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_charset() */
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, ANYOFRb))
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, ANYOFRb)) {
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, ANYOFRb)
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, ANYOFRb)
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);
2145 SvREFCNT_dec(invlist);
2147 /* Make sure is clone-safe */
2148 ssc->invlist = NULL;
2150 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
2151 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
2152 OP(ssc) = ANYOFPOSIXL;
2154 else if (RExC_contains_locale) {
2158 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
2161 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
2162 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
2163 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
2164 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
2165 ? (TRIE_LIST_CUR( idx ) - 1) \
2171 dump_trie(trie,widecharmap,revcharmap)
2172 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
2173 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
2175 These routines dump out a trie in a somewhat readable format.
2176 The _interim_ variants are used for debugging the interim
2177 tables that are used to generate the final compressed
2178 representation which is what dump_trie expects.
2180 Part of the reason for their existence is to provide a form
2181 of documentation as to how the different representations function.
2186 Dumps the final compressed table form of the trie to Perl_debug_log.
2187 Used for debugging make_trie().
2191 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
2192 AV *revcharmap, U32 depth)
2195 SV *sv=sv_newmortal();
2196 int colwidth= widecharmap ? 6 : 4;
2198 GET_RE_DEBUG_FLAGS_DECL;
2200 PERL_ARGS_ASSERT_DUMP_TRIE;
2202 Perl_re_indentf( aTHX_ "Char : %-6s%-6s%-4s ",
2203 depth+1, "Match","Base","Ofs" );
2205 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
2206 SV ** const tmp = av_fetch( revcharmap, state, 0);
2208 Perl_re_printf( aTHX_ "%*s",
2210 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2211 PL_colors[0], PL_colors[1],
2212 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2213 PERL_PV_ESCAPE_FIRSTCHAR
2218 Perl_re_printf( aTHX_ "\n");
2219 Perl_re_indentf( aTHX_ "State|-----------------------", depth+1);
2221 for( state = 0 ; state < trie->uniquecharcount ; state++ )
2222 Perl_re_printf( aTHX_ "%.*s", colwidth, "--------");
2223 Perl_re_printf( aTHX_ "\n");
2225 for( state = 1 ; state < trie->statecount ; state++ ) {
2226 const U32 base = trie->states[ state ].trans.base;
2228 Perl_re_indentf( aTHX_ "#%4" UVXf "|", depth+1, (UV)state);
2230 if ( trie->states[ state ].wordnum ) {
2231 Perl_re_printf( aTHX_ " W%4X", trie->states[ state ].wordnum );
2233 Perl_re_printf( aTHX_ "%6s", "" );
2236 Perl_re_printf( aTHX_ " @%4" UVXf " ", (UV)base );
2241 while( ( base + ofs < trie->uniquecharcount ) ||
2242 ( base + ofs - trie->uniquecharcount < trie->lasttrans
2243 && trie->trans[ base + ofs - trie->uniquecharcount ].check
2247 Perl_re_printf( aTHX_ "+%2" UVXf "[ ", (UV)ofs);
2249 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2250 if ( ( base + ofs >= trie->uniquecharcount )
2251 && ( base + ofs - trie->uniquecharcount
2253 && trie->trans[ base + ofs
2254 - trie->uniquecharcount ].check == state )
2256 Perl_re_printf( aTHX_ "%*" UVXf, colwidth,
2257 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next
2260 Perl_re_printf( aTHX_ "%*s", colwidth," ." );
2264 Perl_re_printf( aTHX_ "]");
2267 Perl_re_printf( aTHX_ "\n" );
2269 Perl_re_indentf( aTHX_ "word_info N:(prev,len)=",
2271 for (word=1; word <= trie->wordcount; word++) {
2272 Perl_re_printf( aTHX_ " %d:(%d,%d)",
2273 (int)word, (int)(trie->wordinfo[word].prev),
2274 (int)(trie->wordinfo[word].len));
2276 Perl_re_printf( aTHX_ "\n" );
2279 Dumps a fully constructed but uncompressed trie in list form.
2280 List tries normally only are used for construction when the number of
2281 possible chars (trie->uniquecharcount) is very high.
2282 Used for debugging make_trie().
2285 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
2286 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2290 SV *sv=sv_newmortal();
2291 int colwidth= widecharmap ? 6 : 4;
2292 GET_RE_DEBUG_FLAGS_DECL;
2294 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
2296 /* print out the table precompression. */
2297 Perl_re_indentf( aTHX_ "State :Word | Transition Data\n",
2299 Perl_re_indentf( aTHX_ "%s",
2300 depth+1, "------:-----+-----------------\n" );
2302 for( state=1 ; state < next_alloc ; state ++ ) {
2305 Perl_re_indentf( aTHX_ " %4" UVXf " :",
2306 depth+1, (UV)state );
2307 if ( ! trie->states[ state ].wordnum ) {
2308 Perl_re_printf( aTHX_ "%5s| ","");
2310 Perl_re_printf( aTHX_ "W%4x| ",
2311 trie->states[ state ].wordnum
2314 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
2315 SV ** const tmp = av_fetch( revcharmap,
2316 TRIE_LIST_ITEM(state, charid).forid, 0);
2318 Perl_re_printf( aTHX_ "%*s:%3X=%4" UVXf " | ",
2320 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
2322 PL_colors[0], PL_colors[1],
2323 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
2324 | PERL_PV_ESCAPE_FIRSTCHAR
2326 TRIE_LIST_ITEM(state, charid).forid,
2327 (UV)TRIE_LIST_ITEM(state, charid).newstate
2330 Perl_re_printf( aTHX_ "\n%*s| ",
2331 (int)((depth * 2) + 14), "");
2334 Perl_re_printf( aTHX_ "\n");
2339 Dumps a fully constructed but uncompressed trie in table form.
2340 This is the normal DFA style state transition table, with a few
2341 twists to facilitate compression later.
2342 Used for debugging make_trie().
2345 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
2346 HV *widecharmap, AV *revcharmap, U32 next_alloc,
2351 SV *sv=sv_newmortal();
2352 int colwidth= widecharmap ? 6 : 4;
2353 GET_RE_DEBUG_FLAGS_DECL;
2355 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
2358 print out the table precompression so that we can do a visual check
2359 that they are identical.
2362 Perl_re_indentf( aTHX_ "Char : ", depth+1 );
2364 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2365 SV ** const tmp = av_fetch( revcharmap, charid, 0);
2367 Perl_re_printf( aTHX_ "%*s",
2369 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
2370 PL_colors[0], PL_colors[1],
2371 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2372 PERL_PV_ESCAPE_FIRSTCHAR
2378 Perl_re_printf( aTHX_ "\n");
2379 Perl_re_indentf( aTHX_ "State+-", depth+1 );
2381 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
2382 Perl_re_printf( aTHX_ "%.*s", colwidth,"--------");
2385 Perl_re_printf( aTHX_ "\n" );
2387 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
2389 Perl_re_indentf( aTHX_ "%4" UVXf " : ",
2391 (UV)TRIE_NODENUM( state ) );
2393 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
2394 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
2396 Perl_re_printf( aTHX_ "%*" UVXf, colwidth, v );
2398 Perl_re_printf( aTHX_ "%*s", colwidth, "." );
2400 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
2401 Perl_re_printf( aTHX_ " (%4" UVXf ")\n",
2402 (UV)trie->trans[ state ].check );
2404 Perl_re_printf( aTHX_ " (%4" UVXf ") W%4X\n",
2405 (UV)trie->trans[ state ].check,
2406 trie->states[ TRIE_NODENUM( state ) ].wordnum );
2414 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
2415 startbranch: the first branch in the whole branch sequence
2416 first : start branch of sequence of branch-exact nodes.
2417 May be the same as startbranch
2418 last : Thing following the last branch.
2419 May be the same as tail.
2420 tail : item following the branch sequence
2421 count : words in the sequence
2422 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
2423 depth : indent depth
2425 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
2427 A trie is an N'ary tree where the branches are determined by digital
2428 decomposition of the key. IE, at the root node you look up the 1st character and
2429 follow that branch repeat until you find the end of the branches. Nodes can be
2430 marked as "accepting" meaning they represent a complete word. Eg:
2434 would convert into the following structure. Numbers represent states, letters
2435 following numbers represent valid transitions on the letter from that state, if
2436 the number is in square brackets it represents an accepting state, otherwise it
2437 will be in parenthesis.
2439 +-h->+-e->[3]-+-r->(8)-+-s->[9]
2443 (1) +-i->(6)-+-s->[7]
2445 +-s->(3)-+-h->(4)-+-e->[5]
2447 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
2449 This shows that when matching against the string 'hers' we will begin at state 1
2450 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
2451 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
2452 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
2453 single traverse. We store a mapping from accepting to state to which word was
2454 matched, and then when we have multiple possibilities we try to complete the
2455 rest of the regex in the order in which they occurred in the alternation.
2457 The only prior NFA like behaviour that would be changed by the TRIE support is
2458 the silent ignoring of duplicate alternations which are of the form:
2460 / (DUPE|DUPE) X? (?{ ... }) Y /x
2462 Thus EVAL blocks following a trie may be called a different number of times with
2463 and without the optimisation. With the optimisations dupes will be silently
2464 ignored. This inconsistent behaviour of EVAL type nodes is well established as
2465 the following demonstrates:
2467 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
2469 which prints out 'word' three times, but
2471 'words'=~/(word|word|word)(?{ print $1 })S/
2473 which doesnt print it out at all. This is due to other optimisations kicking in.
2475 Example of what happens on a structural level:
2477 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2479 1: CURLYM[1] {1,32767}(18)
2490 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2491 and should turn into:
2493 1: CURLYM[1] {1,32767}(18)
2495 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2503 Cases where tail != last would be like /(?foo|bar)baz/:
2513 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2514 and would end up looking like:
2517 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2524 d = uvchr_to_utf8_flags(d, uv, 0);
2526 is the recommended Unicode-aware way of saying
2531 #define TRIE_STORE_REVCHAR(val) \
2534 SV *zlopp = newSV(UTF8_MAXBYTES); \
2535 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2536 unsigned char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2538 SvCUR_set(zlopp, kapow - flrbbbbb); \
2541 av_push(revcharmap, zlopp); \
2543 char ooooff = (char)val; \
2544 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2548 /* This gets the next character from the input, folding it if not already
2550 #define TRIE_READ_CHAR STMT_START { \
2553 /* if it is UTF then it is either already folded, or does not need \
2555 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2557 else if (folder == PL_fold_latin1) { \
2558 /* This folder implies Unicode rules, which in the range expressible \
2559 * by not UTF is the lower case, with the two exceptions, one of \
2560 * which should have been taken care of before calling this */ \
2561 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2562 uvc = toLOWER_L1(*uc); \
2563 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2566 /* raw data, will be folded later if needed */ \
2574 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2575 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2576 U32 ging = TRIE_LIST_LEN( state ) * 2; \
2577 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2578 TRIE_LIST_LEN( state ) = ging; \
2580 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2581 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2582 TRIE_LIST_CUR( state )++; \
2585 #define TRIE_LIST_NEW(state) STMT_START { \
2586 Newx( trie->states[ state ].trans.list, \
2587 4, reg_trie_trans_le ); \
2588 TRIE_LIST_CUR( state ) = 1; \
2589 TRIE_LIST_LEN( state ) = 4; \
2592 #define TRIE_HANDLE_WORD(state) STMT_START { \
2593 U16 dupe= trie->states[ state ].wordnum; \
2594 regnode * const noper_next = regnext( noper ); \
2597 /* store the word for dumping */ \
2599 if (OP(noper) != NOTHING) \
2600 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2602 tmp = newSVpvn_utf8( "", 0, UTF ); \
2603 av_push( trie_words, tmp ); \
2607 trie->wordinfo[curword].prev = 0; \
2608 trie->wordinfo[curword].len = wordlen; \
2609 trie->wordinfo[curword].accept = state; \
2611 if ( noper_next < tail ) { \
2613 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2615 trie->jump[curword] = (U16)(noper_next - convert); \
2617 jumper = noper_next; \
2619 nextbranch= regnext(cur); \
2623 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2624 /* chain, so that when the bits of chain are later */\
2625 /* linked together, the dups appear in the chain */\
2626 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2627 trie->wordinfo[dupe].prev = curword; \
2629 /* we haven't inserted this word yet. */ \
2630 trie->states[ state ].wordnum = curword; \
2635 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2636 ( ( base + charid >= ucharcount \
2637 && base + charid < ubound \
2638 && state == trie->trans[ base - ucharcount + charid ].check \
2639 && trie->trans[ base - ucharcount + charid ].next ) \
2640 ? trie->trans[ base - ucharcount + charid ].next \
2641 : ( state==1 ? special : 0 ) \
2644 #define TRIE_BITMAP_SET_FOLDED(trie, uvc, folder) \
2646 TRIE_BITMAP_SET(trie, uvc); \
2647 /* store the folded codepoint */ \
2649 TRIE_BITMAP_SET(trie, folder[(U8) uvc ]); \
2652 /* store first byte of utf8 representation of */ \
2653 /* variant codepoints */ \
2654 if (! UVCHR_IS_INVARIANT(uvc)) { \
2655 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc)); \
2660 #define MADE_JUMP_TRIE 2
2661 #define MADE_EXACT_TRIE 4
2664 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2665 regnode *first, regnode *last, regnode *tail,
2666 U32 word_count, U32 flags, U32 depth)
2668 /* first pass, loop through and scan words */
2669 reg_trie_data *trie;
2670 HV *widecharmap = NULL;
2671 AV *revcharmap = newAV();
2677 regnode *jumper = NULL;
2678 regnode *nextbranch = NULL;
2679 regnode *convert = NULL;
2680 U32 *prev_states; /* temp array mapping each state to previous one */
2681 /* we just use folder as a flag in utf8 */
2682 const U8 * folder = NULL;
2684 /* in the below add_data call we are storing either 'tu' or 'tuaa'
2685 * which stands for one trie structure, one hash, optionally followed
2688 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuaa"));
2689 AV *trie_words = NULL;
2690 /* along with revcharmap, this only used during construction but both are
2691 * useful during debugging so we store them in the struct when debugging.
2694 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2695 STRLEN trie_charcount=0;
2697 SV *re_trie_maxbuff;
2698 GET_RE_DEBUG_FLAGS_DECL;
2700 PERL_ARGS_ASSERT_MAKE_TRIE;
2702 PERL_UNUSED_ARG(depth);
2706 case EXACT: case EXACT_REQ8: case EXACTL: break;
2710 case EXACTFLU8: folder = PL_fold_latin1; break;
2711 case EXACTF: folder = PL_fold; break;
2712 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2715 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2717 trie->startstate = 1;
2718 trie->wordcount = word_count;
2719 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2720 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2721 if (flags == EXACT || flags == EXACT_REQ8 || flags == EXACTL)
2722 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2723 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2724 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2727 trie_words = newAV();
2730 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, GV_ADD);
2731 assert(re_trie_maxbuff);
2732 if (!SvIOK(re_trie_maxbuff)) {
2733 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2735 DEBUG_TRIE_COMPILE_r({
2736 Perl_re_indentf( aTHX_
2737 "make_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2739 REG_NODE_NUM(startbranch), REG_NODE_NUM(first),
2740 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2743 /* Find the node we are going to overwrite */
2744 if ( first == startbranch && OP( last ) != BRANCH ) {
2745 /* whole branch chain */
2748 /* branch sub-chain */
2749 convert = NEXTOPER( first );
2752 /* -- First loop and Setup --
2754 We first traverse the branches and scan each word to determine if it
2755 contains widechars, and how many unique chars there are, this is
2756 important as we have to build a table with at least as many columns as we
2759 We use an array of integers to represent the character codes 0..255
2760 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2761 the native representation of the character value as the key and IV's for
2764 *TODO* If we keep track of how many times each character is used we can
2765 remap the columns so that the table compression later on is more
2766 efficient in terms of memory by ensuring the most common value is in the
2767 middle and the least common are on the outside. IMO this would be better
2768 than a most to least common mapping as theres a decent chance the most
2769 common letter will share a node with the least common, meaning the node
2770 will not be compressible. With a middle is most common approach the worst
2771 case is when we have the least common nodes twice.
2775 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2776 regnode *noper = NEXTOPER( cur );
2780 U32 wordlen = 0; /* required init */
2781 STRLEN minchars = 0;
2782 STRLEN maxchars = 0;
2783 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2786 if (OP(noper) == NOTHING) {
2787 /* skip past a NOTHING at the start of an alternation
2788 * eg, /(?:)a|(?:b)/ should be the same as /a|b/
2790 regnode *noper_next= regnext(noper);
2791 if (noper_next < tail)
2796 && ( OP(noper) == flags
2797 || (flags == EXACT && OP(noper) == EXACT_REQ8)
2798 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
2799 || OP(noper) == EXACTFUP))))
2801 uc= (U8*)STRING(noper);
2802 e= uc + STR_LEN(noper);
2809 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2810 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2811 regardless of encoding */
2812 if (OP( noper ) == EXACTFUP) {
2813 /* false positives are ok, so just set this */
2814 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2818 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2820 TRIE_CHARCOUNT(trie)++;
2823 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2824 * is in effect. Under /i, this character can match itself, or
2825 * anything that folds to it. If not under /i, it can match just
2826 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2827 * all fold to k, and all are single characters. But some folds
2828 * expand to more than one character, so for example LATIN SMALL
2829 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2830 * the string beginning at 'uc' is 'ffi', it could be matched by
2831 * three characters, or just by the one ligature character. (It
2832 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2833 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2834 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2835 * match.) The trie needs to know the minimum and maximum number
2836 * of characters that could match so that it can use size alone to
2837 * quickly reject many match attempts. The max is simple: it is
2838 * the number of folded characters in this branch (since a fold is
2839 * never shorter than what folds to it. */
2843 /* And the min is equal to the max if not under /i (indicated by
2844 * 'folder' being NULL), or there are no multi-character folds. If
2845 * there is a multi-character fold, the min is incremented just
2846 * once, for the character that folds to the sequence. Each
2847 * character in the sequence needs to be added to the list below of
2848 * characters in the trie, but we count only the first towards the
2849 * min number of characters needed. This is done through the
2850 * variable 'foldlen', which is returned by the macros that look
2851 * for these sequences as the number of bytes the sequence
2852 * occupies. Each time through the loop, we decrement 'foldlen' by
2853 * how many bytes the current char occupies. Only when it reaches
2854 * 0 do we increment 'minchars' or look for another multi-character
2856 if (folder == NULL) {
2859 else if (foldlen > 0) {
2860 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2865 /* See if *uc is the beginning of a multi-character fold. If
2866 * so, we decrement the length remaining to look at, to account
2867 * for the current character this iteration. (We can use 'uc'
2868 * instead of the fold returned by TRIE_READ_CHAR because for
2869 * non-UTF, the latin1_safe macro is smart enough to account
2870 * for all the unfolded characters, and because for UTF, the
2871 * string will already have been folded earlier in the
2872 * compilation process */
2874 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2875 foldlen -= UTF8SKIP(uc);
2878 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2883 /* The current character (and any potential folds) should be added
2884 * to the possible matching characters for this position in this
2888 U8 folded= folder[ (U8) uvc ];
2889 if ( !trie->charmap[ folded ] ) {
2890 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2891 TRIE_STORE_REVCHAR( folded );
2894 if ( !trie->charmap[ uvc ] ) {
2895 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2896 TRIE_STORE_REVCHAR( uvc );
2899 /* store the codepoint in the bitmap, and its folded
2901 TRIE_BITMAP_SET_FOLDED(trie, uvc, folder);
2902 set_bit = 0; /* We've done our bit :-) */
2906 /* XXX We could come up with the list of code points that fold
2907 * to this using PL_utf8_foldclosures, except not for
2908 * multi-char folds, as there may be multiple combinations
2909 * there that could work, which needs to wait until runtime to
2910 * resolve (The comment about LIGATURE FFI above is such an
2915 widecharmap = newHV();
2917 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2920 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%" UVXf, uvc );
2922 if ( !SvTRUE( *svpp ) ) {
2923 sv_setiv( *svpp, ++trie->uniquecharcount );
2924 TRIE_STORE_REVCHAR(uvc);
2927 } /* end loop through characters in this branch of the trie */
2929 /* We take the min and max for this branch and combine to find the min
2930 * and max for all branches processed so far */
2931 if( cur == first ) {
2932 trie->minlen = minchars;
2933 trie->maxlen = maxchars;
2934 } else if (minchars < trie->minlen) {
2935 trie->minlen = minchars;
2936 } else if (maxchars > trie->maxlen) {
2937 trie->maxlen = maxchars;
2939 } /* end first pass */
2940 DEBUG_TRIE_COMPILE_r(
2941 Perl_re_indentf( aTHX_
2942 "TRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2944 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2945 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2946 (int)trie->minlen, (int)trie->maxlen )
2950 We now know what we are dealing with in terms of unique chars and
2951 string sizes so we can calculate how much memory a naive
2952 representation using a flat table will take. If it's over a reasonable
2953 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2954 conservative but potentially much slower representation using an array
2957 At the end we convert both representations into the same compressed
2958 form that will be used in regexec.c for matching with. The latter
2959 is a form that cannot be used to construct with but has memory
2960 properties similar to the list form and access properties similar
2961 to the table form making it both suitable for fast searches and
2962 small enough that its feasable to store for the duration of a program.
2964 See the comment in the code where the compressed table is produced
2965 inplace from the flat tabe representation for an explanation of how
2966 the compression works.
2971 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2974 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2975 > SvIV(re_trie_maxbuff) )
2978 Second Pass -- Array Of Lists Representation
2980 Each state will be represented by a list of charid:state records
2981 (reg_trie_trans_le) the first such element holds the CUR and LEN
2982 points of the allocated array. (See defines above).
2984 We build the initial structure using the lists, and then convert
2985 it into the compressed table form which allows faster lookups
2986 (but cant be modified once converted).
2989 STRLEN transcount = 1;
2991 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using list compiler\n",
2994 trie->states = (reg_trie_state *)
2995 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2996 sizeof(reg_trie_state) );
3000 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3002 regnode *noper = NEXTOPER( cur );
3003 U32 state = 1; /* required init */
3004 U16 charid = 0; /* sanity init */
3005 U32 wordlen = 0; /* required init */
3007 if (OP(noper) == NOTHING) {
3008 regnode *noper_next= regnext(noper);
3009 if (noper_next < tail)
3014 && ( OP(noper) == flags
3015 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3016 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3017 || OP(noper) == EXACTFUP))))
3019 const U8 *uc= (U8*)STRING(noper);
3020 const U8 *e= uc + STR_LEN(noper);
3022 for ( ; uc < e ; uc += len ) {
3027 charid = trie->charmap[ uvc ];
3029 SV** const svpp = hv_fetch( widecharmap,
3036 charid=(U16)SvIV( *svpp );
3039 /* charid is now 0 if we dont know the char read, or
3040 * nonzero if we do */
3047 if ( !trie->states[ state ].trans.list ) {
3048 TRIE_LIST_NEW( state );
3051 check <= TRIE_LIST_USED( state );
3054 if ( TRIE_LIST_ITEM( state, check ).forid
3057 newstate = TRIE_LIST_ITEM( state, check ).newstate;
3062 newstate = next_alloc++;
3063 prev_states[newstate] = state;
3064 TRIE_LIST_PUSH( state, charid, newstate );
3069 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3073 TRIE_HANDLE_WORD(state);
3075 } /* end second pass */
3077 /* next alloc is the NEXT state to be allocated */
3078 trie->statecount = next_alloc;
3079 trie->states = (reg_trie_state *)
3080 PerlMemShared_realloc( trie->states,
3082 * sizeof(reg_trie_state) );
3084 /* and now dump it out before we compress it */
3085 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
3086 revcharmap, next_alloc,
3090 trie->trans = (reg_trie_trans *)
3091 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
3098 for( state=1 ; state < next_alloc ; state ++ ) {
3102 DEBUG_TRIE_COMPILE_MORE_r(
3103 Perl_re_printf( aTHX_ "tp: %d zp: %d ",tp,zp)
3107 if (trie->states[state].trans.list) {
3108 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
3112 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3113 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
3114 if ( forid < minid ) {
3116 } else if ( forid > maxid ) {
3120 if ( transcount < tp + maxid - minid + 1) {
3122 trie->trans = (reg_trie_trans *)
3123 PerlMemShared_realloc( trie->trans,
3125 * sizeof(reg_trie_trans) );
3126 Zero( trie->trans + (transcount / 2),
3130 base = trie->uniquecharcount + tp - minid;
3131 if ( maxid == minid ) {
3133 for ( ; zp < tp ; zp++ ) {
3134 if ( ! trie->trans[ zp ].next ) {
3135 base = trie->uniquecharcount + zp - minid;
3136 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
3138 trie->trans[ zp ].check = state;
3144 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
3146 trie->trans[ tp ].check = state;
3151 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
3152 const U32 tid = base
3153 - trie->uniquecharcount
3154 + TRIE_LIST_ITEM( state, idx ).forid;
3155 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
3157 trie->trans[ tid ].check = state;
3159 tp += ( maxid - minid + 1 );
3161 Safefree(trie->states[ state ].trans.list);
3164 DEBUG_TRIE_COMPILE_MORE_r(
3165 Perl_re_printf( aTHX_ " base: %d\n",base);
3168 trie->states[ state ].trans.base=base;
3170 trie->lasttrans = tp + 1;
3174 Second Pass -- Flat Table Representation.
3176 we dont use the 0 slot of either trans[] or states[] so we add 1 to
3177 each. We know that we will need Charcount+1 trans at most to store
3178 the data (one row per char at worst case) So we preallocate both
3179 structures assuming worst case.
3181 We then construct the trie using only the .next slots of the entry
3184 We use the .check field of the first entry of the node temporarily
3185 to make compression both faster and easier by keeping track of how
3186 many non zero fields are in the node.
3188 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
3191 There are two terms at use here: state as a TRIE_NODEIDX() which is
3192 a number representing the first entry of the node, and state as a
3193 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
3194 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
3195 if there are 2 entrys per node. eg:
3203 The table is internally in the right hand, idx form. However as we
3204 also have to deal with the states array which is indexed by nodenum
3205 we have to use TRIE_NODENUM() to convert.
3208 DEBUG_TRIE_COMPILE_MORE_r( Perl_re_indentf( aTHX_ "Compiling trie using table compiler\n",
3211 trie->trans = (reg_trie_trans *)
3212 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
3213 * trie->uniquecharcount + 1,
3214 sizeof(reg_trie_trans) );
3215 trie->states = (reg_trie_state *)
3216 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
3217 sizeof(reg_trie_state) );
3218 next_alloc = trie->uniquecharcount + 1;
3221 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
3223 regnode *noper = NEXTOPER( cur );
3225 U32 state = 1; /* required init */
3227 U16 charid = 0; /* sanity init */
3228 U32 accept_state = 0; /* sanity init */
3230 U32 wordlen = 0; /* required init */
3232 if (OP(noper) == NOTHING) {
3233 regnode *noper_next= regnext(noper);
3234 if (noper_next < tail)
3239 && ( OP(noper) == flags
3240 || (flags == EXACT && OP(noper) == EXACT_REQ8)
3241 || (flags == EXACTFU && ( OP(noper) == EXACTFU_REQ8
3242 || OP(noper) == EXACTFUP))))
3244 const U8 *uc= (U8*)STRING(noper);
3245 const U8 *e= uc + STR_LEN(noper);
3247 for ( ; uc < e ; uc += len ) {
3252 charid = trie->charmap[ uvc ];
3254 SV* const * const svpp = hv_fetch( widecharmap,
3258 charid = svpp ? (U16)SvIV(*svpp) : 0;
3262 if ( !trie->trans[ state + charid ].next ) {
3263 trie->trans[ state + charid ].next = next_alloc;
3264 trie->trans[ state ].check++;
3265 prev_states[TRIE_NODENUM(next_alloc)]
3266 = TRIE_NODENUM(state);
3267 next_alloc += trie->uniquecharcount;
3269 state = trie->trans[ state + charid ].next;
3271 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %" IVdf, uvc );
3273 /* charid is now 0 if we dont know the char read, or
3274 * nonzero if we do */
3277 accept_state = TRIE_NODENUM( state );
3278 TRIE_HANDLE_WORD(accept_state);
3280 } /* end second pass */
3282 /* and now dump it out before we compress it */
3283 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
3285 next_alloc, depth+1));
3289 * Inplace compress the table.*
3291 For sparse data sets the table constructed by the trie algorithm will
3292 be mostly 0/FAIL transitions or to put it another way mostly empty.
3293 (Note that leaf nodes will not contain any transitions.)
3295 This algorithm compresses the tables by eliminating most such
3296 transitions, at the cost of a modest bit of extra work during lookup:
3298 - Each states[] entry contains a .base field which indicates the
3299 index in the state[] array wheres its transition data is stored.
3301 - If .base is 0 there are no valid transitions from that node.
3303 - If .base is nonzero then charid is added to it to find an entry in
3306 -If trans[states[state].base+charid].check!=state then the
3307 transition is taken to be a 0/Fail transition. Thus if there are fail
3308 transitions at the front of the node then the .base offset will point
3309 somewhere inside the previous nodes data (or maybe even into a node
3310 even earlier), but the .check field determines if the transition is
3314 The following process inplace converts the table to the compressed
3315 table: We first do not compress the root node 1,and mark all its
3316 .check pointers as 1 and set its .base pointer as 1 as well. This
3317 allows us to do a DFA construction from the compressed table later,
3318 and ensures that any .base pointers we calculate later are greater
3321 - We set 'pos' to indicate the first entry of the second node.
3323 - We then iterate over the columns of the node, finding the first and
3324 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
3325 and set the .check pointers accordingly, and advance pos
3326 appropriately and repreat for the next node. Note that when we copy
3327 the next pointers we have to convert them from the original
3328 NODEIDX form to NODENUM form as the former is not valid post
3331 - If a node has no transitions used we mark its base as 0 and do not
3332 advance the pos pointer.
3334 - If a node only has one transition we use a second pointer into the
3335 structure to fill in allocated fail transitions from other states.
3336 This pointer is independent of the main pointer and scans forward
3337 looking for null transitions that are allocated to a state. When it
3338 finds one it writes the single transition into the "hole". If the
3339 pointer doesnt find one the single transition is appended as normal.
3341 - Once compressed we can Renew/realloc the structures to release the
3344 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
3345 specifically Fig 3.47 and the associated pseudocode.
3349 const U32 laststate = TRIE_NODENUM( next_alloc );
3352 trie->statecount = laststate;
3354 for ( state = 1 ; state < laststate ; state++ ) {
3356 const U32 stateidx = TRIE_NODEIDX( state );
3357 const U32 o_used = trie->trans[ stateidx ].check;
3358 U32 used = trie->trans[ stateidx ].check;
3359 trie->trans[ stateidx ].check = 0;
3362 used && charid < trie->uniquecharcount;
3365 if ( flag || trie->trans[ stateidx + charid ].next ) {
3366 if ( trie->trans[ stateidx + charid ].next ) {
3368 for ( ; zp < pos ; zp++ ) {
3369 if ( ! trie->trans[ zp ].next ) {
3373 trie->states[ state ].trans.base
3375 + trie->uniquecharcount
3377 trie->trans[ zp ].next
3378 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
3380 trie->trans[ zp ].check = state;
3381 if ( ++zp > pos ) pos = zp;
3388 trie->states[ state ].trans.base
3389 = pos + trie->uniquecharcount - charid ;
3391 trie->trans[ pos ].next
3392 = SAFE_TRIE_NODENUM(
3393 trie->trans[ stateidx + charid ].next );
3394 trie->trans[ pos ].check = state;
3399 trie->lasttrans = pos + 1;
3400 trie->states = (reg_trie_state *)
3401 PerlMemShared_realloc( trie->states, laststate
3402 * sizeof(reg_trie_state) );
3403 DEBUG_TRIE_COMPILE_MORE_r(
3404 Perl_re_indentf( aTHX_ "Alloc: %d Orig: %" IVdf " elements, Final:%" IVdf ". Savings of %%%5.2f\n",
3406 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
3410 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
3413 } /* end table compress */
3415 DEBUG_TRIE_COMPILE_MORE_r(
3416 Perl_re_indentf( aTHX_ "Statecount:%" UVxf " Lasttrans:%" UVxf "\n",
3418 (UV)trie->statecount,
3419 (UV)trie->lasttrans)
3421 /* resize the trans array to remove unused space */
3422 trie->trans = (reg_trie_trans *)
3423 PerlMemShared_realloc( trie->trans, trie->lasttrans
3424 * sizeof(reg_trie_trans) );
3426 { /* Modify the program and insert the new TRIE node */
3427 U8 nodetype =(U8)(flags & 0xFF);
3431 regnode *optimize = NULL;
3432 #ifdef RE_TRACK_PATTERN_OFFSETS
3435 U32 mjd_nodelen = 0;
3436 #endif /* RE_TRACK_PATTERN_OFFSETS */
3437 #endif /* DEBUGGING */
3439 This means we convert either the first branch or the first Exact,
3440 depending on whether the thing following (in 'last') is a branch
3441 or not and whther first is the startbranch (ie is it a sub part of
3442 the alternation or is it the whole thing.)
3443 Assuming its a sub part we convert the EXACT otherwise we convert
3444 the whole branch sequence, including the first.
3446 /* Find the node we are going to overwrite */
3447 if ( first != startbranch || OP( last ) == BRANCH ) {
3448 /* branch sub-chain */
3449 NEXT_OFF( first ) = (U16)(last - first);
3450 #ifdef RE_TRACK_PATTERN_OFFSETS
3452 mjd_offset= Node_Offset((convert));
3453 mjd_nodelen= Node_Length((convert));
3456 /* whole branch chain */
3458 #ifdef RE_TRACK_PATTERN_OFFSETS
3461 const regnode *nop = NEXTOPER( convert );
3462 mjd_offset= Node_Offset((nop));
3463 mjd_nodelen= Node_Length((nop));
3467 Perl_re_indentf( aTHX_ "MJD offset:%" UVuf " MJD length:%" UVuf "\n",
3469 (UV)mjd_offset, (UV)mjd_nodelen)
3472 /* But first we check to see if there is a common prefix we can
3473 split out as an EXACT and put in front of the TRIE node. */
3474 trie->startstate= 1;
3475 if ( trie->bitmap && !widecharmap && !trie->jump ) {
3476 /* we want to find the first state that has more than
3477 * one transition, if that state is not the first state
3478 * then we have a common prefix which we can remove.
3481 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
3483 I32 first_ofs = -1; /* keeps track of the ofs of the first
3484 transition, -1 means none */
3486 const U32 base = trie->states[ state ].trans.base;
3488 /* does this state terminate an alternation? */
3489 if ( trie->states[state].wordnum )
3492 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
3493 if ( ( base + ofs >= trie->uniquecharcount ) &&
3494 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
3495 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
3497 if ( ++count > 1 ) {
3498 /* we have more than one transition */
3501 /* if this is the first state there is no common prefix
3502 * to extract, so we can exit */
3503 if ( state == 1 ) break;
3504 tmp = av_fetch( revcharmap, ofs, 0);
3505 ch = (U8*)SvPV_nolen_const( *tmp );
3507 /* if we are on count 2 then we need to initialize the
3508 * bitmap, and store the previous char if there was one
3511 /* clear the bitmap */
3512 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
3514 Perl_re_indentf( aTHX_ "New Start State=%" UVuf " Class: [",
3517 if (first_ofs >= 0) {
3518 SV ** const tmp = av_fetch( revcharmap, first_ofs, 0);
3519 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3521 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3523 Perl_re_printf( aTHX_ "%s", (char*)ch)
3527 /* store the current firstchar in the bitmap */
3528 TRIE_BITMAP_SET_FOLDED(trie,*ch, folder);
3529 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "%s", ch));
3535 /* This state has only one transition, its transition is part
3536 * of a common prefix - we need to concatenate the char it
3537 * represents to what we have so far. */
3538 SV **tmp = av_fetch( revcharmap, first_ofs, 0);
3540 char *ch = SvPV( *tmp, len );
3542 SV *sv=sv_newmortal();
3543 Perl_re_indentf( aTHX_ "Prefix State: %" UVuf " Ofs:%" UVuf " Char='%s'\n",
3545 (UV)state, (UV)first_ofs,
3546 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3547 PL_colors[0], PL_colors[1],
3548 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3549 PERL_PV_ESCAPE_FIRSTCHAR
3554 OP( convert ) = nodetype;
3555 str=STRING(convert);
3556 setSTR_LEN(convert, 0);
3558 setSTR_LEN(convert, STR_LEN(convert) + len);
3564 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "]\n"));
3569 trie->prefixlen = (state-1);
3571 regnode *n = convert+NODE_SZ_STR(convert);
3572 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3573 trie->startstate = state;
3574 trie->minlen -= (state - 1);
3575 trie->maxlen -= (state - 1);
3577 /* At least the UNICOS C compiler choked on this
3578 * being argument to DEBUG_r(), so let's just have
3581 #ifdef PERL_EXT_RE_BUILD
3587 regnode *fix = convert;
3588 U32 word = trie->wordcount;
3589 #ifdef RE_TRACK_PATTERN_OFFSETS
3592 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3593 while( ++fix < n ) {
3594 Set_Node_Offset_Length(fix, 0, 0);
3597 SV ** const tmp = av_fetch( trie_words, word, 0 );
3599 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3600 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3602 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3610 NEXT_OFF(convert) = (U16)(tail - convert);
3611 DEBUG_r(optimize= n);
3617 if ( trie->maxlen ) {
3618 NEXT_OFF( convert ) = (U16)(tail - convert);
3619 ARG_SET( convert, data_slot );
3620 /* Store the offset to the first unabsorbed branch in
3621 jump[0], which is otherwise unused by the jump logic.
3622 We use this when dumping a trie and during optimisation. */
3624 trie->jump[0] = (U16)(nextbranch - convert);
3626 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3627 * and there is a bitmap
3628 * and the first "jump target" node we found leaves enough room
3629 * then convert the TRIE node into a TRIEC node, with the bitmap
3630 * embedded inline in the opcode - this is hypothetically faster.
3632 if ( !trie->states[trie->startstate].wordnum
3634 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3636 OP( convert ) = TRIEC;
3637 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3638 PerlMemShared_free(trie->bitmap);
3641 OP( convert ) = TRIE;
3643 /* store the type in the flags */
3644 convert->flags = nodetype;
3648 + regarglen[ OP( convert ) ];
3650 /* XXX We really should free up the resource in trie now,
3651 as we won't use them - (which resources?) dmq */
3653 /* needed for dumping*/
3654 DEBUG_r(if (optimize) {
3655 regnode *opt = convert;
3657 while ( ++opt < optimize) {
3658 Set_Node_Offset_Length(opt, 0, 0);
3661 Try to clean up some of the debris left after the
3664 while( optimize < jumper ) {
3665 Track_Code( mjd_nodelen += Node_Length((optimize)); );
3666 OP( optimize ) = OPTIMIZED;
3667 Set_Node_Offset_Length(optimize, 0, 0);
3670 Set_Node_Offset_Length(convert, mjd_offset, mjd_nodelen);
3672 } /* end node insert */
3674 /* Finish populating the prev field of the wordinfo array. Walk back
3675 * from each accept state until we find another accept state, and if
3676 * so, point the first word's .prev field at the second word. If the
3677 * second already has a .prev field set, stop now. This will be the
3678 * case either if we've already processed that word's accept state,
3679 * or that state had multiple words, and the overspill words were
3680 * already linked up earlier.
3687 for (word=1; word <= trie->wordcount; word++) {
3689 if (trie->wordinfo[word].prev)
3691 state = trie->wordinfo[word].accept;
3693 state = prev_states[state];
3696 prev = trie->states[state].wordnum;
3700 trie->wordinfo[word].prev = prev;
3702 Safefree(prev_states);
3706 /* and now dump out the compressed format */
3707 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3709 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3711 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3712 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3714 SvREFCNT_dec_NN(revcharmap);
3718 : trie->startstate>1
3724 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3726 /* The Trie is constructed and compressed now so we can build a fail array if
3729 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3731 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3735 We find the fail state for each state in the trie, this state is the longest
3736 proper suffix of the current state's 'word' that is also a proper prefix of
3737 another word in our trie. State 1 represents the word '' and is thus the
3738 default fail state. This allows the DFA not to have to restart after its
3739 tried and failed a word at a given point, it simply continues as though it
3740 had been matching the other word in the first place.
3742 'abcdgu'=~/abcdefg|cdgu/
3743 When we get to 'd' we are still matching the first word, we would encounter
3744 'g' which would fail, which would bring us to the state representing 'd' in
3745 the second word where we would try 'g' and succeed, proceeding to match
3748 /* add a fail transition */
3749 const U32 trie_offset = ARG(source);
3750 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3752 const U32 ucharcount = trie->uniquecharcount;
3753 const U32 numstates = trie->statecount;
3754 const U32 ubound = trie->lasttrans + ucharcount;
3758 U32 base = trie->states[ 1 ].trans.base;
3761 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3763 GET_RE_DEBUG_FLAGS_DECL;
3765 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3766 PERL_UNUSED_CONTEXT;
3768 PERL_UNUSED_ARG(depth);
3771 if ( OP(source) == TRIE ) {
3772 struct regnode_1 *op = (struct regnode_1 *)
3773 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3774 StructCopy(source, op, struct regnode_1);
3775 stclass = (regnode *)op;
3777 struct regnode_charclass *op = (struct regnode_charclass *)
3778 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3779 StructCopy(source, op, struct regnode_charclass);
3780 stclass = (regnode *)op;
3782 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3784 ARG_SET( stclass, data_slot );
3785 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3786 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3787 aho->trie=trie_offset;
3788 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3789 Copy( trie->states, aho->states, numstates, reg_trie_state );
3790 Newx( q, numstates, U32);
3791 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3794 /* initialize fail[0..1] to be 1 so that we always have
3795 a valid final fail state */
3796 fail[ 0 ] = fail[ 1 ] = 1;
3798 for ( charid = 0; charid < ucharcount ; charid++ ) {
3799 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3801 q[ q_write ] = newstate;
3802 /* set to point at the root */
3803 fail[ q[ q_write++ ] ]=1;
3806 while ( q_read < q_write) {
3807 const U32 cur = q[ q_read++ % numstates ];
3808 base = trie->states[ cur ].trans.base;
3810 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3811 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3813 U32 fail_state = cur;
3816 fail_state = fail[ fail_state ];
3817 fail_base = aho->states[ fail_state ].trans.base;
3818 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3820 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3821 fail[ ch_state ] = fail_state;
3822 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3824 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3826 q[ q_write++ % numstates] = ch_state;
3830 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3831 when we fail in state 1, this allows us to use the
3832 charclass scan to find a valid start char. This is based on the principle
3833 that theres a good chance the string being searched contains lots of stuff
3834 that cant be a start char.
3836 fail[ 0 ] = fail[ 1 ] = 0;
3837 DEBUG_TRIE_COMPILE_r({
3838 Perl_re_indentf( aTHX_ "Stclass Failtable (%" UVuf " states): 0",
3839 depth, (UV)numstates
3841 for( q_read=1; q_read<numstates; q_read++ ) {
3842 Perl_re_printf( aTHX_ ", %" UVuf, (UV)fail[q_read]);
3844 Perl_re_printf( aTHX_ "\n");
3847 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3852 /* The below joins as many adjacent EXACTish nodes as possible into a single
3853 * one. The regop may be changed if the node(s) contain certain sequences that
3854 * require special handling. The joining is only done if:
3855 * 1) there is room in the current conglomerated node to entirely contain the
3857 * 2) they are compatible node types
3859 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3860 * these get optimized out
3862 * XXX khw thinks this should be enhanced to fill EXACT (at least) nodes as full
3863 * as possible, even if that means splitting an existing node so that its first
3864 * part is moved to the preceeding node. This would maximise the efficiency of
3865 * memEQ during matching.
3867 * If a node is to match under /i (folded), the number of characters it matches
3868 * can be different than its character length if it contains a multi-character
3869 * fold. *min_subtract is set to the total delta number of characters of the
3872 * And *unfolded_multi_char is set to indicate whether or not the node contains
3873 * an unfolded multi-char fold. This happens when it won't be known until
3874 * runtime whether the fold is valid or not; namely
3875 * 1) for EXACTF nodes that contain LATIN SMALL LETTER SHARP S, as only if the
3876 * target string being matched against turns out to be UTF-8 is that fold
3878 * 2) for EXACTFL nodes whose folding rules depend on the locale in force at
3880 * (Multi-char folds whose components are all above the Latin1 range are not
3881 * run-time locale dependent, and have already been folded by the time this
3882 * function is called.)
3884 * This is as good a place as any to discuss the design of handling these
3885 * multi-character fold sequences. It's been wrong in Perl for a very long
3886 * time. There are three code points in Unicode whose multi-character folds
3887 * were long ago discovered to mess things up. The previous designs for
3888 * dealing with these involved assigning a special node for them. This
3889 * approach doesn't always work, as evidenced by this example:
3890 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3891 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3892 * would match just the \xDF, it won't be able to handle the case where a
3893 * successful match would have to cross the node's boundary. The new approach
3894 * that hopefully generally solves the problem generates an EXACTFUP node
3895 * that is "sss" in this case.
3897 * It turns out that there are problems with all multi-character folds, and not
3898 * just these three. Now the code is general, for all such cases. The
3899 * approach taken is:
3900 * 1) This routine examines each EXACTFish node that could contain multi-
3901 * character folded sequences. Since a single character can fold into
3902 * such a sequence, the minimum match length for this node is less than
3903 * the number of characters in the node. This routine returns in
3904 * *min_subtract how many characters to subtract from the the actual
3905 * length of the string to get a real minimum match length; it is 0 if
3906 * there are no multi-char foldeds. This delta is used by the caller to
3907 * adjust the min length of the match, and the delta between min and max,
3908 * so that the optimizer doesn't reject these possibilities based on size
3911 * 2) For the sequence involving the LATIN SMALL LETTER SHARP S (U+00DF)
3912 * under /u, we fold it to 'ss' in regatom(), and in this routine, after
3913 * joining, we scan for occurrences of the sequence 'ss' in non-UTF-8
3914 * EXACTFU nodes. The node type of such nodes is then changed to
3915 * EXACTFUP, indicating it is problematic, and needs careful handling.
3916 * (The procedures in step 1) above are sufficient to handle this case in
3917 * UTF-8 encoded nodes.) The reason this is problematic is that this is
3918 * the only case where there is a possible fold length change in non-UTF-8
3919 * patterns. By reserving a special node type for problematic cases, the
3920 * far more common regular EXACTFU nodes can be processed faster.
3921 * regexec.c takes advantage of this.
3923 * EXACTFUP has been created as a grab-bag for (hopefully uncommon)
3924 * problematic cases. These all only occur when the pattern is not
3925 * UTF-8. In addition to the 'ss' sequence where there is a possible fold
3926 * length change, it handles the situation where the string cannot be
3927 * entirely folded. The strings in an EXACTFish node are folded as much
3928 * as possible during compilation in regcomp.c. This saves effort in
3929 * regex matching. By using an EXACTFUP node when it is not possible to
3930 * fully fold at compile time, regexec.c can know that everything in an
3931 * EXACTFU node is folded, so folding can be skipped at runtime. The only
3932 * case where folding in EXACTFU nodes can't be done at compile time is
3933 * the presumably uncommon MICRO SIGN, when the pattern isn't UTF-8. This
3934 * is because its fold requires UTF-8 to represent. Thus EXACTFUP nodes
3935 * handle two very different cases. Alternatively, there could have been
3936 * a node type where there are length changes, one for unfolded, and one
3937 * for both. If yet another special case needed to be created, the number
3938 * of required node types would have to go to 7. khw figures that even
3939 * though there are plenty of node types to spare, that the maintenance
3940 * cost wasn't worth the small speedup of doing it that way, especially
3941 * since he thinks the MICRO SIGN is rarely encountered in practice.
3943 * There are other cases where folding isn't done at compile time, but
3944 * none of them are under /u, and hence not for EXACTFU nodes. The folds
3945 * in EXACTFL nodes aren't known until runtime, and vary as the locale
3946 * changes. Some folds in EXACTF depend on if the runtime target string
3947 * is UTF-8 or not. (regatom() will create an EXACTFU node even under /di
3948 * when no fold in it depends on the UTF-8ness of the target string.)
3950 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3951 * validity of the fold won't be known until runtime, and so must remain
3952 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFAA
3953 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3954 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3955 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3956 * The reason this is a problem is that the optimizer part of regexec.c
3957 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3958 * that a character in the pattern corresponds to at most a single
3959 * character in the target string. (And I do mean character, and not byte
3960 * here, unlike other parts of the documentation that have never been
3961 * updated to account for multibyte Unicode.) Sharp s in EXACTF and
3962 * EXACTFL nodes can match the two character string 'ss'; in EXACTFAA
3963 * nodes it can match "\x{17F}\x{17F}". These, along with other ones in
3964 * EXACTFL nodes, violate the assumption, and they are the only instances
3965 * where it is violated. I'm reluctant to try to change the assumption,
3966 * as the code involved is impenetrable to me (khw), so instead the code
3967 * here punts. This routine examines EXACTFL nodes, and (when the pattern
3968 * isn't UTF-8) EXACTF and EXACTFAA for such unfolded folds, and returns a
3969 * boolean indicating whether or not the node contains such a fold. When
3970 * it is true, the caller sets a flag that later causes the optimizer in
3971 * this file to not set values for the floating and fixed string lengths,
3972 * and thus avoids the optimizer code in regexec.c that makes the invalid
3973 * assumption. Thus, there is no optimization based on string lengths for
3974 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3975 * EXACTF and EXACTFAA nodes that contain the sharp s. (The reason the
3976 * assumption is wrong only in these cases is that all other non-UTF-8
3977 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3978 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3979 * EXACTF nodes because we don't know at compile time if it actually
3980 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3981 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3982 * always matches; and EXACTFAA where it never does. In an EXACTFAA node
3983 * in a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3984 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3985 * string would require the pattern to be forced into UTF-8, the overhead
3986 * of which we want to avoid. Similarly the unfolded multi-char folds in
3987 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3990 * Similarly, the code that generates tries doesn't currently handle
3991 * not-already-folded multi-char folds, and it looks like a pain to change
3992 * that. Therefore, trie generation of EXACTFAA nodes with the sharp s
3993 * doesn't work. Instead, such an EXACTFAA is turned into a new regnode,
3994 * EXACTFAA_NO_TRIE, which the trie code knows not to handle. Most people
3995 * using /iaa matching will be doing so almost entirely with ASCII
3996 * strings, so this should rarely be encountered in practice */
3998 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3999 if (PL_regkind[OP(scan)] == EXACT && OP(scan) != LEXACT \
4000 && OP(scan) != LEXACT_REQ8) \
4001 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags), NULL, depth+1)
4004 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
4005 UV *min_subtract, bool *unfolded_multi_char,
4006 U32 flags, regnode *val, U32 depth)
4008 /* Merge several consecutive EXACTish nodes into one. */
4010 regnode *n = regnext(scan);
4012 regnode *next = scan + NODE_SZ_STR(scan);
4016 regnode *stop = scan;
4017 GET_RE_DEBUG_FLAGS_DECL;
4019 PERL_UNUSED_ARG(depth);
4022 PERL_ARGS_ASSERT_JOIN_EXACT;
4023 #ifndef EXPERIMENTAL_INPLACESCAN
4024 PERL_UNUSED_ARG(flags);
4025 PERL_UNUSED_ARG(val);
4027 DEBUG_PEEP("join", scan, depth, 0);
4029 assert(PL_regkind[OP(scan)] == EXACT);
4031 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
4032 * EXACT ones that are mergeable to the current one. */
4034 && ( PL_regkind[OP(n)] == NOTHING
4035 || (stringok && PL_regkind[OP(n)] == EXACT))
4037 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
4040 if (OP(n) == TAIL || n > next)
4042 if (PL_regkind[OP(n)] == NOTHING) {
4043 DEBUG_PEEP("skip:", n, depth, 0);
4044 NEXT_OFF(scan) += NEXT_OFF(n);
4045 next = n + NODE_STEP_REGNODE;
4052 else if (stringok) {
4053 const unsigned int oldl = STR_LEN(scan);
4054 regnode * const nnext = regnext(n);
4056 /* XXX I (khw) kind of doubt that this works on platforms (should
4057 * Perl ever run on one) where U8_MAX is above 255 because of lots
4058 * of other assumptions */
4059 /* Don't join if the sum can't fit into a single node */
4060 if (oldl + STR_LEN(n) > U8_MAX)
4063 /* Joining something that requires UTF-8 with something that
4064 * doesn't, means the result requires UTF-8. */
4065 if (OP(scan) == EXACT && (OP(n) == EXACT_REQ8)) {
4066 OP(scan) = EXACT_REQ8;
4068 else if (OP(scan) == EXACT_REQ8 && (OP(n) == EXACT)) {
4069 ; /* join is compatible, no need to change OP */
4071 else if ((OP(scan) == EXACTFU) && (OP(n) == EXACTFU_REQ8)) {
4072 OP(scan) = EXACTFU_REQ8;
4074 else if ((OP(scan) == EXACTFU_REQ8) && (OP(n) == EXACTFU)) {
4075 ; /* join is compatible, no need to change OP */
4077 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU) {
4078 ; /* join is compatible, no need to change OP */
4080 else if (OP(scan) == EXACTFU && OP(n) == EXACTFU_S_EDGE) {
4082 /* Under /di, temporary EXACTFU_S_EDGE nodes are generated,
4083 * which can join with EXACTFU ones. We check for this case
4084 * here. These need to be resolved to either EXACTFU or
4085 * EXACTF at joining time. They have nothing in them that
4086 * would forbid them from being the more desirable EXACTFU
4087 * nodes except that they begin and/or end with a single [Ss].
4088 * The reason this is problematic is because they could be
4089 * joined in this loop with an adjacent node that ends and/or
4090 * begins with [Ss] which would then form the sequence 'ss',
4091 * which matches differently under /di than /ui, in which case
4092 * EXACTFU can't be used. If the 'ss' sequence doesn't get
4093 * formed, the nodes get absorbed into any adjacent EXACTFU
4094 * node. And if the only adjacent node is EXACTF, they get
4095 * absorbed into that, under the theory that a longer node is
4096 * better than two shorter ones, even if one is EXACTFU. Note
4097 * that EXACTFU_REQ8 is generated only for UTF-8 patterns,
4098 * and the EXACTFU_S_EDGE ones only for non-UTF-8. */
4100 if (STRING(n)[STR_LEN(n)-1] == 's') {
4102 /* Here the joined node would end with 's'. If the node
4103 * following the combination is an EXACTF one, it's better to
4104 * join this trailing edge 's' node with that one, leaving the
4105 * current one in 'scan' be the more desirable EXACTFU */
4106 if (OP(nnext) == EXACTF) {
4110 OP(scan) = EXACTFU_S_EDGE;
4112 } /* Otherwise, the beginning 's' of the 2nd node just
4113 becomes an interior 's' in 'scan' */
4115 else if (OP(scan) == EXACTF && OP(n) == EXACTF) {
4116 ; /* join is compatible, no need to change OP */
4118 else if (OP(scan) == EXACTF && OP(n) == EXACTFU_S_EDGE) {
4120 /* EXACTF nodes are compatible for joining with EXACTFU_S_EDGE
4121 * nodes. But the latter nodes can be also joined with EXACTFU
4122 * ones, and that is a better outcome, so if the node following
4123 * 'n' is EXACTFU, quit now so that those two can be joined
4125 if (OP(nnext) == EXACTFU) {
4129 /* The join is compatible, and the combined node will be
4130 * EXACTF. (These don't care if they begin or end with 's' */
4132 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU_S_EDGE) {
4133 if ( STRING(scan)[STR_LEN(scan)-1] == 's'
4134 && STRING(n)[0] == 's')
4136 /* When combined, we have the sequence 'ss', which means we
4137 * have to remain /di */
4141 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTFU) {
4142 if (STRING(n)[0] == 's') {
4143 ; /* Here the join is compatible and the combined node
4144 starts with 's', no need to change OP */
4146 else { /* Now the trailing 's' is in the interior */
4150 else if (OP(scan) == EXACTFU_S_EDGE && OP(n) == EXACTF) {
4152 /* The join is compatible, and the combined node will be
4153 * EXACTF. (These don't care if they begin or end with 's' */
4156 else if (OP(scan) != OP(n)) {
4158 /* The only other compatible joinings are the same node type */
4162 DEBUG_PEEP("merg", n, depth, 0);
4165 NEXT_OFF(scan) += NEXT_OFF(n);
4166 setSTR_LEN(scan, STR_LEN(scan) + STR_LEN(n));
4167 next = n + NODE_SZ_STR(n);
4168 /* Now we can overwrite *n : */
4169 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
4177 #ifdef EXPERIMENTAL_INPLACESCAN
4178 if (flags && !NEXT_OFF(n)) {
4179 DEBUG_PEEP("atch", val, depth, 0);
4180 if (reg_off_by_arg[OP(n)]) {
4181 ARG_SET(n, val - n);
4184 NEXT_OFF(n) = val - n;
4191 /* This temporary node can now be turned into EXACTFU, and must, as
4192 * regexec.c doesn't handle it */
4193 if (OP(scan) == EXACTFU_S_EDGE) {
4198 *unfolded_multi_char = FALSE;
4200 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
4201 * can now analyze for sequences of problematic code points. (Prior to
4202 * this final joining, sequences could have been split over boundaries, and
4203 * hence missed). The sequences only happen in folding, hence for any
4204 * non-EXACT EXACTish node */
4205 if (OP(scan) != EXACT && OP(scan) != EXACT_REQ8 && OP(scan) != EXACTL) {
4206 U8* s0 = (U8*) STRING(scan);
4208 U8* s_end = s0 + STR_LEN(scan);
4210 int total_count_delta = 0; /* Total delta number of characters that
4211 multi-char folds expand to */
4213 /* One pass is made over the node's string looking for all the
4214 * possibilities. To avoid some tests in the loop, there are two main
4215 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
4220 if (OP(scan) == EXACTFL) {
4223 /* An EXACTFL node would already have been changed to another
4224 * node type unless there is at least one character in it that
4225 * is problematic; likely a character whose fold definition
4226 * won't be known until runtime, and so has yet to be folded.
