
11608 Set_Node_Length(REGNODE_p(ret),
11609 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11610 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11612 *flagp |= POSTPONED;
11613 assert(*RExC_parse == ')');
11614 nextchar(pRExC_state);
11619 case '?': /* (??...) */
11621 if (*RExC_parse != '{') {
11622 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11623 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11625 "Sequence (%" UTF8f "...) not recognized",
11626 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11627 NOT_REACHED; /*NOTREACHED*/
11629 *flagp |= POSTPONED;
11633 case '{': /* (?{...}) */
11636 struct reg_code_block *cb;
11639 RExC_seen_zerolen++;
11641 if ( !pRExC_state->code_blocks
11642 || pRExC_state->code_index
11643 >= pRExC_state->code_blocks->count
11644 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11645 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11648 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11649 FAIL("panic: Sequence (?{...}): no code block found\n");
11650 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11652 /* this is a pre-compiled code block (?{...}) */
11653 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11654 RExC_parse = RExC_start + cb->end;
11656 if (cb->src_regex) {
11657 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11658 RExC_rxi->data->data[n] =
11659 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11660 RExC_rxi->data->data[n+1] = (void*)o;
11663 n = add_data(pRExC_state,
11664 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11665 RExC_rxi->data->data[n] = (void*)o;
11667 pRExC_state->code_index++;
11668 nextchar(pRExC_state);
11671 regnode_offset eval;
11672 ret = reg_node(pRExC_state, LOGICAL);
11674 eval = reg2Lanode(pRExC_state, EVAL,
11677 /* for later propagation into (??{})
11679 RExC_flags & RXf_PMf_COMPILETIME
11681 FLAGS(REGNODE_p(ret)) = 2;
11682 if (! REGTAIL(pRExC_state, ret, eval)) {
11683 REQUIRE_BRANCHJ(flagp, 0);
11685 /* deal with the length of this later - MJD */
11688 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11689 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11690 Set_Node_Offset(REGNODE_p(ret), parse_start);
11693 case '(': /* (?(?{...})...) and (?(?=...)...) */
11696 const int DEFINE_len = sizeof("DEFINE") - 1;
11697 if ( RExC_parse < RExC_end - 1
11698 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11699 && ( RExC_parse[1] == '='
11700 || RExC_parse[1] == '!'
11701 || RExC_parse[1] == '<'
11702 || RExC_parse[1] == '{'))
11703 || ( RExC_parse[0] == '*' /* (?(*...)) */
11704 && ( memBEGINs(RExC_parse + 1,
11705 (Size_t) (RExC_end - (RExC_parse + 1)),
11707 || memBEGINs(RExC_parse + 1,
11708 (Size_t) (RExC_end - (RExC_parse + 1)),
11710 || memBEGINs(RExC_parse + 1,
11711 (Size_t) (RExC_end - (RExC_parse + 1)),
11713 || memBEGINs(RExC_parse + 1,
11714 (Size_t) (RExC_end - (RExC_parse + 1)),
11716 || memBEGINs(RExC_parse + 1,
11717 (Size_t) (RExC_end - (RExC_parse + 1)),
11718 "positive_lookahead:")
11719 || memBEGINs(RExC_parse + 1,
11720 (Size_t) (RExC_end - (RExC_parse + 1)),
11721 "positive_lookbehind:")
11722 || memBEGINs(RExC_parse + 1,
11723 (Size_t) (RExC_end - (RExC_parse + 1)),
11724 "negative_lookahead:")
11725 || memBEGINs(RExC_parse + 1,
11726 (Size_t) (RExC_end - (RExC_parse + 1)),
11727 "negative_lookbehind:"))))
11728 ) { /* Lookahead or eval. */
11730 regnode_offset tail;
11732 ret = reg_node(pRExC_state, LOGICAL);
11733 FLAGS(REGNODE_p(ret)) = 1;
11735 tail = reg(pRExC_state, 1, &flag, depth+1);
11736 RETURN_FAIL_ON_RESTART(flag, flagp);
11737 if (! REGTAIL(pRExC_state, ret, tail)) {
11738 REQUIRE_BRANCHJ(flagp, 0);
11742 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11743 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11745 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11746 char *name_start= RExC_parse++;
11748 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11749 if ( RExC_parse == name_start
11750 || RExC_parse >= RExC_end
11751 || *RExC_parse != ch)
11753 vFAIL2("Sequence (?(%c... not terminated",
11754 (ch == '>' ? '<' : ch));
11758 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11759 RExC_rxi->data->data[num]=(void*)sv_dat;
11760 SvREFCNT_inc_simple_void_NN(sv_dat);
11762 ret = reganode(pRExC_state, GROUPPN, num);
11763 goto insert_if_check_paren;
11765 else if (memBEGINs(RExC_parse,
11766 (STRLEN) (RExC_end - RExC_parse),
11769 ret = reganode(pRExC_state, DEFINEP, 0);
11770 RExC_parse += DEFINE_len;
11772 goto insert_if_check_paren;
11774 else if (RExC_parse[0] == 'R') {
11776 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11777 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11778 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11781 if (RExC_parse[0] == '0') {
11785 else if (inRANGE(RExC_parse[0], '1', '9')) {
11788 if (grok_atoUV(RExC_parse, &uv, &endptr)
11791 parno = (I32)uv + 1;
11792 RExC_parse = (char*)endptr;
11794 /* else "Switch condition not recognized" below */
11795 } else if (RExC_parse[0] == '&') {
11798 sv_dat = reg_scan_name(pRExC_state,
11799 REG_RSN_RETURN_DATA);
11801 parno = 1 + *((I32 *)SvPVX(sv_dat));
11803 ret = reganode(pRExC_state, INSUBP, parno);
11804 goto insert_if_check_paren;
11806 else if (inRANGE(RExC_parse[0], '1', '9')) {
11811 if (grok_atoUV(RExC_parse, &uv, &endptr)
11815 RExC_parse = (char*)endptr;
11818 vFAIL("panic: grok_atoUV returned FALSE");
11820 ret = reganode(pRExC_state, GROUPP, parno);
11822 insert_if_check_paren:
11823 if (UCHARAT(RExC_parse) != ')') {
11825 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11827 vFAIL("Switch condition not recognized");
11829 nextchar(pRExC_state);
11831 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
11834 REQUIRE_BRANCHJ(flagp, 0);
11836 br = regbranch(pRExC_state, &flags, 1, depth+1);
11838 RETURN_FAIL_ON_RESTART(flags,flagp);
11839 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11842 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
11845 REQUIRE_BRANCHJ(flagp, 0);
11847 c = UCHARAT(RExC_parse);
11848 nextchar(pRExC_state);
11849 if (flags&HASWIDTH)
11850 *flagp |= HASWIDTH;
11853 vFAIL("(?(DEFINE)....) does not allow branches");
11855 /* Fake one for optimizer. */
11856 lastbr = reganode(pRExC_state, IFTHEN, 0);
11858 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
11859 RETURN_FAIL_ON_RESTART(flags, flagp);
11860 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11863 if (! REGTAIL(pRExC_state, ret, lastbr)) {
11864 REQUIRE_BRANCHJ(flagp, 0);
11866 if (flags&HASWIDTH)
11867 *flagp |= HASWIDTH;
11868 c = UCHARAT(RExC_parse);
11869 nextchar(pRExC_state);
11874 if (RExC_parse >= RExC_end)
11875 vFAIL("Switch (?(condition)... not terminated");
11877 vFAIL("Switch (?(condition)... contains too many branches");
11879 ender = reg_node(pRExC_state, TAIL);
11880 if (! REGTAIL(pRExC_state, br, ender)) {
11881 REQUIRE_BRANCHJ(flagp, 0);
11884 if (! REGTAIL(pRExC_state, lastbr, ender)) {
11885 REQUIRE_BRANCHJ(flagp, 0);
11887 if (! REGTAIL(pRExC_state,
11890 NEXTOPER(REGNODE_p(lastbr)))),
11893 REQUIRE_BRANCHJ(flagp, 0);
11897 if (! REGTAIL(pRExC_state, ret, ender)) {
11898 REQUIRE_BRANCHJ(flagp, 0);
11900 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
11901 RExC_size++; /* XXX WHY do we need this?!!
11902 For large programs it seems to be required
11903 but I can't figure out why. -- dmq*/
11908 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11910 vFAIL("Unknown switch condition (?(...))");
11912 case '[': /* (?[ ... ]) */
11913 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
11915 case 0: /* A NUL */
11916 RExC_parse--; /* for vFAIL to print correctly */
11917 vFAIL("Sequence (? incomplete");
11921 if (RExC_strict) { /* [perl #132851] */
11922 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
11925 default: /* e.g., (?i) */
11926 RExC_parse = (char *) seqstart + 1;
11928 parse_lparen_question_flags(pRExC_state);
11929 if (UCHARAT(RExC_parse) != ':') {
11930 if (RExC_parse < RExC_end)
11931 nextchar(pRExC_state);
11936 nextchar(pRExC_state);
11942 if (*RExC_parse == '{') {
11943 ckWARNregdep(RExC_parse + 1,
11944 "Unescaped left brace in regex is "
11945 "deprecated here (and will be fatal "
11946 "in Perl 5.32), passed through");
11948 /* Not bothering to indent here, as the above 'else' is temporary
11950 if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11954 if (! ALL_PARENS_COUNTED) {
11955 /* If we are in our first pass through (and maybe only pass),
11956 * we need to allocate memory for the capturing parentheses
11960 if (!RExC_parens_buf_size) {
11961 /* first guess at number of parens we might encounter */
11962 RExC_parens_buf_size = 10;
11964 /* setup RExC_open_parens, which holds the address of each
11965 * OPEN tag, and to make things simpler for the 0 index the
11966 * start of the program - this is used later for offsets */
11967 Newxz(RExC_open_parens, RExC_parens_buf_size,
11969 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
11971 /* setup RExC_close_parens, which holds the address of each
11972 * CLOSE tag, and to make things simpler for the 0 index
11973 * the end of the program - this is used later for offsets
11975 Newxz(RExC_close_parens, RExC_parens_buf_size,
11977 /* we dont know where end op starts yet, so we dont need to
11978 * set RExC_close_parens[0] like we do RExC_open_parens[0]
11981 else if (RExC_npar > RExC_parens_buf_size) {
11982 I32 old_size = RExC_parens_buf_size;
11984 RExC_parens_buf_size *= 2;
11986 Renew(RExC_open_parens, RExC_parens_buf_size,
11988 Zero(RExC_open_parens + old_size,
11989 RExC_parens_buf_size - old_size, regnode_offset);
11991 Renew(RExC_close_parens, RExC_parens_buf_size,
11993 Zero(RExC_close_parens + old_size,
11994 RExC_parens_buf_size - old_size, regnode_offset);
11998 ret = reganode(pRExC_state, OPEN, parno);
11999 if (!RExC_nestroot)
12000 RExC_nestroot = parno;
12001 if (RExC_open_parens && !RExC_open_parens[parno])
12003 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12004 "%*s%*s Setting open paren #%" IVdf " to %d\n",
12005 22, "| |", (int)(depth * 2 + 1), "",
12007 RExC_open_parens[parno]= ret;
12010 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12011 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12014 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12024 /* Pick up the branches, linking them together. */
12025 parse_start = RExC_parse; /* MJD */
12026 br = regbranch(pRExC_state, &flags, 1, depth+1);
12028 /* branch_len = (paren != 0); */
12031 RETURN_FAIL_ON_RESTART(flags, flagp);
12032 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12034 if (*RExC_parse == '|') {
12035 if (RExC_use_BRANCHJ) {
12036 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12039 reginsert(pRExC_state, BRANCH, br, depth+1);
12040 Set_Node_Length(REGNODE_p(br), paren != 0);
12041 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12045 else if (paren == ':') {
12046 *flagp |= flags&SIMPLE;
12048 if (is_open) { /* Starts with OPEN. */
12049 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12050 REQUIRE_BRANCHJ(flagp, 0);
12053 else if (paren != '?') /* Not Conditional */
12055 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12057 while (*RExC_parse == '|') {
12058 if (RExC_use_BRANCHJ) {
12061 ender = reganode(pRExC_state, LONGJMP, 0);
12063 /* Append to the previous. */
12064 shut_gcc_up = REGTAIL(pRExC_state,
12065 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12067 PERL_UNUSED_VAR(shut_gcc_up);
12069 nextchar(pRExC_state);
12070 if (freeze_paren) {
12071 if (RExC_npar > after_freeze)
12072 after_freeze = RExC_npar;
12073 RExC_npar = freeze_paren;
12075 br = regbranch(pRExC_state, &flags, 0, depth+1);
12078 RETURN_FAIL_ON_RESTART(flags, flagp);
12079 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12081 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12082 REQUIRE_BRANCHJ(flagp, 0);
12085 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12088 if (have_branch || paren != ':') {
12091 /* Make a closing node, and hook it on the end. */
12094 ender = reg_node(pRExC_state, TAIL);
12097 ender = reganode(pRExC_state, CLOSE, parno);
12098 if ( RExC_close_parens ) {
12099 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12100 "%*s%*s Setting close paren #%" IVdf " to %d\n",
12101 22, "| |", (int)(depth * 2 + 1), "",
12102 (IV)parno, ender));
12103 RExC_close_parens[parno]= ender;
12104 if (RExC_nestroot == parno)
12107 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12108 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12111 ender = reg_node(pRExC_state, SRCLOSE);
12112 RExC_in_script_run = 0;
12122 *flagp &= ~HASWIDTH;
12124 case 't': /* aTomic */
12126 ender = reg_node(pRExC_state, SUCCEED);
12129 ender = reg_node(pRExC_state, END);
12130 assert(!RExC_end_op); /* there can only be one! */
12131 RExC_end_op = REGNODE_p(ender);
12132 if (RExC_close_parens) {
12133 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12134 "%*s%*s Setting close paren #0 (END) to %d\n",
12135 22, "| |", (int)(depth * 2 + 1), "",
12138 RExC_close_parens[0]= ender;
12143 DEBUG_PARSE_MSG("lsbr");
12144 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12145 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12146 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12147 SvPV_nolen_const(RExC_mysv1),
12149 SvPV_nolen_const(RExC_mysv2),
12151 (IV)(ender - lastbr)
12154 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12155 REQUIRE_BRANCHJ(flagp, 0);
12159 char is_nothing= 1;
12161 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12163 /* Hook the tails of the branches to the closing node. */
12164 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12165 const U8 op = PL_regkind[OP(br)];
12166 if (op == BRANCH) {
12167 if (! REGTAIL_STUDY(pRExC_state,
12168 REGNODE_OFFSET(NEXTOPER(br)),
12171 REQUIRE_BRANCHJ(flagp, 0);
12173 if ( OP(NEXTOPER(br)) != NOTHING
12174 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12177 else if (op == BRANCHJ) {
12178 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12179 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12181 PERL_UNUSED_VAR(shut_gcc_up);
12182 /* for now we always disable this optimisation * /
12183 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12184 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12190 regnode * ret_as_regnode = REGNODE_p(ret);
12191 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12192 ? regnext(ret_as_regnode)
12195 DEBUG_PARSE_MSG("NADA");
12196 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12197 NULL, pRExC_state);
12198 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12199 NULL, pRExC_state);
12200 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12201 SvPV_nolen_const(RExC_mysv1),
12202 (IV)REG_NODE_NUM(ret_as_regnode),
12203 SvPV_nolen_const(RExC_mysv2),
12209 if (OP(REGNODE_p(ender)) == TAIL) {
12211 RExC_emit= REGNODE_OFFSET(br) + 1;
12214 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12215 OP(opt)= OPTIMIZED;
12216 NEXT_OFF(br)= REGNODE_p(ender) - br;
12224 /* Even/odd or x=don't care: 010101x10x */
12225 static const char parens[] = "=!aA<,>Bbt";
12226 /* flag below is set to 0 up through 'A'; 1 for larger */
12228 if (paren && (p = strchr(parens, paren))) {
12229 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12230 int flag = (p - parens) > 3;
12232 if (paren == '>' || paren == 't') {
12233 node = SUSPEND, flag = 0;
12236 reginsert(pRExC_state, node, ret, depth+1);
12237 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12238 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12239 FLAGS(REGNODE_p(ret)) = flag;
12240 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12242 REQUIRE_BRANCHJ(flagp, 0);
12247 /* Check for proper termination. */
12249 /* restore original flags, but keep (?p) and, if we've encountered
12250 * something in the parse that changes /d rules into /u, keep the /u */
12251 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12252 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
12253 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12255 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12256 RExC_parse = oregcomp_parse;
12257 vFAIL("Unmatched (");
12259 nextchar(pRExC_state);
12261 else if (!paren && RExC_parse < RExC_end) {
12262 if (*RExC_parse == ')') {
12264 vFAIL("Unmatched )");
12267 FAIL("Junk on end of regexp"); /* "Can't happen". */
12268 NOT_REACHED; /* NOTREACHED */
12271 if (RExC_in_lookbehind) {
12272 RExC_in_lookbehind--;
12274 if (RExC_in_lookahead) {
12275 RExC_in_lookahead--;
12277 if (after_freeze > RExC_npar)
12278 RExC_npar = after_freeze;
12283 - regbranch - one alternative of an | operator
12285 * Implements the concatenation operator.
12287 * On success, returns the offset at which any next node should be placed into
12288 * the regex engine program being compiled.
12290 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12291 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12294 STATIC regnode_offset
12295 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12297 regnode_offset ret;
12298 regnode_offset chain = 0;
12299 regnode_offset latest;
12300 I32 flags = 0, c = 0;
12301 GET_RE_DEBUG_FLAGS_DECL;
12303 PERL_ARGS_ASSERT_REGBRANCH;
12305 DEBUG_PARSE("brnc");
12310 if (RExC_use_BRANCHJ)
12311 ret = reganode(pRExC_state, BRANCHJ, 0);
12313 ret = reg_node(pRExC_state, BRANCH);
12314 Set_Node_Length(REGNODE_p(ret), 1);
12318 *flagp = WORST; /* Tentatively. */
12320 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12321 FALSE /* Don't force to /x */ );
12322 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12323 flags &= ~TRYAGAIN;
12324 latest = regpiece(pRExC_state, &flags, depth+1);
12326 if (flags & TRYAGAIN)
12328 RETURN_FAIL_ON_RESTART(flags, flagp);
12329 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12333 *flagp |= flags&(HASWIDTH|POSTPONED);
12334 if (chain == 0) /* First piece. */
12335 *flagp |= flags&SPSTART;
12337 /* FIXME adding one for every branch after the first is probably
12338 * excessive now we have TRIE support. (hv) */
12340 if (! REGTAIL(pRExC_state, chain, latest)) {
12341 /* XXX We could just redo this branch, but figuring out what
12342 * bookkeeping needs to be reset is a pain, and it's likely
12343 * that other branches that goto END will also be too large */
12344 REQUIRE_BRANCHJ(flagp, 0);
12350 if (chain == 0) { /* Loop ran zero times. */
12351 chain = reg_node(pRExC_state, NOTHING);
12356 *flagp |= flags&SIMPLE;
12363 - regpiece - something followed by possible quantifier * + ? {n,m}
12365 * Note that the branching code sequences used for ? and the general cases
12366 * of * and + are somewhat optimized: they use the same NOTHING node as
12367 * both the endmarker for their branch list and the body of the last branch.
12368 * It might seem that this node could be dispensed with entirely, but the
12369 * endmarker role is not redundant.
12371 * On success, returns the offset at which any next node should be placed into
12372 * the regex engine program being compiled.
12374 * Returns 0 otherwise, with *flagp set to indicate why:
12375 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12376 * RESTART_PARSE if the parse needs to be restarted, or'd with
12377 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12379 STATIC regnode_offset
12380 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12382 regnode_offset ret;
12386 const char * const origparse = RExC_parse;
12388 I32 max = REG_INFTY;
12389 #ifdef RE_TRACK_PATTERN_OFFSETS
12392 const char *maxpos = NULL;
12395 /* Save the original in case we change the emitted regop to a FAIL. */
12396 const regnode_offset orig_emit = RExC_emit;
12398 GET_RE_DEBUG_FLAGS_DECL;
12400 PERL_ARGS_ASSERT_REGPIECE;
12402 DEBUG_PARSE("piec");
12404 ret = regatom(pRExC_state, &flags, depth+1);
12406 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12407 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12412 if (op == '{' && regcurly(RExC_parse)) {
12414 #ifdef RE_TRACK_PATTERN_OFFSETS
12415 parse_start = RExC_parse; /* MJD */
12417 next = RExC_parse + 1;
12418 while (isDIGIT(*next) || *next == ',') {
12419 if (*next == ',') {
12427 if (*next == '}') { /* got one */
12428 const char* endptr;
12432 if (isDIGIT(*RExC_parse)) {
12434 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12435 vFAIL("Invalid quantifier in {,}");
12436 if (uv >= REG_INFTY)
12437 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12442 if (*maxpos == ',')
12445 maxpos = RExC_parse;
12446 if (isDIGIT(*maxpos)) {
12448 if (!grok_atoUV(maxpos, &uv, &endptr))
12449 vFAIL("Invalid quantifier in {,}");
12450 if (uv >= REG_INFTY)
12451 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12454 max = REG_INFTY; /* meaning "infinity" */
12457 nextchar(pRExC_state);
12458 if (max < min) { /* If can't match, warn and optimize to fail
12460 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12461 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12462 NEXT_OFF(REGNODE_p(orig_emit)) =
12463 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12466 else if (min == max && *RExC_parse == '?')
12468 ckWARN2reg(RExC_parse + 1,
12469 "Useless use of greediness modifier '%c'",
12474 if ((flags&SIMPLE)) {
12475 if (min == 0 && max == REG_INFTY) {
12476 reginsert(pRExC_state, STAR, ret, depth+1);
12478 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12481 if (min == 1 && max == REG_INFTY) {
12482 reginsert(pRExC_state, PLUS, ret, depth+1);
12484 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12487 MARK_NAUGHTY_EXP(2, 2);
12488 reginsert(pRExC_state, CURLY, ret, depth+1);
12489 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12490 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12493 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12495 FLAGS(REGNODE_p(w)) = 0;
12496 if (! REGTAIL(pRExC_state, ret, w)) {
12497 REQUIRE_BRANCHJ(flagp, 0);
12499 if (RExC_use_BRANCHJ) {
12500 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12501 reginsert(pRExC_state, NOTHING, ret, depth+1);
12502 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12504 reginsert(pRExC_state, CURLYX, ret, depth+1);
12506 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12507 Set_Node_Length(REGNODE_p(ret),
12508 op == '{' ? (RExC_parse - parse_start) : 1);
12510 if (RExC_use_BRANCHJ)
12511 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12513 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12516 REQUIRE_BRANCHJ(flagp, 0);
12518 RExC_whilem_seen++;
12519 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12521 FLAGS(REGNODE_p(ret)) = 0;
12526 *flagp |= HASWIDTH;
12527 ARG1_SET(REGNODE_p(ret), (U16)min);
12528 ARG2_SET(REGNODE_p(ret), (U16)max);
12529 if (max == REG_INFTY)
12530 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12536 if (!ISMULT1(op)) {
12541 #if 0 /* Now runtime fix should be reliable. */
12543 /* if this is reinstated, don't forget to put this back into perldiag:
12545 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12547 (F) The part of the regexp subject to either the * or + quantifier
12548 could match an empty string. The {#} shows in the regular
12549 expression about where the problem was discovered.
12553 if (!(flags&HASWIDTH) && op != '?')
12554 vFAIL("Regexp *+ operand could be empty");
12557 #ifdef RE_TRACK_PATTERN_OFFSETS
12558 parse_start = RExC_parse;
12560 nextchar(pRExC_state);
12562 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12568 else if (op == '+') {
12572 else if (op == '?') {
12577 if (!(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12578 ckWARN2reg(RExC_parse,
12579 "%" UTF8f " matches null string many times",
12580 UTF8fARG(UTF, (RExC_parse >= origparse
12581 ? RExC_parse - origparse
12586 if (*RExC_parse == '?') {
12587 nextchar(pRExC_state);
12588 reginsert(pRExC_state, MINMOD, ret, depth+1);
12589 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12590 REQUIRE_BRANCHJ(flagp, 0);
12593 else if (*RExC_parse == '+') {
12594 regnode_offset ender;
12595 nextchar(pRExC_state);
12596 ender = reg_node(pRExC_state, SUCCEED);
12597 if (! REGTAIL(pRExC_state, ret, ender)) {
12598 REQUIRE_BRANCHJ(flagp, 0);
12600 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12601 ender = reg_node(pRExC_state, TAIL);
12602 if (! REGTAIL(pRExC_state, ret, ender)) {
12603 REQUIRE_BRANCHJ(flagp, 0);
12607 if (ISMULT2(RExC_parse)) {
12609 vFAIL("Nested quantifiers");
12616 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12617 regnode_offset * node_p,
12625 /* This routine teases apart the various meanings of \N and returns
12626 * accordingly. The input parameters constrain which meaning(s) is/are valid
12627 * in the current context.
12629 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12631 * If <code_point_p> is not NULL, the context is expecting the result to be a
12632 * single code point. If this \N instance turns out to a single code point,
12633 * the function returns TRUE and sets *code_point_p to that code point.
12635 * If <node_p> is not NULL, the context is expecting the result to be one of
12636 * the things representable by a regnode. If this \N instance turns out to be
12637 * one such, the function generates the regnode, returns TRUE and sets *node_p
12638 * to point to the offset of that regnode into the regex engine program being
12641 * If this instance of \N isn't legal in any context, this function will
12642 * generate a fatal error and not return.
12644 * On input, RExC_parse should point to the first char following the \N at the
12645 * time of the call. On successful return, RExC_parse will have been updated
12646 * to point to just after the sequence identified by this routine. Also
12647 * *flagp has been updated as needed.
12649 * When there is some problem with the current context and this \N instance,
12650 * the function returns FALSE, without advancing RExC_parse, nor setting
12651 * *node_p, nor *code_point_p, nor *flagp.
12653 * If <cp_count> is not NULL, the caller wants to know the length (in code
12654 * points) that this \N sequence matches. This is set, and the input is
12655 * parsed for errors, even if the function returns FALSE, as detailed below.
12657 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12659 * Probably the most common case is for the \N to specify a single code point.
12660 * *cp_count will be set to 1, and *code_point_p will be set to that code
12663 * Another possibility is for the input to be an empty \N{}. This is no
12664 * longer accepted, and will generate a fatal error.
12666 * Another possibility is for a custom charnames handler to be in effect which
12667 * translates the input name to an empty string. *cp_count will be set to 0.
12668 * *node_p will be set to a generated NOTHING node.
12670 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12671 * set to 0. *node_p will be set to a generated REG_ANY node.
12673 * The fifth possibility is that \N resolves to a sequence of more than one
12674 * code points. *cp_count will be set to the number of code points in the
12675 * sequence. *node_p will be set to a generated node returned by this
12676 * function calling S_reg().
12678 * The final possibility is that it is premature to be calling this function;
12679 * the parse needs to be restarted. This can happen when this changes from
12680 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12681 * latter occurs only when the fifth possibility would otherwise be in
12682 * effect, and is because one of those code points requires the pattern to be
12683 * recompiled as UTF-8. The function returns FALSE, and sets the
12684 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12685 * happens, the caller needs to desist from continuing parsing, and return
12686 * this information to its caller. This is not set for when there is only one
12687 * code point, as this can be called as part of an ANYOF node, and they can
12688 * store above-Latin1 code points without the pattern having to be in UTF-8.
12690 * For non-single-quoted regexes, the tokenizer has resolved character and
12691 * sequence names inside \N{...} into their Unicode values, normalizing the
12692 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12693 * hex-represented code points in the sequence. This is done there because
12694 * the names can vary based on what charnames pragma is in scope at the time,
12695 * so we need a way to take a snapshot of what they resolve to at the time of
12696 * the original parse. [perl #56444].
12698 * That parsing is skipped for single-quoted regexes, so here we may get
12699 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12700 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12701 * the native character set for non-ASCII platforms. The other possibilities
12702 * are already native, so no translation is done. */
12704 char * endbrace; /* points to '}' following the name */
12705 char* p = RExC_parse; /* Temporary */
12707 SV * substitute_parse = NULL;
12712 GET_RE_DEBUG_FLAGS_DECL;
12714 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12716 GET_RE_DEBUG_FLAGS;
12718 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12719 assert(! (node_p && cp_count)); /* At most 1 should be set */
12721 if (cp_count) { /* Initialize return for the most common case */
12725 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12726 * modifier. The other meanings do not, so use a temporary until we find
12727 * out which we are being called with */
12728 skip_to_be_ignored_text(pRExC_state, &p,
12729 FALSE /* Don't force to /x */ );
12731 /* Disambiguate between \N meaning a named character versus \N meaning
12732 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12733 * quantifier, or if there is no '{' at all */
12734 if (*p != '{' || regcurly(p)) {
12744 *node_p = reg_node(pRExC_state, REG_ANY);
12745 *flagp |= HASWIDTH|SIMPLE;
12747 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
12751 /* The test above made sure that the next real character is a '{', but
12752 * under the /x modifier, it could be separated by space (or a comment and
12753 * \n) and this is not allowed (for consistency with \x{...} and the
12754 * tokenizer handling of \N{NAME}). */
12755 if (*RExC_parse != '{') {
12756 vFAIL("Missing braces on \\N{}");
12759 RExC_parse++; /* Skip past the '{' */
12761 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12762 if (! endbrace) { /* no trailing brace */
12763 vFAIL2("Missing right brace on \\%c{}", 'N');
12766 /* Here, we have decided it should be a named character or sequence. These
12767 * imply Unicode semantics */
12768 REQUIRE_UNI_RULES(flagp, FALSE);
12770 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
12771 * nothing at all (not allowed under strict) */
12772 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
12773 RExC_parse = endbrace;
12775 RExC_parse++; /* Position after the "}" */
12776 vFAIL("Zero length \\N{}");
12782 nextchar(pRExC_state);
12787 *node_p = reg_node(pRExC_state, NOTHING);
12791 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
12793 /* Here, the name isn't of the form U+.... This can happen if the
12794 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
12795 * is the time to find out what the name means */
12797 const STRLEN name_len = endbrace - RExC_parse;
12798 SV * value_sv; /* What does this name evaluate to */
12800 const U8 * value; /* string of name's value */
12801 STRLEN value_len; /* and its length */
12803 /* RExC_unlexed_names is a hash of names that weren't evaluated by
12804 * toke.c, and their values. Make sure is initialized */
12805 if (! RExC_unlexed_names) {
12806 RExC_unlexed_names = newHV();
12809 /* If we have already seen this name in this pattern, use that. This
12810 * allows us to only call the charnames handler once per name per
12811 * pattern. A broken or malicious handler could return something
12812 * different each time, which could cause the results to vary depending
12813 * on if something gets added or subtracted from the pattern that
12814 * causes the number of passes to change, for example */
12815 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
12818 value_sv = *value_svp;
12820 else { /* Otherwise we have to go out and get the name */
12821 const char * error_msg = NULL;
12822 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
12826 RExC_parse = endbrace;
12830 /* If no error message, should have gotten a valid return */
12833 /* Save the name's meaning for later use */
12834 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
12837 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
12841 /* Here, we have the value the name evaluates to in 'value_sv' */
12842 value = (U8 *) SvPV(value_sv, value_len);
12844 /* See if the result is one code point vs 0 or multiple */
12845 if (value_len > 0 && value_len <= (UV) ((SvUTF8(value_sv))
12849 /* Here, exactly one code point. If that isn't what is wanted,
12851 if (! code_point_p) {
12856 /* Convert from string to numeric code point */
12857 *code_point_p = (SvUTF8(value_sv))
12858 ? valid_utf8_to_uvchr(value, NULL)
12861 /* Have parsed this entire single code point \N{...}. *cp_count
12862 * has already been set to 1, so don't do it again. */
12863 RExC_parse = endbrace;
12864 nextchar(pRExC_state);
12866 } /* End of is a single code point */
12868 /* Count the code points, if caller desires. The API says to do this
12869 * even if we will later return FALSE */
12873 *cp_count = (SvUTF8(value_sv))
12874 ? utf8_length(value, value + value_len)
12878 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12879 * But don't back the pointer up if the caller wants to know how many
12880 * code points there are (they need to handle it themselves in this
12889 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
12890 * reg recursively to parse it. That way, it retains its atomicness,
12891 * while not having to worry about any special handling that some code
12892 * points may have. */
12894 substitute_parse = newSVpvs("?:");
12895 sv_catsv(substitute_parse, value_sv);
12896 sv_catpv(substitute_parse, ")");
12898 /* The value should already be native, so no need to convert on EBCDIC
12900 assert(! RExC_recode_x_to_native);
12903 else { /* \N{U+...} */
12904 Size_t count = 0; /* code point count kept internally */
12906 /* We can get to here when the input is \N{U+...} or when toke.c has
12907 * converted a name to the \N{U+...} form. This include changing a
12908 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
12910 RExC_parse += 2; /* Skip past the 'U+' */
12912 /* Code points are separated by dots. The '}' terminates the whole
12915 do { /* Loop until the ending brace */
12917 char * start_digit; /* The first of the current code point */
12918 if (! isXDIGIT(*RExC_parse)) {
12920 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12923 start_digit = RExC_parse;
12926 /* Loop through the hex digits of the current code point */
12928 /* Adding this digit will shift the result 4 bits. If that
12929 * result would be above the legal max, it's overflow */
12930 if (cp > MAX_LEGAL_CP >> 4) {
12932 /* Find the end of the code point */
12935 } while (isXDIGIT(*RExC_parse) || *RExC_parse == '_');
12937 /* Be sure to synchronize this message with the similar one
12939 vFAIL4("Use of code point 0x%.*s is not allowed; the"
12940 " permissible max is 0x%" UVxf,
12941 (int) (RExC_parse - start_digit), start_digit,
12945 /* Accumulate this (valid) digit into the running total */
12946 cp = (cp << 4) + READ_XDIGIT(RExC_parse);
12948 /* READ_XDIGIT advanced the input pointer. Ignore a single
12949 * underscore separator */
12950 if (*RExC_parse == '_' && isXDIGIT(RExC_parse[1])) {
12953 } while (isXDIGIT(*RExC_parse));
12955 /* Here, have accumulated the next code point */
12956 if (RExC_parse >= endbrace) { /* If done ... */
12961 /* Here, is a single code point; fail if doesn't want that */
12962 if (! code_point_p) {
12967 /* A single code point is easy to handle; just return it */
12968 *code_point_p = UNI_TO_NATIVE(cp);
12969 RExC_parse = endbrace;
12970 nextchar(pRExC_state);
12974 /* Here, the only legal thing would be a multiple character
12975 * sequence (of the form "\N{U+c1.c2. ... }". So the next
12976 * character must be a dot (and the one after that can't be the
12977 * endbrace, or we'd have something like \N{U+100.} ) */
12978 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
12979 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12980 ? UTF8SKIP(RExC_parse)
12982 if (RExC_parse >= endbrace) { /* Guard against malformed utf8 */
12983 RExC_parse = endbrace;
12985 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12988 /* Here, looks like its really a multiple character sequence. Fail
12989 * if that's not what the caller wants. But continue with counting
12990 * and error checking if they still want a count */
12991 if (! node_p && ! cp_count) {
12995 /* What is done here is to convert this to a sub-pattern of the
12996 * form \x{char1}\x{char2}... and then call reg recursively to
12997 * parse it (enclosing in "(?: ... )" ). That way, it retains its
12998 * atomicness, while not having to worry about special handling
12999 * that some code points may have. We don't create a subpattern,
13000 * but go through the motions of code point counting and error
13001 * checking, if the caller doesn't want a node returned. */
13003 if (node_p && count == 1) {
13004 substitute_parse = newSVpvs("?:");
13010 /* Convert to notation the rest of the code understands */
13011 sv_catpvs(substitute_parse, "\\x{");
13012 sv_catpvn(substitute_parse, start_digit,
13013 RExC_parse - start_digit);
13014 sv_catpvs(substitute_parse, "}");
13017 /* Move to after the dot (or ending brace the final time through.)
13022 } while (RExC_parse < endbrace);
13024 if (! node_p) { /* Doesn't want the node */
13031 sv_catpvs(substitute_parse, ")");
13033 /* The values are Unicode, and therefore have to be converted to native
13034 * on a non-Unicode (meaning non-ASCII) platform. */
13035 SET_recode_x_to_native(1);
13038 /* Here, we have the string the name evaluates to, ready to be parsed,
13039 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13040 * constructs. This can be called from within a substitute parse already.
13041 * The error reporting mechanism doesn't work for 2 levels of this, but the
13042 * code above has validated this new construct, so there should be no
13043 * errors generated by the below. And this isn' an exact copy, so the
13044 * mechanism to seamlessly deal with this won't work, so turn off warnings
13046 save_start = RExC_start;
13047 orig_end = RExC_end;
13049 RExC_parse = RExC_start = SvPVX(substitute_parse);
13050 RExC_end = RExC_parse + SvCUR(substitute_parse);
13051 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13053 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13055 /* Restore the saved values */
13057 RExC_start = save_start;
13058 RExC_parse = endbrace;
13059 RExC_end = orig_end;
13060 SET_recode_x_to_native(0);
13062 SvREFCNT_dec_NN(substitute_parse);
13065 RETURN_FAIL_ON_RESTART(flags, flagp);
13066 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13069 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13071 nextchar(pRExC_state);
13077 PERL_STATIC_INLINE U8
13078 S_compute_EXACTish(RExC_state_t *pRExC_state)
13082 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13090 op = get_regex_charset(RExC_flags);
13091 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13092 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13093 been, so there is no hole */
13096 return op + EXACTF;
13100 S_new_regcurly(const char *s, const char *e)
13102 /* This is a temporary function designed to match the most lenient form of
13103 * a {m,n} quantifier we ever envision, with either number omitted, and
13104 * spaces anywhere between/before/after them.
13106 * If this function fails, then the string it matches is very unlikely to
13107 * ever be considered a valid quantifier, so we can allow the '{' that
13108 * begins it to be considered as a literal */
13110 bool has_min = FALSE;
13111 bool has_max = FALSE;
13113 PERL_ARGS_ASSERT_NEW_REGCURLY;
13115 if (s >= e || *s++ != '{')
13118 while (s < e && isSPACE(*s)) {
13121 while (s < e && isDIGIT(*s)) {
13125 while (s < e && isSPACE(*s)) {
13131 while (s < e && isSPACE(*s)) {
13134 while (s < e && isDIGIT(*s)) {
13138 while (s < e && isSPACE(*s)) {
13143 return s < e && *s == '}' && (has_min || has_max);
13146 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13147 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13150 S_backref_value(char *p, char *e)
13152 const char* endptr = e;
13154 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13161 - regatom - the lowest level
13163 Try to identify anything special at the start of the current parse position.
13164 If there is, then handle it as required. This may involve generating a
13165 single regop, such as for an assertion; or it may involve recursing, such as
13166 to handle a () structure.
13168 If the string doesn't start with something special then we gobble up
13169 as much literal text as we can. If we encounter a quantifier, we have to
13170 back off the final literal character, as that quantifier applies to just it
13171 and not to the whole string of literals.
13173 Once we have been able to handle whatever type of thing started the
13174 sequence, we return the offset into the regex engine program being compiled
13175 at which any next regnode should be placed.
13177 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13178 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13179 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13180 Otherwise does not return 0.
13182 Note: we have to be careful with escapes, as they can be both literal
13183 and special, and in the case of \10 and friends, context determines which.
13185 A summary of the code structure is:
13187 switch (first_byte) {
13188 cases for each special:
13189 handle this special;
13192 switch (2nd byte) {
13193 cases for each unambiguous special:
13194 handle this special;
13196 cases for each ambigous special/literal:
13198 if (special) handle here
13200 default: // unambiguously literal:
13203 default: // is a literal char
13206 create EXACTish node for literal;
13207 while (more input and node isn't full) {
13208 switch (input_byte) {
13209 cases for each special;
13210 make sure parse pointer is set so that the next call to
13211 regatom will see this special first
13212 goto loopdone; // EXACTish node terminated by prev. char
13214 append char to EXACTISH node;
13216 get next input byte;
13220 return the generated node;
13222 Specifically there are two separate switches for handling
13223 escape sequences, with the one for handling literal escapes requiring
13224 a dummy entry for all of the special escapes that are actually handled
13229 STATIC regnode_offset
13230 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13233 regnode_offset ret = 0;
13239 GET_RE_DEBUG_FLAGS_DECL;
13241 *flagp = WORST; /* Tentatively. */
13243 DEBUG_PARSE("atom");
13245 PERL_ARGS_ASSERT_REGATOM;
13248 parse_start = RExC_parse;
13249 assert(RExC_parse < RExC_end);
13250 switch ((U8)*RExC_parse) {
13252 RExC_seen_zerolen++;
13253 nextchar(pRExC_state);
13254 if (RExC_flags & RXf_PMf_MULTILINE)
13255 ret = reg_node(pRExC_state, MBOL);
13257 ret = reg_node(pRExC_state, SBOL);
13258 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13261 nextchar(pRExC_state);
13263 RExC_seen_zerolen++;
13264 if (RExC_flags & RXf_PMf_MULTILINE)
13265 ret = reg_node(pRExC_state, MEOL);
13267 ret = reg_node(pRExC_state, SEOL);
13268 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13271 nextchar(pRExC_state);
13272 if (RExC_flags & RXf_PMf_SINGLELINE)
13273 ret = reg_node(pRExC_state, SANY);
13275 ret = reg_node(pRExC_state, REG_ANY);
13276 *flagp |= HASWIDTH|SIMPLE;
13278 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13282 char * const oregcomp_parse = ++RExC_parse;
13283 ret = regclass(pRExC_state, flagp, depth+1,
13284 FALSE, /* means parse the whole char class */
13285 TRUE, /* allow multi-char folds */
13286 FALSE, /* don't silence non-portable warnings. */
13287 (bool) RExC_strict,
13288 TRUE, /* Allow an optimized regnode result */
13291 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13292 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13295 if (*RExC_parse != ']') {
13296 RExC_parse = oregcomp_parse;
13297 vFAIL("Unmatched [");
13299 nextchar(pRExC_state);
13300 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13304 nextchar(pRExC_state);
13305 ret = reg(pRExC_state, 2, &flags, depth+1);
13307 if (flags & TRYAGAIN) {
13308 if (RExC_parse >= RExC_end) {
13309 /* Make parent create an empty node if needed. */
13310 *flagp |= TRYAGAIN;
13315 RETURN_FAIL_ON_RESTART(flags, flagp);
13316 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13319 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13323 if (flags & TRYAGAIN) {
13324 *flagp |= TRYAGAIN;
13327 vFAIL("Internal urp");
13328 /* Supposed to be caught earlier. */
13334 vFAIL("Quantifier follows nothing");
13339 This switch handles escape sequences that resolve to some kind
13340 of special regop and not to literal text. Escape sequences that
13341 resolve to literal text are handled below in the switch marked
13344 Every entry in this switch *must* have a corresponding entry
13345 in the literal escape switch. However, the opposite is not
13346 required, as the default for this switch is to jump to the
13347 literal text handling code.
13350 switch ((U8)*RExC_parse) {
13351 /* Special Escapes */
13353 RExC_seen_zerolen++;
13354 ret = reg_node(pRExC_state, SBOL);
13355 /* SBOL is shared with /^/ so we set the flags so we can tell
13356 * /\A/ from /^/ in split. */
13357 FLAGS(REGNODE_p(ret)) = 1;
13359 goto finish_meta_pat;
13361 ret = reg_node(pRExC_state, GPOS);
13362 RExC_seen |= REG_GPOS_SEEN;
13364 goto finish_meta_pat;
13366 if (!RExC_in_lookbehind && !RExC_in_lookahead) {
13367 RExC_seen_zerolen++;
13368 ret = reg_node(pRExC_state, KEEPS);
13370 /* XXX:dmq : disabling in-place substitution seems to
13371 * be necessary here to avoid cases of memory corruption, as
13372 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13374 RExC_seen |= REG_LOOKBEHIND_SEEN;
13375 goto finish_meta_pat;
13378 ++RExC_parse; /* advance past the 'K' */
13379 vFAIL("\\K not permitted in lookahead/lookbehind");
13382 ret = reg_node(pRExC_state, SEOL);
13384 RExC_seen_zerolen++; /* Do not optimize RE away */
13385 goto finish_meta_pat;
13387 ret = reg_node(pRExC_state, EOS);
13389 RExC_seen_zerolen++; /* Do not optimize RE away */
13390 goto finish_meta_pat;
13392 vFAIL("\\C no longer supported");
13394 ret = reg_node(pRExC_state, CLUMP);
13395 *flagp |= HASWIDTH;
13396 goto finish_meta_pat;
13404 regex_charset charset = get_regex_charset(RExC_flags);
13406 RExC_seen_zerolen++;
13407 RExC_seen |= REG_LOOKBEHIND_SEEN;
13408 op = BOUND + charset;
13410 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13411 flags = TRADITIONAL_BOUND;
13412 if (op > BOUNDA) { /* /aa is same as /a */
13418 char name = *RExC_parse;
13419 char * endbrace = NULL;
13421 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13424 vFAIL2("Missing right brace on \\%c{}", name);
13426 /* XXX Need to decide whether to take spaces or not. Should be
13427 * consistent with \p{}, but that currently is SPACE, which
13428 * means vertical too, which seems wrong
13429 * while (isBLANK(*RExC_parse)) {
13432 if (endbrace == RExC_parse) {
13433 RExC_parse++; /* After the '}' */
13434 vFAIL2("Empty \\%c{}", name);
13436 length = endbrace - RExC_parse;
13437 /*while (isBLANK(*(RExC_parse + length - 1))) {
13440 switch (*RExC_parse) {
13443 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13445 goto bad_bound_type;
13450 if (length != 2 || *(RExC_parse + 1) != 'b') {
13451 goto bad_bound_type;
13456 if (length != 2 || *(RExC_parse + 1) != 'b') {
13457 goto bad_bound_type;
13462 if (length != 2 || *(RExC_parse + 1) != 'b') {
13463 goto bad_bound_type;
13469 RExC_parse = endbrace;
13471 "'%" UTF8f "' is an unknown bound type",
13472 UTF8fARG(UTF, length, endbrace - length));
13473 NOT_REACHED; /*NOTREACHED*/
13475 RExC_parse = endbrace;
13476 REQUIRE_UNI_RULES(flagp, 0);
13481 else if (op >= BOUNDA) { /* /aa is same as /a */
13485 /* Don't have to worry about UTF-8, in this message because
13486 * to get here the contents of the \b must be ASCII */
13487 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13488 "Using /u for '%.*s' instead of /%s",
13490 endbrace - length + 1,
13491 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13492 ? ASCII_RESTRICT_PAT_MODS
13493 : ASCII_MORE_RESTRICT_PAT_MODS);
13498 RExC_seen_d_op = TRUE;
13500 else if (op == BOUNDL) {
13501 RExC_contains_locale = 1;
13505 op += NBOUND - BOUND;
13508 ret = reg_node(pRExC_state, op);
13509 FLAGS(REGNODE_p(ret)) = flags;
13513 goto finish_meta_pat;
13517 ret = reg_node(pRExC_state, LNBREAK);
13518 *flagp |= HASWIDTH|SIMPLE;
13519 goto finish_meta_pat;
13533 /* These all have the same meaning inside [brackets], and it knows
13534 * how to do the best optimizations for them. So, pretend we found
13535 * these within brackets, and let it do the work */
13538 ret = regclass(pRExC_state, flagp, depth+1,
13539 TRUE, /* means just parse this element */
13540 FALSE, /* don't allow multi-char folds */
13541 FALSE, /* don't silence non-portable warnings. It
13542 would be a bug if these returned
13544 (bool) RExC_strict,
13545 TRUE, /* Allow an optimized regnode result */
13547 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13548 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13549 * multi-char folds are allowed. */
13551 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13554 RExC_parse--; /* regclass() leaves this one too far ahead */
13557 /* The escapes above that don't take a parameter can't be
13558 * followed by a '{'. But 'pX', 'p{foo}' and
13559 * correspondingly 'P' can be */
13560 if ( RExC_parse - parse_start == 1
13561 && UCHARAT(RExC_parse + 1) == '{'
13562 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13565 vFAIL("Unescaped left brace in regex is illegal here");
13567 Set_Node_Offset(REGNODE_p(ret), parse_start);
13568 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13569 nextchar(pRExC_state);
13572 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13573 * \N{...} evaluates to a sequence of more than one code points).
13574 * The function call below returns a regnode, which is our result.
13575 * The parameters cause it to fail if the \N{} evaluates to a
13576 * single code point; we handle those like any other literal. The
13577 * reason that the multicharacter case is handled here and not as
13578 * part of the EXACtish code is because of quantifiers. In
13579 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13580 * this way makes that Just Happen. dmq.
13581 * join_exact() will join this up with adjacent EXACTish nodes
13582 * later on, if appropriate. */
13584 if (grok_bslash_N(pRExC_state,
13585 &ret, /* Want a regnode returned */
13586 NULL, /* Fail if evaluates to a single code
13588 NULL, /* Don't need a count of how many code
13597 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13599 /* Here, evaluates to a single code point. Go get that */
13600 RExC_parse = parse_start;
13603 case 'k': /* Handle \k<NAME> and \k'NAME' */
13607 if ( RExC_parse >= RExC_end - 1
13608 || (( ch = RExC_parse[1]) != '<'
13613 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13614 vFAIL2("Sequence %.2s... not terminated", parse_start);
13617 ret = handle_named_backref(pRExC_state,
13629 case '1': case '2': case '3': case '4':
13630 case '5': case '6': case '7': case '8': case '9':
13635 if (*RExC_parse == 'g') {
13639 if (*RExC_parse == '{') {
13643 if (*RExC_parse == '-') {
13647 if (hasbrace && !isDIGIT(*RExC_parse)) {
13648 if (isrel) RExC_parse--;
13650 goto parse_named_seq;
13653 if (RExC_parse >= RExC_end) {
13654 goto unterminated_g;
13656 num = S_backref_value(RExC_parse, RExC_end);
13658 vFAIL("Reference to invalid group 0");
13659 else if (num == I32_MAX) {
13660 if (isDIGIT(*RExC_parse))
13661 vFAIL("Reference to nonexistent group");
13664 vFAIL("Unterminated \\g... pattern");
13668 num = RExC_npar - num;
13670 vFAIL("Reference to nonexistent or unclosed group");
13674 num = S_backref_value(RExC_parse, RExC_end);
13675 /* bare \NNN might be backref or octal - if it is larger
13676 * than or equal RExC_npar then it is assumed to be an
13677 * octal escape. Note RExC_npar is +1 from the actual
13678 * number of parens. */
13679 /* Note we do NOT check if num == I32_MAX here, as that is
13680 * handled by the RExC_npar check */
13683 /* any numeric escape < 10 is always a backref */
13685 /* any numeric escape < RExC_npar is a backref */
13686 && num >= RExC_npar
13687 /* cannot be an octal escape if it starts with 8 */
13688 && *RExC_parse != '8'
13689 /* cannot be an octal escape if it starts with 9 */
13690 && *RExC_parse != '9'
13692 /* Probably not meant to be a backref, instead likely
13693 * to be an octal character escape, e.g. \35 or \777.
13694 * The above logic should make it obvious why using
13695 * octal escapes in patterns is problematic. - Yves */
13696 RExC_parse = parse_start;
13701 /* At this point RExC_parse points at a numeric escape like
13702 * \12 or \88 or something similar, which we should NOT treat
13703 * as an octal escape. It may or may not be a valid backref
13704 * escape. For instance \88888888 is unlikely to be a valid
13706 while (isDIGIT(*RExC_parse))
13709 if (*RExC_parse != '}')
13710 vFAIL("Unterminated \\g{...} pattern");
13713 if (num >= (I32)RExC_npar) {
13715 /* It might be a forward reference; we can't fail until we
13716 * know, by completing the parse to get all the groups, and
13717 * then reparsing */
13718 if (ALL_PARENS_COUNTED) {
13719 if (num >= RExC_total_parens) {
13720 vFAIL("Reference to nonexistent group");
13724 REQUIRE_PARENS_PASS;
13728 ret = reganode(pRExC_state,
13731 : (ASCII_FOLD_RESTRICTED)
13733 : (AT_LEAST_UNI_SEMANTICS)
13739 if (OP(REGNODE_p(ret)) == REFF) {
13740 RExC_seen_d_op = TRUE;
13742 *flagp |= HASWIDTH;
13744 /* override incorrect value set in reganode MJD */
13745 Set_Node_Offset(REGNODE_p(ret), parse_start);
13746 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
13747 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13748 FALSE /* Don't force to /x */ );
13752 if (RExC_parse >= RExC_end)
13753 FAIL("Trailing \\");
13756 /* Do not generate "unrecognized" warnings here, we fall
13757 back into the quick-grab loop below */
13758 RExC_parse = parse_start;
13760 } /* end of switch on a \foo sequence */
13765 /* '#' comments should have been spaced over before this function was
13767 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13769 if (RExC_flags & RXf_PMf_EXTENDED) {
13770 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13771 if (RExC_parse < RExC_end)
13781 /* Here, we have determined that the next thing is probably a
13782 * literal character. RExC_parse points to the first byte of its
13783 * definition. (It still may be an escape sequence that evaluates
13784 * to a single character) */
13791 U32 max_string_len = 255;
13793 /* We may have to reparse the node, artificially stopping filling
13794 * it early, based on info gleaned in the first parse. This
13795 * variable gives where we stop. Make it above the normal stopping
13796 * place first time through. */
13797 U32 upper_fill = max_string_len + 1;
13799 /* We start out as an EXACT node, even if under /i, until we find a
13800 * character which is in a fold. The algorithm now segregates into
13801 * separate nodes, characters that fold from those that don't under
13802 * /i. (This hopefully will create nodes that are fixed strings
13803 * even under /i, giving the optimizer something to grab on to.)
13804 * So, if a node has something in it and the next character is in
13805 * the opposite category, that node is closed up, and the function
13806 * returns. Then regatom is called again, and a new node is
13807 * created for the new category. */
13808 U8 node_type = EXACT;
13810 /* Assume the node will be fully used; the excess is given back at
13811 * the end. We can't make any other length assumptions, as a byte
13812 * input sequence could shrink down. */
13813 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len);
13815 bool next_is_quantifier;
13816 char * oldp = NULL;
13817 char * old_oldp = NULL;
13819 /* We can convert EXACTF nodes to EXACTFU if they contain only
13820 * characters that match identically regardless of the target
13821 * string's UTF8ness. The reason to do this is that EXACTF is not
13822 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
13825 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13826 * contain only above-Latin1 characters (hence must be in UTF8),
13827 * which don't participate in folds with Latin1-range characters,
13828 * as the latter's folds aren't known until runtime. */
13829 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13831 /* Single-character EXACTish nodes are almost always SIMPLE. This
13832 * allows us to override this as encountered */
13833 U8 maybe_SIMPLE = SIMPLE;
13835 /* Does this node contain something that can't match unless the
13836 * target string is (also) in UTF-8 */
13837 bool requires_utf8_target = FALSE;
13839 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
13840 bool has_ss = FALSE;
13842 /* So is the MICRO SIGN */
13843 bool has_micro_sign = FALSE;
13845 /* Set when we fill up the current node and there is still more
13846 * text to process */
13849 /* Allocate an EXACT node. The node_type may change below to
13850 * another EXACTish node, but since the size of the node doesn't
13851 * change, it works */
13852 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
13854 FILL_NODE(ret, node_type);
13857 s = STRING(REGNODE_p(ret));
13869 /* This breaks under rare circumstances. If folding, we do not
13870 * want to split a node at a character that is a non-final in a
13871 * multi-char fold, as an input string could just happen to want to
13872 * match across the node boundary. The code at the end of the loop
13873 * looks for this, and backs off until it finds not such a
13874 * character, but it is possible (though extremely, extremely
13875 * unlikely) for all characters in the node to be non-final fold
13876 * ones, in which case we just leave the node fully filled, and
13877 * hope that it doesn't match the string in just the wrong place */
13879 assert( ! UTF /* Is at the beginning of a character */
13880 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13881 || UTF8_IS_START(UCHARAT(RExC_parse)));
13883 overflowed = FALSE;
13885 /* Here, we have a literal character. Find the maximal string of
13886 * them in the input that we can fit into a single EXACTish node.
13887 * We quit at the first non-literal or when the node gets full, or
13888 * under /i the categorization of folding/non-folding character
13890 while (p < RExC_end && len < upper_fill) {
13892 /* In most cases each iteration adds one byte to the output.
13893 * The exceptions override this */
13894 Size_t added_len = 1;
13899 /* White space has already been ignored */
13900 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13901 || ! is_PATWS_safe((p), RExC_end, UTF));
13913 /* Literal Escapes Switch
13915 This switch is meant to handle escape sequences that
13916 resolve to a literal character.
13918 Every escape sequence that represents something
13919 else, like an assertion or a char class, is handled
13920 in the switch marked 'Special Escapes' above in this
13921 routine, but also has an entry here as anything that
13922 isn't explicitly mentioned here will be treated as
13923 an unescaped equivalent literal.
13926 switch ((U8)*++p) {
13928 /* These are all the special escapes. */
13929 case 'A': /* Start assertion */
13930 case 'b': case 'B': /* Word-boundary assertion*/
13931 case 'C': /* Single char !DANGEROUS! */
13932 case 'd': case 'D': /* digit class */
13933 case 'g': case 'G': /* generic-backref, pos assertion */
13934 case 'h': case 'H': /* HORIZWS */
13935 case 'k': case 'K': /* named backref, keep marker */
13936 case 'p': case 'P': /* Unicode property */
13937 case 'R': /* LNBREAK */
13938 case 's': case 'S': /* space class */
13939 case 'v': case 'V': /* VERTWS */
13940 case 'w': case 'W': /* word class */
13941 case 'X': /* eXtended Unicode "combining
13942 character sequence" */
13943 case 'z': case 'Z': /* End of line/string assertion */
13947 /* Anything after here is an escape that resolves to a
13948 literal. (Except digits, which may or may not)
13954 case 'N': /* Handle a single-code point named character. */
13955 RExC_parse = p + 1;
13956 if (! grok_bslash_N(pRExC_state,
13957 NULL, /* Fail if evaluates to
13958 anything other than a
13959 single code point */
13960 &ender, /* The returned single code
13962 NULL, /* Don't need a count of
13963 how many code points */
13968 if (*flagp & NEED_UTF8)
13969 FAIL("panic: grok_bslash_N set NEED_UTF8");
13970 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13972 /* Here, it wasn't a single code point. Go close
13973 * up this EXACTish node. The switch() prior to
13974 * this switch handles the other cases */
13975 RExC_parse = p = oldp;
13979 RExC_parse = parse_start;
13981 /* The \N{} means the pattern, if previously /d,
13982 * becomes /u. That means it can't be an EXACTF node,
13983 * but an EXACTFU */
13984 if (node_type == EXACTF) {
13985 node_type = EXACTFU;
13987 /* If the node already contains something that
13988 * differs between EXACTF and EXACTFU, reparse it
13990 if (! maybe_exactfu) {
14011 ender = ESC_NATIVE;
14021 const char* error_msg;
14023 bool valid = grok_bslash_o(&p,
14027 TO_OUTPUT_WARNINGS(p),
14028 (bool) RExC_strict,
14029 TRUE, /* Output warnings
14034 RExC_parse = p; /* going to die anyway; point
14035 to exact spot of failure */
14038 UPDATE_WARNINGS_LOC(p - 1);
14044 UV result = UV_MAX; /* initialize to erroneous
14046 const char* error_msg;
14048 bool valid = grok_bslash_x(&p,
14052 TO_OUTPUT_WARNINGS(p),
14053 (bool) RExC_strict,
14054 TRUE, /* Silence warnings
14059 RExC_parse = p; /* going to die anyway; point
14060 to exact spot of failure */
14063 UPDATE_WARNINGS_LOC(p - 1);
14067 if (ender < 0x100) {
14068 if (RExC_recode_x_to_native) {
14069 ender = LATIN1_TO_NATIVE(ender);
14077 ender = grok_bslash_c(*p, TO_OUTPUT_WARNINGS(p));
14078 UPDATE_WARNINGS_LOC(p);
14081 case '8': case '9': /* must be a backreference */
14083 /* we have an escape like \8 which cannot be an octal escape
14084 * so we exit the loop, and let the outer loop handle this
14085 * escape which may or may not be a legitimate backref. */
14087 case '1': case '2': case '3':case '4':
14088 case '5': case '6': case '7':
14089 /* When we parse backslash escapes there is ambiguity
14090 * between backreferences and octal escapes. Any escape
14091 * from \1 - \9 is a backreference, any multi-digit
14092 * escape which does not start with 0 and which when
14093 * evaluated as decimal could refer to an already
14094 * parsed capture buffer is a back reference. Anything
14097 * Note this implies that \118 could be interpreted as
14098 * 118 OR as "\11" . "8" depending on whether there
14099 * were 118 capture buffers defined already in the
14102 /* NOTE, RExC_npar is 1 more than the actual number of
14103 * parens we have seen so far, hence the "<" as opposed
14105 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14106 { /* Not to be treated as an octal constant, go
14114 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14116 ender = grok_oct(p, &numlen, &flags, NULL);
14118 if ( isDIGIT(*p) /* like \08, \178 */
14119 && ckWARN(WARN_REGEXP)
14122 reg_warn_non_literal_string(
14124 form_short_octal_warning(p, numlen));
14130 FAIL("Trailing \\");
14133 if (isALPHANUMERIC(*p)) {
14134 /* An alpha followed by '{' is going to fail next
14135 * iteration, so don't output this warning in that
14137 if (! isALPHA(*p) || *(p + 1) != '{') {
14138 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14139 " passed through", p);
14142 goto normal_default;
14143 } /* End of switch on '\' */
14146 /* Trying to gain new uses for '{' without breaking too
14147 * much existing code is hard. The solution currently
14149 * 1) If there is no ambiguity that a '{' should always
14150 * be taken literally, at the start of a construct, we
14152 * 2) If the literal '{' conflicts with our desired use
14153 * of it as a metacharacter, we die. The deprecation
14154 * cycles for this have come and gone.
14155 * 3) If there is ambiguity, we raise a simple warning.
14156 * This could happen, for example, if the user
14157 * intended it to introduce a quantifier, but slightly
14158 * misspelled the quantifier. Without this warning,
14159 * the quantifier would silently be taken as a literal
14160 * string of characters instead of a meta construct */
14161 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14163 || ( p > parse_start + 1
14164 && isALPHA_A(*(p - 1))
14165 && *(p - 2) == '\\')
14166 || new_regcurly(p, RExC_end))
14168 RExC_parse = p + 1;
14169 vFAIL("Unescaped left brace in regex is "
14172 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14173 " passed through");
14175 goto normal_default;
14178 if (p > RExC_parse && RExC_strict) {
14179 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14182 default: /* A literal character */
14184 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14186 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14187 &numlen, UTF8_ALLOW_DEFAULT);
14193 } /* End of switch on the literal */
14195 /* Here, have looked at the literal character, and <ender>
14196 * contains its ordinal; <p> points to the character after it.
14200 REQUIRE_UTF8(flagp);
14203 /* We need to check if the next non-ignored thing is a
14204 * quantifier. Move <p> to after anything that should be
14205 * ignored, which, as a side effect, positions <p> for the next
14206 * loop iteration */
14207 skip_to_be_ignored_text(pRExC_state, &p,
14208 FALSE /* Don't force to /x */ );
14210 /* If the next thing is a quantifier, it applies to this
14211 * character only, which means that this character has to be in
14212 * its own node and can't just be appended to the string in an
14213 * existing node, so if there are already other characters in
14214 * the node, close the node with just them, and set up to do
14215 * this character again next time through, when it will be the
14216 * only thing in its new node */
14218 next_is_quantifier = LIKELY(p < RExC_end)
14219 && UNLIKELY(ISMULT2(p));
14221 if (next_is_quantifier && LIKELY(len)) {
14226 /* Ready to add 'ender' to the node */
14228 if (! FOLD) { /* The simple case, just append the literal */
14231 /* Don't output if it would overflow */
14232 if (UNLIKELY(len > max_string_len - ((UTF)
14233 ? UVCHR_SKIP(ender)
14240 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14241 *(s++) = (char) ender;
14244 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14245 added_len = (char *) new_s - s;
14246 s = (char *) new_s;
14249 requires_utf8_target = TRUE;
14253 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14255 /* Here are folding under /l, and the code point is
14256 * problematic. If this is the first character in the
14257 * node, change the node type to folding. Otherwise, if
14258 * this is the first problematic character, close up the
14259 * existing node, so can start a new node with this one */
14261 node_type = EXACTFL;
14262 RExC_contains_locale = 1;
14264 else if (node_type == EXACT) {
14269 /* This problematic code point means we can't simplify
14271 maybe_exactfu = FALSE;
14273 /* Here, we are adding a problematic fold character.
14274 * "Problematic" in this context means that its fold isn't
14275 * known until runtime. (The non-problematic code points
14276 * are the above-Latin1 ones that fold to also all
14277 * above-Latin1. Their folds don't vary no matter what the
14278 * locale is.) But here we have characters whose fold
14279 * depends on the locale. We just add in the unfolded
14280 * character, and wait until runtime to fold it */
14281 goto not_fold_common;
14283 else /* regular fold; see if actually is in a fold */
14284 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14286 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14288 /* Here, folding, but the character isn't in a fold.
14290 * Start a new node if previous characters in the node were
14292 if (len && node_type != EXACT) {
14297 /* Here, continuing a node with non-folded characters. Add
14299 goto not_fold_common;
14301 else { /* Here, does participate in some fold */
14303 /* If this is the first character in the node, change its
14304 * type to folding. Otherwise, if this is the first
14305 * folding character in the node, close up the existing
14306 * node, so can start a new node with this one. */
14308 node_type = compute_EXACTish(pRExC_state);
14310 else if (node_type == EXACT) {
14315 if (UTF) { /* Use the folded value */
14316 if (UVCHR_IS_INVARIANT(ender)) {
14317 if (UNLIKELY(len + 1 > max_string_len)) {
14322 *(s)++ = (U8) toFOLD(ender);
14325 U8 temp[UTF8_MAXBYTES_CASE+1];
14327 UV folded = _to_uni_fold_flags(
14331 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14332 ? FOLD_FLAGS_NOMIX_ASCII
14334 if (UNLIKELY(len + added_len > max_string_len)) {
14339 Copy(temp, s, added_len, char);
14343 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14345 /* U+B5 folds to the MU, so its possible for a
14346 * non-UTF-8 target to match it */
14347 requires_utf8_target = TRUE;
14353 /* Here is non-UTF8. First, see if the character's
14354 * fold differs between /d and /u. */
14355 if (PL_fold[ender] != PL_fold_latin1[ender]) {
14356 maybe_exactfu = FALSE;
14359 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14360 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14361 || UNICODE_DOT_DOT_VERSION > 0)
14363 /* On non-ancient Unicode versions, this includes the
14364 * multi-char fold SHARP S to 'ss' */
14366 if ( UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)
14367 || ( isALPHA_FOLD_EQ(ender, 's')
14369 && isALPHA_FOLD_EQ(*(s-1), 's')))
14371 /* Here, we have one of the following:
14372 * a) a SHARP S. This folds to 'ss' only under
14373 * /u rules. If we are in that situation,
14374 * fold the SHARP S to 'ss'. See the comments
14375 * for join_exact() as to why we fold this
14376 * non-UTF at compile time, and no others.
14377 * b) 'ss'. When under /u, there's nothing
14378 * special needed to be done here. The
14379 * previous iteration handled the first 's',
14380 * and this iteration will handle the second.
14381 * If, on the otherhand it's not /u, we have
14382 * to exclude the possibility of moving to /u,
14383 * so that we won't generate an unwanted
14384 * match, unless, at runtime, the target
14385 * string is in UTF-8.
14389 maybe_exactfu = FALSE; /* Can't generate an
14390 EXACTFU node (unless we
14391 already are in one) */
14392 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14394 if (node_type == EXACTFU) {
14396 if (UNLIKELY(len + 2 > max_string_len)) {
14403 /* Let the code below add in the extra 's'
14412 else if (UNLIKELY(ender == MICRO_SIGN)) {
14413 has_micro_sign = TRUE;
14416 if (UNLIKELY(len + 1 > max_string_len)) {
14421 *(s++) = (DEPENDS_SEMANTICS)
14422 ? (char) toFOLD(ender)
14424 /* Under /u, the fold of any character in
14425 * the 0-255 range happens to be its
14426 * lowercase equivalent, except for LATIN
14427 * SMALL LETTER SHARP S, which was handled
14428 * above, and the MICRO SIGN, whose fold
14429 * requires UTF-8 to represent. */
14430 : (char) toLOWER_L1(ender);
14432 } /* End of adding current character to the node */
14436 if (next_is_quantifier) {
14438 /* Here, the next input is a quantifier, and to get here,
14439 * the current character is the only one in the node. */
14443 } /* End of loop through literal characters */
14445 /* Here we have either exhausted the input or run out of room in
14446 * the node. If the former, we are done. (If we encountered a
14447 * character that can't be in the node, transfer is made directly
14448 * to <loopdone>, and so we wouldn't have fallen off the end of the
14450 if (LIKELY(! overflowed)) {
14454 /* Here we have run out of room. We can grow plain EXACT and
14455 * LEXACT nodes. If the pattern is gigantic enough, though,
14456 * eventually we'll have to artificially chunk the pattern into
14457 * multiple nodes. */
14458 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14459 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14460 Size_t overhead_expansion = 0;
14462 Size_t max_nodes_for_string;
14466 /* Here we couldn't fit the final character in the current
14467 * node, so it will have to be reparsed, no matter what else we
14471 /* If would have overflowed a regular EXACT node, switch
14472 * instead to an LEXACT. The code below is structured so that
14473 * the actual growing code is common to changing from an EXACT
14474 * or just increasing the LEXACT size. This means that we have
14475 * to save the string in the EXACT case before growing, and
14476 * then copy it afterwards to its new location */
14477 if (node_type == EXACT) {
14478 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14479 RExC_emit += overhead_expansion;
14480 Copy(s0, temp, len, char);
14483 /* Ready to grow. If it was a plain EXACT, the string was
14484 * saved, and the first few bytes of it overwritten by adding
14485 * an argument field. We assume, as we do elsewhere in this
14486 * file, that one byte of remaining input will translate into
14487 * one byte of output, and if that's too small, we grow again,
14488 * if too large the excess memory is freed at the end */
14490 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14491 achievable = MIN(max_nodes_for_string,
14492 current_string_nodes + STR_SZ(RExC_end - p));
14493 delta = achievable - current_string_nodes;
14495 /* If there is just no more room, go finish up this chunk of
14501 change_engine_size(pRExC_state, delta + overhead_expansion);
14502 current_string_nodes += delta;
14504 = sizeof(struct regnode) * current_string_nodes;
14505 upper_fill = max_string_len + 1;
14507 /* If the length was small, we know this was originally an
14508 * EXACT node now converted to LEXACT, and the string has to be
14509 * restored. Otherwise the string was untouched. 260 is just
14510 * a number safely above 255 so don't have to worry about
14511 * getting it precise */
14513 node_type = LEXACT;
14514 FILL_NODE(ret, node_type);
14515 s0 = STRING(REGNODE_p(ret));
14516 Copy(temp, s0, len, char);
14520 goto continue_parse;
14524 /* Here is /i. Running out of room creates a problem if we are
14525 * folding, and the split happens in the middle of a
14526 * multi-character fold, as a match that should have occurred,
14527 * won't, due to the way nodes are matched, and our artificial
14528 * boundary. So back off until we aren't splitting such a
14529 * fold. If there is no such place to back off to, we end up
14530 * taking the entire node as-is. This can happen if the node
14531 * consists entirely of 'f' or entirely of 's' characters (or
14532 * things that fold to them) as 'ff' and 'ss' are
14533 * multi-character folds.
14536 * old_oldp points to the beginning in the input of the
14537 * penultimate character in the node.
14538 * oldp points to the beginning in the input of the
14539 * final character in the node.
14540 * p points to the beginning in the input of the
14541 * next character in the input, the one that won't
14544 * We aren't in the middle of a multi-char fold unless the
14545 * final character in the node can appear in a non-final
14546 * position in such a fold. Very few characters actually
14547 * participate in multi-character folds, and fewer still can be
14548 * in the non-final position. But it's complicated to know
14549 * here if that final character is folded or not, so skip this
14552 /* Make sure enough space for final char of node,
14553 * first char of following node, and the fold of the
14554 * following char (so we don't have to worry about
14555 * that fold running off the end */
14556 U8 foldbuf[UTF8_MAXBYTES_CASE * 5 + 1];
14559 char * const sav_oldp = oldp;
14563 /* The Unicode standard says that multi character folds consist
14564 * of either two or three characters. So we create a buffer
14565 * containing a window of three. The first is the final
14566 * character in the node (folded), and then the two that begin
14567 * the following node. But if the first character of the
14568 * following node can't be in a non-final fold position, there
14569 * is no need to look at its successor character. The macros
14570 * used below to check for multi character folds require folded
14571 * inputs, so we have to fold these. (The fold of p was likely
14572 * calculated in the loop above, but it hasn't beeen saved, and
14573 * khw thinks it would be too entangled to change to do so) */
14575 if (UTF || LIKELY(UCHARAT(p) != MICRO_SIGN)) {
14576 folded = _to_uni_fold_flags(ender,
14582 foldbuf[0] = folded = MICRO_SIGN;
14586 /* Here, foldbuf contains the fold of the first character in
14587 * the next node. We may also need the next one (if there is
14588 * one) to get our third, but if the first character folded to
14589 * more than one, those extra one(s) will serve as the third.
14590 * Also, we don't need a third unless the previous one can
14591 * appear in a non-final position in a fold */
14592 if ( ((RExC_end - p) > ((UTF) ? UVCHR_SKIP(ender) : 1))
14593 && (fold_len == 1 || ( UTF
14594 && UVCHR_SKIP(folded) == fold_len))
14595 && UNLIKELY(_invlist_contains_cp(PL_NonFinalFold, folded)))
14598 STRLEN next_fold_len;
14600 toFOLD_utf8_safe((U8*) p + UTF8SKIP(p),
14601 (U8*) RExC_end, foldbuf + fold_len,
14603 fold_len += next_fold_len;
14606 if (UNLIKELY(p[1] == LATIN_SMALL_LETTER_SHARP_S)) {
14607 foldbuf[fold_len] = 's';
14610 foldbuf[fold_len] = toLOWER_L1(p[1]);
14616 /* Here foldbuf contains the the fold of p, and if appropriate
14617 * that of the character following p in the input. */
14619 /* Search backwards until find a place that doesn't split a
14620 * multi-char fold */
14623 char s_fold_buf[UTF8_MAXBYTES_CASE];
14624 char * s_fold = s_fold_buf;
14628 /* There's no safe place in the node to split. Quit so
14629 * will take the whole node */
14634 /* Backup 1 character. The first time through this moves s
14635 * to point to the final character in the node */
14637 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s0);
14643 /* 's' may or may not be folded; so make sure it is, and
14644 * use just the final character in its fold (should there
14645 * be more than one */
14647 toFOLD_utf8_safe((U8*) s,
14648 (U8*) s + UTF8SKIP(s),
14649 (U8 *) s_fold_buf, &s_len);
14650 while (s_fold + UTF8SKIP(s_fold) < s_fold_buf + s_len)
14652 s_fold += UTF8SKIP(s_fold);
14654 s_len = UTF8SKIP(s_fold);
14657 if (UNLIKELY(UCHARAT(s) == LATIN_SMALL_LETTER_SHARP_S))
14659 s_fold_buf[0] = 's';
14661 else { /* This works for all other non-UTF-8 folds
14663 s_fold_buf[0] = toLOWER_L1(UCHARAT(s));
14668 /* Unshift this character to the beginning of the buffer,
14669 * No longer needed trailing characters are overwritten.
14671 Move(foldbuf, foldbuf + s_len, sizeof(foldbuf) - s_len, U8);
14672 Copy(s_fold, foldbuf, s_len, U8);
14674 /* If this isn't a multi-character fold, we have found a
14675 * splittable place. If this is the final character in the
14676 * node, that means the node is valid as-is, and can quit.
14677 * Otherwise, we note how much we can fill the node before
14678 * coming to a non-splittable position, and go parse it
14679 * again, stopping there. This is done because we know
14680 * where in the output to stop, but we don't have a map to
14681 * where that is in the input. One could be created, but
14682 * it seems like overkill for such a rare event as we are
14683 * dealing with here */
14685 if (! is_MULTI_CHAR_FOLD_utf8_safe(foldbuf,
14686 foldbuf + UTF8_MAXBYTES_CASE))
14688 upper_fill = s + UTF8SKIP(s) - s0;
14689 if (LIKELY(oldp)) {
14695 else if (! is_MULTI_CHAR_FOLD_latin1_safe(foldbuf,
14696 foldbuf + UTF8_MAXBYTES_CASE))
14698 upper_fill = s + 1 - s0;
14699 if (LIKELY(oldp)) {
14708 } /* End of loop backing up through the node */
14710 /* Here the node consists entirely of non-final multi-char
14711 * folds. (Likely it is all 'f's or all 's's.) There's no
14712 * decent place to split it, so give up and just take the whole
14715 } /* End of verifying node ends with an appropriate char */
14719 loopdone: /* Jumped to when encounters something that shouldn't be
14722 /* Free up any over-allocated space; cast is to silence bogus
14723 * warning in MS VC */
14724 change_engine_size(pRExC_state,
14725 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
14727 /* I (khw) don't know if you can get here with zero length, but the
14728 * old code handled this situation by creating a zero-length EXACT
14729 * node. Might as well be NOTHING instead */
14731 OP(REGNODE_p(ret)) = NOTHING;
14735 /* If the node type is EXACT here, check to see if it
14736 * should be EXACTL, or EXACT_REQ8. */
14737 if (node_type == EXACT) {
14739 node_type = EXACTL;
14741 else if (requires_utf8_target) {
14742 node_type = EXACT_REQ8;
14745 else if (node_type == LEXACT) {
14746 if (requires_utf8_target) {
14747 node_type = LEXACT_REQ8;
14751 if ( UNLIKELY(has_micro_sign || has_ss)
14752 && (node_type == EXACTFU || ( node_type == EXACTF
14753 && maybe_exactfu)))
14754 { /* These two conditions are problematic in non-UTF-8
14757 node_type = EXACTFUP;
14759 else if (node_type == EXACTFL) {
14761 /* 'maybe_exactfu' is deliberately set above to
14762 * indicate this node type, where all code points in it
14764 if (maybe_exactfu) {
14765 node_type = EXACTFLU8;
14768 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
14770 /* A character that folds to more than one will
14771 * match multiple characters, so can't be SIMPLE.