4227 * For all but the UTF-8 locale, folds are 1-1 in length, but
4228 * to handle the UTF-8 case, we need to create a temporary
4229 * folded copy using UTF-8 locale rules in order to analyze it.
4230 * This is because our macros that look to see if a sequence is
4231 * a multi-char fold assume everything is folded (otherwise the
4232 * tests in those macros would be too complicated and slow).
4233 * Note that here, the non-problematic folds will have already
4234 * been done, so we can just copy such characters. We actually
4235 * don't completely fold the EXACTFL string. We skip the
4236 * unfolded multi-char folds, as that would just create work
4237 * below to figure out the size they already are */
4239 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
4242 STRLEN s_len = UTF8SKIP(s);
4243 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
4244 Copy(s, d, s_len, U8);
4247 else if (is_FOLDS_TO_MULTI_utf8(s)) {
4248 *unfolded_multi_char = TRUE;
4249 Copy(s, d, s_len, U8);
4252 else if (isASCII(*s)) {
4253 *(d++) = toFOLD(*s);
4257 _toFOLD_utf8_flags(s, s_end, d, &len, FOLD_FLAGS_FULL);
4263 /* Point the remainder of the routine to look at our temporary
4267 } /* End of creating folded copy of EXACTFL string */
4269 /* Examine the string for a multi-character fold sequence. UTF-8
4270 * patterns have all characters pre-folded by the time this code is
4272 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
4273 length sequence we are looking for is 2 */
4275 int count = 0; /* How many characters in a multi-char fold */
4276 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
4277 if (! len) { /* Not a multi-char fold: get next char */
4282 { /* Here is a generic multi-char fold. */
4283 U8* multi_end = s + len;
4285 /* Count how many characters are in it. In the case of
4286 * /aa, no folds which contain ASCII code points are
4287 * allowed, so check for those, and skip if found. */
4288 if (OP(scan) != EXACTFAA && OP(scan) != EXACTFAA_NO_TRIE) {
4289 count = utf8_length(s, multi_end);
4293 while (s < multi_end) {
4296 goto next_iteration;
4306 /* The delta is how long the sequence is minus 1 (1 is how long
4307 * the character that folds to the sequence is) */
4308 total_count_delta += count - 1;
4312 /* We created a temporary folded copy of the string in EXACTFL
4313 * nodes. Therefore we need to be sure it doesn't go below zero,
4314 * as the real string could be shorter */
4315 if (OP(scan) == EXACTFL) {
4316 int total_chars = utf8_length((U8*) STRING(scan),
4317 (U8*) STRING(scan) + STR_LEN(scan));
4318 if (total_count_delta > total_chars) {
4319 total_count_delta = total_chars;
4323 *min_subtract += total_count_delta;
4326 else if (OP(scan) == EXACTFAA) {
4328 /* Non-UTF-8 pattern, EXACTFAA node. There can't be a multi-char
4329 * fold to the ASCII range (and there are no existing ones in the
4330 * upper latin1 range). But, as outlined in the comments preceding
4331 * this function, we need to flag any occurrences of the sharp s.
4332 * This character forbids trie formation (because of added
4334 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
4335 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
4336 || UNICODE_DOT_DOT_VERSION > 0)
4338 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
4339 OP(scan) = EXACTFAA_NO_TRIE;
4340 *unfolded_multi_char = TRUE;
4348 /* Non-UTF-8 pattern, not EXACTFAA node. Look for the multi-char
4349 * folds that are all Latin1. As explained in the comments
4350 * preceding this function, we look also for the sharp s in EXACTF
4351 * and EXACTFL nodes; it can be in the final position. Otherwise
4352 * we can stop looking 1 byte earlier because have to find at least
4353 * two characters for a multi-fold */
4354 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
4359 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
4360 if (! len) { /* Not a multi-char fold. */
4361 if (*s == LATIN_SMALL_LETTER_SHARP_S
4362 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
4364 *unfolded_multi_char = TRUE;
4371 && isALPHA_FOLD_EQ(*s, 's')
4372 && isALPHA_FOLD_EQ(*(s+1), 's'))
4375 /* EXACTF nodes need to know that the minimum length
4376 * changed so that a sharp s in the string can match this
4377 * ss in the pattern, but they remain EXACTF nodes, as they
4378 * won't match this unless the target string is is UTF-8,
4379 * which we don't know until runtime. EXACTFL nodes can't
4380 * transform into EXACTFU nodes */
4381 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
4382 OP(scan) = EXACTFUP;
4386 *min_subtract += len - 1;
4392 if ( STR_LEN(scan) == 1
4393 && isALPHA_A(* STRING(scan))
4394 && ( OP(scan) == EXACTFAA
4395 || ( OP(scan) == EXACTFU
4396 && ! HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(* STRING(scan)))))
4398 U8 mask = ~ ('A' ^ 'a'); /* These differ in just one bit */
4400 /* Replace a length 1 ASCII fold pair node with an ANYOFM node,
4401 * with the mask set to the complement of the bit that differs
4402 * between upper and lower case, and the lowest code point of the
4403 * pair (which the '&' forces) */
4405 ARG_SET(scan, *STRING(scan) & mask);
4411 /* Allow dumping but overwriting the collection of skipped
4412 * ops and/or strings with fake optimized ops */
4413 n = scan + NODE_SZ_STR(scan);
4421 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl", scan, depth, 0);});
4425 /* REx optimizer. Converts nodes into quicker variants "in place".
4426 Finds fixed substrings. */
4428 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
4429 to the position after last scanned or to NULL. */
4431 #define INIT_AND_WITHP \
4432 assert(!and_withp); \
4433 Newx(and_withp, 1, regnode_ssc); \
4434 SAVEFREEPV(and_withp)
4438 S_unwind_scan_frames(pTHX_ const void *p)
4440 scan_frame *f= (scan_frame *)p;
4442 scan_frame *n= f->next_frame;
4448 /* the return from this sub is the minimum length that could possibly match */
4450 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
4451 SSize_t *minlenp, SSize_t *deltap,
4456 regnode_ssc *and_withp,
4457 U32 flags, U32 depth)
4458 /* scanp: Start here (read-write). */
4459 /* deltap: Write maxlen-minlen here. */
4460 /* last: Stop before this one. */
4461 /* data: string data about the pattern */
4462 /* stopparen: treat close N as END */
4463 /* recursed: which subroutines have we recursed into */
4464 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
4467 /* There must be at least this number of characters to match */
4470 regnode *scan = *scanp, *next;
4472 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
4473 int is_inf_internal = 0; /* The studied chunk is infinite */
4474 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
4475 scan_data_t data_fake;
4476 SV *re_trie_maxbuff = NULL;
4477 regnode *first_non_open = scan;
4478 SSize_t stopmin = SSize_t_MAX;
4479 scan_frame *frame = NULL;
4480 GET_RE_DEBUG_FLAGS_DECL;
4482 PERL_ARGS_ASSERT_STUDY_CHUNK;
4483 RExC_study_started= 1;
4485 Zero(&data_fake, 1, scan_data_t);
4488 while (first_non_open && OP(first_non_open) == OPEN)
4489 first_non_open=regnext(first_non_open);
4495 RExC_study_chunk_recursed_count++;
4497 DEBUG_OPTIMISE_MORE_r(
4499 Perl_re_indentf( aTHX_ "study_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
4500 depth, (long)stopparen,
4501 (unsigned long)RExC_study_chunk_recursed_count,
4502 (unsigned long)depth, (unsigned long)recursed_depth,
4505 if (recursed_depth) {
4508 for ( j = 0 ; j < recursed_depth ; j++ ) {
4509 for ( i = 0 ; i < (U32)RExC_total_parens ; i++ ) {
4511 PAREN_TEST(RExC_study_chunk_recursed +
4512 ( j * RExC_study_chunk_recursed_bytes), i )
4515 !PAREN_TEST(RExC_study_chunk_recursed +
4516 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
4519 Perl_re_printf( aTHX_ " %d",(int)i);
4523 if ( j + 1 < recursed_depth ) {
4524 Perl_re_printf( aTHX_ ",");
4528 Perl_re_printf( aTHX_ "\n");
4531 while ( scan && OP(scan) != END && scan < last ){
4532 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
4533 node length to get a real minimum (because
4534 the folded version may be shorter) */
4535 bool unfolded_multi_char = FALSE;
4536 /* Peephole optimizer: */
4537 DEBUG_STUDYDATA("Peep", data, depth, is_inf);
4538 DEBUG_PEEP("Peep", scan, depth, flags);
4541 /* The reason we do this here is that we need to deal with things like
4542 * /(?:f)(?:o)(?:o)/ which cant be dealt with by the normal EXACT
4543 * parsing code, as each (?:..) is handled by a different invocation of
4546 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
4548 /* Follow the next-chain of the current node and optimize
4549 away all the NOTHINGs from it. */
4550 if (OP(scan) != CURLYX) {
4551 const int max = (reg_off_by_arg[OP(scan)]
4553 /* I32 may be smaller than U16 on CRAYs! */
4554 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
4555 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
4559 /* Skip NOTHING and LONGJMP. */
4560 while ( (n = regnext(n))
4561 && ( (PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
4562 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
4563 && off + noff < max)
4565 if (reg_off_by_arg[OP(scan)])
4568 NEXT_OFF(scan) = off;
4571 /* The principal pseudo-switch. Cannot be a switch, since we look into
4572 * several different things. */
4573 if ( OP(scan) == DEFINEP ) {
4575 SSize_t deltanext = 0;
4576 SSize_t fake_last_close = 0;
4577 I32 f = SCF_IN_DEFINE;
4579 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4580 scan = regnext(scan);
4581 assert( OP(scan) == IFTHEN );
4582 DEBUG_PEEP("expect IFTHEN", scan, depth, flags);
4584 data_fake.last_closep= &fake_last_close;
4586 next = regnext(scan);
4587 scan = NEXTOPER(NEXTOPER(scan));
4588 DEBUG_PEEP("scan", scan, depth, flags);
4589 DEBUG_PEEP("next", next, depth, flags);
4591 /* we suppose the run is continuous, last=next...
4592 * NOTE we dont use the return here! */
4593 /* DEFINEP study_chunk() recursion */
4594 (void)study_chunk(pRExC_state, &scan, &minlen,
4595 &deltanext, next, &data_fake, stopparen,
4596 recursed_depth, NULL, f, depth+1);
4601 OP(scan) == BRANCH ||
4602 OP(scan) == BRANCHJ ||
4605 next = regnext(scan);
4608 /* The op(next)==code check below is to see if we
4609 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
4610 * IFTHEN is special as it might not appear in pairs.
4611 * Not sure whether BRANCH-BRANCHJ is possible, regardless
4612 * we dont handle it cleanly. */
4613 if (OP(next) == code || code == IFTHEN) {
4614 /* NOTE - There is similar code to this block below for
4615 * handling TRIE nodes on a re-study. If you change stuff here
4616 * check there too. */
4617 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
4619 regnode * const startbranch=scan;
4621 if (flags & SCF_DO_SUBSTR) {
4622 /* Cannot merge strings after this. */
4623 scan_commit(pRExC_state, data, minlenp, is_inf);
4626 if (flags & SCF_DO_STCLASS)
4627 ssc_init_zero(pRExC_state, &accum);
4629 while (OP(scan) == code) {
4630 SSize_t deltanext, minnext, fake;
4632 regnode_ssc this_class;
4634 DEBUG_PEEP("Branch", scan, depth, flags);
4637 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
4639 data_fake.whilem_c = data->whilem_c;
4640 data_fake.last_closep = data->last_closep;
4643 data_fake.last_closep = &fake;
4645 data_fake.pos_delta = delta;
4646 next = regnext(scan);
4648 scan = NEXTOPER(scan); /* everything */
4649 if (code != BRANCH) /* everything but BRANCH */
4650 scan = NEXTOPER(scan);
4652 if (flags & SCF_DO_STCLASS) {
4653 ssc_init(pRExC_state, &this_class);
4654 data_fake.start_class = &this_class;
4655 f = SCF_DO_STCLASS_AND;
4657 if (flags & SCF_WHILEM_VISITED_POS)
4658 f |= SCF_WHILEM_VISITED_POS;
4660 /* we suppose the run is continuous, last=next...*/
4661 /* recurse study_chunk() for each BRANCH in an alternation */
4662 minnext = study_chunk(pRExC_state, &scan, minlenp,
4663 &deltanext, next, &data_fake, stopparen,
4664 recursed_depth, NULL, f, depth+1);
4668 if (deltanext == SSize_t_MAX) {
4669 is_inf = is_inf_internal = 1;
4671 } else if (max1 < minnext + deltanext)
4672 max1 = minnext + deltanext;
4674 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4676 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4677 if ( stopmin > minnext)
4678 stopmin = min + min1;
4679 flags &= ~SCF_DO_SUBSTR;
4681 data->flags |= SCF_SEEN_ACCEPT;
4684 if (data_fake.flags & SF_HAS_EVAL)
4685 data->flags |= SF_HAS_EVAL;
4686 data->whilem_c = data_fake.whilem_c;
4688 if (flags & SCF_DO_STCLASS)
4689 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4691 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4693 if (flags & SCF_DO_SUBSTR) {
4694 data->pos_min += min1;
4695 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4696 data->pos_delta = SSize_t_MAX;
4698 data->pos_delta += max1 - min1;
4699 if (max1 != min1 || is_inf)
4700 data->cur_is_floating = 1;
4703 if (delta == SSize_t_MAX
4704 || SSize_t_MAX - delta - (max1 - min1) < 0)
4705 delta = SSize_t_MAX;
4707 delta += max1 - min1;
4708 if (flags & SCF_DO_STCLASS_OR) {
4709 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4711 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4712 flags &= ~SCF_DO_STCLASS;
4715 else if (flags & SCF_DO_STCLASS_AND) {
4717 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4718 flags &= ~SCF_DO_STCLASS;
4721 /* Switch to OR mode: cache the old value of
4722 * data->start_class */
4724 StructCopy(data->start_class, and_withp, regnode_ssc);
4725 flags &= ~SCF_DO_STCLASS_AND;
4726 StructCopy(&accum, data->start_class, regnode_ssc);
4727 flags |= SCF_DO_STCLASS_OR;
4731 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4732 OP( startbranch ) == BRANCH )
4736 Assuming this was/is a branch we are dealing with: 'scan'
4737 now points at the item that follows the branch sequence,
4738 whatever it is. We now start at the beginning of the
4739 sequence and look for subsequences of
4745 which would be constructed from a pattern like
4748 If we can find such a subsequence we need to turn the first
4749 element into a trie and then add the subsequent branch exact
4750 strings to the trie.
4754 1. patterns where the whole set of branches can be
4757 2. patterns where only a subset can be converted.
4759 In case 1 we can replace the whole set with a single regop
4760 for the trie. In case 2 we need to keep the start and end
4763 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4764 becomes BRANCH TRIE; BRANCH X;
4766 There is an additional case, that being where there is a
4767 common prefix, which gets split out into an EXACT like node
4768 preceding the TRIE node.
4770 If x(1..n)==tail then we can do a simple trie, if not we make
4771 a "jump" trie, such that when we match the appropriate word
4772 we "jump" to the appropriate tail node. Essentially we turn
4773 a nested if into a case structure of sorts.
4778 if (!re_trie_maxbuff) {
4779 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4780 if (!SvIOK(re_trie_maxbuff))
4781 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4783 if ( SvIV(re_trie_maxbuff)>=0 ) {
4785 regnode *first = (regnode *)NULL;
4786 regnode *prev = (regnode *)NULL;
4787 regnode *tail = scan;
4791 /* var tail is used because there may be a TAIL
4792 regop in the way. Ie, the exacts will point to the
4793 thing following the TAIL, but the last branch will
4794 point at the TAIL. So we advance tail. If we
4795 have nested (?:) we may have to move through several
4799 while ( OP( tail ) == TAIL ) {
4800 /* this is the TAIL generated by (?:) */
4801 tail = regnext( tail );
4805 DEBUG_TRIE_COMPILE_r({
4806 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4807 Perl_re_indentf( aTHX_ "%s %" UVuf ":%s\n",
4809 "Looking for TRIE'able sequences. Tail node is ",
4810 (UV) REGNODE_OFFSET(tail),
4811 SvPV_nolen_const( RExC_mysv )
4817 Step through the branches
4818 cur represents each branch,
4819 noper is the first thing to be matched as part
4821 noper_next is the regnext() of that node.
4823 We normally handle a case like this
4824 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4825 support building with NOJUMPTRIE, which restricts
4826 the trie logic to structures like /FOO|BAR/.
4828 If noper is a trieable nodetype then the branch is
4829 a possible optimization target. If we are building
4830 under NOJUMPTRIE then we require that noper_next is
4831 the same as scan (our current position in the regex
4834 Once we have two or more consecutive such branches
4835 we can create a trie of the EXACT's contents and
4836 stitch it in place into the program.
4838 If the sequence represents all of the branches in
4839 the alternation we replace the entire thing with a
4842 Otherwise when it is a subsequence we need to
4843 stitch it in place and replace only the relevant
4844 branches. This means the first branch has to remain
4845 as it is used by the alternation logic, and its
4846 next pointer, and needs to be repointed at the item
4847 on the branch chain following the last branch we
4848 have optimized away.
4850 This could be either a BRANCH, in which case the
4851 subsequence is internal, or it could be the item
4852 following the branch sequence in which case the
4853 subsequence is at the end (which does not
4854 necessarily mean the first node is the start of the
4857 TRIE_TYPE(X) is a define which maps the optype to a
4861 ----------------+-----------
4866 EXACTFU_REQ8 | EXACTFU
4870 EXACTFLU8 | EXACTFLU8
4874 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4876 : ( EXACT == (X) || EXACT_REQ8 == (X) ) \
4878 : ( EXACTFU == (X) \
4879 || EXACTFU_REQ8 == (X) \
4880 || EXACTFUP == (X) ) \
4882 : ( EXACTFAA == (X) ) \
4884 : ( EXACTL == (X) ) \
4886 : ( EXACTFLU8 == (X) ) \
4890 /* dont use tail as the end marker for this traverse */
4891 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4892 regnode * const noper = NEXTOPER( cur );
4893 U8 noper_type = OP( noper );
4894 U8 noper_trietype = TRIE_TYPE( noper_type );
4895 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4896 regnode * const noper_next = regnext( noper );
4897 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4898 U8 noper_next_trietype = (noper_next && noper_next < tail) ? TRIE_TYPE( noper_next_type ) :0;
4901 DEBUG_TRIE_COMPILE_r({
4902 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4903 Perl_re_indentf( aTHX_ "- %d:%s (%d)",
4905 REG_NODE_NUM(cur), SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4907 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4908 Perl_re_printf( aTHX_ " -> %d:%s",
4909 REG_NODE_NUM(noper), SvPV_nolen_const(RExC_mysv));
4912 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4913 Perl_re_printf( aTHX_ "\t=> %d:%s\t",
4914 REG_NODE_NUM(noper_next), SvPV_nolen_const(RExC_mysv));
4916 Perl_re_printf( aTHX_ "(First==%d,Last==%d,Cur==%d,tt==%s,ntt==%s,nntt==%s)\n",
4917 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
4918 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4922 /* Is noper a trieable nodetype that can be merged
4923 * with the current trie (if there is one)? */
4927 ( noper_trietype == NOTHING )
4928 || ( trietype == NOTHING )
4929 || ( trietype == noper_trietype )
4932 && noper_next >= tail
4936 /* Handle mergable triable node Either we are
4937 * the first node in a new trieable sequence,
4938 * in which case we do some bookkeeping,
4939 * otherwise we update the end pointer. */
4942 if ( noper_trietype == NOTHING ) {
4943 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4944 regnode * const noper_next = regnext( noper );
4945 U8 noper_next_type = (noper_next && noper_next < tail) ? OP(noper_next) : 0;
4946 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4949 if ( noper_next_trietype ) {
4950 trietype = noper_next_trietype;
4951 } else if (noper_next_type) {
4952 /* a NOTHING regop is 1 regop wide.
4953 * We need at least two for a trie
4954 * so we can't merge this in */
4958 trietype = noper_trietype;
4961 if ( trietype == NOTHING )
4962 trietype = noper_trietype;
4967 } /* end handle mergable triable node */
4969 /* handle unmergable node -
4970 * noper may either be a triable node which can
4971 * not be tried together with the current trie,
4972 * or a non triable node */
4974 /* If last is set and trietype is not
4975 * NOTHING then we have found at least two
4976 * triable branch sequences in a row of a
4977 * similar trietype so we can turn them
4978 * into a trie. If/when we allow NOTHING to
4979 * start a trie sequence this condition
4980 * will be required, and it isn't expensive
4981 * so we leave it in for now. */
4982 if ( trietype && trietype != NOTHING )
4983 make_trie( pRExC_state,
4984 startbranch, first, cur, tail,
4985 count, trietype, depth+1 );
4986 prev = NULL; /* note: we clear/update
4987 first, trietype etc below,
4988 so we dont do it here */
4992 && noper_next >= tail
4995 /* noper is triable, so we can start a new
4999 trietype = noper_trietype;
5001 /* if we already saw a first but the
5002 * current node is not triable then we have
5003 * to reset the first information. */
5008 } /* end handle unmergable node */
5009 } /* loop over branches */
5010 DEBUG_TRIE_COMPILE_r({
5011 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5012 Perl_re_indentf( aTHX_ "- %s (%d) <SCAN FINISHED> ",
5013 depth+1, SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5014 Perl_re_printf( aTHX_ "(First==%d, Last==%d, Cur==%d, tt==%s)\n",
5015 REG_NODE_NUM(first), REG_NODE_NUM(prev), REG_NODE_NUM(cur),
5016 PL_reg_name[trietype]
5020 if ( prev && trietype ) {
5021 if ( trietype != NOTHING ) {
5022 /* the last branch of the sequence was part of
5023 * a trie, so we have to construct it here
5024 * outside of the loop */
5025 made= make_trie( pRExC_state, startbranch,
5026 first, scan, tail, count,
5027 trietype, depth+1 );
5028 #ifdef TRIE_STUDY_OPT
5029 if ( ((made == MADE_EXACT_TRIE &&
5030 startbranch == first)
5031 || ( first_non_open == first )) &&
5033 flags |= SCF_TRIE_RESTUDY;
5034 if ( startbranch == first
5037 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
5042 /* at this point we know whatever we have is a
5043 * NOTHING sequence/branch AND if 'startbranch'
5044 * is 'first' then we can turn the whole thing
5047 if ( startbranch == first ) {
5049 /* the entire thing is a NOTHING sequence,
5050 * something like this: (?:|) So we can
5051 * turn it into a plain NOTHING op. */
5052 DEBUG_TRIE_COMPILE_r({
5053 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
5054 Perl_re_indentf( aTHX_ "- %s (%d) <NOTHING BRANCH SEQUENCE>\n",
5056 SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur));
5059 OP(startbranch)= NOTHING;
5060 NEXT_OFF(startbranch)= tail - startbranch;
5061 for ( opt= startbranch + 1; opt < tail ; opt++ )
5065 } /* end if ( prev) */
5066 } /* 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);
5869 if (flags & SCF_DO_STCLASS_AND)
5870 ssc_and(pRExC_state, data->start_class,
5871 (regnode_charclass *) scan);
5873 ssc_or(pRExC_state, data->start_class,
5874 (regnode_charclass *) scan);
5880 SV* cp_list = get_ANYOFM_contents(scan);
5882 if (flags & SCF_DO_STCLASS_OR) {
5883 ssc_union(data->start_class, cp_list, invert);
5885 else if (flags & SCF_DO_STCLASS_AND) {
5886 ssc_intersection(data->start_class, cp_list, invert);
5889 SvREFCNT_dec_NN(cp_list);
5898 cp_list = _add_range_to_invlist(cp_list,
5900 ANYOFRbase(scan) + ANYOFRdelta(scan));
5902 if (flags & SCF_DO_STCLASS_OR) {
5903 ssc_union(data->start_class, cp_list, invert);
5905 else if (flags & SCF_DO_STCLASS_AND) {
5906 ssc_intersection(data->start_class, cp_list, invert);
5909 SvREFCNT_dec_NN(cp_list);
5918 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5919 if (flags & SCF_DO_STCLASS_AND) {
5920 bool was_there = cBOOL(
5921 ANYOF_POSIXL_TEST(data->start_class,
5923 ANYOF_POSIXL_ZERO(data->start_class);
5924 if (was_there) { /* Do an AND */
5925 ANYOF_POSIXL_SET(data->start_class, namedclass);
5927 /* No individual code points can now match */
5928 data->start_class->invlist
5929 = sv_2mortal(_new_invlist(0));
5932 int complement = namedclass + ((invert) ? -1 : 1);
5934 assert(flags & SCF_DO_STCLASS_OR);
5936 /* If the complement of this class was already there,
5937 * the result is that they match all code points,
5938 * (\d + \D == everything). Remove the classes from
5939 * future consideration. Locale is not relevant in
5941 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5942 ssc_match_all_cp(data->start_class);
5943 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5944 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5946 else { /* The usual case; just add this class to the
5948 ANYOF_POSIXL_SET(data->start_class, namedclass);
5953 case NPOSIXA: /* For these, we always know the exact set of
5958 my_invlist = invlist_clone(PL_Posix_ptrs[FLAGS(scan)], NULL);
5959 goto join_posix_and_ascii;
5967 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)], NULL);
5969 /* NPOSIXD matches all upper Latin1 code points unless the
5970 * target string being matched is UTF-8, which is
5971 * unknowable until match time. Since we are going to
5972 * invert, we want to get rid of all of them so that the
5973 * inversion will match all */
5974 if (OP(scan) == NPOSIXD) {
5975 _invlist_subtract(my_invlist, PL_UpperLatin1,
5979 join_posix_and_ascii:
5981 if (flags & SCF_DO_STCLASS_AND) {
5982 ssc_intersection(data->start_class, my_invlist, invert);
5983 ssc_clear_locale(data->start_class);
5986 assert(flags & SCF_DO_STCLASS_OR);
5987 ssc_union(data->start_class, my_invlist, invert);
5989 SvREFCNT_dec(my_invlist);
5991 if (flags & SCF_DO_STCLASS_OR)
5992 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5993 flags &= ~SCF_DO_STCLASS;
5996 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5997 data->flags |= (OP(scan) == MEOL
6000 scan_commit(pRExC_state, data, minlenp, is_inf);
6003 else if ( PL_regkind[OP(scan)] == BRANCHJ
6004 /* Lookbehind, or need to calculate parens/evals/stclass: */
6005 && (scan->flags || data || (flags & SCF_DO_STCLASS))
6006 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
6008 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6009 || OP(scan) == UNLESSM )
6011 /* Negative Lookahead/lookbehind
6012 In this case we can't do fixed string optimisation.
6015 SSize_t deltanext, minnext, fake = 0;
6020 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6022 data_fake.whilem_c = data->whilem_c;
6023 data_fake.last_closep = data->last_closep;
6026 data_fake.last_closep = &fake;
6027 data_fake.pos_delta = delta;
6028 if ( flags & SCF_DO_STCLASS && !scan->flags
6029 && OP(scan) == IFMATCH ) { /* Lookahead */
6030 ssc_init(pRExC_state, &intrnl);
6031 data_fake.start_class = &intrnl;
6032 f |= SCF_DO_STCLASS_AND;
6034 if (flags & SCF_WHILEM_VISITED_POS)
6035 f |= SCF_WHILEM_VISITED_POS;
6036 next = regnext(scan);
6037 nscan = NEXTOPER(NEXTOPER(scan));
6039 /* recurse study_chunk() for lookahead body */
6040 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
6041 last, &data_fake, stopparen,
6042 recursed_depth, NULL, f, depth+1);
6045 || deltanext > (I32) U8_MAX
6046 || minnext > (I32)U8_MAX
6047 || minnext + deltanext > (I32)U8_MAX)
6049 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6053 /* The 'next_off' field has been repurposed to count the
6054 * additional starting positions to try beyond the initial
6055 * one. (This leaves it at 0 for non-variable length
6056 * matches to avoid breakage for those not using this
6059 scan->next_off = deltanext;
6060 ckWARNexperimental(RExC_parse,
6061 WARN_EXPERIMENTAL__VLB,
6062 "Variable length lookbehind is experimental");
6064 scan->flags = (U8)minnext + deltanext;
6067 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6069 if (data_fake.flags & SF_HAS_EVAL)
6070 data->flags |= SF_HAS_EVAL;
6071 data->whilem_c = data_fake.whilem_c;
6073 if (f & SCF_DO_STCLASS_AND) {
6074 if (flags & SCF_DO_STCLASS_OR) {
6075 /* OR before, AND after: ideally we would recurse with
6076 * data_fake to get the AND applied by study of the
6077 * remainder of the pattern, and then derecurse;
6078 * *** HACK *** for now just treat as "no information".
6079 * See [perl #56690].
6081 ssc_init(pRExC_state, data->start_class);
6083 /* AND before and after: combine and continue. These
6084 * assertions are zero-length, so can match an EMPTY
6086 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6087 ANYOF_FLAGS(data->start_class)
6088 |= SSC_MATCHES_EMPTY_STRING;
6092 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
6094 /* Positive Lookahead/lookbehind
6095 In this case we can do fixed string optimisation,
6096 but we must be careful about it. Note in the case of
6097 lookbehind the positions will be offset by the minimum
6098 length of the pattern, something we won't know about
6099 until after the recurse.
6101 SSize_t deltanext, fake = 0;
6105 /* We use SAVEFREEPV so that when the full compile
6106 is finished perl will clean up the allocated
6107 minlens when it's all done. This way we don't
6108 have to worry about freeing them when we know
6109 they wont be used, which would be a pain.
6112 Newx( minnextp, 1, SSize_t );
6113 SAVEFREEPV(minnextp);
6116 StructCopy(data, &data_fake, scan_data_t);
6117 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
6120 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
6121 data_fake.last_found=newSVsv(data->last_found);
6125 data_fake.last_closep = &fake;
6126 data_fake.flags = 0;
6127 data_fake.substrs[0].flags = 0;
6128 data_fake.substrs[1].flags = 0;
6129 data_fake.pos_delta = delta;
6131 data_fake.flags |= SF_IS_INF;
6132 if ( flags & SCF_DO_STCLASS && !scan->flags
6133 && OP(scan) == IFMATCH ) { /* Lookahead */
6134 ssc_init(pRExC_state, &intrnl);
6135 data_fake.start_class = &intrnl;
6136 f |= SCF_DO_STCLASS_AND;
6138 if (flags & SCF_WHILEM_VISITED_POS)
6139 f |= SCF_WHILEM_VISITED_POS;
6140 next = regnext(scan);
6141 nscan = NEXTOPER(NEXTOPER(scan));
6143 /* positive lookahead study_chunk() recursion */
6144 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
6145 &deltanext, last, &data_fake,
6146 stopparen, recursed_depth, NULL,
6149 assert(0); /* This code has never been tested since this
6150 is normally not compiled */
6152 || deltanext > (I32) U8_MAX
6153 || *minnextp > (I32)U8_MAX
6154 || *minnextp + deltanext > (I32)U8_MAX)
6156 FAIL2("Lookbehind longer than %" UVuf " not implemented",
6161 scan->next_off = deltanext;
6163 scan->flags = (U8)*minnextp + deltanext;
6168 if (f & SCF_DO_STCLASS_AND) {
6169 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
6170 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
6173 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6175 if (data_fake.flags & SF_HAS_EVAL)
6176 data->flags |= SF_HAS_EVAL;
6177 data->whilem_c = data_fake.whilem_c;
6178 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
6180 if (RExC_rx->minlen<*minnextp)
6181 RExC_rx->minlen=*minnextp;
6182 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
6183 SvREFCNT_dec_NN(data_fake.last_found);
6185 for (i = 0; i < 2; i++) {
6186 if (data_fake.substrs[i].minlenp != minlenp) {
6187 data->substrs[i].min_offset =
6188 data_fake.substrs[i].min_offset;
6189 data->substrs[i].max_offset =
6190 data_fake.substrs[i].max_offset;
6191 data->substrs[i].minlenp =
6192 data_fake.substrs[i].minlenp;
6193 data->substrs[i].lookbehind += scan->flags;
6201 else if (OP(scan) == OPEN) {
6202 if (stopparen != (I32)ARG(scan))
6205 else if (OP(scan) == CLOSE) {
6206 if (stopparen == (I32)ARG(scan)) {
6209 if ((I32)ARG(scan) == is_par) {
6210 next = regnext(scan);
6212 if ( next && (OP(next) != WHILEM) && next < last)
6213 is_par = 0; /* Disable optimization */
6216 *(data->last_closep) = ARG(scan);
6218 else if (OP(scan) == EVAL) {
6220 data->flags |= SF_HAS_EVAL;
6222 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
6223 if (flags & SCF_DO_SUBSTR) {
6224 scan_commit(pRExC_state, data, minlenp, is_inf);
6225 flags &= ~SCF_DO_SUBSTR;
6227 if (data && OP(scan)==ACCEPT) {
6228 data->flags |= SCF_SEEN_ACCEPT;
6233 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
6235 if (flags & SCF_DO_SUBSTR) {
6236 scan_commit(pRExC_state, data, minlenp, is_inf);
6237 data->cur_is_floating = 1; /* float */
6239 is_inf = is_inf_internal = 1;
6240 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
6241 ssc_anything(data->start_class);
6242 flags &= ~SCF_DO_STCLASS;
6244 else if (OP(scan) == GPOS) {
6245 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
6246 !(delta || is_inf || (data && data->pos_delta)))
6248 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
6249 RExC_rx->intflags |= PREGf_ANCH_GPOS;
6250 if (RExC_rx->gofs < (STRLEN)min)
6251 RExC_rx->gofs = min;
6253 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
6257 #ifdef TRIE_STUDY_OPT
6258 #ifdef FULL_TRIE_STUDY
6259 else if (PL_regkind[OP(scan)] == TRIE) {
6260 /* NOTE - There is similar code to this block above for handling
6261 BRANCH nodes on the initial study. If you change stuff here
6263 regnode *trie_node= scan;
6264 regnode *tail= regnext(scan);
6265 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6266 SSize_t max1 = 0, min1 = SSize_t_MAX;
6269 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
6270 /* Cannot merge strings after this. */
6271 scan_commit(pRExC_state, data, minlenp, is_inf);
6273 if (flags & SCF_DO_STCLASS)
6274 ssc_init_zero(pRExC_state, &accum);
6280 const regnode *nextbranch= NULL;
6283 for ( word=1 ; word <= trie->wordcount ; word++)
6285 SSize_t deltanext=0, minnext=0, f = 0, fake;
6286 regnode_ssc this_class;
6288 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
6290 data_fake.whilem_c = data->whilem_c;
6291 data_fake.last_closep = data->last_closep;
6294 data_fake.last_closep = &fake;
6295 data_fake.pos_delta = delta;
6296 if (flags & SCF_DO_STCLASS) {
6297 ssc_init(pRExC_state, &this_class);
6298 data_fake.start_class = &this_class;
6299 f = SCF_DO_STCLASS_AND;
6301 if (flags & SCF_WHILEM_VISITED_POS)
6302 f |= SCF_WHILEM_VISITED_POS;
6304 if (trie->jump[word]) {
6306 nextbranch = trie_node + trie->jump[0];
6307 scan= trie_node + trie->jump[word];
6308 /* We go from the jump point to the branch that follows
6309 it. Note this means we need the vestigal unused
6310 branches even though they arent otherwise used. */
6311 /* optimise study_chunk() for TRIE */
6312 minnext = study_chunk(pRExC_state, &scan, minlenp,
6313 &deltanext, (regnode *)nextbranch, &data_fake,
6314 stopparen, recursed_depth, NULL, f, depth+1);
6316 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
6317 nextbranch= regnext((regnode*)nextbranch);
6319 if (min1 > (SSize_t)(minnext + trie->minlen))
6320 min1 = minnext + trie->minlen;
6321 if (deltanext == SSize_t_MAX) {
6322 is_inf = is_inf_internal = 1;
6324 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
6325 max1 = minnext + deltanext + trie->maxlen;
6327 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
6329 if (data_fake.flags & SCF_SEEN_ACCEPT) {
6330 if ( stopmin > min + min1)
6331 stopmin = min + min1;
6332 flags &= ~SCF_DO_SUBSTR;
6334 data->flags |= SCF_SEEN_ACCEPT;
6337 if (data_fake.flags & SF_HAS_EVAL)
6338 data->flags |= SF_HAS_EVAL;
6339 data->whilem_c = data_fake.whilem_c;
6341 if (flags & SCF_DO_STCLASS)
6342 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
6345 if (flags & SCF_DO_SUBSTR) {
6346 data->pos_min += min1;
6347 data->pos_delta += max1 - min1;
6348 if (max1 != min1 || is_inf)
6349 data->cur_is_floating = 1; /* float */
6352 if (delta != SSize_t_MAX) {
6353 if (SSize_t_MAX - (max1 - min1) >= delta)
6354 delta += max1 - min1;
6356 delta = SSize_t_MAX;
6358 if (flags & SCF_DO_STCLASS_OR) {
6359 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6361 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6362 flags &= ~SCF_DO_STCLASS;
6365 else if (flags & SCF_DO_STCLASS_AND) {
6367 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
6368 flags &= ~SCF_DO_STCLASS;
6371 /* Switch to OR mode: cache the old value of
6372 * data->start_class */
6374 StructCopy(data->start_class, and_withp, regnode_ssc);
6375 flags &= ~SCF_DO_STCLASS_AND;
6376 StructCopy(&accum, data->start_class, regnode_ssc);
6377 flags |= SCF_DO_STCLASS_OR;
6384 else if (PL_regkind[OP(scan)] == TRIE) {
6385 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
6388 min += trie->minlen;
6389 delta += (trie->maxlen - trie->minlen);
6390 flags &= ~SCF_DO_STCLASS; /* xxx */
6391 if (flags & SCF_DO_SUBSTR) {
6392 /* Cannot expect anything... */
6393 scan_commit(pRExC_state, data, minlenp, is_inf);
6394 data->pos_min += trie->minlen;
6395 data->pos_delta += (trie->maxlen - trie->minlen);
6396 if (trie->maxlen != trie->minlen)
6397 data->cur_is_floating = 1; /* float */
6399 if (trie->jump) /* no more substrings -- for now /grr*/
6400 flags &= ~SCF_DO_SUBSTR;
6402 #endif /* old or new */
6403 #endif /* TRIE_STUDY_OPT */
6405 /* Else: zero-length, ignore. */
6406 scan = regnext(scan);
6411 /* we need to unwind recursion. */
6414 DEBUG_STUDYDATA("frame-end", data, depth, is_inf);
6415 DEBUG_PEEP("fend", scan, depth, flags);
6417 /* restore previous context */
6418 last = frame->last_regnode;
6419 scan = frame->next_regnode;
6420 stopparen = frame->stopparen;
6421 recursed_depth = frame->prev_recursed_depth;
6423 RExC_frame_last = frame->prev_frame;
6424 frame = frame->this_prev_frame;
6425 goto fake_study_recurse;
6429 DEBUG_STUDYDATA("pre-fin", data, depth, is_inf);
6432 *deltap = is_inf_internal ? SSize_t_MAX : delta;
6434 if (flags & SCF_DO_SUBSTR && is_inf)
6435 data->pos_delta = SSize_t_MAX - data->pos_min;
6436 if (is_par > (I32)U8_MAX)
6438 if (is_par && pars==1 && data) {
6439 data->flags |= SF_IN_PAR;
6440 data->flags &= ~SF_HAS_PAR;
6442 else if (pars && data) {
6443 data->flags |= SF_HAS_PAR;
6444 data->flags &= ~SF_IN_PAR;
6446 if (flags & SCF_DO_STCLASS_OR)
6447 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
6448 if (flags & SCF_TRIE_RESTUDY)
6449 data->flags |= SCF_TRIE_RESTUDY;
6451 DEBUG_STUDYDATA("post-fin", data, depth, is_inf);
6454 SSize_t final_minlen= min < stopmin ? min : stopmin;
6456 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
6457 if (final_minlen > SSize_t_MAX - delta)
6458 RExC_maxlen = SSize_t_MAX;
6459 else if (RExC_maxlen < final_minlen + delta)
6460 RExC_maxlen = final_minlen + delta;
6462 return final_minlen;
6464 NOT_REACHED; /* NOTREACHED */
6468 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
6470 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
6472 PERL_ARGS_ASSERT_ADD_DATA;
6474 Renewc(RExC_rxi->data,
6475 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
6476 char, struct reg_data);
6478 Renew(RExC_rxi->data->what, count + n, U8);
6480 Newx(RExC_rxi->data->what, n, U8);
6481 RExC_rxi->data->count = count + n;
6482 Copy(s, RExC_rxi->data->what + count, n, U8);
6486 /*XXX: todo make this not included in a non debugging perl, but appears to be
6487 * used anyway there, in 'use re' */
6488 #ifndef PERL_IN_XSUB_RE
6490 Perl_reginitcolors(pTHX)
6492 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
6494 char *t = savepv(s);
6498 t = strchr(t, '\t');
6504 PL_colors[i] = t = (char *)"";
6509 PL_colors[i++] = (char *)"";
6516 #ifdef TRIE_STUDY_OPT
6517 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
6520 (data.flags & SCF_TRIE_RESTUDY) \
6528 #define CHECK_RESTUDY_GOTO_butfirst
6532 * pregcomp - compile a regular expression into internal code
6534 * Decides which engine's compiler to call based on the hint currently in
6538 #ifndef PERL_IN_XSUB_RE
6540 /* return the currently in-scope regex engine (or the default if none) */
6542 regexp_engine const *
6543 Perl_current_re_engine(pTHX)
6545 if (IN_PERL_COMPILETIME) {
6546 HV * const table = GvHV(PL_hintgv);
6549 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
6550 return &PL_core_reg_engine;
6551 ptr = hv_fetchs(table, "regcomp", FALSE);
6552 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
6553 return &PL_core_reg_engine;
6554 return INT2PTR(regexp_engine*, SvIV(*ptr));
6558 if (!PL_curcop->cop_hints_hash)
6559 return &PL_core_reg_engine;
6560 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
6561 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
6562 return &PL_core_reg_engine;
6563 return INT2PTR(regexp_engine*, SvIV(ptr));
6569 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
6571 regexp_engine const *eng = current_re_engine();
6572 GET_RE_DEBUG_FLAGS_DECL;
6574 PERL_ARGS_ASSERT_PREGCOMP;
6576 /* Dispatch a request to compile a regexp to correct regexp engine. */
6578 Perl_re_printf( aTHX_ "Using engine %" UVxf "\n",
6581 return CALLREGCOMP_ENG(eng, pattern, flags);
6585 /* public(ish) entry point for the perl core's own regex compiling code.
6586 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
6587 * pattern rather than a list of OPs, and uses the internal engine rather
6588 * than the current one */
6591 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
6593 SV *pat = pattern; /* defeat constness! */
6594 PERL_ARGS_ASSERT_RE_COMPILE;
6595 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
6596 #ifdef PERL_IN_XSUB_RE
6599 &PL_core_reg_engine,
6601 NULL, NULL, rx_flags, 0);
6606 S_free_codeblocks(pTHX_ struct reg_code_blocks *cbs)
6610 if (--cbs->refcnt > 0)
6612 for (n = 0; n < cbs->count; n++) {
6613 REGEXP *rx = cbs->cb[n].src_regex;
6615 cbs->cb[n].src_regex = NULL;
6616 SvREFCNT_dec_NN(rx);
6624 static struct reg_code_blocks *
6625 S_alloc_code_blocks(pTHX_ int ncode)
6627 struct reg_code_blocks *cbs;
6628 Newx(cbs, 1, struct reg_code_blocks);
6631 SAVEDESTRUCTOR_X(S_free_codeblocks, cbs);
6633 Newx(cbs->cb, ncode, struct reg_code_block);
6640 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
6641 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
6642 * point to the realloced string and length.
6644 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
6648 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
6649 char **pat_p, STRLEN *plen_p, int num_code_blocks)
6651 U8 *const src = (U8*)*pat_p;
6656 GET_RE_DEBUG_FLAGS_DECL;
6658 DEBUG_PARSE_r(Perl_re_printf( aTHX_
6659 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
6661 /* 1 for each byte + 1 for each byte that expands to two, + trailing NUL */
6662 Newx(dst, *plen_p + variant_under_utf8_count(src, src + *plen_p) + 1, U8);
6665 while (s < *plen_p) {
6666 append_utf8_from_native_byte(src[s], &d);
6668 if (n < num_code_blocks) {
6669 assert(pRExC_state->code_blocks);
6670 if (!do_end && pRExC_state->code_blocks->cb[n].start == s) {
6671 pRExC_state->code_blocks->cb[n].start = d - dst - 1;
6672 assert(*(d - 1) == '(');
6675 else if (do_end && pRExC_state->code_blocks->cb[n].end == s) {
6676 pRExC_state->code_blocks->cb[n].end = d - dst - 1;
6677 assert(*(d - 1) == ')');
6686 *pat_p = (char*) dst;
6688 RExC_orig_utf8 = RExC_utf8 = 1;
6693 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
6694 * while recording any code block indices, and handling overloading,
6695 * nested qr// objects etc. If pat is null, it will allocate a new
6696 * string, or just return the first arg, if there's only one.
6698 * Returns the malloced/updated pat.
6699 * patternp and pat_count is the array of SVs to be concatted;
6700 * oplist is the optional list of ops that generated the SVs;
6701 * recompile_p is a pointer to a boolean that will be set if
6702 * the regex will need to be recompiled.
6703 * delim, if non-null is an SV that will be inserted between each element
6707 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
6708 SV *pat, SV ** const patternp, int pat_count,
6709 OP *oplist, bool *recompile_p, SV *delim)
6713 bool use_delim = FALSE;
6714 bool alloced = FALSE;
6716 /* if we know we have at least two args, create an empty string,
6717 * then concatenate args to that. For no args, return an empty string */
6718 if (!pat && pat_count != 1) {
6724 for (svp = patternp; svp < patternp + pat_count; svp++) {
6727 STRLEN orig_patlen = 0;
6729 SV *msv = use_delim ? delim : *svp;
6730 if (!msv) msv = &PL_sv_undef;
6732 /* if we've got a delimiter, we go round the loop twice for each
6733 * svp slot (except the last), using the delimiter the second
6742 if (SvTYPE(msv) == SVt_PVAV) {
6743 /* we've encountered an interpolated array within
6744 * the pattern, e.g. /...@a..../. Expand the list of elements,
6745 * then recursively append elements.
6746 * The code in this block is based on S_pushav() */
6748 AV *const av = (AV*)msv;
6749 const SSize_t maxarg = AvFILL(av) + 1;
6753 assert(oplist->op_type == OP_PADAV
6754 || oplist->op_type == OP_RV2AV);
6755 oplist = OpSIBLING(oplist);
6758 if (SvRMAGICAL(av)) {
6761 Newx(array, maxarg, SV*);
6763 for (i=0; i < maxarg; i++) {
6764 SV ** const svp = av_fetch(av, i, FALSE);
6765 array[i] = svp ? *svp : &PL_sv_undef;
6769 array = AvARRAY(av);
6771 pat = S_concat_pat(aTHX_ pRExC_state, pat,
6772 array, maxarg, NULL, recompile_p,
6774 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
6780 /* we make the assumption here that each op in the list of
6781 * op_siblings maps to one SV pushed onto the stack,
6782 * except for code blocks, with have both an OP_NULL and
6784 * This allows us to match up the list of SVs against the
6785 * list of OPs to find the next code block.
6787 * Note that PUSHMARK PADSV PADSV ..
6789 * PADRANGE PADSV PADSV ..
6790 * so the alignment still works. */
6793 if (oplist->op_type == OP_NULL
6794 && (oplist->op_flags & OPf_SPECIAL))
6796 assert(n < pRExC_state->code_blocks->count);
6797 pRExC_state->code_blocks->cb[n].start = pat ? SvCUR(pat) : 0;
6798 pRExC_state->code_blocks->cb[n].block = oplist;
6799 pRExC_state->code_blocks->cb[n].src_regex = NULL;
6802 oplist = OpSIBLING(oplist); /* skip CONST */
6805 oplist = OpSIBLING(oplist);;
6808 /* apply magic and QR overloading to arg */
6811 if (SvROK(msv) && SvAMAGIC(msv)) {
6812 SV *sv = AMG_CALLunary(msv, regexp_amg);
6816 if (SvTYPE(sv) != SVt_REGEXP)
6817 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6822 /* try concatenation overload ... */
6823 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6824 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6827 /* overloading involved: all bets are off over literal
6828 * code. Pretend we haven't seen it */
6830 pRExC_state->code_blocks->count -= n;
6834 /* ... or failing that, try "" overload */
6835 while (SvAMAGIC(msv)
6836 && (sv = AMG_CALLunary(msv, string_amg))
6840 && SvRV(msv) == SvRV(sv))
6845 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6849 /* this is a partially unrolled
6850 * sv_catsv_nomg(pat, msv);
6851 * that allows us to adjust code block indices if
6854 char *dst = SvPV_force_nomg(pat, dlen);
6856 if (SvUTF8(msv) && !SvUTF8(pat)) {
6857 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6858 sv_setpvn(pat, dst, dlen);
6861 sv_catsv_nomg(pat, msv);
6865 /* We have only one SV to process, but we need to verify
6866 * it is properly null terminated or we will fail asserts
6867 * later. In theory we probably shouldn't get such SV's,
6868 * but if we do we should handle it gracefully. */
6869 if ( SvTYPE(msv) != SVt_PV || (SvLEN(msv) > SvCUR(msv) && *(SvEND(msv)) == 0) || SvIsCOW_shared_hash(msv) ) {
6870 /* not a string, or a string with a trailing null */
6873 /* a string with no trailing null, we need to copy it
6874 * so it has a trailing null */
6875 pat = sv_2mortal(newSVsv(msv));
6880 pRExC_state->code_blocks->cb[n-1].end = SvCUR(pat)-1;
6883 /* extract any code blocks within any embedded qr//'s */
6884 if (rx && SvTYPE(rx) == SVt_REGEXP
6885 && RX_ENGINE((REGEXP*)rx)->op_comp)
6888 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6889 if (ri->code_blocks && ri->code_blocks->count) {
6891 /* the presence of an embedded qr// with code means
6892 * we should always recompile: the text of the
6893 * qr// may not have changed, but it may be a
6894 * different closure than last time */
6896 if (pRExC_state->code_blocks) {
6897 int new_count = pRExC_state->code_blocks->count
6898 + ri->code_blocks->count;
6899 Renew(pRExC_state->code_blocks->cb,
6900 new_count, struct reg_code_block);
6901 pRExC_state->code_blocks->count = new_count;
6904 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_
6905 ri->code_blocks->count);
6907 for (i=0; i < ri->code_blocks->count; i++) {
6908 struct reg_code_block *src, *dst;
6909 STRLEN offset = orig_patlen
6910 + ReANY((REGEXP *)rx)->pre_prefix;
6911 assert(n < pRExC_state->code_blocks->count);
6912 src = &ri->code_blocks->cb[i];
6913 dst = &pRExC_state->code_blocks->cb[n];
6914 dst->start = src->start + offset;
6915 dst->end = src->end + offset;
6916 dst->block = src->block;
6917 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6926 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6935 /* see if there are any run-time code blocks in the pattern.
6936 * False positives are allowed */
6939 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6940 char *pat, STRLEN plen)
6945 PERL_UNUSED_CONTEXT;
6947 for (s = 0; s < plen; s++) {
6948 if ( pRExC_state->code_blocks
6949 && n < pRExC_state->code_blocks->count
6950 && s == pRExC_state->code_blocks->cb[n].start)
6952 s = pRExC_state->code_blocks->cb[n].end;
6956 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6958 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6960 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6967 /* Handle run-time code blocks. We will already have compiled any direct
6968 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6969 * copy of it, but with any literal code blocks blanked out and
6970 * appropriate chars escaped; then feed it into
6972 * eval "qr'modified_pattern'"
6976 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6980 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6982 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6983 * and merge them with any code blocks of the original regexp.
6985 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6986 * instead, just save the qr and return FALSE; this tells our caller that
6987 * the original pattern needs upgrading to utf8.
6991 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6992 char *pat, STRLEN plen)
6996 GET_RE_DEBUG_FLAGS_DECL;
6998 if (pRExC_state->runtime_code_qr) {
6999 /* this is the second time we've been called; this should
7000 * only happen if the main pattern got upgraded to utf8
7001 * during compilation; re-use the qr we compiled first time
7002 * round (which should be utf8 too)
7004 qr = pRExC_state->runtime_code_qr;
7005 pRExC_state->runtime_code_qr = NULL;
7006 assert(RExC_utf8 && SvUTF8(qr));
7012 int newlen = plen + 7; /* allow for "qr''xx\0" extra chars */
7016 /* determine how many extra chars we need for ' and \ escaping */
7017 for (s = 0; s < plen; s++) {
7018 if (pat[s] == '\'' || pat[s] == '\\')
7022 Newx(newpat, newlen, char);
7024 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
7026 for (s = 0; s < plen; s++) {
7027 if ( pRExC_state->code_blocks
7028 && n < pRExC_state->code_blocks->count
7029 && s == pRExC_state->code_blocks->cb[n].start)
7031 /* blank out literal code block so that they aren't
7032 * recompiled: eg change from/to:
7042 assert(pat[s] == '(');
7043 assert(pat[s+1] == '?');
7047 while (s < pRExC_state->code_blocks->cb[n].end) {
7055 if (pat[s] == '\'' || pat[s] == '\\')
7060 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED) {
7062 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED_MORE) {
7068 Perl_re_printf( aTHX_
7069 "%sre-parsing pattern for runtime code:%s %s\n",
7070 PL_colors[4], PL_colors[5], newpat);
7073 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
7079 PUSHSTACKi(PERLSI_REQUIRE);
7080 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
7081 * parsing qr''; normally only q'' does this. It also alters
7083 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
7084 SvREFCNT_dec_NN(sv);
7089 SV * const errsv = ERRSV;
7090 if (SvTRUE_NN(errsv))
7091 /* use croak_sv ? */
7092 Perl_croak_nocontext("%" SVf, SVfARG(errsv));
7094 assert(SvROK(qr_ref));
7096 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
7097 /* the leaving below frees the tmp qr_ref.