14772 * We don't have to worry about this with EXACTFLU8
14773 * nodes just above, as they have already been
14774 * folded (since the fold doesn't vary at run
14775 * time). Here, if the final character in the node
14776 * folds to multiple, it can't be simple. (This
14777 * only has an effect if the node has only a single
14778 * character, hence the final one, as elsewhere we
14779 * turn off simple for nodes whose length > 1 */
14783 else if (node_type == EXACTF) { /* Means is /di */
14785 /* This intermediate variable is needed solely because
14786 * the asserts in the macro where used exceed Win32's
14787 * literal string capacity */
14788 char first_char = * STRING(REGNODE_p(ret));
14790 /* If 'maybe_exactfu' is clear, then we need to stay
14791 * /di. If it is set, it means there are no code
14792 * points that match differently depending on UTF8ness
14793 * of the target string, so it can become an EXACTFU
14795 if (! maybe_exactfu) {
14796 RExC_seen_d_op = TRUE;
14798 else if ( isALPHA_FOLD_EQ(first_char, 's')
14799 || isALPHA_FOLD_EQ(ender, 's'))
14801 /* But, if the node begins or ends in an 's' we
14802 * have to defer changing it into an EXACTFU, as
14803 * the node could later get joined with another one
14804 * that ends or begins with 's' creating an 'ss'
14805 * sequence which would then wrongly match the
14806 * sharp s without the target being UTF-8. We
14807 * create a special node that we resolve later when
14808 * we join nodes together */
14810 node_type = EXACTFU_S_EDGE;
14813 node_type = EXACTFU;
14817 if (requires_utf8_target && node_type == EXACTFU) {
14818 node_type = EXACTFU_REQ8;
14822 OP(REGNODE_p(ret)) = node_type;
14823 setSTR_LEN(REGNODE_p(ret), len);
14824 RExC_emit += STR_SZ(len);
14826 /* If the node isn't a single character, it can't be SIMPLE */
14827 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
14831 *flagp |= HASWIDTH | maybe_SIMPLE;
14834 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
14838 /* len is STRLEN which is unsigned, need to copy to signed */
14841 vFAIL("Internal disaster");
14844 } /* End of label 'defchar:' */
14846 } /* End of giant switch on input character */
14848 /* Position parse to next real character */
14849 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
14850 FALSE /* Don't force to /x */ );
14851 if ( *RExC_parse == '{'
14852 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
14854 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
14856 vFAIL("Unescaped left brace in regex is illegal here");
14858 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
14859 " passed through");
14867 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
14869 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
14870 * sets up the bitmap and any flags, removing those code points from the
14871 * inversion list, setting it to NULL should it become completely empty */
14875 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
14876 assert(PL_regkind[OP(node)] == ANYOF);
14878 /* There is no bitmap for this node type */
14879 if (inRANGE(OP(node), ANYOFH, ANYOFHr)) {
14883 ANYOF_BITMAP_ZERO(node);
14884 if (*invlist_ptr) {
14886 /* This gets set if we actually need to modify things */
14887 bool change_invlist = FALSE;
14891 /* Start looking through *invlist_ptr */
14892 invlist_iterinit(*invlist_ptr);
14893 while (invlist_iternext(*invlist_ptr, &start, &end)) {
14897 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
14898 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
14901 /* Quit if are above what we should change */
14902 if (start >= NUM_ANYOF_CODE_POINTS) {
14906 change_invlist = TRUE;
14908 /* Set all the bits in the range, up to the max that we are doing */
14909 high = (end < NUM_ANYOF_CODE_POINTS - 1)
14911 : NUM_ANYOF_CODE_POINTS - 1;
14912 for (i = start; i <= (int) high; i++) {
14913 if (! ANYOF_BITMAP_TEST(node, i)) {
14914 ANYOF_BITMAP_SET(node, i);
14918 invlist_iterfinish(*invlist_ptr);
14920 /* Done with loop; remove any code points that are in the bitmap from
14921 * *invlist_ptr; similarly for code points above the bitmap if we have
14922 * a flag to match all of them anyways */
14923 if (change_invlist) {
14924 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
14926 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
14927 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
14930 /* If have completely emptied it, remove it completely */
14931 if (_invlist_len(*invlist_ptr) == 0) {
14932 SvREFCNT_dec_NN(*invlist_ptr);
14933 *invlist_ptr = NULL;
14938 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
14939 Character classes ([:foo:]) can also be negated ([:^foo:]).
14940 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
14941 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
14942 but trigger failures because they are currently unimplemented. */
14944 #define POSIXCC_DONE(c) ((c) == ':')
14945 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
14946 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
14947 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
14949 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
14950 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
14951 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
14953 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
14955 /* 'posix_warnings' and 'warn_text' are names of variables in the following
14957 #define ADD_POSIX_WARNING(p, text) STMT_START { \
14958 if (posix_warnings) { \
14959 if (! RExC_warn_text ) RExC_warn_text = \
14960 (AV *) sv_2mortal((SV *) newAV()); \
14961 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
14965 REPORT_LOCATION_ARGS(p))); \
14968 #define CLEAR_POSIX_WARNINGS() \
14970 if (posix_warnings && RExC_warn_text) \
14971 av_clear(RExC_warn_text); \
14974 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
14976 CLEAR_POSIX_WARNINGS(); \
14981 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
14983 const char * const s, /* Where the putative posix class begins.
14984 Normally, this is one past the '['. This
14985 parameter exists so it can be somewhere
14986 besides RExC_parse. */
14987 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
14989 AV ** posix_warnings, /* Where to place any generated warnings, or
14991 const bool check_only /* Don't die if error */
14994 /* This parses what the caller thinks may be one of the three POSIX
14996 * 1) a character class, like [:blank:]
14997 * 2) a collating symbol, like [. .]
14998 * 3) an equivalence class, like [= =]
14999 * In the latter two cases, it croaks if it finds a syntactically legal
15000 * one, as these are not handled by Perl.
15002 * The main purpose is to look for a POSIX character class. It returns:
15003 * a) the class number
15004 * if it is a completely syntactically and semantically legal class.
15005 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15006 * closing ']' of the class
15007 * b) OOB_NAMEDCLASS
15008 * if it appears that one of the three POSIX constructs was meant, but
15009 * its specification was somehow defective. 'updated_parse_ptr', if
15010 * not NULL, is set to point to the character just after the end
15011 * character of the class. See below for handling of warnings.
15012 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15013 * if it doesn't appear that a POSIX construct was intended.
15014 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15017 * In b) there may be errors or warnings generated. If 'check_only' is
15018 * TRUE, then any errors are discarded. Warnings are returned to the
15019 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15020 * instead it is NULL, warnings are suppressed.
15022 * The reason for this function, and its complexity is that a bracketed
15023 * character class can contain just about anything. But it's easy to
15024 * mistype the very specific posix class syntax but yielding a valid
15025 * regular bracketed class, so it silently gets compiled into something
15026 * quite unintended.
15028 * The solution adopted here maintains backward compatibility except that
15029 * it adds a warning if it looks like a posix class was intended but
15030 * improperly specified. The warning is not raised unless what is input
15031 * very closely resembles one of the 14 legal posix classes. To do this,
15032 * it uses fuzzy parsing. It calculates how many single-character edits it
15033 * would take to transform what was input into a legal posix class. Only
15034 * if that number is quite small does it think that the intention was a
15035 * posix class. Obviously these are heuristics, and there will be cases
15036 * where it errs on one side or another, and they can be tweaked as
15037 * experience informs.
15039 * The syntax for a legal posix class is:
15041 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15043 * What this routine considers syntactically to be an intended posix class
15044 * is this (the comments indicate some restrictions that the pattern
15047 * qr/(?x: \[? # The left bracket, possibly
15049 * \h* # possibly followed by blanks
15050 * (?: \^ \h* )? # possibly a misplaced caret
15051 * [:;]? # The opening class character,
15052 * # possibly omitted. A typo
15053 * # semi-colon can also be used.
15055 * \^? # possibly a correctly placed
15056 * # caret, but not if there was also
15057 * # a misplaced one
15059 * .{3,15} # The class name. If there are
15060 * # deviations from the legal syntax,
15061 * # its edit distance must be close
15062 * # to a real class name in order
15063 * # for it to be considered to be
15064 * # an intended posix class.
15066 * [[:punct:]]? # The closing class character,
15067 * # possibly omitted. If not a colon
15068 * # nor semi colon, the class name
15069 * # must be even closer to a valid
15072 * \]? # The right bracket, possibly
15076 * In the above, \h must be ASCII-only.
15078 * These are heuristics, and can be tweaked as field experience dictates.
15079 * There will be cases when someone didn't intend to specify a posix class
15080 * that this warns as being so. The goal is to minimize these, while
15081 * maximizing the catching of things intended to be a posix class that
15082 * aren't parsed as such.
15086 const char * const e = RExC_end;
15087 unsigned complement = 0; /* If to complement the class */
15088 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15089 bool has_opening_bracket = FALSE;
15090 bool has_opening_colon = FALSE;
15091 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15093 const char * possible_end = NULL; /* used for a 2nd parse pass */
15094 const char* name_start; /* ptr to class name first char */
15096 /* If the number of single-character typos the input name is away from a
15097 * legal name is no more than this number, it is considered to have meant
15098 * the legal name */
15099 int max_distance = 2;
15101 /* to store the name. The size determines the maximum length before we
15102 * decide that no posix class was intended. Should be at least
15103 * sizeof("alphanumeric") */
15105 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15107 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15109 CLEAR_POSIX_WARNINGS();
15112 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15115 if (*(p - 1) != '[') {
15116 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15117 found_problem = TRUE;
15120 has_opening_bracket = TRUE;
15123 /* They could be confused and think you can put spaces between the
15126 found_problem = TRUE;
15130 } while (p < e && isBLANK(*p));
15132 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15135 /* For [. .] and [= =]. These are quite different internally from [: :],
15136 * so they are handled separately. */
15137 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15138 and 1 for at least one char in it
15141 const char open_char = *p;
15142 const char * temp_ptr = p + 1;
15144 /* These two constructs are not handled by perl, and if we find a
15145 * syntactically valid one, we croak. khw, who wrote this code, finds
15146 * this explanation of them very unclear:
15147 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15148 * And searching the rest of the internet wasn't very helpful either.
15149 * It looks like just about any byte can be in these constructs,
15150 * depending on the locale. But unless the pattern is being compiled
15151 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15152 * In that case, it looks like [= =] isn't allowed at all, and that
15153 * [. .] could be any single code point, but for longer strings the
15154 * constituent characters would have to be the ASCII alphabetics plus
15155 * the minus-hyphen. Any sensible locale definition would limit itself
15156 * to these. And any portable one definitely should. Trying to parse
15157 * the general case is a nightmare (see [perl #127604]). So, this code
15158 * looks only for interiors of these constructs that match:
15160 * Using \w relaxes the apparent rules a little, without adding much
15161 * danger of mistaking something else for one of these constructs.
15163 * [. .] in some implementations described on the internet is usable to
15164 * escape a character that otherwise is special in bracketed character
15165 * classes. For example [.].] means a literal right bracket instead of
15166 * the ending of the class
15168 * [= =] can legitimately contain a [. .] construct, but we don't
15169 * handle this case, as that [. .] construct will later get parsed
15170 * itself and croak then. And [= =] is checked for even when not under
15171 * /l, as Perl has long done so.
15173 * The code below relies on there being a trailing NUL, so it doesn't
15174 * have to keep checking if the parse ptr < e.
15176 if (temp_ptr[1] == open_char) {
15179 else while ( temp_ptr < e
15180 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15185 if (*temp_ptr == open_char) {
15187 if (*temp_ptr == ']') {
15189 if (! found_problem && ! check_only) {
15190 RExC_parse = (char *) temp_ptr;
15191 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15192 "extensions", open_char, open_char);
15195 /* Here, the syntax wasn't completely valid, or else the call
15196 * is to check-only */
15197 if (updated_parse_ptr) {
15198 *updated_parse_ptr = (char *) temp_ptr;
15201 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15205 /* If we find something that started out to look like one of these
15206 * constructs, but isn't, we continue below so that it can be checked
15207 * for being a class name with a typo of '.' or '=' instead of a colon.
15211 /* Here, we think there is a possibility that a [: :] class was meant, and
15212 * we have the first real character. It could be they think the '^' comes
15215 found_problem = TRUE;
15216 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15221 found_problem = TRUE;
15225 } while (p < e && isBLANK(*p));
15227 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15231 /* But the first character should be a colon, which they could have easily
15232 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15233 * distinguish from a colon, so treat that as a colon). */
15236 has_opening_colon = TRUE;
15238 else if (*p == ';') {
15239 found_problem = TRUE;
15241 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15242 has_opening_colon = TRUE;
15245 found_problem = TRUE;
15246 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15248 /* Consider an initial punctuation (not one of the recognized ones) to
15249 * be a left terminator */
15250 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15255 /* They may think that you can put spaces between the components */
15257 found_problem = TRUE;
15261 } while (p < e && isBLANK(*p));
15263 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15268 /* We consider something like [^:^alnum:]] to not have been intended to
15269 * be a posix class, but XXX maybe we should */
15271 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15278 /* Again, they may think that you can put spaces between the components */
15280 found_problem = TRUE;
15284 } while (p < e && isBLANK(*p));
15286 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15291 /* XXX This ']' may be a typo, and something else was meant. But
15292 * treating it as such creates enough complications, that that
15293 * possibility isn't currently considered here. So we assume that the
15294 * ']' is what is intended, and if we've already found an initial '[',
15295 * this leaves this construct looking like [:] or [:^], which almost
15296 * certainly weren't intended to be posix classes */
15297 if (has_opening_bracket) {
15298 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15301 /* But this function can be called when we parse the colon for
15302 * something like qr/[alpha:]]/, so we back up to look for the
15307 found_problem = TRUE;
15308 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15310 else if (*p != ':') {
15312 /* XXX We are currently very restrictive here, so this code doesn't
15313 * consider the possibility that, say, /[alpha.]]/ was intended to
15314 * be a posix class. */
15315 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15318 /* Here we have something like 'foo:]'. There was no initial colon,
15319 * and we back up over 'foo. XXX Unlike the going forward case, we
15320 * don't handle typos of non-word chars in the middle */
15321 has_opening_colon = FALSE;
15324 while (p > RExC_start && isWORDCHAR(*p)) {
15329 /* Here, we have positioned ourselves to where we think the first
15330 * character in the potential class is */
15333 /* Now the interior really starts. There are certain key characters that
15334 * can end the interior, or these could just be typos. To catch both
15335 * cases, we may have to do two passes. In the first pass, we keep on
15336 * going unless we come to a sequence that matches
15337 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15338 * This means it takes a sequence to end the pass, so two typos in a row if
15339 * that wasn't what was intended. If the class is perfectly formed, just
15340 * this one pass is needed. We also stop if there are too many characters
15341 * being accumulated, but this number is deliberately set higher than any
15342 * real class. It is set high enough so that someone who thinks that
15343 * 'alphanumeric' is a correct name would get warned that it wasn't.
15344 * While doing the pass, we keep track of where the key characters were in
15345 * it. If we don't find an end to the class, and one of the key characters
15346 * was found, we redo the pass, but stop when we get to that character.
15347 * Thus the key character was considered a typo in the first pass, but a
15348 * terminator in the second. If two key characters are found, we stop at
15349 * the second one in the first pass. Again this can miss two typos, but
15350 * catches a single one
15352 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15353 * point to the first key character. For the second pass, it starts as -1.
15359 bool has_blank = FALSE;
15360 bool has_upper = FALSE;
15361 bool has_terminating_colon = FALSE;
15362 bool has_terminating_bracket = FALSE;
15363 bool has_semi_colon = FALSE;
15364 unsigned int name_len = 0;
15365 int punct_count = 0;
15369 /* Squeeze out blanks when looking up the class name below */
15370 if (isBLANK(*p) ) {
15372 found_problem = TRUE;
15377 /* The name will end with a punctuation */
15379 const char * peek = p + 1;
15381 /* Treat any non-']' punctuation followed by a ']' (possibly
15382 * with intervening blanks) as trying to terminate the class.
15383 * ']]' is very likely to mean a class was intended (but
15384 * missing the colon), but the warning message that gets
15385 * generated shows the error position better if we exit the
15386 * loop at the bottom (eventually), so skip it here. */
15388 if (peek < e && isBLANK(*peek)) {
15390 found_problem = TRUE;
15393 } while (peek < e && isBLANK(*peek));
15396 if (peek < e && *peek == ']') {
15397 has_terminating_bracket = TRUE;
15399 has_terminating_colon = TRUE;
15401 else if (*p == ';') {
15402 has_semi_colon = TRUE;
15403 has_terminating_colon = TRUE;
15406 found_problem = TRUE;
15413 /* Here we have punctuation we thought didn't end the class.
15414 * Keep track of the position of the key characters that are
15415 * more likely to have been class-enders */
15416 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15418 /* Allow just one such possible class-ender not actually
15419 * ending the class. */
15420 if (possible_end) {
15426 /* If we have too many punctuation characters, no use in
15428 if (++punct_count > max_distance) {
15432 /* Treat the punctuation as a typo. */
15433 input_text[name_len++] = *p;
15436 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15437 input_text[name_len++] = toLOWER(*p);
15439 found_problem = TRUE;
15441 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15442 input_text[name_len++] = *p;
15446 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15450 /* The declaration of 'input_text' is how long we allow a potential
15451 * class name to be, before saying they didn't mean a class name at
15453 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15458 /* We get to here when the possible class name hasn't been properly
15459 * terminated before:
15460 * 1) we ran off the end of the pattern; or
15461 * 2) found two characters, each of which might have been intended to
15462 * be the name's terminator
15463 * 3) found so many punctuation characters in the purported name,
15464 * that the edit distance to a valid one is exceeded
15465 * 4) we decided it was more characters than anyone could have
15466 * intended to be one. */
15468 found_problem = TRUE;
15470 /* In the final two cases, we know that looking up what we've
15471 * accumulated won't lead to a match, even a fuzzy one. */
15472 if ( name_len >= C_ARRAY_LENGTH(input_text)
15473 || punct_count > max_distance)
15475 /* If there was an intermediate key character that could have been
15476 * an intended end, redo the parse, but stop there */
15477 if (possible_end && possible_end != (char *) -1) {
15478 possible_end = (char *) -1; /* Special signal value to say
15479 we've done a first pass */
15484 /* Otherwise, it can't have meant to have been a class */
15485 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15488 /* If we ran off the end, and the final character was a punctuation
15489 * one, back up one, to look at that final one just below. Later, we
15490 * will restore the parse pointer if appropriate */
15491 if (name_len && p == e && isPUNCT(*(p-1))) {
15496 if (p < e && isPUNCT(*p)) {
15498 has_terminating_bracket = TRUE;
15500 /* If this is a 2nd ']', and the first one is just below this
15501 * one, consider that to be the real terminator. This gives a
15502 * uniform and better positioning for the warning message */
15504 && possible_end != (char *) -1
15505 && *possible_end == ']'
15506 && name_len && input_text[name_len - 1] == ']')
15511 /* And this is actually equivalent to having done the 2nd
15512 * pass now, so set it to not try again */
15513 possible_end = (char *) -1;
15518 has_terminating_colon = TRUE;
15520 else if (*p == ';') {
15521 has_semi_colon = TRUE;
15522 has_terminating_colon = TRUE;
15530 /* Here, we have a class name to look up. We can short circuit the
15531 * stuff below for short names that can't possibly be meant to be a
15532 * class name. (We can do this on the first pass, as any second pass
15533 * will yield an even shorter name) */
15534 if (name_len < 3) {
15535 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15538 /* Find which class it is. Initially switch on the length of the name.
15540 switch (name_len) {
15542 if (memEQs(name_start, 4, "word")) {
15543 /* this is not POSIX, this is the Perl \w */
15544 class_number = ANYOF_WORDCHAR;
15548 /* Names all of length 5: alnum alpha ascii blank cntrl digit
15549 * graph lower print punct space upper
15550 * Offset 4 gives the best switch position. */
15551 switch (name_start[4]) {
15553 if (memBEGINs(name_start, 5, "alph")) /* alpha */
15554 class_number = ANYOF_ALPHA;
15557 if (memBEGINs(name_start, 5, "spac")) /* space */
15558 class_number = ANYOF_SPACE;
15561 if (memBEGINs(name_start, 5, "grap")) /* graph */
15562 class_number = ANYOF_GRAPH;
15565 if (memBEGINs(name_start, 5, "asci")) /* ascii */
15566 class_number = ANYOF_ASCII;
15569 if (memBEGINs(name_start, 5, "blan")) /* blank */
15570 class_number = ANYOF_BLANK;
15573 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
15574 class_number = ANYOF_CNTRL;
15577 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
15578 class_number = ANYOF_ALPHANUMERIC;
15581 if (memBEGINs(name_start, 5, "lowe")) /* lower */
15582 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
15583 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
15584 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
15587 if (memBEGINs(name_start, 5, "digi")) /* digit */
15588 class_number = ANYOF_DIGIT;
15589 else if (memBEGINs(name_start, 5, "prin")) /* print */
15590 class_number = ANYOF_PRINT;
15591 else if (memBEGINs(name_start, 5, "punc")) /* punct */
15592 class_number = ANYOF_PUNCT;
15597 if (memEQs(name_start, 6, "xdigit"))
15598 class_number = ANYOF_XDIGIT;
15602 /* If the name exactly matches a posix class name the class number will
15603 * here be set to it, and the input almost certainly was meant to be a
15604 * posix class, so we can skip further checking. If instead the syntax
15605 * is exactly correct, but the name isn't one of the legal ones, we
15606 * will return that as an error below. But if neither of these apply,
15607 * it could be that no posix class was intended at all, or that one
15608 * was, but there was a typo. We tease these apart by doing fuzzy
15609 * matching on the name */
15610 if (class_number == OOB_NAMEDCLASS && found_problem) {
15611 const UV posix_names[][6] = {
15612 { 'a', 'l', 'n', 'u', 'm' },
15613 { 'a', 'l', 'p', 'h', 'a' },
15614 { 'a', 's', 'c', 'i', 'i' },
15615 { 'b', 'l', 'a', 'n', 'k' },
15616 { 'c', 'n', 't', 'r', 'l' },
15617 { 'd', 'i', 'g', 'i', 't' },
15618 { 'g', 'r', 'a', 'p', 'h' },
15619 { 'l', 'o', 'w', 'e', 'r' },
15620 { 'p', 'r', 'i', 'n', 't' },
15621 { 'p', 'u', 'n', 'c', 't' },
15622 { 's', 'p', 'a', 'c', 'e' },
15623 { 'u', 'p', 'p', 'e', 'r' },
15624 { 'w', 'o', 'r', 'd' },
15625 { 'x', 'd', 'i', 'g', 'i', 't' }
15627 /* The names of the above all have added NULs to make them the same
15628 * size, so we need to also have the real lengths */
15629 const UV posix_name_lengths[] = {
15630 sizeof("alnum") - 1,
15631 sizeof("alpha") - 1,
15632 sizeof("ascii") - 1,
15633 sizeof("blank") - 1,
15634 sizeof("cntrl") - 1,
15635 sizeof("digit") - 1,
15636 sizeof("graph") - 1,
15637 sizeof("lower") - 1,
15638 sizeof("print") - 1,
15639 sizeof("punct") - 1,
15640 sizeof("space") - 1,
15641 sizeof("upper") - 1,
15642 sizeof("word") - 1,
15643 sizeof("xdigit")- 1
15646 int temp_max = max_distance; /* Use a temporary, so if we
15647 reparse, we haven't changed the
15650 /* Use a smaller max edit distance if we are missing one of the
15652 if ( has_opening_bracket + has_opening_colon < 2
15653 || has_terminating_bracket + has_terminating_colon < 2)
15658 /* See if the input name is close to a legal one */
15659 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
15661 /* Short circuit call if the lengths are too far apart to be
15663 if (abs( (int) (name_len - posix_name_lengths[i]))
15669 if (edit_distance(input_text,
15672 posix_name_lengths[i],
15676 { /* If it is close, it probably was intended to be a class */
15677 goto probably_meant_to_be;
15681 /* Here the input name is not close enough to a valid class name
15682 * for us to consider it to be intended to be a posix class. If
15683 * we haven't already done so, and the parse found a character that
15684 * could have been terminators for the name, but which we absorbed
15685 * as typos during the first pass, repeat the parse, signalling it
15686 * to stop at that character */
15687 if (possible_end && possible_end != (char *) -1) {
15688 possible_end = (char *) -1;
15693 /* Here neither pass found a close-enough class name */
15694 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15697 probably_meant_to_be:
15699 /* Here we think that a posix specification was intended. Update any
15701 if (updated_parse_ptr) {
15702 *updated_parse_ptr = (char *) p;
15705 /* If a posix class name was intended but incorrectly specified, we
15706 * output or return the warnings */
15707 if (found_problem) {
15709 /* We set flags for these issues in the parse loop above instead of
15710 * adding them to the list of warnings, because we can parse it
15711 * twice, and we only want one warning instance */
15713 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
15716 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15718 if (has_semi_colon) {
15719 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15721 else if (! has_terminating_colon) {
15722 ADD_POSIX_WARNING(p, "there is no terminating ':'");
15724 if (! has_terminating_bracket) {
15725 ADD_POSIX_WARNING(p, "there is no terminating ']'");
15728 if ( posix_warnings
15730 && av_top_index(RExC_warn_text) > -1)
15732 *posix_warnings = RExC_warn_text;
15735 else if (class_number != OOB_NAMEDCLASS) {
15736 /* If it is a known class, return the class. The class number
15737 * #defines are structured so each complement is +1 to the normal
15739 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
15741 else if (! check_only) {
15743 /* Here, it is an unrecognized class. This is an error (unless the
15744 * call is to check only, which we've already handled above) */
15745 const char * const complement_string = (complement)
15748 RExC_parse = (char *) p;
15749 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
15751 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
15755 return OOB_NAMEDCLASS;
15757 #undef ADD_POSIX_WARNING
15759 STATIC unsigned int
15760 S_regex_set_precedence(const U8 my_operator) {
15762 /* Returns the precedence in the (?[...]) construct of the input operator,
15763 * specified by its character representation. The precedence follows
15764 * general Perl rules, but it extends this so that ')' and ']' have (low)
15765 * precedence even though they aren't really operators */
15767 switch (my_operator) {
15783 NOT_REACHED; /* NOTREACHED */
15784 return 0; /* Silence compiler warning */
15787 STATIC regnode_offset
15788 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
15789 I32 *flagp, U32 depth,
15790 char * const oregcomp_parse)
15792 /* Handle the (?[...]) construct to do set operations */
15794 U8 curchar; /* Current character being parsed */
15795 UV start, end; /* End points of code point ranges */
15796 SV* final = NULL; /* The end result inversion list */
15797 SV* result_string; /* 'final' stringified */
15798 AV* stack; /* stack of operators and operands not yet
15800 AV* fence_stack = NULL; /* A stack containing the positions in
15801 'stack' of where the undealt-with left
15802 parens would be if they were actually
15804 /* The 'volatile' is a workaround for an optimiser bug
15805 * in Solaris Studio 12.3. See RT #127455 */
15806 volatile IV fence = 0; /* Position of where most recent undealt-
15807 with left paren in stack is; -1 if none.
15809 STRLEN len; /* Temporary */
15810 regnode_offset node; /* Temporary, and final regnode returned by
15812 const bool save_fold = FOLD; /* Temporary */
15813 char *save_end, *save_parse; /* Temporaries */
15814 const bool in_locale = LOC; /* we turn off /l during processing */
15816 GET_RE_DEBUG_FLAGS_DECL;
15818 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
15820 DEBUG_PARSE("xcls");
15823 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
15826 /* The use of this operator implies /u. This is required so that the
15827 * compile time values are valid in all runtime cases */
15828 REQUIRE_UNI_RULES(flagp, 0);
15830 ckWARNexperimental(RExC_parse,
15831 WARN_EXPERIMENTAL__REGEX_SETS,
15832 "The regex_sets feature is experimental");
15834 /* Everything in this construct is a metacharacter. Operands begin with
15835 * either a '\' (for an escape sequence), or a '[' for a bracketed
15836 * character class. Any other character should be an operator, or
15837 * parenthesis for grouping. Both types of operands are handled by calling
15838 * regclass() to parse them. It is called with a parameter to indicate to
15839 * return the computed inversion list. The parsing here is implemented via
15840 * a stack. Each entry on the stack is a single character representing one
15841 * of the operators; or else a pointer to an operand inversion list. */
15843 #define IS_OPERATOR(a) SvIOK(a)
15844 #define IS_OPERAND(a) (! IS_OPERATOR(a))
15846 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
15847 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
15848 * with pronouncing it called it Reverse Polish instead, but now that YOU
15849 * know how to pronounce it you can use the correct term, thus giving due
15850 * credit to the person who invented it, and impressing your geek friends.
15851 * Wikipedia says that the pronounciation of "Ł" has been changing so that
15852 * it is now more like an English initial W (as in wonk) than an L.)
15854 * This means that, for example, 'a | b & c' is stored on the stack as
15862 * where the numbers in brackets give the stack [array] element number.
15863 * In this implementation, parentheses are not stored on the stack.
15864 * Instead a '(' creates a "fence" so that the part of the stack below the
15865 * fence is invisible except to the corresponding ')' (this allows us to
15866 * replace testing for parens, by using instead subtraction of the fence
15867 * position). As new operands are processed they are pushed onto the stack
15868 * (except as noted in the next paragraph). New operators of higher
15869 * precedence than the current final one are inserted on the stack before
15870 * the lhs operand (so that when the rhs is pushed next, everything will be
15871 * in the correct positions shown above. When an operator of equal or
15872 * lower precedence is encountered in parsing, all the stacked operations
15873 * of equal or higher precedence are evaluated, leaving the result as the
15874 * top entry on the stack. This makes higher precedence operations
15875 * evaluate before lower precedence ones, and causes operations of equal
15876 * precedence to left associate.
15878 * The only unary operator '!' is immediately pushed onto the stack when
15879 * encountered. When an operand is encountered, if the top of the stack is
15880 * a '!", the complement is immediately performed, and the '!' popped. The
15881 * resulting value is treated as a new operand, and the logic in the
15882 * previous paragraph is executed. Thus in the expression
15884 * the stack looks like
15890 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
15897 * A ')' is treated as an operator with lower precedence than all the
15898 * aforementioned ones, which causes all operations on the stack above the
15899 * corresponding '(' to be evaluated down to a single resultant operand.
15900 * Then the fence for the '(' is removed, and the operand goes through the
15901 * algorithm above, without the fence.
15903 * A separate stack is kept of the fence positions, so that the position of
15904 * the latest so-far unbalanced '(' is at the top of it.
15906 * The ']' ending the construct is treated as the lowest operator of all,
15907 * so that everything gets evaluated down to a single operand, which is the
15910 sv_2mortal((SV *)(stack = newAV()));
15911 sv_2mortal((SV *)(fence_stack = newAV()));
15913 while (RExC_parse < RExC_end) {
15914 I32 top_index; /* Index of top-most element in 'stack' */
15915 SV** top_ptr; /* Pointer to top 'stack' element */
15916 SV* current = NULL; /* To contain the current inversion list
15918 SV* only_to_avoid_leaks;
15920 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15921 TRUE /* Force /x */ );
15922 if (RExC_parse >= RExC_end) { /* Fail */
15926 curchar = UCHARAT(RExC_parse);
15930 #ifdef ENABLE_REGEX_SETS_DEBUGGING
15931 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
15932 DEBUG_U(dump_regex_sets_structures(pRExC_state,
15933 stack, fence, fence_stack));
15936 top_index = av_tindex_skip_len_mg(stack);
15939 SV** stacked_ptr; /* Ptr to something already on 'stack' */
15940 char stacked_operator; /* The topmost operator on the 'stack'. */
15941 SV* lhs; /* Operand to the left of the operator */
15942 SV* rhs; /* Operand to the right of the operator */
15943 SV* fence_ptr; /* Pointer to top element of the fence
15948 if ( RExC_parse < RExC_end - 2
15949 && UCHARAT(RExC_parse + 1) == '?'
15950 && UCHARAT(RExC_parse + 2) == '^')
15952 /* If is a '(?', could be an embedded '(?^flags:(?[...])'.
15953 * This happens when we have some thing like
15955 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
15957 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
15959 * Here we would be handling the interpolated
15960 * '$thai_or_lao'. We handle this by a recursive call to
15961 * ourselves which returns the inversion list the
15962 * interpolated expression evaluates to. We use the flags
15963 * from the interpolated pattern. */
15964 U32 save_flags = RExC_flags;
15965 const char * save_parse;
15967 RExC_parse += 2; /* Skip past the '(?' */
15968 save_parse = RExC_parse;
15970 /* Parse the flags for the '(?'. We already know the first
15971 * flag to parse is a '^' */
15972 parse_lparen_question_flags(pRExC_state);
15974 if ( RExC_parse >= RExC_end - 4
15975 || UCHARAT(RExC_parse) != ':'
15976 || UCHARAT(++RExC_parse) != '('
15977 || UCHARAT(++RExC_parse) != '?'