7098 * Give qr a life of its own */
7106 if (!RExC_utf8 && SvUTF8(qr)) {
7107 /* first time through; the pattern got upgraded; save the
7108 * qr for the next time through */
7109 assert(!pRExC_state->runtime_code_qr);
7110 pRExC_state->runtime_code_qr = qr;
7115 /* extract any code blocks within the returned qr// */
7118 /* merge the main (r1) and run-time (r2) code blocks into one */
7120 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
7121 struct reg_code_block *new_block, *dst;
7122 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
7126 if (!r2->code_blocks || !r2->code_blocks->count) /* we guessed wrong */
7128 SvREFCNT_dec_NN(qr);
7132 if (!r1->code_blocks)
7133 r1->code_blocks = S_alloc_code_blocks(aTHX_ 0);
7135 r1c = r1->code_blocks->count;
7136 r2c = r2->code_blocks->count;
7138 Newx(new_block, r1c + r2c, struct reg_code_block);
7142 while (i1 < r1c || i2 < r2c) {
7143 struct reg_code_block *src;
7147 src = &r2->code_blocks->cb[i2++];
7151 src = &r1->code_blocks->cb[i1++];
7152 else if ( r1->code_blocks->cb[i1].start
7153 < r2->code_blocks->cb[i2].start)
7155 src = &r1->code_blocks->cb[i1++];
7156 assert(src->end < r2->code_blocks->cb[i2].start);
7159 assert( r1->code_blocks->cb[i1].start
7160 > r2->code_blocks->cb[i2].start);
7161 src = &r2->code_blocks->cb[i2++];
7163 assert(src->end < r1->code_blocks->cb[i1].start);
7166 assert(pat[src->start] == '(');
7167 assert(pat[src->end] == ')');
7168 dst->start = src->start;
7169 dst->end = src->end;
7170 dst->block = src->block;
7171 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
7175 r1->code_blocks->count += r2c;
7176 Safefree(r1->code_blocks->cb);
7177 r1->code_blocks->cb = new_block;
7180 SvREFCNT_dec_NN(qr);
7186 S_setup_longest(pTHX_ RExC_state_t *pRExC_state,
7187 struct reg_substr_datum *rsd,
7188 struct scan_data_substrs *sub,
7189 STRLEN longest_length)
7191 /* This is the common code for setting up the floating and fixed length
7192 * string data extracted from Perl_re_op_compile() below. Returns a boolean
7193 * as to whether succeeded or not */
7197 bool eol = cBOOL(sub->flags & SF_BEFORE_EOL);
7198 bool meol = cBOOL(sub->flags & SF_BEFORE_MEOL);
7200 if (! (longest_length
7201 || (eol /* Can't have SEOL and MULTI */
7202 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
7204 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
7205 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
7210 /* copy the information about the longest from the reg_scan_data
7211 over to the program. */
7212 if (SvUTF8(sub->str)) {
7214 rsd->utf8_substr = sub->str;
7216 rsd->substr = sub->str;
7217 rsd->utf8_substr = NULL;
7219 /* end_shift is how many chars that must be matched that
7220 follow this item. We calculate it ahead of time as once the
7221 lookbehind offset is added in we lose the ability to correctly
7223 ml = sub->minlenp ? *(sub->minlenp) : (SSize_t)longest_length;
7224 rsd->end_shift = ml - sub->min_offset
7226 /* XXX SvTAIL is always false here - did you mean FBMcf_TAIL
7228 + (SvTAIL(sub->str) != 0)
7232 t = (eol/* Can't have SEOL and MULTI */
7233 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
7234 fbm_compile(sub->str, t ? FBMcf_TAIL : 0);
7240 S_set_regex_pv(pTHX_ RExC_state_t *pRExC_state, REGEXP *Rx)
7242 /* Calculates and sets in the compiled pattern 'Rx' the string to compile,
7243 * properly wrapped with the right modifiers */
7245 bool has_p = ((RExC_rx->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
7246 bool has_charset = RExC_utf8 || (get_regex_charset(RExC_rx->extflags)
7247 != REGEX_DEPENDS_CHARSET);
7249 /* The caret is output if there are any defaults: if not all the STD
7250 * flags are set, or if no character set specifier is needed */
7252 (((RExC_rx->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
7254 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
7255 == REG_RUN_ON_COMMENT_SEEN);
7256 U8 reganch = (U8)((RExC_rx->extflags & RXf_PMf_STD_PMMOD)
7257 >> RXf_PMf_STD_PMMOD_SHIFT);
7258 const char *fptr = STD_PAT_MODS; /*"msixxn"*/
7260 STRLEN pat_len = RExC_precomp_end - RExC_precomp;
7262 /* We output all the necessary flags; we never output a minus, as all
7263 * those are defaults, so are
7264 * covered by the caret */
7265 const STRLEN wraplen = pat_len + has_p + has_runon
7266 + has_default /* If needs a caret */
7267 + PL_bitcount[reganch] /* 1 char for each set standard flag */
7269 /* If needs a character set specifier */
7270 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
7271 + (sizeof("(?:)") - 1);
7273 PERL_ARGS_ASSERT_SET_REGEX_PV;
7275 /* make sure PL_bitcount bounds not exceeded */
7276 assert(sizeof(STD_PAT_MODS) <= 8);
7278 p = sv_grow(MUTABLE_SV(Rx), wraplen + 1); /* +1 for the ending NUL */
7281 SvFLAGS(Rx) |= SVf_UTF8;
7284 /* If a default, cover it using the caret */
7286 *p++= DEFAULT_PAT_MOD;
7292 name = get_regex_charset_name(RExC_rx->extflags, &len);
7293 if (strEQ(name, DEPENDS_PAT_MODS)) { /* /d under UTF-8 => /u */
7295 name = UNICODE_PAT_MODS;
7296 len = sizeof(UNICODE_PAT_MODS) - 1;
7298 Copy(name, p, len, char);
7302 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
7305 while((ch = *fptr++)) {
7313 Copy(RExC_precomp, p, pat_len, char);
7314 assert ((RX_WRAPPED(Rx) - p) < 16);
7315 RExC_rx->pre_prefix = p - RX_WRAPPED(Rx);
7318 /* Adding a trailing \n causes this to compile properly:
7319 my $R = qr / A B C # D E/x; /($R)/
7320 Otherwise the parens are considered part of the comment */
7325 SvCUR_set(Rx, p - RX_WRAPPED(Rx));
7329 * Perl_re_op_compile - the perl internal RE engine's function to compile a
7330 * regular expression into internal code.
7331 * The pattern may be passed either as:
7332 * a list of SVs (patternp plus pat_count)
7333 * a list of OPs (expr)
7334 * If both are passed, the SV list is used, but the OP list indicates
7335 * which SVs are actually pre-compiled code blocks
7337 * The SVs in the list have magic and qr overloading applied to them (and
7338 * the list may be modified in-place with replacement SVs in the latter
7341 * If the pattern hasn't changed from old_re, then old_re will be
7344 * eng is the current engine. If that engine has an op_comp method, then
7345 * handle directly (i.e. we assume that op_comp was us); otherwise, just
7346 * do the initial concatenation of arguments and pass on to the external
7349 * If is_bare_re is not null, set it to a boolean indicating whether the
7350 * arg list reduced (after overloading) to a single bare regex which has
7351 * been returned (i.e. /$qr/).
7353 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
7355 * pm_flags contains the PMf_* flags, typically based on those from the
7356 * pm_flags field of the related PMOP. Currently we're only interested in
7357 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
7359 * For many years this code had an initial sizing pass that calculated
7360 * (sometimes incorrectly, leading to security holes) the size needed for the
7361 * compiled pattern. That was changed by commit
7362 * 7c932d07cab18751bfc7515b4320436273a459e2 in 5.29, which reallocs the size, a
7363 * node at a time, as parsing goes along. Patches welcome to fix any obsolete
7364 * references to this sizing pass.
7366 * Now, an initial crude guess as to the size needed is made, based on the
7367 * length of the pattern. Patches welcome to improve that guess. That amount
7368 * of space is malloc'd and then immediately freed, and then clawed back node
7369 * by node. This design is to minimze, to the extent possible, memory churn
7370 * when doing the the reallocs.
7372 * A separate parentheses counting pass may be needed in some cases.
7373 * (Previously the sizing pass did this.) Patches welcome to reduce the number
7376 * The existence of a sizing pass necessitated design decisions that are no
7377 * longer needed. There are potential areas of simplification.
7379 * Beware that the optimization-preparation code in here knows about some
7380 * of the structure of the compiled regexp. [I'll say.]
7384 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
7385 OP *expr, const regexp_engine* eng, REGEXP *old_re,
7386 bool *is_bare_re, const U32 orig_rx_flags, const U32 pm_flags)
7389 REGEXP *Rx; /* Capital 'R' means points to a REGEXP */
7397 SV** new_patternp = patternp;
7399 /* these are all flags - maybe they should be turned
7400 * into a single int with different bit masks */
7401 I32 sawlookahead = 0;
7406 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
7408 bool runtime_code = 0;
7410 RExC_state_t RExC_state;
7411 RExC_state_t * const pRExC_state = &RExC_state;
7412 #ifdef TRIE_STUDY_OPT
7414 RExC_state_t copyRExC_state;
7416 GET_RE_DEBUG_FLAGS_DECL;
7418 PERL_ARGS_ASSERT_RE_OP_COMPILE;
7420 DEBUG_r(if (!PL_colorset) reginitcolors());
7423 pRExC_state->warn_text = NULL;
7424 pRExC_state->unlexed_names = NULL;
7425 pRExC_state->code_blocks = NULL;
7428 *is_bare_re = FALSE;
7430 if (expr && (expr->op_type == OP_LIST ||
7431 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
7432 /* allocate code_blocks if needed */
7436 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
7437 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
7438 ncode++; /* count of DO blocks */
7441 pRExC_state->code_blocks = S_alloc_code_blocks(aTHX_ ncode);
7445 /* compile-time pattern with just OP_CONSTs and DO blocks */
7450 /* find how many CONSTs there are */
7453 if (expr->op_type == OP_CONST)
7456 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7457 if (o->op_type == OP_CONST)
7461 /* fake up an SV array */
7463 assert(!new_patternp);
7464 Newx(new_patternp, n, SV*);
7465 SAVEFREEPV(new_patternp);
7469 if (expr->op_type == OP_CONST)
7470 new_patternp[n] = cSVOPx_sv(expr);
7472 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
7473 if (o->op_type == OP_CONST)
7474 new_patternp[n++] = cSVOPo_sv;
7479 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7480 "Assembling pattern from %d elements%s\n", pat_count,
7481 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7483 /* set expr to the first arg op */
7485 if (pRExC_state->code_blocks && pRExC_state->code_blocks->count
7486 && expr->op_type != OP_CONST)
7488 expr = cLISTOPx(expr)->op_first;
7489 assert( expr->op_type == OP_PUSHMARK
7490 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
7491 || expr->op_type == OP_PADRANGE);
7492 expr = OpSIBLING(expr);
7495 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
7496 expr, &recompile, NULL);
7498 /* handle bare (possibly after overloading) regex: foo =~ $re */
7503 if (SvTYPE(re) == SVt_REGEXP) {
7507 DEBUG_PARSE_r(Perl_re_printf( aTHX_
7508 "Precompiled pattern%s\n",
7509 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
7515 exp = SvPV_nomg(pat, plen);
7517 if (!eng->op_comp) {
7518 if ((SvUTF8(pat) && IN_BYTES)
7519 || SvGMAGICAL(pat) || SvAMAGIC(pat))
7521 /* make a temporary copy; either to convert to bytes,
7522 * or to avoid repeating get-magic / overloaded stringify */
7523 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
7524 (IN_BYTES ? 0 : SvUTF8(pat)));
7526 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
7529 /* ignore the utf8ness if the pattern is 0 length */
7530 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
7531 RExC_uni_semantics = 0;
7532 RExC_contains_locale = 0;
7533 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
7534 RExC_in_script_run = 0;
7535 RExC_study_started = 0;
7536 pRExC_state->runtime_code_qr = NULL;
7537 RExC_frame_head= NULL;
7538 RExC_frame_last= NULL;
7539 RExC_frame_count= 0;
7540 RExC_latest_warn_offset = 0;
7541 RExC_use_BRANCHJ = 0;
7542 RExC_total_parens = 0;
7543 RExC_open_parens = NULL;
7544 RExC_close_parens = NULL;
7545 RExC_paren_names = NULL;
7547 RExC_seen_d_op = FALSE;
7549 RExC_paren_name_list = NULL;
7553 RExC_mysv1= sv_newmortal();
7554 RExC_mysv2= sv_newmortal();
7558 SV *dsv= sv_newmortal();
7559 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7560 Perl_re_printf( aTHX_ "%sCompiling REx%s %s\n",
7561 PL_colors[4], PL_colors[5], s);
7564 /* we jump here if we have to recompile, e.g., from upgrading the pattern
7567 if ((pm_flags & PMf_USE_RE_EVAL)
7568 /* this second condition covers the non-regex literal case,
7569 * i.e. $foo =~ '(?{})'. */
7570 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
7572 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
7575 /* return old regex if pattern hasn't changed */
7576 /* XXX: note in the below we have to check the flags as well as the
7579 * Things get a touch tricky as we have to compare the utf8 flag
7580 * independently from the compile flags. */
7584 && !!RX_UTF8(old_re) == !!RExC_utf8
7585 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
7586 && RX_PRECOMP(old_re)
7587 && RX_PRELEN(old_re) == plen
7588 && memEQ(RX_PRECOMP(old_re), exp, plen)
7589 && !runtime_code /* with runtime code, always recompile */ )
7592 SV *dsv= sv_newmortal();
7593 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, PL_dump_re_max_len);
7594 Perl_re_printf( aTHX_ "%sSkipping recompilation of unchanged REx%s %s\n",
7595 PL_colors[4], PL_colors[5], s);
7600 /* Allocate the pattern's SV */
7601 RExC_rx_sv = Rx = (REGEXP*) newSV_type(SVt_REGEXP);
7602 RExC_rx = ReANY(Rx);
7603 if ( RExC_rx == NULL )
7604 FAIL("Regexp out of space");
7606 rx_flags = orig_rx_flags;
7608 if ( (UTF || RExC_uni_semantics)
7609 && initial_charset == REGEX_DEPENDS_CHARSET)
7612 /* Set to use unicode semantics if the pattern is in utf8 and has the
7613 * 'depends' charset specified, as it means unicode when utf8 */
7614 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
7615 RExC_uni_semantics = 1;
7618 RExC_pm_flags = pm_flags;
7621 assert(TAINTING_get || !TAINT_get);
7623 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
7625 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
7626 /* whoops, we have a non-utf8 pattern, whilst run-time code
7627 * got compiled as utf8. Try again with a utf8 pattern */
7628 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7629 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7633 assert(!pRExC_state->runtime_code_qr);
7639 RExC_in_lookbehind = 0;
7640 RExC_in_lookahead = 0;
7641 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
7642 RExC_recode_x_to_native = 0;
7643 RExC_in_multi_char_class = 0;
7645 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = RExC_precomp = exp;
7646 RExC_precomp_end = RExC_end = exp + plen;
7648 RExC_whilem_seen = 0;
7650 RExC_recurse = NULL;
7651 RExC_study_chunk_recursed = NULL;
7652 RExC_study_chunk_recursed_bytes= 0;
7653 RExC_recurse_count = 0;
7654 pRExC_state->code_index = 0;
7656 /* Initialize the string in the compiled pattern. This is so that there is
7657 * something to output if necessary */
7658 set_regex_pv(pRExC_state, Rx);
7661 Perl_re_printf( aTHX_
7662 "Starting parse and generation\n");
7664 RExC_lastparse=NULL;
7667 /* Allocate space and zero-initialize. Note, the two step process
7668 of zeroing when in debug mode, thus anything assigned has to
7669 happen after that */
7672 /* On the first pass of the parse, we guess how big this will be. Then
7673 * we grow in one operation to that amount and then give it back. As
7674 * we go along, we re-allocate what we need.
7676 * XXX Currently the guess is essentially that the pattern will be an
7677 * EXACT node with one byte input, one byte output. This is crude, and
7678 * better heuristics are welcome.
7680 * On any subsequent passes, we guess what we actually computed in the
7681 * latest earlier pass. Such a pass probably didn't complete so is
7682 * missing stuff. We could improve those guesses by knowing where the
7683 * parse stopped, and use the length so far plus apply the above
7684 * assumption to what's left. */
7685 RExC_size = STR_SZ(RExC_end - RExC_start);
7688 Newxc(RExC_rxi, sizeof(regexp_internal) + RExC_size, char, regexp_internal);
7689 if ( RExC_rxi == NULL )
7690 FAIL("Regexp out of space");
7692 Zero(RExC_rxi, sizeof(regexp_internal) + RExC_size, char);
7693 RXi_SET( RExC_rx, RExC_rxi );
7695 /* We start from 0 (over from 0 in the case this is a reparse. The first
7696 * node parsed will give back any excess memory we have allocated so far).
7700 /* non-zero initialization begins here */
7701 RExC_rx->engine= eng;
7702 RExC_rx->extflags = rx_flags;
7703 RXp_COMPFLAGS(RExC_rx) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
7705 if (pm_flags & PMf_IS_QR) {
7706 RExC_rxi->code_blocks = pRExC_state->code_blocks;
7707 if (RExC_rxi->code_blocks) {
7708 RExC_rxi->code_blocks->refcnt++;
7712 RExC_rx->intflags = 0;
7714 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
7717 /* This NUL is guaranteed because the pattern comes from an SV*, and the sv
7718 * code makes sure the final byte is an uncounted NUL. But should this
7719 * ever not be the case, lots of things could read beyond the end of the
7720 * buffer: loops like
7721 * while(isFOO(*RExC_parse)) RExC_parse++;
7722 * strchr(RExC_parse, "foo");
7723 * etc. So it is worth noting. */
7724 assert(*RExC_end == '\0');
7728 RExC_parens_buf_size = 0;
7729 RExC_emit_start = RExC_rxi->program;
7730 pRExC_state->code_index = 0;
7732 *((char*) RExC_emit_start) = (char) REG_MAGIC;
7736 if (reg(pRExC_state, 0, &flags, 1)) {
7738 /* Success!, But we may need to redo the parse knowing how many parens
7739 * there actually are */
7740 if (IN_PARENS_PASS) {
7741 flags |= RESTART_PARSE;
7744 /* We have that number in RExC_npar */
7745 RExC_total_parens = RExC_npar;
7747 else if (! MUST_RESTART(flags)) {
7749 Perl_croak(aTHX_ "panic: reg returned failure to re_op_compile, flags=%#" UVxf, (UV) flags);
7752 /* Here, we either have success, or we have to redo the parse for some reason */
7753 if (MUST_RESTART(flags)) {
7755 /* It's possible to write a regexp in ascii that represents Unicode
7756 codepoints outside of the byte range, such as via \x{100}. If we
7757 detect such a sequence we have to convert the entire pattern to utf8
7758 and then recompile, as our sizing calculation will have been based
7759 on 1 byte == 1 character, but we will need to use utf8 to encode
7760 at least some part of the pattern, and therefore must convert the whole
7763 if (flags & NEED_UTF8) {
7765 /* We have stored the offset of the final warning output so far.
7766 * That must be adjusted. Any variant characters between the start
7767 * of the pattern and this warning count for 2 bytes in the final,
7768 * so just add them again */
7769 if (UNLIKELY(RExC_latest_warn_offset > 0)) {
7770 RExC_latest_warn_offset +=
7771 variant_under_utf8_count((U8 *) exp, (U8 *) exp
7772 + RExC_latest_warn_offset);
7774 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
7775 pRExC_state->code_blocks ? pRExC_state->code_blocks->count : 0);
7776 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse after upgrade\n"));
7779 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "Need to redo parse\n"));
7782 if (ALL_PARENS_COUNTED) {
7783 /* Make enough room for all the known parens, and zero it */
7784 Renew(RExC_open_parens, RExC_total_parens, regnode_offset);
7785 Zero(RExC_open_parens, RExC_total_parens, regnode_offset);
7786 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
7788 Renew(RExC_close_parens, RExC_total_parens, regnode_offset);
7789 Zero(RExC_close_parens, RExC_total_parens, regnode_offset);
7791 else { /* Parse did not complete. Reinitialize the parentheses
7793 RExC_total_parens = 0;
7794 if (RExC_open_parens) {
7795 Safefree(RExC_open_parens);
7796 RExC_open_parens = NULL;
7798 if (RExC_close_parens) {
7799 Safefree(RExC_close_parens);
7800 RExC_close_parens = NULL;
7804 /* Clean up what we did in this parse */
7805 SvREFCNT_dec_NN(RExC_rx_sv);
7810 /* Here, we have successfully parsed and generated the pattern's program
7811 * for the regex engine. We are ready to finish things up and look for
7814 /* Update the string to compile, with correct modifiers, etc */
7815 set_regex_pv(pRExC_state, Rx);
7817 RExC_rx->nparens = RExC_total_parens - 1;
7819 /* Uses the upper 4 bits of the FLAGS field, so keep within that size */
7820 if (RExC_whilem_seen > 15)
7821 RExC_whilem_seen = 15;
7824 Perl_re_printf( aTHX_
7825 "Required size %" IVdf " nodes\n", (IV)RExC_size);
7827 RExC_lastparse=NULL;
7830 #ifdef RE_TRACK_PATTERN_OFFSETS
7831 DEBUG_OFFSETS_r(Perl_re_printf( aTHX_
7832 "%s %" UVuf " bytes for offset annotations.\n",
7833 RExC_offsets ? "Got" : "Couldn't get",
7834 (UV)((RExC_offsets[0] * 2 + 1))));
7835 DEBUG_OFFSETS_r(if (RExC_offsets) {
7836 const STRLEN len = RExC_offsets[0];
7838 GET_RE_DEBUG_FLAGS_DECL;
7839 Perl_re_printf( aTHX_
7840 "Offsets: [%" UVuf "]\n\t", (UV)RExC_offsets[0]);
7841 for (i = 1; i <= len; i++) {
7842 if (RExC_offsets[i*2-1] || RExC_offsets[i*2])
7843 Perl_re_printf( aTHX_ "%" UVuf ":%" UVuf "[%" UVuf "] ",
7844 (UV)i, (UV)RExC_offsets[i*2-1], (UV)RExC_offsets[i*2]);
7846 Perl_re_printf( aTHX_ "\n");
7850 SetProgLen(RExC_rxi,RExC_size);
7853 DEBUG_DUMP_PRE_OPTIMIZE_r({
7854 SV * const sv = sv_newmortal();
7855 RXi_GET_DECL(RExC_rx, ri);
7857 Perl_re_printf( aTHX_ "Program before optimization:\n");
7859 (void)dumpuntil(RExC_rx, ri->program, ri->program + 1, NULL, NULL,
7864 Perl_re_printf( aTHX_ "Starting post parse optimization\n");
7867 /* XXXX To minimize changes to RE engine we always allocate
7868 3-units-long substrs field. */
7869 Newx(RExC_rx->substrs, 1, struct reg_substr_data);
7870 if (RExC_recurse_count) {
7871 Newx(RExC_recurse, RExC_recurse_count, regnode *);
7872 SAVEFREEPV(RExC_recurse);
7875 if (RExC_seen & REG_RECURSE_SEEN) {
7876 /* Note, RExC_total_parens is 1 + the number of parens in a pattern.
7877 * So its 1 if there are no parens. */
7878 RExC_study_chunk_recursed_bytes= (RExC_total_parens >> 3) +
7879 ((RExC_total_parens & 0x07) != 0);
7880 Newx(RExC_study_chunk_recursed,
7881 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7882 SAVEFREEPV(RExC_study_chunk_recursed);
7886 RExC_rx->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
7888 RExC_study_chunk_recursed_count= 0;
7890 Zero(RExC_rx->substrs, 1, struct reg_substr_data);
7891 if (RExC_study_chunk_recursed) {
7892 Zero(RExC_study_chunk_recursed,
7893 RExC_study_chunk_recursed_bytes * RExC_total_parens, U8);
7897 #ifdef TRIE_STUDY_OPT
7899 StructCopy(&zero_scan_data, &data, scan_data_t);
7900 copyRExC_state = RExC_state;
7903 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ "Restudying\n"));
7905 RExC_state = copyRExC_state;
7906 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
7907 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
7909 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
7910 StructCopy(&zero_scan_data, &data, scan_data_t);
7913 StructCopy(&zero_scan_data, &data, scan_data_t);
7916 /* Dig out information for optimizations. */
7917 RExC_rx->extflags = RExC_flags; /* was pm_op */
7918 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
7921 SvUTF8_on(Rx); /* Unicode in it? */
7922 RExC_rxi->regstclass = NULL;
7923 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
7924 RExC_rx->intflags |= PREGf_NAUGHTY;
7925 scan = RExC_rxi->program + 1; /* First BRANCH. */
7927 /* testing for BRANCH here tells us whether there is "must appear"
7928 data in the pattern. If there is then we can use it for optimisations */
7929 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7932 STRLEN longest_length[2];
7933 regnode_ssc ch_class; /* pointed to by data */
7935 SSize_t last_close = 0; /* pointed to by data */
7936 regnode *first= scan;
7937 regnode *first_next= regnext(first);
7941 * Skip introductions and multiplicators >= 1
7942 * so that we can extract the 'meat' of the pattern that must
7943 * match in the large if() sequence following.
7944 * NOTE that EXACT is NOT covered here, as it is normally
7945 * picked up by the optimiser separately.
7947 * This is unfortunate as the optimiser isnt handling lookahead
7948 * properly currently.
7951 while ((OP(first) == OPEN && (sawopen = 1)) ||
7952 /* An OR of *one* alternative - should not happen now. */
7953 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7954 /* for now we can't handle lookbehind IFMATCH*/
7955 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7956 (OP(first) == PLUS) ||
7957 (OP(first) == MINMOD) ||
7958 /* An {n,m} with n>0 */
7959 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7960 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7963 * the only op that could be a regnode is PLUS, all the rest
7964 * will be regnode_1 or regnode_2.
7966 * (yves doesn't think this is true)
7968 if (OP(first) == PLUS)
7971 if (OP(first) == MINMOD)
7973 first += regarglen[OP(first)];
7975 first = NEXTOPER(first);
7976 first_next= regnext(first);
7979 /* Starting-point info. */
7981 DEBUG_PEEP("first:", first, 0, 0);
7982 /* Ignore EXACT as we deal with it later. */
7983 if (PL_regkind[OP(first)] == EXACT) {
7984 if ( OP(first) == EXACT
7985 || OP(first) == LEXACT
7986 || OP(first) == EXACT_REQ8
7987 || OP(first) == LEXACT_REQ8
7988 || OP(first) == EXACTL)
7990 NOOP; /* Empty, get anchored substr later. */
7993 RExC_rxi->regstclass = first;
7996 else if (PL_regkind[OP(first)] == TRIE &&
7997 ((reg_trie_data *)RExC_rxi->data->data[ ARG(first) ])->minlen>0)
7999 /* this can happen only on restudy */
8000 RExC_rxi->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
8003 else if (REGNODE_SIMPLE(OP(first)))
8004 RExC_rxi->regstclass = first;
8005 else if (PL_regkind[OP(first)] == BOUND ||
8006 PL_regkind[OP(first)] == NBOUND)
8007 RExC_rxi->regstclass = first;
8008 else if (PL_regkind[OP(first)] == BOL) {
8009 RExC_rx->intflags |= (OP(first) == MBOL
8012 first = NEXTOPER(first);
8015 else if (OP(first) == GPOS) {
8016 RExC_rx->intflags |= PREGf_ANCH_GPOS;
8017 first = NEXTOPER(first);
8020 else if ((!sawopen || !RExC_sawback) &&
8022 (OP(first) == STAR &&
8023 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
8024 !(RExC_rx->intflags & PREGf_ANCH) && !pRExC_state->code_blocks)
8026 /* turn .* into ^.* with an implied $*=1 */
8028 (OP(NEXTOPER(first)) == REG_ANY)
8031 RExC_rx->intflags |= (type | PREGf_IMPLICIT);
8032 first = NEXTOPER(first);
8035 if (sawplus && !sawminmod && !sawlookahead
8036 && (!sawopen || !RExC_sawback)
8037 && !pRExC_state->code_blocks) /* May examine pos and $& */
8038 /* x+ must match at the 1st pos of run of x's */
8039 RExC_rx->intflags |= PREGf_SKIP;
8041 /* Scan is after the zeroth branch, first is atomic matcher. */
8042 #ifdef TRIE_STUDY_OPT
8045 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8046 (IV)(first - scan + 1))
8050 Perl_re_printf( aTHX_ "first at %" IVdf "\n",
8051 (IV)(first - scan + 1))
8057 * If there's something expensive in the r.e., find the
8058 * longest literal string that must appear and make it the
8059 * regmust. Resolve ties in favor of later strings, since
8060 * the regstart check works with the beginning of the r.e.
8061 * and avoiding duplication strengthens checking. Not a
8062 * strong reason, but sufficient in the absence of others.
8063 * [Now we resolve ties in favor of the earlier string if
8064 * it happens that c_offset_min has been invalidated, since the
8065 * earlier string may buy us something the later one won't.]
8068 data.substrs[0].str = newSVpvs("");
8069 data.substrs[1].str = newSVpvs("");
8070 data.last_found = newSVpvs("");
8071 data.cur_is_floating = 0; /* initially any found substring is fixed */
8072 ENTER_with_name("study_chunk");
8073 SAVEFREESV(data.substrs[0].str);
8074 SAVEFREESV(data.substrs[1].str);
8075 SAVEFREESV(data.last_found);
8077 if (!RExC_rxi->regstclass) {
8078 ssc_init(pRExC_state, &ch_class);
8079 data.start_class = &ch_class;
8080 stclass_flag = SCF_DO_STCLASS_AND;
8081 } else /* XXXX Check for BOUND? */
8083 data.last_closep = &last_close;
8087 * MAIN ENTRY FOR study_chunk() FOR m/PATTERN/
8088 * (NO top level branches)
8090 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
8091 scan + RExC_size, /* Up to end */
8093 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
8094 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
8098 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
8101 if ( RExC_total_parens == 1 && !data.cur_is_floating
8102 && data.last_start_min == 0 && data.last_end > 0
8103 && !RExC_seen_zerolen
8104 && !(RExC_seen & REG_VERBARG_SEEN)
8105 && !(RExC_seen & REG_GPOS_SEEN)
8107 RExC_rx->extflags |= RXf_CHECK_ALL;
8109 scan_commit(pRExC_state, &data,&minlen, 0);
8112 /* XXX this is done in reverse order because that's the way the
8113 * code was before it was parameterised. Don't know whether it
8114 * actually needs doing in reverse order. DAPM */
8115 for (i = 1; i >= 0; i--) {
8116 longest_length[i] = CHR_SVLEN(data.substrs[i].str);
8119 && SvCUR(data.substrs[0].str) /* ok to leave SvCUR */
8120 && data.substrs[0].min_offset
8121 == data.substrs[1].min_offset
8122 && SvCUR(data.substrs[0].str)
8123 == SvCUR(data.substrs[1].str)
8125 && S_setup_longest (aTHX_ pRExC_state,
8126 &(RExC_rx->substrs->data[i]),
8130 RExC_rx->substrs->data[i].min_offset =
8131 data.substrs[i].min_offset - data.substrs[i].lookbehind;
8133 RExC_rx->substrs->data[i].max_offset = data.substrs[i].max_offset;
8134 /* Don't offset infinity */
8135 if (data.substrs[i].max_offset < SSize_t_MAX)
8136 RExC_rx->substrs->data[i].max_offset -= data.substrs[i].lookbehind;
8137 SvREFCNT_inc_simple_void_NN(data.substrs[i].str);
8140 RExC_rx->substrs->data[i].substr = NULL;
8141 RExC_rx->substrs->data[i].utf8_substr = NULL;
8142 longest_length[i] = 0;
8146 LEAVE_with_name("study_chunk");
8148 if (RExC_rxi->regstclass
8149 && (OP(RExC_rxi->regstclass) == REG_ANY || OP(RExC_rxi->regstclass) == SANY))
8150 RExC_rxi->regstclass = NULL;
8152 if ((!(RExC_rx->substrs->data[0].substr || RExC_rx->substrs->data[0].utf8_substr)
8153 || RExC_rx->substrs->data[0].min_offset)
8155 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8156 && is_ssc_worth_it(pRExC_state, data.start_class))
8158 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8160 ssc_finalize(pRExC_state, data.start_class);
8162 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8163 StructCopy(data.start_class,
8164 (regnode_ssc*)RExC_rxi->data->data[n],
8166 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8167 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8168 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
8169 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8170 Perl_re_printf( aTHX_
8171 "synthetic stclass \"%s\".\n",
8172 SvPVX_const(sv));});
8173 data.start_class = NULL;
8176 /* A temporary algorithm prefers floated substr to fixed one of
8177 * same length to dig more info. */
8178 i = (longest_length[0] <= longest_length[1]);
8179 RExC_rx->substrs->check_ix = i;
8180 RExC_rx->check_end_shift = RExC_rx->substrs->data[i].end_shift;
8181 RExC_rx->check_substr = RExC_rx->substrs->data[i].substr;
8182 RExC_rx->check_utf8 = RExC_rx->substrs->data[i].utf8_substr;
8183 RExC_rx->check_offset_min = RExC_rx->substrs->data[i].min_offset;
8184 RExC_rx->check_offset_max = RExC_rx->substrs->data[i].max_offset;
8185 if (!i && (RExC_rx->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS)))
8186 RExC_rx->intflags |= PREGf_NOSCAN;
8188 if ((RExC_rx->check_substr || RExC_rx->check_utf8) ) {
8189 RExC_rx->extflags |= RXf_USE_INTUIT;
8190 if (SvTAIL(RExC_rx->check_substr ? RExC_rx->check_substr : RExC_rx->check_utf8))
8191 RExC_rx->extflags |= RXf_INTUIT_TAIL;
8194 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
8195 if ( (STRLEN)minlen < longest_length[1] )
8196 minlen= longest_length[1];
8197 if ( (STRLEN)minlen < longest_length[0] )
8198 minlen= longest_length[0];
8202 /* Several toplevels. Best we can is to set minlen. */
8204 regnode_ssc ch_class;
8205 SSize_t last_close = 0;
8207 DEBUG_PARSE_r(Perl_re_printf( aTHX_ "\nMulti Top Level\n"));
8209 scan = RExC_rxi->program + 1;
8210 ssc_init(pRExC_state, &ch_class);
8211 data.start_class = &ch_class;
8212 data.last_closep = &last_close;
8216 * MAIN ENTRY FOR study_chunk() FOR m/P1|P2|.../
8217 * (patterns WITH top level branches)
8219 minlen = study_chunk(pRExC_state,
8220 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
8221 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
8222 ? SCF_TRIE_DOING_RESTUDY
8226 CHECK_RESTUDY_GOTO_butfirst(NOOP);
8228 RExC_rx->check_substr = NULL;
8229 RExC_rx->check_utf8 = NULL;
8230 RExC_rx->substrs->data[0].substr = NULL;
8231 RExC_rx->substrs->data[0].utf8_substr = NULL;
8232 RExC_rx->substrs->data[1].substr = NULL;
8233 RExC_rx->substrs->data[1].utf8_substr = NULL;
8235 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
8236 && is_ssc_worth_it(pRExC_state, data.start_class))
8238 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
8240 ssc_finalize(pRExC_state, data.start_class);
8242 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
8243 StructCopy(data.start_class,
8244 (regnode_ssc*)RExC_rxi->data->data[n],
8246 RExC_rxi->regstclass = (regnode*)RExC_rxi->data->data[n];
8247 RExC_rx->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
8248 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
8249 regprop(RExC_rx, sv, (regnode*)data.start_class, NULL, pRExC_state);
8250 Perl_re_printf( aTHX_
8251 "synthetic stclass \"%s\".\n",
8252 SvPVX_const(sv));});
8253 data.start_class = NULL;
8257 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
8258 RExC_rx->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
8259 RExC_rx->maxlen = REG_INFTY;
8262 RExC_rx->maxlen = RExC_maxlen;
8265 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
8266 the "real" pattern. */
8268 Perl_re_printf( aTHX_ "minlen: %" IVdf " RExC_rx->minlen:%" IVdf " maxlen:%" IVdf "\n",
8269 (IV)minlen, (IV)RExC_rx->minlen, (IV)RExC_maxlen);
8271 RExC_rx->minlenret = minlen;
8272 if (RExC_rx->minlen < minlen)
8273 RExC_rx->minlen = minlen;
8275 if (RExC_seen & REG_RECURSE_SEEN ) {
8276 RExC_rx->intflags |= PREGf_RECURSE_SEEN;
8277 Newx(RExC_rx->recurse_locinput, RExC_rx->nparens + 1, char *);
8279 if (RExC_seen & REG_GPOS_SEEN)
8280 RExC_rx->intflags |= PREGf_GPOS_SEEN;
8281 if (RExC_seen & REG_LOOKBEHIND_SEEN)
8282 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
8284 if (pRExC_state->code_blocks)
8285 RExC_rx->extflags |= RXf_EVAL_SEEN;
8286 if (RExC_seen & REG_VERBARG_SEEN)
8288 RExC_rx->intflags |= PREGf_VERBARG_SEEN;
8289 RExC_rx->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
8291 if (RExC_seen & REG_CUTGROUP_SEEN)
8292 RExC_rx->intflags |= PREGf_CUTGROUP_SEEN;
8293 if (pm_flags & PMf_USE_RE_EVAL)
8294 RExC_rx->intflags |= PREGf_USE_RE_EVAL;
8295 if (RExC_paren_names)
8296 RXp_PAREN_NAMES(RExC_rx) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
8298 RXp_PAREN_NAMES(RExC_rx) = NULL;
8300 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
8301 * so it can be used in pp.c */
8302 if (RExC_rx->intflags & PREGf_ANCH)
8303 RExC_rx->extflags |= RXf_IS_ANCHORED;
8307 /* this is used to identify "special" patterns that might result
8308 * in Perl NOT calling the regex engine and instead doing the match "itself",
8309 * particularly special cases in split//. By having the regex compiler
8310 * do this pattern matching at a regop level (instead of by inspecting the pattern)
8311 * we avoid weird issues with equivalent patterns resulting in different behavior,
8312 * AND we allow non Perl engines to get the same optimizations by the setting the
8313 * flags appropriately - Yves */
8314 regnode *first = RExC_rxi->program + 1;
8316 regnode *next = regnext(first);
8319 if (PL_regkind[fop] == NOTHING && nop == END)
8320 RExC_rx->extflags |= RXf_NULL;
8321 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
8322 /* when fop is SBOL first->flags will be true only when it was
8323 * produced by parsing /\A/, and not when parsing /^/. This is
8324 * very important for the split code as there we want to
8325 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
8326 * See rt #122761 for more details. -- Yves */
8327 RExC_rx->extflags |= RXf_START_ONLY;
8328 else if (fop == PLUS
8329 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
8331 RExC_rx->extflags |= RXf_WHITE;
8332 else if ( RExC_rx->extflags & RXf_SPLIT
8333 && ( fop == EXACT || fop == LEXACT
8334 || fop == EXACT_REQ8 || fop == LEXACT_REQ8
8336 && STR_LEN(first) == 1
8337 && *(STRING(first)) == ' '
8339 RExC_rx->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
8343 if (RExC_contains_locale) {
8344 RXp_EXTFLAGS(RExC_rx) |= RXf_TAINTED;
8348 if (RExC_paren_names) {
8349 RExC_rxi->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
8350 RExC_rxi->data->data[RExC_rxi->name_list_idx]
8351 = (void*)SvREFCNT_inc(RExC_paren_name_list);
8354 RExC_rxi->name_list_idx = 0;
8356 while ( RExC_recurse_count > 0 ) {
8357 const regnode *scan = RExC_recurse[ --RExC_recurse_count ];
8359 * This data structure is set up in study_chunk() and is used
8360 * to calculate the distance between a GOSUB regopcode and
8361 * the OPEN/CURLYM (CURLYM's are special and can act like OPEN's)
8364 * If for some reason someone writes code that optimises
8365 * away a GOSUB opcode then the assert should be changed to
8366 * an if(scan) to guard the ARG2L_SET() - Yves
8369 assert(scan && OP(scan) == GOSUB);
8370 ARG2L_SET( scan, RExC_open_parens[ARG(scan)] - REGNODE_OFFSET(scan));
8373 Newxz(RExC_rx->offs, RExC_total_parens, regexp_paren_pair);
8374 /* assume we don't need to swap parens around before we match */
8376 Perl_re_printf( aTHX_ "study_chunk_recursed_count: %lu\n",
8377 (unsigned long)RExC_study_chunk_recursed_count);
8381 Perl_re_printf( aTHX_ "Final program:\n");
8385 if (RExC_open_parens) {
8386 Safefree(RExC_open_parens);
8387 RExC_open_parens = NULL;
8389 if (RExC_close_parens) {
8390 Safefree(RExC_close_parens);
8391 RExC_close_parens = NULL;
8395 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
8396 * by setting the regexp SV to readonly-only instead. If the
8397 * pattern's been recompiled, the USEDness should remain. */
8398 if (old_re && SvREADONLY(old_re))
8406 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
8409 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
8411 PERL_UNUSED_ARG(value);
8413 if (flags & RXapif_FETCH) {
8414 return reg_named_buff_fetch(rx, key, flags);
8415 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
8416 Perl_croak_no_modify();
8418 } else if (flags & RXapif_EXISTS) {
8419 return reg_named_buff_exists(rx, key, flags)
8422 } else if (flags & RXapif_REGNAMES) {
8423 return reg_named_buff_all(rx, flags);
8424 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
8425 return reg_named_buff_scalar(rx, flags);
8427 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
8433 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
8436 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
8437 PERL_UNUSED_ARG(lastkey);
8439 if (flags & RXapif_FIRSTKEY)
8440 return reg_named_buff_firstkey(rx, flags);
8441 else if (flags & RXapif_NEXTKEY)
8442 return reg_named_buff_nextkey(rx, flags);
8444 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
8451 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
8455 struct regexp *const rx = ReANY(r);
8457 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
8459 if (rx && RXp_PAREN_NAMES(rx)) {
8460 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
8463 SV* sv_dat=HeVAL(he_str);
8464 I32 *nums=(I32*)SvPVX(sv_dat);
8465 AV * const retarray = (flags & RXapif_ALL) ? newAV() : NULL;
8466 for ( i=0; i<SvIVX(sv_dat); i++ ) {
8467 if ((I32)(rx->nparens) >= nums[i]
8468 && rx->offs[nums[i]].start != -1
8469 && rx->offs[nums[i]].end != -1)
8472 CALLREG_NUMBUF_FETCH(r, nums[i], ret);
8477 ret = newSVsv(&PL_sv_undef);
8480 av_push(retarray, ret);
8483 return newRV_noinc(MUTABLE_SV(retarray));
8490 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
8493 struct regexp *const rx = ReANY(r);
8495 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
8497 if (rx && RXp_PAREN_NAMES(rx)) {
8498 if (flags & RXapif_ALL) {
8499 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
8501 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
8503 SvREFCNT_dec_NN(sv);
8515 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
8517 struct regexp *const rx = ReANY(r);
8519 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
8521 if ( rx && RXp_PAREN_NAMES(rx) ) {
8522 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
8524 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
8531 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
8533 struct regexp *const rx = ReANY(r);
8534 GET_RE_DEBUG_FLAGS_DECL;
8536 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
8538 if (rx && RXp_PAREN_NAMES(rx)) {
8539 HV *hv = RXp_PAREN_NAMES(rx);
8541 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8544 SV* sv_dat = HeVAL(temphe);
8545 I32 *nums = (I32*)SvPVX(sv_dat);
8546 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8547 if ((I32)(rx->lastparen) >= nums[i] &&
8548 rx->offs[nums[i]].start != -1 &&
8549 rx->offs[nums[i]].end != -1)
8555 if (parno || flags & RXapif_ALL) {
8556 return newSVhek(HeKEY_hek(temphe));
8564 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
8569 struct regexp *const rx = ReANY(r);
8571 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
8573 if (rx && RXp_PAREN_NAMES(rx)) {
8574 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
8575 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
8576 } else if (flags & RXapif_ONE) {
8577 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
8578 av = MUTABLE_AV(SvRV(ret));
8579 length = av_tindex(av);
8580 SvREFCNT_dec_NN(ret);
8581 return newSViv(length + 1);
8583 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
8588 return &PL_sv_undef;
8592 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
8594 struct regexp *const rx = ReANY(r);
8597 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
8599 if (rx && RXp_PAREN_NAMES(rx)) {
8600 HV *hv= RXp_PAREN_NAMES(rx);
8602 (void)hv_iterinit(hv);
8603 while ( (temphe = hv_iternext_flags(hv, 0)) ) {
8606 SV* sv_dat = HeVAL(temphe);
8607 I32 *nums = (I32*)SvPVX(sv_dat);
8608 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
8609 if ((I32)(rx->lastparen) >= nums[i] &&
8610 rx->offs[nums[i]].start != -1 &&
8611 rx->offs[nums[i]].end != -1)
8617 if (parno || flags & RXapif_ALL) {
8618 av_push(av, newSVhek(HeKEY_hek(temphe)));
8623 return newRV_noinc(MUTABLE_SV(av));
8627 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
8630 struct regexp *const rx = ReANY(r);
8636 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
8638 if ( n == RX_BUFF_IDX_CARET_PREMATCH
8639 || n == RX_BUFF_IDX_CARET_FULLMATCH
8640 || n == RX_BUFF_IDX_CARET_POSTMATCH
8643 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8645 /* on something like
8648 * the KEEPCOPY is set on the PMOP rather than the regex */
8649 if (PL_curpm && r == PM_GETRE(PL_curpm))
8650 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8659 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
8660 /* no need to distinguish between them any more */
8661 n = RX_BUFF_IDX_FULLMATCH;
8663 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
8664 && rx->offs[0].start != -1)
8666 /* $`, ${^PREMATCH} */
8667 i = rx->offs[0].start;
8671 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
8672 && rx->offs[0].end != -1)
8674 /* $', ${^POSTMATCH} */
8675 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
8676 i = rx->sublen + rx->suboffset - rx->offs[0].end;
8679 if (inRANGE(n, 0, (I32)rx->nparens) &&
8680 (s1 = rx->offs[n].start) != -1 &&
8681 (t1 = rx->offs[n].end) != -1)
8683 /* $&, ${^MATCH}, $1 ... */
8685 s = rx->subbeg + s1 - rx->suboffset;
8690 assert(s >= rx->subbeg);
8691 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
8693 #ifdef NO_TAINT_SUPPORT
8694 sv_setpvn(sv, s, i);
8696 const int oldtainted = TAINT_get;
8698 sv_setpvn(sv, s, i);
8699 TAINT_set(oldtainted);
8701 if (RXp_MATCH_UTF8(rx))
8706 if (RXp_MATCH_TAINTED(rx)) {
8707 if (SvTYPE(sv) >= SVt_PVMG) {
8708 MAGIC* const mg = SvMAGIC(sv);
8711 SvMAGIC_set(sv, mg->mg_moremagic);
8713 if ((mgt = SvMAGIC(sv))) {
8714 mg->mg_moremagic = mgt;
8715 SvMAGIC_set(sv, mg);
8732 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
8733 SV const * const value)
8735 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
8737 PERL_UNUSED_ARG(rx);
8738 PERL_UNUSED_ARG(paren);
8739 PERL_UNUSED_ARG(value);
8742 Perl_croak_no_modify();
8746 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
8749 struct regexp *const rx = ReANY(r);
8753 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
8755 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
8756 || paren == RX_BUFF_IDX_CARET_FULLMATCH
8757 || paren == RX_BUFF_IDX_CARET_POSTMATCH
8760 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
8762 /* on something like
8765 * the KEEPCOPY is set on the PMOP rather than the regex */
8766 if (PL_curpm && r == PM_GETRE(PL_curpm))
8767 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
8773 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
8775 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
8776 case RX_BUFF_IDX_PREMATCH: /* $` */
8777 if (rx->offs[0].start != -1) {
8778 i = rx->offs[0].start;
8787 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
8788 case RX_BUFF_IDX_POSTMATCH: /* $' */
8789 if (rx->offs[0].end != -1) {
8790 i = rx->sublen - rx->offs[0].end;
8792 s1 = rx->offs[0].end;
8799 default: /* $& / ${^MATCH}, $1, $2, ... */
8800 if (paren <= (I32)rx->nparens &&
8801 (s1 = rx->offs[paren].start) != -1 &&
8802 (t1 = rx->offs[paren].end) != -1)
8808 if (ckWARN(WARN_UNINITIALIZED))
8809 report_uninit((const SV *)sv);
8814 if (i > 0 && RXp_MATCH_UTF8(rx)) {
8815 const char * const s = rx->subbeg - rx->suboffset + s1;
8820 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
8827 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
8829 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
8830 PERL_UNUSED_ARG(rx);
8834 return newSVpvs("Regexp");
8837 /* Scans the name of a named buffer from the pattern.
8838 * If flags is REG_RSN_RETURN_NULL returns null.
8839 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
8840 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
8841 * to the parsed name as looked up in the RExC_paren_names hash.
8842 * If there is an error throws a vFAIL().. type exception.
8845 #define REG_RSN_RETURN_NULL 0
8846 #define REG_RSN_RETURN_NAME 1
8847 #define REG_RSN_RETURN_DATA 2
8850 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
8852 char *name_start = RExC_parse;
8855 PERL_ARGS_ASSERT_REG_SCAN_NAME;
8857 assert (RExC_parse <= RExC_end);
8858 if (RExC_parse == RExC_end) NOOP;
8859 else if (isIDFIRST_lazy_if_safe(RExC_parse, RExC_end, UTF)) {
8860 /* Note that the code here assumes well-formed UTF-8. Skip IDFIRST by
8861 * using do...while */
8864 RExC_parse += UTF8SKIP(RExC_parse);
8865 } while ( RExC_parse < RExC_end
8866 && isWORDCHAR_utf8_safe((U8*)RExC_parse, (U8*) RExC_end));
8870 } while (RExC_parse < RExC_end && isWORDCHAR(*RExC_parse));
8872 RExC_parse++; /* so the <- from the vFAIL is after the offending
8874 vFAIL("Group name must start with a non-digit word character");
8876 sv_name = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
8877 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
8878 if ( flags == REG_RSN_RETURN_NAME)
8880 else if (flags==REG_RSN_RETURN_DATA) {
8883 if ( ! sv_name ) /* should not happen*/
8884 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
8885 if (RExC_paren_names)
8886 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
8888 sv_dat = HeVAL(he_str);
8889 if ( ! sv_dat ) { /* Didn't find group */
8891 /* It might be a forward reference; we can't fail until we
8892 * know, by completing the parse to get all the groups, and
8894 if (ALL_PARENS_COUNTED) {
8895 vFAIL("Reference to nonexistent named group");
8898 REQUIRE_PARENS_PASS;
8904 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
8905 (unsigned long) flags);
8908 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
8909 if (RExC_lastparse!=RExC_parse) { \
8910 Perl_re_printf( aTHX_ "%s", \
8911 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
8912 RExC_end - RExC_parse, 16, \
8914 PERL_PV_ESCAPE_UNI_DETECT | \
8915 PERL_PV_PRETTY_ELLIPSES | \
8916 PERL_PV_PRETTY_LTGT | \
8917 PERL_PV_ESCAPE_RE | \
8918 PERL_PV_PRETTY_EXACTSIZE \
8922 Perl_re_printf( aTHX_ "%16s",""); \
8924 if (RExC_lastnum!=RExC_emit) \
8925 Perl_re_printf( aTHX_ "|%4d", RExC_emit); \
8927 Perl_re_printf( aTHX_ "|%4s",""); \
8928 Perl_re_printf( aTHX_ "|%*s%-4s", \
8929 (int)((depth*2)), "", \
8932 RExC_lastnum=RExC_emit; \
8933 RExC_lastparse=RExC_parse; \
8938 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8939 DEBUG_PARSE_MSG((funcname)); \
8940 Perl_re_printf( aTHX_ "%4s","\n"); \
8942 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({\
8943 DEBUG_PARSE_MSG((funcname)); \
8944 Perl_re_printf( aTHX_ fmt "\n",args); \
8947 /* This section of code defines the inversion list object and its methods. The
8948 * interfaces are highly subject to change, so as much as possible is static to
8949 * this file. An inversion list is here implemented as a malloc'd C UV array
8950 * as an SVt_INVLIST scalar.
8952 * An inversion list for Unicode is an array of code points, sorted by ordinal
8953 * number. Each element gives the code point that begins a range that extends
8954 * up-to but not including the code point given by the next element. The final
8955 * element gives the first code point of a range that extends to the platform's
8956 * infinity. The even-numbered elements (invlist[0], invlist[2], invlist[4],
8957 * ...) give ranges whose code points are all in the inversion list. We say
8958 * that those ranges are in the set. The odd-numbered elements give ranges
8959 * whose code points are not in the inversion list, and hence not in the set.
8960 * Thus, element [0] is the first code point in the list. Element [1]
8961 * is the first code point beyond that not in the list; and element [2] is the
8962 * first code point beyond that that is in the list. In other words, the first
8963 * range is invlist[0]..(invlist[1]-1), and all code points in that range are
8964 * in the inversion list. The second range is invlist[1]..(invlist[2]-1), and
8965 * all code points in that range are not in the inversion list. The third
8966 * range invlist[2]..(invlist[3]-1) gives code points that are in the inversion
8967 * list, and so forth. Thus every element whose index is divisible by two
8968 * gives the beginning of a range that is in the list, and every element whose
8969 * index is not divisible by two gives the beginning of a range not in the
8970 * list. If the final element's index is divisible by two, the inversion list
8971 * extends to the platform's infinity; otherwise the highest code point in the
8972 * inversion list is the contents of that element minus 1.
8974 * A range that contains just a single code point N will look like
8976 * invlist[i+1] == N+1
8978 * If N is UV_MAX (the highest representable code point on the machine), N+1 is
8979 * impossible to represent, so element [i+1] is omitted. The single element
8981 * invlist[0] == UV_MAX
8982 * contains just UV_MAX, but is interpreted as matching to infinity.
8984 * Taking the complement (inverting) an inversion list is quite simple, if the
8985 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8986 * This implementation reserves an element at the beginning of each inversion
8987 * list to always contain 0; there is an additional flag in the header which
8988 * indicates if the list begins at the 0, or is offset to begin at the next
8989 * element. This means that the inversion list can be inverted without any
8990 * copying; just flip the flag.
8992 * More about inversion lists can be found in "Unicode Demystified"
8993 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8995 * The inversion list data structure is currently implemented as an SV pointing
8996 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8997 * array of UV whose memory management is automatically handled by the existing
8998 * facilities for SV's.
9000 * Some of the methods should always be private to the implementation, and some
9001 * should eventually be made public */
9003 /* The header definitions are in F<invlist_inline.h> */
9005 #ifndef PERL_IN_XSUB_RE
9007 PERL_STATIC_INLINE UV*
9008 S__invlist_array_init(SV* const invlist, const bool will_have_0)
9010 /* Returns a pointer to the first element in the inversion list's array.