15978 || UCHARAT(++RExC_parse) != '[')
15981 /* In combination with the above, this moves the
15982 * pointer to the point just after the first erroneous
15984 if (RExC_parse >= RExC_end - 4) {
15985 RExC_parse = RExC_end;
15987 else if (RExC_parse != save_parse) {
15988 RExC_parse += (UTF)
15989 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
15992 vFAIL("Expecting '(?flags:(?[...'");
15995 /* Recurse, with the meat of the embedded expression */
15997 if (! handle_regex_sets(pRExC_state, ¤t, flagp,
15998 depth+1, oregcomp_parse))
16000 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16003 /* Here, 'current' contains the embedded expression's
16004 * inversion list, and RExC_parse points to the trailing
16005 * ']'; the next character should be the ')' */
16007 if (UCHARAT(RExC_parse) != ')')
16008 vFAIL("Expecting close paren for nested extended charclass");
16010 /* Then the ')' matching the original '(' handled by this
16011 * case: statement */
16013 if (UCHARAT(RExC_parse) != ')')
16014 vFAIL("Expecting close paren for wrapper for nested extended charclass");
16016 RExC_flags = save_flags;
16017 goto handle_operand;
16020 /* A regular '('. Look behind for illegal syntax */
16021 if (top_index - fence >= 0) {
16022 /* If the top entry on the stack is an operator, it had
16023 * better be a '!', otherwise the entry below the top
16024 * operand should be an operator */
16025 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16026 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16027 || ( IS_OPERAND(*top_ptr)
16028 && ( top_index - fence < 1
16029 || ! (stacked_ptr = av_fetch(stack,
16032 || ! IS_OPERATOR(*stacked_ptr))))
16035 vFAIL("Unexpected '(' with no preceding operator");
16039 /* Stack the position of this undealt-with left paren */
16040 av_push(fence_stack, newSViv(fence));
16041 fence = top_index + 1;
16045 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16046 * multi-char folds are allowed. */
16047 if (!regclass(pRExC_state, flagp, depth+1,
16048 TRUE, /* means parse just the next thing */
16049 FALSE, /* don't allow multi-char folds */
16050 FALSE, /* don't silence non-portable warnings. */
16052 FALSE, /* Require return to be an ANYOF */
16055 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16056 goto regclass_failed;
16059 /* regclass() will return with parsing just the \ sequence,
16060 * leaving the parse pointer at the next thing to parse */
16062 goto handle_operand;
16064 case '[': /* Is a bracketed character class */
16066 /* See if this is a [:posix:] class. */
16067 bool is_posix_class = (OOB_NAMEDCLASS
16068 < handle_possible_posix(pRExC_state,
16072 TRUE /* checking only */));
16073 /* If it is a posix class, leave the parse pointer at the '['
16074 * to fool regclass() into thinking it is part of a
16075 * '[[:posix:]]'. */
16076 if (! is_posix_class) {
16080 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16081 * multi-char folds are allowed. */
16082 if (!regclass(pRExC_state, flagp, depth+1,
16083 is_posix_class, /* parse the whole char
16084 class only if not a
16086 FALSE, /* don't allow multi-char folds */
16087 TRUE, /* silence non-portable warnings. */
16089 FALSE, /* Require return to be an ANYOF */
16092 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16093 goto regclass_failed;
16100 /* function call leaves parse pointing to the ']', except if we
16102 if (is_posix_class) {
16106 goto handle_operand;
16110 if (top_index >= 1) {
16111 goto join_operators;
16114 /* Only a single operand on the stack: are done */
16118 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16119 if (UCHARAT(RExC_parse - 1) == ']') {
16123 vFAIL("Unexpected ')'");
16126 /* If nothing after the fence, is missing an operand */
16127 if (top_index - fence < 0) {
16131 /* If at least two things on the stack, treat this as an
16133 if (top_index - fence >= 1) {
16134 goto join_operators;
16137 /* Here only a single thing on the fenced stack, and there is a
16138 * fence. Get rid of it */
16139 fence_ptr = av_pop(fence_stack);
16141 fence = SvIV(fence_ptr);
16142 SvREFCNT_dec_NN(fence_ptr);
16149 /* Having gotten rid of the fence, we pop the operand at the
16150 * stack top and process it as a newly encountered operand */
16151 current = av_pop(stack);
16152 if (IS_OPERAND(current)) {
16153 goto handle_operand;
16165 /* These binary operators should have a left operand already
16167 if ( top_index - fence < 0
16168 || top_index - fence == 1
16169 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16170 || ! IS_OPERAND(*top_ptr))
16172 goto unexpected_binary;
16175 /* If only the one operand is on the part of the stack visible
16176 * to us, we just place this operator in the proper position */
16177 if (top_index - fence < 2) {
16179 /* Place the operator before the operand */
16181 SV* lhs = av_pop(stack);
16182 av_push(stack, newSVuv(curchar));
16183 av_push(stack, lhs);
16187 /* But if there is something else on the stack, we need to
16188 * process it before this new operator if and only if the
16189 * stacked operation has equal or higher precedence than the
16194 /* The operator on the stack is supposed to be below both its
16196 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16197 || IS_OPERAND(*stacked_ptr))
16199 /* But if not, it's legal and indicates we are completely
16200 * done if and only if we're currently processing a ']',
16201 * which should be the final thing in the expression */
16202 if (curchar == ']') {
16208 vFAIL2("Unexpected binary operator '%c' with no "
16209 "preceding operand", curchar);
16211 stacked_operator = (char) SvUV(*stacked_ptr);
16213 if (regex_set_precedence(curchar)
16214 > regex_set_precedence(stacked_operator))
16216 /* Here, the new operator has higher precedence than the
16217 * stacked one. This means we need to add the new one to
16218 * the stack to await its rhs operand (and maybe more
16219 * stuff). We put it before the lhs operand, leaving
16220 * untouched the stacked operator and everything below it
16222 lhs = av_pop(stack);
16223 assert(IS_OPERAND(lhs));
16225 av_push(stack, newSVuv(curchar));
16226 av_push(stack, lhs);
16230 /* Here, the new operator has equal or lower precedence than
16231 * what's already there. This means the operation already
16232 * there should be performed now, before the new one. */
16234 rhs = av_pop(stack);
16235 if (! IS_OPERAND(rhs)) {
16237 /* This can happen when a ! is not followed by an operand,
16238 * like in /(?[\t &!])/ */
16242 lhs = av_pop(stack);
16244 if (! IS_OPERAND(lhs)) {
16246 /* This can happen when there is an empty (), like in
16247 * /(?[[0]+()+])/ */
16251 switch (stacked_operator) {
16253 _invlist_intersection(lhs, rhs, &rhs);
16258 _invlist_union(lhs, rhs, &rhs);
16262 _invlist_subtract(lhs, rhs, &rhs);
16265 case '^': /* The union minus the intersection */
16270 _invlist_union(lhs, rhs, &u);
16271 _invlist_intersection(lhs, rhs, &i);
16272 _invlist_subtract(u, i, &rhs);
16273 SvREFCNT_dec_NN(i);
16274 SvREFCNT_dec_NN(u);
16280 /* Here, the higher precedence operation has been done, and the
16281 * result is in 'rhs'. We overwrite the stacked operator with
16282 * the result. Then we redo this code to either push the new
16283 * operator onto the stack or perform any higher precedence
16284 * stacked operation */
16285 only_to_avoid_leaks = av_pop(stack);
16286 SvREFCNT_dec(only_to_avoid_leaks);
16287 av_push(stack, rhs);
16290 case '!': /* Highest priority, right associative */
16292 /* If what's already at the top of the stack is another '!",
16293 * they just cancel each other out */
16294 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16295 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16297 only_to_avoid_leaks = av_pop(stack);
16298 SvREFCNT_dec(only_to_avoid_leaks);
16300 else { /* Otherwise, since it's right associative, just push
16302 av_push(stack, newSVuv(curchar));
16307 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16308 if (RExC_parse >= RExC_end) {
16311 vFAIL("Unexpected character");
16315 /* Here 'current' is the operand. If something is already on the
16316 * stack, we have to check if it is a !. But first, the code above
16317 * may have altered the stack in the time since we earlier set
16320 top_index = av_tindex_skip_len_mg(stack);
16321 if (top_index - fence >= 0) {
16322 /* If the top entry on the stack is an operator, it had better
16323 * be a '!', otherwise the entry below the top operand should
16324 * be an operator */
16325 top_ptr = av_fetch(stack, top_index, FALSE);
16327 if (IS_OPERATOR(*top_ptr)) {
16329 /* The only permissible operator at the top of the stack is
16330 * '!', which is applied immediately to this operand. */
16331 curchar = (char) SvUV(*top_ptr);
16332 if (curchar != '!') {
16333 SvREFCNT_dec(current);
16334 vFAIL2("Unexpected binary operator '%c' with no "
16335 "preceding operand", curchar);
16338 _invlist_invert(current);
16340 only_to_avoid_leaks = av_pop(stack);
16341 SvREFCNT_dec(only_to_avoid_leaks);
16343 /* And we redo with the inverted operand. This allows
16344 * handling multiple ! in a row */
16345 goto handle_operand;
16347 /* Single operand is ok only for the non-binary ')'
16349 else if ((top_index - fence == 0 && curchar != ')')
16350 || (top_index - fence > 0
16351 && (! (stacked_ptr = av_fetch(stack,
16354 || IS_OPERAND(*stacked_ptr))))
16356 SvREFCNT_dec(current);
16357 vFAIL("Operand with no preceding operator");
16361 /* Here there was nothing on the stack or the top element was
16362 * another operand. Just add this new one */
16363 av_push(stack, current);
16365 } /* End of switch on next parse token */
16367 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16368 } /* End of loop parsing through the construct */
16370 vFAIL("Syntax error in (?[...])");
16374 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16375 if (RExC_parse < RExC_end) {
16379 vFAIL("Unexpected ']' with no following ')' in (?[...");
16382 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16383 vFAIL("Unmatched (");
16386 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16387 || ((final = av_pop(stack)) == NULL)
16388 || ! IS_OPERAND(final)
16389 || ! is_invlist(final)
16390 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16393 SvREFCNT_dec(final);
16394 vFAIL("Incomplete expression within '(?[ ])'");
16397 /* Here, 'final' is the resultant inversion list from evaluating the
16398 * expression. Return it if so requested */
16399 if (return_invlist) {
16400 *return_invlist = final;
16404 /* Otherwise generate a resultant node, based on 'final'. regclass() is
16405 * expecting a string of ranges and individual code points */
16406 invlist_iterinit(final);
16407 result_string = newSVpvs("");
16408 while (invlist_iternext(final, &start, &end)) {
16409 if (start == end) {
16410 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16413 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
16418 /* About to generate an ANYOF (or similar) node from the inversion list we
16419 * have calculated */
16420 save_parse = RExC_parse;
16421 RExC_parse = SvPV(result_string, len);
16422 save_end = RExC_end;
16423 RExC_end = RExC_parse + len;
16424 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16426 /* We turn off folding around the call, as the class we have constructed
16427 * already has all folding taken into consideration, and we don't want
16428 * regclass() to add to that */
16429 RExC_flags &= ~RXf_PMf_FOLD;
16430 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16431 * folds are allowed. */
16432 node = regclass(pRExC_state, flagp, depth+1,
16433 FALSE, /* means parse the whole char class */
16434 FALSE, /* don't allow multi-char folds */
16435 TRUE, /* silence non-portable warnings. The above may very
16436 well have generated non-portable code points, but
16437 they're valid on this machine */
16438 FALSE, /* similarly, no need for strict */
16439 FALSE, /* Require return to be an ANYOF */
16444 RExC_parse = save_parse + 1;
16445 RExC_end = save_end;
16446 SvREFCNT_dec_NN(final);
16447 SvREFCNT_dec_NN(result_string);
16450 RExC_flags |= RXf_PMf_FOLD;
16454 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16455 goto regclass_failed;
16458 /* Fix up the node type if we are in locale. (We have pretended we are
16459 * under /u for the purposes of regclass(), as this construct will only
16460 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
16461 * as to cause any warnings about bad locales to be output in regexec.c),
16462 * and add the flag that indicates to check if not in a UTF-8 locale. The
16463 * reason we above forbid optimization into something other than an ANYOF
16464 * node is simply to minimize the number of code changes in regexec.c.
16465 * Otherwise we would have to create new EXACTish node types and deal with
16466 * them. This decision could be revisited should this construct become
16469 * (One might think we could look at the resulting ANYOF node and suppress
16470 * the flag if everything is above 255, as those would be UTF-8 only,
16471 * but this isn't true, as the components that led to that result could
16472 * have been locale-affected, and just happen to cancel each other out
16473 * under UTF-8 locales.) */
16475 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16477 assert(OP(REGNODE_p(node)) == ANYOF);
16479 OP(REGNODE_p(node)) = ANYOFL;
16480 ANYOF_FLAGS(REGNODE_p(node))
16481 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16484 nextchar(pRExC_state);
16485 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
16489 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
16493 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16496 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16497 AV * stack, const IV fence, AV * fence_stack)
16498 { /* Dumps the stacks in handle_regex_sets() */
16500 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16501 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16504 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16506 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16508 if (stack_top < 0) {
16509 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16512 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16513 for (i = stack_top; i >= 0; i--) {
16514 SV ** element_ptr = av_fetch(stack, i, FALSE);
16515 if (! element_ptr) {
16518 if (IS_OPERATOR(*element_ptr)) {
16519 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16520 (int) i, (int) SvIV(*element_ptr));
16523 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16524 sv_dump(*element_ptr);
16529 if (fence_stack_top < 0) {
16530 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16533 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16534 for (i = fence_stack_top; i >= 0; i--) {
16535 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16536 if (! element_ptr) {
16539 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16540 (int) i, (int) SvIV(*element_ptr));
16551 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16553 /* This adds the Latin1/above-Latin1 folding rules.
16555 * This should be called only for a Latin1-range code points, cp, which is
16556 * known to be involved in a simple fold with other code points above
16557 * Latin1. It would give false results if /aa has been specified.
16558 * Multi-char folds are outside the scope of this, and must be handled
16561 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
16563 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
16565 /* The rules that are valid for all Unicode versions are hard-coded in */
16570 add_cp_to_invlist(*invlist, KELVIN_SIGN);
16574 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
16577 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
16578 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
16580 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
16581 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
16582 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
16584 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
16585 *invlist = add_cp_to_invlist(*invlist,
16586 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
16589 default: /* Other code points are checked against the data for the
16590 current Unicode version */
16592 Size_t folds_count;
16593 unsigned int first_fold;
16594 const unsigned int * remaining_folds;
16598 folded_cp = toFOLD(cp);
16601 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
16603 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
16606 if (folded_cp > 255) {
16607 *invlist = add_cp_to_invlist(*invlist, folded_cp);
16610 folds_count = _inverse_folds(folded_cp, &first_fold,
16612 if (folds_count == 0) {
16614 /* Use deprecated warning to increase the chances of this being
16616 ckWARN2reg_d(RExC_parse,
16617 "Perl folding rules are not up-to-date for 0x%02X;"
16618 " please use the perlbug utility to report;", cp);
16623 if (first_fold > 255) {
16624 *invlist = add_cp_to_invlist(*invlist, first_fold);
16626 for (i = 0; i < folds_count - 1; i++) {
16627 if (remaining_folds[i] > 255) {
16628 *invlist = add_cp_to_invlist(*invlist,
16629 remaining_folds[i]);
16639 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
16641 /* Output the elements of the array given by '*posix_warnings' as REGEXP
16645 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
16647 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
16649 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
16653 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
16654 if (first_is_fatal) { /* Avoid leaking this */
16655 av_undef(posix_warnings); /* This isn't necessary if the
16656 array is mortal, but is a
16658 (void) sv_2mortal(msg);
16661 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
16662 SvREFCNT_dec_NN(msg);
16665 UPDATE_WARNINGS_LOC(RExC_parse);
16669 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
16671 /* This adds the string scalar <multi_string> to the array
16672 * <multi_char_matches>. <multi_string> is known to have exactly
16673 * <cp_count> code points in it. This is used when constructing a
16674 * bracketed character class and we find something that needs to match more
16675 * than a single character.
16677 * <multi_char_matches> is actually an array of arrays. Each top-level
16678 * element is an array that contains all the strings known so far that are
16679 * the same length. And that length (in number of code points) is the same
16680 * as the index of the top-level array. Hence, the [2] element is an
16681 * array, each element thereof is a string containing TWO code points;
16682 * while element [3] is for strings of THREE characters, and so on. Since
16683 * this is for multi-char strings there can never be a [0] nor [1] element.
16685 * When we rewrite the character class below, we will do so such that the
16686 * longest strings are written first, so that it prefers the longest
16687 * matching strings first. This is done even if it turns out that any
16688 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
16689 * Christiansen has agreed that this is ok. This makes the test for the
16690 * ligature 'ffi' come before the test for 'ff', for example */
16693 AV** this_array_ptr;
16695 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
16697 if (! multi_char_matches) {
16698 multi_char_matches = newAV();
16701 if (av_exists(multi_char_matches, cp_count)) {
16702 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
16703 this_array = *this_array_ptr;
16706 this_array = newAV();
16707 av_store(multi_char_matches, cp_count,
16710 av_push(this_array, multi_string);
16712 return multi_char_matches;
16715 /* The names of properties whose definitions are not known at compile time are
16716 * stored in this SV, after a constant heading. So if the length has been
16717 * changed since initialization, then there is a run-time definition. */
16718 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
16719 (SvCUR(listsv) != initial_listsv_len)
16721 /* There is a restricted set of white space characters that are legal when
16722 * ignoring white space in a bracketed character class. This generates the
16723 * code to skip them.
16725 * There is a line below that uses the same white space criteria but is outside
16726 * this macro. Both here and there must use the same definition */
16727 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
16730 while (isBLANK_A(UCHARAT(p))) \
16737 STATIC regnode_offset
16738 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
16739 const bool stop_at_1, /* Just parse the next thing, don't
16740 look for a full character class */
16741 bool allow_mutiple_chars,
16742 const bool silence_non_portable, /* Don't output warnings
16746 bool optimizable, /* ? Allow a non-ANYOF return
16748 SV** ret_invlist /* Return an inversion list, not a node */
16751 /* parse a bracketed class specification. Most of these will produce an
16752 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
16753 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
16754 * under /i with multi-character folds: it will be rewritten following the
16755 * paradigm of this example, where the <multi-fold>s are characters which
16756 * fold to multiple character sequences:
16757 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
16758 * gets effectively rewritten as:
16759 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
16760 * reg() gets called (recursively) on the rewritten version, and this
16761 * function will return what it constructs. (Actually the <multi-fold>s
16762 * aren't physically removed from the [abcdefghi], it's just that they are
16763 * ignored in the recursion by means of a flag:
16764 * <RExC_in_multi_char_class>.)
16766 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
16767 * characters, with the corresponding bit set if that character is in the
16768 * list. For characters above this, an inversion list is used. There
16769 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
16770 * determinable at compile time
16772 * On success, returns the offset at which any next node should be placed
16773 * into the regex engine program being compiled.
16775 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
16776 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
16781 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
16783 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
16784 regnode_offset ret = -1; /* Initialized to an illegal value */
16786 int namedclass = OOB_NAMEDCLASS;
16787 char *rangebegin = NULL;
16788 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
16789 aren't available at the time this was called */
16790 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
16791 than just initialized. */
16792 SV* properties = NULL; /* Code points that match \p{} \P{} */
16793 SV* posixes = NULL; /* Code points that match classes like [:word:],
16794 extended beyond the Latin1 range. These have to
16795 be kept separate from other code points for much
16796 of this function because their handling is
16797 different under /i, and for most classes under
16799 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
16800 separate for a while from the non-complemented
16801 versions because of complications with /d
16803 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
16804 treated more simply than the general case,
16805 leading to less compilation and execution
16807 UV element_count = 0; /* Number of distinct elements in the class.
16808 Optimizations may be possible if this is tiny */
16809 AV * multi_char_matches = NULL; /* Code points that fold to more than one
16810 character; used under /i */
16812 char * stop_ptr = RExC_end; /* where to stop parsing */
16814 /* ignore unescaped whitespace? */
16815 const bool skip_white = cBOOL( ret_invlist
16816 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
16818 /* inversion list of code points this node matches only when the target
16819 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
16821 SV* upper_latin1_only_utf8_matches = NULL;
16823 /* Inversion list of code points this node matches regardless of things
16824 * like locale, folding, utf8ness of the target string */
16825 SV* cp_list = NULL;
16827 /* Like cp_list, but code points on this list need to be checked for things
16828 * that fold to/from them under /i */
16829 SV* cp_foldable_list = NULL;
16831 /* Like cp_list, but code points on this list are valid only when the
16832 * runtime locale is UTF-8 */
16833 SV* only_utf8_locale_list = NULL;
16835 /* In a range, if one of the endpoints is non-character-set portable,
16836 * meaning that it hard-codes a code point that may mean a different
16837 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
16838 * mnemonic '\t' which each mean the same character no matter which
16839 * character set the platform is on. */
16840 unsigned int non_portable_endpoint = 0;
16842 /* Is the range unicode? which means on a platform that isn't 1-1 native
16843 * to Unicode (i.e. non-ASCII), each code point in it should be considered
16844 * to be a Unicode value. */
16845 bool unicode_range = FALSE;
16846 bool invert = FALSE; /* Is this class to be complemented */
16848 bool warn_super = ALWAYS_WARN_SUPER;
16850 const char * orig_parse = RExC_parse;
16852 /* This variable is used to mark where the end in the input is of something
16853 * that looks like a POSIX construct but isn't. During the parse, when
16854 * something looks like it could be such a construct is encountered, it is
16855 * checked for being one, but not if we've already checked this area of the
16856 * input. Only after this position is reached do we check again */
16857 char *not_posix_region_end = RExC_parse - 1;
16859 AV* posix_warnings = NULL;
16860 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
16861 U8 op = END; /* The returned node-type, initialized to an impossible
16863 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
16864 U32 posixl = 0; /* bit field of posix classes matched under /l */
16867 /* Flags as to what things aren't knowable until runtime. (Note that these are
16868 * mutually exclusive.) */
16869 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
16870 haven't been defined as of yet */
16871 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
16873 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
16874 what gets folded */
16875 U32 has_runtime_dependency = 0; /* OR of the above flags */
16877 GET_RE_DEBUG_FLAGS_DECL;
16879 PERL_ARGS_ASSERT_REGCLASS;
16881 PERL_UNUSED_ARG(depth);
16885 /* If wants an inversion list returned, we can't optimize to something
16888 optimizable = FALSE;
16891 DEBUG_PARSE("clas");
16893 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
16894 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
16895 && UNICODE_DOT_DOT_VERSION == 0)
16896 allow_mutiple_chars = FALSE;
16899 /* We include the /i status at the beginning of this so that we can
16900 * know it at runtime */
16901 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
16902 initial_listsv_len = SvCUR(listsv);
16903 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
16905 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16907 assert(RExC_parse <= RExC_end);
16909 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
16912 allow_mutiple_chars = FALSE;
16914 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16917 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
16918 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
16919 int maybe_class = handle_possible_posix(pRExC_state,
16921 ¬_posix_region_end,
16923 TRUE /* checking only */);
16924 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
16925 ckWARN4reg(not_posix_region_end,
16926 "POSIX syntax [%c %c] belongs inside character classes%s",
16927 *RExC_parse, *RExC_parse,
16928 (maybe_class == OOB_NAMEDCLASS)
16929 ? ((POSIXCC_NOTYET(*RExC_parse))
16930 ? " (but this one isn't implemented)"
16931 : " (but this one isn't fully valid)")
16937 /* If the caller wants us to just parse a single element, accomplish this
16938 * by faking the loop ending condition */
16939 if (stop_at_1 && RExC_end > RExC_parse) {
16940 stop_ptr = RExC_parse + 1;
16943 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
16944 if (UCHARAT(RExC_parse) == ']')
16945 goto charclassloop;
16949 if ( posix_warnings
16950 && av_tindex_skip_len_mg(posix_warnings) >= 0
16951 && RExC_parse > not_posix_region_end)
16953 /* Warnings about posix class issues are considered tentative until
16954 * we are far enough along in the parse that we can no longer
16955 * change our mind, at which point we output them. This is done
16956 * each time through the loop so that a later class won't zap them
16957 * before they have been dealt with. */
16958 output_posix_warnings(pRExC_state, posix_warnings);
16961 if (RExC_parse >= stop_ptr) {
16965 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
16967 if (UCHARAT(RExC_parse) == ']') {
16973 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
16974 save_value = value;
16975 save_prevvalue = prevvalue;
16978 rangebegin = RExC_parse;
16980 non_portable_endpoint = 0;
16982 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
16983 value = utf8n_to_uvchr((U8*)RExC_parse,
16984 RExC_end - RExC_parse,
16985 &numlen, UTF8_ALLOW_DEFAULT);
16986 RExC_parse += numlen;
16989 value = UCHARAT(RExC_parse++);
16991 if (value == '[') {
16992 char * posix_class_end;
16993 namedclass = handle_possible_posix(pRExC_state,
16996 do_posix_warnings ? &posix_warnings : NULL,
16997 FALSE /* die if error */);
16998 if (namedclass > OOB_NAMEDCLASS) {
17000 /* If there was an earlier attempt to parse this particular
17001 * posix class, and it failed, it was a false alarm, as this
17002 * successful one proves */
17003 if ( posix_warnings
17004 && av_tindex_skip_len_mg(posix_warnings) >= 0
17005 && not_posix_region_end >= RExC_parse
17006 && not_posix_region_end <= posix_class_end)
17008 av_undef(posix_warnings);
17011 RExC_parse = posix_class_end;
17013 else if (namedclass == OOB_NAMEDCLASS) {
17014 not_posix_region_end = posix_class_end;
17017 namedclass = OOB_NAMEDCLASS;
17020 else if ( RExC_parse - 1 > not_posix_region_end
17021 && MAYBE_POSIXCC(value))
17023 (void) handle_possible_posix(
17025 RExC_parse - 1, /* -1 because parse has already been
17027 ¬_posix_region_end,
17028 do_posix_warnings ? &posix_warnings : NULL,
17029 TRUE /* checking only */);
17031 else if ( strict && ! skip_white
17032 && ( _generic_isCC(value, _CC_VERTSPACE)
17033 || is_VERTWS_cp_high(value)))
17035 vFAIL("Literal vertical space in [] is illegal except under /x");
17037 else if (value == '\\') {
17038 /* Is a backslash; get the code point of the char after it */
17040 if (RExC_parse >= RExC_end) {
17041 vFAIL("Unmatched [");
17044 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17045 value = utf8n_to_uvchr((U8*)RExC_parse,
17046 RExC_end - RExC_parse,
17047 &numlen, UTF8_ALLOW_DEFAULT);
17048 RExC_parse += numlen;
17051 value = UCHARAT(RExC_parse++);
17053 /* Some compilers cannot handle switching on 64-bit integer
17054 * values, therefore value cannot be an UV. Yes, this will
17055 * be a problem later if we want switch on Unicode.
17056 * A similar issue a little bit later when switching on
17057 * namedclass. --jhi */
17059 /* If the \ is escaping white space when white space is being
17060 * skipped, it means that that white space is wanted literally, and
17061 * is already in 'value'. Otherwise, need to translate the escape
17062 * into what it signifies. */
17063 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17065 case 'w': namedclass = ANYOF_WORDCHAR; break;
17066 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17067 case 's': namedclass = ANYOF_SPACE; break;
17068 case 'S': namedclass = ANYOF_NSPACE; break;
17069 case 'd': namedclass = ANYOF_DIGIT; break;
17070 case 'D': namedclass = ANYOF_NDIGIT; break;
17071 case 'v': namedclass = ANYOF_VERTWS; break;
17072 case 'V': namedclass = ANYOF_NVERTWS; break;
17073 case 'h': namedclass = ANYOF_HORIZWS; break;
17074 case 'H': namedclass = ANYOF_NHORIZWS; break;
17075 case 'N': /* Handle \N{NAME} in class */
17077 const char * const backslash_N_beg = RExC_parse - 2;
17080 if (! grok_bslash_N(pRExC_state,
17081 NULL, /* No regnode */
17082 &value, /* Yes single value */
17083 &cp_count, /* Multiple code pt count */
17089 if (*flagp & NEED_UTF8)
17090 FAIL("panic: grok_bslash_N set NEED_UTF8");
17092 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17094 if (cp_count < 0) {
17095 vFAIL("\\N in a character class must be a named character: \\N{...}");
17097 else if (cp_count == 0) {
17098 ckWARNreg(RExC_parse,
17099 "Ignoring zero length \\N{} in character class");
17101 else { /* cp_count > 1 */
17102 assert(cp_count > 1);
17103 if (! RExC_in_multi_char_class) {
17104 if ( ! allow_mutiple_chars
17107 || *RExC_parse == '-')
17111 vFAIL("\\N{} here is restricted to one character");
17113 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17114 break; /* <value> contains the first code
17115 point. Drop out of the switch to
17119 SV * multi_char_N = newSVpvn(backslash_N_beg,
17120 RExC_parse - backslash_N_beg);
17122 = add_multi_match(multi_char_matches,
17127 } /* End of cp_count != 1 */
17129 /* This element should not be processed further in this
17132 value = save_value;
17133 prevvalue = save_prevvalue;
17134 continue; /* Back to top of loop to get next char */
17137 /* Here, is a single code point, and <value> contains it */
17138 unicode_range = TRUE; /* \N{} are Unicode */
17146 /* \p means they want Unicode semantics */
17147 REQUIRE_UNI_RULES(flagp, 0);
17149 if (RExC_parse >= RExC_end)
17150 vFAIL2("Empty \\%c", (U8)value);
17151 if (*RExC_parse == '{') {
17152 const U8 c = (U8)value;
17153 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17156 vFAIL2("Missing right brace on \\%c{}", c);
17161 /* White space is allowed adjacent to the braces and after
17162 * any '^', even when not under /x */
17163 while (isSPACE(*RExC_parse)) {
17167 if (UCHARAT(RExC_parse) == '^') {
17169 /* toggle. (The rhs xor gets the single bit that
17170 * differs between P and p; the other xor inverts just
17172 value ^= 'P' ^ 'p';
17175 while (isSPACE(*RExC_parse)) {
17180 if (e == RExC_parse)
17181 vFAIL2("Empty \\%c{}", c);
17183 n = e - RExC_parse;
17184 while (isSPACE(*(RExC_parse + n - 1)))
17187 } /* The \p isn't immediately followed by a '{' */
17188 else if (! isALPHA(*RExC_parse)) {
17189 RExC_parse += (UTF)
17190 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17192 vFAIL2("Character following \\%c must be '{' or a "
17193 "single-character Unicode property name",
17201 char* name = RExC_parse;
17203 /* Any message returned about expanding the definition */
17204 SV* msg = newSVpvs_flags("", SVs_TEMP);
17206 /* If set TRUE, the property is user-defined as opposed to
17207 * official Unicode */
17208 bool user_defined = FALSE;
17210 SV * prop_definition = parse_uniprop_string(
17211 name, n, UTF, FOLD,
17212 FALSE, /* This is compile-time */
17214 /* We can't defer this defn when
17215 * the full result is required in
17217 ! cBOOL(ret_invlist),
17223 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17224 assert(prop_definition == NULL);
17225 RExC_parse = e + 1;
17226 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17227 thing so, or else the display is
17231 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17232 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17233 SvCUR(msg), SvPVX(msg)));
17236 if (! is_invlist(prop_definition)) {
17238 /* Here, the definition isn't known, so we have gotten
17239 * returned a string that will be evaluated if and when
17240 * encountered at runtime. We add it to the list of
17241 * such properties, along with whether it should be
17242 * complemented or not */
17243 if (value == 'P') {
17244 sv_catpvs(listsv, "!");
17247 sv_catpvs(listsv, "+");
17249 sv_catsv(listsv, prop_definition);
17251 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17253 /* We don't know yet what this matches, so have to flag
17255 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17258 assert (prop_definition && is_invlist(prop_definition));
17260 /* Here we do have the complete property definition
17262 * Temporary workaround for [perl #133136]. For this
17263 * precise input that is in the .t that is failing,
17264 * load utf8.pm, which is what the test wants, so that
17265 * that .t passes */
17266 if ( memEQs(RExC_start, e + 1 - RExC_start,
17268 && ! hv_common(GvHVn(PL_incgv),
17270 "utf8.pm", sizeof("utf8.pm") - 1,
17271 0, HV_FETCH_ISEXISTS, NULL, 0))
17273 require_pv("utf8.pm");
17276 if (! user_defined &&
17277 /* We warn on matching an above-Unicode code point
17278 * if the match would return true, except don't
17279 * warn for \p{All}, which has exactly one element
17281 (_invlist_contains_cp(prop_definition, 0x110000)
17282 && (! (_invlist_len(prop_definition) == 1
17283 && *invlist_array(prop_definition) == 0))))
17288 /* Invert if asking for the complement */
17289 if (value == 'P') {
17290 _invlist_union_complement_2nd(properties,
17295 _invlist_union(properties, prop_definition, &properties);
17300 RExC_parse = e + 1;
17301 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17305 case 'n': value = '\n'; break;
17306 case 'r': value = '\r'; break;
17307 case 't': value = '\t'; break;
17308 case 'f': value = '\f'; break;
17309 case 'b': value = '\b'; break;
17310 case 'e': value = ESC_NATIVE; break;
17311 case 'a': value = '\a'; break;
17313 RExC_parse--; /* function expects to be pointed at the 'o' */
17315 const char* error_msg;
17316 bool valid = grok_bslash_o(&RExC_parse,
17320 TO_OUTPUT_WARNINGS(RExC_parse),
17322 silence_non_portable,
17327 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17329 non_portable_endpoint++;
17332 RExC_parse--; /* function expects to be pointed at the 'x' */
17334 const char* error_msg;
17335 bool valid = grok_bslash_x(&RExC_parse,
17339 TO_OUTPUT_WARNINGS(RExC_parse),
17341 silence_non_portable,
17346 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17348 non_portable_endpoint++;
17351 value = grok_bslash_c(*RExC_parse, TO_OUTPUT_WARNINGS(RExC_parse));
17352 UPDATE_WARNINGS_LOC(RExC_parse);
17354 non_portable_endpoint++;
17356 case '0': case '1': case '2': case '3': case '4':
17357 case '5': case '6': case '7':
17359 /* Take 1-3 octal digits */
17360 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
17361 numlen = (strict) ? 4 : 3;
17362 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17363 RExC_parse += numlen;
17366 RExC_parse += (UTF)
17367 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17369 vFAIL("Need exactly 3 octal digits");
17371 else if ( numlen < 3 /* like \08, \178 */
17372 && RExC_parse < RExC_end
17373 && isDIGIT(*RExC_parse)
17374 && ckWARN(WARN_REGEXP))
17376 reg_warn_non_literal_string(
17378 form_short_octal_warning(RExC_parse, numlen));
17381 non_portable_endpoint++;
17385 /* Allow \_ to not give an error */
17386 if (isWORDCHAR(value) && value != '_') {
17388 vFAIL2("Unrecognized escape \\%c in character class",
17392 ckWARN2reg(RExC_parse,
17393 "Unrecognized escape \\%c in character class passed through",
17398 } /* End of switch on char following backslash */
17399 } /* end of handling backslash escape sequences */
17401 /* Here, we have the current token in 'value' */
17403 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17406 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17407 * literal, as is the character that began the false range, i.e.
17408 * the 'a' in the examples */
17410 const int w = (RExC_parse >= rangebegin)
17411 ? RExC_parse - rangebegin
17415 "False [] range \"%" UTF8f "\"",
17416 UTF8fARG(UTF, w, rangebegin));
17419 ckWARN2reg(RExC_parse,
17420 "False [] range \"%" UTF8f "\"",
17421 UTF8fARG(UTF, w, rangebegin));
17422 cp_list = add_cp_to_invlist(cp_list, '-');
17423 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17427 range = 0; /* this was not a true range */
17428 element_count += 2; /* So counts for three values */
17431 classnum = namedclass_to_classnum(namedclass);
17433 if (LOC && namedclass < ANYOF_POSIXL_MAX
17434 #ifndef HAS_ISASCII
17435 && classnum != _CC_ASCII
17438 SV* scratch_list = NULL;
17440 /* What the Posix classes (like \w, [:space:]) match isn't
17441 * generally knowable under locale until actual match time. A
17442 * special node is used for these which has extra space for a
17443 * bitmap, with a bit reserved for each named class that is to
17444 * be matched against. (This isn't needed for \p{} and
17445 * pseudo-classes, as they are not affected by locale, and
17446 * hence are dealt with separately.) However, if a named class
17447 * and its complement are both present, then it matches
17448 * everything, and there is no runtime dependency. Odd numbers
17449 * are the complements of the next lower number, so xor works.
17450 * (Note that something like [\w\D] should match everything,
17451 * because \d should be a proper subset of \w. But rather than
17452 * trust that the locale is well behaved, we leave this to
17453 * runtime to sort out) */
17454 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
17455 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
17456 POSIXL_ZERO(posixl);
17457 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
17458 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
17459 continue; /* We could ignore the rest of the class, but
17460 best to parse it for any errors */
17462 else { /* Here, isn't the complement of any already parsed
17464 POSIXL_SET(posixl, namedclass);
17465 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
17466 anyof_flags |= ANYOF_MATCHES_POSIXL;
17468 /* The above-Latin1 characters are not subject to locale
17469 * rules. Just add them to the unconditionally-matched
17472 /* Get the list of the above-Latin1 code points this
17474 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17475 PL_XPosix_ptrs[classnum],
17477 /* Odd numbers are complements,
17478 * like NDIGIT, NASCII, ... */
17479 namedclass % 2 != 0,
17481 /* Checking if 'cp_list' is NULL first saves an extra
17482 * clone. Its reference count will be decremented at the
17483 * next union, etc, or if this is the only instance, at the
17484 * end of the routine */
17486 cp_list = scratch_list;
17489 _invlist_union(cp_list, scratch_list, &cp_list);
17490 SvREFCNT_dec_NN(scratch_list);
17492 continue; /* Go get next character */
17497 /* Here, is not /l, or is a POSIX class for which /l doesn't
17498 * matter (or is a Unicode property, which is skipped here). */
17499 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17500 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17502 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17503 * nor /l make a difference in what these match,
17504 * therefore we just add what they match to cp_list. */
17505 if (classnum != _CC_VERTSPACE) {
17506 assert( namedclass == ANYOF_HORIZWS
17507 || namedclass == ANYOF_NHORIZWS);
17509 /* It turns out that \h is just a synonym for
17511 classnum = _CC_BLANK;
17514 _invlist_union_maybe_complement_2nd(
17516 PL_XPosix_ptrs[classnum],
17517 namedclass % 2 != 0, /* Complement if odd
17518 (NHORIZWS, NVERTWS)
17523 else if ( AT_LEAST_UNI_SEMANTICS
17524 || classnum == _CC_ASCII
17525 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
17526 || classnum == _CC_XDIGIT)))
17528 /* We usually have to worry about /d affecting what POSIX
17529 * classes match, with special code needed because we won't
17530 * know until runtime what all matches. But there is no
17531 * extra work needed under /u and /a; and [:ascii:] is
17532 * unaffected by /d; and :digit: and :xdigit: don't have
17533 * runtime differences under /d. So we can special case
17534 * these, and avoid some extra work below, and at runtime.
17536 _invlist_union_maybe_complement_2nd(
17538 ((AT_LEAST_ASCII_RESTRICTED)
17539 ? PL_Posix_ptrs[classnum]
17540 : PL_XPosix_ptrs[classnum]),
17541 namedclass % 2 != 0,
17544 else { /* Garden variety class. If is NUPPER, NALPHA, ...
17545 complement and use nposixes */
17546 SV** posixes_ptr = namedclass % 2 == 0
17549 _invlist_union_maybe_complement_2nd(
17551 PL_XPosix_ptrs[classnum],
17552 namedclass % 2 != 0,
17556 } /* end of namedclass \blah */
17558 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17560 /* If 'range' is set, 'value' is the ending of a range--check its
17561 * validity. (If value isn't a single code point in the case of a
17562 * range, we should have figured that out above in the code that
17563 * catches false ranges). Later, we will handle each individual code
17564 * point in the range. If 'range' isn't set, this could be the
17565 * beginning of a range, so check for that by looking ahead to see if
17566 * the next real character to be processed is the range indicator--the
17571 /* For unicode ranges, we have to test that the Unicode as opposed
17572 * to the native values are not decreasing. (Above 255, there is
17573 * no difference between native and Unicode) */
17574 if (unicode_range && prevvalue < 255 && value < 255) {
17575 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
17576 goto backwards_range;
17581 if (prevvalue > value) /* b-a */ {
17586 w = RExC_parse - rangebegin;
17588 "Invalid [] range \"%" UTF8f "\"",
17589 UTF8fARG(UTF, w, rangebegin));
17590 NOT_REACHED; /* NOTREACHED */
17594 prevvalue = value; /* save the beginning of the potential range */
17595 if (! stop_at_1 /* Can't be a range if parsing just one thing */
17596 && *RExC_parse == '-')
17598 char* next_char_ptr = RExC_parse + 1;
17600 /* Get the next real char after the '-' */
17601 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
17603 /* If the '-' is at the end of the class (just before the ']',
17604 * it is a literal minus; otherwise it is a range */
17605 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
17606 RExC_parse = next_char_ptr;
17608 /* a bad range like \w-, [:word:]- ? */
17609 if (namedclass > OOB_NAMEDCLASS) {
17610 if (strict || ckWARN(WARN_REGEXP)) {
17611 const int w = RExC_parse >= rangebegin
17612 ? RExC_parse - rangebegin
17615 vFAIL4("False [] range \"%*.*s\"",
17620 "False [] range \"%*.*s\"",
17624 cp_list = add_cp_to_invlist(cp_list, '-');
17627 range = 1; /* yeah, it's a range! */
17628 continue; /* but do it the next time */
17633 if (namedclass > OOB_NAMEDCLASS) {
17637 /* Here, we have a single value this time through the loop, and
17638 * <prevvalue> is the beginning of the range, if any; or <value> if
17641 /* non-Latin1 code point implies unicode semantics. */
17643 REQUIRE_UNI_RULES(flagp, 0);
17646 /* Ready to process either the single value, or the completed range.
17647 * For single-valued non-inverted ranges, we consider the possibility
17648 * of multi-char folds. (We made a conscious decision to not do this
17649 * for the other cases because it can often lead to non-intuitive
17650 * results. For example, you have the peculiar case that:
17651 * "s s" =~ /^[^\xDF]+$/i => Y
17652 * "ss" =~ /^[^\xDF]+$/i => N
17654 * See [perl #89750] */
17655 if (FOLD && allow_mutiple_chars && value == prevvalue) {
17656 if ( value == LATIN_SMALL_LETTER_SHARP_S
17657 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
17660 /* Here <value> is indeed a multi-char fold. Get what it is */
17662 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17665 UV folded = _to_uni_fold_flags(
17669 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
17670 ? FOLD_FLAGS_NOMIX_ASCII
17674 /* Here, <folded> should be the first character of the
17675 * multi-char fold of <value>, with <foldbuf> containing the
17676 * whole thing. But, if this fold is not allowed (because of
17677 * the flags), <fold> will be the same as <value>, and should
17678 * be processed like any other character, so skip the special
17680 if (folded != value) {
17682 /* Skip if we are recursed, currently parsing the class
17683 * again. Otherwise add this character to the list of
17684 * multi-char folds. */
17685 if (! RExC_in_multi_char_class) {
17686 STRLEN cp_count = utf8_length(foldbuf,
17687 foldbuf + foldlen);
17688 SV* multi_fold = sv_2mortal(newSVpvs(""));
17690 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
17693 = add_multi_match(multi_char_matches,
17699 /* This element should not be processed further in this
17702 value = save_value;
17703 prevvalue = save_prevvalue;
17709 if (strict && ckWARN(WARN_REGEXP)) {
17712 /* If the range starts above 255, everything is portable and
17713 * likely to be so for any forseeable character set, so don't
17715 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
17716 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
17718 else if (prevvalue != value) {
17720 /* Under strict, ranges that stop and/or end in an ASCII
17721 * printable should have each end point be a portable value
17722 * for it (preferably like 'A', but we don't warn if it is
17723 * a (portable) Unicode name or code point), and the range
17724 * must be be all digits or all letters of the same case.
17725 * Otherwise, the range is non-portable and unclear as to
17726 * what it contains */
17727 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
17728 && ( non_portable_endpoint
17729 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
17730 || (isLOWER_A(prevvalue) && isLOWER_A(value))
17731 || (isUPPER_A(prevvalue) && isUPPER_A(value))
17733 vWARN(RExC_parse, "Ranges of ASCII printables should"
17734 " be some subset of \"0-9\","
17735 " \"A-Z\", or \"a-z\"");
17737 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
17738 SSize_t index_start;
17739 SSize_t index_final;
17741 /* But the nature of Unicode and languages mean we
17742 * can't do the same checks for above-ASCII ranges,
17743 * except in the case of digit ones. These should
17744 * contain only digits from the same group of 10. The
17745 * ASCII case is handled just above. Hence here, the
17746 * range could be a range of digits. First some
17747 * unlikely special cases. Grandfather in that a range
17748 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
17749 * if its starting value is one of the 10 digits prior
17750 * to it. This is because it is an alternate way of
17751 * writing 19D1, and some people may expect it to be in
17752 * that group. But it is bad, because it won't give
17753 * the expected results. In Unicode 5.2 it was
17754 * considered to be in that group (of 11, hence), but
17755 * this was fixed in the next version */
17757 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
17758 goto warn_bad_digit_range;
17760 else if (UNLIKELY( prevvalue >= 0x1D7CE
17761 && value <= 0x1D7FF))
17763 /* This is the only other case currently in Unicode
17764 * where the algorithm below fails. The code
17765 * points just above are the end points of a single
17766 * range containing only decimal digits. It is 5
17767 * different series of 0-9. All other ranges of
17768 * digits currently in Unicode are just a single
17769 * series. (And mktables will notify us if a later
17770 * Unicode version breaks this.)
17772 * If the range being checked is at most 9 long,
17773 * and the digit values represented are in
17774 * numerical order, they are from the same series.
17776 if ( value - prevvalue > 9
17777 || ((( value - 0x1D7CE) % 10)
17778 <= (prevvalue - 0x1D7CE) % 10))
17780 goto warn_bad_digit_range;
17785 /* For all other ranges of digits in Unicode, the
17786 * algorithm is just to check if both end points
17787 * are in the same series, which is the same range.
17789 index_start = _invlist_search(
17790 PL_XPosix_ptrs[_CC_DIGIT],
17793 /* Warn if the range starts and ends with a digit,
17794 * and they are not in the same group of 10. */
17795 if ( index_start >= 0
17796 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
17798 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
17799 value)) != index_start
17800 && index_final >= 0
17801 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
17803 warn_bad_digit_range:
17804 vWARN(RExC_parse, "Ranges of digits should be"
17805 " from the same group of"
17812 if ((! range || prevvalue == value) && non_portable_endpoint) {
17813 if (isPRINT_A(value)) {
17816 if (isBACKSLASHED_PUNCT(value)) {
17817 literal[d++] = '\\';
17819 literal[d++] = (char) value;
17820 literal[d++] = '\0';
17823 "\"%.*s\" is more clearly written simply as \"%s\"",
17824 (int) (RExC_parse - rangebegin),
17829 else if (isMNEMONIC_CNTRL(value)) {
17831 "\"%.*s\" is more clearly written simply as \"%s\"",
17832 (int) (RExC_parse - rangebegin),
17834 cntrl_to_mnemonic((U8) value)
17840 /* Deal with this element of the class */
17843 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17846 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
17847 * that don't require special handling, we can just add the range like
17848 * we do for ASCII platforms */
17849 if ((UNLIKELY(prevvalue == 0) && value >= 255)
17850 || ! (prevvalue < 256
17852 || (! non_portable_endpoint
17853 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
17854 || (isUPPER_A(prevvalue)
17855 && isUPPER_A(value)))))))
17857 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17861 /* Here, requires special handling. This can be because it is a
17862 * range whose code points are considered to be Unicode, and so
17863 * must be individually translated into native, or because its a
17864 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
17865 * EBCDIC, but we have defined them to include only the "expected"
17866 * upper or lower case ASCII alphabetics. Subranges above 255 are
17867 * the same in native and Unicode, so can be added as a range */
17868 U8 start = NATIVE_TO_LATIN1(prevvalue);
17870 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
17871 for (j = start; j <= end; j++) {
17872 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
17875 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
17881 range = 0; /* this range (if it was one) is done now */
17882 } /* End of loop through all the text within the brackets */
17884 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
17885 output_posix_warnings(pRExC_state, posix_warnings);
17888 /* If anything in the class expands to more than one character, we have to
17889 * deal with them by building up a substitute parse string, and recursively
17890 * calling reg() on it, instead of proceeding */
17891 if (multi_char_matches) {
17892 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
17895 char *save_end = RExC_end;
17896 char *save_parse = RExC_parse;
17897 char *save_start = RExC_start;
17898 Size_t constructed_prefix_len = 0; /* This gives the length of the
17899 constructed portion of the
17900 substitute parse. */
17901 bool first_time = TRUE; /* First multi-char occurrence doesn't get
17906 /* Only one level of recursion allowed */
17907 assert(RExC_copy_start_in_constructed == RExC_precomp);
17909 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
17910 because too confusing */
17912 sv_catpvs(substitute_parse, "(?:");
17916 /* Look at the longest folds first */
17917 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
17922 if (av_exists(multi_char_matches, cp_count)) {
17923 AV** this_array_ptr;
17926 this_array_ptr = (AV**) av_fetch(multi_char_matches,
17928 while ((this_sequence = av_pop(*this_array_ptr)) !=
17931 if (! first_time) {
17932 sv_catpvs(substitute_parse, "|");
17934 first_time = FALSE;
17936 sv_catpv(substitute_parse, SvPVX(this_sequence));
17941 /* If the character class contains anything else besides these
17942 * multi-character folds, have to include it in recursive parsing */
17943 if (element_count) {
17944 sv_catpvs(substitute_parse, "|[");
17945 constructed_prefix_len = SvCUR(substitute_parse);
17946 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
17948 /* Put in a closing ']' only if not going off the end, as otherwise
17949 * we are adding something that really isn't there */
17950 if (RExC_parse < RExC_end) {
17951 sv_catpvs(substitute_parse, "]");
17955 sv_catpvs(substitute_parse, ")");
17958 /* This is a way to get the parse to skip forward a whole named
17959 * sequence instead of matching the 2nd character when it fails the
17961 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
17965 /* Set up the data structure so that any errors will be properly
17966 * reported. See the comments at the definition of
17967 * REPORT_LOCATION_ARGS for details */
17968 RExC_copy_start_in_input = (char *) orig_parse;
17969 RExC_start = RExC_parse = SvPV(substitute_parse, len);
17970 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
17971 RExC_end = RExC_parse + len;
17972 RExC_in_multi_char_class = 1;
17974 ret = reg(pRExC_state, 1, ®_flags, depth+1);
17976 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PARSE|NEED_UTF8);
17978 /* And restore so can parse the rest of the pattern */
17979 RExC_parse = save_parse;
17980 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
17981 RExC_end = save_end;
17982 RExC_in_multi_char_class = 0;
17983 SvREFCNT_dec_NN(multi_char_matches);
17987 /* If folding, we calculate all characters that could fold to or from the
17988 * ones already on the list */
17989 if (cp_foldable_list) {
17991 UV start, end; /* End points of code point ranges */
17993 SV* fold_intersection = NULL;
17996 /* Our calculated list will be for Unicode rules. For locale
17997 * matching, we have to keep a separate list that is consulted at
17998 * runtime only when the locale indicates Unicode rules (and we
17999 * don't include potential matches in the ASCII/Latin1 range, as
18000 * any code point could fold to any other, based on the run-time
18001 * locale). For non-locale, we just use the general list */
18003 use_list = &only_utf8_locale_list;
18006 use_list = &cp_list;
18009 /* Only the characters in this class that participate in folds need
18010 * be checked. Get the intersection of this class and all the
18011 * possible characters that are foldable. This can quickly narrow
18012 * down a large class */
18013 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18014 &fold_intersection);
18016 /* Now look at the foldable characters in this class individually */
18017 invlist_iterinit(fold_intersection);
18018 while (invlist_iternext(fold_intersection, &start, &end)) {
18022 /* Look at every character in the range */
18023 for (j = start; j <= end; j++) {
18024 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18027 Size_t folds_count;
18028 unsigned int first_fold;
18029 const unsigned int * remaining_folds;
18033 /* Under /l, we don't know what code points below 256
18034 * fold to, except we do know the MICRO SIGN folds to
18035 * an above-255 character if the locale is UTF-8, so we
18036 * add it to the special list (in *use_list) Otherwise
18037 * we know now what things can match, though some folds
18038 * are valid under /d only if the target is UTF-8.
18039 * Those go in a separate list */
18040 if ( IS_IN_SOME_FOLD_L1(j)
18041 && ! (LOC && j != MICRO_SIGN))
18044 /* ASCII is always matched; non-ASCII is matched
18045 * only under Unicode rules (which could happen
18046 * under /l if the locale is a UTF-8 one */
18047 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18048 *use_list = add_cp_to_invlist(*use_list,
18049 PL_fold_latin1[j]);
18051 else if (j != PL_fold_latin1[j]) {
18052 upper_latin1_only_utf8_matches
18053 = add_cp_to_invlist(
18054 upper_latin1_only_utf8_matches,
18055 PL_fold_latin1[j]);
18059 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18060 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18062 add_above_Latin1_folds(pRExC_state,
18069 /* Here is an above Latin1 character. We don't have the
18070 * rules hard-coded for it. First, get its fold. This is
18071 * the simple fold, as the multi-character folds have been
18072 * handled earlier and separated out */
18073 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18074 (ASCII_FOLD_RESTRICTED)
18075 ? FOLD_FLAGS_NOMIX_ASCII
18078 /* Single character fold of above Latin1. Add everything
18079 * in its fold closure to the list that this node should
18081 folds_count = _inverse_folds(folded, &first_fold,
18083 for (k = 0; k <= folds_count; k++) {
18084 UV c = (k == 0) /* First time through use itself */
18086 : (k == 1) /* 2nd time use, the first fold */
18089 /* Then the remaining ones */
18090 : remaining_folds[k-2];
18092 /* /aa doesn't allow folds between ASCII and non- */
18093 if (( ASCII_FOLD_RESTRICTED
18094 && (isASCII(c) != isASCII(j))))
18099 /* Folds under /l which cross the 255/256 boundary are
18100 * added to a separate list. (These are valid only
18101 * when the locale is UTF-8.) */
18102 if (c < 256 && LOC) {
18103 *use_list = add_cp_to_invlist(*use_list, c);
18107 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18109 cp_list = add_cp_to_invlist(cp_list, c);
18112 /* Similarly folds involving non-ascii Latin1
18113 * characters under /d are added to their list */
18114 upper_latin1_only_utf8_matches
18115 = add_cp_to_invlist(
18116 upper_latin1_only_utf8_matches,
18122 SvREFCNT_dec_NN(fold_intersection);
18125 /* Now that we have finished adding all the folds, there is no reason
18126 * to keep the foldable list separate */
18127 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18128 SvREFCNT_dec_NN(cp_foldable_list);
18131 /* And combine the result (if any) with any inversion lists from posix
18132 * classes. The lists are kept separate up to now because we don't want to
18133 * fold the classes */
18134 if (simple_posixes) { /* These are the classes known to be unaffected by
18137 _invlist_union(cp_list, simple_posixes, &cp_list);
18138 SvREFCNT_dec_NN(simple_posixes);
18141 cp_list = simple_posixes;
18144 if (posixes || nposixes) {
18145 if (! DEPENDS_SEMANTICS) {
18147 /* For everything but /d, we can just add the current 'posixes' and
18148 * 'nposixes' to the main list */
18151 _invlist_union(cp_list, posixes, &cp_list);
18152 SvREFCNT_dec_NN(posixes);
18160 _invlist_union(cp_list, nposixes, &cp_list);
18161 SvREFCNT_dec_NN(nposixes);
18164 cp_list = nposixes;
18169 /* Under /d, things like \w match upper Latin1 characters only if
18170 * the target string is in UTF-8. But things like \W match all the
18171 * upper Latin1 characters if the target string is not in UTF-8.
18173 * Handle the case with something like \W separately */
18175 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18177 /* A complemented posix class matches all upper Latin1
18178 * characters if not in UTF-8. And it matches just certain
18179 * ones when in UTF-8. That means those certain ones are
18180 * matched regardless, so can just be added to the
18181 * unconditional list */
18183 _invlist_union(cp_list, nposixes, &cp_list);
18184 SvREFCNT_dec_NN(nposixes);
18188 cp_list = nposixes;
18191 /* Likewise for 'posixes' */
18192 _invlist_union(posixes, cp_list, &cp_list);
18194 /* Likewise for anything else in the range that matched only
18196 if (upper_latin1_only_utf8_matches) {
18197 _invlist_union(cp_list,
18198 upper_latin1_only_utf8_matches,
18200 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18201 upper_latin1_only_utf8_matches = NULL;
18204 /* If we don't match all the upper Latin1 characters regardless
18205 * of UTF-8ness, we have to set a flag to match the rest when
18207 _invlist_subtract(only_non_utf8_list, cp_list,
18208 &only_non_utf8_list);
18209 if (_invlist_len(only_non_utf8_list) != 0) {
18210 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18212 SvREFCNT_dec_NN(only_non_utf8_list);
18215 /* Here there were no complemented posix classes. That means
18216 * the upper Latin1 characters in 'posixes' match only when the
18217 * target string is in UTF-8. So we have to add them to the
18218 * list of those types of code points, while adding the
18219 * remainder to the unconditional list.
18221 * First calculate what they are */
18222 SV* nonascii_but_latin1_properties = NULL;
18223 _invlist_intersection(posixes, PL_UpperLatin1,
18224 &nonascii_but_latin1_properties);
18226 /* And add them to the final list of such characters. */
18227 _invlist_union(upper_latin1_only_utf8_matches,
18228 nonascii_but_latin1_properties,
18229 &upper_latin1_only_utf8_matches);
18231 /* Remove them from what now becomes the unconditional list */
18232 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18235 /* And add those unconditional ones to the final list */
18237 _invlist_union(cp_list, posixes, &cp_list);
18238 SvREFCNT_dec_NN(posixes);
18245 SvREFCNT_dec(nonascii_but_latin1_properties);
18247 /* Get rid of any characters from the conditional list that we
18248 * now know are matched unconditionally, which may make that
18250 _invlist_subtract(upper_latin1_only_utf8_matches,
18252 &upper_latin1_only_utf8_matches);
18253 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18254 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18255 upper_latin1_only_utf8_matches = NULL;
18261 /* And combine the result (if any) with any inversion list from properties.
18262 * The lists are kept separate up to now so that we can distinguish the two
18263 * in regards to matching above-Unicode. A run-time warning is generated
18264 * if a Unicode property is matched against a non-Unicode code point. But,
18265 * we allow user-defined properties to match anything, without any warning,
18266 * and we also suppress the warning if there is a portion of the character
18267 * class that isn't a Unicode property, and which matches above Unicode, \W
18268 * or [\x{110000}] for example.
18269 * (Note that in this case, unlike the Posix one above, there is no
18270 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18271 * forces Unicode semantics */
18275 /* If it matters to the final outcome, see if a non-property
18276 * component of the class matches above Unicode. If so, the
18277 * warning gets suppressed. This is true even if just a single
18278 * such code point is specified, as, though not strictly correct if
18279 * another such code point is matched against, the fact that they
18280 * are using above-Unicode code points indicates they should know
18281 * the issues involved */
18283 warn_super = ! (invert
18284 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18287 _invlist_union(properties, cp_list, &cp_list);
18288 SvREFCNT_dec_NN(properties);
18291 cp_list = properties;
18296 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18298 /* Because an ANYOF node is the only one that warns, this node
18299 * can't be optimized into something else */
18300 optimizable = FALSE;
18304 /* Here, we have calculated what code points should be in the character
18307 * Now we can see about various optimizations. Fold calculation (which we
18308 * did above) needs to take place before inversion. Otherwise /[^k]/i
18309 * would invert to include K, which under /i would match k, which it
18310 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18311 * folded until runtime */
18313 /* If we didn't do folding, it's because some information isn't available
18314 * until runtime; set the run-time fold flag for these We know to set the
18315 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18316 * at least one 0-255 range code point */
18319 /* Some things on the list might be unconditionally included because of
18320 * other components. Remove them, and clean up the list if it goes to
18322 if (only_utf8_locale_list && cp_list) {
18323 _invlist_subtract(only_utf8_locale_list, cp_list,
18324 &only_utf8_locale_list);
18326 if (_invlist_len(only_utf8_locale_list) == 0) {
18327 SvREFCNT_dec_NN(only_utf8_locale_list);
18328 only_utf8_locale_list = NULL;
18331 if ( only_utf8_locale_list
18332 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18333 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18335 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18338 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18340 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
18342 invlist_iterinit(cp_list);
18343 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
18344 anyof_flags |= ANYOFL_FOLD;
18345 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18347 invlist_iterfinish(cp_list);
18350 else if ( DEPENDS_SEMANTICS
18351 && ( upper_latin1_only_utf8_matches
18352 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18354 RExC_seen_d_op = TRUE;
18355 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
18358 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
18362 && ! has_runtime_dependency)
18364 _invlist_invert(cp_list);
18366 /* Clear the invert flag since have just done it here */
18371 *ret_invlist = cp_list;
18376 /* All possible optimizations below still have these characteristics.
18377 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
18379 *flagp |= HASWIDTH|SIMPLE;
18381 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
18382 RExC_contains_locale = 1;
18385 /* Some character classes are equivalent to other nodes. Such nodes take
18386 * up less room, and some nodes require fewer operations to execute, than
18387 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
18388 * improve efficiency. */
18391 PERL_UINT_FAST8_T i;
18392 Size_t partial_cp_count = 0;
18393 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
18394 UV end[MAX_FOLD_FROMS+1] = { 0 };
18396 if (cp_list) { /* Count the code points in enough ranges that we would
18397 see all the ones possible in any fold in this version
18400 invlist_iterinit(cp_list);
18401 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
18402 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
18405 partial_cp_count += end[i] - start[i] + 1;
18408 invlist_iterfinish(cp_list);
18411 /* If we know at compile time that this matches every possible code
18412 * point, any run-time dependencies don't matter */
18413 if (start[0] == 0 && end[0] == UV_MAX) {
18415 ret = reganode(pRExC_state, OPFAIL, 0);
18418 ret = reg_node(pRExC_state, SANY);
18424 /* Similarly, for /l posix classes, if both a class and its
18425 * complement match, any run-time dependencies don't matter */
18427 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
18430 if ( POSIXL_TEST(posixl, namedclass) /* class */
18431 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
18434 ret = reganode(pRExC_state, OPFAIL, 0);
18437 ret = reg_node(pRExC_state, SANY);
18443 /* For well-behaved locales, some classes are subsets of others,
18444 * so complementing the subset and including the non-complemented
18445 * superset should match everything, like [\D[:alnum:]], and
18446 * [[:^alpha:][:alnum:]], but some implementations of locales are
18447 * buggy, and khw thinks its a bad idea to have optimization change
18448 * behavior, even if it avoids an OS bug in a given case */
18450 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
18452 /* If is a single posix /l class, can optimize to just that op.
18453 * Such a node will not match anything in the Latin1 range, as that
18454 * is not determinable until runtime, but will match whatever the
18455 * class does outside that range. (Note that some classes won't
18456 * match anything outside the range, like [:ascii:]) */
18457 if ( isSINGLE_BIT_SET(posixl)
18458 && (partial_cp_count == 0 || start[0] > 255))
18461 SV * class_above_latin1 = NULL;
18462 bool already_inverted;
18463 bool are_equivalent;
18465 /* Compute which bit is set, which is the same thing as, e.g.,
18466 * ANYOF_CNTRL. From
18467 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
18469 static const int MultiplyDeBruijnBitPosition2[32] =
18471 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
18472 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
18475 namedclass = MultiplyDeBruijnBitPosition2[(posixl
18476 * 0x077CB531U) >> 27];
18477 classnum = namedclass_to_classnum(namedclass);
18479 /* The named classes are such that the inverted number is one
18480 * larger than the non-inverted one */
18481 already_inverted = namedclass
18482 - classnum_to_namedclass(classnum);
18484 /* Create an inversion list of the official property, inverted
18485 * if the constructed node list is inverted, and restricted to
18486 * only the above latin1 code points, which are the only ones
18487 * known at compile time */
18488 _invlist_intersection_maybe_complement_2nd(
18490 PL_XPosix_ptrs[classnum],
18492 &class_above_latin1);
18493 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
18495 SvREFCNT_dec_NN(class_above_latin1);
18497 if (are_equivalent) {
18499 /* Resolve the run-time inversion flag with this possibly
18500 * inverted class */
18501 invert = invert ^ already_inverted;
18503 ret = reg_node(pRExC_state,
18504 POSIXL + invert * (NPOSIXL - POSIXL));
18505 FLAGS(REGNODE_p(ret)) = classnum;
18511 /* khw can't think of any other possible transformation involving
18513 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
18517 if (! has_runtime_dependency) {
18519 /* If the list is empty, nothing matches. This happens, for
18520 * example, when a Unicode property that doesn't match anything is
18521 * the only element in the character class (perluniprops.pod notes
18522 * such properties). */
18523 if (partial_cp_count == 0) {
18525 ret = reg_node(pRExC_state, SANY);
18528 ret = reganode(pRExC_state, OPFAIL, 0);
18534 /* If matches everything but \n */
18535 if ( start[0] == 0 && end[0] == '\n' - 1
18536 && start[1] == '\n' + 1 && end[1] == UV_MAX)
18539 ret = reg_node(pRExC_state, REG_ANY);
18545 /* Next see if can optimize classes that contain just a few code points
18546 * into an EXACTish node. The reason to do this is to let the
18547 * optimizer join this node with adjacent EXACTish ones.
18549 * An EXACTFish node can be generated even if not under /i, and vice
18550 * versa. But care must be taken. An EXACTFish node has to be such
18551 * that it only matches precisely the code points in the class, but we
18552 * want to generate the least restrictive one that does that, to
18553 * increase the odds of being able to join with an adjacent node. For
18554 * example, if the class contains [kK], we have to make it an EXACTFAA
18555 * node to prevent the KELVIN SIGN from matching. Whether we are under
18556 * /i or not is irrelevant in this case. Less obvious is the pattern
18557 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
18558 * supposed to match the single character U+0149 LATIN SMALL LETTER N
18559 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
18560 * that includes \X{02BC}, there is a multi-char fold that does, and so
18561 * the node generated for it must be an EXACTFish one. On the other
18562 * hand qr/:/i should generate a plain EXACT node since the colon
18563 * participates in no fold whatsoever, and having it EXACT tells the
18564 * optimizer the target string cannot match unless it has a colon in
18567 * We don't typically generate an EXACTish node if doing so would
18568 * require changing the pattern to UTF-8, as that affects /d and
18569 * otherwise is slower. However, under /i, not changing to UTF-8 can
18570 * miss some potential multi-character folds. We calculate the
18571 * EXACTish node, and then decide if something would be missed if we
18576 /* Only try if there are no more code points in the class than
18577 * in the max possible fold */
18578 && partial_cp_count > 0 && partial_cp_count <= MAX_FOLD_FROMS + 1
18580 && (start[0] < 256 || UTF || FOLD))
18582 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
18584 /* We can always make a single code point class into an
18585 * EXACTish node. */
18589 /* Here is /l: Use EXACTL, except /li indicates EXACTFL,
18590 * as that means there is a fold not known until runtime so
18591 * shows as only a single code point here. */
18592 op = (FOLD) ? EXACTFL : EXACTL;
18594 else if (! FOLD) { /* Not /l and not /i */
18595 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
18597 else if (start[0] < 256) { /* /i, not /l, and the code point is
18600 /* Under /i, it gets a little tricky. A code point that
18601 * doesn't participate in a fold should be an EXACT node.
18602 * We know this one isn't the result of a simple fold, or
18603 * there'd be more than one code point in the list, but it
18604 * could be part of a multi- character fold. In that case
18605 * we better not create an EXACT node, as we would wrongly
18606 * be telling the optimizer that this code point must be in
18607 * the target string, and that is wrong. This is because
18608 * if the sequence around this code point forms a
18609 * multi-char fold, what needs to be in the string could be
18610 * the code point that folds to the sequence.
18612 * This handles the case of below-255 code points, as we
18613 * have an easy look up for those. The next clause handles
18614 * the above-256 one */
18615 op = IS_IN_SOME_FOLD_L1(start[0])
18619 else { /* /i, larger code point. Since we are under /i, and
18620 have just this code point, we know that it can't
18621 fold to something else, so PL_InMultiCharFold
18623 op = _invlist_contains_cp(PL_InMultiCharFold,
18631 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
18632 && _invlist_contains_cp(PL_in_some_fold, start[0]))
18634 /* Here, the only runtime dependency, if any, is from /d, and
18635 * the class matches more than one code point, and the lowest
18636 * code point participates in some fold. It might be that the
18637 * other code points are /i equivalent to this one, and hence
18638 * they would representable by an EXACTFish node. Above, we
18639 * eliminated classes that contain too many code points to be
18640 * EXACTFish, with the test for MAX_FOLD_FROMS
18642 * First, special case the ASCII fold pairs, like 'B' and 'b'.
18643 * We do this because we have EXACTFAA at our disposal for the
18645 if (partial_cp_count == 2 && isASCII(start[0])) {
18647 /* The only ASCII characters that participate in folds are
18649 assert(isALPHA(start[0]));
18650 if ( end[0] == start[0] /* First range is a single
18651 character, so 2nd exists */
18652 && isALPHA_FOLD_EQ(start[0], start[1]))
18655 /* Here, is part of an ASCII fold pair */
18657 if ( ASCII_FOLD_RESTRICTED
18658 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
18660 /* If the second clause just above was true, it
18661 * means we can't be under /i, or else the list
18662 * would have included more than this fold pair.
18663 * Therefore we have to exclude the possibility of
18664 * whatever else it is that folds to these, by
18665 * using EXACTFAA */
18668 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
18670 /* Here, there's no simple fold that start[0] is part
18671 * of, but there is a multi-character one. If we
18672 * are not under /i, we want to exclude that
18673 * possibility; if under /i, we want to include it
18675 op = (FOLD) ? EXACTFU : EXACTFAA;
18679 /* Here, the only possible fold start[0] particpates in
18680 * is with start[1]. /i or not isn't relevant */
18684 value = toFOLD(start[0]);
18687 else if ( ! upper_latin1_only_utf8_matches
18688 || ( _invlist_len(upper_latin1_only_utf8_matches)
18691 invlist_highest(upper_latin1_only_utf8_matches)]
18694 /* Here, the smallest character is non-ascii or there are
18695 * more than 2 code points matched by this node. Also, we
18696 * either don't have /d UTF-8 dependent matches, or if we
18697 * do, they look like they could be a single character that
18698 * is the fold of the lowest one in the always-match list.
18699 * This test quickly excludes most of the false positives
18700 * when there are /d UTF-8 depdendent matches. These are
18701 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
18702 * SMALL LETTER A WITH GRAVE iff the target string is
18703 * UTF-8. (We don't have to worry above about exceeding
18704 * the array bounds of PL_fold_latin1[] because any code
18705 * point in 'upper_latin1_only_utf8_matches' is below 256.)
18707 * EXACTFAA would apply only to pairs (hence exactly 2 code
18708 * points) in the ASCII range, so we can't use it here to
18709 * artificially restrict the fold domain, so we check if
18710 * the class does or does not match some EXACTFish node.
18711 * Further, if we aren't under /i, and and the folded-to
18712 * character is part of a multi-character fold, we can't do
18713 * this optimization, as the sequence around it could be
18714 * that multi-character fold, and we don't here know the
18715 * context, so we have to assume it is that multi-char
18716 * fold, to prevent potential bugs.
18718 * To do the general case, we first find the fold of the
18719 * lowest code point (which may be higher than the lowest
18720 * one), then find everything that folds to it. (The data
18721 * structure we have only maps from the folded code points,
18722 * so we have to do the earlier step.) */
18725 U8 foldbuf[UTF8_MAXBYTES_CASE];
18726 UV folded = _to_uni_fold_flags(start[0],
18727 foldbuf, &foldlen, 0);
18728 unsigned int first_fold;
18729 const unsigned int * remaining_folds;
18730 Size_t folds_to_this_cp_count = _inverse_folds(
18734 Size_t folds_count = folds_to_this_cp_count + 1;
18735 SV * fold_list = _new_invlist(folds_count);
18738 /* If there are UTF-8 dependent matches, create a temporary
18739 * list of what this node matches, including them. */
18740 SV * all_cp_list = NULL;
18741 SV ** use_this_list = &cp_list;
18743 if (upper_latin1_only_utf8_matches) {
18744 all_cp_list = _new_invlist(0);
18745 use_this_list = &all_cp_list;
18746 _invlist_union(cp_list,
18747 upper_latin1_only_utf8_matches,
18751 /* Having gotten everything that participates in the fold
18752 * containing the lowest code point, we turn that into an
18753 * inversion list, making sure everything is included. */
18754 fold_list = add_cp_to_invlist(fold_list, start[0]);
18755 fold_list = add_cp_to_invlist(fold_list, folded);
18756 if (folds_to_this_cp_count > 0) {
18757 fold_list = add_cp_to_invlist(fold_list, first_fold);
18758 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
18759 fold_list = add_cp_to_invlist(fold_list,
18760 remaining_folds[i]);
18764 /* If the fold list is identical to what's in this ANYOF
18765 * node, the node can be represented by an EXACTFish one
18767 if (_invlistEQ(*use_this_list, fold_list,
18768 0 /* Don't complement */ )
18771 /* But, we have to be careful, as mentioned above.
18772 * Just the right sequence of characters could match
18773 * this if it is part of a multi-character fold. That
18774 * IS what we want if we are under /i. But it ISN'T
18775 * what we want if not under /i, as it could match when
18776 * it shouldn't. So, when we aren't under /i and this
18777 * character participates in a multi-char fold, we
18778 * don't optimize into an EXACTFish node. So, for each
18779 * case below we have to check if we are folding
18780 * and if not, if it is not part of a multi-char fold.
18782 if (start[0] > 255) { /* Highish code point */
18783 if (FOLD || ! _invlist_contains_cp(
18784 PL_InMultiCharFold, folded))
18788 : (ASCII_FOLD_RESTRICTED)
18793 } /* Below, the lowest code point < 256 */
18796 && DEPENDS_SEMANTICS)
18797 { /* An EXACTF node containing a single character
18798 's', can be an EXACTFU if it doesn't get
18799 joined with an adjacent 's' */
18800 op = EXACTFU_S_EDGE;
18804 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
18806 if (upper_latin1_only_utf8_matches) {
18809 /* We can't use the fold, as that only matches
18813 else if ( UNLIKELY(start[0] == MICRO_SIGN)
18815 { /* EXACTFUP is a special node for this
18817 op = (ASCII_FOLD_RESTRICTED)
18820 value = MICRO_SIGN;
18822 else if ( ASCII_FOLD_RESTRICTED
18823 && ! isASCII(start[0]))
18824 { /* For ASCII under /iaa, we can use EXACTFU
18836 SvREFCNT_dec_NN(fold_list);
18837 SvREFCNT_dec(all_cp_list);
18843 /* Here, we have calculated what EXACTish node we would use.
18844 * But we don't use it if it would require converting the
18845 * pattern to UTF-8, unless not using it could cause us to miss
18846 * some folds (hence be buggy) */
18848 if (! UTF && value > 255) {
18849 SV * in_multis = NULL;
18853 /* If there is no code point that is part of a multi-char
18854 * fold, then there aren't any matches, so we don't do this
18855 * optimization. Otherwise, it could match depending on
18856 * the context around us, so we do upgrade */
18857 _invlist_intersection(PL_InMultiCharFold, cp_list, &in_multis);
18858 if (UNLIKELY(_invlist_len(in_multis) != 0)) {
18859 REQUIRE_UTF8(flagp);
18867 U8 len = (UTF) ? UVCHR_SKIP(value) : 1;
18869 ret = regnode_guts(pRExC_state, op, len, "exact");
18870 FILL_NODE(ret, op);
18871 RExC_emit += 1 + STR_SZ(len);
18872 setSTR_LEN(REGNODE_p(ret), len);
18874 *STRING(REGNODE_p(ret)) = (U8) value;
18877 uvchr_to_utf8((U8 *) STRING(REGNODE_p(ret)), value);
18884 if (! has_runtime_dependency) {
18886 /* See if this can be turned into an ANYOFM node. Think about the
18887 * bit patterns in two different bytes. In some positions, the
18888 * bits in each will be 1; and in other positions both will be 0;
18889 * and in some positions the bit will be 1 in one byte, and 0 in
18890 * the other. Let 'n' be the number of positions where the bits
18891 * differ. We create a mask which has exactly 'n' 0 bits, each in
18892 * a position where the two bytes differ. Now take the set of all
18893 * bytes that when ANDed with the mask yield the same result. That
18894 * set has 2**n elements, and is representable by just two 8 bit
18895 * numbers: the result and the mask. Importantly, matching the set
18896 * can be vectorized by creating a word full of the result bytes,
18897 * and a word full of the mask bytes, yielding a significant speed
18898 * up. Here, see if this node matches such a set. As a concrete
18899 * example consider [01], and the byte representing '0' which is
18900 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
18901 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
18902 * 0x30. Any other bytes ANDed yield something else. So [01],
18903 * which is a common usage, is optimizable into ANYOFM, and can
18904 * benefit from the speed up. We can only do this on UTF-8
18905 * invariant bytes, because they have the same bit patterns under
18907 PERL_UINT_FAST8_T inverted = 0;
18909 const PERL_UINT_FAST8_T max_permissible = 0xFF;
18911 const PERL_UINT_FAST8_T max_permissible = 0x7F;
18913 /* If doesn't fit the criteria for ANYOFM, invert and try again.
18914 * If that works we will instead later generate an NANYOFM, and
18915 * invert back when through */
18916 if (invlist_highest(cp_list) > max_permissible) {
18917 _invlist_invert(cp_list);
18921 if (invlist_highest(cp_list) <= max_permissible) {
18922 UV this_start, this_end;
18923 UV lowest_cp = UV_MAX; /* inited to suppress compiler warn */
18924 U8 bits_differing = 0;
18925 Size_t full_cp_count = 0;
18926 bool first_time = TRUE;
18928 /* Go through the bytes and find the bit positions that differ
18930 invlist_iterinit(cp_list);
18931 while (invlist_iternext(cp_list, &this_start, &this_end)) {
18932 unsigned int i = this_start;
18935 if (! UVCHR_IS_INVARIANT(i)) {
18939 first_time = FALSE;
18940 lowest_cp = this_start;
18942 /* We have set up the code point to compare with.
18943 * Don't compare it with itself */
18947 /* Find the bit positions that differ from the lowest code
18948 * point in the node. Keep track of all such positions by
18950 for (; i <= this_end; i++) {
18951 if (! UVCHR_IS_INVARIANT(i)) {
18955 bits_differing |= i ^ lowest_cp;
18958 full_cp_count += this_end - this_start + 1;
18961 /* At the end of the loop, we count how many bits differ from
18962 * the bits in lowest code point, call the count 'd'. If the
18963 * set we found contains 2**d elements, it is the closure of
18964 * all code points that differ only in those bit positions. To
18965 * convince yourself of that, first note that the number in the
18966 * closure must be a power of 2, which we test for. The only
18967 * way we could have that count and it be some differing set,
18968 * is if we got some code points that don't differ from the
18969 * lowest code point in any position, but do differ from each
18970 * other in some other position. That means one code point has
18971 * a 1 in that position, and another has a 0. But that would
18972 * mean that one of them differs from the lowest code point in
18973 * that position, which possibility we've already excluded. */
18974 if ( (inverted || full_cp_count > 1)
18975 && full_cp_count == 1U << PL_bitcount[bits_differing])
18979 op = ANYOFM + inverted;;
18981 /* We need to make the bits that differ be 0's */
18982 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
18984 /* The argument is the lowest code point */
18985 ret = reganode(pRExC_state, op, lowest_cp);
18986 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
18990 invlist_iterfinish(cp_list);
18994 _invlist_invert(cp_list);
19001 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19002 * all were invariants, it wasn't inverted, and there is a single
19003 * range. This would be faster than some of the posix nodes we
19004 * create below like /\d/a, but would be twice the size. Without
19005 * having actually measured the gain, khw doesn't think the
19006 * tradeoff is really worth it */
19009 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
19010 PERL_UINT_FAST8_T type;
19011 SV * intersection = NULL;
19012 SV* d_invlist = NULL;
19014 /* See if this matches any of the POSIX classes. The POSIXA and
19015 * POSIXD ones are about the same speed as ANYOF ops, but take less
19016 * room; the ones that have above-Latin1 code point matches are
19017 * somewhat faster than ANYOF. */
19019 for (type = POSIXA; type >= POSIXD; type--) {
19022 if (type == POSIXL) { /* But not /l posix classes */
19026 for (posix_class = 0;
19027 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19030 SV** our_code_points = &cp_list;
19031 SV** official_code_points;
19034 if (type == POSIXA) {
19035 official_code_points = &PL_Posix_ptrs[posix_class];
19038 official_code_points = &PL_XPosix_ptrs[posix_class];
19041 /* Skip non-existent classes of this type. e.g. \v only
19042 * has an entry in PL_XPosix_ptrs */
19043 if (! *official_code_points) {
19047 /* Try both the regular class, and its inversion */
19048 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19049 bool this_inverted = invert ^ try_inverted;
19051 if (type != POSIXD) {
19053 /* This class that isn't /d can't match if we have
19054 * /d dependencies */
19055 if (has_runtime_dependency
19056 & HAS_D_RUNTIME_DEPENDENCY)
19061 else /* is /d */ if (! this_inverted) {
19063 /* /d classes don't match anything non-ASCII below
19064 * 256 unconditionally (which cp_list contains) */
19065 _invlist_intersection(cp_list, PL_UpperLatin1,
19067 if (_invlist_len(intersection) != 0) {
19071 SvREFCNT_dec(d_invlist);
19072 d_invlist = invlist_clone(cp_list, NULL);
19074 /* But under UTF-8 it turns into using /u rules.