9011 * This is called upon initialization of an inversion list. Where the
9012 * array begins depends on whether the list has the code point U+0000 in it
9013 * or not. The other parameter tells it whether the code that follows this
9014 * call is about to put a 0 in the inversion list or not. The first
9015 * element is either the element reserved for 0, if TRUE, or the element
9016 * after it, if FALSE */
9018 bool* offset = get_invlist_offset_addr(invlist);
9019 UV* zero_addr = (UV *) SvPVX(invlist);
9021 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
9024 assert(! _invlist_len(invlist));
9028 /* 1^1 = 0; 1^0 = 1 */
9029 *offset = 1 ^ will_have_0;
9030 return zero_addr + *offset;
9034 S_invlist_replace_list_destroys_src(pTHX_ SV * dest, SV * src)
9036 /* Replaces the inversion list in 'dest' with the one from 'src'. It
9037 * steals the list from 'src', so 'src' is made to have a NULL list. This
9038 * is similar to what SvSetMagicSV() would do, if it were implemented on
9039 * inversion lists, though this routine avoids a copy */
9041 const UV src_len = _invlist_len(src);
9042 const bool src_offset = *get_invlist_offset_addr(src);
9043 const STRLEN src_byte_len = SvLEN(src);
9044 char * array = SvPVX(src);
9046 const int oldtainted = TAINT_get;
9048 PERL_ARGS_ASSERT_INVLIST_REPLACE_LIST_DESTROYS_SRC;
9050 assert(is_invlist(src));
9051 assert(is_invlist(dest));
9052 assert(! invlist_is_iterating(src));
9053 assert(SvCUR(src) == 0 || SvCUR(src) < SvLEN(src));
9055 /* Make sure it ends in the right place with a NUL, as our inversion list
9056 * manipulations aren't careful to keep this true, but sv_usepvn_flags()
9058 array[src_byte_len - 1] = '\0';
9060 TAINT_NOT; /* Otherwise it breaks */
9061 sv_usepvn_flags(dest,
9065 /* This flag is documented to cause a copy to be avoided */
9066 SV_HAS_TRAILING_NUL);
9067 TAINT_set(oldtainted);
9072 /* Finish up copying over the other fields in an inversion list */
9073 *get_invlist_offset_addr(dest) = src_offset;
9074 invlist_set_len(dest, src_len, src_offset);
9075 *get_invlist_previous_index_addr(dest) = 0;
9076 invlist_iterfinish(dest);
9079 PERL_STATIC_INLINE IV*
9080 S_get_invlist_previous_index_addr(SV* invlist)
9082 /* Return the address of the IV that is reserved to hold the cached index
9084 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
9086 assert(is_invlist(invlist));
9088 return &(((XINVLIST*) SvANY(invlist))->prev_index);
9091 PERL_STATIC_INLINE IV
9092 S_invlist_previous_index(SV* const invlist)
9094 /* Returns cached index of previous search */
9096 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
9098 return *get_invlist_previous_index_addr(invlist);
9101 PERL_STATIC_INLINE void
9102 S_invlist_set_previous_index(SV* const invlist, const IV index)
9104 /* Caches <index> for later retrieval */
9106 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
9108 assert(index == 0 || index < (int) _invlist_len(invlist));
9110 *get_invlist_previous_index_addr(invlist) = index;
9113 PERL_STATIC_INLINE void
9114 S_invlist_trim(SV* invlist)
9116 /* Free the not currently-being-used space in an inversion list */
9118 /* But don't free up the space needed for the 0 UV that is always at the
9119 * beginning of the list, nor the trailing NUL */
9120 const UV min_size = TO_INTERNAL_SIZE(1) + 1;
9122 PERL_ARGS_ASSERT_INVLIST_TRIM;
9124 assert(is_invlist(invlist));
9126 SvPV_renew(invlist, MAX(min_size, SvCUR(invlist) + 1));
9129 PERL_STATIC_INLINE void
9130 S_invlist_clear(pTHX_ SV* invlist) /* Empty the inversion list */
9132 PERL_ARGS_ASSERT_INVLIST_CLEAR;
9134 assert(is_invlist(invlist));
9136 invlist_set_len(invlist, 0, 0);
9137 invlist_trim(invlist);
9140 #endif /* ifndef PERL_IN_XSUB_RE */
9142 PERL_STATIC_INLINE bool
9143 S_invlist_is_iterating(SV* const invlist)
9145 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9147 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9150 #ifndef PERL_IN_XSUB_RE
9152 PERL_STATIC_INLINE UV
9153 S_invlist_max(SV* const invlist)
9155 /* Returns the maximum number of elements storable in the inversion list's
9156 * array, without having to realloc() */
9158 PERL_ARGS_ASSERT_INVLIST_MAX;
9160 assert(is_invlist(invlist));
9162 /* Assumes worst case, in which the 0 element is not counted in the
9163 * inversion list, so subtracts 1 for that */
9164 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
9165 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
9166 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
9170 S_initialize_invlist_guts(pTHX_ SV* invlist, const Size_t initial_size)
9172 PERL_ARGS_ASSERT_INITIALIZE_INVLIST_GUTS;
9174 /* First 1 is in case the zero element isn't in the list; second 1 is for
9176 SvGROW(invlist, TO_INTERNAL_SIZE(initial_size + 1) + 1);
9177 invlist_set_len(invlist, 0, 0);
9179 /* Force iterinit() to be used to get iteration to work */
9180 invlist_iterfinish(invlist);
9182 *get_invlist_previous_index_addr(invlist) = 0;
9183 SvPOK_on(invlist); /* This allows B to extract the PV */
9187 Perl__new_invlist(pTHX_ IV initial_size)
9190 /* Return a pointer to a newly constructed inversion list, with enough
9191 * space to store 'initial_size' elements. If that number is negative, a
9192 * system default is used instead */
9196 if (initial_size < 0) {
9200 new_list = newSV_type(SVt_INVLIST);
9201 initialize_invlist_guts(new_list, initial_size);
9207 Perl__new_invlist_C_array(pTHX_ const UV* const list)
9209 /* Return a pointer to a newly constructed inversion list, initialized to
9210 * point to <list>, which has to be in the exact correct inversion list
9211 * form, including internal fields. Thus this is a dangerous routine that
9212 * should not be used in the wrong hands. The passed in 'list' contains
9213 * several header fields at the beginning that are not part of the
9214 * inversion list body proper */
9216 const STRLEN length = (STRLEN) list[0];
9217 const UV version_id = list[1];
9218 const bool offset = cBOOL(list[2]);
9219 #define HEADER_LENGTH 3
9220 /* If any of the above changes in any way, you must change HEADER_LENGTH
9221 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
9222 * perl -E 'say int(rand 2**31-1)'
9224 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
9225 data structure type, so that one being
9226 passed in can be validated to be an
9227 inversion list of the correct vintage.
9230 SV* invlist = newSV_type(SVt_INVLIST);
9232 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
9234 if (version_id != INVLIST_VERSION_ID) {
9235 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
9238 /* The generated array passed in includes header elements that aren't part
9239 * of the list proper, so start it just after them */
9240 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
9242 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
9243 shouldn't touch it */
9245 *(get_invlist_offset_addr(invlist)) = offset;
9247 /* The 'length' passed to us is the physical number of elements in the
9248 * inversion list. But if there is an offset the logical number is one
9250 invlist_set_len(invlist, length - offset, offset);
9252 invlist_set_previous_index(invlist, 0);
9254 /* Initialize the iteration pointer. */
9255 invlist_iterfinish(invlist);
9257 SvREADONLY_on(invlist);
9264 S__append_range_to_invlist(pTHX_ SV* const invlist,
9265 const UV start, const UV end)
9267 /* Subject to change or removal. Append the range from 'start' to 'end' at
9268 * the end of the inversion list. The range must be above any existing
9272 UV max = invlist_max(invlist);
9273 UV len = _invlist_len(invlist);
9276 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
9278 if (len == 0) { /* Empty lists must be initialized */
9279 offset = start != 0;
9280 array = _invlist_array_init(invlist, ! offset);
9283 /* Here, the existing list is non-empty. The current max entry in the
9284 * list is generally the first value not in the set, except when the
9285 * set extends to the end of permissible values, in which case it is
9286 * the first entry in that final set, and so this call is an attempt to
9287 * append out-of-order */
9289 UV final_element = len - 1;
9290 array = invlist_array(invlist);
9291 if ( array[final_element] > start
9292 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
9294 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",
9295 array[final_element], start,
9296 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
9299 /* Here, it is a legal append. If the new range begins 1 above the end
9300 * of the range below it, it is extending the range below it, so the
9301 * new first value not in the set is one greater than the newly
9302 * extended range. */
9303 offset = *get_invlist_offset_addr(invlist);
9304 if (array[final_element] == start) {
9305 if (end != UV_MAX) {
9306 array[final_element] = end + 1;
9309 /* But if the end is the maximum representable on the machine,
9310 * assume that infinity was actually what was meant. Just let
9311 * the range that this would extend to have no end */
9312 invlist_set_len(invlist, len - 1, offset);
9318 /* Here the new range doesn't extend any existing set. Add it */
9320 len += 2; /* Includes an element each for the start and end of range */
9322 /* If wll overflow the existing space, extend, which may cause the array to
9325 invlist_extend(invlist, len);
9327 /* Have to set len here to avoid assert failure in invlist_array() */
9328 invlist_set_len(invlist, len, offset);
9330 array = invlist_array(invlist);
9333 invlist_set_len(invlist, len, offset);
9336 /* The next item on the list starts the range, the one after that is
9337 * one past the new range. */
9338 array[len - 2] = start;
9339 if (end != UV_MAX) {
9340 array[len - 1] = end + 1;
9343 /* But if the end is the maximum representable on the machine, just let
9344 * the range have no end */
9345 invlist_set_len(invlist, len - 1, offset);
9350 Perl__invlist_search(SV* const invlist, const UV cp)
9352 /* Searches the inversion list for the entry that contains the input code
9353 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
9354 * return value is the index into the list's array of the range that
9355 * contains <cp>, that is, 'i' such that
9356 * array[i] <= cp < array[i+1]
9361 IV high = _invlist_len(invlist);
9362 const IV highest_element = high - 1;
9365 PERL_ARGS_ASSERT__INVLIST_SEARCH;
9367 /* If list is empty, return failure. */
9372 /* (We can't get the array unless we know the list is non-empty) */
9373 array = invlist_array(invlist);
9375 mid = invlist_previous_index(invlist);
9377 if (mid > highest_element) {
9378 mid = highest_element;
9381 /* <mid> contains the cache of the result of the previous call to this
9382 * function (0 the first time). See if this call is for the same result,
9383 * or if it is for mid-1. This is under the theory that calls to this
9384 * function will often be for related code points that are near each other.
9385 * And benchmarks show that caching gives better results. We also test
9386 * here if the code point is within the bounds of the list. These tests
9387 * replace others that would have had to be made anyway to make sure that
9388 * the array bounds were not exceeded, and these give us extra information
9389 * at the same time */
9390 if (cp >= array[mid]) {
9391 if (cp >= array[highest_element]) {
9392 return highest_element;
9395 /* Here, array[mid] <= cp < array[highest_element]. This means that
9396 * the final element is not the answer, so can exclude it; it also
9397 * means that <mid> is not the final element, so can refer to 'mid + 1'
9399 if (cp < array[mid + 1]) {
9405 else { /* cp < aray[mid] */
9406 if (cp < array[0]) { /* Fail if outside the array */
9410 if (cp >= array[mid - 1]) {
9415 /* Binary search. What we are looking for is <i> such that
9416 * array[i] <= cp < array[i+1]
9417 * The loop below converges on the i+1. Note that there may not be an
9418 * (i+1)th element in the array, and things work nonetheless */
9419 while (low < high) {
9420 mid = (low + high) / 2;
9421 assert(mid <= highest_element);
9422 if (array[mid] <= cp) { /* cp >= array[mid] */
9425 /* We could do this extra test to exit the loop early.
9426 if (cp < array[low]) {
9431 else { /* cp < array[mid] */
9438 invlist_set_previous_index(invlist, high);
9443 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9444 const bool complement_b, SV** output)
9446 /* Take the union of two inversion lists and point '*output' to it. On
9447 * input, '*output' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9448 * even 'a' or 'b'). If to an inversion list, the contents of the original
9449 * list will be replaced by the union. The first list, 'a', may be
9450 * NULL, in which case a copy of the second list is placed in '*output'.
9451 * If 'complement_b' is TRUE, the union is taken of the complement
9452 * (inversion) of 'b' instead of b itself.
9454 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9455 * Richard Gillam, published by Addison-Wesley, and explained at some
9456 * length there. The preface says to incorporate its examples into your
9457 * code at your own risk.
9459 * The algorithm is like a merge sort. */
9461 const UV* array_a; /* a's array */
9463 UV len_a; /* length of a's array */
9466 SV* u; /* the resulting union */
9470 UV i_a = 0; /* current index into a's array */
9474 /* running count, as explained in the algorithm source book; items are
9475 * stopped accumulating and are output when the count changes to/from 0.
9476 * The count is incremented when we start a range that's in an input's set,
9477 * and decremented when we start a range that's not in a set. So this
9478 * variable can be 0, 1, or 2. When it is 0 neither input is in their set,
9479 * and hence nothing goes into the union; 1, just one of the inputs is in
9480 * its set (and its current range gets added to the union); and 2 when both
9481 * inputs are in their sets. */
9484 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
9486 assert(*output == NULL || is_invlist(*output));
9488 len_b = _invlist_len(b);
9491 /* Here, 'b' is empty, hence it's complement is all possible code
9492 * points. So if the union includes the complement of 'b', it includes
9493 * everything, and we need not even look at 'a'. It's easiest to
9494 * create a new inversion list that matches everything. */
9496 SV* everything = _add_range_to_invlist(NULL, 0, UV_MAX);
9498 if (*output == NULL) { /* If the output didn't exist, just point it
9500 *output = everything;
9502 else { /* Otherwise, replace its contents with the new list */
9503 invlist_replace_list_destroys_src(*output, everything);
9504 SvREFCNT_dec_NN(everything);
9510 /* Here, we don't want the complement of 'b', and since 'b' is empty,
9511 * the union will come entirely from 'a'. If 'a' is NULL or empty, the
9512 * output will be empty */
9514 if (a == NULL || _invlist_len(a) == 0) {
9515 if (*output == NULL) {
9516 *output = _new_invlist(0);
9519 invlist_clear(*output);
9524 /* Here, 'a' is not empty, but 'b' is, so 'a' entirely determines the
9525 * union. We can just return a copy of 'a' if '*output' doesn't point
9526 * to an existing list */
9527 if (*output == NULL) {
9528 *output = invlist_clone(a, NULL);
9532 /* If the output is to overwrite 'a', we have a no-op, as it's
9538 /* Here, '*output' is to be overwritten by 'a' */
9539 u = invlist_clone(a, NULL);
9540 invlist_replace_list_destroys_src(*output, u);
9546 /* Here 'b' is not empty. See about 'a' */
9548 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
9550 /* Here, 'a' is empty (and b is not). That means the union will come
9551 * entirely from 'b'. If '*output' is NULL, we can directly return a
9552 * clone of 'b'. Otherwise, we replace the contents of '*output' with
9555 SV ** dest = (*output == NULL) ? output : &u;
9556 *dest = invlist_clone(b, NULL);
9558 _invlist_invert(*dest);
9562 invlist_replace_list_destroys_src(*output, u);
9569 /* Here both lists exist and are non-empty */
9570 array_a = invlist_array(a);
9571 array_b = invlist_array(b);
9573 /* If are to take the union of 'a' with the complement of b, set it
9574 * up so are looking at b's complement. */
9577 /* To complement, we invert: if the first element is 0, remove it. To
9578 * do this, we just pretend the array starts one later */
9579 if (array_b[0] == 0) {
9585 /* But if the first element is not zero, we pretend the list starts
9586 * at the 0 that is always stored immediately before the array. */
9592 /* Size the union for the worst case: that the sets are completely
9594 u = _new_invlist(len_a + len_b);
9596 /* Will contain U+0000 if either component does */
9597 array_u = _invlist_array_init(u, ( len_a > 0 && array_a[0] == 0)
9598 || (len_b > 0 && array_b[0] == 0));
9600 /* Go through each input list item by item, stopping when have exhausted
9602 while (i_a < len_a && i_b < len_b) {
9603 UV cp; /* The element to potentially add to the union's array */
9604 bool cp_in_set; /* is it in the the input list's set or not */
9606 /* We need to take one or the other of the two inputs for the union.
9607 * Since we are merging two sorted lists, we take the smaller of the
9608 * next items. In case of a tie, we take first the one that is in its
9609 * set. If we first took the one not in its set, it would decrement
9610 * the count, possibly to 0 which would cause it to be output as ending
9611 * the range, and the next time through we would take the same number,
9612 * and output it again as beginning the next range. By doing it the
9613 * opposite way, there is no possibility that the count will be
9614 * momentarily decremented to 0, and thus the two adjoining ranges will
9615 * be seamlessly merged. (In a tie and both are in the set or both not
9616 * in the set, it doesn't matter which we take first.) */
9617 if ( array_a[i_a] < array_b[i_b]
9618 || ( array_a[i_a] == array_b[i_b]
9619 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9621 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9622 cp = array_a[i_a++];
9625 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9626 cp = array_b[i_b++];
9629 /* Here, have chosen which of the two inputs to look at. Only output
9630 * if the running count changes to/from 0, which marks the
9631 * beginning/end of a range that's in the set */
9634 array_u[i_u++] = cp;
9641 array_u[i_u++] = cp;
9647 /* The loop above increments the index into exactly one of the input lists
9648 * each iteration, and ends when either index gets to its list end. That
9649 * means the other index is lower than its end, and so something is
9650 * remaining in that one. We decrement 'count', as explained below, if
9651 * that list is in its set. (i_a and i_b each currently index the element
9652 * beyond the one we care about.) */
9653 if ( (i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9654 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9659 /* Above we decremented 'count' if the list that had unexamined elements in
9660 * it was in its set. This has made it so that 'count' being non-zero
9661 * means there isn't anything left to output; and 'count' equal to 0 means
9662 * that what is left to output is precisely that which is left in the
9663 * non-exhausted input list.
9665 * To see why, note first that the exhausted input obviously has nothing
9666 * left to add to the union. If it was in its set at its end, that means
9667 * the set extends from here to the platform's infinity, and hence so does
9668 * the union and the non-exhausted set is irrelevant. The exhausted set
9669 * also contributed 1 to 'count'. If 'count' was 2, it got decremented to
9670 * 1, but if it was 1, the non-exhausted set wasn't in its set, and so
9671 * 'count' remains at 1. This is consistent with the decremented 'count'
9672 * != 0 meaning there's nothing left to add to the union.
9674 * But if the exhausted input wasn't in its set, it contributed 0 to
9675 * 'count', and the rest of the union will be whatever the other input is.
9676 * If 'count' was 0, neither list was in its set, and 'count' remains 0;
9677 * otherwise it gets decremented to 0. This is consistent with 'count'
9678 * == 0 meaning the remainder of the union is whatever is left in the
9679 * non-exhausted list. */
9684 IV copy_count = len_a - i_a;
9685 if (copy_count > 0) { /* The non-exhausted input is 'a' */
9686 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
9688 else { /* The non-exhausted input is b */
9689 copy_count = len_b - i_b;
9690 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
9692 len_u = i_u + copy_count;
9695 /* Set the result to the final length, which can change the pointer to
9696 * array_u, so re-find it. (Note that it is unlikely that this will
9697 * change, as we are shrinking the space, not enlarging it) */
9698 if (len_u != _invlist_len(u)) {
9699 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
9701 array_u = invlist_array(u);
9704 if (*output == NULL) { /* Simply return the new inversion list */
9708 /* Otherwise, overwrite the inversion list that was in '*output'. We
9709 * could instead free '*output', and then set it to 'u', but experience
9710 * has shown [perl #127392] that if the input is a mortal, we can get a
9711 * huge build-up of these during regex compilation before they get
9713 invlist_replace_list_destroys_src(*output, u);
9721 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
9722 const bool complement_b, SV** i)
9724 /* Take the intersection of two inversion lists and point '*i' to it. On
9725 * input, '*i' MUST POINT TO NULL OR TO AN SV* INVERSION LIST (possibly
9726 * even 'a' or 'b'). If to an inversion list, the contents of the original
9727 * list will be replaced by the intersection. The first list, 'a', may be
9728 * NULL, in which case '*i' will be an empty list. If 'complement_b' is
9729 * TRUE, the result will be the intersection of 'a' and the complement (or
9730 * inversion) of 'b' instead of 'b' directly.
9732 * The basis for this comes from "Unicode Demystified" Chapter 13 by
9733 * Richard Gillam, published by Addison-Wesley, and explained at some
9734 * length there. The preface says to incorporate its examples into your
9735 * code at your own risk. In fact, it had bugs
9737 * The algorithm is like a merge sort, and is essentially the same as the
9741 const UV* array_a; /* a's array */
9743 UV len_a; /* length of a's array */
9746 SV* r; /* the resulting intersection */
9750 UV i_a = 0; /* current index into a's array */
9754 /* running count of how many of the two inputs are postitioned at ranges
9755 * that are in their sets. As explained in the algorithm source book,
9756 * items are stopped accumulating and are output when the count changes
9757 * to/from 2. The count is incremented when we start a range that's in an
9758 * input's set, and decremented when we start a range that's not in a set.
9759 * Only when it is 2 are we in the intersection. */
9762 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
9764 assert(*i == NULL || is_invlist(*i));
9766 /* Special case if either one is empty */
9767 len_a = (a == NULL) ? 0 : _invlist_len(a);
9768 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
9769 if (len_a != 0 && complement_b) {
9771 /* Here, 'a' is not empty, therefore from the enclosing 'if', 'b'
9772 * must be empty. Here, also we are using 'b's complement, which
9773 * hence must be every possible code point. Thus the intersection
9776 if (*i == a) { /* No-op */
9781 *i = invlist_clone(a, NULL);
9785 r = invlist_clone(a, NULL);
9786 invlist_replace_list_destroys_src(*i, r);
9791 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
9792 * intersection must be empty */
9794 *i = _new_invlist(0);
9802 /* Here both lists exist and are non-empty */
9803 array_a = invlist_array(a);
9804 array_b = invlist_array(b);
9806 /* If are to take the intersection of 'a' with the complement of b, set it
9807 * up so are looking at b's complement. */
9810 /* To complement, we invert: if the first element is 0, remove it. To
9811 * do this, we just pretend the array starts one later */
9812 if (array_b[0] == 0) {
9818 /* But if the first element is not zero, we pretend the list starts
9819 * at the 0 that is always stored immediately before the array. */
9825 /* Size the intersection for the worst case: that the intersection ends up
9826 * fragmenting everything to be completely disjoint */
9827 r= _new_invlist(len_a + len_b);
9829 /* Will contain U+0000 iff both components do */
9830 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
9831 && len_b > 0 && array_b[0] == 0);
9833 /* Go through each list item by item, stopping when have exhausted one of
9835 while (i_a < len_a && i_b < len_b) {
9836 UV cp; /* The element to potentially add to the intersection's
9838 bool cp_in_set; /* Is it in the input list's set or not */
9840 /* We need to take one or the other of the two inputs for the
9841 * intersection. Since we are merging two sorted lists, we take the
9842 * smaller of the next items. In case of a tie, we take first the one
9843 * that is not in its set (a difference from the union algorithm). If
9844 * we first took the one in its set, it would increment the count,
9845 * possibly to 2 which would cause it to be output as starting a range
9846 * in the intersection, and the next time through we would take that
9847 * same number, and output it again as ending the set. By doing the
9848 * opposite of this, there is no possibility that the count will be
9849 * momentarily incremented to 2. (In a tie and both are in the set or
9850 * both not in the set, it doesn't matter which we take first.) */
9851 if ( array_a[i_a] < array_b[i_b]
9852 || ( array_a[i_a] == array_b[i_b]
9853 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
9855 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
9856 cp = array_a[i_a++];
9859 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
9863 /* Here, have chosen which of the two inputs to look at. Only output
9864 * if the running count changes to/from 2, which marks the
9865 * beginning/end of a range that's in the intersection */
9869 array_r[i_r++] = cp;
9874 array_r[i_r++] = cp;
9881 /* The loop above increments the index into exactly one of the input lists
9882 * each iteration, and ends when either index gets to its list end. That
9883 * means the other index is lower than its end, and so something is
9884 * remaining in that one. We increment 'count', as explained below, if the
9885 * exhausted list was in its set. (i_a and i_b each currently index the
9886 * element beyond the one we care about.) */
9887 if ( (i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
9888 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
9893 /* Above we incremented 'count' if the exhausted list was in its set. This
9894 * has made it so that 'count' being below 2 means there is nothing left to
9895 * output; otheriwse what's left to add to the intersection is precisely
9896 * that which is left in the non-exhausted input list.
9898 * To see why, note first that the exhausted input obviously has nothing
9899 * left to affect the intersection. If it was in its set at its end, that
9900 * means the set extends from here to the platform's infinity, and hence
9901 * anything in the non-exhausted's list will be in the intersection, and
9902 * anything not in it won't be. Hence, the rest of the intersection is
9903 * precisely what's in the non-exhausted list The exhausted set also
9904 * contributed 1 to 'count', meaning 'count' was at least 1. Incrementing
9905 * it means 'count' is now at least 2. This is consistent with the
9906 * incremented 'count' being >= 2 means to add the non-exhausted list to
9909 * But if the exhausted input wasn't in its set, it contributed 0 to
9910 * 'count', and the intersection can't include anything further; the
9911 * non-exhausted set is irrelevant. 'count' was at most 1, and doesn't get
9912 * incremented. This is consistent with 'count' being < 2 meaning nothing
9913 * further to add to the intersection. */
9914 if (count < 2) { /* Nothing left to put in the intersection. */
9917 else { /* copy the non-exhausted list, unchanged. */
9918 IV copy_count = len_a - i_a;
9919 if (copy_count > 0) { /* a is the one with stuff left */
9920 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9922 else { /* b is the one with stuff left */
9923 copy_count = len_b - i_b;
9924 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9926 len_r = i_r + copy_count;
9929 /* Set the result to the final length, which can change the pointer to
9930 * array_r, so re-find it. (Note that it is unlikely that this will
9931 * change, as we are shrinking the space, not enlarging it) */
9932 if (len_r != _invlist_len(r)) {
9933 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9935 array_r = invlist_array(r);
9938 if (*i == NULL) { /* Simply return the calculated intersection */
9941 else { /* Otherwise, replace the existing inversion list in '*i'. We could
9942 instead free '*i', and then set it to 'r', but experience has
9943 shown [perl #127392] that if the input is a mortal, we can get a
9944 huge build-up of these during regex compilation before they get
9947 invlist_replace_list_destroys_src(*i, r);
9959 Perl__add_range_to_invlist(pTHX_ SV* invlist, UV start, UV end)
9961 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9962 * set. A pointer to the inversion list is returned. This may actually be
9963 * a new list, in which case the passed in one has been destroyed. The
9964 * passed-in inversion list can be NULL, in which case a new one is created
9965 * with just the one range in it. The new list is not necessarily
9966 * NUL-terminated. Space is not freed if the inversion list shrinks as a
9967 * result of this function. The gain would not be large, and in many
9968 * cases, this is called multiple times on a single inversion list, so
9969 * anything freed may almost immediately be needed again.
9971 * This used to mostly call the 'union' routine, but that is much more
9972 * heavyweight than really needed for a single range addition */
9974 UV* array; /* The array implementing the inversion list */
9975 UV len; /* How many elements in 'array' */
9976 SSize_t i_s; /* index into the invlist array where 'start'
9978 SSize_t i_e = 0; /* And the index where 'end' should go */
9979 UV cur_highest; /* The highest code point in the inversion list
9980 upon entry to this function */
9982 /* This range becomes the whole inversion list if none already existed */
9983 if (invlist == NULL) {
9984 invlist = _new_invlist(2);
9985 _append_range_to_invlist(invlist, start, end);
9989 /* Likewise, if the inversion list is currently empty */
9990 len = _invlist_len(invlist);
9992 _append_range_to_invlist(invlist, start, end);
9996 /* Starting here, we have to know the internals of the list */
9997 array = invlist_array(invlist);
9999 /* If the new range ends higher than the current highest ... */
10000 cur_highest = invlist_highest(invlist);
10001 if (end > cur_highest) {
10003 /* If the whole range is higher, we can just append it */
10004 if (start > cur_highest) {
10005 _append_range_to_invlist(invlist, start, end);
10009 /* Otherwise, add the portion that is higher ... */
10010 _append_range_to_invlist(invlist, cur_highest + 1, end);
10012 /* ... and continue on below to handle the rest. As a result of the
10013 * above append, we know that the index of the end of the range is the
10014 * final even numbered one of the array. Recall that the final element
10015 * always starts a range that extends to infinity. If that range is in
10016 * the set (meaning the set goes from here to infinity), it will be an
10017 * even index, but if it isn't in the set, it's odd, and the final
10018 * range in the set is one less, which is even. */
10019 if (end == UV_MAX) {
10027 /* We have dealt with appending, now see about prepending. If the new
10028 * range starts lower than the current lowest ... */
10029 if (start < array[0]) {
10031 /* Adding something which has 0 in it is somewhat tricky, and uncommon.
10032 * Let the union code handle it, rather than having to know the
10033 * trickiness in two code places. */
10034 if (UNLIKELY(start == 0)) {
10037 range_invlist = _new_invlist(2);
10038 _append_range_to_invlist(range_invlist, start, end);
10040 _invlist_union(invlist, range_invlist, &invlist);
10042 SvREFCNT_dec_NN(range_invlist);
10047 /* If the whole new range comes before the first entry, and doesn't
10048 * extend it, we have to insert it as an additional range */
10049 if (end < array[0] - 1) {
10051 goto splice_in_new_range;
10054 /* Here the new range adjoins the existing first range, extending it
10058 /* And continue on below to handle the rest. We know that the index of
10059 * the beginning of the range is the first one of the array */
10062 else { /* Not prepending any part of the new range to the existing list.
10063 * Find where in the list it should go. This finds i_s, such that:
10064 * invlist[i_s] <= start < array[i_s+1]
10066 i_s = _invlist_search(invlist, start);
10069 /* At this point, any extending before the beginning of the inversion list
10070 * and/or after the end has been done. This has made it so that, in the
10071 * code below, each endpoint of the new range is either in a range that is
10072 * in the set, or is in a gap between two ranges that are. This means we
10073 * don't have to worry about exceeding the array bounds.
10075 * Find where in the list the new range ends (but we can skip this if we
10076 * have already determined what it is, or if it will be the same as i_s,
10077 * which we already have computed) */
10079 i_e = (start == end)
10081 : _invlist_search(invlist, end);
10084 /* Here generally invlist[i_e] <= end < array[i_e+1]. But if invlist[i_e]
10085 * is a range that goes to infinity there is no element at invlist[i_e+1],
10086 * so only the first relation holds. */
10088 if ( ! ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10090 /* Here, the ranges on either side of the beginning of the new range
10091 * are in the set, and this range starts in the gap between them.
10093 * The new range extends the range above it downwards if the new range
10094 * ends at or above that range's start */
10095 const bool extends_the_range_above = ( end == UV_MAX
10096 || end + 1 >= array[i_s+1]);
10098 /* The new range extends the range below it upwards if it begins just
10099 * after where that range ends */
10100 if (start == array[i_s]) {
10102 /* If the new range fills the entire gap between the other ranges,
10103 * they will get merged together. Other ranges may also get
10104 * merged, depending on how many of them the new range spans. In
10105 * the general case, we do the merge later, just once, after we
10106 * figure out how many to merge. But in the case where the new
10107 * range exactly spans just this one gap (possibly extending into
10108 * the one above), we do the merge here, and an early exit. This
10109 * is done here to avoid having to special case later. */
10110 if (i_e - i_s <= 1) {
10112 /* If i_e - i_s == 1, it means that the new range terminates
10113 * within the range above, and hence 'extends_the_range_above'
10114 * must be true. (If the range above it extends to infinity,
10115 * 'i_s+2' will be above the array's limit, but 'len-i_s-2'
10116 * will be 0, so no harm done.) */
10117 if (extends_the_range_above) {
10118 Move(array + i_s + 2, array + i_s, len - i_s - 2, UV);
10119 invlist_set_len(invlist,
10121 *(get_invlist_offset_addr(invlist)));
10125 /* Here, i_e must == i_s. We keep them in sync, as they apply
10126 * to the same range, and below we are about to decrement i_s
10131 /* Here, the new range is adjacent to the one below. (It may also
10132 * span beyond the range above, but that will get resolved later.)
10133 * Extend the range below to include this one. */
10134 array[i_s] = (end == UV_MAX) ? UV_MAX : end + 1;
10136 start = array[i_s];
10138 else if (extends_the_range_above) {
10140 /* Here the new range only extends the range above it, but not the
10141 * one below. It merges with the one above. Again, we keep i_e
10142 * and i_s in sync if they point to the same range */
10147 array[i_s] = start;
10151 /* Here, we've dealt with the new range start extending any adjoining
10154 * If the new range extends to infinity, it is now the final one,
10155 * regardless of what was there before */
10156 if (UNLIKELY(end == UV_MAX)) {
10157 invlist_set_len(invlist, i_s + 1, *(get_invlist_offset_addr(invlist)));
10161 /* If i_e started as == i_s, it has also been dealt with,
10162 * and been updated to the new i_s, which will fail the following if */
10163 if (! ELEMENT_RANGE_MATCHES_INVLIST(i_e)) {
10165 /* Here, the ranges on either side of the end of the new range are in
10166 * the set, and this range ends in the gap between them.
10168 * If this range is adjacent to (hence extends) the range above it, it
10169 * becomes part of that range; likewise if it extends the range below,
10170 * it becomes part of that range */
10171 if (end + 1 == array[i_e+1]) {
10173 array[i_e] = start;
10175 else if (start <= array[i_e]) {
10176 array[i_e] = end + 1;
10183 /* If the range fits entirely in an existing range (as possibly already
10184 * extended above), it doesn't add anything new */
10185 if (ELEMENT_RANGE_MATCHES_INVLIST(i_s)) {
10189 /* Here, no part of the range is in the list. Must add it. It will
10190 * occupy 2 more slots */
10191 splice_in_new_range:
10193 invlist_extend(invlist, len + 2);
10194 array = invlist_array(invlist);
10195 /* Move the rest of the array down two slots. Don't include any
10197 Move(array + i_e + 1, array + i_e + 3, len - i_e - 1, UV);
10199 /* Do the actual splice */
10200 array[i_e+1] = start;
10201 array[i_e+2] = end + 1;
10202 invlist_set_len(invlist, len + 2, *(get_invlist_offset_addr(invlist)));
10206 /* Here the new range crossed the boundaries of a pre-existing range. The
10207 * code above has adjusted things so that both ends are in ranges that are
10208 * in the set. This means everything in between must also be in the set.
10209 * Just squash things together */
10210 Move(array + i_e + 1, array + i_s + 1, len - i_e - 1, UV);
10211 invlist_set_len(invlist,
10213 *(get_invlist_offset_addr(invlist)));
10219 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
10220 UV** other_elements_ptr)
10222 /* Create and return an inversion list whose contents are to be populated
10223 * by the caller. The caller gives the number of elements (in 'size') and
10224 * the very first element ('element0'). This function will set
10225 * '*other_elements_ptr' to an array of UVs, where the remaining elements
10226 * are to be placed.
10228 * Obviously there is some trust involved that the caller will properly
10229 * fill in the other elements of the array.
10231 * (The first element needs to be passed in, as the underlying code does
10232 * things differently depending on whether it is zero or non-zero) */
10234 SV* invlist = _new_invlist(size);
10237 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
10239 invlist = add_cp_to_invlist(invlist, element0);
10240 offset = *get_invlist_offset_addr(invlist);
10242 invlist_set_len(invlist, size, offset);
10243 *other_elements_ptr = invlist_array(invlist) + 1;
10249 #ifndef PERL_IN_XSUB_RE
10251 Perl__invlist_invert(pTHX_ SV* const invlist)
10253 /* Complement the input inversion list. This adds a 0 if the list didn't
10254 * have a zero; removes it otherwise. As described above, the data
10255 * structure is set up so that this is very efficient */
10257 PERL_ARGS_ASSERT__INVLIST_INVERT;
10259 assert(! invlist_is_iterating(invlist));
10261 /* The inverse of matching nothing is matching everything */
10262 if (_invlist_len(invlist) == 0) {
10263 _append_range_to_invlist(invlist, 0, UV_MAX);
10267 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
10271 Perl_invlist_clone(pTHX_ SV* const invlist, SV* new_invlist)
10273 /* Return a new inversion list that is a copy of the input one, which is
10274 * unchanged. The new list will not be mortal even if the old one was. */
10276 const STRLEN nominal_length = _invlist_len(invlist);
10277 const STRLEN physical_length = SvCUR(invlist);
10278 const bool offset = *(get_invlist_offset_addr(invlist));
10280 PERL_ARGS_ASSERT_INVLIST_CLONE;
10282 if (new_invlist == NULL) {
10283 new_invlist = _new_invlist(nominal_length);
10286 sv_upgrade(new_invlist, SVt_INVLIST);
10287 initialize_invlist_guts(new_invlist, nominal_length);
10290 *(get_invlist_offset_addr(new_invlist)) = offset;
10291 invlist_set_len(new_invlist, nominal_length, offset);
10292 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
10294 return new_invlist;
10299 PERL_STATIC_INLINE UV
10300 S_invlist_lowest(SV* const invlist)
10302 /* Returns the lowest code point that matches an inversion list. This API
10303 * has an ambiguity, as it returns 0 under either the lowest is actually
10304 * 0, or if the list is empty. If this distinction matters to you, check
10305 * for emptiness before calling this function */
10307 UV len = _invlist_len(invlist);
10310 PERL_ARGS_ASSERT_INVLIST_LOWEST;
10316 array = invlist_array(invlist);
10322 S_invlist_contents(pTHX_ SV* const invlist, const bool traditional_style)
10324 /* Get the contents of an inversion list into a string SV so that they can
10325 * be printed out. If 'traditional_style' is TRUE, it uses the format
10326 * traditionally done for debug tracing; otherwise it uses a format
10327 * suitable for just copying to the output, with blanks between ranges and
10328 * a dash between range components */
10332 const char intra_range_delimiter = (traditional_style ? '\t' : '-');
10333 const char inter_range_delimiter = (traditional_style ? '\n' : ' ');
10335 if (traditional_style) {
10336 output = newSVpvs("\n");
10339 output = newSVpvs("");
10342 PERL_ARGS_ASSERT_INVLIST_CONTENTS;
10344 assert(! invlist_is_iterating(invlist));
10346 invlist_iterinit(invlist);
10347 while (invlist_iternext(invlist, &start, &end)) {
10348 if (end == UV_MAX) {
10349 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%cINFTY%c",
10350 start, intra_range_delimiter,
10351 inter_range_delimiter);
10353 else if (end != start) {
10354 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c%04" UVXf "%c",
10356 intra_range_delimiter,
10357 end, inter_range_delimiter);
10360 Perl_sv_catpvf(aTHX_ output, "%04" UVXf "%c",
10361 start, inter_range_delimiter);
10365 if (SvCUR(output) && ! traditional_style) {/* Get rid of trailing blank */
10366 SvCUR_set(output, SvCUR(output) - 1);
10372 #ifndef PERL_IN_XSUB_RE
10374 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
10375 const char * const indent, SV* const invlist)
10377 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
10378 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
10379 * the string 'indent'. The output looks like this:
10380 [0] 0x000A .. 0x000D
10382 [4] 0x2028 .. 0x2029
10383 [6] 0x3104 .. INFTY
10384 * This means that the first range of code points matched by the list are
10385 * 0xA through 0xD; the second range contains only the single code point
10386 * 0x85, etc. An inversion list is an array of UVs. Two array elements
10387 * are used to define each range (except if the final range extends to
10388 * infinity, only a single element is needed). The array index of the
10389 * first element for the corresponding range is given in brackets. */
10394 PERL_ARGS_ASSERT__INVLIST_DUMP;
10396 if (invlist_is_iterating(invlist)) {
10397 Perl_dump_indent(aTHX_ level, file,
10398 "%sCan't dump inversion list because is in middle of iterating\n",
10403 invlist_iterinit(invlist);
10404 while (invlist_iternext(invlist, &start, &end)) {
10405 if (end == UV_MAX) {
10406 Perl_dump_indent(aTHX_ level, file,
10407 "%s[%" UVuf "] 0x%04" UVXf " .. INFTY\n",
10408 indent, (UV)count, start);
10410 else if (end != start) {
10411 Perl_dump_indent(aTHX_ level, file,
10412 "%s[%" UVuf "] 0x%04" UVXf " .. 0x%04" UVXf "\n",
10413 indent, (UV)count, start, end);
10416 Perl_dump_indent(aTHX_ level, file, "%s[%" UVuf "] 0x%04" UVXf "\n",
10417 indent, (UV)count, start);
10425 #if defined(PERL_ARGS_ASSERT__INVLISTEQ) && !defined(PERL_IN_XSUB_RE)
10427 Perl__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
10429 /* Return a boolean as to if the two passed in inversion lists are
10430 * identical. The final argument, if TRUE, says to take the complement of
10431 * the second inversion list before doing the comparison */
10433 const UV len_a = _invlist_len(a);
10434 UV len_b = _invlist_len(b);
10436 const UV* array_a = NULL;
10437 const UV* array_b = NULL;
10439 PERL_ARGS_ASSERT__INVLISTEQ;
10441 /* This code avoids accessing the arrays unless it knows the length is
10446 return ! complement_b;
10450 array_a = invlist_array(a);
10454 array_b = invlist_array(b);
10457 /* If are to compare 'a' with the complement of b, set it
10458 * up so are looking at b's complement. */
10459 if (complement_b) {
10461 /* The complement of nothing is everything, so <a> would have to have
10462 * just one element, starting at zero (ending at infinity) */
10464 return (len_a == 1 && array_a[0] == 0);
10466 if (array_b[0] == 0) {
10468 /* Otherwise, to complement, we invert. Here, the first element is
10469 * 0, just remove it. To do this, we just pretend the array starts
10477 /* But if the first element is not zero, we pretend the list starts
10478 * at the 0 that is always stored immediately before the array. */
10484 return len_a == len_b
10485 && memEQ(array_a, array_b, len_a * sizeof(array_a[0]));
10491 * As best we can, determine the characters that can match the start of
10492 * the given EXACTF-ish node. This is for use in creating ssc nodes, so there
10493 * can be false positive matches
10495 * Returns the invlist as a new SV*; it is the caller's responsibility to
10496 * call SvREFCNT_dec() when done with it.
10499 S_make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
10502 const U8 * s = (U8*)STRING(node);
10503 SSize_t bytelen = STR_LEN(node);
10505 /* Start out big enough for 2 separate code points */
10506 SV* invlist = _new_invlist(4);
10508 PERL_ARGS_ASSERT_MAKE_EXACTF_INVLIST;
10513 /* We punt and assume can match anything if the node begins
10514 * with a multi-character fold. Things are complicated. For
10515 * example, /ffi/i could match any of:
10516 * "\N{LATIN SMALL LIGATURE FFI}"
10517 * "\N{LATIN SMALL LIGATURE FF}I"
10518 * "F\N{LATIN SMALL LIGATURE FI}"
10519 * plus several other things; and making sure we have all the
10520 * possibilities is hard. */
10521 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
10522 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10525 /* Any Latin1 range character can potentially match any
10526 * other depending on the locale, and in Turkic locales, U+130 and
10528 if (OP(node) == EXACTFL) {
10529 _invlist_union(invlist, PL_Latin1, &invlist);
10530 invlist = add_cp_to_invlist(invlist,
10531 LATIN_SMALL_LETTER_DOTLESS_I);
10532 invlist = add_cp_to_invlist(invlist,
10533 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10536 /* But otherwise, it matches at least itself. We can
10537 * quickly tell if it has a distinct fold, and if so,
10538 * it matches that as well */
10539 invlist = add_cp_to_invlist(invlist, uc);
10540 if (IS_IN_SOME_FOLD_L1(uc))
10541 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
10544 /* Some characters match above-Latin1 ones under /i. This
10545 * is true of EXACTFL ones when the locale is UTF-8 */
10546 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
10547 && (! isASCII(uc) || (OP(node) != EXACTFAA
10548 && OP(node) != EXACTFAA_NO_TRIE)))
10550 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
10554 else { /* Pattern is UTF-8 */
10555 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
10556 const U8* e = s + bytelen;
10559 fc = uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
10561 /* The only code points that aren't folded in a UTF EXACTFish
10562 * node are are the problematic ones in EXACTFL nodes */
10563 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
10564 /* We need to check for the possibility that this EXACTFL
10565 * node begins with a multi-char fold. Therefore we fold
10566 * the first few characters of it so that we can make that
10572 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
10574 *(d++) = (U8) toFOLD(*s);
10575 if (fc < 0) { /* Save the first fold */
10582 UV fold = toFOLD_utf8_safe(s, e, d, &len);
10583 if (fc < 0) { /* Save the first fold */
10591 /* And set up so the code below that looks in this folded
10592 * buffer instead of the node's string */
10597 /* When we reach here 's' points to the fold of the first
10598 * character(s) of the node; and 'e' points to far enough along
10599 * the folded string to be just past any possible multi-char
10602 * Unlike the non-UTF-8 case, the macro for determining if a
10603 * string is a multi-char fold requires all the characters to
10604 * already be folded. This is because of all the complications
10605 * if not. Note that they are folded anyway, except in EXACTFL
10606 * nodes. Like the non-UTF case above, we punt if the node
10607 * begins with a multi-char fold */
10609 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
10610 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
10612 else { /* Single char fold */
10614 unsigned int first_fold;
10615 const unsigned int * remaining_folds;
10616 Size_t folds_count;
10618 /* It matches itself */
10619 invlist = add_cp_to_invlist(invlist, fc);
10621 /* ... plus all the things that fold to it, which are found in
10622 * PL_utf8_foldclosures */
10623 folds_count = _inverse_folds(fc, &first_fold,
10625 for (k = 0; k < folds_count; k++) {
10626 UV c = (k == 0) ? first_fold : remaining_folds[k-1];
10628 /* /aa doesn't allow folds between ASCII and non- */
10629 if ( (OP(node) == EXACTFAA || OP(node) == EXACTFAA_NO_TRIE)
10630 && isASCII(c) != isASCII(fc))
10635 invlist = add_cp_to_invlist(invlist, c);
10638 if (OP(node) == EXACTFL) {
10640 /* If either [iI] are present in an EXACTFL node the above code
10641 * should have added its normal case pair, but under a Turkish
10642 * locale they could match instead the case pairs from it. Add
10643 * those as potential matches as well */
10644 if (isALPHA_FOLD_EQ(fc, 'I')) {
10645 invlist = add_cp_to_invlist(invlist,
10646 LATIN_SMALL_LETTER_DOTLESS_I);
10647 invlist = add_cp_to_invlist(invlist,
10648 LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
10650 else if (fc == LATIN_SMALL_LETTER_DOTLESS_I) {
10651 invlist = add_cp_to_invlist(invlist, 'I');
10653 else if (fc == LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE) {
10654 invlist = add_cp_to_invlist(invlist, 'i');
10663 #undef HEADER_LENGTH
10664 #undef TO_INTERNAL_SIZE
10665 #undef FROM_INTERNAL_SIZE
10666 #undef INVLIST_VERSION_ID
10668 /* End of inversion list object */
10671 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
10673 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
10674 * constructs, and updates RExC_flags with them. On input, RExC_parse
10675 * should point to the first flag; it is updated on output to point to the
10676 * final ')' or ':'. There needs to be at least one flag, or this will
10679 /* for (?g), (?gc), and (?o) warnings; warning
10680 about (?c) will warn about (?g) -- japhy */
10682 #define WASTED_O 0x01
10683 #define WASTED_G 0x02
10684 #define WASTED_C 0x04
10685 #define WASTED_GC (WASTED_G|WASTED_C)
10686 I32 wastedflags = 0x00;
10687 U32 posflags = 0, negflags = 0;
10688 U32 *flagsp = &posflags;
10689 char has_charset_modifier = '\0';
10691 bool has_use_defaults = FALSE;
10692 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
10693 int x_mod_count = 0;
10695 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
10697 /* '^' as an initial flag sets certain defaults */
10698 if (UCHARAT(RExC_parse) == '^') {
10700 has_use_defaults = TRUE;
10701 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
10702 cs = (RExC_uni_semantics)
10703 ? REGEX_UNICODE_CHARSET
10704 : REGEX_DEPENDS_CHARSET;
10705 set_regex_charset(&RExC_flags, cs);
10708 cs = get_regex_charset(RExC_flags);
10709 if ( cs == REGEX_DEPENDS_CHARSET
10710 && RExC_uni_semantics)
10712 cs = REGEX_UNICODE_CHARSET;
10716 while (RExC_parse < RExC_end) {
10717 /* && strchr("iogcmsx", *RExC_parse) */
10718 /* (?g), (?gc) and (?o) are useless here
10719 and must be globally applied -- japhy */
10720 switch (*RExC_parse) {
10722 /* Code for the imsxn flags */
10723 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
10725 case LOCALE_PAT_MOD:
10726 if (has_charset_modifier) {
10727 goto excess_modifier;
10729 else if (flagsp == &negflags) {
10732 cs = REGEX_LOCALE_CHARSET;
10733 has_charset_modifier = LOCALE_PAT_MOD;
10735 case UNICODE_PAT_MOD:
10736 if (has_charset_modifier) {
10737 goto excess_modifier;
10739 else if (flagsp == &negflags) {
10742 cs = REGEX_UNICODE_CHARSET;
10743 has_charset_modifier = UNICODE_PAT_MOD;
10745 case ASCII_RESTRICT_PAT_MOD:
10746 if (flagsp == &negflags) {
10749 if (has_charset_modifier) {
10750 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
10751 goto excess_modifier;
10753 /* Doubled modifier implies more restricted */
10754 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
10757 cs = REGEX_ASCII_RESTRICTED_CHARSET;
10759 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
10761 case DEPENDS_PAT_MOD:
10762 if (has_use_defaults) {
10763 goto fail_modifiers;
10765 else if (flagsp == &negflags) {
10768 else if (has_charset_modifier) {
10769 goto excess_modifier;
10772 /* The dual charset means unicode semantics if the
10773 * pattern (or target, not known until runtime) are
10774 * utf8, or something in the pattern indicates unicode
10776 cs = (RExC_uni_semantics)
10777 ? REGEX_UNICODE_CHARSET
10778 : REGEX_DEPENDS_CHARSET;
10779 has_charset_modifier = DEPENDS_PAT_MOD;
10783 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
10784 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
10786 else if (has_charset_modifier == *(RExC_parse - 1)) {
10787 vFAIL2("Regexp modifier \"%c\" may not appear twice",
10788 *(RExC_parse - 1));
10791 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
10793 NOT_REACHED; /*NOTREACHED*/
10796 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
10797 *(RExC_parse - 1));
10798 NOT_REACHED; /*NOTREACHED*/
10799 case ONCE_PAT_MOD: /* 'o' */
10800 case GLOBAL_PAT_MOD: /* 'g' */
10801 if (ckWARN(WARN_REGEXP)) {
10802 const I32 wflagbit = *RExC_parse == 'o'
10805 if (! (wastedflags & wflagbit) ) {
10806 wastedflags |= wflagbit;
10807 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10810 "Useless (%s%c) - %suse /%c modifier",
10811 flagsp == &negflags ? "?-" : "?",
10813 flagsp == &negflags ? "don't " : "",
10820 case CONTINUE_PAT_MOD: /* 'c' */
10821 if (ckWARN(WARN_REGEXP)) {
10822 if (! (wastedflags & WASTED_C) ) {
10823 wastedflags |= WASTED_GC;
10824 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
10827 "Useless (%sc) - %suse /gc modifier",
10828 flagsp == &negflags ? "?-" : "?",
10829 flagsp == &negflags ? "don't " : ""
10834 case KEEPCOPY_PAT_MOD: /* 'p' */
10835 if (flagsp == &negflags) {
10836 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
10838 *flagsp |= RXf_PMf_KEEPCOPY;
10842 /* A flag is a default iff it is following a minus, so
10843 * if there is a minus, it means will be trying to
10844 * re-specify a default which is an error */
10845 if (has_use_defaults || flagsp == &negflags) {
10846 goto fail_modifiers;
10848 flagsp = &negflags;
10849 wastedflags = 0; /* reset so (?g-c) warns twice */
10855 if ((posflags & (RXf_PMf_EXTENDED|RXf_PMf_EXTENDED_MORE)) == RXf_PMf_EXTENDED) {
10856 negflags |= RXf_PMf_EXTENDED_MORE;
10858 RExC_flags |= posflags;
10860 if (negflags & RXf_PMf_EXTENDED) {
10861 negflags |= RXf_PMf_EXTENDED_MORE;
10863 RExC_flags &= ~negflags;
10864 set_regex_charset(&RExC_flags, cs);
10869 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
10870 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10871 vFAIL2utf8f("Sequence (%" UTF8f "...) not recognized",
10872 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10873 NOT_REACHED; /*NOTREACHED*/
10876 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10879 vFAIL("Sequence (?... not terminated");
10883 - reg - regular expression, i.e. main body or parenthesized thing
10885 * Caller must absorb opening parenthesis.
10887 * Combining parenthesis handling with the base level of regular expression
10888 * is a trifle forced, but the need to tie the tails of the branches to what
10889 * follows makes it hard to avoid.
10891 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
10893 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
10895 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
10898 PERL_STATIC_INLINE regnode_offset
10899 S_handle_named_backref(pTHX_ RExC_state_t *pRExC_state,
10901 char * parse_start,
10905 regnode_offset ret;
10906 char* name_start = RExC_parse;
10908 SV *sv_dat = reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
10909 GET_RE_DEBUG_FLAGS_DECL;
10911 PERL_ARGS_ASSERT_HANDLE_NAMED_BACKREF;
10913 if (RExC_parse == name_start || *RExC_parse != ch) {
10914 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
10915 vFAIL2("Sequence %.3s... not terminated", parse_start);
10919 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10920 RExC_rxi->data->data[num]=(void*)sv_dat;
10921 SvREFCNT_inc_simple_void_NN(sv_dat);
10924 ret = reganode(pRExC_state,
10927 : (ASCII_FOLD_RESTRICTED)
10929 : (AT_LEAST_UNI_SEMANTICS)
10935 *flagp |= HASWIDTH;
10937 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
10938 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
10940 nextchar(pRExC_state);
10944 /* On success, returns the offset at which any next node should be placed into
10945 * the regex engine program being compiled.
10947 * Returns 0 otherwise, with *flagp set to indicate why:
10948 * TRYAGAIN at the end of (?) that only sets flags.