19075 * Add the things it matches under these conditions
19076 * so that we check below that these are identical
19077 * to what the tested class should match */
19078 if (upper_latin1_only_utf8_matches) {
19081 upper_latin1_only_utf8_matches,
19084 our_code_points = &d_invlist;
19086 else { /* POSIXD, inverted. If this doesn't have this
19087 flag set, it isn't /d. */
19088 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
19092 our_code_points = &cp_list;
19095 /* Here, have weeded out some things. We want to see
19096 * if the list of characters this node contains
19097 * ('*our_code_points') precisely matches those of the
19098 * class we are currently checking against
19099 * ('*official_code_points'). */
19100 if (_invlistEQ(*our_code_points,
19101 *official_code_points,
19104 /* Here, they precisely match. Optimize this ANYOF
19105 * node into its equivalent POSIX one of the
19106 * correct type, possibly inverted */
19107 ret = reg_node(pRExC_state, (try_inverted)
19111 FLAGS(REGNODE_p(ret)) = posix_class;
19112 SvREFCNT_dec(d_invlist);
19113 SvREFCNT_dec(intersection);
19119 SvREFCNT_dec(d_invlist);
19120 SvREFCNT_dec(intersection);
19123 /* If didn't find an optimization and there is no need for a bitmap,
19124 * optimize to indicate that */
19125 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19127 && ! upper_latin1_only_utf8_matches
19128 && anyof_flags == 0)
19130 U8 low_utf8[UTF8_MAXBYTES+1];
19131 UV highest_cp = invlist_highest(cp_list);
19135 /* Currently the maximum allowed code point by the system is
19136 * IV_MAX. Higher ones are reserved for future internal use. This
19137 * particular regnode can be used for higher ones, but we can't
19138 * calculate the code point of those. IV_MAX suffices though, as
19139 * it will be a large first byte */
19140 (void) uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX));
19142 /* We store the lowest possible first byte of the UTF-8
19143 * representation, using the flags field. This allows for quick
19144 * ruling out of some inputs without having to convert from UTF-8
19145 * to code point. For EBCDIC, this has to be I8. */
19146 anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
19148 /* If the first UTF-8 start byte for the highest code point in the
19149 * range is suitably small, we may be able to get an upper bound as
19151 if (highest_cp <= IV_MAX) {
19152 U8 high_utf8[UTF8_MAXBYTES+1];
19154 (void) uvchr_to_utf8(high_utf8, highest_cp);
19156 /* If the lowest and highest are the same, we can get an exact
19157 * first byte instead of a just minimum. We signal this with a
19158 * different regnode */
19159 if (low_utf8[0] == high_utf8[0]) {
19161 /* No need to convert to I8 for EBCDIC as this is an exact
19163 anyof_flags = low_utf8[0];
19166 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
19169 /* Here, the high byte is not the same as the low, but is
19170 * small enough that its reasonable to have a loose upper
19171 * bound, which is packed in with the strict lower bound.
19172 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
19173 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
19174 * is the same thing as UTF-8 */
19177 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - anyof_flags;
19178 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
19181 if (range_diff <= max_range_diff / 8) {
19184 else if (range_diff <= max_range_diff / 4) {
19187 else if (range_diff <= max_range_diff / 2) {
19190 anyof_flags = (anyof_flags - 0xC0) << 2 | bits;
19195 goto done_finding_op;
19197 } /* End of seeing if can optimize it into a different node */
19199 is_anyof: /* It's going to be an ANYOF node. */
19200 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19210 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19211 FILL_NODE(ret, op); /* We set the argument later */
19212 RExC_emit += 1 + regarglen[op];
19213 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19215 /* Here, <cp_list> contains all the code points we can determine at
19216 * compile time that match under all conditions. Go through it, and
19217 * for things that belong in the bitmap, put them there, and delete from
19218 * <cp_list>. While we are at it, see if everything above 255 is in the
19219 * list, and if so, set a flag to speed up execution */
19221 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19224 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19228 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19231 /* Here, the bitmap has been populated with all the Latin1 code points that
19232 * always match. Can now add to the overall list those that match only
19233 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19235 if (upper_latin1_only_utf8_matches) {
19237 _invlist_union(cp_list,
19238 upper_latin1_only_utf8_matches,
19240 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19243 cp_list = upper_latin1_only_utf8_matches;
19245 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19248 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19249 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19251 only_utf8_locale_list);
19256 /* Here, the node is getting optimized into something that's not an ANYOF
19257 * one. Finish up. */
19259 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19260 RExC_parse - orig_parse);;
19261 SvREFCNT_dec(cp_list);;
19265 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19268 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19269 regnode* const node,
19271 SV* const runtime_defns,
19272 SV* const only_utf8_locale_list)
19274 /* Sets the arg field of an ANYOF-type node 'node', using information about
19275 * the node passed-in. If there is nothing outside the node's bitmap, the
19276 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
19277 * the count returned by add_data(), having allocated and stored an array,
19280 * av[0] stores the inversion list defining this class as far as known at
19281 * this time, or PL_sv_undef if nothing definite is now known.
19282 * av[1] stores the inversion list of code points that match only if the
19283 * current locale is UTF-8, or if none, PL_sv_undef if there is an
19284 * av[2], or no entry otherwise.
19285 * av[2] stores the list of user-defined properties whose subroutine
19286 * definitions aren't known at this time, or no entry if none. */
19290 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
19292 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
19293 assert(! (ANYOF_FLAGS(node)
19294 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
19295 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
19298 AV * const av = newAV();
19302 av_store(av, INVLIST_INDEX, cp_list);
19305 if (only_utf8_locale_list) {
19306 av_store(av, ONLY_LOCALE_MATCHES_INDEX, only_utf8_locale_list);
19309 if (runtime_defns) {
19310 av_store(av, DEFERRED_USER_DEFINED_INDEX, SvREFCNT_inc(runtime_defns));
19313 rv = newRV_noinc(MUTABLE_SV(av));
19314 n = add_data(pRExC_state, STR_WITH_LEN("s"));
19315 RExC_rxi->data->data[n] = (void*)rv;
19320 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
19322 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
19323 const regnode* node,
19326 SV** only_utf8_locale_ptr,
19327 SV** output_invlist)
19330 /* For internal core use only.
19331 * Returns the inversion list for the input 'node' in the regex 'prog'.
19332 * If <doinit> is 'true', will attempt to create the inversion list if not
19334 * If <listsvp> is non-null, will return the printable contents of the
19335 * property definition. This can be used to get debugging information
19336 * even before the inversion list exists, by calling this function with
19337 * 'doinit' set to false, in which case the components that will be used
19338 * to eventually create the inversion list are returned (in a printable
19340 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
19341 * store an inversion list of code points that should match only if the
19342 * execution-time locale is a UTF-8 one.
19343 * If <output_invlist> is not NULL, it is where this routine is to store an
19344 * inversion list of the code points that would be instead returned in
19345 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
19346 * when this parameter is used, is just the non-code point data that
19347 * will go into creating the inversion list. This currently should be just
19348 * user-defined properties whose definitions were not known at compile
19349 * time. Using this parameter allows for easier manipulation of the
19350 * inversion list's data by the caller. It is illegal to call this
19351 * function with this parameter set, but not <listsvp>
19353 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
19354 * that, in spite of this function's name, the inversion list it returns
19355 * may include the bitmap data as well */
19357 SV *si = NULL; /* Input initialization string */
19358 SV* invlist = NULL;
19360 RXi_GET_DECL(prog, progi);
19361 const struct reg_data * const data = prog ? progi->data : NULL;
19363 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
19364 assert(! output_invlist || listsvp);
19366 if (data && data->count) {
19367 const U32 n = ARG(node);
19369 if (data->what[n] == 's') {
19370 SV * const rv = MUTABLE_SV(data->data[n]);
19371 AV * const av = MUTABLE_AV(SvRV(rv));
19372 SV **const ary = AvARRAY(av);
19374 invlist = ary[INVLIST_INDEX];
19376 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
19377 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
19380 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
19381 si = ary[DEFERRED_USER_DEFINED_INDEX];
19384 if (doinit && (si || invlist)) {
19387 SV * msg = newSVpvs_flags("", SVs_TEMP);
19389 SV * prop_definition = handle_user_defined_property(
19390 "", 0, FALSE, /* There is no \p{}, \P{} */
19391 SvPVX_const(si)[1] - '0', /* /i or not has been
19392 stored here for just
19394 TRUE, /* run time */
19395 FALSE, /* This call must find the defn */
19396 si, /* The property definition */
19399 0 /* base level call */
19403 assert(prop_definition == NULL);
19405 Perl_croak(aTHX_ "%" UTF8f,
19406 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
19410 _invlist_union(invlist, prop_definition, &invlist);
19411 SvREFCNT_dec_NN(prop_definition);
19414 invlist = prop_definition;
19417 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
19418 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
19420 av_store(av, INVLIST_INDEX, invlist);
19421 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
19422 ? ONLY_LOCALE_MATCHES_INDEX:
19430 /* If requested, return a printable version of what this ANYOF node matches
19433 SV* matches_string = NULL;
19435 /* This function can be called at compile-time, before everything gets
19436 * resolved, in which case we return the currently best available
19437 * information, which is the string that will eventually be used to do
19438 * that resolving, 'si' */
19440 /* Here, we only have 'si' (and possibly some passed-in data in
19441 * 'invlist', which is handled below) If the caller only wants
19442 * 'si', use that. */
19443 if (! output_invlist) {
19444 matches_string = newSVsv(si);
19447 /* But if the caller wants an inversion list of the node, we
19448 * need to parse 'si' and place as much as possible in the
19449 * desired output inversion list, making 'matches_string' only
19450 * contain the currently unresolvable things */
19451 const char *si_string = SvPVX(si);
19452 STRLEN remaining = SvCUR(si);
19456 /* Ignore everything before the first new-line */
19457 while (*si_string != '\n' && remaining > 0) {
19461 assert(remaining > 0);
19466 while (remaining > 0) {
19468 /* The data consists of just strings defining user-defined
19469 * property names, but in prior incarnations, and perhaps
19470 * somehow from pluggable regex engines, it could still
19471 * hold hex code point definitions. Each component of a
19472 * range would be separated by a tab, and each range by a
19473 * new-line. If these are found, instead add them to the
19474 * inversion list */
19475 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
19476 |PERL_SCAN_SILENT_NON_PORTABLE;
19477 STRLEN len = remaining;
19478 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
19480 /* If the hex decode routine found something, it should go
19481 * up to the next \n */
19482 if ( *(si_string + len) == '\n') {
19483 if (count) { /* 2nd code point on line */
19484 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
19487 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
19490 goto prepare_for_next_iteration;
19493 /* If the hex decode was instead for the lower range limit,
19494 * save it, and go parse the upper range limit */
19495 if (*(si_string + len) == '\t') {
19496 assert(count == 0);
19500 prepare_for_next_iteration:
19501 si_string += len + 1;
19502 remaining -= len + 1;
19506 /* Here, didn't find a legal hex number. Just add it from
19507 * here to the next \n */
19510 while (*(si_string + len) != '\n' && remaining > 0) {
19514 if (*(si_string + len) == '\n') {
19518 if (matches_string) {
19519 sv_catpvn(matches_string, si_string, len - 1);
19522 matches_string = newSVpvn(si_string, len - 1);
19525 sv_catpvs(matches_string, " ");
19526 } /* end of loop through the text */
19528 assert(matches_string);
19529 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
19530 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
19532 } /* end of has an 'si' */
19535 /* Add the stuff that's already known */
19538 /* Again, if the caller doesn't want the output inversion list, put
19539 * everything in 'matches-string' */
19540 if (! output_invlist) {
19541 if ( ! matches_string) {
19542 matches_string = newSVpvs("\n");
19544 sv_catsv(matches_string, invlist_contents(invlist,
19545 TRUE /* traditional style */
19548 else if (! *output_invlist) {
19549 *output_invlist = invlist_clone(invlist, NULL);
19552 _invlist_union(*output_invlist, invlist, output_invlist);
19556 *listsvp = matches_string;
19561 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
19563 /* reg_skipcomment()
19565 Absorbs an /x style # comment from the input stream,
19566 returning a pointer to the first character beyond the comment, or if the
19567 comment terminates the pattern without anything following it, this returns
19568 one past the final character of the pattern (in other words, RExC_end) and
19569 sets the REG_RUN_ON_COMMENT_SEEN flag.
19571 Note it's the callers responsibility to ensure that we are
19572 actually in /x mode
19576 PERL_STATIC_INLINE char*
19577 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
19579 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
19583 while (p < RExC_end) {
19584 if (*(++p) == '\n') {
19589 /* we ran off the end of the pattern without ending the comment, so we have
19590 * to add an \n when wrapping */
19591 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
19596 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
19598 const bool force_to_xmod
19601 /* If the text at the current parse position '*p' is a '(?#...)' comment,
19602 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
19603 * is /x whitespace, advance '*p' so that on exit it points to the first
19604 * byte past all such white space and comments */
19606 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
19608 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
19610 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
19613 if (RExC_end - (*p) >= 3
19615 && *(*p + 1) == '?'
19616 && *(*p + 2) == '#')
19618 while (*(*p) != ')') {
19619 if ((*p) == RExC_end)
19620 FAIL("Sequence (?#... not terminated");
19628 const char * save_p = *p;
19629 while ((*p) < RExC_end) {
19631 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
19634 else if (*(*p) == '#') {
19635 (*p) = reg_skipcomment(pRExC_state, (*p));
19641 if (*p != save_p) {
19654 Advances the parse position by one byte, unless that byte is the beginning
19655 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
19656 those two cases, the parse position is advanced beyond all such comments and
19659 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
19663 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
19665 PERL_ARGS_ASSERT_NEXTCHAR;
19667 if (RExC_parse < RExC_end) {
19669 || UTF8_IS_INVARIANT(*RExC_parse)
19670 || UTF8_IS_START(*RExC_parse));
19672 RExC_parse += (UTF)
19673 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
19676 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
19677 FALSE /* Don't force /x */ );
19682 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
19684 /* 'size' is the delta number of smallest regnode equivalents to add or
19685 * subtract from the current memory allocated to the regex engine being
19688 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
19693 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
19694 /* +1 for REG_MAGIC */
19697 if ( RExC_rxi == NULL )
19698 FAIL("Regexp out of space");
19699 RXi_SET(RExC_rx, RExC_rxi);
19701 RExC_emit_start = RExC_rxi->program;
19703 Zero(REGNODE_p(RExC_emit), size, regnode);
19706 #ifdef RE_TRACK_PATTERN_OFFSETS
19707 Renew(RExC_offsets, 2*RExC_size+1, U32);
19709 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
19711 RExC_offsets[0] = RExC_size;
19715 STATIC regnode_offset
19716 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
19718 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
19719 * equivalents space. It aligns and increments RExC_size and RExC_emit
19721 * It returns the regnode's offset into the regex engine program */
19723 const regnode_offset ret = RExC_emit;
19725 GET_RE_DEBUG_FLAGS_DECL;
19727 PERL_ARGS_ASSERT_REGNODE_GUTS;
19729 SIZE_ALIGN(RExC_size);
19730 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
19731 NODE_ALIGN_FILL(REGNODE_p(ret));
19732 #ifndef RE_TRACK_PATTERN_OFFSETS
19733 PERL_UNUSED_ARG(name);
19734 PERL_UNUSED_ARG(op);
19736 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
19738 if (RExC_offsets) { /* MJD */
19740 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
19743 (UV)(RExC_emit) > RExC_offsets[0]
19744 ? "Overwriting end of array!\n" : "OK",
19746 (UV)(RExC_parse - RExC_start),
19747 (UV)RExC_offsets[0]));
19748 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
19755 - reg_node - emit a node
19757 STATIC regnode_offset /* Location. */
19758 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
19760 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
19761 regnode_offset ptr = ret;
19763 PERL_ARGS_ASSERT_REG_NODE;
19765 assert(regarglen[op] == 0);
19767 FILL_ADVANCE_NODE(ptr, op);
19773 - reganode - emit a node with an argument
19775 STATIC regnode_offset /* Location. */
19776 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
19778 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
19779 regnode_offset ptr = ret;
19781 PERL_ARGS_ASSERT_REGANODE;
19783 /* ANYOF are special cased to allow non-length 1 args */
19784 assert(regarglen[op] == 1);
19786 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
19791 STATIC regnode_offset
19792 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
19794 /* emit a node with U32 and I32 arguments */
19796 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
19797 regnode_offset ptr = ret;
19799 PERL_ARGS_ASSERT_REG2LANODE;
19801 assert(regarglen[op] == 2);
19803 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
19809 - reginsert - insert an operator in front of already-emitted operand
19811 * That means that on exit 'operand' is the offset of the newly inserted
19812 * operator, and the original operand has been relocated.
19814 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
19815 * set up NEXT_OFF() of the inserted node if needed. Something like this:
19817 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
19818 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
19820 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
19823 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
19824 const regnode_offset operand, const U32 depth)
19829 const int offset = regarglen[(U8)op];
19830 const int size = NODE_STEP_REGNODE + offset;
19831 GET_RE_DEBUG_FLAGS_DECL;
19833 PERL_ARGS_ASSERT_REGINSERT;
19834 PERL_UNUSED_CONTEXT;
19835 PERL_UNUSED_ARG(depth);
19836 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
19837 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
19838 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
19839 studying. If this is wrong then we need to adjust RExC_recurse
19840 below like we do with RExC_open_parens/RExC_close_parens. */
19841 change_engine_size(pRExC_state, (Ptrdiff_t) size);
19842 src = REGNODE_p(RExC_emit);
19844 dst = REGNODE_p(RExC_emit);
19846 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
19847 * and [perl #133871] shows this can lead to problems, so skip this
19848 * realignment of parens until a later pass when they are reliable */
19849 if (! IN_PARENS_PASS && RExC_open_parens) {
19851 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
19852 /* remember that RExC_npar is rex->nparens + 1,
19853 * iow it is 1 more than the number of parens seen in
19854 * the pattern so far. */
19855 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
19856 /* note, RExC_open_parens[0] is the start of the
19857 * regex, it can't move. RExC_close_parens[0] is the end
19858 * of the regex, it *can* move. */
19859 if ( paren && RExC_open_parens[paren] >= operand ) {
19860 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
19861 RExC_open_parens[paren] += size;
19863 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
19865 if ( RExC_close_parens[paren] >= operand ) {
19866 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
19867 RExC_close_parens[paren] += size;
19869 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
19874 RExC_end_op += size;
19876 while (src > REGNODE_p(operand)) {
19877 StructCopy(--src, --dst, regnode);
19878 #ifdef RE_TRACK_PATTERN_OFFSETS
19879 if (RExC_offsets) { /* MJD 20010112 */
19881 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
19885 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
19886 ? "Overwriting end of array!\n" : "OK",
19887 (UV)REGNODE_OFFSET(src),
19888 (UV)REGNODE_OFFSET(dst),
19889 (UV)RExC_offsets[0]));
19890 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
19891 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
19896 place = REGNODE_p(operand); /* Op node, where operand used to be. */
19897 #ifdef RE_TRACK_PATTERN_OFFSETS
19898 if (RExC_offsets) { /* MJD */
19900 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
19904 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
19905 ? "Overwriting end of array!\n" : "OK",
19906 (UV)REGNODE_OFFSET(place),
19907 (UV)(RExC_parse - RExC_start),
19908 (UV)RExC_offsets[0]));
19909 Set_Node_Offset(place, RExC_parse);
19910 Set_Node_Length(place, 1);
19913 src = NEXTOPER(place);
19915 FILL_NODE(operand, op);
19917 /* Zero out any arguments in the new node */
19918 Zero(src, offset, regnode);
19922 - regtail - set the next-pointer at the end of a node chain of p to val. If
19923 that value won't fit in the space available, instead returns FALSE.
19924 (Except asserts if we can't fit in the largest space the regex
19925 engine is designed for.)
19926 - SEE ALSO: regtail_study
19929 S_regtail(pTHX_ RExC_state_t * pRExC_state,
19930 const regnode_offset p,
19931 const regnode_offset val,
19934 regnode_offset scan;
19935 GET_RE_DEBUG_FLAGS_DECL;
19937 PERL_ARGS_ASSERT_REGTAIL;
19939 PERL_UNUSED_ARG(depth);
19942 /* Find last node. */
19943 scan = (regnode_offset) p;
19945 regnode * const temp = regnext(REGNODE_p(scan));
19947 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
19948 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
19949 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
19950 SvPV_nolen_const(RExC_mysv), scan,
19951 (temp == NULL ? "->" : ""),
19952 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
19957 scan = REGNODE_OFFSET(temp);
19960 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
19961 assert((UV) (val - scan) <= U32_MAX);
19962 ARG_SET(REGNODE_p(scan), val - scan);
19965 if (val - scan > U16_MAX) {
19966 /* Populate this with something that won't loop and will likely
19967 * lead to a crash if the caller ignores the failure return, and
19968 * execution continues */
19969 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
19972 NEXT_OFF(REGNODE_p(scan)) = val - scan;
19980 - regtail_study - set the next-pointer at the end of a node chain of p to val.
19981 - Look for optimizable sequences at the same time.
19982 - currently only looks for EXACT chains.
19984 This is experimental code. The idea is to use this routine to perform
19985 in place optimizations on branches and groups as they are constructed,
19986 with the long term intention of removing optimization from study_chunk so
19987 that it is purely analytical.
19989 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
19990 to control which is which.
19992 This used to return a value that was ignored. It was a problem that it is
19993 #ifdef'd to be another function that didn't return a value. khw has changed it
19994 so both currently return a pass/fail return.
19997 /* TODO: All four parms should be const */
20000 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20001 const regnode_offset val, U32 depth)
20003 regnode_offset scan;
20005 #ifdef EXPERIMENTAL_INPLACESCAN
20008 GET_RE_DEBUG_FLAGS_DECL;
20010 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20013 /* Find last node. */
20017 regnode * const temp = regnext(REGNODE_p(scan));
20018 #ifdef EXPERIMENTAL_INPLACESCAN
20019 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20020 bool unfolded_multi_char; /* Unexamined in this routine */
20021 if (join_exact(pRExC_state, scan, &min,
20022 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20023 return TRUE; /* Was return EXACT */
20027 switch (OP(REGNODE_p(scan))) {
20034 case EXACTFU_S_EDGE:
20035 case EXACTFAA_NO_TRIE:
20042 if( exact == PSEUDO )
20043 exact= OP(REGNODE_p(scan));
20044 else if ( exact != OP(REGNODE_p(scan)) )
20053 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20054 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20055 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
20056 SvPV_nolen_const(RExC_mysv),
20058 PL_reg_name[exact]);
20062 scan = REGNODE_OFFSET(temp);
20065 DEBUG_PARSE_MSG("");
20066 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20067 Perl_re_printf( aTHX_
20068 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20069 SvPV_nolen_const(RExC_mysv),
20074 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20075 assert((UV) (val - scan) <= U32_MAX);
20076 ARG_SET(REGNODE_p(scan), val - scan);
20079 if (val - scan > U16_MAX) {
20080 /* Populate this with something that won't loop and will likely
20081 * lead to a crash if the caller ignores the failure return, and
20082 * execution continues */
20083 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20086 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20089 return TRUE; /* Was 'return exact' */
20094 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
20096 /* Returns an inversion list of all the code points matched by the
20097 * ANYOFM/NANYOFM node 'n' */
20099 SV * cp_list = _new_invlist(-1);
20100 const U8 lowest = (U8) ARG(n);
20103 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
20105 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
20107 /* Starting with the lowest code point, any code point that ANDed with the
20108 * mask yields the lowest code point is in the set */
20109 for (i = lowest; i <= 0xFF; i++) {
20110 if ((i & FLAGS(n)) == ARG(n)) {
20111 cp_list = add_cp_to_invlist(cp_list, i);
20114 /* We know how many code points (a power of two) that are in the
20115 * set. No use looking once we've got that number */
20116 if (count >= needed) break;
20120 if (OP(n) == NANYOFM) {
20121 _invlist_invert(cp_list);
20127 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
20132 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
20137 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20139 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
20140 if (flags & (1<<bit)) {
20141 if (!set++ && lead)
20142 Perl_re_printf( aTHX_ "%s", lead);
20143 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
20148 Perl_re_printf( aTHX_ "\n");
20150 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20155 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20161 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20163 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20164 if (flags & (1<<bit)) {
20165 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20168 if (!set++ && lead)
20169 Perl_re_printf( aTHX_ "%s", lead);
20170 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20173 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20174 if (!set++ && lead) {
20175 Perl_re_printf( aTHX_ "%s", lead);
20178 case REGEX_UNICODE_CHARSET:
20179 Perl_re_printf( aTHX_ "UNICODE");
20181 case REGEX_LOCALE_CHARSET:
20182 Perl_re_printf( aTHX_ "LOCALE");
20184 case REGEX_ASCII_RESTRICTED_CHARSET:
20185 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20187 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20188 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20191 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20197 Perl_re_printf( aTHX_ "\n");
20199 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20205 Perl_regdump(pTHX_ const regexp *r)
20209 SV * const sv = sv_newmortal();
20210 SV *dsv= sv_newmortal();
20211 RXi_GET_DECL(r, ri);
20212 GET_RE_DEBUG_FLAGS_DECL;
20214 PERL_ARGS_ASSERT_REGDUMP;
20216 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20218 /* Header fields of interest. */
20219 for (i = 0; i < 2; i++) {
20220 if (r->substrs->data[i].substr) {
20221 RE_PV_QUOTED_DECL(s, 0, dsv,
20222 SvPVX_const(r->substrs->data[i].substr),
20223 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20224 PL_dump_re_max_len);
20225 Perl_re_printf( aTHX_
20226 "%s %s%s at %" IVdf "..%" UVuf " ",
20227 i ? "floating" : "anchored",
20229 RE_SV_TAIL(r->substrs->data[i].substr),
20230 (IV)r->substrs->data[i].min_offset,
20231 (UV)r->substrs->data[i].max_offset);
20233 else if (r->substrs->data[i].utf8_substr) {
20234 RE_PV_QUOTED_DECL(s, 1, dsv,
20235 SvPVX_const(r->substrs->data[i].utf8_substr),
20236 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20238 Perl_re_printf( aTHX_
20239 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20240 i ? "floating" : "anchored",
20242 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20243 (IV)r->substrs->data[i].min_offset,
20244 (UV)r->substrs->data[i].max_offset);
20248 if (r->check_substr || r->check_utf8)
20249 Perl_re_printf( aTHX_
20251 ( r->check_substr == r->substrs->data[1].substr
20252 && r->check_utf8 == r->substrs->data[1].utf8_substr
20253 ? "(checking floating" : "(checking anchored"));
20254 if (r->intflags & PREGf_NOSCAN)
20255 Perl_re_printf( aTHX_ " noscan");
20256 if (r->extflags & RXf_CHECK_ALL)
20257 Perl_re_printf( aTHX_ " isall");
20258 if (r->check_substr || r->check_utf8)
20259 Perl_re_printf( aTHX_ ") ");
20261 if (ri->regstclass) {
20262 regprop(r, sv, ri->regstclass, NULL, NULL);
20263 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
20265 if (r->intflags & PREGf_ANCH) {
20266 Perl_re_printf( aTHX_ "anchored");
20267 if (r->intflags & PREGf_ANCH_MBOL)
20268 Perl_re_printf( aTHX_ "(MBOL)");
20269 if (r->intflags & PREGf_ANCH_SBOL)
20270 Perl_re_printf( aTHX_ "(SBOL)");
20271 if (r->intflags & PREGf_ANCH_GPOS)
20272 Perl_re_printf( aTHX_ "(GPOS)");
20273 Perl_re_printf( aTHX_ " ");
20275 if (r->intflags & PREGf_GPOS_SEEN)
20276 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
20277 if (r->intflags & PREGf_SKIP)
20278 Perl_re_printf( aTHX_ "plus ");
20279 if (r->intflags & PREGf_IMPLICIT)
20280 Perl_re_printf( aTHX_ "implicit ");
20281 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
20282 if (r->extflags & RXf_EVAL_SEEN)
20283 Perl_re_printf( aTHX_ "with eval ");
20284 Perl_re_printf( aTHX_ "\n");
20286 regdump_extflags("r->extflags: ", r->extflags);
20287 regdump_intflags("r->intflags: ", r->intflags);
20290 PERL_ARGS_ASSERT_REGDUMP;
20291 PERL_UNUSED_CONTEXT;
20292 PERL_UNUSED_ARG(r);
20293 #endif /* DEBUGGING */
20296 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
20299 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
20300 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
20301 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
20302 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
20303 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
20304 || _CC_VERTSPACE != 15
20305 # error Need to adjust order of anyofs[]
20307 static const char * const anyofs[] = {
20344 - regprop - printable representation of opcode, with run time support
20348 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
20353 RXi_GET_DECL(prog, progi);
20354 GET_RE_DEBUG_FLAGS_DECL;
20356 PERL_ARGS_ASSERT_REGPROP;
20360 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
20361 if (pRExC_state) { /* This gives more info, if we have it */
20362 FAIL3("panic: corrupted regexp opcode %d > %d",
20363 (int)OP(o), (int)REGNODE_MAX);
20366 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
20367 (int)OP(o), (int)REGNODE_MAX);
20370 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
20372 k = PL_regkind[OP(o)];
20375 sv_catpvs(sv, " ");
20376 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
20377 * is a crude hack but it may be the best for now since
20378 * we have no flag "this EXACTish node was UTF-8"
20380 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
20381 PL_colors[0], PL_colors[1],
20382 PERL_PV_ESCAPE_UNI_DETECT |
20383 PERL_PV_ESCAPE_NONASCII |
20384 PERL_PV_PRETTY_ELLIPSES |
20385 PERL_PV_PRETTY_LTGT |
20386 PERL_PV_PRETTY_NOCLEAR
20388 } else if (k == TRIE) {
20389 /* print the details of the trie in dumpuntil instead, as
20390 * progi->data isn't available here */
20391 const char op = OP(o);
20392 const U32 n = ARG(o);
20393 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
20394 (reg_ac_data *)progi->data->data[n] :
20396 const reg_trie_data * const trie
20397 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
20399 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
20400 DEBUG_TRIE_COMPILE_r({
20402 sv_catpvs(sv, "(JUMP)");
20403 Perl_sv_catpvf(aTHX_ sv,
20404 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
20405 (UV)trie->startstate,
20406 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
20407 (UV)trie->wordcount,
20410 (UV)TRIE_CHARCOUNT(trie),
20411 (UV)trie->uniquecharcount
20414 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
20415 sv_catpvs(sv, "[");
20416 (void) put_charclass_bitmap_innards(sv,
20417 ((IS_ANYOF_TRIE(op))
20419 : TRIE_BITMAP(trie)),
20425 sv_catpvs(sv, "]");
20427 } else if (k == CURLY) {
20428 U32 lo = ARG1(o), hi = ARG2(o);
20429 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
20430 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
20431 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
20432 if (hi == REG_INFTY)
20433 sv_catpvs(sv, "INFTY");
20435 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
20436 sv_catpvs(sv, "}");
20438 else if (k == WHILEM && o->flags) /* Ordinal/of */
20439 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
20440 else if (k == REF || k == OPEN || k == CLOSE
20441 || k == GROUPP || OP(o)==ACCEPT)
20443 AV *name_list= NULL;
20444 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
20445 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
20446 if ( RXp_PAREN_NAMES(prog) ) {
20447 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20448 } else if ( pRExC_state ) {
20449 name_list= RExC_paren_name_list;
20452 if ( k != REF || (OP(o) < REFN)) {
20453 SV **name= av_fetch(name_list, parno, 0 );
20455 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20458 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
20459 I32 *nums=(I32*)SvPVX(sv_dat);
20460 SV **name= av_fetch(name_list, nums[0], 0 );
20463 for ( n=0; n<SvIVX(sv_dat); n++ ) {
20464 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
20465 (n ? "," : ""), (IV)nums[n]);
20467 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20471 if ( k == REF && reginfo) {
20472 U32 n = ARG(o); /* which paren pair */
20473 I32 ln = prog->offs[n].start;
20474 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
20475 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
20476 else if (ln == prog->offs[n].end)
20477 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
20479 const char *s = reginfo->strbeg + ln;
20480 Perl_sv_catpvf(aTHX_ sv, ": ");
20481 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
20482 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
20485 } else if (k == GOSUB) {
20486 AV *name_list= NULL;
20487 if ( RXp_PAREN_NAMES(prog) ) {
20488 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20489 } else if ( pRExC_state ) {
20490 name_list= RExC_paren_name_list;
20493 /* Paren and offset */
20494 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
20495 (int)((o + (int)ARG2L(o)) - progi->program) );
20497 SV **name= av_fetch(name_list, ARG(o), 0 );
20499 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20502 else if (k == LOGICAL)
20503 /* 2: embedded, otherwise 1 */
20504 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
20505 else if (k == ANYOF) {
20506 const U8 flags = inRANGE(OP(o), ANYOFH, ANYOFHr)
20509 bool do_sep = FALSE; /* Do we need to separate various components of
20511 /* Set if there is still an unresolved user-defined property */
20512 SV *unresolved = NULL;
20514 /* Things that are ignored except when the runtime locale is UTF-8 */
20515 SV *only_utf8_locale_invlist = NULL;
20517 /* Code points that don't fit in the bitmap */
20518 SV *nonbitmap_invlist = NULL;
20520 /* And things that aren't in the bitmap, but are small enough to be */
20521 SV* bitmap_range_not_in_bitmap = NULL;
20523 const bool inverted = flags & ANYOF_INVERT;
20525 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
20526 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
20527 sv_catpvs(sv, "{utf8-locale-reqd}");
20529 if (flags & ANYOFL_FOLD) {
20530 sv_catpvs(sv, "{i}");
20534 /* If there is stuff outside the bitmap, get it */
20535 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
20536 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
20538 &only_utf8_locale_invlist,
20539 &nonbitmap_invlist);
20540 /* The non-bitmap data may contain stuff that could fit in the
20541 * bitmap. This could come from a user-defined property being
20542 * finally resolved when this call was done; or much more likely
20543 * because there are matches that require UTF-8 to be valid, and so
20544 * aren't in the bitmap. This is teased apart later */
20545 _invlist_intersection(nonbitmap_invlist,
20547 &bitmap_range_not_in_bitmap);
20548 /* Leave just the things that don't fit into the bitmap */
20549 _invlist_subtract(nonbitmap_invlist,
20551 &nonbitmap_invlist);
20554 /* Obey this flag to add all above-the-bitmap code points */
20555 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
20556 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
20557 NUM_ANYOF_CODE_POINTS,
20561 /* Ready to start outputting. First, the initial left bracket */
20562 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20564 if (! inRANGE(OP(o), ANYOFH, ANYOFHr)) {
20565 /* Then all the things that could fit in the bitmap */
20566 do_sep = put_charclass_bitmap_innards(sv,
20568 bitmap_range_not_in_bitmap,
20569 only_utf8_locale_invlist,
20572 /* Can't try inverting for a
20573 * better display if there
20574 * are things that haven't
20576 unresolved != NULL);
20577 SvREFCNT_dec(bitmap_range_not_in_bitmap);
20579 /* If there are user-defined properties which haven't been defined
20580 * yet, output them. If the result is not to be inverted, it is
20581 * clearest to output them in a separate [] from the bitmap range
20582 * stuff. If the result is to be complemented, we have to show
20583 * everything in one [], as the inversion applies to the whole
20584 * thing. Use {braces} to separate them from anything in the
20585 * bitmap and anything above the bitmap. */
20588 if (! do_sep) { /* If didn't output anything in the bitmap
20590 sv_catpvs(sv, "^");
20592 sv_catpvs(sv, "{");
20595 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
20598 sv_catsv(sv, unresolved);
20600 sv_catpvs(sv, "}");
20602 do_sep = ! inverted;
20606 /* And, finally, add the above-the-bitmap stuff */
20607 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
20610 /* See if truncation size is overridden */
20611 const STRLEN dump_len = (PL_dump_re_max_len > 256)
20612 ? PL_dump_re_max_len
20615 /* This is output in a separate [] */
20617 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
20620 /* And, for easy of understanding, it is shown in the
20621 * uncomplemented form if possible. The one exception being if
20622 * there are unresolved items, where the inversion has to be
20623 * delayed until runtime */
20624 if (inverted && ! unresolved) {
20625 _invlist_invert(nonbitmap_invlist);
20626 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
20629 contents = invlist_contents(nonbitmap_invlist,
20630 FALSE /* output suitable for catsv */
20633 /* If the output is shorter than the permissible maximum, just do it. */
20634 if (SvCUR(contents) <= dump_len) {
20635 sv_catsv(sv, contents);
20638 const char * contents_string = SvPVX(contents);
20639 STRLEN i = dump_len;
20641 /* Otherwise, start at the permissible max and work back to the
20642 * first break possibility */
20643 while (i > 0 && contents_string[i] != ' ') {
20646 if (i == 0) { /* Fail-safe. Use the max if we couldn't
20647 find a legal break */
20651 sv_catpvn(sv, contents_string, i);
20652 sv_catpvs(sv, "...");
20655 SvREFCNT_dec_NN(contents);
20656 SvREFCNT_dec_NN(nonbitmap_invlist);
20659 /* And finally the matching, closing ']' */
20660 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20662 if (inRANGE(OP(o), ANYOFH, ANYOFHr)) {
20663 U8 lowest = (OP(o) != ANYOFHr)
20665 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
20666 U8 highest = (OP(o) == ANYOFHb)
20670 : HIGHEST_ANYOF_HRx_BYTE(FLAGS(o));
20671 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
20672 if (lowest != highest) {
20673 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
20675 Perl_sv_catpvf(aTHX_ sv, ")");
20678 SvREFCNT_dec(unresolved);
20680 else if (k == ANYOFM) {
20681 SV * cp_list = get_ANYOFM_contents(o);
20683 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20684 if (OP(o) == NANYOFM) {
20685 _invlist_invert(cp_list);
20688 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, TRUE);
20689 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
20691 SvREFCNT_dec(cp_list);
20693 else if (k == POSIXD || k == NPOSIXD) {
20694 U8 index = FLAGS(o) * 2;
20695 if (index < C_ARRAY_LENGTH(anyofs)) {
20696 if (*anyofs[index] != '[') {
20697 sv_catpvs(sv, "[");
20699 sv_catpv(sv, anyofs[index]);
20700 if (*anyofs[index] != '[') {
20701 sv_catpvs(sv, "]");
20705 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
20708 else if (k == BOUND || k == NBOUND) {
20709 /* Must be synced with order of 'bound_type' in regcomp.h */
20710 const char * const bounds[] = {
20711 "", /* Traditional */
20717 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
20718 sv_catpv(sv, bounds[FLAGS(o)]);
20720 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
20721 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
20723 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
20725 Perl_sv_catpvf(aTHX_ sv, "]");
20727 else if (OP(o) == SBOL)
20728 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
20730 /* add on the verb argument if there is one */
20731 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
20733 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
20734 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
20736 sv_catpvs(sv, ":NULL");
20739 PERL_UNUSED_CONTEXT;
20740 PERL_UNUSED_ARG(sv);
20741 PERL_UNUSED_ARG(o);
20742 PERL_UNUSED_ARG(prog);
20743 PERL_UNUSED_ARG(reginfo);
20744 PERL_UNUSED_ARG(pRExC_state);
20745 #endif /* DEBUGGING */
20751 Perl_re_intuit_string(pTHX_ REGEXP * const r)
20752 { /* Assume that RE_INTUIT is set */
20753 struct regexp *const prog = ReANY(r);
20754 GET_RE_DEBUG_FLAGS_DECL;
20756 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
20757 PERL_UNUSED_CONTEXT;
20761 const char * const s = SvPV_nolen_const(RX_UTF8(r)
20762 ? prog->check_utf8 : prog->check_substr);
20764 if (!PL_colorset) reginitcolors();
20765 Perl_re_printf( aTHX_
20766 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
20768 RX_UTF8(r) ? "utf8 " : "",
20769 PL_colors[5], PL_colors[0],
20772 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
20775 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
20776 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
20782 handles refcounting and freeing the perl core regexp structure. When
20783 it is necessary to actually free the structure the first thing it
20784 does is call the 'free' method of the regexp_engine associated to
20785 the regexp, allowing the handling of the void *pprivate; member
20786 first. (This routine is not overridable by extensions, which is why
20787 the extensions free is called first.)