10949 * RESTART_PARSE if the parse needs to be restarted, or'd with
10950 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
10951 * Otherwise would only return 0 if regbranch() returns 0, which cannot
10953 STATIC regnode_offset
10954 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp, U32 depth)
10955 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
10956 * 2 is like 1, but indicates that nextchar() has been called to advance
10957 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
10958 * this flag alerts us to the need to check for that */
10960 regnode_offset ret = 0; /* Will be the head of the group. */
10962 regnode_offset lastbr;
10963 regnode_offset ender = 0;
10966 U32 oregflags = RExC_flags;
10967 bool have_branch = 0;
10969 I32 freeze_paren = 0;
10970 I32 after_freeze = 0;
10971 I32 num; /* numeric backreferences */
10972 SV * max_open; /* Max number of unclosed parens */
10974 char * parse_start = RExC_parse; /* MJD */
10975 char * const oregcomp_parse = RExC_parse;
10977 GET_RE_DEBUG_FLAGS_DECL;
10979 PERL_ARGS_ASSERT_REG;
10980 DEBUG_PARSE("reg ");
10982 max_open = get_sv(RE_COMPILE_RECURSION_LIMIT, GV_ADD);
10984 if (!SvIOK(max_open)) {
10985 sv_setiv(max_open, RE_COMPILE_RECURSION_INIT);
10987 if (depth > 4 * (UV) SvIV(max_open)) { /* We increase depth by 4 for each
10989 vFAIL("Too many nested open parens");
10992 *flagp = 0; /* Tentatively. */
10994 if (RExC_in_lookbehind) {
10995 RExC_in_lookbehind++;
10997 if (RExC_in_lookahead) {
10998 RExC_in_lookahead++;
11001 /* Having this true makes it feasible to have a lot fewer tests for the
11002 * parse pointer being in scope. For example, we can write
11003 * while(isFOO(*RExC_parse)) RExC_parse++;
11005 * while(RExC_parse < RExC_end && isFOO(*RExC_parse)) RExC_parse++;
11007 assert(*RExC_end == '\0');
11009 /* Make an OPEN node, if parenthesized. */
11012 /* Under /x, space and comments can be gobbled up between the '(' and
11013 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
11014 * intervening space, as the sequence is a token, and a token should be
11016 bool has_intervening_patws = (paren == 2)
11017 && *(RExC_parse - 1) != '(';
11019 if (RExC_parse >= RExC_end) {
11020 vFAIL("Unmatched (");
11023 if (paren == 'r') { /* Atomic script run */
11027 else if ( *RExC_parse == '*') { /* (*VERB:ARG), (*construct:...) */
11028 char *start_verb = RExC_parse + 1;
11030 char *start_arg = NULL;
11031 unsigned char op = 0;
11032 int arg_required = 0;
11033 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
11034 bool has_upper = FALSE;
11036 if (has_intervening_patws) {
11037 RExC_parse++; /* past the '*' */
11039 /* For strict backwards compatibility, don't change the message
11040 * now that we also have lowercase operands */
11041 if (isUPPER(*RExC_parse)) {
11042 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
11045 vFAIL("In '(*...)', the '(' and '*' must be adjacent");
11048 while (RExC_parse < RExC_end && *RExC_parse != ')' ) {
11049 if ( *RExC_parse == ':' ) {
11050 start_arg = RExC_parse + 1;
11054 if (isUPPER(*RExC_parse)) {
11060 RExC_parse += UTF8SKIP(RExC_parse);
11063 verb_len = RExC_parse - start_verb;
11065 if (RExC_parse >= RExC_end) {
11066 goto unterminated_verb_pattern;
11069 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11070 while ( RExC_parse < RExC_end && *RExC_parse != ')' ) {
11071 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11073 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11074 unterminated_verb_pattern:
11076 vFAIL("Unterminated verb pattern argument");
11079 vFAIL("Unterminated '(*...' argument");
11083 if ( RExC_parse >= RExC_end || *RExC_parse != ')' ) {
11085 vFAIL("Unterminated verb pattern");
11088 vFAIL("Unterminated '(*...' construct");
11093 /* Here, we know that RExC_parse < RExC_end */
11095 switch ( *start_verb ) {
11096 case 'A': /* (*ACCEPT) */
11097 if ( memEQs(start_verb, verb_len,"ACCEPT") ) {
11099 internal_argval = RExC_nestroot;
11102 case 'C': /* (*COMMIT) */
11103 if ( memEQs(start_verb, verb_len,"COMMIT") )
11106 case 'F': /* (*FAIL) */
11107 if ( verb_len==1 || memEQs(start_verb, verb_len,"FAIL") ) {
11111 case ':': /* (*:NAME) */
11112 case 'M': /* (*MARK:NAME) */
11113 if ( verb_len==0 || memEQs(start_verb, verb_len,"MARK") ) {
11118 case 'P': /* (*PRUNE) */
11119 if ( memEQs(start_verb, verb_len,"PRUNE") )
11122 case 'S': /* (*SKIP) */
11123 if ( memEQs(start_verb, verb_len,"SKIP") )
11126 case 'T': /* (*THEN) */
11127 /* [19:06] <TimToady> :: is then */
11128 if ( memEQs(start_verb, verb_len,"THEN") ) {
11130 RExC_seen |= REG_CUTGROUP_SEEN;
11134 if ( memEQs(start_verb, verb_len, "asr")
11135 || memEQs(start_verb, verb_len, "atomic_script_run"))
11137 paren = 'r'; /* Mnemonic: recursed run */
11140 else if (memEQs(start_verb, verb_len, "atomic")) {
11141 paren = 't'; /* AtOMIC */
11142 goto alpha_assertions;
11146 if ( memEQs(start_verb, verb_len, "plb")
11147 || memEQs(start_verb, verb_len, "positive_lookbehind"))
11150 goto lookbehind_alpha_assertions;
11152 else if ( memEQs(start_verb, verb_len, "pla")
11153 || memEQs(start_verb, verb_len, "positive_lookahead"))
11156 goto alpha_assertions;
11160 if ( memEQs(start_verb, verb_len, "nlb")
11161 || memEQs(start_verb, verb_len, "negative_lookbehind"))
11164 goto lookbehind_alpha_assertions;
11166 else if ( memEQs(start_verb, verb_len, "nla")
11167 || memEQs(start_verb, verb_len, "negative_lookahead"))
11170 goto alpha_assertions;
11174 if ( memEQs(start_verb, verb_len, "sr")
11175 || memEQs(start_verb, verb_len, "script_run"))
11177 regnode_offset atomic;
11183 /* This indicates Unicode rules. */
11184 REQUIRE_UNI_RULES(flagp, 0);
11190 RExC_parse = start_arg;
11192 if (RExC_in_script_run) {
11194 /* Nested script runs are treated as no-ops, because
11195 * if the nested one fails, the outer one must as
11196 * well. It could fail sooner, and avoid (??{} with
11197 * side effects, but that is explicitly documented as
11198 * undefined behavior. */
11202 if (paren == 's') {
11207 /* But, the atomic part of a nested atomic script run
11208 * isn't a no-op, but can be treated just like a '(?>'
11214 if (paren == 's') {
11215 /* Here, we're starting a new regular script run */
11216 ret = reg_node(pRExC_state, SROPEN);
11217 RExC_in_script_run = 1;
11222 /* Here, we are starting an atomic script run. This is
11223 * handled by recursing to deal with the atomic portion
11224 * separately, enclosed in SROPEN ... SRCLOSE nodes */
11226 ret = reg_node(pRExC_state, SROPEN);
11228 RExC_in_script_run = 1;
11230 atomic = reg(pRExC_state, 'r', &flags, depth);
11231 if (flags & (RESTART_PARSE|NEED_UTF8)) {
11232 *flagp = flags & (RESTART_PARSE|NEED_UTF8);
11236 if (! REGTAIL(pRExC_state, ret, atomic)) {
11237 REQUIRE_BRANCHJ(flagp, 0);
11240 if (! REGTAIL(pRExC_state, atomic, reg_node(pRExC_state,
11243 REQUIRE_BRANCHJ(flagp, 0);
11246 RExC_in_script_run = 0;
11252 lookbehind_alpha_assertions:
11253 RExC_seen |= REG_LOOKBEHIND_SEEN;
11254 RExC_in_lookbehind++;
11259 RExC_seen_zerolen++;
11265 /* An empty negative lookahead assertion simply is failure */
11266 if (paren == 'A' && RExC_parse == start_arg) {
11267 ret=reganode(pRExC_state, OPFAIL, 0);
11268 nextchar(pRExC_state);
11272 RExC_parse = start_arg;
11277 "'(*%" UTF8f "' requires a terminating ':'",
11278 UTF8fARG(UTF, verb_len, start_verb));
11279 NOT_REACHED; /*NOTREACHED*/
11281 } /* End of switch */
11284 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11286 if (has_upper || verb_len == 0) {
11288 "Unknown verb pattern '%" UTF8f "'",
11289 UTF8fARG(UTF, verb_len, start_verb));
11293 "Unknown '(*...)' construct '%" UTF8f "'",
11294 UTF8fARG(UTF, verb_len, start_verb));
11297 if ( RExC_parse == start_arg ) {
11300 if ( arg_required && !start_arg ) {
11301 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
11302 verb_len, start_verb);
11304 if (internal_argval == -1) {
11305 ret = reganode(pRExC_state, op, 0);
11307 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
11309 RExC_seen |= REG_VERBARG_SEEN;
11311 SV *sv = newSVpvn( start_arg,
11312 RExC_parse - start_arg);
11313 ARG(REGNODE_p(ret)) = add_data( pRExC_state,
11314 STR_WITH_LEN("S"));
11315 RExC_rxi->data->data[ARG(REGNODE_p(ret))]=(void*)sv;
11316 FLAGS(REGNODE_p(ret)) = 1;
11318 FLAGS(REGNODE_p(ret)) = 0;
11320 if ( internal_argval != -1 )
11321 ARG2L_SET(REGNODE_p(ret), internal_argval);
11322 nextchar(pRExC_state);
11325 else if (*RExC_parse == '?') { /* (?...) */
11326 bool is_logical = 0;
11327 const char * const seqstart = RExC_parse;
11328 const char * endptr;
11329 if (has_intervening_patws) {
11331 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
11334 RExC_parse++; /* past the '?' */
11335 paren = *RExC_parse; /* might be a trailing NUL, if not
11337 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
11338 if (RExC_parse > RExC_end) {
11341 ret = 0; /* For look-ahead/behind. */
11344 case 'P': /* (?P...) variants for those used to PCRE/Python */
11345 paren = *RExC_parse;
11346 if ( paren == '<') { /* (?P<...>) named capture */
11348 if (RExC_parse >= RExC_end) {
11349 vFAIL("Sequence (?P<... not terminated");
11351 goto named_capture;
11353 else if (paren == '>') { /* (?P>name) named recursion */
11355 if (RExC_parse >= RExC_end) {
11356 vFAIL("Sequence (?P>... not terminated");
11358 goto named_recursion;
11360 else if (paren == '=') { /* (?P=...) named backref */
11362 return handle_named_backref(pRExC_state, flagp,
11365 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11366 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11367 vFAIL3("Sequence (%.*s...) not recognized",
11368 RExC_parse-seqstart, seqstart);
11369 NOT_REACHED; /*NOTREACHED*/
11370 case '<': /* (?<...) */
11371 if (*RExC_parse == '!')
11373 else if (*RExC_parse != '=')
11380 case '\'': /* (?'...') */
11381 name_start = RExC_parse;
11382 svname = reg_scan_name(pRExC_state, REG_RSN_RETURN_NAME);
11383 if ( RExC_parse == name_start
11384 || RExC_parse >= RExC_end
11385 || *RExC_parse != paren)
11387 vFAIL2("Sequence (?%c... not terminated",
11388 paren=='>' ? '<' : paren);
11393 if (!svname) /* shouldn't happen */
11395 "panic: reg_scan_name returned NULL");
11396 if (!RExC_paren_names) {
11397 RExC_paren_names= newHV();
11398 sv_2mortal(MUTABLE_SV(RExC_paren_names));
11400 RExC_paren_name_list= newAV();
11401 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
11404 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
11406 sv_dat = HeVAL(he_str);
11408 /* croak baby croak */
11410 "panic: paren_name hash element allocation failed");
11411 } else if ( SvPOK(sv_dat) ) {
11412 /* (?|...) can mean we have dupes so scan to check
11413 its already been stored. Maybe a flag indicating
11414 we are inside such a construct would be useful,
11415 but the arrays are likely to be quite small, so
11416 for now we punt -- dmq */
11417 IV count = SvIV(sv_dat);
11418 I32 *pv = (I32*)SvPVX(sv_dat);
11420 for ( i = 0 ; i < count ; i++ ) {
11421 if ( pv[i] == RExC_npar ) {
11427 pv = (I32*)SvGROW(sv_dat,
11428 SvCUR(sv_dat) + sizeof(I32)+1);
11429 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
11430 pv[count] = RExC_npar;
11431 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
11434 (void)SvUPGRADE(sv_dat, SVt_PVNV);
11435 sv_setpvn(sv_dat, (char *)&(RExC_npar),
11438 SvIV_set(sv_dat, 1);
11441 /* Yes this does cause a memory leak in debugging Perls
11443 if (!av_store(RExC_paren_name_list,
11444 RExC_npar, SvREFCNT_inc_NN(svname)))
11445 SvREFCNT_dec_NN(svname);
11448 /*sv_dump(sv_dat);*/
11450 nextchar(pRExC_state);
11452 goto capturing_parens;
11455 RExC_seen |= REG_LOOKBEHIND_SEEN;
11456 RExC_in_lookbehind++;
11458 if (RExC_parse >= RExC_end) {
11459 vFAIL("Sequence (?... not terminated");
11461 RExC_seen_zerolen++;
11463 case '=': /* (?=...) */
11464 RExC_seen_zerolen++;
11465 RExC_in_lookahead++;
11467 case '!': /* (?!...) */
11468 RExC_seen_zerolen++;
11469 /* check if we're really just a "FAIL" assertion */
11470 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11471 FALSE /* Don't force to /x */ );
11472 if (*RExC_parse == ')') {
11473 ret=reganode(pRExC_state, OPFAIL, 0);
11474 nextchar(pRExC_state);
11478 case '|': /* (?|...) */
11479 /* branch reset, behave like a (?:...) except that
11480 buffers in alternations share the same numbers */
11482 after_freeze = freeze_paren = RExC_npar;
11484 /* XXX This construct currently requires an extra pass.
11485 * Investigation would be required to see if that could be
11487 REQUIRE_PARENS_PASS;
11489 case ':': /* (?:...) */
11490 case '>': /* (?>...) */
11492 case '$': /* (?$...) */
11493 case '@': /* (?@...) */
11494 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
11496 case '0' : /* (?0) */
11497 case 'R' : /* (?R) */
11498 if (RExC_parse == RExC_end || *RExC_parse != ')')
11499 FAIL("Sequence (?R) not terminated");
11501 RExC_seen |= REG_RECURSE_SEEN;
11503 /* XXX These constructs currently require an extra pass.
11504 * It probably could be changed */
11505 REQUIRE_PARENS_PASS;
11507 *flagp |= POSTPONED;
11508 goto gen_recurse_regop;
11510 /* named and numeric backreferences */
11511 case '&': /* (?&NAME) */
11512 parse_start = RExC_parse - 1;
11515 SV *sv_dat = reg_scan_name(pRExC_state,
11516 REG_RSN_RETURN_DATA);
11517 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
11519 if (RExC_parse >= RExC_end || *RExC_parse != ')')
11520 vFAIL("Sequence (?&... not terminated");
11521 goto gen_recurse_regop;
11524 if (! inRANGE(RExC_parse[0], '1', '9')) {
11526 vFAIL("Illegal pattern");
11528 goto parse_recursion;
11530 case '-': /* (?-1) */
11531 if (! inRANGE(RExC_parse[0], '1', '9')) {
11532 RExC_parse--; /* rewind to let it be handled later */
11536 case '1': case '2': case '3': case '4': /* (?1) */
11537 case '5': case '6': case '7': case '8': case '9':
11538 RExC_parse = (char *) seqstart + 1; /* Point to the digit */
11541 bool is_neg = FALSE;
11543 parse_start = RExC_parse - 1; /* MJD */
11544 if (*RExC_parse == '-') {
11549 if (grok_atoUV(RExC_parse, &unum, &endptr)
11553 RExC_parse = (char*)endptr;
11557 /* Some limit for num? */
11561 if (*RExC_parse!=')')
11562 vFAIL("Expecting close bracket");
11565 if ( paren == '-' ) {
11567 Diagram of capture buffer numbering.
11568 Top line is the normal capture buffer numbers
11569 Bottom line is the negative indexing as from
11573 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
11577 num = RExC_npar + num;
11580 /* It might be a forward reference; we can't fail until
11581 * we know, by completing the parse to get all the
11582 * groups, and then reparsing */
11583 if (ALL_PARENS_COUNTED) {
11585 vFAIL("Reference to nonexistent group");
11588 REQUIRE_PARENS_PASS;
11591 } else if ( paren == '+' ) {
11592 num = RExC_npar + num - 1;
11594 /* We keep track how many GOSUB items we have produced.
11595 To start off the ARG2L() of the GOSUB holds its "id",
11596 which is used later in conjunction with RExC_recurse
11597 to calculate the offset we need to jump for the GOSUB,
11598 which it will store in the final representation.
11599 We have to defer the actual calculation until much later
11600 as the regop may move.
11603 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
11604 if (num >= RExC_npar) {
11606 /* It might be a forward reference; we can't fail until we
11607 * know, by completing the parse to get all the groups, and
11608 * then reparsing */
11609 if (ALL_PARENS_COUNTED) {
11610 if (num >= RExC_total_parens) {
11612 vFAIL("Reference to nonexistent group");
11616 REQUIRE_PARENS_PASS;
11619 RExC_recurse_count++;
11620 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
11621 "%*s%*s Recurse #%" UVuf " to %" IVdf "\n",
11622 22, "| |", (int)(depth * 2 + 1), "",
11623 (UV)ARG(REGNODE_p(ret)),
11624 (IV)ARG2L(REGNODE_p(ret))));
11625 RExC_seen |= REG_RECURSE_SEEN;
11627 Set_Node_Length(REGNODE_p(ret),
11628 1 + regarglen[OP(REGNODE_p(ret))]); /* MJD */
11629 Set_Node_Offset(REGNODE_p(ret), parse_start); /* MJD */
11631 *flagp |= POSTPONED;
11632 assert(*RExC_parse == ')');
11633 nextchar(pRExC_state);
11638 case '?': /* (??...) */
11640 if (*RExC_parse != '{') {
11641 RExC_parse += SKIP_IF_CHAR(RExC_parse, RExC_end);
11642 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
11644 "Sequence (%" UTF8f "...) not recognized",
11645 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
11646 NOT_REACHED; /*NOTREACHED*/
11648 *flagp |= POSTPONED;
11652 case '{': /* (?{...}) */
11655 struct reg_code_block *cb;
11658 RExC_seen_zerolen++;
11660 if ( !pRExC_state->code_blocks
11661 || pRExC_state->code_index
11662 >= pRExC_state->code_blocks->count
11663 || pRExC_state->code_blocks->cb[pRExC_state->code_index].start
11664 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
11667 if (RExC_pm_flags & PMf_USE_RE_EVAL)
11668 FAIL("panic: Sequence (?{...}): no code block found\n");
11669 FAIL("Eval-group not allowed at runtime, use re 'eval'");
11671 /* this is a pre-compiled code block (?{...}) */
11672 cb = &pRExC_state->code_blocks->cb[pRExC_state->code_index];
11673 RExC_parse = RExC_start + cb->end;
11675 if (cb->src_regex) {
11676 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
11677 RExC_rxi->data->data[n] =
11678 (void*)SvREFCNT_inc((SV*)cb->src_regex);
11679 RExC_rxi->data->data[n+1] = (void*)o;
11682 n = add_data(pRExC_state,
11683 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
11684 RExC_rxi->data->data[n] = (void*)o;
11686 pRExC_state->code_index++;
11687 nextchar(pRExC_state);
11690 regnode_offset eval;
11691 ret = reg_node(pRExC_state, LOGICAL);
11693 eval = reg2Lanode(pRExC_state, EVAL,
11696 /* for later propagation into (??{})
11698 RExC_flags & RXf_PMf_COMPILETIME
11700 FLAGS(REGNODE_p(ret)) = 2;
11701 if (! REGTAIL(pRExC_state, ret, eval)) {
11702 REQUIRE_BRANCHJ(flagp, 0);
11704 /* deal with the length of this later - MJD */
11707 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
11708 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1);
11709 Set_Node_Offset(REGNODE_p(ret), parse_start);
11712 case '(': /* (?(?{...})...) and (?(?=...)...) */
11715 const int DEFINE_len = sizeof("DEFINE") - 1;
11716 if ( RExC_parse < RExC_end - 1
11717 && ( ( RExC_parse[0] == '?' /* (?(?...)) */
11718 && ( RExC_parse[1] == '='
11719 || RExC_parse[1] == '!'
11720 || RExC_parse[1] == '<'
11721 || RExC_parse[1] == '{'))
11722 || ( RExC_parse[0] == '*' /* (?(*...)) */
11723 && ( memBEGINs(RExC_parse + 1,
11724 (Size_t) (RExC_end - (RExC_parse + 1)),
11726 || memBEGINs(RExC_parse + 1,
11727 (Size_t) (RExC_end - (RExC_parse + 1)),
11729 || memBEGINs(RExC_parse + 1,
11730 (Size_t) (RExC_end - (RExC_parse + 1)),
11732 || memBEGINs(RExC_parse + 1,
11733 (Size_t) (RExC_end - (RExC_parse + 1)),
11735 || memBEGINs(RExC_parse + 1,
11736 (Size_t) (RExC_end - (RExC_parse + 1)),
11737 "positive_lookahead:")
11738 || memBEGINs(RExC_parse + 1,
11739 (Size_t) (RExC_end - (RExC_parse + 1)),
11740 "positive_lookbehind:")
11741 || memBEGINs(RExC_parse + 1,
11742 (Size_t) (RExC_end - (RExC_parse + 1)),
11743 "negative_lookahead:")
11744 || memBEGINs(RExC_parse + 1,
11745 (Size_t) (RExC_end - (RExC_parse + 1)),
11746 "negative_lookbehind:"))))
11747 ) { /* Lookahead or eval. */
11749 regnode_offset tail;
11751 ret = reg_node(pRExC_state, LOGICAL);
11752 FLAGS(REGNODE_p(ret)) = 1;
11754 tail = reg(pRExC_state, 1, &flag, depth+1);
11755 RETURN_FAIL_ON_RESTART(flag, flagp);
11756 if (! REGTAIL(pRExC_state, ret, tail)) {
11757 REQUIRE_BRANCHJ(flagp, 0);
11761 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
11762 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
11764 char ch = RExC_parse[0] == '<' ? '>' : '\'';
11765 char *name_start= RExC_parse++;
11767 SV *sv_dat=reg_scan_name(pRExC_state, REG_RSN_RETURN_DATA);
11768 if ( RExC_parse == name_start
11769 || RExC_parse >= RExC_end
11770 || *RExC_parse != ch)
11772 vFAIL2("Sequence (?(%c... not terminated",
11773 (ch == '>' ? '<' : ch));
11777 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11778 RExC_rxi->data->data[num]=(void*)sv_dat;
11779 SvREFCNT_inc_simple_void_NN(sv_dat);
11781 ret = reganode(pRExC_state, GROUPPN, num);
11782 goto insert_if_check_paren;
11784 else if (memBEGINs(RExC_parse,
11785 (STRLEN) (RExC_end - RExC_parse),
11788 ret = reganode(pRExC_state, DEFINEP, 0);
11789 RExC_parse += DEFINE_len;
11791 goto insert_if_check_paren;
11793 else if (RExC_parse[0] == 'R') {
11795 /* parno == 0 => /(?(R)YES|NO)/ "in any form of recursion OR eval"
11796 * parno == 1 => /(?(R0)YES|NO)/ "in GOSUB (?0) / (?R)"
11797 * parno == 2 => /(?(R1)YES|NO)/ "in GOSUB (?1) (parno-1)"
11800 if (RExC_parse[0] == '0') {
11804 else if (inRANGE(RExC_parse[0], '1', '9')) {
11807 if (grok_atoUV(RExC_parse, &uv, &endptr)
11810 parno = (I32)uv + 1;
11811 RExC_parse = (char*)endptr;
11813 /* else "Switch condition not recognized" below */
11814 } else if (RExC_parse[0] == '&') {
11817 sv_dat = reg_scan_name(pRExC_state,
11818 REG_RSN_RETURN_DATA);
11820 parno = 1 + *((I32 *)SvPVX(sv_dat));
11822 ret = reganode(pRExC_state, INSUBP, parno);
11823 goto insert_if_check_paren;
11825 else if (inRANGE(RExC_parse[0], '1', '9')) {
11830 if (grok_atoUV(RExC_parse, &uv, &endptr)
11834 RExC_parse = (char*)endptr;
11837 vFAIL("panic: grok_atoUV returned FALSE");
11839 ret = reganode(pRExC_state, GROUPP, parno);
11841 insert_if_check_paren:
11842 if (UCHARAT(RExC_parse) != ')') {
11844 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11846 vFAIL("Switch condition not recognized");
11848 nextchar(pRExC_state);
11850 if (! REGTAIL(pRExC_state, ret, reganode(pRExC_state,
11853 REQUIRE_BRANCHJ(flagp, 0);
11855 br = regbranch(pRExC_state, &flags, 1, depth+1);
11857 RETURN_FAIL_ON_RESTART(flags,flagp);
11858 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11861 if (! REGTAIL(pRExC_state, br, reganode(pRExC_state,
11864 REQUIRE_BRANCHJ(flagp, 0);
11866 c = UCHARAT(RExC_parse);
11867 nextchar(pRExC_state);
11868 if (flags&HASWIDTH)
11869 *flagp |= HASWIDTH;
11872 vFAIL("(?(DEFINE)....) does not allow branches");
11874 /* Fake one for optimizer. */
11875 lastbr = reganode(pRExC_state, IFTHEN, 0);
11877 if (!regbranch(pRExC_state, &flags, 1, depth+1)) {
11878 RETURN_FAIL_ON_RESTART(flags, flagp);
11879 FAIL2("panic: regbranch returned failure, flags=%#" UVxf,
11882 if (! REGTAIL(pRExC_state, ret, lastbr)) {
11883 REQUIRE_BRANCHJ(flagp, 0);
11885 if (flags&HASWIDTH)
11886 *flagp |= HASWIDTH;
11887 c = UCHARAT(RExC_parse);
11888 nextchar(pRExC_state);
11893 if (RExC_parse >= RExC_end)
11894 vFAIL("Switch (?(condition)... not terminated");
11896 vFAIL("Switch (?(condition)... contains too many branches");
11898 ender = reg_node(pRExC_state, TAIL);
11899 if (! REGTAIL(pRExC_state, br, ender)) {
11900 REQUIRE_BRANCHJ(flagp, 0);
11903 if (! REGTAIL(pRExC_state, lastbr, ender)) {
11904 REQUIRE_BRANCHJ(flagp, 0);
11906 if (! REGTAIL(pRExC_state,
11909 NEXTOPER(REGNODE_p(lastbr)))),
11912 REQUIRE_BRANCHJ(flagp, 0);
11916 if (! REGTAIL(pRExC_state, ret, ender)) {
11917 REQUIRE_BRANCHJ(flagp, 0);
11919 #if 0 /* Removing this doesn't cause failures in the test suite -- khw */
11920 RExC_size++; /* XXX WHY do we need this?!!
11921 For large programs it seems to be required
11922 but I can't figure out why. -- dmq*/
11927 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
11929 vFAIL("Unknown switch condition (?(...))");
11931 case '[': /* (?[ ... ]) */
11932 return handle_regex_sets(pRExC_state, NULL, flagp, depth+1,
11934 case 0: /* A NUL */
11935 RExC_parse--; /* for vFAIL to print correctly */
11936 vFAIL("Sequence (? incomplete");
11940 if (RExC_strict) { /* [perl #132851] */
11941 ckWARNreg(RExC_parse, "Empty (?) without any modifiers");
11944 default: /* e.g., (?i) */
11945 RExC_parse = (char *) seqstart + 1;
11947 parse_lparen_question_flags(pRExC_state);
11948 if (UCHARAT(RExC_parse) != ':') {
11949 if (RExC_parse < RExC_end)
11950 nextchar(pRExC_state);
11955 nextchar(pRExC_state);
11960 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
11964 if (! ALL_PARENS_COUNTED) {
11965 /* If we are in our first pass through (and maybe only pass),
11966 * we need to allocate memory for the capturing parentheses
11970 if (!RExC_parens_buf_size) {
11971 /* first guess at number of parens we might encounter */
11972 RExC_parens_buf_size = 10;
11974 /* setup RExC_open_parens, which holds the address of each
11975 * OPEN tag, and to make things simpler for the 0 index the
11976 * start of the program - this is used later for offsets */
11977 Newxz(RExC_open_parens, RExC_parens_buf_size,
11979 RExC_open_parens[0] = 1; /* +1 for REG_MAGIC */
11981 /* setup RExC_close_parens, which holds the address of each
11982 * CLOSE tag, and to make things simpler for the 0 index
11983 * the end of the program - this is used later for offsets
11985 Newxz(RExC_close_parens, RExC_parens_buf_size,
11987 /* we dont know where end op starts yet, so we dont need to
11988 * set RExC_close_parens[0] like we do RExC_open_parens[0]
11991 else if (RExC_npar > RExC_parens_buf_size) {
11992 I32 old_size = RExC_parens_buf_size;
11994 RExC_parens_buf_size *= 2;
11996 Renew(RExC_open_parens, RExC_parens_buf_size,
11998 Zero(RExC_open_parens + old_size,
11999 RExC_parens_buf_size - old_size, regnode_offset);
12001 Renew(RExC_close_parens, RExC_parens_buf_size,
12003 Zero(RExC_close_parens + old_size,
12004 RExC_parens_buf_size - old_size, regnode_offset);
12008 ret = reganode(pRExC_state, OPEN, parno);
12009 if (!RExC_nestroot)
12010 RExC_nestroot = parno;
12011 if (RExC_open_parens && !RExC_open_parens[parno])
12013 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12014 "%*s%*s Setting open paren #%" IVdf " to %d\n",
12015 22, "| |", (int)(depth * 2 + 1), "",
12017 RExC_open_parens[parno]= ret;
12020 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
12021 Set_Node_Offset(REGNODE_p(ret), RExC_parse); /* MJD */
12024 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
12033 /* Pick up the branches, linking them together. */
12034 parse_start = RExC_parse; /* MJD */
12035 br = regbranch(pRExC_state, &flags, 1, depth+1);
12037 /* branch_len = (paren != 0); */
12040 RETURN_FAIL_ON_RESTART(flags, flagp);
12041 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12043 if (*RExC_parse == '|') {
12044 if (RExC_use_BRANCHJ) {
12045 reginsert(pRExC_state, BRANCHJ, br, depth+1);
12048 reginsert(pRExC_state, BRANCH, br, depth+1);
12049 Set_Node_Length(REGNODE_p(br), paren != 0);
12050 Set_Node_Offset_To_R(br, parse_start-RExC_start);
12054 else if (paren == ':') {
12055 *flagp |= flags&SIMPLE;
12057 if (is_open) { /* Starts with OPEN. */
12058 if (! REGTAIL(pRExC_state, ret, br)) { /* OPEN -> first. */
12059 REQUIRE_BRANCHJ(flagp, 0);
12062 else if (paren != '?') /* Not Conditional */
12064 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12066 while (*RExC_parse == '|') {
12067 if (RExC_use_BRANCHJ) {
12070 ender = reganode(pRExC_state, LONGJMP, 0);
12072 /* Append to the previous. */
12073 shut_gcc_up = REGTAIL(pRExC_state,
12074 REGNODE_OFFSET(NEXTOPER(NEXTOPER(REGNODE_p(lastbr)))),
12076 PERL_UNUSED_VAR(shut_gcc_up);
12078 nextchar(pRExC_state);
12079 if (freeze_paren) {
12080 if (RExC_npar > after_freeze)
12081 after_freeze = RExC_npar;
12082 RExC_npar = freeze_paren;
12084 br = regbranch(pRExC_state, &flags, 0, depth+1);
12087 RETURN_FAIL_ON_RESTART(flags, flagp);
12088 FAIL2("panic: regbranch returned failure, flags=%#" UVxf, (UV) flags);
12090 if (! REGTAIL(pRExC_state, lastbr, br)) { /* BRANCH -> BRANCH. */
12091 REQUIRE_BRANCHJ(flagp, 0);
12094 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
12097 if (have_branch || paren != ':') {
12100 /* Make a closing node, and hook it on the end. */
12103 ender = reg_node(pRExC_state, TAIL);
12106 ender = reganode(pRExC_state, CLOSE, parno);
12107 if ( RExC_close_parens ) {
12108 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12109 "%*s%*s Setting close paren #%" IVdf " to %d\n",
12110 22, "| |", (int)(depth * 2 + 1), "",
12111 (IV)parno, ender));
12112 RExC_close_parens[parno]= ender;
12113 if (RExC_nestroot == parno)
12116 Set_Node_Offset(REGNODE_p(ender), RExC_parse+1); /* MJD */
12117 Set_Node_Length(REGNODE_p(ender), 1); /* MJD */
12120 ender = reg_node(pRExC_state, SRCLOSE);
12121 RExC_in_script_run = 0;
12131 *flagp &= ~HASWIDTH;
12133 case 't': /* aTomic */
12135 ender = reg_node(pRExC_state, SUCCEED);
12138 ender = reg_node(pRExC_state, END);
12139 assert(!RExC_end_op); /* there can only be one! */
12140 RExC_end_op = REGNODE_p(ender);
12141 if (RExC_close_parens) {
12142 DEBUG_OPTIMISE_MORE_r(Perl_re_printf( aTHX_
12143 "%*s%*s Setting close paren #0 (END) to %d\n",
12144 22, "| |", (int)(depth * 2 + 1), "",
12147 RExC_close_parens[0]= ender;
12152 DEBUG_PARSE_MSG("lsbr");
12153 regprop(RExC_rx, RExC_mysv1, REGNODE_p(lastbr), NULL, pRExC_state);
12154 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender), NULL, pRExC_state);
12155 Perl_re_printf( aTHX_ "~ tying lastbr %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12156 SvPV_nolen_const(RExC_mysv1),
12158 SvPV_nolen_const(RExC_mysv2),
12160 (IV)(ender - lastbr)
12163 if (! REGTAIL(pRExC_state, lastbr, ender)) {
12164 REQUIRE_BRANCHJ(flagp, 0);
12168 char is_nothing= 1;
12170 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
12172 /* Hook the tails of the branches to the closing node. */
12173 for (br = REGNODE_p(ret); br; br = regnext(br)) {
12174 const U8 op = PL_regkind[OP(br)];
12175 if (op == BRANCH) {
12176 if (! REGTAIL_STUDY(pRExC_state,
12177 REGNODE_OFFSET(NEXTOPER(br)),
12180 REQUIRE_BRANCHJ(flagp, 0);
12182 if ( OP(NEXTOPER(br)) != NOTHING
12183 || regnext(NEXTOPER(br)) != REGNODE_p(ender))
12186 else if (op == BRANCHJ) {
12187 bool shut_gcc_up = REGTAIL_STUDY(pRExC_state,
12188 REGNODE_OFFSET(NEXTOPER(NEXTOPER(br))),
12190 PERL_UNUSED_VAR(shut_gcc_up);
12191 /* for now we always disable this optimisation * /
12192 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
12193 || regnext(NEXTOPER(NEXTOPER(br))) != REGNODE_p(ender))
12199 regnode * ret_as_regnode = REGNODE_p(ret);
12200 br= PL_regkind[OP(ret_as_regnode)] != BRANCH
12201 ? regnext(ret_as_regnode)
12204 DEBUG_PARSE_MSG("NADA");
12205 regprop(RExC_rx, RExC_mysv1, ret_as_regnode,
12206 NULL, pRExC_state);
12207 regprop(RExC_rx, RExC_mysv2, REGNODE_p(ender),
12208 NULL, pRExC_state);
12209 Perl_re_printf( aTHX_ "~ converting ret %s (%" IVdf ") to ender %s (%" IVdf ") offset %" IVdf "\n",
12210 SvPV_nolen_const(RExC_mysv1),
12211 (IV)REG_NODE_NUM(ret_as_regnode),
12212 SvPV_nolen_const(RExC_mysv2),
12218 if (OP(REGNODE_p(ender)) == TAIL) {
12220 RExC_emit= REGNODE_OFFSET(br) + 1;
12223 for ( opt= br + 1; opt < REGNODE_p(ender) ; opt++ )
12224 OP(opt)= OPTIMIZED;
12225 NEXT_OFF(br)= REGNODE_p(ender) - br;
12233 /* Even/odd or x=don't care: 010101x10x */
12234 static const char parens[] = "=!aA<,>Bbt";
12235 /* flag below is set to 0 up through 'A'; 1 for larger */
12237 if (paren && (p = strchr(parens, paren))) {
12238 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
12239 int flag = (p - parens) > 3;
12241 if (paren == '>' || paren == 't') {
12242 node = SUSPEND, flag = 0;
12245 reginsert(pRExC_state, node, ret, depth+1);
12246 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12247 Set_Node_Offset(REGNODE_p(ret), parse_start + 1);
12248 FLAGS(REGNODE_p(ret)) = flag;
12249 if (! REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL)))
12251 REQUIRE_BRANCHJ(flagp, 0);
12256 /* Check for proper termination. */
12258 /* restore original flags, but keep (?p) and, if we've encountered
12259 * something in the parse that changes /d rules into /u, keep the /u */
12260 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
12261 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
12262 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
12264 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
12265 RExC_parse = oregcomp_parse;
12266 vFAIL("Unmatched (");
12268 nextchar(pRExC_state);
12270 else if (!paren && RExC_parse < RExC_end) {
12271 if (*RExC_parse == ')') {
12273 vFAIL("Unmatched )");
12276 FAIL("Junk on end of regexp"); /* "Can't happen". */
12277 NOT_REACHED; /* NOTREACHED */
12280 if (RExC_in_lookbehind) {
12281 RExC_in_lookbehind--;
12283 if (RExC_in_lookahead) {
12284 RExC_in_lookahead--;
12286 if (after_freeze > RExC_npar)
12287 RExC_npar = after_freeze;
12292 - regbranch - one alternative of an | operator
12294 * Implements the concatenation operator.
12296 * On success, returns the offset at which any next node should be placed into
12297 * the regex engine program being compiled.
12299 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
12300 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
12303 STATIC regnode_offset
12304 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
12306 regnode_offset ret;
12307 regnode_offset chain = 0;
12308 regnode_offset latest;
12309 I32 flags = 0, c = 0;
12310 GET_RE_DEBUG_FLAGS_DECL;
12312 PERL_ARGS_ASSERT_REGBRANCH;
12314 DEBUG_PARSE("brnc");
12319 if (RExC_use_BRANCHJ)
12320 ret = reganode(pRExC_state, BRANCHJ, 0);
12322 ret = reg_node(pRExC_state, BRANCH);
12323 Set_Node_Length(REGNODE_p(ret), 1);
12327 *flagp = WORST; /* Tentatively. */
12329 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12330 FALSE /* Don't force to /x */ );
12331 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
12332 flags &= ~TRYAGAIN;
12333 latest = regpiece(pRExC_state, &flags, depth+1);
12335 if (flags & TRYAGAIN)
12337 RETURN_FAIL_ON_RESTART(flags, flagp);
12338 FAIL2("panic: regpiece returned failure, flags=%#" UVxf, (UV) flags);
12342 *flagp |= flags&(HASWIDTH|POSTPONED);
12343 if (chain == 0) /* First piece. */
12344 *flagp |= flags&SPSTART;
12346 /* FIXME adding one for every branch after the first is probably
12347 * excessive now we have TRIE support. (hv) */
12349 if (! REGTAIL(pRExC_state, chain, latest)) {
12350 /* XXX We could just redo this branch, but figuring out what
12351 * bookkeeping needs to be reset is a pain, and it's likely
12352 * that other branches that goto END will also be too large */
12353 REQUIRE_BRANCHJ(flagp, 0);
12359 if (chain == 0) { /* Loop ran zero times. */
12360 chain = reg_node(pRExC_state, NOTHING);
12365 *flagp |= flags&SIMPLE;
12372 - regpiece - something followed by possible quantifier * + ? {n,m}
12374 * Note that the branching code sequences used for ? and the general cases
12375 * of * and + are somewhat optimized: they use the same NOTHING node as
12376 * both the endmarker for their branch list and the body of the last branch.
12377 * It might seem that this node could be dispensed with entirely, but the
12378 * endmarker role is not redundant.
12380 * On success, returns the offset at which any next node should be placed into
12381 * the regex engine program being compiled.
12383 * Returns 0 otherwise, with *flagp set to indicate why:
12384 * TRYAGAIN if regatom() returns 0 with TRYAGAIN.
12385 * RESTART_PARSE if the parse needs to be restarted, or'd with
12386 * NEED_UTF8 if the pattern needs to be upgraded to UTF-8.
12388 STATIC regnode_offset
12389 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
12391 regnode_offset ret;
12395 const char * const origparse = RExC_parse;
12397 I32 max = REG_INFTY;
12398 #ifdef RE_TRACK_PATTERN_OFFSETS
12401 const char *maxpos = NULL;
12404 /* Save the original in case we change the emitted regop to a FAIL. */
12405 const regnode_offset orig_emit = RExC_emit;
12407 GET_RE_DEBUG_FLAGS_DECL;
12409 PERL_ARGS_ASSERT_REGPIECE;
12411 DEBUG_PARSE("piec");
12413 ret = regatom(pRExC_state, &flags, depth+1);
12415 RETURN_FAIL_ON_RESTART_OR_FLAGS(flags, flagp, TRYAGAIN);
12416 FAIL2("panic: regatom returned failure, flags=%#" UVxf, (UV) flags);
12421 if (op == '{' && regcurly(RExC_parse)) {
12423 #ifdef RE_TRACK_PATTERN_OFFSETS
12424 parse_start = RExC_parse; /* MJD */
12426 next = RExC_parse + 1;
12427 while (isDIGIT(*next) || *next == ',') {
12428 if (*next == ',') {
12436 if (*next == '}') { /* got one */
12437 const char* endptr;
12441 if (isDIGIT(*RExC_parse)) {
12443 if (!grok_atoUV(RExC_parse, &uv, &endptr))
12444 vFAIL("Invalid quantifier in {,}");
12445 if (uv >= REG_INFTY)
12446 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12451 if (*maxpos == ',')
12454 maxpos = RExC_parse;
12455 if (isDIGIT(*maxpos)) {
12457 if (!grok_atoUV(maxpos, &uv, &endptr))
12458 vFAIL("Invalid quantifier in {,}");
12459 if (uv >= REG_INFTY)
12460 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
12463 max = REG_INFTY; /* meaning "infinity" */
12466 nextchar(pRExC_state);
12467 if (max < min) { /* If can't match, warn and optimize to fail
12469 reginsert(pRExC_state, OPFAIL, orig_emit, depth+1);
12470 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
12471 NEXT_OFF(REGNODE_p(orig_emit)) =
12472 regarglen[OPFAIL] + NODE_STEP_REGNODE;
12475 else if (min == max && *RExC_parse == '?')
12477 ckWARN2reg(RExC_parse + 1,
12478 "Useless use of greediness modifier '%c'",
12483 if ((flags&SIMPLE)) {
12484 if (min == 0 && max == REG_INFTY) {
12485 reginsert(pRExC_state, STAR, ret, depth+1);
12487 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12490 if (min == 1 && max == REG_INFTY) {
12491 reginsert(pRExC_state, PLUS, ret, depth+1);
12493 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12496 MARK_NAUGHTY_EXP(2, 2);
12497 reginsert(pRExC_state, CURLY, ret, depth+1);
12498 Set_Node_Offset(REGNODE_p(ret), parse_start+1); /* MJD */
12499 Set_Node_Cur_Length(REGNODE_p(ret), parse_start);
12502 const regnode_offset w = reg_node(pRExC_state, WHILEM);
12504 FLAGS(REGNODE_p(w)) = 0;
12505 if (! REGTAIL(pRExC_state, ret, w)) {
12506 REQUIRE_BRANCHJ(flagp, 0);
12508 if (RExC_use_BRANCHJ) {
12509 reginsert(pRExC_state, LONGJMP, ret, depth+1);
12510 reginsert(pRExC_state, NOTHING, ret, depth+1);
12511 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over LONGJMP. */
12513 reginsert(pRExC_state, CURLYX, ret, depth+1);
12515 Set_Node_Offset(REGNODE_p(ret), parse_start+1);
12516 Set_Node_Length(REGNODE_p(ret),
12517 op == '{' ? (RExC_parse - parse_start) : 1);
12519 if (RExC_use_BRANCHJ)
12520 NEXT_OFF(REGNODE_p(ret)) = 3; /* Go over NOTHING to
12522 if (! REGTAIL(pRExC_state, ret, reg_node(pRExC_state,
12525 REQUIRE_BRANCHJ(flagp, 0);
12527 RExC_whilem_seen++;
12528 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
12530 FLAGS(REGNODE_p(ret)) = 0;
12535 *flagp |= HASWIDTH;
12536 ARG1_SET(REGNODE_p(ret), (U16)min);
12537 ARG2_SET(REGNODE_p(ret), (U16)max);
12538 if (max == REG_INFTY)
12539 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
12545 if (!ISMULT1(op)) {
12550 #if 0 /* Now runtime fix should be reliable. */
12552 /* if this is reinstated, don't forget to put this back into perldiag:
12554 =item Regexp *+ operand could be empty at {#} in regex m/%s/
12556 (F) The part of the regexp subject to either the * or + quantifier
12557 could match an empty string. The {#} shows in the regular
12558 expression about where the problem was discovered.
12562 if (!(flags&HASWIDTH) && op != '?')
12563 vFAIL("Regexp *+ operand could be empty");
12566 #ifdef RE_TRACK_PATTERN_OFFSETS
12567 parse_start = RExC_parse;
12569 nextchar(pRExC_state);
12571 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
12577 else if (op == '+') {
12581 else if (op == '?') {
12586 if (!(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
12587 ckWARN2reg(RExC_parse,
12588 "%" UTF8f " matches null string many times",
12589 UTF8fARG(UTF, (RExC_parse >= origparse
12590 ? RExC_parse - origparse
12595 if (*RExC_parse == '?') {
12596 nextchar(pRExC_state);
12597 reginsert(pRExC_state, MINMOD, ret, depth+1);
12598 if (! REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE)) {
12599 REQUIRE_BRANCHJ(flagp, 0);
12602 else if (*RExC_parse == '+') {
12603 regnode_offset ender;
12604 nextchar(pRExC_state);
12605 ender = reg_node(pRExC_state, SUCCEED);
12606 if (! REGTAIL(pRExC_state, ret, ender)) {
12607 REQUIRE_BRANCHJ(flagp, 0);
12609 reginsert(pRExC_state, SUSPEND, ret, depth+1);
12610 ender = reg_node(pRExC_state, TAIL);
12611 if (! REGTAIL(pRExC_state, ret, ender)) {
12612 REQUIRE_BRANCHJ(flagp, 0);
12616 if (ISMULT2(RExC_parse)) {
12618 vFAIL("Nested quantifiers");
12625 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
12626 regnode_offset * node_p,
12634 /* This routine teases apart the various meanings of \N and returns
12635 * accordingly. The input parameters constrain which meaning(s) is/are valid
12636 * in the current context.
12638 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
12640 * If <code_point_p> is not NULL, the context is expecting the result to be a
12641 * single code point. If this \N instance turns out to a single code point,
12642 * the function returns TRUE and sets *code_point_p to that code point.
12644 * If <node_p> is not NULL, the context is expecting the result to be one of
12645 * the things representable by a regnode. If this \N instance turns out to be
12646 * one such, the function generates the regnode, returns TRUE and sets *node_p
12647 * to point to the offset of that regnode into the regex engine program being
12650 * If this instance of \N isn't legal in any context, this function will
12651 * generate a fatal error and not return.
12653 * On input, RExC_parse should point to the first char following the \N at the
12654 * time of the call. On successful return, RExC_parse will have been updated
12655 * to point to just after the sequence identified by this routine. Also
12656 * *flagp has been updated as needed.
12658 * When there is some problem with the current context and this \N instance,
12659 * the function returns FALSE, without advancing RExC_parse, nor setting
12660 * *node_p, nor *code_point_p, nor *flagp.
12662 * If <cp_count> is not NULL, the caller wants to know the length (in code
12663 * points) that this \N sequence matches. This is set, and the input is
12664 * parsed for errors, even if the function returns FALSE, as detailed below.
12666 * There are 6 possibilities here, as detailed in the next 6 paragraphs.
12668 * Probably the most common case is for the \N to specify a single code point.
12669 * *cp_count will be set to 1, and *code_point_p will be set to that code
12672 * Another possibility is for the input to be an empty \N{}. This is no
12673 * longer accepted, and will generate a fatal error.
12675 * Another possibility is for a custom charnames handler to be in effect which
12676 * translates the input name to an empty string. *cp_count will be set to 0.
12677 * *node_p will be set to a generated NOTHING node.
12679 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
12680 * set to 0. *node_p will be set to a generated REG_ANY node.
12682 * The fifth possibility is that \N resolves to a sequence of more than one
12683 * code points. *cp_count will be set to the number of code points in the
12684 * sequence. *node_p will be set to a generated node returned by this
12685 * function calling S_reg().
12687 * The final possibility is that it is premature to be calling this function;
12688 * the parse needs to be restarted. This can happen when this changes from
12689 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
12690 * latter occurs only when the fifth possibility would otherwise be in
12691 * effect, and is because one of those code points requires the pattern to be
12692 * recompiled as UTF-8. The function returns FALSE, and sets the
12693 * RESTART_PARSE and NEED_UTF8 flags in *flagp, as appropriate. When this
12694 * happens, the caller needs to desist from continuing parsing, and return
12695 * this information to its caller. This is not set for when there is only one
12696 * code point, as this can be called as part of an ANYOF node, and they can
12697 * store above-Latin1 code points without the pattern having to be in UTF-8.
12699 * For non-single-quoted regexes, the tokenizer has resolved character and
12700 * sequence names inside \N{...} into their Unicode values, normalizing the
12701 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
12702 * hex-represented code points in the sequence. This is done there because
12703 * the names can vary based on what charnames pragma is in scope at the time,
12704 * so we need a way to take a snapshot of what they resolve to at the time of
12705 * the original parse. [perl #56444].
12707 * That parsing is skipped for single-quoted regexes, so here we may get
12708 * '\N{NAME}', which is parsed now. If the single-quoted regex is something
12709 * like '\N{U+41}', that code point is Unicode, and has to be translated into
12710 * the native character set for non-ASCII platforms. The other possibilities
12711 * are already native, so no translation is done. */
12713 char * endbrace; /* points to '}' following the name */
12714 char* p = RExC_parse; /* Temporary */
12716 SV * substitute_parse = NULL;
12721 GET_RE_DEBUG_FLAGS_DECL;
12723 PERL_ARGS_ASSERT_GROK_BSLASH_N;
12725 GET_RE_DEBUG_FLAGS;
12727 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
12728 assert(! (node_p && cp_count)); /* At most 1 should be set */
12730 if (cp_count) { /* Initialize return for the most common case */
12734 /* The [^\n] meaning of \N ignores spaces and comments under the /x
12735 * modifier. The other meanings do not, so use a temporary until we find
12736 * out which we are being called with */
12737 skip_to_be_ignored_text(pRExC_state, &p,
12738 FALSE /* Don't force to /x */ );
12740 /* Disambiguate between \N meaning a named character versus \N meaning
12741 * [^\n]. The latter is assumed when the {...} following the \N is a legal
12742 * quantifier, or if there is no '{' at all */
12743 if (*p != '{' || regcurly(p)) {
12753 *node_p = reg_node(pRExC_state, REG_ANY);
12754 *flagp |= HASWIDTH|SIMPLE;
12756 Set_Node_Length(REGNODE_p(*(node_p)), 1); /* MJD */
12760 /* The test above made sure that the next real character is a '{', but
12761 * under the /x modifier, it could be separated by space (or a comment and
12762 * \n) and this is not allowed (for consistency with \x{...} and the
12763 * tokenizer handling of \N{NAME}). */
12764 if (*RExC_parse != '{') {
12765 vFAIL("Missing braces on \\N{}");
12768 RExC_parse++; /* Skip past the '{' */
12770 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
12771 if (! endbrace) { /* no trailing brace */
12772 vFAIL2("Missing right brace on \\%c{}", 'N');
12775 /* Here, we have decided it should be a named character or sequence. These
12776 * imply Unicode semantics */
12777 REQUIRE_UNI_RULES(flagp, FALSE);
12779 /* \N{_} is what toke.c returns to us to indicate a name that evaluates to
12780 * nothing at all (not allowed under strict) */
12781 if (endbrace - RExC_parse == 1 && *RExC_parse == '_') {
12782 RExC_parse = endbrace;
12784 RExC_parse++; /* Position after the "}" */
12785 vFAIL("Zero length \\N{}");
12791 nextchar(pRExC_state);
12796 *node_p = reg_node(pRExC_state, NOTHING);
12800 if (endbrace - RExC_parse < 2 || ! strBEGINs(RExC_parse, "U+")) {
12802 /* Here, the name isn't of the form U+.... This can happen if the
12803 * pattern is single-quoted, so didn't get evaluated in toke.c. Now
12804 * is the time to find out what the name means */
12806 const STRLEN name_len = endbrace - RExC_parse;
12807 SV * value_sv; /* What does this name evaluate to */
12809 const U8 * value; /* string of name's value */
12810 STRLEN value_len; /* and its length */
12812 /* RExC_unlexed_names is a hash of names that weren't evaluated by
12813 * toke.c, and their values. Make sure is initialized */
12814 if (! RExC_unlexed_names) {
12815 RExC_unlexed_names = newHV();
12818 /* If we have already seen this name in this pattern, use that. This
12819 * allows us to only call the charnames handler once per name per
12820 * pattern. A broken or malicious handler could return something
12821 * different each time, which could cause the results to vary depending
12822 * on if something gets added or subtracted from the pattern that
12823 * causes the number of passes to change, for example */
12824 if ((value_svp = hv_fetch(RExC_unlexed_names, RExC_parse,
12827 value_sv = *value_svp;
12829 else { /* Otherwise we have to go out and get the name */
12830 const char * error_msg = NULL;
12831 value_sv = get_and_check_backslash_N_name(RExC_parse, endbrace,
12835 RExC_parse = endbrace;
12839 /* If no error message, should have gotten a valid return */
12842 /* Save the name's meaning for later use */
12843 if (! hv_store(RExC_unlexed_names, RExC_parse, name_len,
12846 Perl_croak(aTHX_ "panic: hv_store() unexpectedly failed");
12850 /* Here, we have the value the name evaluates to in 'value_sv' */
12851 value = (U8 *) SvPV(value_sv, value_len);
12853 /* See if the result is one code point vs 0 or multiple */
12854 if (inRANGE(value_len, 1, ((UV) SvUTF8(value_sv)
12858 /* Here, exactly one code point. If that isn't what is wanted,
12860 if (! code_point_p) {
12865 /* Convert from string to numeric code point */
12866 *code_point_p = (SvUTF8(value_sv))
12867 ? valid_utf8_to_uvchr(value, NULL)
12870 /* Have parsed this entire single code point \N{...}. *cp_count
12871 * has already been set to 1, so don't do it again. */
12872 RExC_parse = endbrace;
12873 nextchar(pRExC_state);
12875 } /* End of is a single code point */
12877 /* Count the code points, if caller desires. The API says to do this
12878 * even if we will later return FALSE */
12882 *cp_count = (SvUTF8(value_sv))
12883 ? utf8_length(value, value + value_len)
12887 /* Fail if caller doesn't want to handle a multi-code-point sequence.
12888 * But don't back the pointer up if the caller wants to know how many
12889 * code points there are (they need to handle it themselves in this
12898 /* Convert this to a sub-pattern of the form "(?: ... )", and then call
12899 * reg recursively to parse it. That way, it retains its atomicness,
12900 * while not having to worry about any special handling that some code
12901 * points may have. */
12903 substitute_parse = newSVpvs("?:");
12904 sv_catsv(substitute_parse, value_sv);
12905 sv_catpv(substitute_parse, ")");
12907 /* The value should already be native, so no need to convert on EBCDIC
12909 assert(! RExC_recode_x_to_native);
12912 else { /* \N{U+...} */
12913 Size_t count = 0; /* code point count kept internally */
12915 /* We can get to here when the input is \N{U+...} or when toke.c has
12916 * converted a name to the \N{U+...} form. This include changing a
12917 * name that evaluates to multiple code points to \N{U+c1.c2.c3 ...} */
12919 RExC_parse += 2; /* Skip past the 'U+' */
12921 /* Code points are separated by dots. The '}' terminates the whole
12924 do { /* Loop until the ending brace */
12926 char * start_digit; /* The first of the current code point */
12927 if (! isXDIGIT(*RExC_parse)) {
12929 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12932 start_digit = RExC_parse;
12935 /* Loop through the hex digits of the current code point */
12937 /* Adding this digit will shift the result 4 bits. If that
12938 * result would be above the legal max, it's overflow */
12939 if (cp > MAX_LEGAL_CP >> 4) {
12941 /* Find the end of the code point */
12944 } while (isXDIGIT(*RExC_parse) || *RExC_parse == '_');
12946 /* Be sure to synchronize this message with the similar one
12948 vFAIL4("Use of code point 0x%.*s is not allowed; the"
12949 " permissible max is 0x%" UVxf,
12950 (int) (RExC_parse - start_digit), start_digit,
12954 /* Accumulate this (valid) digit into the running total */
12955 cp = (cp << 4) + READ_XDIGIT(RExC_parse);
12957 /* READ_XDIGIT advanced the input pointer. Ignore a single
12958 * underscore separator */
12959 if (*RExC_parse == '_' && isXDIGIT(RExC_parse[1])) {
12962 } while (isXDIGIT(*RExC_parse));
12964 /* Here, have accumulated the next code point */
12965 if (RExC_parse >= endbrace) { /* If done ... */
12970 /* Here, is a single code point; fail if doesn't want that */
12971 if (! code_point_p) {
12976 /* A single code point is easy to handle; just return it */
12977 *code_point_p = UNI_TO_NATIVE(cp);
12978 RExC_parse = endbrace;
12979 nextchar(pRExC_state);
12983 /* Here, the only legal thing would be a multiple character
12984 * sequence (of the form "\N{U+c1.c2. ... }". So the next
12985 * character must be a dot (and the one after that can't be the
12986 * endbrace, or we'd have something like \N{U+100.} ) */
12987 if (*RExC_parse != '.' || RExC_parse + 1 >= endbrace) {
12988 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
12989 ? UTF8SKIP(RExC_parse)
12991 if (RExC_parse >= endbrace) { /* Guard against malformed utf8 */
12992 RExC_parse = endbrace;
12994 vFAIL("Invalid hexadecimal number in \\N{U+...}");
12997 /* Here, looks like its really a multiple character sequence. Fail
12998 * if that's not what the caller wants. But continue with counting
12999 * and error checking if they still want a count */
13000 if (! node_p && ! cp_count) {
13004 /* What is done here is to convert this to a sub-pattern of the
13005 * form \x{char1}\x{char2}... and then call reg recursively to
13006 * parse it (enclosing in "(?: ... )" ). That way, it retains its
13007 * atomicness, while not having to worry about special handling
13008 * that some code points may have. We don't create a subpattern,
13009 * but go through the motions of code point counting and error
13010 * checking, if the caller doesn't want a node returned. */
13012 if (node_p && count == 1) {
13013 substitute_parse = newSVpvs("?:");
13019 /* Convert to notation the rest of the code understands */
13020 sv_catpvs(substitute_parse, "\\x{");
13021 sv_catpvn(substitute_parse, start_digit,
13022 RExC_parse - start_digit);
13023 sv_catpvs(substitute_parse, "}");
13026 /* Move to after the dot (or ending brace the final time through.)
13031 } while (RExC_parse < endbrace);
13033 if (! node_p) { /* Doesn't want the node */
13040 sv_catpvs(substitute_parse, ")");
13042 /* The values are Unicode, and therefore have to be converted to native
13043 * on a non-Unicode (meaning non-ASCII) platform. */
13044 SET_recode_x_to_native(1);
13047 /* Here, we have the string the name evaluates to, ready to be parsed,
13048 * stored in 'substitute_parse' as a series of valid "\x{...}\x{...}"
13049 * constructs. This can be called from within a substitute parse already.
13050 * The error reporting mechanism doesn't work for 2 levels of this, but the
13051 * code above has validated this new construct, so there should be no
13052 * errors generated by the below. And this isn' an exact copy, so the
13053 * mechanism to seamlessly deal with this won't work, so turn off warnings
13055 save_start = RExC_start;
13056 orig_end = RExC_end;
13058 RExC_parse = RExC_start = SvPVX(substitute_parse);
13059 RExC_end = RExC_parse + SvCUR(substitute_parse);
13060 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
13062 *node_p = reg(pRExC_state, 1, &flags, depth+1);
13064 /* Restore the saved values */
13066 RExC_start = save_start;
13067 RExC_parse = endbrace;
13068 RExC_end = orig_end;
13069 SET_recode_x_to_native(0);
13071 SvREFCNT_dec_NN(substitute_parse);
13074 RETURN_FAIL_ON_RESTART(flags, flagp);
13075 FAIL2("panic: reg returned failure to grok_bslash_N, flags=%#" UVxf,
13078 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13080 nextchar(pRExC_state);
13086 PERL_STATIC_INLINE U8
13087 S_compute_EXACTish(RExC_state_t *pRExC_state)
13091 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
13099 op = get_regex_charset(RExC_flags);
13100 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
13101 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
13102 been, so there is no hole */
13105 return op + EXACTF;
13109 S_new_regcurly(const char *s, const char *e)
13111 /* This is a temporary function designed to match the most lenient form of
13112 * a {m,n} quantifier we ever envision, with either number omitted, and
13113 * spaces anywhere between/before/after them.
13115 * If this function fails, then the string it matches is very unlikely to
13116 * ever be considered a valid quantifier, so we can allow the '{' that
13117 * begins it to be considered as a literal */
13119 bool has_min = FALSE;
13120 bool has_max = FALSE;
13122 PERL_ARGS_ASSERT_NEW_REGCURLY;
13124 if (s >= e || *s++ != '{')
13127 while (s < e && isSPACE(*s)) {
13130 while (s < e && isDIGIT(*s)) {
13134 while (s < e && isSPACE(*s)) {
13140 while (s < e && isSPACE(*s)) {
13143 while (s < e && isDIGIT(*s)) {
13147 while (s < e && isSPACE(*s)) {
13152 return s < e && *s == '}' && (has_min || has_max);
13155 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
13156 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
13159 S_backref_value(char *p, char *e)
13161 const char* endptr = e;
13163 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
13170 - regatom - the lowest level
13172 Try to identify anything special at the start of the current parse position.
13173 If there is, then handle it as required. This may involve generating a
13174 single regop, such as for an assertion; or it may involve recursing, such as
13175 to handle a () structure.
13177 If the string doesn't start with something special then we gobble up
13178 as much literal text as we can. If we encounter a quantifier, we have to
13179 back off the final literal character, as that quantifier applies to just it
13180 and not to the whole string of literals.
13182 Once we have been able to handle whatever type of thing started the
13183 sequence, we return the offset into the regex engine program being compiled
13184 at which any next regnode should be placed.
13186 Returns 0, setting *flagp to TRYAGAIN if reg() returns 0 with TRYAGAIN.
13187 Returns 0, setting *flagp to RESTART_PARSE if the parse needs to be
13188 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
13189 Otherwise does not return 0.
13191 Note: we have to be careful with escapes, as they can be both literal
13192 and special, and in the case of \10 and friends, context determines which.
13194 A summary of the code structure is:
13196 switch (first_byte) {
13197 cases for each special:
13198 handle this special;
13201 switch (2nd byte) {
13202 cases for each unambiguous special:
13203 handle this special;
13205 cases for each ambigous special/literal:
13207 if (special) handle here
13209 default: // unambiguously literal:
13212 default: // is a literal char
13215 create EXACTish node for literal;
13216 while (more input and node isn't full) {
13217 switch (input_byte) {
13218 cases for each special;
13219 make sure parse pointer is set so that the next call to
13220 regatom will see this special first
13221 goto loopdone; // EXACTish node terminated by prev. char
13223 append char to EXACTISH node;
13225 get next input byte;
13229 return the generated node;
13231 Specifically there are two separate switches for handling
13232 escape sequences, with the one for handling literal escapes requiring
13233 a dummy entry for all of the special escapes that are actually handled
13238 STATIC regnode_offset
13239 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
13242 regnode_offset ret = 0;
13248 GET_RE_DEBUG_FLAGS_DECL;
13250 *flagp = WORST; /* Tentatively. */
13252 DEBUG_PARSE("atom");
13254 PERL_ARGS_ASSERT_REGATOM;
13257 parse_start = RExC_parse;
13258 assert(RExC_parse < RExC_end);
13259 switch ((U8)*RExC_parse) {
13261 RExC_seen_zerolen++;
13262 nextchar(pRExC_state);
13263 if (RExC_flags & RXf_PMf_MULTILINE)
13264 ret = reg_node(pRExC_state, MBOL);
13266 ret = reg_node(pRExC_state, SBOL);
13267 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13270 nextchar(pRExC_state);
13272 RExC_seen_zerolen++;
13273 if (RExC_flags & RXf_PMf_MULTILINE)
13274 ret = reg_node(pRExC_state, MEOL);
13276 ret = reg_node(pRExC_state, SEOL);
13277 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13280 nextchar(pRExC_state);
13281 if (RExC_flags & RXf_PMf_SINGLELINE)
13282 ret = reg_node(pRExC_state, SANY);
13284 ret = reg_node(pRExC_state, REG_ANY);
13285 *flagp |= HASWIDTH|SIMPLE;
13287 Set_Node_Length(REGNODE_p(ret), 1); /* MJD */
13291 char * const oregcomp_parse = ++RExC_parse;
13292 ret = regclass(pRExC_state, flagp, depth+1,
13293 FALSE, /* means parse the whole char class */
13294 TRUE, /* allow multi-char folds */
13295 FALSE, /* don't silence non-portable warnings. */
13296 (bool) RExC_strict,
13297 TRUE, /* Allow an optimized regnode result */
13300 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13301 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13304 if (*RExC_parse != ']') {
13305 RExC_parse = oregcomp_parse;
13306 vFAIL("Unmatched [");
13308 nextchar(pRExC_state);
13309 Set_Node_Length(REGNODE_p(ret), RExC_parse - oregcomp_parse + 1); /* MJD */
13313 nextchar(pRExC_state);
13314 ret = reg(pRExC_state, 2, &flags, depth+1);
13316 if (flags & TRYAGAIN) {
13317 if (RExC_parse >= RExC_end) {
13318 /* Make parent create an empty node if needed. */
13319 *flagp |= TRYAGAIN;
13324 RETURN_FAIL_ON_RESTART(flags, flagp);
13325 FAIL2("panic: reg returned failure to regatom, flags=%#" UVxf,
13328 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
13332 if (flags & TRYAGAIN) {
13333 *flagp |= TRYAGAIN;
13336 vFAIL("Internal urp");
13337 /* Supposed to be caught earlier. */
13343 vFAIL("Quantifier follows nothing");
13348 This switch handles escape sequences that resolve to some kind
13349 of special regop and not to literal text. Escape sequences that
13350 resolve to literal text are handled below in the switch marked
13353 Every entry in this switch *must* have a corresponding entry
13354 in the literal escape switch. However, the opposite is not
13355 required, as the default for this switch is to jump to the
13356 literal text handling code.
13359 switch ((U8)*RExC_parse) {
13360 /* Special Escapes */
13362 RExC_seen_zerolen++;
13363 ret = reg_node(pRExC_state, SBOL);
13364 /* SBOL is shared with /^/ so we set the flags so we can tell
13365 * /\A/ from /^/ in split. */
13366 FLAGS(REGNODE_p(ret)) = 1;
13368 goto finish_meta_pat;
13370 ret = reg_node(pRExC_state, GPOS);
13371 RExC_seen |= REG_GPOS_SEEN;
13373 goto finish_meta_pat;
13375 if (!RExC_in_lookbehind && !RExC_in_lookahead) {
13376 RExC_seen_zerolen++;
13377 ret = reg_node(pRExC_state, KEEPS);
13379 /* XXX:dmq : disabling in-place substitution seems to
13380 * be necessary here to avoid cases of memory corruption, as
13381 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
13383 RExC_seen |= REG_LOOKBEHIND_SEEN;
13384 goto finish_meta_pat;
13387 ++RExC_parse; /* advance past the 'K' */
13388 vFAIL("\\K not permitted in lookahead/lookbehind");
13391 ret = reg_node(pRExC_state, SEOL);
13393 RExC_seen_zerolen++; /* Do not optimize RE away */
13394 goto finish_meta_pat;
13396 ret = reg_node(pRExC_state, EOS);
13398 RExC_seen_zerolen++; /* Do not optimize RE away */
13399 goto finish_meta_pat;
13401 vFAIL("\\C no longer supported");
13403 ret = reg_node(pRExC_state, CLUMP);
13404 *flagp |= HASWIDTH;
13405 goto finish_meta_pat;
13413 regex_charset charset = get_regex_charset(RExC_flags);
13415 RExC_seen_zerolen++;
13416 RExC_seen |= REG_LOOKBEHIND_SEEN;
13417 op = BOUND + charset;
13419 if (RExC_parse >= RExC_end || *(RExC_parse + 1) != '{') {
13420 flags = TRADITIONAL_BOUND;
13421 if (op > BOUNDA) { /* /aa is same as /a */
13427 char name = *RExC_parse;
13428 char * endbrace = NULL;
13430 endbrace = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
13433 vFAIL2("Missing right brace on \\%c{}", name);
13435 /* XXX Need to decide whether to take spaces or not. Should be
13436 * consistent with \p{}, but that currently is SPACE, which
13437 * means vertical too, which seems wrong
13438 * while (isBLANK(*RExC_parse)) {
13441 if (endbrace == RExC_parse) {
13442 RExC_parse++; /* After the '}' */
13443 vFAIL2("Empty \\%c{}", name);
13445 length = endbrace - RExC_parse;
13446 /*while (isBLANK(*(RExC_parse + length - 1))) {
13449 switch (*RExC_parse) {
13452 && (memNEs(RExC_parse + 1, length - 1, "cb")))
13454 goto bad_bound_type;
13459 if (length != 2 || *(RExC_parse + 1) != 'b') {
13460 goto bad_bound_type;
13465 if (length != 2 || *(RExC_parse + 1) != 'b') {
13466 goto bad_bound_type;
13471 if (length != 2 || *(RExC_parse + 1) != 'b') {
13472 goto bad_bound_type;
13478 RExC_parse = endbrace;
13480 "'%" UTF8f "' is an unknown bound type",
13481 UTF8fARG(UTF, length, endbrace - length));
13482 NOT_REACHED; /*NOTREACHED*/
13484 RExC_parse = endbrace;
13485 REQUIRE_UNI_RULES(flagp, 0);
13490 else if (op >= BOUNDA) { /* /aa is same as /a */
13494 /* Don't have to worry about UTF-8, in this message because
13495 * to get here the contents of the \b must be ASCII */
13496 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
13497 "Using /u for '%.*s' instead of /%s",
13499 endbrace - length + 1,
13500 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
13501 ? ASCII_RESTRICT_PAT_MODS
13502 : ASCII_MORE_RESTRICT_PAT_MODS);
13507 RExC_seen_d_op = TRUE;
13509 else if (op == BOUNDL) {
13510 RExC_contains_locale = 1;
13514 op += NBOUND - BOUND;
13517 ret = reg_node(pRExC_state, op);
13518 FLAGS(REGNODE_p(ret)) = flags;
13522 goto finish_meta_pat;
13526 ret = reg_node(pRExC_state, LNBREAK);
13527 *flagp |= HASWIDTH|SIMPLE;
13528 goto finish_meta_pat;
13542 /* These all have the same meaning inside [brackets], and it knows
13543 * how to do the best optimizations for them. So, pretend we found
13544 * these within brackets, and let it do the work */
13547 ret = regclass(pRExC_state, flagp, depth+1,
13548 TRUE, /* means just parse this element */
13549 FALSE, /* don't allow multi-char folds */
13550 FALSE, /* don't silence non-portable warnings. It
13551 would be a bug if these returned
13553 (bool) RExC_strict,
13554 TRUE, /* Allow an optimized regnode result */
13556 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13557 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
13558 * multi-char folds are allowed. */
13560 FAIL2("panic: regclass returned failure to regatom, flags=%#" UVxf,
13563 RExC_parse--; /* regclass() leaves this one too far ahead */
13566 /* The escapes above that don't take a parameter can't be
13567 * followed by a '{'. But 'pX', 'p{foo}' and
13568 * correspondingly 'P' can be */
13569 if ( RExC_parse - parse_start == 1
13570 && UCHARAT(RExC_parse + 1) == '{'
13571 && UNLIKELY(! new_regcurly(RExC_parse + 1, RExC_end)))
13574 vFAIL("Unescaped left brace in regex is illegal here");
13576 Set_Node_Offset(REGNODE_p(ret), parse_start);
13577 Set_Node_Length(REGNODE_p(ret), RExC_parse - parse_start + 1); /* MJD */
13578 nextchar(pRExC_state);
13581 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
13582 * \N{...} evaluates to a sequence of more than one code points).
13583 * The function call below returns a regnode, which is our result.
13584 * The parameters cause it to fail if the \N{} evaluates to a
13585 * single code point; we handle those like any other literal. The
13586 * reason that the multicharacter case is handled here and not as
13587 * part of the EXACtish code is because of quantifiers. In
13588 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
13589 * this way makes that Just Happen. dmq.
13590 * join_exact() will join this up with adjacent EXACTish nodes
13591 * later on, if appropriate. */
13593 if (grok_bslash_N(pRExC_state,
13594 &ret, /* Want a regnode returned */
13595 NULL, /* Fail if evaluates to a single code
13597 NULL, /* Don't need a count of how many code
13606 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13608 /* Here, evaluates to a single code point. Go get that */
13609 RExC_parse = parse_start;
13612 case 'k': /* Handle \k<NAME> and \k'NAME' */
13616 if ( RExC_parse >= RExC_end - 1
13617 || (( ch = RExC_parse[1]) != '<'
13622 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
13623 vFAIL2("Sequence %.2s... not terminated", parse_start);
13626 ret = handle_named_backref(pRExC_state,
13638 case '1': case '2': case '3': case '4':
13639 case '5': case '6': case '7': case '8': case '9':
13644 if (*RExC_parse == 'g') {
13648 if (*RExC_parse == '{') {
13652 if (*RExC_parse == '-') {
13656 if (hasbrace && !isDIGIT(*RExC_parse)) {
13657 if (isrel) RExC_parse--;
13659 goto parse_named_seq;
13662 if (RExC_parse >= RExC_end) {
13663 goto unterminated_g;
13665 num = S_backref_value(RExC_parse, RExC_end);
13667 vFAIL("Reference to invalid group 0");
13668 else if (num == I32_MAX) {
13669 if (isDIGIT(*RExC_parse))
13670 vFAIL("Reference to nonexistent group");
13673 vFAIL("Unterminated \\g... pattern");
13677 num = RExC_npar - num;
13679 vFAIL("Reference to nonexistent or unclosed group");
13683 num = S_backref_value(RExC_parse, RExC_end);
13684 /* bare \NNN might be backref or octal - if it is larger
13685 * than or equal RExC_npar then it is assumed to be an
13686 * octal escape. Note RExC_npar is +1 from the actual
13687 * number of parens. */
13688 /* Note we do NOT check if num == I32_MAX here, as that is
13689 * handled by the RExC_npar check */
13692 /* any numeric escape < 10 is always a backref */
13694 /* any numeric escape < RExC_npar is a backref */
13695 && num >= RExC_npar
13696 /* cannot be an octal escape if it starts with 8 */
13697 && *RExC_parse != '8'
13698 /* cannot be an octal escape if it starts with 9 */
13699 && *RExC_parse != '9'
13701 /* Probably not meant to be a backref, instead likely
13702 * to be an octal character escape, e.g. \35 or \777.
13703 * The above logic should make it obvious why using
13704 * octal escapes in patterns is problematic. - Yves */
13705 RExC_parse = parse_start;
13710 /* At this point RExC_parse points at a numeric escape like
13711 * \12 or \88 or something similar, which we should NOT treat
13712 * as an octal escape. It may or may not be a valid backref
13713 * escape. For instance \88888888 is unlikely to be a valid
13715 while (isDIGIT(*RExC_parse))
13718 if (*RExC_parse != '}')
13719 vFAIL("Unterminated \\g{...} pattern");
13722 if (num >= (I32)RExC_npar) {
13724 /* It might be a forward reference; we can't fail until we
13725 * know, by completing the parse to get all the groups, and
13726 * then reparsing */
13727 if (ALL_PARENS_COUNTED) {
13728 if (num >= RExC_total_parens) {
13729 vFAIL("Reference to nonexistent group");
13733 REQUIRE_PARENS_PASS;
13737 ret = reganode(pRExC_state,
13740 : (ASCII_FOLD_RESTRICTED)
13742 : (AT_LEAST_UNI_SEMANTICS)
13748 if (OP(REGNODE_p(ret)) == REFF) {
13749 RExC_seen_d_op = TRUE;
13751 *flagp |= HASWIDTH;
13753 /* override incorrect value set in reganode MJD */
13754 Set_Node_Offset(REGNODE_p(ret), parse_start);
13755 Set_Node_Cur_Length(REGNODE_p(ret), parse_start-1);
13756 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13757 FALSE /* Don't force to /x */ );
13761 if (RExC_parse >= RExC_end)
13762 FAIL("Trailing \\");
13765 /* Do not generate "unrecognized" warnings here, we fall
13766 back into the quick-grab loop below */
13767 RExC_parse = parse_start;
13769 } /* end of switch on a \foo sequence */
13774 /* '#' comments should have been spaced over before this function was
13776 assert((RExC_flags & RXf_PMf_EXTENDED) == 0);
13778 if (RExC_flags & RXf_PMf_EXTENDED) {
13779 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
13780 if (RExC_parse < RExC_end)
13790 /* Here, we have determined that the next thing is probably a
13791 * literal character. RExC_parse points to the first byte of its
13792 * definition. (It still may be an escape sequence that evaluates
13793 * to a single character) */
13798 char *s, *old_s = NULL, *old_old_s = NULL;
13800 U32 max_string_len = 255;
13802 /* We may have to reparse the node, artificially stopping filling
13803 * it early, based on info gleaned in the first parse. This
13804 * variable gives where we stop. Make it above the normal stopping
13805 * place first time through; otherwise it would stop too early */
13806 U32 upper_fill = max_string_len + 1;
13808 /* We start out as an EXACT node, even if under /i, until we find a
13809 * character which is in a fold. The algorithm now segregates into
13810 * separate nodes, characters that fold from those that don't under
13811 * /i. (This hopefully will create nodes that are fixed strings
13812 * even under /i, giving the optimizer something to grab on to.)
13813 * So, if a node has something in it and the next character is in
13814 * the opposite category, that node is closed up, and the function
13815 * returns. Then regatom is called again, and a new node is
13816 * created for the new category. */
13817 U8 node_type = EXACT;
13819 /* Assume the node will be fully used; the excess is given back at
13820 * the end. Under /i, we may need to temporarily add the fold of
13821 * an extra character or two at the end to check for splitting
13822 * multi-char folds, so allocate extra space for that. We can't
13823 * make any other length assumptions, as a byte input sequence
13824 * could shrink down. */
13825 Ptrdiff_t current_string_nodes = STR_SZ(max_string_len
13829 ? UTF8_MAXBYTES_CASE
13830 /* Max non-UTF-8 expansion is 2 */ : 2)));
13832 bool next_is_quantifier;
13833 char * oldp = NULL;
13835 /* We can convert EXACTF nodes to EXACTFU if they contain only
13836 * characters that match identically regardless of the target
13837 * string's UTF8ness. The reason to do this is that EXACTF is not
13838 * trie-able, EXACTFU is, and EXACTFU requires fewer operations at
13841 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
13842 * contain only above-Latin1 characters (hence must be in UTF8),
13843 * which don't participate in folds with Latin1-range characters,
13844 * as the latter's folds aren't known until runtime. */
13845 bool maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13847 /* Single-character EXACTish nodes are almost always SIMPLE. This
13848 * allows us to override this as encountered */
13849 U8 maybe_SIMPLE = SIMPLE;
13851 /* Does this node contain something that can't match unless the
13852 * target string is (also) in UTF-8 */
13853 bool requires_utf8_target = FALSE;
13855 /* The sequence 'ss' is problematic in non-UTF-8 patterns. */
13856 bool has_ss = FALSE;
13858 /* So is the MICRO SIGN */
13859 bool has_micro_sign = FALSE;
13861 /* Set when we fill up the current node and there is still more
13862 * text to process */
13865 /* Allocate an EXACT node. The node_type may change below to
13866 * another EXACTish node, but since the size of the node doesn't
13867 * change, it works */
13868 ret = regnode_guts(pRExC_state, node_type, current_string_nodes,
13870 FILL_NODE(ret, node_type);
13873 s = STRING(REGNODE_p(ret));
13884 maybe_exactfu = FOLD && (DEPENDS_SEMANTICS || LOC);
13885 maybe_SIMPLE = SIMPLE;
13886 requires_utf8_target = FALSE;
13888 has_micro_sign = FALSE;
13892 /* This breaks under rare circumstances. If folding, we do not
13893 * want to split a node at a character that is a non-final in a
13894 * multi-char fold, as an input string could just happen to want to
13895 * match across the node boundary. The code at the end of the loop
13896 * looks for this, and backs off until it finds not such a
13897 * character, but it is possible (though extremely, extremely
13898 * unlikely) for all characters in the node to be non-final fold
13899 * ones, in which case we just leave the node fully filled, and
13900 * hope that it doesn't match the string in just the wrong place */
13902 assert( ! UTF /* Is at the beginning of a character */
13903 || UTF8_IS_INVARIANT(UCHARAT(RExC_parse))
13904 || UTF8_IS_START(UCHARAT(RExC_parse)));
13906 overflowed = FALSE;
13908 /* Here, we have a literal character. Find the maximal string of
13909 * them in the input that we can fit into a single EXACTish node.
13910 * We quit at the first non-literal or when the node gets full, or
13911 * under /i the categorization of folding/non-folding character
13913 while (p < RExC_end && len < upper_fill) {
13915 /* In most cases each iteration adds one byte to the output.
13916 * The exceptions override this */
13917 Size_t added_len = 1;
13923 /* White space has already been ignored */
13924 assert( (RExC_flags & RXf_PMf_EXTENDED) == 0
13925 || ! is_PATWS_safe((p), RExC_end, UTF));
13937 /* Literal Escapes Switch
13939 This switch is meant to handle escape sequences that
13940 resolve to a literal character.
13942 Every escape sequence that represents something
13943 else, like an assertion or a char class, is handled
13944 in the switch marked 'Special Escapes' above in this
13945 routine, but also has an entry here as anything that
13946 isn't explicitly mentioned here will be treated as
13947 an unescaped equivalent literal.
13950 switch ((U8)*++p) {
13952 /* These are all the special escapes. */
13953 case 'A': /* Start assertion */
13954 case 'b': case 'B': /* Word-boundary assertion*/
13955 case 'C': /* Single char !DANGEROUS! */
13956 case 'd': case 'D': /* digit class */
13957 case 'g': case 'G': /* generic-backref, pos assertion */
13958 case 'h': case 'H': /* HORIZWS */
13959 case 'k': case 'K': /* named backref, keep marker */
13960 case 'p': case 'P': /* Unicode property */
13961 case 'R': /* LNBREAK */
13962 case 's': case 'S': /* space class */
13963 case 'v': case 'V': /* VERTWS */
13964 case 'w': case 'W': /* word class */
13965 case 'X': /* eXtended Unicode "combining
13966 character sequence" */
13967 case 'z': case 'Z': /* End of line/string assertion */
13971 /* Anything after here is an escape that resolves to a
13972 literal. (Except digits, which may or may not)
13978 case 'N': /* Handle a single-code point named character. */
13979 RExC_parse = p + 1;
13980 if (! grok_bslash_N(pRExC_state,
13981 NULL, /* Fail if evaluates to
13982 anything other than a
13983 single code point */
13984 &ender, /* The returned single code
13986 NULL, /* Don't need a count of
13987 how many code points */
13992 if (*flagp & NEED_UTF8)
13993 FAIL("panic: grok_bslash_N set NEED_UTF8");
13994 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
13996 /* Here, it wasn't a single code point. Go close
13997 * up this EXACTish node. The switch() prior to
13998 * this switch handles the other cases */
13999 RExC_parse = p = oldp;
14003 RExC_parse = parse_start;
14005 /* The \N{} means the pattern, if previously /d,
14006 * becomes /u. That means it can't be an EXACTF node,
14007 * but an EXACTFU */
14008 if (node_type == EXACTF) {
14009 node_type = EXACTFU;
14011 /* If the node already contains something that
14012 * differs between EXACTF and EXACTFU, reparse it
14014 if (! maybe_exactfu) {
14035 ender = ESC_NATIVE;
14045 const char* error_msg;
14047 bool valid = grok_bslash_o(&p,
14051 TO_OUTPUT_WARNINGS(p),
14052 (bool) RExC_strict,
14053 TRUE, /* Output warnings
14058 RExC_parse = p; /* going to die anyway; point
14059 to exact spot of failure */
14062 UPDATE_WARNINGS_LOC(p - 1);
14068 UV result = UV_MAX; /* initialize to erroneous
14070 const char* error_msg;
14072 bool valid = grok_bslash_x(&p,
14076 TO_OUTPUT_WARNINGS(p),
14077 (bool) RExC_strict,
14078 TRUE, /* Silence warnings
14083 RExC_parse = p; /* going to die anyway; point
14084 to exact spot of failure */
14087 UPDATE_WARNINGS_LOC(p - 1);
14091 if (ender < 0x100) {
14092 if (RExC_recode_x_to_native) {
14093 ender = LATIN1_TO_NATIVE(ender);
14101 ender = grok_bslash_c(*p, TO_OUTPUT_WARNINGS(p));
14102 UPDATE_WARNINGS_LOC(p);
14105 case '8': case '9': /* must be a backreference */
14107 /* we have an escape like \8 which cannot be an octal escape
14108 * so we exit the loop, and let the outer loop handle this
14109 * escape which may or may not be a legitimate backref. */
14111 case '1': case '2': case '3':case '4':
14112 case '5': case '6': case '7':
14113 /* When we parse backslash escapes there is ambiguity
14114 * between backreferences and octal escapes. Any escape
14115 * from \1 - \9 is a backreference, any multi-digit
14116 * escape which does not start with 0 and which when
14117 * evaluated as decimal could refer to an already
14118 * parsed capture buffer is a back reference. Anything
14121 * Note this implies that \118 could be interpreted as
14122 * 118 OR as "\11" . "8" depending on whether there
14123 * were 118 capture buffers defined already in the
14126 /* NOTE, RExC_npar is 1 more than the actual number of
14127 * parens we have seen so far, hence the "<" as opposed
14129 if ( !isDIGIT(p[1]) || S_backref_value(p, RExC_end) < RExC_npar)
14130 { /* Not to be treated as an octal constant, go
14138 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14140 ender = grok_oct(p, &numlen, &flags, NULL);
14142 if ( isDIGIT(*p) /* like \08, \178 */
14143 && ckWARN(WARN_REGEXP)
14146 reg_warn_non_literal_string(
14148 form_short_octal_warning(p, numlen));
14154 FAIL("Trailing \\");
14157 if (isALPHANUMERIC(*p)) {
14158 /* An alpha followed by '{' is going to fail next
14159 * iteration, so don't output this warning in that
14161 if (! isALPHA(*p) || *(p + 1) != '{') {
14162 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s"
14163 " passed through", p);
14166 goto normal_default;
14167 } /* End of switch on '\' */
14170 /* Trying to gain new uses for '{' without breaking too
14171 * much existing code is hard. The solution currently
14173 * 1) If there is no ambiguity that a '{' should always
14174 * be taken literally, at the start of a construct, we
14176 * 2) If the literal '{' conflicts with our desired use
14177 * of it as a metacharacter, we die. The deprecation
14178 * cycles for this have come and gone.
14179 * 3) If there is ambiguity, we raise a simple warning.
14180 * This could happen, for example, if the user
14181 * intended it to introduce a quantifier, but slightly
14182 * misspelled the quantifier. Without this warning,
14183 * the quantifier would silently be taken as a literal
14184 * string of characters instead of a meta construct */
14185 if (len || (p > RExC_start && isALPHA_A(*(p - 1)))) {
14187 || ( p > parse_start + 1
14188 && isALPHA_A(*(p - 1))
14189 && *(p - 2) == '\\')
14190 || new_regcurly(p, RExC_end))
14192 RExC_parse = p + 1;
14193 vFAIL("Unescaped left brace in regex is "
14196 ckWARNreg(p + 1, "Unescaped left brace in regex is"
14197 " passed through");
14199 goto normal_default;
14202 if (p > RExC_parse && RExC_strict) {
14203 ckWARN2reg(p + 1, "Unescaped literal '%c'", *p);
14206 default: /* A literal character */
14208 if (! UTF8_IS_INVARIANT(*p) && UTF) {
14210 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
14211 &numlen, UTF8_ALLOW_DEFAULT);
14217 } /* End of switch on the literal */
14219 /* Here, have looked at the literal character, and <ender>
14220 * contains its ordinal; <p> points to the character after it.
14224 REQUIRE_UTF8(flagp);
14227 /* We need to check if the next non-ignored thing is a
14228 * quantifier. Move <p> to after anything that should be
14229 * ignored, which, as a side effect, positions <p> for the next
14230 * loop iteration */
14231 skip_to_be_ignored_text(pRExC_state, &p,
14232 FALSE /* Don't force to /x */ );
14234 /* If the next thing is a quantifier, it applies to this
14235 * character only, which means that this character has to be in
14236 * its own node and can't just be appended to the string in an
14237 * existing node, so if there are already other characters in
14238 * the node, close the node with just them, and set up to do
14239 * this character again next time through, when it will be the
14240 * only thing in its new node */
14242 next_is_quantifier = LIKELY(p < RExC_end)
14243 && UNLIKELY(ISMULT2(p));
14245 if (next_is_quantifier && LIKELY(len)) {
14250 /* Ready to add 'ender' to the node */
14252 if (! FOLD) { /* The simple case, just append the literal */
14255 /* Don't output if it would overflow */
14256 if (UNLIKELY(len > max_string_len - ((UTF)
14257 ? UVCHR_SKIP(ender)
14264 if (UVCHR_IS_INVARIANT(ender) || ! UTF) {
14265 *(s++) = (char) ender;
14268 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
14269 added_len = (char *) new_s - s;
14270 s = (char *) new_s;
14273 requires_utf8_target = TRUE;
14277 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
14279 /* Here are folding under /l, and the code point is
14280 * problematic. If this is the first character in the
14281 * node, change the node type to folding. Otherwise, if
14282 * this is the first problematic character, close up the
14283 * existing node, so can start a new node with this one */
14285 node_type = EXACTFL;
14286 RExC_contains_locale = 1;
14288 else if (node_type == EXACT) {
14293 /* This problematic code point means we can't simplify
14295 maybe_exactfu = FALSE;
14297 /* Here, we are adding a problematic fold character.
14298 * "Problematic" in this context means that its fold isn't
14299 * known until runtime. (The non-problematic code points
14300 * are the above-Latin1 ones that fold to also all
14301 * above-Latin1. Their folds don't vary no matter what the
14302 * locale is.) But here we have characters whose fold
14303 * depends on the locale. We just add in the unfolded
14304 * character, and wait until runtime to fold it */
14305 goto not_fold_common;
14307 else /* regular fold; see if actually is in a fold */
14308 if ( (ender < 256 && ! IS_IN_SOME_FOLD_L1(ender))
14310 && ! _invlist_contains_cp(PL_in_some_fold, ender)))
14312 /* Here, folding, but the character isn't in a fold.
14314 * Start a new node if previous characters in the node were
14316 if (len && node_type != EXACT) {
14321 /* Here, continuing a node with non-folded characters. Add
14323 goto not_fold_common;
14325 else { /* Here, does participate in some fold */
14327 /* If this is the first character in the node, change its
14328 * type to folding. Otherwise, if this is the first
14329 * folding character in the node, close up the existing
14330 * node, so can start a new node with this one. */
14332 node_type = compute_EXACTish(pRExC_state);
14334 else if (node_type == EXACT) {
14339 if (UTF) { /* Alway use the folded value for UTF-8
14341 if (UVCHR_IS_INVARIANT(ender)) {
14342 if (UNLIKELY(len + 1 > max_string_len)) {
14347 *(s)++ = (U8) toFOLD(ender);
14350 UV folded = _to_uni_fold_flags(
14352 (U8 *) s, /* We have allocated extra space
14353 in 's' so can't run off the
14356 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14357 ? FOLD_FLAGS_NOMIX_ASCII
14359 if (UNLIKELY(len + added_len > max_string_len)) {
14367 && LIKELY(folded != GREEK_SMALL_LETTER_MU))
14369 /* U+B5 folds to the MU, so its possible for a
14370 * non-UTF-8 target to match it */
14371 requires_utf8_target = TRUE;
14375 else { /* Here is non-UTF8. */
14377 /* The fold will be one or (rarely) two characters.
14378 * Check that there's room for at least a single one
14379 * before setting any flags, etc. Because otherwise an
14380 * overflowing character could cause a flag to be set
14381 * even though it doesn't end up in this node. (For
14382 * the two character fold, we check again, before
14383 * setting any flags) */
14384 if (UNLIKELY(len + 1 > max_string_len)) {
14389 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
14390 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
14391 || UNICODE_DOT_DOT_VERSION > 0)
14393 /* On non-ancient Unicodes, check for the only possible
14394 * multi-char fold */
14395 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
14397 /* This potential multi-char fold means the node
14398 * can't be simple (because it could match more
14399 * than a single char). And in some cases it will
14400 * match 'ss', so set that flag */
14404 /* It can't change to be an EXACTFU (unless already
14405 * is one). We fold it iff under /u rules. */
14406 if (node_type != EXACTFU) {
14407 maybe_exactfu = FALSE;
14410 if (UNLIKELY(len + 2 > max_string_len)) {
14419 goto done_with_this_char;
14422 else if ( UNLIKELY(isALPHA_FOLD_EQ(ender, 's'))
14424 && UNLIKELY(isALPHA_FOLD_EQ(*(s-1), 's')))
14426 /* Also, the sequence 'ss' is special when not
14427 * under /u. If the target string is UTF-8, it
14428 * should match SHARP S; otherwise it won't. So,
14429 * here we have to exclude the possibility of this
14430 * node moving to /u.*/
14432 maybe_exactfu = FALSE;
14435 /* Here, the fold will be a single character */
14437 if (UNLIKELY(ender == MICRO_SIGN)) {
14438 has_micro_sign = TRUE;
14440 else if (PL_fold[ender] != PL_fold_latin1[ender]) {
14442 /* If the character's fold differs between /d and
14443 * /u, this can't change to be an EXACTFU node */
14444 maybe_exactfu = FALSE;
14447 *(s++) = (DEPENDS_SEMANTICS)
14448 ? (char) toFOLD(ender)
14450 /* Under /u, the fold of any character in
14451 * the 0-255 range happens to be its
14452 * lowercase equivalent, except for LATIN
14453 * SMALL LETTER SHARP S, which was handled
14454 * above, and the MICRO SIGN, whose fold
14455 * requires UTF-8 to represent. */
14456 : (char) toLOWER_L1(ender);
14458 } /* End of adding current character to the node */
14460 done_with_this_char:
14464 if (next_is_quantifier) {
14466 /* Here, the next input is a quantifier, and to get here,
14467 * the current character is the only one in the node. */
14471 } /* End of loop through literal characters */
14473 /* Here we have either exhausted the input or run out of room in
14474 * the node. If the former, we are done. (If we encountered a
14475 * character that can't be in the node, transfer is made directly
14476 * to <loopdone>, and so we wouldn't have fallen off the end of the
14478 if (LIKELY(! overflowed)) {
14482 /* Here we have run out of room. We can grow plain EXACT and
14483 * LEXACT nodes. If the pattern is gigantic enough, though,
14484 * eventually we'll have to artificially chunk the pattern into
14485 * multiple nodes. */
14486 if (! LOC && (node_type == EXACT || node_type == LEXACT)) {
14487 Size_t overhead = 1 + regarglen[OP(REGNODE_p(ret))];
14488 Size_t overhead_expansion = 0;
14490 Size_t max_nodes_for_string;
14494 /* Here we couldn't fit the final character in the current
14495 * node, so it will have to be reparsed, no matter what else we
14499 /* If would have overflowed a regular EXACT node, switch
14500 * instead to an LEXACT. The code below is structured so that
14501 * the actual growing code is common to changing from an EXACT
14502 * or just increasing the LEXACT size. This means that we have
14503 * to save the string in the EXACT case before growing, and
14504 * then copy it afterwards to its new location */
14505 if (node_type == EXACT) {
14506 overhead_expansion = regarglen[LEXACT] - regarglen[EXACT];
14507 RExC_emit += overhead_expansion;
14508 Copy(s0, temp, len, char);
14511 /* Ready to grow. If it was a plain EXACT, the string was
14512 * saved, and the first few bytes of it overwritten by adding
14513 * an argument field. We assume, as we do elsewhere in this
14514 * file, that one byte of remaining input will translate into
14515 * one byte of output, and if that's too small, we grow again,
14516 * if too large the excess memory is freed at the end */
14518 max_nodes_for_string = U16_MAX - overhead - overhead_expansion;
14519 achievable = MIN(max_nodes_for_string,
14520 current_string_nodes + STR_SZ(RExC_end - p));
14521 delta = achievable - current_string_nodes;
14523 /* If there is just no more room, go finish up this chunk of
14529 change_engine_size(pRExC_state, delta + overhead_expansion);
14530 current_string_nodes += delta;
14532 = sizeof(struct regnode) * current_string_nodes;
14533 upper_fill = max_string_len + 1;
14535 /* If the length was small, we know this was originally an
14536 * EXACT node now converted to LEXACT, and the string has to be
14537 * restored. Otherwise the string was untouched. 260 is just
14538 * a number safely above 255 so don't have to worry about
14539 * getting it precise */
14541 node_type = LEXACT;
14542 FILL_NODE(ret, node_type);
14543 s0 = STRING(REGNODE_p(ret));
14544 Copy(temp, s0, len, char);
14548 goto continue_parse;
14551 bool splittable = FALSE;
14552 bool backed_up = FALSE;
14556 /* Here is /i. Running out of room creates a problem if we are
14557 * folding, and the split happens in the middle of a
14558 * multi-character fold, as a match that should have occurred,
14559 * won't, due to the way nodes are matched, and our artificial
14560 * boundary. So back off until we aren't splitting such a
14561 * fold. If there is no such place to back off to, we end up
14562 * taking the entire node as-is. This can happen if the node
14563 * consists entirely of 'f' or entirely of 's' characters (or
14564 * things that fold to them) as 'ff' and 'ss' are
14565 * multi-character folds.
14567 * The Unicode standard says that multi character folds consist
14568 * of either two or three characters. That means we would be
14569 * splitting one if the final character in the node is at the
14570 * beginning of either type, or is the second of a three
14574 * ender is the code point of the character that won't fit
14576 * s points to just beyond the final byte in the node.
14577 * It's where we would place ender if there were
14578 * room, and where in fact we do place ender's fold
14579 * in the code below, as we've over-allocated space
14580 * for s0 (hence s) to allow for this
14581 * e starts at 's' and advances as we append things.
14582 * old_s is the same as 's'. (If ender had fit, 's' would
14583 * have been advanced to beyond it).
14584 * old_old_s points to the beginning byte of the final
14585 * character in the node
14586 * p points to the beginning byte in the input of the
14587 * character beyond 'ender'.
14588 * oldp points to the beginning byte in the input of
14591 * In the case of /il, we haven't folded anything that could be
14592 * affected by the locale. That means only above-Latin1
14593 * characters that fold to other above-latin1 characters get
14594 * folded at compile time. To check where a good place to
14595 * split nodes is, everything in it will have to be folded.
14596 * The boolean 'maybe_exactfu' keeps track in /il if there are
14597 * any unfolded characters in the node. */
14598 bool need_to_fold_loc = LOC && ! maybe_exactfu;
14600 /* If we do need to fold the node, we need a place to store the
14601 * folded copy, and a way to map back to the unfolded original
14603 char * locfold_buf = NULL;
14604 Size_t * loc_correspondence = NULL;
14606 if (! need_to_fold_loc) { /* The normal case. Just
14607 initialize to the actual node */
14610 s = old_old_s; /* Point to the beginning of the final char
14611 that fits in the node */
14615 /* Here, we have filled a /il node, and there are unfolded
14616 * characters in it. If the runtime locale turns out to be
14617 * UTF-8, there are possible multi-character folds, just
14618 * like when not under /l. The node hence can't terminate
14619 * in the middle of such a fold. To determine this, we
14620 * have to create a folded copy of this node. That means
14621 * reparsing the node, folding everything assuming a UTF-8
14622 * locale. (If at runtime it isn't such a locale, the
14623 * actions here wouldn't have been necessary, but we have
14624 * to assume the worst case.) If we find we need to back
14625 * off the folded string, we do so, and then map that
14626 * position back to the original unfolded node, which then
14627 * gets output, truncated at that spot */
14629 char * redo_p = RExC_parse;
14633 /* Allow enough space assuming a single byte input folds to
14634 * a single byte output, plus assume that the two unparsed
14635 * characters (that we may need) fold to the largest number
14636 * of bytes possible, plus extra for one more worst case
14637 * scenario. In the loop below, if we start eating into
14638 * that final spare space, we enlarge this initial space */
14639 Size_t size = max_string_len + (3 * UTF8_MAXBYTES_CASE) + 1;
14641 Newxz(locfold_buf, size, char);
14642 Newxz(loc_correspondence, size, Size_t);
14644 /* Redo this node's parse, folding into 'locfold_buf' */
14645 redo_p = RExC_parse;
14646 old_redo_e = redo_e = locfold_buf;
14647 while (redo_p <= oldp) {
14649 old_redo_e = redo_e;
14650 loc_correspondence[redo_e - locfold_buf]
14651 = redo_p - RExC_parse;
14656 (void) _to_utf8_fold_flags((U8 *) redo_p,
14661 redo_e += added_len;
14662 redo_p += UTF8SKIP(redo_p);
14666 /* Note that if this code is run on some ancient
14667 * Unicode versions, SHARP S doesn't fold to 'ss',
14668 * but rather than clutter the code with #ifdef's,
14669 * as is done above, we ignore that possibility.
14670 * This is ok because this code doesn't affect what
14671 * gets matched, but merely where the node gets
14673 if (UCHARAT(redo_p) != LATIN_SMALL_LETTER_SHARP_S) {
14674 *redo_e++ = toLOWER_L1(UCHARAT(redo_p));
14684 /* If we're getting so close to the end that a
14685 * worst-case fold in the next character would cause us
14686 * to overflow, increase, assuming one byte output byte
14687 * per one byte input one, plus room for another worst
14689 if ( redo_p <= oldp
14690 && redo_e > locfold_buf + size
14691 - (UTF8_MAXBYTES_CASE + 1))
14693 Size_t new_size = size
14695 + UTF8_MAXBYTES_CASE + 1;
14696 Ptrdiff_t e_offset = redo_e - locfold_buf;
14698 Renew(locfold_buf, new_size, char);
14699 Renew(loc_correspondence, new_size, Size_t);
14702 redo_e = locfold_buf + e_offset;
14706 /* Set so that things are in terms of the folded, temporary
14709 s_start = locfold_buf;
14714 /* Here, we have 's', 's_start' and 'e' set up to point to the
14715 * input that goes into the node, folded.
14717 * If the final character of the node and the fold of ender
14718 * form the first two characters of a three character fold, we
14719 * need to peek ahead at the next (unparsed) character in the
14720 * input to determine if the three actually do form such a
14721 * fold. Just looking at that character is not generally
14722 * sufficient, as it could be, for example, an escape sequence
14723 * that evaluates to something else, and it needs to be folded.
14725 * khw originally thought to just go through the parse loop one
14726 * extra time, but that doesn't work easily as that iteration
14727 * could cause things to think that the parse is over and to
14728 * goto loopdone. The character could be a '$' for example, or
14729 * the character beyond could be a quantifier, and other
14730 * glitches as well.
14732 * The solution used here for peeking ahead is to look at that
14733 * next character. If it isn't ASCII punctuation, then it will
14734 * be something that continues in an EXACTish node if there
14735 * were space. We append the fold of it to s, having reserved
14736 * enough room in s0 for the purpose. If we can't reasonably
14737 * peek ahead, we instead assume the worst case: that it is
14738 * something that would form the completion of a multi-char
14741 * If we can't split between s and ender, we work backwards
14742 * character-by-character down to s0. At each current point
14743 * see if we are at the beginning of a multi-char fold. If so,
14744 * that means we would be splitting the fold across nodes, and
14745 * so we back up one and try again.
14747 * If we're not at the beginning, we still could be at the
14748 * final two characters of a (rare) three character fold. We
14749 * check if the sequence starting at the character before the
14750 * current position (and including the current and next
14751 * characters) is a three character fold. If not, the node can
14752 * be split here. If it is, we have to backup two characters
14755 * Otherwise, the node can be split at the current position.
14757 * The same logic is used for UTF-8 patterns and not */
14761 /* Append the fold of ender */
14762 (void) _to_uni_fold_flags(
14766 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14767 ? FOLD_FLAGS_NOMIX_ASCII
14771 /* 's' and the character folded to by ender may be the
14772 * first two of a three-character fold, in which case the
14773 * node should not be split here. That may mean examining
14774 * the so-far unparsed character starting at 'p'. But if
14775 * ender folded to more than one character, we already have
14776 * three characters to look at. Also, we first check if
14777 * the sequence consisting of s and the next character form
14778 * the first two of some three character fold. If not,
14779 * there's no need to peek ahead. */
14780 if ( added_len <= UTF8SKIP(e - added_len)
14781 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_utf8_safe(s, e)))
14783 /* Here, the two do form the beginning of a potential
14784 * three character fold. The unexamined character may
14785 * or may not complete it. Peek at it. It might be
14786 * something that ends the node or an escape sequence,
14787 * in which case we don't know without a lot of work
14788 * what it evaluates to, so we have to assume the worst
14789 * case: that it does complete the fold, and so we
14790 * can't split here. All such instances will have
14791 * that character be an ASCII punctuation character,
14792 * like a backslash. So, for that case, backup one and
14793 * drop down to try at that position */
14795 s = (char *) utf8_hop_back((U8 *) s, -1,
14800 /* Here, since it's not punctuation, it must be a
14801 * real character, and we can append its fold to
14802 * 'e' (having deliberately reserved enough space
14803 * for this eventuality) and drop down to check if
14804 * the three actually do form a folded sequence */
14805 (void) _to_utf8_fold_flags(
14806 (U8 *) p, (U8 *) RExC_end,
14809 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
14810 ? FOLD_FLAGS_NOMIX_ASCII
14816 /* Here, we either have three characters available in
14817 * sequence starting at 's', or we have two characters and
14818 * know that the following one can't possibly be part of a
14819 * three character fold. We go through the node backwards
14820 * until we find a place where we can split it without
14821 * breaking apart a multi-character fold. At any given
14822 * point we have to worry about if such a fold begins at
14823 * the current 's', and also if a three-character fold
14824 * begins at s-1, (containing s and s+1). Splitting in
14825 * either case would break apart a fold */
14827 char *prev_s = (char *) utf8_hop_back((U8 *) s, -1,
14830 /* If is a multi-char fold, can't split here. Backup
14831 * one char and try again */
14832 if (UNLIKELY(is_MULTI_CHAR_FOLD_utf8_safe(s, e))) {
14838 /* If the two characters beginning at 's' are part of a
14839 * three character fold starting at the character
14840 * before s, we can't split either before or after s.
14841 * Backup two chars and try again */
14842 if ( LIKELY(s > s_start)
14843 && UNLIKELY(is_THREE_CHAR_FOLD_utf8_safe(prev_s, e)))
14846 s = (char *) utf8_hop_back((U8 *) s, -1, (U8 *) s_start);
14851 /* Here there's no multi-char fold between s and the
14852 * next character following it. We can split */
14856 } while (s > s_start); /* End of loops backing up through the node */
14858 /* Here we either couldn't find a place to split the node,
14859 * or else we broke out of the loop setting 'splittable' to
14860 * true. In the latter case, the place to split is between
14861 * the first and second characters in the sequence starting
14867 else { /* Pattern not UTF-8 */
14868 if ( ender != LATIN_SMALL_LETTER_SHARP_S
14869 || ASCII_FOLD_RESTRICTED)
14871 *e++ = toLOWER_L1(ender);
14879 && UNLIKELY(is_THREE_CHAR_FOLD_HEAD_latin1_safe(s, e)))
14886 if ( UCHARAT(p) != LATIN_SMALL_LETTER_SHARP_S
14887 || ASCII_FOLD_RESTRICTED)
14889 *e++ = toLOWER_L1(ender);
14899 if (UNLIKELY(is_MULTI_CHAR_FOLD_latin1_safe(s, e))) {
14905 if ( LIKELY(s > s_start)
14906 && UNLIKELY(is_THREE_CHAR_FOLD_latin1_safe(s - 1, e)))
14916 } while (s > s_start);
14923 /* Here, we are done backing up. If we didn't backup at all
14924 * (the likely case), just proceed */
14927 /* If we did find a place to split, reparse the entire node
14928 * stopping where we have calculated. */
14931 /* If we created a temporary folded string under /l, we
14932 * have to map that back to the original */
14933 if (need_to_fold_loc) {
14934 upper_fill = loc_correspondence[s - s_start];
14935 Safefree(locfold_buf);
14936 Safefree(loc_correspondence);
14938 if (upper_fill == 0) {
14939 FAIL2("panic: loc_correspondence[%d] is 0",
14940 (int) (s - s_start));
14944 upper_fill = s - s0;
14948 else if (need_to_fold_loc) {
14949 Safefree(locfold_buf);
14950 Safefree(loc_correspondence);
14953 /* Here the node consists entirely of non-final multi-char
14954 * folds. (Likely it is all 'f's or all 's's.) There's no
14955 * decent place to split it, so give up and just take the
14959 } /* End of verifying node ends with an appropriate char */
14961 /* We need to start the next node at the character that didn't fit
14965 loopdone: /* Jumped to when encounters something that shouldn't be
14968 /* Free up any over-allocated space; cast is to silence bogus
14969 * warning in MS VC */
14970 change_engine_size(pRExC_state,
14971 - (Ptrdiff_t) (current_string_nodes - STR_SZ(len)));
14973 /* I (khw) don't know if you can get here with zero length, but the
14974 * old code handled this situation by creating a zero-length EXACT
14975 * node. Might as well be NOTHING instead */
14977 OP(REGNODE_p(ret)) = NOTHING;
14981 /* If the node type is EXACT here, check to see if it
14982 * should be EXACTL, or EXACT_REQ8. */
14983 if (node_type == EXACT) {
14985 node_type = EXACTL;
14987 else if (requires_utf8_target) {
14988 node_type = EXACT_REQ8;
14991 else if (node_type == LEXACT) {
14992 if (requires_utf8_target) {
14993 node_type = LEXACT_REQ8;
14997 if ( UNLIKELY(has_micro_sign || has_ss)
14998 && (node_type == EXACTFU || ( node_type == EXACTF
14999 && maybe_exactfu)))
15000 { /* These two conditions are problematic in non-UTF-8
15003 node_type = EXACTFUP;
15005 else if (node_type == EXACTFL) {
15007 /* 'maybe_exactfu' is deliberately set above to
15008 * indicate this node type, where all code points in it
15010 if (maybe_exactfu) {
15011 node_type = EXACTFLU8;
15014 _invlist_contains_cp(PL_HasMultiCharFold, ender)))
15016 /* A character that folds to more than one will
15017 * match multiple characters, so can't be SIMPLE.
15018 * We don't have to worry about this with EXACTFLU8
15019 * nodes just above, as they have already been
15020 * folded (since the fold doesn't vary at run
15021 * time). Here, if the final character in the node
15022 * folds to multiple, it can't be simple. (This
15023 * only has an effect if the node has only a single
15024 * character, hence the final one, as elsewhere we
15025 * turn off simple for nodes whose length > 1 */
15029 else if (node_type == EXACTF) { /* Means is /di */
15031 /* This intermediate variable is needed solely because
15032 * the asserts in the macro where used exceed Win32's
15033 * literal string capacity */
15034 char first_char = * STRING(REGNODE_p(ret));
15036 /* If 'maybe_exactfu' is clear, then we need to stay
15037 * /di. If it is set, it means there are no code
15038 * points that match differently depending on UTF8ness
15039 * of the target string, so it can become an EXACTFU
15041 if (! maybe_exactfu) {
15042 RExC_seen_d_op = TRUE;
15044 else if ( isALPHA_FOLD_EQ(first_char, 's')
15045 || isALPHA_FOLD_EQ(ender, 's'))
15047 /* But, if the node begins or ends in an 's' we
15048 * have to defer changing it into an EXACTFU, as
15049 * the node could later get joined with another one
15050 * that ends or begins with 's' creating an 'ss'
15051 * sequence which would then wrongly match the
15052 * sharp s without the target being UTF-8. We
15053 * create a special node that we resolve later when
15054 * we join nodes together */
15056 node_type = EXACTFU_S_EDGE;
15059 node_type = EXACTFU;
15063 if (requires_utf8_target && node_type == EXACTFU) {
15064 node_type = EXACTFU_REQ8;
15068 OP(REGNODE_p(ret)) = node_type;
15069 setSTR_LEN(REGNODE_p(ret), len);
15070 RExC_emit += STR_SZ(len);
15072 /* If the node isn't a single character, it can't be SIMPLE */
15073 if (len > (Size_t) ((UTF) ? UTF8SKIP(STRING(REGNODE_p(ret))) : 1)) {
15077 *flagp |= HASWIDTH | maybe_SIMPLE;
15080 Set_Node_Length(REGNODE_p(ret), p - parse_start - 1);
15084 /* len is STRLEN which is unsigned, need to copy to signed */
15087 vFAIL("Internal disaster");
15090 } /* End of label 'defchar:' */
15092 } /* End of giant switch on input character */
15094 /* Position parse to next real character */
15095 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
15096 FALSE /* Don't force to /x */ );
15097 if ( *RExC_parse == '{'
15098 && OP(REGNODE_p(ret)) != SBOL && ! regcurly(RExC_parse))
15100 if (RExC_strict || new_regcurly(RExC_parse, RExC_end)) {
15102 vFAIL("Unescaped left brace in regex is illegal here");
15104 ckWARNreg(RExC_parse + 1, "Unescaped left brace in regex is"
15105 " passed through");
15113 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
15115 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
15116 * sets up the bitmap and any flags, removing those code points from the
15117 * inversion list, setting it to NULL should it become completely empty */
15121 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
15122 assert(PL_regkind[OP(node)] == ANYOF);
15124 /* There is no bitmap for this node type */
15125 if (inRANGE(OP(node), ANYOFH, ANYOFRb)) {
15129 ANYOF_BITMAP_ZERO(node);
15130 if (*invlist_ptr) {
15132 /* This gets set if we actually need to modify things */
15133 bool change_invlist = FALSE;
15137 /* Start looking through *invlist_ptr */
15138 invlist_iterinit(*invlist_ptr);
15139 while (invlist_iternext(*invlist_ptr, &start, &end)) {
15143 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
15144 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
15147 /* Quit if are above what we should change */
15148 if (start >= NUM_ANYOF_CODE_POINTS) {
15152 change_invlist = TRUE;
15154 /* Set all the bits in the range, up to the max that we are doing */
15155 high = (end < NUM_ANYOF_CODE_POINTS - 1)
15157 : NUM_ANYOF_CODE_POINTS - 1;
15158 for (i = start; i <= (int) high; i++) {
15159 if (! ANYOF_BITMAP_TEST(node, i)) {
15160 ANYOF_BITMAP_SET(node, i);
15164 invlist_iterfinish(*invlist_ptr);
15166 /* Done with loop; remove any code points that are in the bitmap from
15167 * *invlist_ptr; similarly for code points above the bitmap if we have
15168 * a flag to match all of them anyways */
15169 if (change_invlist) {
15170 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
15172 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
15173 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
15176 /* If have completely emptied it, remove it completely */
15177 if (_invlist_len(*invlist_ptr) == 0) {
15178 SvREFCNT_dec_NN(*invlist_ptr);
15179 *invlist_ptr = NULL;
15184 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
15185 Character classes ([:foo:]) can also be negated ([:^foo:]).
15186 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
15187 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
15188 but trigger failures because they are currently unimplemented. */
15190 #define POSIXCC_DONE(c) ((c) == ':')
15191 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
15192 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
15193 #define MAYBE_POSIXCC(c) (POSIXCC(c) || (c) == '^' || (c) == ';')
15195 #define WARNING_PREFIX "Assuming NOT a POSIX class since "
15196 #define NO_BLANKS_POSIX_WARNING "no blanks are allowed in one"
15197 #define SEMI_COLON_POSIX_WARNING "a semi-colon was found instead of a colon"
15199 #define NOT_MEANT_TO_BE_A_POSIX_CLASS (OOB_NAMEDCLASS - 1)
15201 /* 'posix_warnings' and 'warn_text' are names of variables in the following
15203 #define ADD_POSIX_WARNING(p, text) STMT_START { \
15204 if (posix_warnings) { \
15205 if (! RExC_warn_text ) RExC_warn_text = \
15206 (AV *) sv_2mortal((SV *) newAV()); \
15207 av_push(RExC_warn_text, Perl_newSVpvf(aTHX_ \
15211 REPORT_LOCATION_ARGS(p))); \
15214 #define CLEAR_POSIX_WARNINGS() \
15216 if (posix_warnings && RExC_warn_text) \
15217 av_clear(RExC_warn_text); \
15220 #define CLEAR_POSIX_WARNINGS_AND_RETURN(ret) \
15222 CLEAR_POSIX_WARNINGS(); \
15227 S_handle_possible_posix(pTHX_ RExC_state_t *pRExC_state,
15229 const char * const s, /* Where the putative posix class begins.
15230 Normally, this is one past the '['. This
15231 parameter exists so it can be somewhere
15232 besides RExC_parse. */
15233 char ** updated_parse_ptr, /* Where to set the updated parse pointer, or
15235 AV ** posix_warnings, /* Where to place any generated warnings, or
15237 const bool check_only /* Don't die if error */
15240 /* This parses what the caller thinks may be one of the three POSIX
15242 * 1) a character class, like [:blank:]
15243 * 2) a collating symbol, like [. .]
15244 * 3) an equivalence class, like [= =]
15245 * In the latter two cases, it croaks if it finds a syntactically legal
15246 * one, as these are not handled by Perl.
15248 * The main purpose is to look for a POSIX character class. It returns:
15249 * a) the class number
15250 * if it is a completely syntactically and semantically legal class.
15251 * 'updated_parse_ptr', if not NULL, is set to point to just after the
15252 * closing ']' of the class
15253 * b) OOB_NAMEDCLASS
15254 * if it appears that one of the three POSIX constructs was meant, but
15255 * its specification was somehow defective. 'updated_parse_ptr', if
15256 * not NULL, is set to point to the character just after the end
15257 * character of the class. See below for handling of warnings.
15258 * c) NOT_MEANT_TO_BE_A_POSIX_CLASS
15259 * if it doesn't appear that a POSIX construct was intended.
15260 * 'updated_parse_ptr' is not changed. No warnings nor errors are
15263 * In b) there may be errors or warnings generated. If 'check_only' is
15264 * TRUE, then any errors are discarded. Warnings are returned to the
15265 * caller via an AV* created into '*posix_warnings' if it is not NULL. If
15266 * instead it is NULL, warnings are suppressed.
15268 * The reason for this function, and its complexity is that a bracketed
15269 * character class can contain just about anything. But it's easy to
15270 * mistype the very specific posix class syntax but yielding a valid
15271 * regular bracketed class, so it silently gets compiled into something
15272 * quite unintended.
15274 * The solution adopted here maintains backward compatibility except that
15275 * it adds a warning if it looks like a posix class was intended but
15276 * improperly specified. The warning is not raised unless what is input
15277 * very closely resembles one of the 14 legal posix classes. To do this,
15278 * it uses fuzzy parsing. It calculates how many single-character edits it
15279 * would take to transform what was input into a legal posix class. Only
15280 * if that number is quite small does it think that the intention was a
15281 * posix class. Obviously these are heuristics, and there will be cases
15282 * where it errs on one side or another, and they can be tweaked as
15283 * experience informs.
15285 * The syntax for a legal posix class is:
15287 * qr/(?xa: \[ : \^? [[:lower:]]{4,6} : \] )/
15289 * What this routine considers syntactically to be an intended posix class
15290 * is this (the comments indicate some restrictions that the pattern
15293 * qr/(?x: \[? # The left bracket, possibly
15295 * \h* # possibly followed by blanks
15296 * (?: \^ \h* )? # possibly a misplaced caret
15297 * [:;]? # The opening class character,
15298 * # possibly omitted. A typo
15299 * # semi-colon can also be used.
15301 * \^? # possibly a correctly placed
15302 * # caret, but not if there was also
15303 * # a misplaced one
15305 * .{3,15} # The class name. If there are
15306 * # deviations from the legal syntax,
15307 * # its edit distance must be close
15308 * # to a real class name in order
15309 * # for it to be considered to be
15310 * # an intended posix class.
15312 * [[:punct:]]? # The closing class character,
15313 * # possibly omitted. If not a colon
15314 * # nor semi colon, the class name
15315 * # must be even closer to a valid
15318 * \]? # The right bracket, possibly
15322 * In the above, \h must be ASCII-only.
15324 * These are heuristics, and can be tweaked as field experience dictates.
15325 * There will be cases when someone didn't intend to specify a posix class
15326 * that this warns as being so. The goal is to minimize these, while
15327 * maximizing the catching of things intended to be a posix class that
15328 * aren't parsed as such.
15332 const char * const e = RExC_end;
15333 unsigned complement = 0; /* If to complement the class */
15334 bool found_problem = FALSE; /* Assume OK until proven otherwise */
15335 bool has_opening_bracket = FALSE;
15336 bool has_opening_colon = FALSE;
15337 int class_number = OOB_NAMEDCLASS; /* Out-of-bounds until find
15339 const char * possible_end = NULL; /* used for a 2nd parse pass */
15340 const char* name_start; /* ptr to class name first char */
15342 /* If the number of single-character typos the input name is away from a
15343 * legal name is no more than this number, it is considered to have meant
15344 * the legal name */
15345 int max_distance = 2;
15347 /* to store the name. The size determines the maximum length before we
15348 * decide that no posix class was intended. Should be at least
15349 * sizeof("alphanumeric") */
15351 STATIC_ASSERT_DECL(C_ARRAY_LENGTH(input_text) >= sizeof "alphanumeric");
15353 PERL_ARGS_ASSERT_HANDLE_POSSIBLE_POSIX;
15355 CLEAR_POSIX_WARNINGS();
15358 return NOT_MEANT_TO_BE_A_POSIX_CLASS;
15361 if (*(p - 1) != '[') {
15362 ADD_POSIX_WARNING(p, "it doesn't start with a '['");
15363 found_problem = TRUE;
15366 has_opening_bracket = TRUE;
15369 /* They could be confused and think you can put spaces between the
15372 found_problem = TRUE;
15376 } while (p < e && isBLANK(*p));
15378 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15381 /* For [. .] and [= =]. These are quite different internally from [: :],
15382 * so they are handled separately. */
15383 if (POSIXCC_NOTYET(*p) && p < e - 3) /* 1 for the close, and 1 for the ']'
15384 and 1 for at least one char in it
15387 const char open_char = *p;
15388 const char * temp_ptr = p + 1;
15390 /* These two constructs are not handled by perl, and if we find a
15391 * syntactically valid one, we croak. khw, who wrote this code, finds
15392 * this explanation of them very unclear:
15393 * http://pubs.opengroup.org/onlinepubs/009696899/basedefs/xbd_chap09.html
15394 * And searching the rest of the internet wasn't very helpful either.
15395 * It looks like just about any byte can be in these constructs,
15396 * depending on the locale. But unless the pattern is being compiled
15397 * under /l, which is very rare, Perl runs under the C or POSIX locale.
15398 * In that case, it looks like [= =] isn't allowed at all, and that
15399 * [. .] could be any single code point, but for longer strings the
15400 * constituent characters would have to be the ASCII alphabetics plus
15401 * the minus-hyphen. Any sensible locale definition would limit itself
15402 * to these. And any portable one definitely should. Trying to parse
15403 * the general case is a nightmare (see [perl #127604]). So, this code
15404 * looks only for interiors of these constructs that match:
15406 * Using \w relaxes the apparent rules a little, without adding much
15407 * danger of mistaking something else for one of these constructs.
15409 * [. .] in some implementations described on the internet is usable to
15410 * escape a character that otherwise is special in bracketed character
15411 * classes. For example [.].] means a literal right bracket instead of
15412 * the ending of the class
15414 * [= =] can legitimately contain a [. .] construct, but we don't
15415 * handle this case, as that [. .] construct will later get parsed
15416 * itself and croak then. And [= =] is checked for even when not under
15417 * /l, as Perl has long done so.
15419 * The code below relies on there being a trailing NUL, so it doesn't
15420 * have to keep checking if the parse ptr < e.
15422 if (temp_ptr[1] == open_char) {
15425 else while ( temp_ptr < e
15426 && (isWORDCHAR(*temp_ptr) || *temp_ptr == '-'))
15431 if (*temp_ptr == open_char) {
15433 if (*temp_ptr == ']') {
15435 if (! found_problem && ! check_only) {
15436 RExC_parse = (char *) temp_ptr;
15437 vFAIL3("POSIX syntax [%c %c] is reserved for future "
15438 "extensions", open_char, open_char);
15441 /* Here, the syntax wasn't completely valid, or else the call
15442 * is to check-only */
15443 if (updated_parse_ptr) {
15444 *updated_parse_ptr = (char *) temp_ptr;
15447 CLEAR_POSIX_WARNINGS_AND_RETURN(OOB_NAMEDCLASS);
15451 /* If we find something that started out to look like one of these
15452 * constructs, but isn't, we continue below so that it can be checked
15453 * for being a class name with a typo of '.' or '=' instead of a colon.
15457 /* Here, we think there is a possibility that a [: :] class was meant, and
15458 * we have the first real character. It could be they think the '^' comes
15461 found_problem = TRUE;
15462 ADD_POSIX_WARNING(p + 1, "the '^' must come after the colon");
15467 found_problem = TRUE;
15471 } while (p < e && isBLANK(*p));
15473 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15477 /* But the first character should be a colon, which they could have easily
15478 * mistyped on a qwerty keyboard as a semi-colon (and which may be hard to
15479 * distinguish from a colon, so treat that as a colon). */
15482 has_opening_colon = TRUE;
15484 else if (*p == ';') {
15485 found_problem = TRUE;
15487 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15488 has_opening_colon = TRUE;
15491 found_problem = TRUE;
15492 ADD_POSIX_WARNING(p, "there must be a starting ':'");
15494 /* Consider an initial punctuation (not one of the recognized ones) to
15495 * be a left terminator */
15496 if (*p != '^' && *p != ']' && isPUNCT(*p)) {
15501 /* They may think that you can put spaces between the components */
15503 found_problem = TRUE;
15507 } while (p < e && isBLANK(*p));
15509 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15514 /* We consider something like [^:^alnum:]] to not have been intended to
15515 * be a posix class, but XXX maybe we should */
15517 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15524 /* Again, they may think that you can put spaces between the components */
15526 found_problem = TRUE;
15530 } while (p < e && isBLANK(*p));
15532 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15537 /* XXX This ']' may be a typo, and something else was meant. But
15538 * treating it as such creates enough complications, that that
15539 * possibility isn't currently considered here. So we assume that the
15540 * ']' is what is intended, and if we've already found an initial '[',
15541 * this leaves this construct looking like [:] or [:^], which almost
15542 * certainly weren't intended to be posix classes */
15543 if (has_opening_bracket) {
15544 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15547 /* But this function can be called when we parse the colon for
15548 * something like qr/[alpha:]]/, so we back up to look for the
15553 found_problem = TRUE;
15554 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15556 else if (*p != ':') {
15558 /* XXX We are currently very restrictive here, so this code doesn't
15559 * consider the possibility that, say, /[alpha.]]/ was intended to
15560 * be a posix class. */
15561 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15564 /* Here we have something like 'foo:]'. There was no initial colon,
15565 * and we back up over 'foo. XXX Unlike the going forward case, we
15566 * don't handle typos of non-word chars in the middle */
15567 has_opening_colon = FALSE;
15570 while (p > RExC_start && isWORDCHAR(*p)) {
15575 /* Here, we have positioned ourselves to where we think the first
15576 * character in the potential class is */
15579 /* Now the interior really starts. There are certain key characters that
15580 * can end the interior, or these could just be typos. To catch both
15581 * cases, we may have to do two passes. In the first pass, we keep on
15582 * going unless we come to a sequence that matches
15583 * qr/ [[:punct:]] [[:blank:]]* \] /xa
15584 * This means it takes a sequence to end the pass, so two typos in a row if
15585 * that wasn't what was intended. If the class is perfectly formed, just
15586 * this one pass is needed. We also stop if there are too many characters
15587 * being accumulated, but this number is deliberately set higher than any
15588 * real class. It is set high enough so that someone who thinks that
15589 * 'alphanumeric' is a correct name would get warned that it wasn't.
15590 * While doing the pass, we keep track of where the key characters were in
15591 * it. If we don't find an end to the class, and one of the key characters
15592 * was found, we redo the pass, but stop when we get to that character.
15593 * Thus the key character was considered a typo in the first pass, but a
15594 * terminator in the second. If two key characters are found, we stop at
15595 * the second one in the first pass. Again this can miss two typos, but
15596 * catches a single one
15598 * In the first pass, 'possible_end' starts as NULL, and then gets set to
15599 * point to the first key character. For the second pass, it starts as -1.
15605 bool has_blank = FALSE;
15606 bool has_upper = FALSE;
15607 bool has_terminating_colon = FALSE;
15608 bool has_terminating_bracket = FALSE;
15609 bool has_semi_colon = FALSE;
15610 unsigned int name_len = 0;
15611 int punct_count = 0;
15615 /* Squeeze out blanks when looking up the class name below */
15616 if (isBLANK(*p) ) {
15618 found_problem = TRUE;
15623 /* The name will end with a punctuation */
15625 const char * peek = p + 1;
15627 /* Treat any non-']' punctuation followed by a ']' (possibly
15628 * with intervening blanks) as trying to terminate the class.
15629 * ']]' is very likely to mean a class was intended (but
15630 * missing the colon), but the warning message that gets
15631 * generated shows the error position better if we exit the
15632 * loop at the bottom (eventually), so skip it here. */
15634 if (peek < e && isBLANK(*peek)) {
15636 found_problem = TRUE;
15639 } while (peek < e && isBLANK(*peek));
15642 if (peek < e && *peek == ']') {
15643 has_terminating_bracket = TRUE;
15645 has_terminating_colon = TRUE;
15647 else if (*p == ';') {
15648 has_semi_colon = TRUE;
15649 has_terminating_colon = TRUE;
15652 found_problem = TRUE;
15659 /* Here we have punctuation we thought didn't end the class.
15660 * Keep track of the position of the key characters that are
15661 * more likely to have been class-enders */
15662 if (*p == ']' || *p == '[' || *p == ':' || *p == ';') {
15664 /* Allow just one such possible class-ender not actually
15665 * ending the class. */
15666 if (possible_end) {
15672 /* If we have too many punctuation characters, no use in
15674 if (++punct_count > max_distance) {
15678 /* Treat the punctuation as a typo. */
15679 input_text[name_len++] = *p;
15682 else if (isUPPER(*p)) { /* Use lowercase for lookup */
15683 input_text[name_len++] = toLOWER(*p);
15685 found_problem = TRUE;
15687 } else if (! UTF || UTF8_IS_INVARIANT(*p)) {
15688 input_text[name_len++] = *p;
15692 input_text[name_len++] = utf8_to_uvchr_buf((U8 *) p, e, NULL);
15696 /* The declaration of 'input_text' is how long we allow a potential
15697 * class name to be, before saying they didn't mean a class name at
15699 if (name_len >= C_ARRAY_LENGTH(input_text)) {
15704 /* We get to here when the possible class name hasn't been properly
15705 * terminated before:
15706 * 1) we ran off the end of the pattern; or
15707 * 2) found two characters, each of which might have been intended to
15708 * be the name's terminator
15709 * 3) found so many punctuation characters in the purported name,
15710 * that the edit distance to a valid one is exceeded
15711 * 4) we decided it was more characters than anyone could have
15712 * intended to be one. */
15714 found_problem = TRUE;
15716 /* In the final two cases, we know that looking up what we've
15717 * accumulated won't lead to a match, even a fuzzy one. */
15718 if ( name_len >= C_ARRAY_LENGTH(input_text)
15719 || punct_count > max_distance)
15721 /* If there was an intermediate key character that could have been
15722 * an intended end, redo the parse, but stop there */
15723 if (possible_end && possible_end != (char *) -1) {
15724 possible_end = (char *) -1; /* Special signal value to say
15725 we've done a first pass */
15730 /* Otherwise, it can't have meant to have been a class */
15731 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15734 /* If we ran off the end, and the final character was a punctuation
15735 * one, back up one, to look at that final one just below. Later, we
15736 * will restore the parse pointer if appropriate */
15737 if (name_len && p == e && isPUNCT(*(p-1))) {
15742 if (p < e && isPUNCT(*p)) {
15744 has_terminating_bracket = TRUE;
15746 /* If this is a 2nd ']', and the first one is just below this
15747 * one, consider that to be the real terminator. This gives a
15748 * uniform and better positioning for the warning message */
15750 && possible_end != (char *) -1
15751 && *possible_end == ']'
15752 && name_len && input_text[name_len - 1] == ']')
15757 /* And this is actually equivalent to having done the 2nd
15758 * pass now, so set it to not try again */
15759 possible_end = (char *) -1;
15764 has_terminating_colon = TRUE;
15766 else if (*p == ';') {
15767 has_semi_colon = TRUE;
15768 has_terminating_colon = TRUE;
15776 /* Here, we have a class name to look up. We can short circuit the
15777 * stuff below for short names that can't possibly be meant to be a
15778 * class name. (We can do this on the first pass, as any second pass
15779 * will yield an even shorter name) */
15780 if (name_len < 3) {
15781 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15784 /* Find which class it is. Initially switch on the length of the name.
15786 switch (name_len) {
15788 if (memEQs(name_start, 4, "word")) {
15789 /* this is not POSIX, this is the Perl \w */
15790 class_number = ANYOF_WORDCHAR;
15794 /* Names all of length 5: alnum alpha ascii blank cntrl digit
15795 * graph lower print punct space upper
15796 * Offset 4 gives the best switch position. */
15797 switch (name_start[4]) {
15799 if (memBEGINs(name_start, 5, "alph")) /* alpha */
15800 class_number = ANYOF_ALPHA;
15803 if (memBEGINs(name_start, 5, "spac")) /* space */
15804 class_number = ANYOF_SPACE;
15807 if (memBEGINs(name_start, 5, "grap")) /* graph */
15808 class_number = ANYOF_GRAPH;
15811 if (memBEGINs(name_start, 5, "asci")) /* ascii */
15812 class_number = ANYOF_ASCII;
15815 if (memBEGINs(name_start, 5, "blan")) /* blank */
15816 class_number = ANYOF_BLANK;
15819 if (memBEGINs(name_start, 5, "cntr")) /* cntrl */
15820 class_number = ANYOF_CNTRL;
15823 if (memBEGINs(name_start, 5, "alnu")) /* alnum */
15824 class_number = ANYOF_ALPHANUMERIC;
15827 if (memBEGINs(name_start, 5, "lowe")) /* lower */
15828 class_number = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
15829 else if (memBEGINs(name_start, 5, "uppe")) /* upper */
15830 class_number = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
15833 if (memBEGINs(name_start, 5, "digi")) /* digit */
15834 class_number = ANYOF_DIGIT;
15835 else if (memBEGINs(name_start, 5, "prin")) /* print */
15836 class_number = ANYOF_PRINT;
15837 else if (memBEGINs(name_start, 5, "punc")) /* punct */
15838 class_number = ANYOF_PUNCT;
15843 if (memEQs(name_start, 6, "xdigit"))
15844 class_number = ANYOF_XDIGIT;
15848 /* If the name exactly matches a posix class name the class number will
15849 * here be set to it, and the input almost certainly was meant to be a
15850 * posix class, so we can skip further checking. If instead the syntax
15851 * is exactly correct, but the name isn't one of the legal ones, we
15852 * will return that as an error below. But if neither of these apply,
15853 * it could be that no posix class was intended at all, or that one
15854 * was, but there was a typo. We tease these apart by doing fuzzy
15855 * matching on the name */
15856 if (class_number == OOB_NAMEDCLASS && found_problem) {
15857 const UV posix_names[][6] = {
15858 { 'a', 'l', 'n', 'u', 'm' },
15859 { 'a', 'l', 'p', 'h', 'a' },
15860 { 'a', 's', 'c', 'i', 'i' },
15861 { 'b', 'l', 'a', 'n', 'k' },
15862 { 'c', 'n', 't', 'r', 'l' },
15863 { 'd', 'i', 'g', 'i', 't' },
15864 { 'g', 'r', 'a', 'p', 'h' },
15865 { 'l', 'o', 'w', 'e', 'r' },
15866 { 'p', 'r', 'i', 'n', 't' },
15867 { 'p', 'u', 'n', 'c', 't' },
15868 { 's', 'p', 'a', 'c', 'e' },
15869 { 'u', 'p', 'p', 'e', 'r' },
15870 { 'w', 'o', 'r', 'd' },
15871 { 'x', 'd', 'i', 'g', 'i', 't' }
15873 /* The names of the above all have added NULs to make them the same
15874 * size, so we need to also have the real lengths */
15875 const UV posix_name_lengths[] = {
15876 sizeof("alnum") - 1,
15877 sizeof("alpha") - 1,
15878 sizeof("ascii") - 1,
15879 sizeof("blank") - 1,
15880 sizeof("cntrl") - 1,
15881 sizeof("digit") - 1,
15882 sizeof("graph") - 1,
15883 sizeof("lower") - 1,
15884 sizeof("print") - 1,
15885 sizeof("punct") - 1,
15886 sizeof("space") - 1,
15887 sizeof("upper") - 1,
15888 sizeof("word") - 1,
15889 sizeof("xdigit")- 1
15892 int temp_max = max_distance; /* Use a temporary, so if we
15893 reparse, we haven't changed the
15896 /* Use a smaller max edit distance if we are missing one of the
15898 if ( has_opening_bracket + has_opening_colon < 2
15899 || has_terminating_bracket + has_terminating_colon < 2)
15904 /* See if the input name is close to a legal one */
15905 for (i = 0; i < C_ARRAY_LENGTH(posix_names); i++) {
15907 /* Short circuit call if the lengths are too far apart to be
15909 if (abs( (int) (name_len - posix_name_lengths[i]))
15915 if (edit_distance(input_text,
15918 posix_name_lengths[i],
15922 { /* If it is close, it probably was intended to be a class */
15923 goto probably_meant_to_be;
15927 /* Here the input name is not close enough to a valid class name
15928 * for us to consider it to be intended to be a posix class. If
15929 * we haven't already done so, and the parse found a character that
15930 * could have been terminators for the name, but which we absorbed
15931 * as typos during the first pass, repeat the parse, signalling it
15932 * to stop at that character */
15933 if (possible_end && possible_end != (char *) -1) {
15934 possible_end = (char *) -1;
15939 /* Here neither pass found a close-enough class name */
15940 CLEAR_POSIX_WARNINGS_AND_RETURN(NOT_MEANT_TO_BE_A_POSIX_CLASS);
15943 probably_meant_to_be:
15945 /* Here we think that a posix specification was intended. Update any
15947 if (updated_parse_ptr) {
15948 *updated_parse_ptr = (char *) p;
15951 /* If a posix class name was intended but incorrectly specified, we
15952 * output or return the warnings */
15953 if (found_problem) {
15955 /* We set flags for these issues in the parse loop above instead of
15956 * adding them to the list of warnings, because we can parse it
15957 * twice, and we only want one warning instance */
15959 ADD_POSIX_WARNING(p, "the name must be all lowercase letters");
15962 ADD_POSIX_WARNING(p, NO_BLANKS_POSIX_WARNING);
15964 if (has_semi_colon) {
15965 ADD_POSIX_WARNING(p, SEMI_COLON_POSIX_WARNING);
15967 else if (! has_terminating_colon) {
15968 ADD_POSIX_WARNING(p, "there is no terminating ':'");
15970 if (! has_terminating_bracket) {
15971 ADD_POSIX_WARNING(p, "there is no terminating ']'");
15974 if ( posix_warnings
15976 && av_top_index(RExC_warn_text) > -1)
15978 *posix_warnings = RExC_warn_text;
15981 else if (class_number != OOB_NAMEDCLASS) {
15982 /* If it is a known class, return the class. The class number
15983 * #defines are structured so each complement is +1 to the normal
15985 CLEAR_POSIX_WARNINGS_AND_RETURN(class_number + complement);
15987 else if (! check_only) {
15989 /* Here, it is an unrecognized class. This is an error (unless the
15990 * call is to check only, which we've already handled above) */
15991 const char * const complement_string = (complement)
15994 RExC_parse = (char *) p;
15995 vFAIL3utf8f("POSIX class [:%s%" UTF8f ":] unknown",
15997 UTF8fARG(UTF, RExC_parse - name_start - 2, name_start));
16001 return OOB_NAMEDCLASS;
16003 #undef ADD_POSIX_WARNING
16005 STATIC unsigned int
16006 S_regex_set_precedence(const U8 my_operator) {
16008 /* Returns the precedence in the (?[...]) construct of the input operator,
16009 * specified by its character representation. The precedence follows
16010 * general Perl rules, but it extends this so that ')' and ']' have (low)
16011 * precedence even though they aren't really operators */
16013 switch (my_operator) {
16029 NOT_REACHED; /* NOTREACHED */
16030 return 0; /* Silence compiler warning */
16033 STATIC regnode_offset
16034 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
16035 I32 *flagp, U32 depth,
16036 char * const oregcomp_parse)
16038 /* Handle the (?[...]) construct to do set operations */
16040 U8 curchar; /* Current character being parsed */
16041 UV start, end; /* End points of code point ranges */
16042 SV* final = NULL; /* The end result inversion list */
16043 SV* result_string; /* 'final' stringified */
16044 AV* stack; /* stack of operators and operands not yet
16046 AV* fence_stack = NULL; /* A stack containing the positions in
16047 'stack' of where the undealt-with left
16048 parens would be if they were actually
16050 /* The 'volatile' is a workaround for an optimiser bug
16051 * in Solaris Studio 12.3. See RT #127455 */
16052 volatile IV fence = 0; /* Position of where most recent undealt-
16053 with left paren in stack is; -1 if none.
16055 STRLEN len; /* Temporary */
16056 regnode_offset node; /* Temporary, and final regnode returned by
16058 const bool save_fold = FOLD; /* Temporary */
16059 char *save_end, *save_parse; /* Temporaries */
16060 const bool in_locale = LOC; /* we turn off /l during processing */
16062 GET_RE_DEBUG_FLAGS_DECL;
16064 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
16066 DEBUG_PARSE("xcls");
16069 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
16072 /* The use of this operator implies /u. This is required so that the
16073 * compile time values are valid in all runtime cases */
16074 REQUIRE_UNI_RULES(flagp, 0);
16076 ckWARNexperimental(RExC_parse,
16077 WARN_EXPERIMENTAL__REGEX_SETS,
16078 "The regex_sets feature is experimental");
16080 /* Everything in this construct is a metacharacter. Operands begin with
16081 * either a '\' (for an escape sequence), or a '[' for a bracketed
16082 * character class. Any other character should be an operator, or
16083 * parenthesis for grouping. Both types of operands are handled by calling
16084 * regclass() to parse them. It is called with a parameter to indicate to
16085 * return the computed inversion list. The parsing here is implemented via
16086 * a stack. Each entry on the stack is a single character representing one
16087 * of the operators; or else a pointer to an operand inversion list. */
16089 #define IS_OPERATOR(a) SvIOK(a)
16090 #define IS_OPERAND(a) (! IS_OPERATOR(a))
16092 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
16093 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
16094 * with pronouncing it called it Reverse Polish instead, but now that YOU
16095 * know how to pronounce it you can use the correct term, thus giving due
16096 * credit to the person who invented it, and impressing your geek friends.
16097 * Wikipedia says that the pronounciation of "Ł" has been changing so that
16098 * it is now more like an English initial W (as in wonk) than an L.)
16100 * This means that, for example, 'a | b & c' is stored on the stack as
16108 * where the numbers in brackets give the stack [array] element number.
16109 * In this implementation, parentheses are not stored on the stack.
16110 * Instead a '(' creates a "fence" so that the part of the stack below the
16111 * fence is invisible except to the corresponding ')' (this allows us to
16112 * replace testing for parens, by using instead subtraction of the fence
16113 * position). As new operands are processed they are pushed onto the stack
16114 * (except as noted in the next paragraph). New operators of higher
16115 * precedence than the current final one are inserted on the stack before
16116 * the lhs operand (so that when the rhs is pushed next, everything will be
16117 * in the correct positions shown above. When an operator of equal or
16118 * lower precedence is encountered in parsing, all the stacked operations
16119 * of equal or higher precedence are evaluated, leaving the result as the
16120 * top entry on the stack. This makes higher precedence operations
16121 * evaluate before lower precedence ones, and causes operations of equal
16122 * precedence to left associate.
16124 * The only unary operator '!' is immediately pushed onto the stack when
16125 * encountered. When an operand is encountered, if the top of the stack is
16126 * a '!", the complement is immediately performed, and the '!' popped. The
16127 * resulting value is treated as a new operand, and the logic in the
16128 * previous paragraph is executed. Thus in the expression
16130 * the stack looks like
16136 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
16143 * A ')' is treated as an operator with lower precedence than all the
16144 * aforementioned ones, which causes all operations on the stack above the
16145 * corresponding '(' to be evaluated down to a single resultant operand.
16146 * Then the fence for the '(' is removed, and the operand goes through the
16147 * algorithm above, without the fence.
16149 * A separate stack is kept of the fence positions, so that the position of
16150 * the latest so-far unbalanced '(' is at the top of it.
16152 * The ']' ending the construct is treated as the lowest operator of all,
16153 * so that everything gets evaluated down to a single operand, which is the
16156 sv_2mortal((SV *)(stack = newAV()));
16157 sv_2mortal((SV *)(fence_stack = newAV()));
16159 while (RExC_parse < RExC_end) {
16160 I32 top_index; /* Index of top-most element in 'stack' */
16161 SV** top_ptr; /* Pointer to top 'stack' element */
16162 SV* current = NULL; /* To contain the current inversion list
16164 SV* only_to_avoid_leaks;
16166 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16167 TRUE /* Force /x */ );
16168 if (RExC_parse >= RExC_end) { /* Fail */
16172 curchar = UCHARAT(RExC_parse);
16176 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16177 /* Enable with -Accflags=-DENABLE_REGEX_SETS_DEBUGGING */
16178 DEBUG_U(dump_regex_sets_structures(pRExC_state,
16179 stack, fence, fence_stack));
16182 top_index = av_tindex_skip_len_mg(stack);
16185 SV** stacked_ptr; /* Ptr to something already on 'stack' */
16186 char stacked_operator; /* The topmost operator on the 'stack'. */
16187 SV* lhs; /* Operand to the left of the operator */
16188 SV* rhs; /* Operand to the right of the operator */
16189 SV* fence_ptr; /* Pointer to top element of the fence
16194 if ( RExC_parse < RExC_end - 2
16195 && UCHARAT(RExC_parse + 1) == '?'
16196 && UCHARAT(RExC_parse + 2) == '^')
16198 /* If is a '(?', could be an embedded '(?^flags:(?[...])'.
16199 * This happens when we have some thing like
16201 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
16203 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
16205 * Here we would be handling the interpolated
16206 * '$thai_or_lao'. We handle this by a recursive call to
16207 * ourselves which returns the inversion list the
16208 * interpolated expression evaluates to. We use the flags
16209 * from the interpolated pattern. */
16210 U32 save_flags = RExC_flags;
16211 const char * save_parse;
16213 RExC_parse += 2; /* Skip past the '(?' */
16214 save_parse = RExC_parse;
16216 /* Parse the flags for the '(?'. We already know the first
16217 * flag to parse is a '^' */
16218 parse_lparen_question_flags(pRExC_state);
16220 if ( RExC_parse >= RExC_end - 4
16221 || UCHARAT(RExC_parse) != ':'
16222 || UCHARAT(++RExC_parse) != '('
16223 || UCHARAT(++RExC_parse) != '?'
16224 || UCHARAT(++RExC_parse) != '[')
16227 /* In combination with the above, this moves the
16228 * pointer to the point just after the first erroneous
16230 if (RExC_parse >= RExC_end - 4) {
16231 RExC_parse = RExC_end;
16233 else if (RExC_parse != save_parse) {
16234 RExC_parse += (UTF)
16235 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
16238 vFAIL("Expecting '(?flags:(?[...'");
16241 /* Recurse, with the meat of the embedded expression */
16243 if (! handle_regex_sets(pRExC_state, ¤t, flagp,
16244 depth+1, oregcomp_parse))
16246 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16249 /* Here, 'current' contains the embedded expression's
16250 * inversion list, and RExC_parse points to the trailing
16251 * ']'; the next character should be the ')' */
16253 if (UCHARAT(RExC_parse) != ')')
16254 vFAIL("Expecting close paren for nested extended charclass");
16256 /* Then the ')' matching the original '(' handled by this
16257 * case: statement */
16259 if (UCHARAT(RExC_parse) != ')')
16260 vFAIL("Expecting close paren for wrapper for nested extended charclass");
16262 RExC_flags = save_flags;
16263 goto handle_operand;
16266 /* A regular '('. Look behind for illegal syntax */
16267 if (top_index - fence >= 0) {
16268 /* If the top entry on the stack is an operator, it had
16269 * better be a '!', otherwise the entry below the top
16270 * operand should be an operator */
16271 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
16272 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
16273 || ( IS_OPERAND(*top_ptr)
16274 && ( top_index - fence < 1
16275 || ! (stacked_ptr = av_fetch(stack,
16278 || ! IS_OPERATOR(*stacked_ptr))))
16281 vFAIL("Unexpected '(' with no preceding operator");
16285 /* Stack the position of this undealt-with left paren */
16286 av_push(fence_stack, newSViv(fence));
16287 fence = top_index + 1;
16291 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16292 * multi-char folds are allowed. */
16293 if (!regclass(pRExC_state, flagp, depth+1,
16294 TRUE, /* means parse just the next thing */
16295 FALSE, /* don't allow multi-char folds */
16296 FALSE, /* don't silence non-portable warnings. */
16298 FALSE, /* Require return to be an ANYOF */
16301 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16302 goto regclass_failed;
16305 /* regclass() will return with parsing just the \ sequence,
16306 * leaving the parse pointer at the next thing to parse */
16308 goto handle_operand;
16310 case '[': /* Is a bracketed character class */
16312 /* See if this is a [:posix:] class. */
16313 bool is_posix_class = (OOB_NAMEDCLASS
16314 < handle_possible_posix(pRExC_state,
16318 TRUE /* checking only */));
16319 /* If it is a posix class, leave the parse pointer at the '['
16320 * to fool regclass() into thinking it is part of a
16321 * '[[:posix:]]'. */
16322 if (! is_posix_class) {
16326 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if
16327 * multi-char folds are allowed. */
16328 if (!regclass(pRExC_state, flagp, depth+1,
16329 is_posix_class, /* parse the whole char
16330 class only if not a
16332 FALSE, /* don't allow multi-char folds */
16333 TRUE, /* silence non-portable warnings. */
16335 FALSE, /* Require return to be an ANYOF */
16338 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16339 goto regclass_failed;
16346 /* function call leaves parse pointing to the ']', except if we
16348 if (is_posix_class) {
16352 goto handle_operand;
16356 if (top_index >= 1) {
16357 goto join_operators;
16360 /* Only a single operand on the stack: are done */
16364 if (av_tindex_skip_len_mg(fence_stack) < 0) {
16365 if (UCHARAT(RExC_parse - 1) == ']') {
16369 vFAIL("Unexpected ')'");
16372 /* If nothing after the fence, is missing an operand */
16373 if (top_index - fence < 0) {
16377 /* If at least two things on the stack, treat this as an
16379 if (top_index - fence >= 1) {
16380 goto join_operators;
16383 /* Here only a single thing on the fenced stack, and there is a
16384 * fence. Get rid of it */
16385 fence_ptr = av_pop(fence_stack);
16387 fence = SvIV(fence_ptr);
16388 SvREFCNT_dec_NN(fence_ptr);
16395 /* Having gotten rid of the fence, we pop the operand at the
16396 * stack top and process it as a newly encountered operand */
16397 current = av_pop(stack);
16398 if (IS_OPERAND(current)) {
16399 goto handle_operand;
16411 /* These binary operators should have a left operand already
16413 if ( top_index - fence < 0
16414 || top_index - fence == 1
16415 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
16416 || ! IS_OPERAND(*top_ptr))
16418 goto unexpected_binary;
16421 /* If only the one operand is on the part of the stack visible
16422 * to us, we just place this operator in the proper position */
16423 if (top_index - fence < 2) {
16425 /* Place the operator before the operand */
16427 SV* lhs = av_pop(stack);
16428 av_push(stack, newSVuv(curchar));
16429 av_push(stack, lhs);
16433 /* But if there is something else on the stack, we need to
16434 * process it before this new operator if and only if the
16435 * stacked operation has equal or higher precedence than the
16440 /* The operator on the stack is supposed to be below both its
16442 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
16443 || IS_OPERAND(*stacked_ptr))
16445 /* But if not, it's legal and indicates we are completely
16446 * done if and only if we're currently processing a ']',
16447 * which should be the final thing in the expression */
16448 if (curchar == ']') {
16454 vFAIL2("Unexpected binary operator '%c' with no "
16455 "preceding operand", curchar);
16457 stacked_operator = (char) SvUV(*stacked_ptr);
16459 if (regex_set_precedence(curchar)
16460 > regex_set_precedence(stacked_operator))
16462 /* Here, the new operator has higher precedence than the
16463 * stacked one. This means we need to add the new one to
16464 * the stack to await its rhs operand (and maybe more
16465 * stuff). We put it before the lhs operand, leaving
16466 * untouched the stacked operator and everything below it
16468 lhs = av_pop(stack);
16469 assert(IS_OPERAND(lhs));
16471 av_push(stack, newSVuv(curchar));
16472 av_push(stack, lhs);
16476 /* Here, the new operator has equal or lower precedence than
16477 * what's already there. This means the operation already
16478 * there should be performed now, before the new one. */
16480 rhs = av_pop(stack);
16481 if (! IS_OPERAND(rhs)) {
16483 /* This can happen when a ! is not followed by an operand,
16484 * like in /(?[\t &!])/ */
16488 lhs = av_pop(stack);
16490 if (! IS_OPERAND(lhs)) {
16492 /* This can happen when there is an empty (), like in
16493 * /(?[[0]+()+])/ */
16497 switch (stacked_operator) {
16499 _invlist_intersection(lhs, rhs, &rhs);
16504 _invlist_union(lhs, rhs, &rhs);
16508 _invlist_subtract(lhs, rhs, &rhs);
16511 case '^': /* The union minus the intersection */
16516 _invlist_union(lhs, rhs, &u);
16517 _invlist_intersection(lhs, rhs, &i);
16518 _invlist_subtract(u, i, &rhs);
16519 SvREFCNT_dec_NN(i);
16520 SvREFCNT_dec_NN(u);
16526 /* Here, the higher precedence operation has been done, and the
16527 * result is in 'rhs'. We overwrite the stacked operator with
16528 * the result. Then we redo this code to either push the new
16529 * operator onto the stack or perform any higher precedence
16530 * stacked operation */
16531 only_to_avoid_leaks = av_pop(stack);
16532 SvREFCNT_dec(only_to_avoid_leaks);
16533 av_push(stack, rhs);
16536 case '!': /* Highest priority, right associative */
16538 /* If what's already at the top of the stack is another '!",
16539 * they just cancel each other out */
16540 if ( (top_ptr = av_fetch(stack, top_index, FALSE))
16541 && (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) == '!'))
16543 only_to_avoid_leaks = av_pop(stack);
16544 SvREFCNT_dec(only_to_avoid_leaks);
16546 else { /* Otherwise, since it's right associative, just push
16548 av_push(stack, newSVuv(curchar));
16553 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16554 if (RExC_parse >= RExC_end) {
16557 vFAIL("Unexpected character");
16561 /* Here 'current' is the operand. If something is already on the
16562 * stack, we have to check if it is a !. But first, the code above
16563 * may have altered the stack in the time since we earlier set
16566 top_index = av_tindex_skip_len_mg(stack);
16567 if (top_index - fence >= 0) {
16568 /* If the top entry on the stack is an operator, it had better
16569 * be a '!', otherwise the entry below the top operand should
16570 * be an operator */
16571 top_ptr = av_fetch(stack, top_index, FALSE);
16573 if (IS_OPERATOR(*top_ptr)) {
16575 /* The only permissible operator at the top of the stack is
16576 * '!', which is applied immediately to this operand. */
16577 curchar = (char) SvUV(*top_ptr);
16578 if (curchar != '!') {
16579 SvREFCNT_dec(current);
16580 vFAIL2("Unexpected binary operator '%c' with no "
16581 "preceding operand", curchar);
16584 _invlist_invert(current);
16586 only_to_avoid_leaks = av_pop(stack);
16587 SvREFCNT_dec(only_to_avoid_leaks);
16589 /* And we redo with the inverted operand. This allows
16590 * handling multiple ! in a row */
16591 goto handle_operand;
16593 /* Single operand is ok only for the non-binary ')'
16595 else if ((top_index - fence == 0 && curchar != ')')
16596 || (top_index - fence > 0
16597 && (! (stacked_ptr = av_fetch(stack,
16600 || IS_OPERAND(*stacked_ptr))))
16602 SvREFCNT_dec(current);
16603 vFAIL("Operand with no preceding operator");
16607 /* Here there was nothing on the stack or the top element was
16608 * another operand. Just add this new one */
16609 av_push(stack, current);
16611 } /* End of switch on next parse token */
16613 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16614 } /* End of loop parsing through the construct */
16616 vFAIL("Syntax error in (?[...])");
16620 if (RExC_parse >= RExC_end || RExC_parse[1] != ')') {
16621 if (RExC_parse < RExC_end) {
16625 vFAIL("Unexpected ']' with no following ')' in (?[...");
16628 if (av_tindex_skip_len_mg(fence_stack) >= 0) {
16629 vFAIL("Unmatched (");
16632 if (av_tindex_skip_len_mg(stack) < 0 /* Was empty */
16633 || ((final = av_pop(stack)) == NULL)
16634 || ! IS_OPERAND(final)
16635 || ! is_invlist(final)
16636 || av_tindex_skip_len_mg(stack) >= 0) /* More left on stack */
16639 SvREFCNT_dec(final);
16640 vFAIL("Incomplete expression within '(?[ ])'");
16643 /* Here, 'final' is the resultant inversion list from evaluating the
16644 * expression. Return it if so requested */
16645 if (return_invlist) {
16646 *return_invlist = final;
16650 /* Otherwise generate a resultant node, based on 'final'. regclass() is
16651 * expecting a string of ranges and individual code points */
16652 invlist_iterinit(final);
16653 result_string = newSVpvs("");
16654 while (invlist_iternext(final, &start, &end)) {
16655 if (start == end) {
16656 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}", start);
16659 Perl_sv_catpvf(aTHX_ result_string, "\\x{%" UVXf "}-\\x{%" UVXf "}",
16664 /* About to generate an ANYOF (or similar) node from the inversion list we
16665 * have calculated */
16666 save_parse = RExC_parse;
16667 RExC_parse = SvPV(result_string, len);
16668 save_end = RExC_end;
16669 RExC_end = RExC_parse + len;
16670 TURN_OFF_WARNINGS_IN_SUBSTITUTE_PARSE;
16672 /* We turn off folding around the call, as the class we have constructed
16673 * already has all folding taken into consideration, and we don't want
16674 * regclass() to add to that */
16675 RExC_flags &= ~RXf_PMf_FOLD;
16676 /* regclass() can only return RESTART_PARSE and NEED_UTF8 if multi-char
16677 * folds are allowed. */
16678 node = regclass(pRExC_state, flagp, depth+1,
16679 FALSE, /* means parse the whole char class */
16680 FALSE, /* don't allow multi-char folds */
16681 TRUE, /* silence non-portable warnings. The above may very
16682 well have generated non-portable code points, but
16683 they're valid on this machine */
16684 FALSE, /* similarly, no need for strict */
16686 /* We can optimize into something besides an ANYOF, except
16687 * under /l, which needs to be ANYOF because of runtime
16688 * checks for locale sanity, etc */
16694 RExC_parse = save_parse + 1;
16695 RExC_end = save_end;
16696 SvREFCNT_dec_NN(final);
16697 SvREFCNT_dec_NN(result_string);
16700 RExC_flags |= RXf_PMf_FOLD;
16704 RETURN_FAIL_ON_RESTART(*flagp, flagp);
16705 goto regclass_failed;
16708 /* Fix up the node type if we are in locale. (We have pretended we are
16709 * under /u for the purposes of regclass(), as this construct will only
16710 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
16711 * as to cause any warnings about bad locales to be output in regexec.c),
16712 * and add the flag that indicates to check if not in a UTF-8 locale. The
16713 * reason we above forbid optimization into something other than an ANYOF
16714 * node is simply to minimize the number of code changes in regexec.c.
16715 * Otherwise we would have to create new EXACTish node types and deal with
16716 * them. This decision could be revisited should this construct become
16719 * (One might think we could look at the resulting ANYOF node and suppress
16720 * the flag if everything is above 255, as those would be UTF-8 only,
16721 * but this isn't true, as the components that led to that result could
16722 * have been locale-affected, and just happen to cancel each other out
16723 * under UTF-8 locales.) */
16725 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
16727 assert(OP(REGNODE_p(node)) == ANYOF);
16729 OP(REGNODE_p(node)) = ANYOFL;
16730 ANYOF_FLAGS(REGNODE_p(node))
16731 |= ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
16734 nextchar(pRExC_state);
16735 Set_Node_Length(REGNODE_p(node), RExC_parse - oregcomp_parse + 1); /* MJD */
16739 FAIL2("panic: regclass returned failure to handle_sets, " "flags=%#" UVxf,
16743 #ifdef ENABLE_REGEX_SETS_DEBUGGING
16746 S_dump_regex_sets_structures(pTHX_ RExC_state_t *pRExC_state,
16747 AV * stack, const IV fence, AV * fence_stack)
16748 { /* Dumps the stacks in handle_regex_sets() */
16750 const SSize_t stack_top = av_tindex_skip_len_mg(stack);
16751 const SSize_t fence_stack_top = av_tindex_skip_len_mg(fence_stack);
16754 PERL_ARGS_ASSERT_DUMP_REGEX_SETS_STRUCTURES;
16756 PerlIO_printf(Perl_debug_log, "\nParse position is:%s\n", RExC_parse);
16758 if (stack_top < 0) {
16759 PerlIO_printf(Perl_debug_log, "Nothing on stack\n");
16762 PerlIO_printf(Perl_debug_log, "Stack: (fence=%d)\n", (int) fence);
16763 for (i = stack_top; i >= 0; i--) {
16764 SV ** element_ptr = av_fetch(stack, i, FALSE);
16765 if (! element_ptr) {
16768 if (IS_OPERATOR(*element_ptr)) {
16769 PerlIO_printf(Perl_debug_log, "[%d]: %c\n",
16770 (int) i, (int) SvIV(*element_ptr));
16773 PerlIO_printf(Perl_debug_log, "[%d] ", (int) i);
16774 sv_dump(*element_ptr);
16779 if (fence_stack_top < 0) {
16780 PerlIO_printf(Perl_debug_log, "Nothing on fence_stack\n");
16783 PerlIO_printf(Perl_debug_log, "Fence_stack: \n");
16784 for (i = fence_stack_top; i >= 0; i--) {
16785 SV ** element_ptr = av_fetch(fence_stack, i, FALSE);
16786 if (! element_ptr) {
16789 PerlIO_printf(Perl_debug_log, "[%d]: %d\n",
16790 (int) i, (int) SvIV(*element_ptr));
16801 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
16803 /* This adds the Latin1/above-Latin1 folding rules.
16805 * This should be called only for a Latin1-range code points, cp, which is
16806 * known to be involved in a simple fold with other code points above
16807 * Latin1. It would give false results if /aa has been specified.
16808 * Multi-char folds are outside the scope of this, and must be handled
16811 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
16813 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
16815 /* The rules that are valid for all Unicode versions are hard-coded in */
16820 add_cp_to_invlist(*invlist, KELVIN_SIGN);
16824 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
16827 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
16828 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
16830 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
16831 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
16832 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
16834 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
16835 *invlist = add_cp_to_invlist(*invlist,
16836 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
16839 default: /* Other code points are checked against the data for the
16840 current Unicode version */
16842 Size_t folds_count;
16843 unsigned int first_fold;
16844 const unsigned int * remaining_folds;
16848 folded_cp = toFOLD(cp);
16851 U8 dummy_fold[UTF8_MAXBYTES_CASE+1];
16853 folded_cp = _to_fold_latin1(cp, dummy_fold, &dummy_len, 0);
16856 if (folded_cp > 255) {
16857 *invlist = add_cp_to_invlist(*invlist, folded_cp);
16860 folds_count = _inverse_folds(folded_cp, &first_fold,
16862 if (folds_count == 0) {
16864 /* Use deprecated warning to increase the chances of this being
16866 ckWARN2reg_d(RExC_parse,
16867 "Perl folding rules are not up-to-date for 0x%02X;"
16868 " please use the perlbug utility to report;", cp);
16873 if (first_fold > 255) {
16874 *invlist = add_cp_to_invlist(*invlist, first_fold);
16876 for (i = 0; i < folds_count - 1; i++) {
16877 if (remaining_folds[i] > 255) {
16878 *invlist = add_cp_to_invlist(*invlist,
16879 remaining_folds[i]);
16889 S_output_posix_warnings(pTHX_ RExC_state_t *pRExC_state, AV* posix_warnings)
16891 /* Output the elements of the array given by '*posix_warnings' as REGEXP
16895 const bool first_is_fatal = ckDEAD(packWARN(WARN_REGEXP));
16897 PERL_ARGS_ASSERT_OUTPUT_POSIX_WARNINGS;
16899 if (! TO_OUTPUT_WARNINGS(RExC_parse)) {
16903 while ((msg = av_shift(posix_warnings)) != &PL_sv_undef) {
16904 if (first_is_fatal) { /* Avoid leaking this */
16905 av_undef(posix_warnings); /* This isn't necessary if the
16906 array is mortal, but is a
16908 (void) sv_2mortal(msg);
16911 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s", SvPVX(msg));
16912 SvREFCNT_dec_NN(msg);
16915 UPDATE_WARNINGS_LOC(RExC_parse);
16918 PERL_STATIC_INLINE Size_t
16919 S_find_first_differing_byte_pos(const U8 * s1, const U8 * s2, const Size_t max)
16921 const U8 * const start = s1;
16922 const U8 * const send = start + max;
16924 PERL_ARGS_ASSERT_FIND_FIRST_DIFFERING_BYTE_POS;
16926 while (s1 < send && *s1 == *s2) {
16935 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
16937 /* This adds the string scalar <multi_string> to the array
16938 * <multi_char_matches>. <multi_string> is known to have exactly
16939 * <cp_count> code points in it. This is used when constructing a
16940 * bracketed character class and we find something that needs to match more
16941 * than a single character.
16943 * <multi_char_matches> is actually an array of arrays. Each top-level
16944 * element is an array that contains all the strings known so far that are
16945 * the same length. And that length (in number of code points) is the same
16946 * as the index of the top-level array. Hence, the [2] element is an
16947 * array, each element thereof is a string containing TWO code points;
16948 * while element [3] is for strings of THREE characters, and so on. Since
16949 * this is for multi-char strings there can never be a [0] nor [1] element.
16951 * When we rewrite the character class below, we will do so such that the
16952 * longest strings are written first, so that it prefers the longest
16953 * matching strings first. This is done even if it turns out that any
16954 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
16955 * Christiansen has agreed that this is ok. This makes the test for the
16956 * ligature 'ffi' come before the test for 'ff', for example */
16959 AV** this_array_ptr;
16961 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
16963 if (! multi_char_matches) {
16964 multi_char_matches = newAV();
16967 if (av_exists(multi_char_matches, cp_count)) {
16968 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
16969 this_array = *this_array_ptr;
16972 this_array = newAV();
16973 av_store(multi_char_matches, cp_count,
16976 av_push(this_array, multi_string);
16978 return multi_char_matches;
16981 /* The names of properties whose definitions are not known at compile time are
16982 * stored in this SV, after a constant heading. So if the length has been
16983 * changed since initialization, then there is a run-time definition. */
16984 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
16985 (SvCUR(listsv) != initial_listsv_len)
16987 /* There is a restricted set of white space characters that are legal when
16988 * ignoring white space in a bracketed character class. This generates the
16989 * code to skip them.
16991 * There is a line below that uses the same white space criteria but is outside
16992 * this macro. Both here and there must use the same definition */
16993 #define SKIP_BRACKETED_WHITE_SPACE(do_skip, p) \
16996 while (isBLANK_A(UCHARAT(p))) \
17003 STATIC regnode_offset
17004 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
17005 const bool stop_at_1, /* Just parse the next thing, don't
17006 look for a full character class */
17007 bool allow_mutiple_chars,
17008 const bool silence_non_portable, /* Don't output warnings
17012 bool optimizable, /* ? Allow a non-ANYOF return
17014 SV** ret_invlist /* Return an inversion list, not a node */
17017 /* parse a bracketed class specification. Most of these will produce an
17018 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
17019 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
17020 * under /i with multi-character folds: it will be rewritten following the
17021 * paradigm of this example, where the <multi-fold>s are characters which
17022 * fold to multiple character sequences:
17023 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
17024 * gets effectively rewritten as:
17025 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
17026 * reg() gets called (recursively) on the rewritten version, and this
17027 * function will return what it constructs. (Actually the <multi-fold>s
17028 * aren't physically removed from the [abcdefghi], it's just that they are
17029 * ignored in the recursion by means of a flag:
17030 * <RExC_in_multi_char_class>.)
17032 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
17033 * characters, with the corresponding bit set if that character is in the
17034 * list. For characters above this, an inversion list is used. There
17035 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
17036 * determinable at compile time
17038 * On success, returns the offset at which any next node should be placed
17039 * into the regex engine program being compiled.