20789 See regdupe and regdupe_internal if you change anything here.
20791 #ifndef PERL_IN_XSUB_RE
20793 Perl_pregfree(pTHX_ REGEXP *r)
20799 Perl_pregfree2(pTHX_ REGEXP *rx)
20801 struct regexp *const r = ReANY(rx);
20802 GET_RE_DEBUG_FLAGS_DECL;
20804 PERL_ARGS_ASSERT_PREGFREE2;
20809 if (r->mother_re) {
20810 ReREFCNT_dec(r->mother_re);
20812 CALLREGFREE_PVT(rx); /* free the private data */
20813 SvREFCNT_dec(RXp_PAREN_NAMES(r));
20817 for (i = 0; i < 2; i++) {
20818 SvREFCNT_dec(r->substrs->data[i].substr);
20819 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
20821 Safefree(r->substrs);
20823 RX_MATCH_COPY_FREE(rx);
20824 #ifdef PERL_ANY_COW
20825 SvREFCNT_dec(r->saved_copy);
20828 SvREFCNT_dec(r->qr_anoncv);
20829 if (r->recurse_locinput)
20830 Safefree(r->recurse_locinput);
20836 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
20837 except that dsv will be created if NULL.
20839 This function is used in two main ways. First to implement
20840 $r = qr/....; $s = $$r;
20842 Secondly, it is used as a hacky workaround to the structural issue of
20844 being stored in the regexp structure which is in turn stored in
20845 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
20846 could be PL_curpm in multiple contexts, and could require multiple
20847 result sets being associated with the pattern simultaneously, such
20848 as when doing a recursive match with (??{$qr})
20850 The solution is to make a lightweight copy of the regexp structure
20851 when a qr// is returned from the code executed by (??{$qr}) this
20852 lightweight copy doesn't actually own any of its data except for
20853 the starp/end and the actual regexp structure itself.
20859 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
20861 struct regexp *drx;
20862 struct regexp *const srx = ReANY(ssv);
20863 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
20865 PERL_ARGS_ASSERT_REG_TEMP_COPY;
20868 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
20870 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
20872 /* our only valid caller, sv_setsv_flags(), should have done
20873 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
20874 assert(!SvOOK(dsv));
20875 assert(!SvIsCOW(dsv));
20876 assert(!SvROK(dsv));
20878 if (SvPVX_const(dsv)) {
20880 Safefree(SvPVX(dsv));
20885 SvOK_off((SV *)dsv);
20888 /* For PVLVs, the head (sv_any) points to an XPVLV, while
20889 * the LV's xpvlenu_rx will point to a regexp body, which
20890 * we allocate here */
20891 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
20892 assert(!SvPVX(dsv));
20893 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
20894 temp->sv_any = NULL;
20895 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
20896 SvREFCNT_dec_NN(temp);
20897 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
20898 ing below will not set it. */
20899 SvCUR_set(dsv, SvCUR(ssv));
20902 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
20903 sv_force_normal(sv) is called. */
20907 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
20908 SvPV_set(dsv, RX_WRAPPED(ssv));
20909 /* We share the same string buffer as the original regexp, on which we
20910 hold a reference count, incremented when mother_re is set below.
20911 The string pointer is copied here, being part of the regexp struct.
20913 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
20914 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
20918 const I32 npar = srx->nparens+1;
20919 Newx(drx->offs, npar, regexp_paren_pair);
20920 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
20922 if (srx->substrs) {
20924 Newx(drx->substrs, 1, struct reg_substr_data);
20925 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
20927 for (i = 0; i < 2; i++) {
20928 SvREFCNT_inc_void(drx->substrs->data[i].substr);
20929 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
20932 /* check_substr and check_utf8, if non-NULL, point to either their
20933 anchored or float namesakes, and don't hold a second reference. */
20935 RX_MATCH_COPIED_off(dsv);
20936 #ifdef PERL_ANY_COW
20937 drx->saved_copy = NULL;
20939 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
20940 SvREFCNT_inc_void(drx->qr_anoncv);
20941 if (srx->recurse_locinput)
20942 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
20949 /* regfree_internal()
20951 Free the private data in a regexp. This is overloadable by
20952 extensions. Perl takes care of the regexp structure in pregfree(),
20953 this covers the *pprivate pointer which technically perl doesn't
20954 know about, however of course we have to handle the
20955 regexp_internal structure when no extension is in use.
20957 Note this is called before freeing anything in the regexp
20962 Perl_regfree_internal(pTHX_ REGEXP * const rx)
20964 struct regexp *const r = ReANY(rx);
20965 RXi_GET_DECL(r, ri);
20966 GET_RE_DEBUG_FLAGS_DECL;
20968 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
20978 SV *dsv= sv_newmortal();
20979 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
20980 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
20981 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
20982 PL_colors[4], PL_colors[5], s);
20986 #ifdef RE_TRACK_PATTERN_OFFSETS
20988 Safefree(ri->u.offsets); /* 20010421 MJD */
20990 if (ri->code_blocks)
20991 S_free_codeblocks(aTHX_ ri->code_blocks);
20994 int n = ri->data->count;
20997 /* If you add a ->what type here, update the comment in regcomp.h */
20998 switch (ri->data->what[n]) {
21004 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21007 Safefree(ri->data->data[n]);
21013 { /* Aho Corasick add-on structure for a trie node.
21014 Used in stclass optimization only */
21016 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21017 #ifdef USE_ITHREADS
21021 refcount = --aho->refcount;
21024 PerlMemShared_free(aho->states);
21025 PerlMemShared_free(aho->fail);
21026 /* do this last!!!! */
21027 PerlMemShared_free(ri->data->data[n]);
21028 /* we should only ever get called once, so
21029 * assert as much, and also guard the free
21030 * which /might/ happen twice. At the least
21031 * it will make code anlyzers happy and it
21032 * doesn't cost much. - Yves */
21033 assert(ri->regstclass);
21034 if (ri->regstclass) {
21035 PerlMemShared_free(ri->regstclass);
21036 ri->regstclass = 0;
21043 /* trie structure. */
21045 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
21046 #ifdef USE_ITHREADS
21050 refcount = --trie->refcount;
21053 PerlMemShared_free(trie->charmap);
21054 PerlMemShared_free(trie->states);
21055 PerlMemShared_free(trie->trans);
21057 PerlMemShared_free(trie->bitmap);
21059 PerlMemShared_free(trie->jump);
21060 PerlMemShared_free(trie->wordinfo);
21061 /* do this last!!!! */
21062 PerlMemShared_free(ri->data->data[n]);
21067 Perl_croak(aTHX_ "panic: regfree data code '%c'",
21068 ri->data->what[n]);
21071 Safefree(ri->data->what);
21072 Safefree(ri->data);
21078 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
21079 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
21080 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
21083 re_dup_guts - duplicate a regexp.
21085 This routine is expected to clone a given regexp structure. It is only
21086 compiled under USE_ITHREADS.
21088 After all of the core data stored in struct regexp is duplicated
21089 the regexp_engine.dupe method is used to copy any private data
21090 stored in the *pprivate pointer. This allows extensions to handle
21091 any duplication it needs to do.
21093 See pregfree() and regfree_internal() if you change anything here.
21095 #if defined(USE_ITHREADS)
21096 #ifndef PERL_IN_XSUB_RE
21098 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
21102 const struct regexp *r = ReANY(sstr);
21103 struct regexp *ret = ReANY(dstr);
21105 PERL_ARGS_ASSERT_RE_DUP_GUTS;
21107 npar = r->nparens+1;
21108 Newx(ret->offs, npar, regexp_paren_pair);
21109 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
21111 if (ret->substrs) {
21112 /* Do it this way to avoid reading from *r after the StructCopy().
21113 That way, if any of the sv_dup_inc()s dislodge *r from the L1
21114 cache, it doesn't matter. */
21116 const bool anchored = r->check_substr
21117 ? r->check_substr == r->substrs->data[0].substr
21118 : r->check_utf8 == r->substrs->data[0].utf8_substr;
21119 Newx(ret->substrs, 1, struct reg_substr_data);
21120 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
21122 for (i = 0; i < 2; i++) {
21123 ret->substrs->data[i].substr =
21124 sv_dup_inc(ret->substrs->data[i].substr, param);
21125 ret->substrs->data[i].utf8_substr =
21126 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
21129 /* check_substr and check_utf8, if non-NULL, point to either their
21130 anchored or float namesakes, and don't hold a second reference. */
21132 if (ret->check_substr) {
21134 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
21136 ret->check_substr = ret->substrs->data[0].substr;
21137 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21139 assert(r->check_substr == r->substrs->data[1].substr);
21140 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
21142 ret->check_substr = ret->substrs->data[1].substr;
21143 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21145 } else if (ret->check_utf8) {
21147 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21149 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21154 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
21155 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
21156 if (r->recurse_locinput)
21157 Newx(ret->recurse_locinput, r->nparens + 1, char *);
21160 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
21162 if (RX_MATCH_COPIED(dstr))
21163 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
21165 ret->subbeg = NULL;
21166 #ifdef PERL_ANY_COW
21167 ret->saved_copy = NULL;
21170 /* Whether mother_re be set or no, we need to copy the string. We
21171 cannot refrain from copying it when the storage points directly to
21172 our mother regexp, because that's
21173 1: a buffer in a different thread
21174 2: something we no longer hold a reference on
21175 so we need to copy it locally. */
21176 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21177 /* set malloced length to a non-zero value so it will be freed
21178 * (otherwise in combination with SVf_FAKE it looks like an alien
21179 * buffer). It doesn't have to be the actual malloced size, since it
21180 * should never be grown */
21181 SvLEN_set(dstr, SvCUR(sstr)+1);
21182 ret->mother_re = NULL;
21184 #endif /* PERL_IN_XSUB_RE */
21189 This is the internal complement to regdupe() which is used to copy
21190 the structure pointed to by the *pprivate pointer in the regexp.
21191 This is the core version of the extension overridable cloning hook.
21192 The regexp structure being duplicated will be copied by perl prior
21193 to this and will be provided as the regexp *r argument, however
21194 with the /old/ structures pprivate pointer value. Thus this routine
21195 may override any copying normally done by perl.
21197 It returns a pointer to the new regexp_internal structure.
21201 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21204 struct regexp *const r = ReANY(rx);
21205 regexp_internal *reti;
21207 RXi_GET_DECL(r, ri);
21209 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21213 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
21214 char, regexp_internal);
21215 Copy(ri->program, reti->program, len+1, regnode);
21218 if (ri->code_blocks) {
21220 Newx(reti->code_blocks, 1, struct reg_code_blocks);
21221 Newx(reti->code_blocks->cb, ri->code_blocks->count,
21222 struct reg_code_block);
21223 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
21224 ri->code_blocks->count, struct reg_code_block);
21225 for (n = 0; n < ri->code_blocks->count; n++)
21226 reti->code_blocks->cb[n].src_regex = (REGEXP*)
21227 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
21228 reti->code_blocks->count = ri->code_blocks->count;
21229 reti->code_blocks->refcnt = 1;
21232 reti->code_blocks = NULL;
21234 reti->regstclass = NULL;
21237 struct reg_data *d;
21238 const int count = ri->data->count;
21241 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
21242 char, struct reg_data);
21243 Newx(d->what, count, U8);
21246 for (i = 0; i < count; i++) {
21247 d->what[i] = ri->data->what[i];
21248 switch (d->what[i]) {
21249 /* see also regcomp.h and regfree_internal() */
21250 case 'a': /* actually an AV, but the dup function is identical.
21251 values seem to be "plain sv's" generally. */
21252 case 'r': /* a compiled regex (but still just another SV) */
21253 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
21254 this use case should go away, the code could have used
21255 'a' instead - see S_set_ANYOF_arg() for array contents. */
21256 case 'S': /* actually an SV, but the dup function is identical. */
21257 case 'u': /* actually an HV, but the dup function is identical.
21258 values are "plain sv's" */
21259 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
21262 /* Synthetic Start Class - "Fake" charclass we generate to optimize
21263 * patterns which could start with several different things. Pre-TRIE
21264 * this was more important than it is now, however this still helps
21265 * in some places, for instance /x?a+/ might produce a SSC equivalent
21266 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
21269 /* This is cheating. */
21270 Newx(d->data[i], 1, regnode_ssc);
21271 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
21272 reti->regstclass = (regnode*)d->data[i];
21275 /* AHO-CORASICK fail table */
21276 /* Trie stclasses are readonly and can thus be shared
21277 * without duplication. We free the stclass in pregfree
21278 * when the corresponding reg_ac_data struct is freed.
21280 reti->regstclass= ri->regstclass;
21283 /* TRIE transition table */
21285 ((reg_trie_data*)ri->data->data[i])->refcount++;
21288 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
21289 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
21290 is not from another regexp */
21291 d->data[i] = ri->data->data[i];
21294 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
21295 ri->data->what[i]);
21304 reti->name_list_idx = ri->name_list_idx;
21306 #ifdef RE_TRACK_PATTERN_OFFSETS
21307 if (ri->u.offsets) {
21308 Newx(reti->u.offsets, 2*len+1, U32);
21309 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
21312 SetProgLen(reti, len);
21315 return (void*)reti;
21318 #endif /* USE_ITHREADS */
21320 #ifndef PERL_IN_XSUB_RE
21323 - regnext - dig the "next" pointer out of a node
21326 Perl_regnext(pTHX_ regnode *p)
21333 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
21334 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
21335 (int)OP(p), (int)REGNODE_MAX);
21338 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
21348 S_re_croak2(pTHX_ bool utf8, const char* pat1, const char* pat2,...)
21351 STRLEN l1 = strlen(pat1);
21352 STRLEN l2 = strlen(pat2);
21355 const char *message;
21357 PERL_ARGS_ASSERT_RE_CROAK2;
21363 Copy(pat1, buf, l1 , char);
21364 Copy(pat2, buf + l1, l2 , char);
21365 buf[l1 + l2] = '\n';
21366 buf[l1 + l2 + 1] = '\0';
21367 va_start(args, pat2);
21368 msv = vmess(buf, &args);
21370 message = SvPV_const(msv, l1);
21373 Copy(message, buf, l1 , char);
21374 /* l1-1 to avoid \n */
21375 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
21378 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
21380 #ifndef PERL_IN_XSUB_RE
21382 Perl_save_re_context(pTHX)
21387 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
21390 const REGEXP * const rx = PM_GETRE(PL_curpm);
21392 nparens = RX_NPARENS(rx);
21395 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
21396 * that PL_curpm will be null, but that utf8.pm and the modules it
21397 * loads will only use $1..$3.
21398 * The t/porting/re_context.t test file checks this assumption.
21403 for (i = 1; i <= nparens; i++) {
21404 char digits[TYPE_CHARS(long)];
21405 const STRLEN len = my_snprintf(digits, sizeof(digits),
21407 GV *const *const gvp
21408 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
21411 GV * const gv = *gvp;
21412 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
21422 S_put_code_point(pTHX_ SV *sv, UV c)
21424 PERL_ARGS_ASSERT_PUT_CODE_POINT;
21427 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
21429 else if (isPRINT(c)) {
21430 const char string = (char) c;
21432 /* We use {phrase} as metanotation in the class, so also escape literal
21434 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
21435 sv_catpvs(sv, "\\");
21436 sv_catpvn(sv, &string, 1);
21438 else if (isMNEMONIC_CNTRL(c)) {
21439 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
21442 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
21446 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
21449 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
21451 /* Appends to 'sv' a displayable version of the range of code points from
21452 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
21453 * that have them, when they occur at the beginning or end of the range.
21454 * It uses hex to output the remaining code points, unless 'allow_literals'
21455 * is true, in which case the printable ASCII ones are output as-is (though
21456 * some of these will be escaped by put_code_point()).
21458 * NOTE: This is designed only for printing ranges of code points that fit
21459 * inside an ANYOF bitmap. Higher code points are simply suppressed
21462 const unsigned int min_range_count = 3;
21464 assert(start <= end);
21466 PERL_ARGS_ASSERT_PUT_RANGE;
21468 while (start <= end) {
21470 const char * format;
21472 if (end - start < min_range_count) {
21474 /* Output chars individually when they occur in short ranges */
21475 for (; start <= end; start++) {
21476 put_code_point(sv, start);
21481 /* If permitted by the input options, and there is a possibility that
21482 * this range contains a printable literal, look to see if there is
21484 if (allow_literals && start <= MAX_PRINT_A) {
21486 /* If the character at the beginning of the range isn't an ASCII
21487 * printable, effectively split the range into two parts:
21488 * 1) the portion before the first such printable,
21490 * and output them separately. */
21491 if (! isPRINT_A(start)) {
21492 UV temp_end = start + 1;
21494 /* There is no point looking beyond the final possible
21495 * printable, in MAX_PRINT_A */
21496 UV max = MIN(end, MAX_PRINT_A);
21498 while (temp_end <= max && ! isPRINT_A(temp_end)) {
21502 /* Here, temp_end points to one beyond the first printable if
21503 * found, or to one beyond 'max' if not. If none found, make
21504 * sure that we use the entire range */
21505 if (temp_end > MAX_PRINT_A) {
21506 temp_end = end + 1;
21509 /* Output the first part of the split range: the part that
21510 * doesn't have printables, with the parameter set to not look
21511 * for literals (otherwise we would infinitely recurse) */
21512 put_range(sv, start, temp_end - 1, FALSE);
21514 /* The 2nd part of the range (if any) starts here. */
21517 /* We do a continue, instead of dropping down, because even if
21518 * the 2nd part is non-empty, it could be so short that we want
21519 * to output it as individual characters, as tested for at the
21520 * top of this loop. */
21524 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
21525 * output a sub-range of just the digits or letters, then process
21526 * the remaining portion as usual. */
21527 if (isALPHANUMERIC_A(start)) {
21528 UV mask = (isDIGIT_A(start))
21533 UV temp_end = start + 1;
21535 /* Find the end of the sub-range that includes just the
21536 * characters in the same class as the first character in it */
21537 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
21542 /* For short ranges, don't duplicate the code above to output
21543 * them; just call recursively */
21544 if (temp_end - start < min_range_count) {
21545 put_range(sv, start, temp_end, FALSE);
21547 else { /* Output as a range */
21548 put_code_point(sv, start);
21549 sv_catpvs(sv, "-");
21550 put_code_point(sv, temp_end);
21552 start = temp_end + 1;
21556 /* We output any other printables as individual characters */
21557 if (isPUNCT_A(start) || isSPACE_A(start)) {
21558 while (start <= end && (isPUNCT_A(start)
21559 || isSPACE_A(start)))
21561 put_code_point(sv, start);
21566 } /* End of looking for literals */
21568 /* Here is not to output as a literal. Some control characters have
21569 * mnemonic names. Split off any of those at the beginning and end of
21570 * the range to print mnemonically. It isn't possible for many of
21571 * these to be in a row, so this won't overwhelm with output */
21573 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
21575 while (isMNEMONIC_CNTRL(start) && start <= end) {
21576 put_code_point(sv, start);
21580 /* If this didn't take care of the whole range ... */
21581 if (start <= end) {
21583 /* Look backwards from the end to find the final non-mnemonic
21586 while (isMNEMONIC_CNTRL(temp_end)) {
21590 /* And separately output the interior range that doesn't start
21591 * or end with mnemonics */
21592 put_range(sv, start, temp_end, FALSE);
21594 /* Then output the mnemonic trailing controls */
21595 start = temp_end + 1;
21596 while (start <= end) {
21597 put_code_point(sv, start);
21604 /* As a final resort, output the range or subrange as hex. */
21606 if (start >= NUM_ANYOF_CODE_POINTS) {
21610 this_end = (end < NUM_ANYOF_CODE_POINTS)
21612 : NUM_ANYOF_CODE_POINTS - 1;
21614 #if NUM_ANYOF_CODE_POINTS > 256
21615 format = (this_end < 256)
21616 ? "\\x%02" UVXf "-\\x%02" UVXf
21617 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
21619 format = "\\x%02" UVXf "-\\x%02" UVXf;
21621 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
21622 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
21623 GCC_DIAG_RESTORE_STMT;
21629 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
21631 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
21635 bool allow_literals = TRUE;
21637 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
21639 /* Generally, it is more readable if printable characters are output as
21640 * literals, but if a range (nearly) spans all of them, it's best to output
21641 * it as a single range. This code will use a single range if all but 2
21642 * ASCII printables are in it */
21643 invlist_iterinit(invlist);
21644 while (invlist_iternext(invlist, &start, &end)) {
21646 /* If the range starts beyond the final printable, it doesn't have any
21648 if (start > MAX_PRINT_A) {
21652 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
21653 * all but two, the range must start and end no later than 2 from
21655 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
21656 if (end > MAX_PRINT_A) {
21662 if (end - start >= MAX_PRINT_A - ' ' - 2) {
21663 allow_literals = FALSE;
21668 invlist_iterfinish(invlist);
21670 /* Here we have figured things out. Output each range */
21671 invlist_iterinit(invlist);
21672 while (invlist_iternext(invlist, &start, &end)) {
21673 if (start >= NUM_ANYOF_CODE_POINTS) {
21676 put_range(sv, start, end, allow_literals);
21678 invlist_iterfinish(invlist);
21684 S_put_charclass_bitmap_innards_common(pTHX_
21685 SV* invlist, /* The bitmap */
21686 SV* posixes, /* Under /l, things like [:word:], \S */
21687 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
21688 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
21689 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
21690 const bool invert /* Is the result to be inverted? */
21693 /* Create and return an SV containing a displayable version of the bitmap
21694 * and associated information determined by the input parameters. If the
21695 * output would have been only the inversion indicator '^', NULL is instead
21701 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
21704 output = newSVpvs("^");
21707 output = newSVpvs("");
21710 /* First, the code points in the bitmap that are unconditionally there */
21711 put_charclass_bitmap_innards_invlist(output, invlist);
21713 /* Traditionally, these have been placed after the main code points */
21715 sv_catsv(output, posixes);
21718 if (only_utf8 && _invlist_len(only_utf8)) {
21719 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
21720 put_charclass_bitmap_innards_invlist(output, only_utf8);
21723 if (not_utf8 && _invlist_len(not_utf8)) {
21724 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
21725 put_charclass_bitmap_innards_invlist(output, not_utf8);
21728 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
21729 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
21730 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
21732 /* This is the only list in this routine that can legally contain code
21733 * points outside the bitmap range. The call just above to
21734 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
21735 * output them here. There's about a half-dozen possible, and none in
21736 * contiguous ranges longer than 2 */
21737 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21739 SV* above_bitmap = NULL;
21741 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
21743 invlist_iterinit(above_bitmap);
21744 while (invlist_iternext(above_bitmap, &start, &end)) {
21747 for (i = start; i <= end; i++) {
21748 put_code_point(output, i);
21751 invlist_iterfinish(above_bitmap);
21752 SvREFCNT_dec_NN(above_bitmap);
21756 if (invert && SvCUR(output) == 1) {
21764 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
21766 SV *nonbitmap_invlist,
21767 SV *only_utf8_locale_invlist,
21768 const regnode * const node,
21769 const bool force_as_is_display)
21771 /* Appends to 'sv' a displayable version of the innards of the bracketed
21772 * character class defined by the other arguments:
21773 * 'bitmap' points to the bitmap, or NULL if to ignore that.
21774 * 'nonbitmap_invlist' is an inversion list of the code points that are in
21775 * the bitmap range, but for some reason aren't in the bitmap; NULL if
21776 * none. The reasons for this could be that they require some
21777 * condition such as the target string being or not being in UTF-8
21778 * (under /d), or because they came from a user-defined property that
21779 * was not resolved at the time of the regex compilation (under /u)
21780 * 'only_utf8_locale_invlist' is an inversion list of the code points that
21781 * are valid only if the runtime locale is a UTF-8 one; NULL if none
21782 * 'node' is the regex pattern ANYOF node. It is needed only when the
21783 * above two parameters are not null, and is passed so that this
21784 * routine can tease apart the various reasons for them.
21785 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
21786 * to invert things to see if that leads to a cleaner display. If
21787 * FALSE, this routine is free to use its judgment about doing this.
21789 * It returns TRUE if there was actually something output. (It may be that
21790 * the bitmap, etc is empty.)
21792 * When called for outputting the bitmap of a non-ANYOF node, just pass the
21793 * bitmap, with the succeeding parameters set to NULL, and the final one to
21797 /* In general, it tries to display the 'cleanest' representation of the
21798 * innards, choosing whether to display them inverted or not, regardless of
21799 * whether the class itself is to be inverted. However, there are some
21800 * cases where it can't try inverting, as what actually matches isn't known
21801 * until runtime, and hence the inversion isn't either. */
21804 bool inverting_allowed = ! force_as_is_display;
21807 STRLEN orig_sv_cur = SvCUR(sv);
21809 SV* invlist; /* Inversion list we accumulate of code points that
21810 are unconditionally matched */
21811 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
21813 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
21815 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
21816 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
21819 SV* as_is_display; /* The output string when we take the inputs
21821 SV* inverted_display; /* The output string when we invert the inputs */
21823 U8 flags = (node) ? ANYOF_FLAGS(node) : 0;
21825 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
21827 /* We are biased in favor of displaying things without them being inverted,
21828 * as that is generally easier to understand */
21829 const int bias = 5;
21831 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
21833 /* Start off with whatever code points are passed in. (We clone, so we
21834 * don't change the caller's list) */
21835 if (nonbitmap_invlist) {
21836 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
21837 invlist = invlist_clone(nonbitmap_invlist, NULL);
21839 else { /* Worst case size is every other code point is matched */
21840 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
21844 if (OP(node) == ANYOFD) {
21846 /* This flag indicates that the code points below 0x100 in the
21847 * nonbitmap list are precisely the ones that match only when the
21848 * target is UTF-8 (they should all be non-ASCII). */
21849 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
21851 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
21852 _invlist_subtract(invlist, only_utf8, &invlist);
21855 /* And this flag for matching all non-ASCII 0xFF and below */
21856 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
21858 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
21861 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
21863 /* If either of these flags are set, what matches isn't
21864 * determinable except during execution, so don't know enough here
21866 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
21867 inverting_allowed = FALSE;
21870 /* What the posix classes match also varies at runtime, so these
21871 * will be output symbolically. */
21872 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
21875 posixes = newSVpvs("");
21876 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
21877 if (ANYOF_POSIXL_TEST(node, i)) {
21878 sv_catpv(posixes, anyofs[i]);
21885 /* Accumulate the bit map into the unconditional match list */
21887 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
21888 if (BITMAP_TEST(bitmap, i)) {
21891 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
21894 invlist = _add_range_to_invlist(invlist, start, i-1);
21899 /* Make sure that the conditional match lists don't have anything in them
21900 * that match unconditionally; otherwise the output is quite confusing.
21901 * This could happen if the code that populates these misses some
21904 _invlist_subtract(only_utf8, invlist, &only_utf8);
21907 _invlist_subtract(not_utf8, invlist, ¬_utf8);
21910 if (only_utf8_locale_invlist) {
21912 /* Since this list is passed in, we have to make a copy before
21914 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
21916 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
21918 /* And, it can get really weird for us to try outputting an inverted
21919 * form of this list when it has things above the bitmap, so don't even
21921 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
21922 inverting_allowed = FALSE;
21926 /* Calculate what the output would be if we take the input as-is */
21927 as_is_display = put_charclass_bitmap_innards_common(invlist,
21934 /* If have to take the output as-is, just do that */
21935 if (! inverting_allowed) {
21936 if (as_is_display) {
21937 sv_catsv(sv, as_is_display);
21938 SvREFCNT_dec_NN(as_is_display);
21941 else { /* But otherwise, create the output again on the inverted input, and
21942 use whichever version is shorter */
21944 int inverted_bias, as_is_bias;
21946 /* We will apply our bias to whichever of the the results doesn't have
21956 inverted_bias = bias;
21959 /* Now invert each of the lists that contribute to the output,
21960 * excluding from the result things outside the possible range */
21962 /* For the unconditional inversion list, we have to add in all the
21963 * conditional code points, so that when inverted, they will be gone
21965 _invlist_union(only_utf8, invlist, &invlist);
21966 _invlist_union(not_utf8, invlist, &invlist);
21967 _invlist_union(only_utf8_locale, invlist, &invlist);
21968 _invlist_invert(invlist);
21969 _invlist_intersection(invlist, PL_InBitmap, &invlist);
21972 _invlist_invert(only_utf8);
21973 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
21975 else if (not_utf8) {
21977 /* If a code point matches iff the target string is not in UTF-8,
21978 * then complementing the result has it not match iff not in UTF-8,
21979 * which is the same thing as matching iff it is UTF-8. */
21980 only_utf8 = not_utf8;
21984 if (only_utf8_locale) {
21985 _invlist_invert(only_utf8_locale);
21986 _invlist_intersection(only_utf8_locale,
21988 &only_utf8_locale);
21991 inverted_display = put_charclass_bitmap_innards_common(
21996 only_utf8_locale, invert);
21998 /* Use the shortest representation, taking into account our bias
21999 * against showing it inverted */
22000 if ( inverted_display
22001 && ( ! as_is_display
22002 || ( SvCUR(inverted_display) + inverted_bias
22003 < SvCUR(as_is_display) + as_is_bias)))
22005 sv_catsv(sv, inverted_display);
22007 else if (as_is_display) {
22008 sv_catsv(sv, as_is_display);
22011 SvREFCNT_dec(as_is_display);
22012 SvREFCNT_dec(inverted_display);
22015 SvREFCNT_dec_NN(invlist);
22016 SvREFCNT_dec(only_utf8);
22017 SvREFCNT_dec(not_utf8);
22018 SvREFCNT_dec(posixes);
22019 SvREFCNT_dec(only_utf8_locale);
22021 return SvCUR(sv) > orig_sv_cur;
22024 #define CLEAR_OPTSTART \
22025 if (optstart) STMT_START { \
22026 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22027 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22031 #define DUMPUNTIL(b,e) \
22033 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
22035 STATIC const regnode *
22036 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
22037 const regnode *last, const regnode *plast,
22038 SV* sv, I32 indent, U32 depth)
22040 U8 op = PSEUDO; /* Arbitrary non-END op. */
22041 const regnode *next;
22042 const regnode *optstart= NULL;
22044 RXi_GET_DECL(r, ri);
22045 GET_RE_DEBUG_FLAGS_DECL;
22047 PERL_ARGS_ASSERT_DUMPUNTIL;
22049 #ifdef DEBUG_DUMPUNTIL
22050 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
22051 last ? last-start : 0, plast ? plast-start : 0);
22054 if (plast && plast < last)
22057 while (PL_regkind[op] != END && (!last || node < last)) {
22059 /* While that wasn't END last time... */
22062 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
22064 next = regnext((regnode *)node);
22067 if (OP(node) == OPTIMIZED) {
22068 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
22075 regprop(r, sv, node, NULL, NULL);
22076 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
22077 (int)(2*indent + 1), "", SvPVX_const(sv));
22079 if (OP(node) != OPTIMIZED) {
22080 if (next == NULL) /* Next ptr. */
22081 Perl_re_printf( aTHX_ " (0)");
22082 else if (PL_regkind[(U8)op] == BRANCH
22083 && PL_regkind[OP(next)] != BRANCH )
22084 Perl_re_printf( aTHX_ " (FAIL)");
22086 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
22087 Perl_re_printf( aTHX_ "\n");
22091 if (PL_regkind[(U8)op] == BRANCHJ) {
22094 const regnode *nnode = (OP(next) == LONGJMP
22095 ? regnext((regnode *)next)
22097 if (last && nnode > last)
22099 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
22102 else if (PL_regkind[(U8)op] == BRANCH) {
22104 DUMPUNTIL(NEXTOPER(node), next);
22106 else if ( PL_regkind[(U8)op] == TRIE ) {
22107 const regnode *this_trie = node;
22108 const char op = OP(node);
22109 const U32 n = ARG(node);
22110 const reg_ac_data * const ac = op>=AHOCORASICK ?
22111 (reg_ac_data *)ri->data->data[n] :
22113 const reg_trie_data * const trie =
22114 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
22116 AV *const trie_words
22117 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
22119 const regnode *nextbranch= NULL;
22122 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
22123 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
22125 Perl_re_indentf( aTHX_ "%s ",
22128 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
22129 SvCUR(*elem_ptr), PL_dump_re_max_len,
22130 PL_colors[0], PL_colors[1],
22132 ? PERL_PV_ESCAPE_UNI
22134 | PERL_PV_PRETTY_ELLIPSES
22135 | PERL_PV_PRETTY_LTGT
22140 U16 dist= trie->jump[word_idx+1];
22141 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
22142 (UV)((dist ? this_trie + dist : next) - start));
22145 nextbranch= this_trie + trie->jump[0];
22146 DUMPUNTIL(this_trie + dist, nextbranch);
22148 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
22149 nextbranch= regnext((regnode *)nextbranch);
22151 Perl_re_printf( aTHX_ "\n");
22154 if (last && next > last)
22159 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
22160 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
22161 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
22163 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
22165 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
22167 else if ( op == PLUS || op == STAR) {
22168 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
22170 else if (PL_regkind[(U8)op] == EXACT) {
22171 /* Literal string, where present. */
22172 node += NODE_SZ_STR(node) - 1;
22173 node = NEXTOPER(node);
22176 node = NEXTOPER(node);
22177 node += regarglen[(U8)op];
22179 if (op == CURLYX || op == OPEN || op == SROPEN)
22183 #ifdef DEBUG_DUMPUNTIL
22184 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22189 #endif /* DEBUGGING */
22191 #ifndef PERL_IN_XSUB_RE
22193 #include "uni_keywords.h"
22196 Perl_init_uniprops(pTHX)
22200 PL_user_def_props = newHV();
22202 #ifdef USE_ITHREADS
22204 HvSHAREKEYS_off(PL_user_def_props);
22205 PL_user_def_props_aTHX = aTHX;
22209 /* Set up the inversion list global variables */
22211 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22212 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
22213 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
22214 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
22215 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
22216 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
22217 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
22218 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
22219 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
22220 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
22221 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
22222 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
22223 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
22224 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
22225 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
22226 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
22228 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22229 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
22230 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
22231 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
22232 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
22233 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
22234 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
22235 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
22236 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
22237 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
22238 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
22239 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
22240 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
22241 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
22242 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
22243 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
22245 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
22246 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
22247 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
22248 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
22249 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
22251 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
22252 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
22253 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
22255 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
22257 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
22258 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
22260 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
22261 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
22263 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
22264 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22265 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
22266 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22267 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
22268 PL_NonFinalFold = _new_invlist_C_array(uni_prop_ptrs[
22269 UNI__PERL_NON_FINAL_FOLDS]);
22271 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
22272 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
22273 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
22274 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
22275 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
22276 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
22277 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
22278 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
22279 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
22282 /* The below are used only by deprecated functions. They could be removed */
22283 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
22284 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
22285 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
22291 This code was mainly added for backcompat to give a warning for non-portable
22292 code points in user-defined properties. But experiments showed that the
22293 warning in earlier perls were only omitted on overflow, which should be an
22294 error, so there really isnt a backcompat issue, and actually adding the
22295 warning when none was present before might cause breakage, for little gain. So
22296 khw left this code in, but not enabled. Tests were never added.
22299 Ei |const char *|get_extended_utf8_msg|const UV cp
22301 PERL_STATIC_INLINE const char *
22302 S_get_extended_utf8_msg(pTHX_ const UV cp)
22304 U8 dummy[UTF8_MAXBYTES + 1];
22308 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
22311 msg = hv_fetchs(msgs, "text", 0);
22314 (void) sv_2mortal((SV *) msgs);
22316 return SvPVX(*msg);
22322 Perl_handle_user_defined_property(pTHX_
22324 /* Parses the contents of a user-defined property definition; returning the
22325 * expanded definition if possible. If so, the return is an inversion
22328 * If there are subroutines that are part of the expansion and which aren't
22329 * known at the time of the call to this function, this returns what
22330 * parse_uniprop_string() returned for the first one encountered.
22332 * If an error was found, NULL is returned, and 'msg' gets a suitable
22333 * message appended to it. (Appending allows the back trace of how we got
22334 * to the faulty definition to be displayed through nested calls of
22335 * user-defined subs.)
22337 * The caller IS responsible for freeing any returned SV.