17041 * Returns 0 otherwise, setting flagp to RESTART_PARSE if the parse needs
17042 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to
17047 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
17049 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
17050 regnode_offset ret = -1; /* Initialized to an illegal value */
17052 int namedclass = OOB_NAMEDCLASS;
17053 char *rangebegin = NULL;
17054 SV *listsv = NULL; /* List of \p{user-defined} whose definitions
17055 aren't available at the time this was called */
17056 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
17057 than just initialized. */
17058 SV* properties = NULL; /* Code points that match \p{} \P{} */
17059 SV* posixes = NULL; /* Code points that match classes like [:word:],
17060 extended beyond the Latin1 range. These have to
17061 be kept separate from other code points for much
17062 of this function because their handling is
17063 different under /i, and for most classes under
17065 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
17066 separate for a while from the non-complemented
17067 versions because of complications with /d
17069 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
17070 treated more simply than the general case,
17071 leading to less compilation and execution
17073 UV element_count = 0; /* Number of distinct elements in the class.
17074 Optimizations may be possible if this is tiny */
17075 AV * multi_char_matches = NULL; /* Code points that fold to more than one
17076 character; used under /i */
17078 char * stop_ptr = RExC_end; /* where to stop parsing */
17080 /* ignore unescaped whitespace? */
17081 const bool skip_white = cBOOL( ret_invlist
17082 || (RExC_flags & RXf_PMf_EXTENDED_MORE));
17084 /* inversion list of code points this node matches only when the target
17085 * string is in UTF-8. These are all non-ASCII, < 256. (Because is under
17087 SV* upper_latin1_only_utf8_matches = NULL;
17089 /* Inversion list of code points this node matches regardless of things
17090 * like locale, folding, utf8ness of the target string */
17091 SV* cp_list = NULL;
17093 /* Like cp_list, but code points on this list need to be checked for things
17094 * that fold to/from them under /i */
17095 SV* cp_foldable_list = NULL;
17097 /* Like cp_list, but code points on this list are valid only when the
17098 * runtime locale is UTF-8 */
17099 SV* only_utf8_locale_list = NULL;
17101 /* In a range, if one of the endpoints is non-character-set portable,
17102 * meaning that it hard-codes a code point that may mean a different
17103 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
17104 * mnemonic '\t' which each mean the same character no matter which
17105 * character set the platform is on. */
17106 unsigned int non_portable_endpoint = 0;
17108 /* Is the range unicode? which means on a platform that isn't 1-1 native
17109 * to Unicode (i.e. non-ASCII), each code point in it should be considered
17110 * to be a Unicode value. */
17111 bool unicode_range = FALSE;
17112 bool invert = FALSE; /* Is this class to be complemented */
17114 bool warn_super = ALWAYS_WARN_SUPER;
17116 const char * orig_parse = RExC_parse;
17118 /* This variable is used to mark where the end in the input is of something
17119 * that looks like a POSIX construct but isn't. During the parse, when
17120 * something looks like it could be such a construct is encountered, it is
17121 * checked for being one, but not if we've already checked this area of the
17122 * input. Only after this position is reached do we check again */
17123 char *not_posix_region_end = RExC_parse - 1;
17125 AV* posix_warnings = NULL;
17126 const bool do_posix_warnings = ckWARN(WARN_REGEXP);
17127 U8 op = END; /* The returned node-type, initialized to an impossible
17129 U8 anyof_flags = 0; /* flag bits if the node is an ANYOF-type */
17130 U32 posixl = 0; /* bit field of posix classes matched under /l */
17133 /* Flags as to what things aren't knowable until runtime. (Note that these are
17134 * mutually exclusive.) */
17135 #define HAS_USER_DEFINED_PROPERTY 0x01 /* /u any user-defined properties that
17136 haven't been defined as of yet */
17137 #define HAS_D_RUNTIME_DEPENDENCY 0x02 /* /d if the target being matched is
17139 #define HAS_L_RUNTIME_DEPENDENCY 0x04 /* /l what the posix classes match and
17140 what gets folded */
17141 U32 has_runtime_dependency = 0; /* OR of the above flags */
17143 GET_RE_DEBUG_FLAGS_DECL;
17145 PERL_ARGS_ASSERT_REGCLASS;
17147 PERL_UNUSED_ARG(depth);
17151 /* If wants an inversion list returned, we can't optimize to something
17154 optimizable = FALSE;
17157 DEBUG_PARSE("clas");
17159 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
17160 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
17161 && UNICODE_DOT_DOT_VERSION == 0)
17162 allow_mutiple_chars = FALSE;
17165 /* We include the /i status at the beginning of this so that we can
17166 * know it at runtime */
17167 listsv = sv_2mortal(Perl_newSVpvf(aTHX_ "#%d\n", cBOOL(FOLD)));
17168 initial_listsv_len = SvCUR(listsv);
17169 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
17171 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17173 assert(RExC_parse <= RExC_end);
17175 if (UCHARAT(RExC_parse) == '^') { /* Complement the class */
17178 allow_mutiple_chars = FALSE;
17180 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17183 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
17184 if (! ret_invlist && MAYBE_POSIXCC(UCHARAT(RExC_parse))) {
17185 int maybe_class = handle_possible_posix(pRExC_state,
17187 ¬_posix_region_end,
17189 TRUE /* checking only */);
17190 if (maybe_class >= OOB_NAMEDCLASS && do_posix_warnings) {
17191 ckWARN4reg(not_posix_region_end,
17192 "POSIX syntax [%c %c] belongs inside character classes%s",
17193 *RExC_parse, *RExC_parse,
17194 (maybe_class == OOB_NAMEDCLASS)
17195 ? ((POSIXCC_NOTYET(*RExC_parse))
17196 ? " (but this one isn't implemented)"
17197 : " (but this one isn't fully valid)")
17203 /* If the caller wants us to just parse a single element, accomplish this
17204 * by faking the loop ending condition */
17205 if (stop_at_1 && RExC_end > RExC_parse) {
17206 stop_ptr = RExC_parse + 1;
17209 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
17210 if (UCHARAT(RExC_parse) == ']')
17211 goto charclassloop;
17215 if ( posix_warnings
17216 && av_tindex_skip_len_mg(posix_warnings) >= 0
17217 && RExC_parse > not_posix_region_end)
17219 /* Warnings about posix class issues are considered tentative until
17220 * we are far enough along in the parse that we can no longer
17221 * change our mind, at which point we output them. This is done
17222 * each time through the loop so that a later class won't zap them
17223 * before they have been dealt with. */
17224 output_posix_warnings(pRExC_state, posix_warnings);
17227 if (RExC_parse >= stop_ptr) {
17231 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17233 if (UCHARAT(RExC_parse) == ']') {
17239 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
17240 save_value = value;
17241 save_prevvalue = prevvalue;
17244 rangebegin = RExC_parse;
17246 non_portable_endpoint = 0;
17248 if (UTF && ! UTF8_IS_INVARIANT(* RExC_parse)) {
17249 value = utf8n_to_uvchr((U8*)RExC_parse,
17250 RExC_end - RExC_parse,
17251 &numlen, UTF8_ALLOW_DEFAULT);
17252 RExC_parse += numlen;
17255 value = UCHARAT(RExC_parse++);
17257 if (value == '[') {
17258 char * posix_class_end;
17259 namedclass = handle_possible_posix(pRExC_state,
17262 do_posix_warnings ? &posix_warnings : NULL,
17263 FALSE /* die if error */);
17264 if (namedclass > OOB_NAMEDCLASS) {
17266 /* If there was an earlier attempt to parse this particular
17267 * posix class, and it failed, it was a false alarm, as this
17268 * successful one proves */
17269 if ( posix_warnings
17270 && av_tindex_skip_len_mg(posix_warnings) >= 0
17271 && not_posix_region_end >= RExC_parse
17272 && not_posix_region_end <= posix_class_end)
17274 av_undef(posix_warnings);
17277 RExC_parse = posix_class_end;
17279 else if (namedclass == OOB_NAMEDCLASS) {
17280 not_posix_region_end = posix_class_end;
17283 namedclass = OOB_NAMEDCLASS;
17286 else if ( RExC_parse - 1 > not_posix_region_end
17287 && MAYBE_POSIXCC(value))
17289 (void) handle_possible_posix(
17291 RExC_parse - 1, /* -1 because parse has already been
17293 ¬_posix_region_end,
17294 do_posix_warnings ? &posix_warnings : NULL,
17295 TRUE /* checking only */);
17297 else if ( strict && ! skip_white
17298 && ( _generic_isCC(value, _CC_VERTSPACE)
17299 || is_VERTWS_cp_high(value)))
17301 vFAIL("Literal vertical space in [] is illegal except under /x");
17303 else if (value == '\\') {
17304 /* Is a backslash; get the code point of the char after it */
17306 if (RExC_parse >= RExC_end) {
17307 vFAIL("Unmatched [");
17310 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
17311 value = utf8n_to_uvchr((U8*)RExC_parse,
17312 RExC_end - RExC_parse,
17313 &numlen, UTF8_ALLOW_DEFAULT);
17314 RExC_parse += numlen;
17317 value = UCHARAT(RExC_parse++);
17319 /* Some compilers cannot handle switching on 64-bit integer
17320 * values, therefore value cannot be an UV. Yes, this will
17321 * be a problem later if we want switch on Unicode.
17322 * A similar issue a little bit later when switching on
17323 * namedclass. --jhi */
17325 /* If the \ is escaping white space when white space is being
17326 * skipped, it means that that white space is wanted literally, and
17327 * is already in 'value'. Otherwise, need to translate the escape
17328 * into what it signifies. */
17329 if (! skip_white || ! isBLANK_A(value)) switch ((I32)value) {
17331 case 'w': namedclass = ANYOF_WORDCHAR; break;
17332 case 'W': namedclass = ANYOF_NWORDCHAR; break;
17333 case 's': namedclass = ANYOF_SPACE; break;
17334 case 'S': namedclass = ANYOF_NSPACE; break;
17335 case 'd': namedclass = ANYOF_DIGIT; break;
17336 case 'D': namedclass = ANYOF_NDIGIT; break;
17337 case 'v': namedclass = ANYOF_VERTWS; break;
17338 case 'V': namedclass = ANYOF_NVERTWS; break;
17339 case 'h': namedclass = ANYOF_HORIZWS; break;
17340 case 'H': namedclass = ANYOF_NHORIZWS; break;
17341 case 'N': /* Handle \N{NAME} in class */
17343 const char * const backslash_N_beg = RExC_parse - 2;
17346 if (! grok_bslash_N(pRExC_state,
17347 NULL, /* No regnode */
17348 &value, /* Yes single value */
17349 &cp_count, /* Multiple code pt count */
17355 if (*flagp & NEED_UTF8)
17356 FAIL("panic: grok_bslash_N set NEED_UTF8");
17358 RETURN_FAIL_ON_RESTART_FLAGP(flagp);
17360 if (cp_count < 0) {
17361 vFAIL("\\N in a character class must be a named character: \\N{...}");
17363 else if (cp_count == 0) {
17364 ckWARNreg(RExC_parse,
17365 "Ignoring zero length \\N{} in character class");
17367 else { /* cp_count > 1 */
17368 assert(cp_count > 1);
17369 if (! RExC_in_multi_char_class) {
17370 if ( ! allow_mutiple_chars
17373 || *RExC_parse == '-')
17377 vFAIL("\\N{} here is restricted to one character");
17379 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
17380 break; /* <value> contains the first code
17381 point. Drop out of the switch to
17385 SV * multi_char_N = newSVpvn(backslash_N_beg,
17386 RExC_parse - backslash_N_beg);
17388 = add_multi_match(multi_char_matches,
17393 } /* End of cp_count != 1 */
17395 /* This element should not be processed further in this
17398 value = save_value;
17399 prevvalue = save_prevvalue;
17400 continue; /* Back to top of loop to get next char */
17403 /* Here, is a single code point, and <value> contains it */
17404 unicode_range = TRUE; /* \N{} are Unicode */
17412 /* \p means they want Unicode semantics */
17413 REQUIRE_UNI_RULES(flagp, 0);
17415 if (RExC_parse >= RExC_end)
17416 vFAIL2("Empty \\%c", (U8)value);
17417 if (*RExC_parse == '{') {
17418 const U8 c = (U8)value;
17419 e = (char *) memchr(RExC_parse, '}', RExC_end - RExC_parse);
17422 vFAIL2("Missing right brace on \\%c{}", c);
17427 /* White space is allowed adjacent to the braces and after
17428 * any '^', even when not under /x */
17429 while (isSPACE(*RExC_parse)) {
17433 if (UCHARAT(RExC_parse) == '^') {
17435 /* toggle. (The rhs xor gets the single bit that
17436 * differs between P and p; the other xor inverts just
17438 value ^= 'P' ^ 'p';
17441 while (isSPACE(*RExC_parse)) {
17446 if (e == RExC_parse)
17447 vFAIL2("Empty \\%c{}", c);
17449 n = e - RExC_parse;
17450 while (isSPACE(*(RExC_parse + n - 1)))
17453 } /* The \p isn't immediately followed by a '{' */
17454 else if (! isALPHA(*RExC_parse)) {
17455 RExC_parse += (UTF)
17456 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17458 vFAIL2("Character following \\%c must be '{' or a "
17459 "single-character Unicode property name",
17467 char* name = RExC_parse;
17469 /* Any message returned about expanding the definition */
17470 SV* msg = newSVpvs_flags("", SVs_TEMP);
17472 /* If set TRUE, the property is user-defined as opposed to
17473 * official Unicode */
17474 bool user_defined = FALSE;
17476 SV * prop_definition = parse_uniprop_string(
17477 name, n, UTF, FOLD,
17478 FALSE, /* This is compile-time */
17480 /* We can't defer this defn when
17481 * the full result is required in
17483 ! cBOOL(ret_invlist),
17489 if (SvCUR(msg)) { /* Assumes any error causes a msg */
17490 assert(prop_definition == NULL);
17491 RExC_parse = e + 1;
17492 if (SvUTF8(msg)) { /* msg being UTF-8 makes the whole
17493 thing so, or else the display is
17497 /* diag_listed_as: Can't find Unicode property definition "%s" in regex; marked by <-- HERE in m/%s/ */
17498 vFAIL2utf8f("%" UTF8f, UTF8fARG(SvUTF8(msg),
17499 SvCUR(msg), SvPVX(msg)));
17502 if (! is_invlist(prop_definition)) {
17504 /* Here, the definition isn't known, so we have gotten
17505 * returned a string that will be evaluated if and when
17506 * encountered at runtime. We add it to the list of
17507 * such properties, along with whether it should be
17508 * complemented or not */
17509 if (value == 'P') {
17510 sv_catpvs(listsv, "!");
17513 sv_catpvs(listsv, "+");
17515 sv_catsv(listsv, prop_definition);
17517 has_runtime_dependency |= HAS_USER_DEFINED_PROPERTY;
17519 /* We don't know yet what this matches, so have to flag
17521 anyof_flags |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
17524 assert (prop_definition && is_invlist(prop_definition));
17526 /* Here we do have the complete property definition
17528 * Temporary workaround for [perl #133136]. For this
17529 * precise input that is in the .t that is failing,
17530 * load utf8.pm, which is what the test wants, so that
17531 * that .t passes */
17532 if ( memEQs(RExC_start, e + 1 - RExC_start,
17534 && ! hv_common(GvHVn(PL_incgv),
17536 "utf8.pm", sizeof("utf8.pm") - 1,
17537 0, HV_FETCH_ISEXISTS, NULL, 0))
17539 require_pv("utf8.pm");
17542 if (! user_defined &&
17543 /* We warn on matching an above-Unicode code point
17544 * if the match would return true, except don't
17545 * warn for \p{All}, which has exactly one element
17547 (_invlist_contains_cp(prop_definition, 0x110000)
17548 && (! (_invlist_len(prop_definition) == 1
17549 && *invlist_array(prop_definition) == 0))))
17554 /* Invert if asking for the complement */
17555 if (value == 'P') {
17556 _invlist_union_complement_2nd(properties,
17561 _invlist_union(properties, prop_definition, &properties);
17566 RExC_parse = e + 1;
17567 namedclass = ANYOF_UNIPROP; /* no official name, but it's
17571 case 'n': value = '\n'; break;
17572 case 'r': value = '\r'; break;
17573 case 't': value = '\t'; break;
17574 case 'f': value = '\f'; break;
17575 case 'b': value = '\b'; break;
17576 case 'e': value = ESC_NATIVE; break;
17577 case 'a': value = '\a'; break;
17579 RExC_parse--; /* function expects to be pointed at the 'o' */
17581 const char* error_msg;
17582 bool valid = grok_bslash_o(&RExC_parse,
17586 TO_OUTPUT_WARNINGS(RExC_parse),
17588 silence_non_portable,
17593 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17595 non_portable_endpoint++;
17598 RExC_parse--; /* function expects to be pointed at the 'x' */
17600 const char* error_msg;
17601 bool valid = grok_bslash_x(&RExC_parse,
17605 TO_OUTPUT_WARNINGS(RExC_parse),
17607 silence_non_portable,
17612 UPDATE_WARNINGS_LOC(RExC_parse - 1);
17614 non_portable_endpoint++;
17617 value = grok_bslash_c(*RExC_parse, TO_OUTPUT_WARNINGS(RExC_parse));
17618 UPDATE_WARNINGS_LOC(RExC_parse);
17620 non_portable_endpoint++;
17622 case '0': case '1': case '2': case '3': case '4':
17623 case '5': case '6': case '7':
17625 /* Take 1-3 octal digits */
17626 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
17627 numlen = (strict) ? 4 : 3;
17628 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
17629 RExC_parse += numlen;
17632 RExC_parse += (UTF)
17633 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
17635 vFAIL("Need exactly 3 octal digits");
17637 else if ( numlen < 3 /* like \08, \178 */
17638 && RExC_parse < RExC_end
17639 && isDIGIT(*RExC_parse)
17640 && ckWARN(WARN_REGEXP))
17642 reg_warn_non_literal_string(
17644 form_short_octal_warning(RExC_parse, numlen));
17647 non_portable_endpoint++;
17651 /* Allow \_ to not give an error */
17652 if (isWORDCHAR(value) && value != '_') {
17654 vFAIL2("Unrecognized escape \\%c in character class",
17658 ckWARN2reg(RExC_parse,
17659 "Unrecognized escape \\%c in character class passed through",
17664 } /* End of switch on char following backslash */
17665 } /* end of handling backslash escape sequences */
17667 /* Here, we have the current token in 'value' */
17669 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
17672 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
17673 * literal, as is the character that began the false range, i.e.
17674 * the 'a' in the examples */
17676 const int w = (RExC_parse >= rangebegin)
17677 ? RExC_parse - rangebegin
17681 "False [] range \"%" UTF8f "\"",
17682 UTF8fARG(UTF, w, rangebegin));
17685 ckWARN2reg(RExC_parse,
17686 "False [] range \"%" UTF8f "\"",
17687 UTF8fARG(UTF, w, rangebegin));
17688 cp_list = add_cp_to_invlist(cp_list, '-');
17689 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
17693 range = 0; /* this was not a true range */
17694 element_count += 2; /* So counts for three values */
17697 classnum = namedclass_to_classnum(namedclass);
17699 if (LOC && namedclass < ANYOF_POSIXL_MAX
17700 #ifndef HAS_ISASCII
17701 && classnum != _CC_ASCII
17704 SV* scratch_list = NULL;
17706 /* What the Posix classes (like \w, [:space:]) match isn't
17707 * generally knowable under locale until actual match time. A
17708 * special node is used for these which has extra space for a
17709 * bitmap, with a bit reserved for each named class that is to
17710 * be matched against. (This isn't needed for \p{} and
17711 * pseudo-classes, as they are not affected by locale, and
17712 * hence are dealt with separately.) However, if a named class
17713 * and its complement are both present, then it matches
17714 * everything, and there is no runtime dependency. Odd numbers
17715 * are the complements of the next lower number, so xor works.
17716 * (Note that something like [\w\D] should match everything,
17717 * because \d should be a proper subset of \w. But rather than
17718 * trust that the locale is well behaved, we leave this to
17719 * runtime to sort out) */
17720 if (POSIXL_TEST(posixl, namedclass ^ 1)) {
17721 cp_list = _add_range_to_invlist(cp_list, 0, UV_MAX);
17722 POSIXL_ZERO(posixl);
17723 has_runtime_dependency &= ~HAS_L_RUNTIME_DEPENDENCY;
17724 anyof_flags &= ~ANYOF_MATCHES_POSIXL;
17725 continue; /* We could ignore the rest of the class, but
17726 best to parse it for any errors */
17728 else { /* Here, isn't the complement of any already parsed
17730 POSIXL_SET(posixl, namedclass);
17731 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
17732 anyof_flags |= ANYOF_MATCHES_POSIXL;
17734 /* The above-Latin1 characters are not subject to locale
17735 * rules. Just add them to the unconditionally-matched
17738 /* Get the list of the above-Latin1 code points this
17740 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
17741 PL_XPosix_ptrs[classnum],
17743 /* Odd numbers are complements,
17744 * like NDIGIT, NASCII, ... */
17745 namedclass % 2 != 0,
17747 /* Checking if 'cp_list' is NULL first saves an extra
17748 * clone. Its reference count will be decremented at the
17749 * next union, etc, or if this is the only instance, at the
17750 * end of the routine */
17752 cp_list = scratch_list;
17755 _invlist_union(cp_list, scratch_list, &cp_list);
17756 SvREFCNT_dec_NN(scratch_list);
17758 continue; /* Go get next character */
17763 /* Here, is not /l, or is a POSIX class for which /l doesn't
17764 * matter (or is a Unicode property, which is skipped here). */
17765 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
17766 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
17768 /* Here, should be \h, \H, \v, or \V. None of /d, /i
17769 * nor /l make a difference in what these match,
17770 * therefore we just add what they match to cp_list. */
17771 if (classnum != _CC_VERTSPACE) {
17772 assert( namedclass == ANYOF_HORIZWS
17773 || namedclass == ANYOF_NHORIZWS);
17775 /* It turns out that \h is just a synonym for
17777 classnum = _CC_BLANK;
17780 _invlist_union_maybe_complement_2nd(
17782 PL_XPosix_ptrs[classnum],
17783 namedclass % 2 != 0, /* Complement if odd
17784 (NHORIZWS, NVERTWS)
17789 else if ( AT_LEAST_UNI_SEMANTICS
17790 || classnum == _CC_ASCII
17791 || (DEPENDS_SEMANTICS && ( classnum == _CC_DIGIT
17792 || classnum == _CC_XDIGIT)))
17794 /* We usually have to worry about /d affecting what POSIX
17795 * classes match, with special code needed because we won't
17796 * know until runtime what all matches. But there is no
17797 * extra work needed under /u and /a; and [:ascii:] is
17798 * unaffected by /d; and :digit: and :xdigit: don't have
17799 * runtime differences under /d. So we can special case
17800 * these, and avoid some extra work below, and at runtime.
17802 _invlist_union_maybe_complement_2nd(
17804 ((AT_LEAST_ASCII_RESTRICTED)
17805 ? PL_Posix_ptrs[classnum]
17806 : PL_XPosix_ptrs[classnum]),
17807 namedclass % 2 != 0,
17810 else { /* Garden variety class. If is NUPPER, NALPHA, ...
17811 complement and use nposixes */
17812 SV** posixes_ptr = namedclass % 2 == 0
17815 _invlist_union_maybe_complement_2nd(
17817 PL_XPosix_ptrs[classnum],
17818 namedclass % 2 != 0,
17822 } /* end of namedclass \blah */
17824 SKIP_BRACKETED_WHITE_SPACE(skip_white, RExC_parse);
17826 /* If 'range' is set, 'value' is the ending of a range--check its
17827 * validity. (If value isn't a single code point in the case of a
17828 * range, we should have figured that out above in the code that
17829 * catches false ranges). Later, we will handle each individual code
17830 * point in the range. If 'range' isn't set, this could be the
17831 * beginning of a range, so check for that by looking ahead to see if
17832 * the next real character to be processed is the range indicator--the
17837 /* For unicode ranges, we have to test that the Unicode as opposed
17838 * to the native values are not decreasing. (Above 255, there is
17839 * no difference between native and Unicode) */
17840 if (unicode_range && prevvalue < 255 && value < 255) {
17841 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
17842 goto backwards_range;
17847 if (prevvalue > value) /* b-a */ {
17852 w = RExC_parse - rangebegin;
17854 "Invalid [] range \"%" UTF8f "\"",
17855 UTF8fARG(UTF, w, rangebegin));
17856 NOT_REACHED; /* NOTREACHED */
17860 prevvalue = value; /* save the beginning of the potential range */
17861 if (! stop_at_1 /* Can't be a range if parsing just one thing */
17862 && *RExC_parse == '-')
17864 char* next_char_ptr = RExC_parse + 1;
17866 /* Get the next real char after the '-' */
17867 SKIP_BRACKETED_WHITE_SPACE(skip_white, next_char_ptr);
17869 /* If the '-' is at the end of the class (just before the ']',
17870 * it is a literal minus; otherwise it is a range */
17871 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
17872 RExC_parse = next_char_ptr;
17874 /* a bad range like \w-, [:word:]- ? */
17875 if (namedclass > OOB_NAMEDCLASS) {
17876 if (strict || ckWARN(WARN_REGEXP)) {
17877 const int w = RExC_parse >= rangebegin
17878 ? RExC_parse - rangebegin
17881 vFAIL4("False [] range \"%*.*s\"",
17886 "False [] range \"%*.*s\"",
17890 cp_list = add_cp_to_invlist(cp_list, '-');
17893 range = 1; /* yeah, it's a range! */
17894 continue; /* but do it the next time */
17899 if (namedclass > OOB_NAMEDCLASS) {
17903 /* Here, we have a single value this time through the loop, and
17904 * <prevvalue> is the beginning of the range, if any; or <value> if
17907 /* non-Latin1 code point implies unicode semantics. */
17909 REQUIRE_UNI_RULES(flagp, 0);
17912 /* Ready to process either the single value, or the completed range.
17913 * For single-valued non-inverted ranges, we consider the possibility
17914 * of multi-char folds. (We made a conscious decision to not do this
17915 * for the other cases because it can often lead to non-intuitive
17916 * results. For example, you have the peculiar case that:
17917 * "s s" =~ /^[^\xDF]+$/i => Y
17918 * "ss" =~ /^[^\xDF]+$/i => N
17920 * See [perl #89750] */
17921 if (FOLD && allow_mutiple_chars && value == prevvalue) {
17922 if ( value == LATIN_SMALL_LETTER_SHARP_S
17923 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
17926 /* Here <value> is indeed a multi-char fold. Get what it is */
17928 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
17931 UV folded = _to_uni_fold_flags(
17935 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
17936 ? FOLD_FLAGS_NOMIX_ASCII
17940 /* Here, <folded> should be the first character of the
17941 * multi-char fold of <value>, with <foldbuf> containing the
17942 * whole thing. But, if this fold is not allowed (because of
17943 * the flags), <fold> will be the same as <value>, and should
17944 * be processed like any other character, so skip the special
17946 if (folded != value) {
17948 /* Skip if we are recursed, currently parsing the class
17949 * again. Otherwise add this character to the list of
17950 * multi-char folds. */
17951 if (! RExC_in_multi_char_class) {
17952 STRLEN cp_count = utf8_length(foldbuf,
17953 foldbuf + foldlen);
17954 SV* multi_fold = sv_2mortal(newSVpvs(""));
17956 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%" UVXf "}", value);
17959 = add_multi_match(multi_char_matches,
17965 /* This element should not be processed further in this
17968 value = save_value;
17969 prevvalue = save_prevvalue;
17975 if (strict && ckWARN(WARN_REGEXP)) {
17978 /* If the range starts above 255, everything is portable and
17979 * likely to be so for any forseeable character set, so don't
17981 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
17982 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
17984 else if (prevvalue != value) {
17986 /* Under strict, ranges that stop and/or end in an ASCII
17987 * printable should have each end point be a portable value
17988 * for it (preferably like 'A', but we don't warn if it is
17989 * a (portable) Unicode name or code point), and the range
17990 * must be be all digits or all letters of the same case.
17991 * Otherwise, the range is non-portable and unclear as to
17992 * what it contains */
17993 if ( (isPRINT_A(prevvalue) || isPRINT_A(value))
17994 && ( non_portable_endpoint
17995 || ! ( (isDIGIT_A(prevvalue) && isDIGIT_A(value))
17996 || (isLOWER_A(prevvalue) && isLOWER_A(value))
17997 || (isUPPER_A(prevvalue) && isUPPER_A(value))
17999 vWARN(RExC_parse, "Ranges of ASCII printables should"
18000 " be some subset of \"0-9\","
18001 " \"A-Z\", or \"a-z\"");
18003 else if (prevvalue >= FIRST_NON_ASCII_DECIMAL_DIGIT) {
18004 SSize_t index_start;
18005 SSize_t index_final;
18007 /* But the nature of Unicode and languages mean we
18008 * can't do the same checks for above-ASCII ranges,
18009 * except in the case of digit ones. These should
18010 * contain only digits from the same group of 10. The
18011 * ASCII case is handled just above. Hence here, the
18012 * range could be a range of digits. First some
18013 * unlikely special cases. Grandfather in that a range
18014 * ending in 19DA (NEW TAI LUE THAM DIGIT ONE) is bad
18015 * if its starting value is one of the 10 digits prior
18016 * to it. This is because it is an alternate way of
18017 * writing 19D1, and some people may expect it to be in
18018 * that group. But it is bad, because it won't give
18019 * the expected results. In Unicode 5.2 it was
18020 * considered to be in that group (of 11, hence), but
18021 * this was fixed in the next version */
18023 if (UNLIKELY(value == 0x19DA && prevvalue >= 0x19D0)) {
18024 goto warn_bad_digit_range;
18026 else if (UNLIKELY( prevvalue >= 0x1D7CE
18027 && value <= 0x1D7FF))
18029 /* This is the only other case currently in Unicode
18030 * where the algorithm below fails. The code
18031 * points just above are the end points of a single
18032 * range containing only decimal digits. It is 5
18033 * different series of 0-9. All other ranges of
18034 * digits currently in Unicode are just a single
18035 * series. (And mktables will notify us if a later
18036 * Unicode version breaks this.)
18038 * If the range being checked is at most 9 long,
18039 * and the digit values represented are in
18040 * numerical order, they are from the same series.
18042 if ( value - prevvalue > 9
18043 || ((( value - 0x1D7CE) % 10)
18044 <= (prevvalue - 0x1D7CE) % 10))
18046 goto warn_bad_digit_range;
18051 /* For all other ranges of digits in Unicode, the
18052 * algorithm is just to check if both end points
18053 * are in the same series, which is the same range.
18055 index_start = _invlist_search(
18056 PL_XPosix_ptrs[_CC_DIGIT],
18059 /* Warn if the range starts and ends with a digit,
18060 * and they are not in the same group of 10. */
18061 if ( index_start >= 0
18062 && ELEMENT_RANGE_MATCHES_INVLIST(index_start)
18064 _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
18065 value)) != index_start
18066 && index_final >= 0
18067 && ELEMENT_RANGE_MATCHES_INVLIST(index_final))
18069 warn_bad_digit_range:
18070 vWARN(RExC_parse, "Ranges of digits should be"
18071 " from the same group of"
18078 if ((! range || prevvalue == value) && non_portable_endpoint) {
18079 if (isPRINT_A(value)) {
18082 if (isBACKSLASHED_PUNCT(value)) {
18083 literal[d++] = '\\';
18085 literal[d++] = (char) value;
18086 literal[d++] = '\0';
18089 "\"%.*s\" is more clearly written simply as \"%s\"",
18090 (int) (RExC_parse - rangebegin),
18095 else if (isMNEMONIC_CNTRL(value)) {
18097 "\"%.*s\" is more clearly written simply as \"%s\"",
18098 (int) (RExC_parse - rangebegin),
18100 cntrl_to_mnemonic((U8) value)
18106 /* Deal with this element of the class */
18109 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18112 /* On non-ASCII platforms, for ranges that span all of 0..255, and ones
18113 * that don't require special handling, we can just add the range like
18114 * we do for ASCII platforms */
18115 if ((UNLIKELY(prevvalue == 0) && value >= 255)
18116 || ! (prevvalue < 256
18118 || (! non_portable_endpoint
18119 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
18120 || (isUPPER_A(prevvalue)
18121 && isUPPER_A(value)))))))
18123 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18127 /* Here, requires special handling. This can be because it is a
18128 * range whose code points are considered to be Unicode, and so
18129 * must be individually translated into native, or because its a
18130 * subrange of 'A-Z' or 'a-z' which each aren't contiguous in
18131 * EBCDIC, but we have defined them to include only the "expected"
18132 * upper or lower case ASCII alphabetics. Subranges above 255 are
18133 * the same in native and Unicode, so can be added as a range */
18134 U8 start = NATIVE_TO_LATIN1(prevvalue);
18136 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
18137 for (j = start; j <= end; j++) {
18138 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
18141 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
18147 range = 0; /* this range (if it was one) is done now */
18148 } /* End of loop through all the text within the brackets */
18150 if ( posix_warnings && av_tindex_skip_len_mg(posix_warnings) >= 0) {
18151 output_posix_warnings(pRExC_state, posix_warnings);
18154 /* If anything in the class expands to more than one character, we have to
18155 * deal with them by building up a substitute parse string, and recursively
18156 * calling reg() on it, instead of proceeding */
18157 if (multi_char_matches) {
18158 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
18161 char *save_end = RExC_end;
18162 char *save_parse = RExC_parse;
18163 char *save_start = RExC_start;
18164 Size_t constructed_prefix_len = 0; /* This gives the length of the
18165 constructed portion of the
18166 substitute parse. */
18167 bool first_time = TRUE; /* First multi-char occurrence doesn't get
18172 /* Only one level of recursion allowed */
18173 assert(RExC_copy_start_in_constructed == RExC_precomp);
18175 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
18176 because too confusing */
18178 sv_catpvs(substitute_parse, "(?:");
18182 /* Look at the longest folds first */
18183 for (cp_count = av_tindex_skip_len_mg(multi_char_matches);
18188 if (av_exists(multi_char_matches, cp_count)) {
18189 AV** this_array_ptr;
18192 this_array_ptr = (AV**) av_fetch(multi_char_matches,
18194 while ((this_sequence = av_pop(*this_array_ptr)) !=
18197 if (! first_time) {
18198 sv_catpvs(substitute_parse, "|");
18200 first_time = FALSE;
18202 sv_catpv(substitute_parse, SvPVX(this_sequence));
18207 /* If the character class contains anything else besides these
18208 * multi-character folds, have to include it in recursive parsing */
18209 if (element_count) {
18210 sv_catpvs(substitute_parse, "|[");
18211 constructed_prefix_len = SvCUR(substitute_parse);
18212 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
18214 /* Put in a closing ']' only if not going off the end, as otherwise
18215 * we are adding something that really isn't there */
18216 if (RExC_parse < RExC_end) {
18217 sv_catpvs(substitute_parse, "]");
18221 sv_catpvs(substitute_parse, ")");
18224 /* This is a way to get the parse to skip forward a whole named
18225 * sequence instead of matching the 2nd character when it fails the
18227 sv_catpvs(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
18231 /* Set up the data structure so that any errors will be properly
18232 * reported. See the comments at the definition of
18233 * REPORT_LOCATION_ARGS for details */
18234 RExC_copy_start_in_input = (char *) orig_parse;
18235 RExC_start = RExC_parse = SvPV(substitute_parse, len);
18236 RExC_copy_start_in_constructed = RExC_start + constructed_prefix_len;
18237 RExC_end = RExC_parse + len;
18238 RExC_in_multi_char_class = 1;
18240 ret = reg(pRExC_state, 1, ®_flags, depth+1);
18242 *flagp |= reg_flags & (HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PARSE|NEED_UTF8);
18244 /* And restore so can parse the rest of the pattern */
18245 RExC_parse = save_parse;
18246 RExC_start = RExC_copy_start_in_constructed = RExC_copy_start_in_input = save_start;
18247 RExC_end = save_end;
18248 RExC_in_multi_char_class = 0;
18249 SvREFCNT_dec_NN(multi_char_matches);
18253 /* If folding, we calculate all characters that could fold to or from the
18254 * ones already on the list */
18255 if (cp_foldable_list) {
18257 UV start, end; /* End points of code point ranges */
18259 SV* fold_intersection = NULL;
18262 /* Our calculated list will be for Unicode rules. For locale
18263 * matching, we have to keep a separate list that is consulted at
18264 * runtime only when the locale indicates Unicode rules (and we
18265 * don't include potential matches in the ASCII/Latin1 range, as
18266 * any code point could fold to any other, based on the run-time
18267 * locale). For non-locale, we just use the general list */
18269 use_list = &only_utf8_locale_list;
18272 use_list = &cp_list;
18275 /* Only the characters in this class that participate in folds need
18276 * be checked. Get the intersection of this class and all the
18277 * possible characters that are foldable. This can quickly narrow
18278 * down a large class */
18279 _invlist_intersection(PL_in_some_fold, cp_foldable_list,
18280 &fold_intersection);
18282 /* Now look at the foldable characters in this class individually */
18283 invlist_iterinit(fold_intersection);
18284 while (invlist_iternext(fold_intersection, &start, &end)) {
18288 /* Look at every character in the range */
18289 for (j = start; j <= end; j++) {
18290 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
18293 Size_t folds_count;
18294 unsigned int first_fold;
18295 const unsigned int * remaining_folds;
18299 /* Under /l, we don't know what code points below 256
18300 * fold to, except we do know the MICRO SIGN folds to
18301 * an above-255 character if the locale is UTF-8, so we
18302 * add it to the special list (in *use_list) Otherwise
18303 * we know now what things can match, though some folds
18304 * are valid under /d only if the target is UTF-8.
18305 * Those go in a separate list */
18306 if ( IS_IN_SOME_FOLD_L1(j)
18307 && ! (LOC && j != MICRO_SIGN))
18310 /* ASCII is always matched; non-ASCII is matched
18311 * only under Unicode rules (which could happen
18312 * under /l if the locale is a UTF-8 one */
18313 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
18314 *use_list = add_cp_to_invlist(*use_list,
18315 PL_fold_latin1[j]);
18317 else if (j != PL_fold_latin1[j]) {
18318 upper_latin1_only_utf8_matches
18319 = add_cp_to_invlist(
18320 upper_latin1_only_utf8_matches,
18321 PL_fold_latin1[j]);
18325 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
18326 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
18328 add_above_Latin1_folds(pRExC_state,
18335 /* Here is an above Latin1 character. We don't have the
18336 * rules hard-coded for it. First, get its fold. This is
18337 * the simple fold, as the multi-character folds have been
18338 * handled earlier and separated out */
18339 folded = _to_uni_fold_flags(j, foldbuf, &foldlen,
18340 (ASCII_FOLD_RESTRICTED)
18341 ? FOLD_FLAGS_NOMIX_ASCII
18344 /* Single character fold of above Latin1. Add everything
18345 * in its fold closure to the list that this node should
18347 folds_count = _inverse_folds(folded, &first_fold,
18349 for (k = 0; k <= folds_count; k++) {
18350 UV c = (k == 0) /* First time through use itself */
18352 : (k == 1) /* 2nd time use, the first fold */
18355 /* Then the remaining ones */
18356 : remaining_folds[k-2];
18358 /* /aa doesn't allow folds between ASCII and non- */
18359 if (( ASCII_FOLD_RESTRICTED
18360 && (isASCII(c) != isASCII(j))))
18365 /* Folds under /l which cross the 255/256 boundary are
18366 * added to a separate list. (These are valid only
18367 * when the locale is UTF-8.) */
18368 if (c < 256 && LOC) {
18369 *use_list = add_cp_to_invlist(*use_list, c);
18373 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
18375 cp_list = add_cp_to_invlist(cp_list, c);
18378 /* Similarly folds involving non-ascii Latin1
18379 * characters under /d are added to their list */
18380 upper_latin1_only_utf8_matches
18381 = add_cp_to_invlist(
18382 upper_latin1_only_utf8_matches,
18388 SvREFCNT_dec_NN(fold_intersection);
18391 /* Now that we have finished adding all the folds, there is no reason
18392 * to keep the foldable list separate */
18393 _invlist_union(cp_list, cp_foldable_list, &cp_list);
18394 SvREFCNT_dec_NN(cp_foldable_list);
18397 /* And combine the result (if any) with any inversion lists from posix
18398 * classes. The lists are kept separate up to now because we don't want to
18399 * fold the classes */
18400 if (simple_posixes) { /* These are the classes known to be unaffected by
18403 _invlist_union(cp_list, simple_posixes, &cp_list);
18404 SvREFCNT_dec_NN(simple_posixes);
18407 cp_list = simple_posixes;
18410 if (posixes || nposixes) {
18411 if (! DEPENDS_SEMANTICS) {
18413 /* For everything but /d, we can just add the current 'posixes' and
18414 * 'nposixes' to the main list */
18417 _invlist_union(cp_list, posixes, &cp_list);
18418 SvREFCNT_dec_NN(posixes);
18426 _invlist_union(cp_list, nposixes, &cp_list);
18427 SvREFCNT_dec_NN(nposixes);
18430 cp_list = nposixes;
18435 /* Under /d, things like \w match upper Latin1 characters only if
18436 * the target string is in UTF-8. But things like \W match all the
18437 * upper Latin1 characters if the target string is not in UTF-8.
18439 * Handle the case with something like \W separately */
18441 SV* only_non_utf8_list = invlist_clone(PL_UpperLatin1, NULL);
18443 /* A complemented posix class matches all upper Latin1
18444 * characters if not in UTF-8. And it matches just certain
18445 * ones when in UTF-8. That means those certain ones are
18446 * matched regardless, so can just be added to the
18447 * unconditional list */
18449 _invlist_union(cp_list, nposixes, &cp_list);
18450 SvREFCNT_dec_NN(nposixes);
18454 cp_list = nposixes;
18457 /* Likewise for 'posixes' */
18458 _invlist_union(posixes, cp_list, &cp_list);
18459 SvREFCNT_dec(posixes);
18461 /* Likewise for anything else in the range that matched only
18463 if (upper_latin1_only_utf8_matches) {
18464 _invlist_union(cp_list,
18465 upper_latin1_only_utf8_matches,
18467 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18468 upper_latin1_only_utf8_matches = NULL;
18471 /* If we don't match all the upper Latin1 characters regardless
18472 * of UTF-8ness, we have to set a flag to match the rest when
18474 _invlist_subtract(only_non_utf8_list, cp_list,
18475 &only_non_utf8_list);
18476 if (_invlist_len(only_non_utf8_list) != 0) {
18477 anyof_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18479 SvREFCNT_dec_NN(only_non_utf8_list);
18482 /* Here there were no complemented posix classes. That means
18483 * the upper Latin1 characters in 'posixes' match only when the
18484 * target string is in UTF-8. So we have to add them to the
18485 * list of those types of code points, while adding the
18486 * remainder to the unconditional list.
18488 * First calculate what they are */
18489 SV* nonascii_but_latin1_properties = NULL;
18490 _invlist_intersection(posixes, PL_UpperLatin1,
18491 &nonascii_but_latin1_properties);
18493 /* And add them to the final list of such characters. */
18494 _invlist_union(upper_latin1_only_utf8_matches,
18495 nonascii_but_latin1_properties,
18496 &upper_latin1_only_utf8_matches);
18498 /* Remove them from what now becomes the unconditional list */
18499 _invlist_subtract(posixes, nonascii_but_latin1_properties,
18502 /* And add those unconditional ones to the final list */
18504 _invlist_union(cp_list, posixes, &cp_list);
18505 SvREFCNT_dec_NN(posixes);
18512 SvREFCNT_dec(nonascii_but_latin1_properties);
18514 /* Get rid of any characters from the conditional list that we
18515 * now know are matched unconditionally, which may make that
18517 _invlist_subtract(upper_latin1_only_utf8_matches,
18519 &upper_latin1_only_utf8_matches);
18520 if (_invlist_len(upper_latin1_only_utf8_matches) == 0) {
18521 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
18522 upper_latin1_only_utf8_matches = NULL;
18528 /* And combine the result (if any) with any inversion list from properties.
18529 * The lists are kept separate up to now so that we can distinguish the two
18530 * in regards to matching above-Unicode. A run-time warning is generated
18531 * if a Unicode property is matched against a non-Unicode code point. But,
18532 * we allow user-defined properties to match anything, without any warning,
18533 * and we also suppress the warning if there is a portion of the character
18534 * class that isn't a Unicode property, and which matches above Unicode, \W
18535 * or [\x{110000}] for example.
18536 * (Note that in this case, unlike the Posix one above, there is no
18537 * <upper_latin1_only_utf8_matches>, because having a Unicode property
18538 * forces Unicode semantics */
18542 /* If it matters to the final outcome, see if a non-property
18543 * component of the class matches above Unicode. If so, the
18544 * warning gets suppressed. This is true even if just a single
18545 * such code point is specified, as, though not strictly correct if
18546 * another such code point is matched against, the fact that they
18547 * are using above-Unicode code points indicates they should know
18548 * the issues involved */
18550 warn_super = ! (invert
18551 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
18554 _invlist_union(properties, cp_list, &cp_list);
18555 SvREFCNT_dec_NN(properties);
18558 cp_list = properties;
18563 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
18565 /* Because an ANYOF node is the only one that warns, this node
18566 * can't be optimized into something else */
18567 optimizable = FALSE;
18571 /* Here, we have calculated what code points should be in the character
18574 * Now we can see about various optimizations. Fold calculation (which we
18575 * did above) needs to take place before inversion. Otherwise /[^k]/i
18576 * would invert to include K, which under /i would match k, which it
18577 * shouldn't. Therefore we can't invert folded locale now, as it won't be
18578 * folded until runtime */
18580 /* If we didn't do folding, it's because some information isn't available
18581 * until runtime; set the run-time fold flag for these We know to set the
18582 * flag if we have a non-NULL list for UTF-8 locales, or the class matches
18583 * at least one 0-255 range code point */
18586 /* Some things on the list might be unconditionally included because of
18587 * other components. Remove them, and clean up the list if it goes to
18589 if (only_utf8_locale_list && cp_list) {
18590 _invlist_subtract(only_utf8_locale_list, cp_list,
18591 &only_utf8_locale_list);
18593 if (_invlist_len(only_utf8_locale_list) == 0) {
18594 SvREFCNT_dec_NN(only_utf8_locale_list);
18595 only_utf8_locale_list = NULL;
18598 if ( only_utf8_locale_list
18599 || (cp_list && ( _invlist_contains_cp(cp_list, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE)
18600 || _invlist_contains_cp(cp_list, LATIN_SMALL_LETTER_DOTLESS_I))))
18602 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18605 | ANYOFL_SHARED_UTF8_LOCALE_fold_HAS_MATCHES_nonfold_REQD;
18607 else if (cp_list && invlist_lowest(cp_list) < 256) {
18608 /* If nothing is below 256, has no locale dependency; otherwise it
18610 anyof_flags |= ANYOFL_FOLD;
18611 has_runtime_dependency |= HAS_L_RUNTIME_DEPENDENCY;
18614 else if ( DEPENDS_SEMANTICS
18615 && ( upper_latin1_only_utf8_matches
18616 || (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)))
18618 RExC_seen_d_op = TRUE;
18619 has_runtime_dependency |= HAS_D_RUNTIME_DEPENDENCY;
18622 /* Optimize inverted patterns (e.g. [^a-z]) when everything is known at
18626 && ! has_runtime_dependency)
18628 _invlist_invert(cp_list);
18630 /* Clear the invert flag since have just done it here */
18635 *ret_invlist = cp_list;
18640 /* All possible optimizations below still have these characteristics.
18641 * (Multi-char folds aren't SIMPLE, but they don't get this far in this
18643 *flagp |= HASWIDTH|SIMPLE;
18645 if (anyof_flags & ANYOF_LOCALE_FLAGS) {
18646 RExC_contains_locale = 1;
18649 /* Some character classes are equivalent to other nodes. Such nodes take
18650 * up less room, and some nodes require fewer operations to execute, than
18651 * ANYOF nodes. EXACTish nodes may be joinable with adjacent nodes to
18652 * improve efficiency. */
18655 PERL_UINT_FAST8_T i;
18656 UV partial_cp_count = 0;
18657 UV start[MAX_FOLD_FROMS+1] = { 0 }; /* +1 for the folded-to char */
18658 UV end[MAX_FOLD_FROMS+1] = { 0 };
18659 bool single_range = FALSE;
18661 if (cp_list) { /* Count the code points in enough ranges that we would
18662 see all the ones possible in any fold in this version
18665 invlist_iterinit(cp_list);
18666 for (i = 0; i <= MAX_FOLD_FROMS; i++) {
18667 if (! invlist_iternext(cp_list, &start[i], &end[i])) {
18670 partial_cp_count += end[i] - start[i] + 1;
18674 single_range = TRUE;
18676 invlist_iterfinish(cp_list);
18679 /* If we know at compile time that this matches every possible code
18680 * point, any run-time dependencies don't matter */
18681 if (start[0] == 0 && end[0] == UV_MAX) {
18683 ret = reganode(pRExC_state, OPFAIL, 0);
18686 ret = reg_node(pRExC_state, SANY);
18692 /* Similarly, for /l posix classes, if both a class and its
18693 * complement match, any run-time dependencies don't matter */
18695 for (namedclass = 0; namedclass < ANYOF_POSIXL_MAX;
18698 if ( POSIXL_TEST(posixl, namedclass) /* class */
18699 && POSIXL_TEST(posixl, namedclass + 1)) /* its complement */
18702 ret = reganode(pRExC_state, OPFAIL, 0);
18705 ret = reg_node(pRExC_state, SANY);
18712 /* For well-behaved locales, some classes are subsets of others,
18713 * so complementing the subset and including the non-complemented
18714 * superset should match everything, like [\D[:alnum:]], and
18715 * [[:^alpha:][:alnum:]], but some implementations of locales are
18716 * buggy, and khw thinks its a bad idea to have optimization change
18717 * behavior, even if it avoids an OS bug in a given case */
18719 #define isSINGLE_BIT_SET(n) isPOWER_OF_2(n)
18721 /* If is a single posix /l class, can optimize to just that op.
18722 * Such a node will not match anything in the Latin1 range, as that
18723 * is not determinable until runtime, but will match whatever the
18724 * class does outside that range. (Note that some classes won't
18725 * match anything outside the range, like [:ascii:]) */
18726 if ( isSINGLE_BIT_SET(posixl)
18727 && (partial_cp_count == 0 || start[0] > 255))
18730 SV * class_above_latin1 = NULL;
18731 bool already_inverted;
18732 bool are_equivalent;
18734 /* Compute which bit is set, which is the same thing as, e.g.,
18735 * ANYOF_CNTRL. From
18736 * https://graphics.stanford.edu/~seander/bithacks.html#IntegerLogDeBruijn
18738 static const int MultiplyDeBruijnBitPosition2[32] =
18740 0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
18741 31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
18744 namedclass = MultiplyDeBruijnBitPosition2[(posixl
18745 * 0x077CB531U) >> 27];
18746 classnum = namedclass_to_classnum(namedclass);
18748 /* The named classes are such that the inverted number is one
18749 * larger than the non-inverted one */
18750 already_inverted = namedclass
18751 - classnum_to_namedclass(classnum);
18753 /* Create an inversion list of the official property, inverted
18754 * if the constructed node list is inverted, and restricted to
18755 * only the above latin1 code points, which are the only ones
18756 * known at compile time */
18757 _invlist_intersection_maybe_complement_2nd(
18759 PL_XPosix_ptrs[classnum],
18761 &class_above_latin1);
18762 are_equivalent = _invlistEQ(class_above_latin1, cp_list,
18764 SvREFCNT_dec_NN(class_above_latin1);
18766 if (are_equivalent) {
18768 /* Resolve the run-time inversion flag with this possibly
18769 * inverted class */
18770 invert = invert ^ already_inverted;
18772 ret = reg_node(pRExC_state,
18773 POSIXL + invert * (NPOSIXL - POSIXL));
18774 FLAGS(REGNODE_p(ret)) = classnum;
18780 /* khw can't think of any other possible transformation involving
18782 if (has_runtime_dependency & HAS_USER_DEFINED_PROPERTY) {
18786 if (! has_runtime_dependency) {
18788 /* If the list is empty, nothing matches. This happens, for
18789 * example, when a Unicode property that doesn't match anything is
18790 * the only element in the character class (perluniprops.pod notes
18791 * such properties). */
18792 if (partial_cp_count == 0) {
18794 ret = reg_node(pRExC_state, SANY);
18797 ret = reganode(pRExC_state, OPFAIL, 0);
18803 /* If matches everything but \n */
18804 if ( start[0] == 0 && end[0] == '\n' - 1
18805 && start[1] == '\n' + 1 && end[1] == UV_MAX)
18808 ret = reg_node(pRExC_state, REG_ANY);
18814 /* Next see if can optimize classes that contain just a few code points
18815 * into an EXACTish node. The reason to do this is to let the
18816 * optimizer join this node with adjacent EXACTish ones, and ANYOF
18817 * nodes require conversion to code point from UTF-8.
18819 * An EXACTFish node can be generated even if not under /i, and vice
18820 * versa. But care must be taken. An EXACTFish node has to be such
18821 * that it only matches precisely the code points in the class, but we
18822 * want to generate the least restrictive one that does that, to
18823 * increase the odds of being able to join with an adjacent node. For
18824 * example, if the class contains [kK], we have to make it an EXACTFAA
18825 * node to prevent the KELVIN SIGN from matching. Whether we are under
18826 * /i or not is irrelevant in this case. Less obvious is the pattern
18827 * qr/[\x{02BC}]n/i. U+02BC is MODIFIER LETTER APOSTROPHE. That is
18828 * supposed to match the single character U+0149 LATIN SMALL LETTER N
18829 * PRECEDED BY APOSTROPHE. And so even though there is no simple fold
18830 * that includes \X{02BC}, there is a multi-char fold that does, and so
18831 * the node generated for it must be an EXACTFish one. On the other
18832 * hand qr/:/i should generate a plain EXACT node since the colon
18833 * participates in no fold whatsoever, and having it EXACT tells the
18834 * optimizer the target string cannot match unless it has a colon in
18840 /* Only try if there are no more code points in the class than
18841 * in the max possible fold */
18842 && inRANGE(partial_cp_count, 1, MAX_FOLD_FROMS + 1))
18844 if (partial_cp_count == 1 && ! upper_latin1_only_utf8_matches)
18846 /* We can always make a single code point class into an
18847 * EXACTish node. */
18851 /* Here is /l: Use EXACTL, except if there is a fold not
18852 * known until runtime so shows as only a single code point
18853 * here. For code points above 255, we know which can
18854 * cause problems by having a potential fold to the Latin1
18857 || ( start[0] > 255
18858 && ! is_PROBLEMATIC_LOCALE_FOLD_cp(start[0])))
18866 else if (! FOLD) { /* Not /l and not /i */
18867 op = (start[0] < 256) ? EXACT : EXACT_REQ8;
18869 else if (start[0] < 256) { /* /i, not /l, and the code point is
18872 /* Under /i, it gets a little tricky. A code point that
18873 * doesn't participate in a fold should be an EXACT node.
18874 * We know this one isn't the result of a simple fold, or
18875 * there'd be more than one code point in the list, but it
18876 * could be part of a multi- character fold. In that case
18877 * we better not create an EXACT node, as we would wrongly
18878 * be telling the optimizer that this code point must be in
18879 * the target string, and that is wrong. This is because
18880 * if the sequence around this code point forms a
18881 * multi-char fold, what needs to be in the string could be
18882 * the code point that folds to the sequence.
18884 * This handles the case of below-255 code points, as we
18885 * have an easy look up for those. The next clause handles
18886 * the above-256 one */
18887 op = IS_IN_SOME_FOLD_L1(start[0])
18891 else { /* /i, larger code point. Since we are under /i, and
18892 have just this code point, we know that it can't
18893 fold to something else, so PL_InMultiCharFold
18895 op = _invlist_contains_cp(PL_InMultiCharFold,
18903 else if ( ! (has_runtime_dependency & ~HAS_D_RUNTIME_DEPENDENCY)
18904 && _invlist_contains_cp(PL_in_some_fold, start[0]))
18906 /* Here, the only runtime dependency, if any, is from /d, and
18907 * the class matches more than one code point, and the lowest
18908 * code point participates in some fold. It might be that the
18909 * other code points are /i equivalent to this one, and hence
18910 * they would representable by an EXACTFish node. Above, we
18911 * eliminated classes that contain too many code points to be
18912 * EXACTFish, with the test for MAX_FOLD_FROMS
18914 * First, special case the ASCII fold pairs, like 'B' and 'b'.
18915 * We do this because we have EXACTFAA at our disposal for the
18917 if (partial_cp_count == 2 && isASCII(start[0])) {
18919 /* The only ASCII characters that participate in folds are
18921 assert(isALPHA(start[0]));
18922 if ( end[0] == start[0] /* First range is a single
18923 character, so 2nd exists */
18924 && isALPHA_FOLD_EQ(start[0], start[1]))
18927 /* Here, is part of an ASCII fold pair */
18929 if ( ASCII_FOLD_RESTRICTED
18930 || HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(start[0]))
18932 /* If the second clause just above was true, it
18933 * means we can't be under /i, or else the list
18934 * would have included more than this fold pair.
18935 * Therefore we have to exclude the possibility of
18936 * whatever else it is that folds to these, by
18937 * using EXACTFAA */
18940 else if (HAS_NONLATIN1_FOLD_CLOSURE(start[0])) {
18942 /* Here, there's no simple fold that start[0] is part
18943 * of, but there is a multi-character one. If we
18944 * are not under /i, we want to exclude that
18945 * possibility; if under /i, we want to include it
18947 op = (FOLD) ? EXACTFU : EXACTFAA;
18951 /* Here, the only possible fold start[0] particpates in
18952 * is with start[1]. /i or not isn't relevant */
18956 value = toFOLD(start[0]);
18959 else if ( ! upper_latin1_only_utf8_matches
18960 || ( _invlist_len(upper_latin1_only_utf8_matches)
18963 invlist_highest(upper_latin1_only_utf8_matches)]
18966 /* Here, the smallest character is non-ascii or there are
18967 * more than 2 code points matched by this node. Also, we
18968 * either don't have /d UTF-8 dependent matches, or if we
18969 * do, they look like they could be a single character that
18970 * is the fold of the lowest one in the always-match list.
18971 * This test quickly excludes most of the false positives
18972 * when there are /d UTF-8 depdendent matches. These are
18973 * like LATIN CAPITAL LETTER A WITH GRAVE matching LATIN
18974 * SMALL LETTER A WITH GRAVE iff the target string is
18975 * UTF-8. (We don't have to worry above about exceeding
18976 * the array bounds of PL_fold_latin1[] because any code
18977 * point in 'upper_latin1_only_utf8_matches' is below 256.)
18979 * EXACTFAA would apply only to pairs (hence exactly 2 code
18980