22339 * The syntax of the contents is pretty much described in perlunicode.pod,
22340 * but we also allow comments on each line */
22342 const char * name, /* Name of property */
22343 const STRLEN name_len, /* The name's length in bytes */
22344 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22345 const bool to_fold, /* ? Is this under /i */
22346 const bool runtime, /* ? Are we in compile- or run-time */
22347 const bool deferrable, /* Is it ok for this property's full definition
22348 to be deferred until later? */
22349 SV* contents, /* The property's definition */
22350 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
22351 getting called unless this is thought to be
22352 a user-defined property */
22353 SV * msg, /* Any error or warning msg(s) are appended to
22355 const STRLEN level) /* Recursion level of this call */
22358 const char * string = SvPV_const(contents, len);
22359 const char * const e = string + len;
22360 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
22361 const STRLEN msgs_length_on_entry = SvCUR(msg);
22363 const char * s0 = string; /* Points to first byte in the current line
22364 being parsed in 'string' */
22365 const char overflow_msg[] = "Code point too large in \"";
22366 SV* running_definition = NULL;
22368 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
22370 *user_defined_ptr = TRUE;
22372 /* Look at each line */
22374 const char * s; /* Current byte */
22375 char op = '+'; /* Default operation is 'union' */
22376 IV min = 0; /* range begin code point */
22377 IV max = -1; /* and range end */
22378 SV* this_definition;
22380 /* Skip comment lines */
22382 s0 = strchr(s0, '\n');
22390 /* For backcompat, allow an empty first line */
22396 /* First character in the line may optionally be the operation */
22405 /* If the line is one or two hex digits separated by blank space, its
22406 * a range; otherwise it is either another user-defined property or an
22411 if (! isXDIGIT(*s)) {
22412 goto check_if_property;
22415 do { /* Each new hex digit will add 4 bits. */
22416 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
22417 s = strchr(s, '\n');
22421 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22422 sv_catpv(msg, overflow_msg);
22423 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22424 UTF8fARG(is_contents_utf8, s - s0, s0));
22425 sv_catpvs(msg, "\"");
22426 goto return_failure;
22429 /* Accumulate this digit into the value */
22430 min = (min << 4) + READ_XDIGIT(s);
22431 } while (isXDIGIT(*s));
22433 while (isBLANK(*s)) { s++; }
22435 /* We allow comments at the end of the line */
22437 s = strchr(s, '\n');
22443 else if (s < e && *s != '\n') {
22444 if (! isXDIGIT(*s)) {
22445 goto check_if_property;
22448 /* Look for the high point of the range */
22451 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
22452 s = strchr(s, '\n');
22456 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22457 sv_catpv(msg, overflow_msg);
22458 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22459 UTF8fARG(is_contents_utf8, s - s0, s0));
22460 sv_catpvs(msg, "\"");
22461 goto return_failure;
22464 max = (max << 4) + READ_XDIGIT(s);
22465 } while (isXDIGIT(*s));
22467 while (isBLANK(*s)) { s++; }
22470 s = strchr(s, '\n');
22475 else if (s < e && *s != '\n') {
22476 goto check_if_property;
22480 if (max == -1) { /* The line only had one entry */
22483 else if (max < min) {
22484 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22485 sv_catpvs(msg, "Illegal range in \"");
22486 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22487 UTF8fARG(is_contents_utf8, s - s0, s0));
22488 sv_catpvs(msg, "\"");
22489 goto return_failure;
22492 #if 0 /* See explanation at definition above of get_extended_utf8_msg() */
22494 if ( UNICODE_IS_PERL_EXTENDED(min)
22495 || UNICODE_IS_PERL_EXTENDED(max))
22497 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22499 /* If both code points are non-portable, warn only on the lower
22501 sv_catpv(msg, get_extended_utf8_msg(
22502 (UNICODE_IS_PERL_EXTENDED(min))
22504 sv_catpvs(msg, " in \"");
22505 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22506 UTF8fARG(is_contents_utf8, s - s0, s0));
22507 sv_catpvs(msg, "\"");
22512 /* Here, this line contains a legal range */
22513 this_definition = sv_2mortal(_new_invlist(2));
22514 this_definition = _add_range_to_invlist(this_definition, min, max);
22519 /* Here it isn't a legal range line. See if it is a legal property
22520 * line. First find the end of the meat of the line */
22521 s = strpbrk(s, "#\n");
22526 /* Ignore trailing blanks in keeping with the requirements of
22527 * parse_uniprop_string() */
22529 while (s > s0 && isBLANK_A(*s)) {
22534 this_definition = parse_uniprop_string(s0, s - s0,
22535 is_utf8, to_fold, runtime,
22537 user_defined_ptr, msg,
22539 ? level /* Don't increase level
22540 if input is empty */
22543 if (this_definition == NULL) {
22544 goto return_failure; /* 'msg' should have had the reason
22545 appended to it by the above call */
22548 if (! is_invlist(this_definition)) { /* Unknown at this time */
22549 return newSVsv(this_definition);
22553 s = strchr(s, '\n');
22563 _invlist_union(running_definition, this_definition,
22564 &running_definition);
22567 _invlist_subtract(running_definition, this_definition,
22568 &running_definition);
22571 _invlist_intersection(running_definition, this_definition,
22572 &running_definition);
22575 _invlist_union_complement_2nd(running_definition,
22576 this_definition, &running_definition);
22579 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
22580 __FILE__, __LINE__, op);
22584 /* Position past the '\n' */
22586 } /* End of loop through the lines of 'contents' */
22588 /* Here, we processed all the lines in 'contents' without error. If we
22589 * didn't add any warnings, simply return success */
22590 if (msgs_length_on_entry == SvCUR(msg)) {
22592 /* If the expansion was empty, the answer isn't nothing: its an empty
22593 * inversion list */
22594 if (running_definition == NULL) {
22595 running_definition = _new_invlist(1);
22598 return running_definition;
22601 /* Otherwise, add some explanatory text, but we will return success */
22605 running_definition = NULL;
22609 if (name_len > 0) {
22610 sv_catpvs(msg, " in expansion of ");
22611 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
22614 return running_definition;
22617 /* As explained below, certain operations need to take place in the first
22618 * thread created. These macros switch contexts */
22619 #ifdef USE_ITHREADS
22620 # define DECLARATION_FOR_GLOBAL_CONTEXT \
22621 PerlInterpreter * save_aTHX = aTHX;
22622 # define SWITCH_TO_GLOBAL_CONTEXT \
22623 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
22624 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
22625 # define CUR_CONTEXT aTHX
22626 # define ORIGINAL_CONTEXT save_aTHX
22628 # define DECLARATION_FOR_GLOBAL_CONTEXT
22629 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
22630 # define RESTORE_CONTEXT NOOP
22631 # define CUR_CONTEXT NULL
22632 # define ORIGINAL_CONTEXT NULL
22636 S_delete_recursion_entry(pTHX_ void *key)
22638 /* Deletes the entry used to detect recursion when expanding user-defined
22639 * properties. This is a function so it can be set up to be called even if
22640 * the program unexpectedly quits */
22643 SV ** current_entry;
22644 const STRLEN key_len = strlen((const char *) key);
22645 DECLARATION_FOR_GLOBAL_CONTEXT;
22647 SWITCH_TO_GLOBAL_CONTEXT;
22649 /* If the entry is one of these types, it is a permanent entry, and not the
22650 * one used to detect recursions. This function should delete only the
22651 * recursion entry */
22652 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
22654 && ! is_invlist(*current_entry)
22655 && ! SvPOK(*current_entry))
22657 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
22665 S_get_fq_name(pTHX_
22666 const char * const name, /* The first non-blank in the \p{}, \P{} */
22667 const Size_t name_len, /* Its length in bytes, not including any trailing space */
22668 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22669 const bool has_colon_colon
22672 /* Returns a mortal SV containing the fully qualified version of the input
22677 fq_name = newSVpvs_flags("", SVs_TEMP);
22679 /* Use the current package if it wasn't included in our input */
22680 if (! has_colon_colon) {
22681 const HV * pkg = (IN_PERL_COMPILETIME)
22683 : CopSTASH(PL_curcop);
22684 const char* pkgname = HvNAME(pkg);
22686 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22687 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
22688 sv_catpvs(fq_name, "::");
22691 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
22692 UTF8fARG(is_utf8, name_len, name));
22697 Perl_parse_uniprop_string(pTHX_
22699 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
22700 * now. If so, the return is an inversion list.
22702 * If the property is user-defined, it is a subroutine, which in turn
22703 * may call other subroutines. This function will call the whole nest of
22704 * them to get the definition they return; if some aren't known at the time
22705 * of the call to this function, the fully qualified name of the highest
22706 * level sub is returned. It is an error to call this function at runtime
22707 * without every sub defined.
22709 * If an error was found, NULL is returned, and 'msg' gets a suitable
22710 * message appended to it. (Appending allows the back trace of how we got
22711 * to the faulty definition to be displayed through nested calls of
22712 * user-defined subs.)
22714 * The caller should NOT try to free any returned inversion list.
22716 * Other parameters will be set on return as described below */
22718 const char * const name, /* The first non-blank in the \p{}, \P{} */
22719 const Size_t name_len, /* Its length in bytes, not including any
22721 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22722 const bool to_fold, /* ? Is this under /i */
22723 const bool runtime, /* TRUE if this is being called at run time */
22724 const bool deferrable, /* TRUE if it's ok for the definition to not be
22725 known at this call */
22726 bool *user_defined_ptr, /* Upon return from this function it will be
22727 set to TRUE if any component is a
22728 user-defined property */
22729 SV * msg, /* Any error or warning msg(s) are appended to
22731 const STRLEN level) /* Recursion level of this call */
22734 char* lookup_name; /* normalized name for lookup in our tables */
22735 unsigned lookup_len; /* Its length */
22736 bool stricter = FALSE; /* Some properties have stricter name
22737 normalization rules, which we decide upon
22738 based on parsing */
22740 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
22741 * (though it requires extra effort to download them from Unicode and
22742 * compile perl to know about them) */
22743 bool is_nv_type = FALSE;
22745 unsigned int i, j = 0;
22746 int equals_pos = -1; /* Where the '=' is found, or negative if none */
22747 int slash_pos = -1; /* Where the '/' is found, or negative if none */
22748 int table_index = 0; /* The entry number for this property in the table
22749 of all Unicode property names */
22750 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
22751 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
22752 the normalized name in certain situations */
22753 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
22754 part of a package name */
22755 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
22756 property rather than a Unicode
22758 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
22759 if an error. If it is an inversion list,
22760 it is the definition. Otherwise it is a
22761 string containing the fully qualified sub
22763 SV * fq_name = NULL; /* For user-defined properties, the fully
22765 bool invert_return = FALSE; /* ? Do we need to complement the result before
22768 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
22770 /* The input will be normalized into 'lookup_name' */
22771 Newx(lookup_name, name_len, char);
22772 SAVEFREEPV(lookup_name);
22774 /* Parse the input. */
22775 for (i = 0; i < name_len; i++) {
22776 char cur = name[i];
22778 /* Most of the characters in the input will be of this ilk, being parts
22780 if (isIDCONT_A(cur)) {
22782 /* Case differences are ignored. Our lookup routine assumes
22783 * everything is lowercase, so normalize to that */
22784 if (isUPPER_A(cur)) {
22785 lookup_name[j++] = toLOWER_A(cur);
22789 if (cur == '_') { /* Don't include these in the normalized name */
22793 lookup_name[j++] = cur;
22795 /* The first character in a user-defined name must be of this type.
22797 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
22798 could_be_user_defined = FALSE;
22804 /* Here, the character is not something typically in a name, But these
22805 * two types of characters (and the '_' above) can be freely ignored in
22806 * most situations. Later it may turn out we shouldn't have ignored
22807 * them, and we have to reparse, but we don't have enough information
22808 * yet to make that decision */
22809 if (cur == '-' || isSPACE_A(cur)) {
22810 could_be_user_defined = FALSE;
22814 /* An equals sign or single colon mark the end of the first part of
22815 * the property name */
22817 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
22819 lookup_name[j++] = '='; /* Treat the colon as an '=' */
22820 equals_pos = j; /* Note where it occurred in the input */
22821 could_be_user_defined = FALSE;
22825 /* Otherwise, this character is part of the name. */
22826 lookup_name[j++] = cur;
22828 /* Here it isn't a single colon, so if it is a colon, it must be a
22832 /* A double colon should be a package qualifier. We note its
22833 * position and continue. Note that one could have
22834 * pkg1::pkg2::...::foo
22835 * so that the position at the end of the loop will be just after
22836 * the final qualifier */
22839 non_pkg_begin = i + 1;
22840 lookup_name[j++] = ':';
22842 else { /* Only word chars (and '::') can be in a user-defined name */
22843 could_be_user_defined = FALSE;
22845 } /* End of parsing through the lhs of the property name (or all of it if
22848 #define STRLENs(s) (sizeof("" s "") - 1)
22850 /* If there is a single package name 'utf8::', it is ambiguous. It could
22851 * be for a user-defined property, or it could be a Unicode property, as
22852 * all of them are considered to be for that package. For the purposes of
22853 * parsing the rest of the property, strip it off */
22854 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
22855 lookup_name += STRLENs("utf8::");
22856 j -= STRLENs("utf8::");
22857 equals_pos -= STRLENs("utf8::");
22860 /* Here, we are either done with the whole property name, if it was simple;
22861 * or are positioned just after the '=' if it is compound. */
22863 if (equals_pos >= 0) {
22864 assert(! stricter); /* We shouldn't have set this yet */
22866 /* Space immediately after the '=' is ignored */
22868 for (; i < name_len; i++) {
22869 if (! isSPACE_A(name[i])) {
22874 /* Most punctuation after the equals indicates a subpattern, like
22876 if ( isPUNCT_A(name[i])
22882 /* Find the property. The table includes the equals sign, so we
22884 table_index = match_uniprop((U8 *) lookup_name, j);
22886 const char * const * prop_values
22887 = UNI_prop_value_ptrs[table_index];
22889 Size_t subpattern_len;
22890 REGEXP * subpattern_re;
22891 char open = name[i++];
22893 const char * pos_in_brackets;
22896 /* A backslash means the real delimitter is the next character.
22898 if (open == '\\') {
22903 /* This data structure is constructed so that the matching
22904 * closing bracket is 3 past its matching opening. The second
22905 * set of closing is so that if the opening is something like
22906 * ']', the closing will be that as well. Something similar is
22907 * done in toke.c */
22908 pos_in_brackets = strchr("([<)]>)]>", open);
22909 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
22912 || name[name_len-1] != close
22913 || (escaped && name[name_len-2] != '\\'))
22915 sv_catpvs(msg, "Unicode property wildcard not terminated");
22916 goto append_name_to_msg;
22919 Perl_ck_warner_d(aTHX_
22920 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
22921 "The Unicode property wildcards feature is experimental");
22923 /* Now create and compile the wildcard subpattern. Use /iaa
22924 * because nothing outside of ASCII will match, and it the
22925 * property values should all match /i. Note that when the
22926 * pattern fails to compile, our added text to the user's
22927 * pattern will be displayed to the user, which is not so
22929 subpattern_len = name_len - i - 1 - escaped;
22930 subpattern = Perl_newSVpvf(aTHX_ "(?iaa:%.*s)",
22931 (unsigned) subpattern_len,
22933 subpattern = sv_2mortal(subpattern);
22934 subpattern_re = re_compile(subpattern, 0);
22935 assert(subpattern_re); /* Should have died if didn't compile
22938 /* For each legal property value, see if the supplied pattern
22940 while (*prop_values) {
22941 const char * const entry = *prop_values;
22942 const Size_t len = strlen(entry);
22943 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
22945 if (pregexec(subpattern_re,
22947 (char *) entry + len,
22951 { /* Here, matched. Add to the returned list */
22952 Size_t total_len = j + len;
22953 SV * sub_invlist = NULL;
22954 char * this_string;
22956 /* We know this is a legal \p{property=value}. Call
22957 * the function to return the list of code points that
22959 Newxz(this_string, total_len + 1, char);
22960 Copy(lookup_name, this_string, j, char);
22961 my_strlcat(this_string, entry, total_len + 1);
22962 SAVEFREEPV(this_string);
22963 sub_invlist = parse_uniprop_string(this_string,
22972 _invlist_union(prop_definition, sub_invlist,
22976 prop_values++; /* Next iteration, look at next propvalue */
22977 } /* End of looking through property values; (the data
22978 structure is terminated by a NULL ptr) */
22980 SvREFCNT_dec_NN(subpattern_re);
22982 if (prop_definition) {
22983 return prop_definition;
22986 sv_catpvs(msg, "No Unicode property value wildcard matches:");
22987 goto append_name_to_msg;
22990 /* Here's how khw thinks we should proceed to handle the properties
22991 * not yet done: Bidi Mirroring Glyph
22992 Bidi Paired Bracket
22993 Case Folding (both full and simple)
22994 Decomposition Mapping
22995 Equivalent Unified Ideograph
22998 Lowercase Mapping (both full and simple)
23000 Titlecase Mapping (both full and simple)
23001 Uppercase Mapping (both full and simple)
23002 * Move the part that looks at the property values into a perl
23003 * script, like utf8_heavy.pl was done. This makes things somewhat
23004 * easier, but most importantly, it avoids always adding all these
23005 * strings to the memory usage when the feature is little-used.
23007 * The property values would all be concatenated into a single
23008 * string per property with each value on a separate line, and the
23009 * code point it's for on alternating lines. Then we match the
23010 * user's input pattern m//mg, without having to worry about their
23011 * uses of '^' and '$'. Only the values that aren't the default
23012 * would be in the strings. Code points would be in UTF-8. The
23013 * search pattern that we would construct would look like
23014 * (?: \n (code-point_re) \n (?aam: user-re ) \n )
23015 * And so $1 would contain the code point that matched the user-re.
23016 * For properties where the default is the code point itself, such
23017 * as any of the case changing mappings, the string would otherwise
23018 * consist of all Unicode code points in UTF-8 strung together.
23019 * This would be impractical. So instead, examine their compiled
23020 * pattern, looking at the ssc. If none, reject the pattern as an
23021 * error. Otherwise run the pattern against every code point in
23022 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
23023 * And it might be good to create an API to return the ssc.
23025 * For the name properties, a new function could be created in
23026 * charnames which essentially does the same thing as above,
23027 * sharing Name.pl with the other charname functions. Don't know
23028 * about loose name matching, or algorithmically determined names.
23029 * Decomposition.pl similarly.
23031 * It might be that a new pattern modifier would have to be
23032 * created, like /t for resTricTed, which changed the behavior of
23033 * some constructs in their subpattern, like \A. */
23034 } /* End of is a wildcard subppattern */
23037 /* Certain properties whose values are numeric need special handling.
23038 * They may optionally be prefixed by 'is'. Ignore that prefix for the
23039 * purposes of checking if this is one of those properties */
23040 if (memBEGINPs(lookup_name, j, "is")) {
23044 /* Then check if it is one of these specially-handled properties. The
23045 * possibilities are hard-coded because easier this way, and the list
23046 * is unlikely to change.
23048 * All numeric value type properties are of this ilk, and are also
23049 * special in a different way later on. So find those first. There
23050 * are several numeric value type properties in the Unihan DB (which is
23051 * unlikely to be compiled with perl, but we handle it here in case it
23052 * does get compiled). They all end with 'numeric'. The interiors
23053 * aren't checked for the precise property. This would stop working if
23054 * a cjk property were to be created that ended with 'numeric' and
23055 * wasn't a numeric type */
23056 is_nv_type = memEQs(lookup_name + lookup_offset,
23057 j - 1 - lookup_offset, "numericvalue")
23058 || memEQs(lookup_name + lookup_offset,
23059 j - 1 - lookup_offset, "nv")
23060 || ( memENDPs(lookup_name + lookup_offset,
23061 j - 1 - lookup_offset, "numeric")
23062 && ( memBEGINPs(lookup_name + lookup_offset,
23063 j - 1 - lookup_offset, "cjk")
23064 || memBEGINPs(lookup_name + lookup_offset,
23065 j - 1 - lookup_offset, "k")));
23067 || memEQs(lookup_name + lookup_offset,
23068 j - 1 - lookup_offset, "canonicalcombiningclass")
23069 || memEQs(lookup_name + lookup_offset,
23070 j - 1 - lookup_offset, "ccc")
23071 || memEQs(lookup_name + lookup_offset,
23072 j - 1 - lookup_offset, "age")
23073 || memEQs(lookup_name + lookup_offset,
23074 j - 1 - lookup_offset, "in")
23075 || memEQs(lookup_name + lookup_offset,
23076 j - 1 - lookup_offset, "presentin"))
23080 /* Since the stuff after the '=' is a number, we can't throw away
23081 * '-' willy-nilly, as those could be a minus sign. Other stricter
23082 * rules also apply. However, these properties all can have the
23083 * rhs not be a number, in which case they contain at least one
23084 * alphabetic. In those cases, the stricter rules don't apply.
23085 * But the numeric type properties can have the alphas [Ee] to
23086 * signify an exponent, and it is still a number with stricter
23087 * rules. So look for an alpha that signifies not-strict */
23089 for (k = i; k < name_len; k++) {
23090 if ( isALPHA_A(name[k])
23091 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
23101 /* A number may have a leading '+' or '-'. The latter is retained
23103 if (name[i] == '+') {
23106 else if (name[i] == '-') {
23107 lookup_name[j++] = '-';
23111 /* Skip leading zeros including single underscores separating the
23112 * zeros, or between the final leading zero and the first other
23114 for (; i < name_len - 1; i++) {
23115 if ( name[i] != '0'
23116 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23123 else { /* No '=' */
23125 /* Only a few properties without an '=' should be parsed with stricter
23126 * rules. The list is unlikely to change. */
23127 if ( memBEGINPs(lookup_name, j, "perl")
23128 && memNEs(lookup_name + 4, j - 4, "space")
23129 && memNEs(lookup_name + 4, j - 4, "word"))
23133 /* We set the inputs back to 0 and the code below will reparse,
23139 /* Here, we have either finished the property, or are positioned to parse
23140 * the remainder, and we know if stricter rules apply. Finish out, if not
23142 for (; i < name_len; i++) {
23143 char cur = name[i];
23145 /* In all instances, case differences are ignored, and we normalize to
23147 if (isUPPER_A(cur)) {
23148 lookup_name[j++] = toLOWER(cur);
23152 /* An underscore is skipped, but not under strict rules unless it
23153 * separates two digits */
23156 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
23157 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
23159 lookup_name[j++] = '_';
23164 /* Hyphens are skipped except under strict */
23165 if (cur == '-' && ! stricter) {
23169 /* XXX Bug in documentation. It says white space skipped adjacent to
23170 * non-word char. Maybe we should, but shouldn't skip it next to a dot
23172 if (isSPACE_A(cur) && ! stricter) {
23176 lookup_name[j++] = cur;
23178 /* Unless this is a non-trailing slash, we are done with it */
23179 if (i >= name_len - 1 || cur != '/') {
23185 /* A slash in the 'numeric value' property indicates that what follows
23186 * is a denominator. It can have a leading '+' and '0's that should be
23187 * skipped. But we have never allowed a negative denominator, so treat
23188 * a minus like every other character. (No need to rule out a second
23189 * '/', as that won't match anything anyway */
23192 if (i < name_len && name[i] == '+') {
23196 /* Skip leading zeros including underscores separating digits */
23197 for (; i < name_len - 1; i++) {
23198 if ( name[i] != '0'
23199 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23205 /* Store the first real character in the denominator */
23206 if (i < name_len) {
23207 lookup_name[j++] = name[i];
23212 /* Here are completely done parsing the input 'name', and 'lookup_name'
23213 * contains a copy, normalized.
23215 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
23216 * different from without the underscores. */
23217 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
23218 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
23219 && UNLIKELY(name[name_len-1] == '_'))
23221 lookup_name[j++] = '&';
23224 /* If the original input began with 'In' or 'Is', it could be a subroutine
23225 * call to a user-defined property instead of a Unicode property name. */
23226 if ( name_len - non_pkg_begin > 2
23227 && name[non_pkg_begin+0] == 'I'
23228 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
23230 /* Names that start with In have different characterstics than those
23231 * that start with Is */
23232 if (name[non_pkg_begin+1] == 's') {
23233 starts_with_Is = TRUE;
23237 could_be_user_defined = FALSE;
23240 if (could_be_user_defined) {
23243 /* If the user defined property returns the empty string, it could
23244 * easily be because the pattern is being compiled before the data it
23245 * actually needs to compile is available. This could be argued to be
23246 * a bug in the perl code, but this is a change of behavior for Perl,
23247 * so we handle it. This means that intentionally returning nothing
23248 * will not be resolved until runtime */
23249 bool empty_return = FALSE;
23251 /* Here, the name could be for a user defined property, which are
23252 * implemented as subs. */
23253 user_sub = get_cvn_flags(name, name_len, 0);
23255 const char insecure[] = "Insecure user-defined property";
23257 /* Here, there is a sub by the correct name. Normally we call it
23258 * to get the property definition */
23260 SV * user_sub_sv = MUTABLE_SV(user_sub);
23261 SV * error; /* Any error returned by calling 'user_sub' */
23262 SV * key; /* The key into the hash of user defined sub names
23265 SV ** saved_user_prop_ptr; /* Hash entry for this property */
23267 /* How many times to retry when another thread is in the middle of
23268 * expanding the same definition we want */
23269 PERL_INT_FAST8_T retry_countdown = 10;
23271 DECLARATION_FOR_GLOBAL_CONTEXT;
23273 /* If we get here, we know this property is user-defined */
23274 *user_defined_ptr = TRUE;
23276 /* We refuse to call a potentially tainted subroutine; returning an
23279 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23280 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23281 goto append_name_to_msg;
23284 /* In principal, we only call each subroutine property definition
23285 * once during the life of the program. This guarantees that the
23286 * property definition never changes. The results of the single
23287 * sub call are stored in a hash, which is used instead for future
23288 * references to this property. The property definition is thus
23289 * immutable. But, to allow the user to have a /i-dependent
23290 * definition, we call the sub once for non-/i, and once for /i,
23291 * should the need arise, passing the /i status as a parameter.
23293 * We start by constructing the hash key name, consisting of the
23294 * fully qualified subroutine name, preceded by the /i status, so
23295 * that there is a key for /i and a different key for non-/i */
23296 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
23297 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23298 non_pkg_begin != 0);
23299 sv_catsv(key, fq_name);
23302 /* We only call the sub once throughout the life of the program
23303 * (with the /i, non-/i exception noted above). That means the
23304 * hash must be global and accessible to all threads. It is
23305 * created at program start-up, before any threads are created, so
23306 * is accessible to all children. But this creates some
23309 * 1) The keys can't be shared, or else problems arise; sharing is
23310 * turned off at hash creation time
23311 * 2) All SVs in it are there for the remainder of the life of the
23312 * program, and must be created in the same interpreter context
23313 * as the hash, or else they will be freed from the wrong pool
23314 * at global destruction time. This is handled by switching to
23315 * the hash's context to create each SV going into it, and then
23316 * immediately switching back
23317 * 3) All accesses to the hash must be controlled by a mutex, to
23318 * prevent two threads from getting an unstable state should
23319 * they simultaneously be accessing it. The code below is
23320 * crafted so that the mutex is locked whenever there is an
23321 * access and unlocked only when the next stable state is
23324 * The hash stores either the definition of the property if it was
23325 * valid, or, if invalid, the error message that was raised. We
23326 * use the type of SV to distinguish.
23328 * There's also the need to guard against the definition expansion
23329 * from infinitely recursing. This is handled by storing the aTHX
23330 * of the expanding thread during the expansion. Again the SV type
23331 * is used to distinguish this from the other two cases. If we
23332 * come to here and the hash entry for this property is our aTHX,
23333 * it means we have recursed, and the code assumes that we would
23334 * infinitely recurse, so instead stops and raises an error.
23335 * (Any recursion has always been treated as infinite recursion in
23338 * If instead, the entry is for a different aTHX, it means that
23339 * that thread has gotten here first, and hasn't finished expanding
23340 * the definition yet. We just have to wait until it is done. We
23341 * sleep and retry a few times, returning an error if the other
23342 * thread doesn't complete. */
23345 USER_PROP_MUTEX_LOCK;
23347 /* If we have an entry for this key, the subroutine has already
23348 * been called once with this /i status. */
23349 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
23350 SvPVX(key), SvCUR(key), 0);
23351 if (saved_user_prop_ptr) {
23353 /* If the saved result is an inversion list, it is the valid
23354 * definition of this property */
23355 if (is_invlist(*saved_user_prop_ptr)) {
23356 prop_definition = *saved_user_prop_ptr;
23358 /* The SV in the hash won't be removed until global
23359 * destruction, so it is stable and we can unlock */
23360 USER_PROP_MUTEX_UNLOCK;
23362 /* The caller shouldn't try to free this SV */
23363 return prop_definition;
23366 /* Otherwise, if it is a string, it is the error message
23367 * that was returned when we first tried to evaluate this
23368 * property. Fail, and append the message */
23369 if (SvPOK(*saved_user_prop_ptr)) {
23370 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23371 sv_catsv(msg, *saved_user_prop_ptr);
23373 /* The SV in the hash won't be removed until global
23374 * destruction, so it is stable and we can unlock */
23375 USER_PROP_MUTEX_UNLOCK;
23380 assert(SvIOK(*saved_user_prop_ptr));
23382 /* Here, we have an unstable entry in the hash. Either another
23383 * thread is in the middle of expanding the property's
23384 * definition, or we are ourselves recursing. We use the aTHX
23385 * in it to distinguish */
23386 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
23388 /* Here, it's another thread doing the expanding. We've
23389 * looked as much as we are going to at the contents of the
23390 * hash entry. It's safe to unlock. */
23391 USER_PROP_MUTEX_UNLOCK;
23393 /* Retry a few times */
23394 if (retry_countdown-- > 0) {
23399 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23400 sv_catpvs(msg, "Timeout waiting for another thread to "
23402 goto append_name_to_msg;
23405 /* Here, we are recursing; don't dig any deeper */
23406 USER_PROP_MUTEX_UNLOCK;
23408 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23410 "Infinite recursion in user-defined property");
23411 goto append_name_to_msg;
23414 /* Here, this thread has exclusive control, and there is no entry
23415 * for this property in the hash. So we have the go ahead to
23416 * expand the definition ourselves. */
23418 PUSHSTACKi(PERLSI_MAGIC);
23421 /* Create a temporary placeholder in the hash to detect recursion
23423 SWITCH_TO_GLOBAL_CONTEXT;
23424 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
23425 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
23428 /* Now that we have a placeholder, we can let other threads
23430 USER_PROP_MUTEX_UNLOCK;
23432 /* Make sure the placeholder always gets destroyed */
23433 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
23438 /* Call the user's function, with the /i status as a parameter.
23439 * Note that we have gone to a lot of trouble to keep this call
23440 * from being within the locked mutex region. */
23441 XPUSHs(boolSV(to_fold));
23444 /* The following block was taken from swash_init(). Presumably
23445 * they apply to here as well, though we no longer use a swash --
23449 /* We might get here via a subroutine signature which uses a utf8
23450 * parameter name, at which point PL_subname will have been set
23451 * but not yet used. */
23452 save_item(PL_subname);
23454 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
23459 if (TAINT_get || SvTRUE(error)) {
23460 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23461 if (SvTRUE(error)) {
23462 sv_catpvs(msg, "Error \"");
23463 sv_catsv(msg, error);
23464 sv_catpvs(msg, "\"");
23467 if (SvTRUE(error)) sv_catpvs(msg, "; ");
23468 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23471 if (name_len > 0) {
23472 sv_catpvs(msg, " in expansion of ");
23473 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
23479 prop_definition = NULL;
23481 else { /* G_SCALAR guarantees a single return value */
23482 SV * contents = POPs;
23484 /* The contents is supposed to be the expansion of the property
23485 * definition. If the definition is deferrable, and we got an
23486 * empty string back, set a flag to later defer it (after clean
23489 && (! SvPOK(contents) || SvCUR(contents) == 0))
23491 empty_return = TRUE;
23493 else { /* Otherwise, call a function to check for valid syntax,
23496 prop_definition = handle_user_defined_property(
23498 is_utf8, to_fold, runtime,
23500 contents, user_defined_ptr,
23506 /* Here, we have the results of the expansion. Delete the
23507 * placeholder, and if the definition is now known, replace it with
23508 * that definition. We need exclusive access to the hash, and we
23509 * can't let anyone else in, between when we delete the placeholder
23510 * and add the permanent entry */
23511 USER_PROP_MUTEX_LOCK;
23513 S_delete_recursion_entry(aTHX_ SvPVX(key));
23515 if ( ! empty_return
23516 && (! prop_definition || is_invlist(prop_definition)))
23518 /* If we got success we use the inversion list defining the
23519 * property; otherwise use the error message */
23520 SWITCH_TO_GLOBAL_CONTEXT;
23521 (void) hv_store_ent(PL_user_def_props,
23524 ? newSVsv(prop_definition)
23530 /* All done, and the hash now has a permanent entry for this
23531 * property. Give up exclusive control */
23532 USER_PROP_MUTEX_UNLOCK;
23538 if (empty_return) {
23539 goto definition_deferred;
23542 if (prop_definition) {
23544 /* If the definition is for something not known at this time,
23545 * we toss it, and go return the main property name, as that's
23546 * the one the user will be aware of */
23547 if (! is_invlist(prop_definition)) {
23548 SvREFCNT_dec_NN(prop_definition);
23549 goto definition_deferred;
23552 sv_2mortal(prop_definition);
23556 return prop_definition;
23558 } /* End of calling the subroutine for the user-defined property */
23559 } /* End of it could be a user-defined property */
23561 /* Here it wasn't a user-defined property that is known at this time. See
23562 * if it is a Unicode property */
23564 lookup_len = j; /* This is a more mnemonic name than 'j' */
23566 /* Get the index into our pointer table of the inversion list corresponding
23567 * to the property */
23568 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23570 /* If it didn't find the property ... */
23571 if (table_index == 0) {
23573 /* Try again stripping off any initial 'Is'. This is because we
23574 * promise that an initial Is is optional. The same isn't true of
23575 * names that start with 'In'. Those can match only blocks, and the
23576 * lookup table already has those accounted for. */
23577 if (starts_with_Is) {
23583 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23586 if (table_index == 0) {
23589 /* Here, we didn't find it. If not a numeric type property, and
23590 * can't be a user-defined one, it isn't a legal property */
23591 if (! is_nv_type) {
23592 if (! could_be_user_defined) {
23596 /* Here, the property name is legal as a user-defined one. At
23597 * compile time, it might just be that the subroutine for that
23598 * property hasn't been encountered yet, but at runtime, it's
23599 * an error to try to use an undefined one */
23600 if (! deferrable) {
23601 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23602 sv_catpvs(msg, "Unknown user-defined property name");
23603 goto append_name_to_msg;
23606 goto definition_deferred;
23607 } /* End of isn't a numeric type property */
23609 /* The numeric type properties need more work to decide. What we
23610 * do is make sure we have the number in canonical form and look
23613 if (slash_pos < 0) { /* No slash */
23615 /* When it isn't a rational, take the input, convert it to a
23616 * NV, then create a canonical string representation of that
23620 SSize_t value_len = lookup_len - equals_pos;
23622 /* Get the value */
23623 if ( value_len <= 0
23624 || my_atof3(lookup_name + equals_pos, &value,
23626 != lookup_name + lookup_len)
23631 /* If the value is an integer, the canonical value is integral
23633 if (Perl_ceil(value) == value) {
23634 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
23635 equals_pos, lookup_name, value);
23637 else { /* Otherwise, it is %e with a known precision */
23640 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
23641 equals_pos, lookup_name,
23642 PL_E_FORMAT_PRECISION, value);
23644 /* The exponent generated is expecting two digits, whereas
23645 * %e on some systems will generate three. Remove leading
23646 * zeros in excess of 2 from the exponent. We start
23647 * looking for them after the '=' */
23648 exp_ptr = strchr(canonical + equals_pos, 'e');
23650 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
23651 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
23653 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
23655 if (excess_exponent_len > 0) {
23656 SSize_t leading_zeros = strspn(cur_ptr, "0");
23657 SSize_t excess_leading_zeros
23658 = MIN(leading_zeros, excess_exponent_len);
23659 if (excess_leading_zeros > 0) {
23660 Move(cur_ptr + excess_leading_zeros,
23662 strlen(cur_ptr) - excess_leading_zeros
23663 + 1, /* Copy the NUL as well */
23670 else { /* Has a slash. Create a rational in canonical form */
23671 UV numerator, denominator, gcd, trial;
23672 const char * end_ptr;
23673 const char * sign = "";
23675 /* We can't just find the numerator, denominator, and do the
23676 * division, then use the method above, because that is
23677 * inexact. And the input could be a rational that is within
23678 * epsilon (given our precision) of a valid rational, and would
23679 * then incorrectly compare valid.
23681 * We're only interested in the part after the '=' */
23682 const char * this_lookup_name = lookup_name + equals_pos;
23683 lookup_len -= equals_pos;
23684 slash_pos -= equals_pos;
23686 /* Handle any leading minus */
23687 if (this_lookup_name[0] == '-') {
23689 this_lookup_name++;
23694 /* Convert the numerator to numeric */
23695 end_ptr = this_lookup_name + slash_pos;
23696 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
23700 /* It better have included all characters before the slash */
23701 if (*end_ptr != '/') {
23705 /* Set to look at just the denominator */
23706 this_lookup_name += slash_pos;
23707 lookup_len -= slash_pos;
23708 end_ptr = this_lookup_name + lookup_len;
23710 /* Convert the denominator to numeric */
23711 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
23715 /* It better be the rest of the characters, and don't divide by
23717 if ( end_ptr != this_lookup_name + lookup_len
23718 || denominator == 0)
23723 /* Get the greatest common denominator using
23724 http://en.wikipedia.org/wiki/Euclidean_algorithm */
23726 trial = denominator;
23727 while (trial != 0) {
23729 trial = gcd % trial;
23733 /* If already in lowest possible terms, we have already tried
23734 * looking this up */
23739 /* Reduce the rational, which should put it in canonical form
23742 denominator /= gcd;
23744 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
23745 equals_pos, lookup_name, sign, numerator, denominator);
23748 /* Here, we have the number in canonical form. Try that */
23749 table_index = match_uniprop((U8 *) canonical, strlen(canonical));
23750 if (table_index == 0) {
23753 } /* End of still didn't find the property in our table */
23754 } /* End of didn't find the property in our table */
23756 /* Here, we have a non-zero return, which is an index into a table of ptrs.
23757 * A negative return signifies that the real index is the absolute value,
23758 * but the result needs to be inverted */
23759 if (table_index < 0) {
23760 invert_return = TRUE;
23761 table_index = -table_index;
23764 /* Out-of band indices indicate a deprecated property. The proper index is
23765 * modulo it with the table size. And dividing by the table size yields
23766 * an offset into a table constructed by regen/mk_invlists.pl to contain
23767 * the corresponding warning message */
23768 if (table_index > MAX_UNI_KEYWORD_INDEX) {
23769 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
23770 table_index %= MAX_UNI_KEYWORD_INDEX;
23771 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
23772 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
23773 (int) name_len, name, deprecated_property_msgs[warning_offset]);
23776 /* In a few properties, a different property is used under /i. These are
23777 * unlikely to change, so are hard-coded here. */
23779 if ( table_index == UNI_XPOSIXUPPER
23780 || table_index == UNI_XPOSIXLOWER
23781 || table_index == UNI_TITLE)
23783 table_index = UNI_CASED;
23785 else if ( table_index == UNI_UPPERCASELETTER
23786 || table_index == UNI_LOWERCASELETTER
23787 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
23788 || table_index == UNI_TITLECASELETTER
23791 table_index = UNI_CASEDLETTER;
23793 else if ( table_index == UNI_POSIXUPPER
23794 || table_index == UNI_POSIXLOWER)
23796 table_index = UNI_POSIXALPHA;
23800 /* Create and return the inversion list */
23801 prop_definition =_new_invlist_C_array(uni_prop_ptrs[table_index]);
23802 sv_2mortal(prop_definition);
23805 /* See if there is a private use override to add to this definition */
23807 COPHH * hinthash = (IN_PERL_COMPILETIME)
23808 ? CopHINTHASH_get(&PL_compiling)
23809 : CopHINTHASH_get(PL_curcop);
23810 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
23812 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
23814 /* See if there is an element in the hints hash for this table */
23815 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
23816 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
23820 SV * pu_definition;
23822 SV * expanded_prop_definition =
23823 sv_2mortal(invlist_clone(prop_definition, NULL));
23825 /* If so, it's definition is the string from here to the next
23826 * \a character. And its format is the same as a user-defined
23828 pos += SvCUR(pu_lookup);
23829 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
23830 pu_invlist = handle_user_defined_property(lookup_name,
23833 0, /* Not folded */
23841 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23842 sv_catpvs(msg, "Insecure private-use override");
23843 goto append_name_to_msg;
23846 /* For now, as a safety measure, make sure that it doesn't
23847 * override non-private use code points */
23848 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
23850 /* Add it to the list to be returned */
23851 _invlist_union(prop_definition, pu_invlist,
23852 &expanded_prop_definition);
23853 prop_definition = expanded_prop_definition;
23854 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
23859 if (invert_return) {
23860 _invlist_invert(prop_definition);
23862 return prop_definition;
23866 if (non_pkg_begin != 0) {
23867 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23868 sv_catpvs(msg, "Illegal user-defined property name");
23871 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23872 sv_catpvs(msg, "Can't find Unicode property definition");
23876 append_name_to_msg:
23878 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
23879 const char * suffix = (runtime && level == 0) ? "}" : "\"";
23881 sv_catpv(msg, prefix);
23882 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23883 sv_catpv(msg, suffix);
23888 definition_deferred:
23890 /* Here it could yet to be defined, so defer evaluation of this
23891 * until its needed at runtime. We need the fully qualified property name
23892 * to avoid ambiguity, and a trailing newline */
23894 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23895 non_pkg_begin != 0 /* If has "::" */
23898 sv_catpvs(fq_name, "\n");
23900 *user_defined_ptr = TRUE;
23907 * ex: set ts=8 sts=4 sw=4 et: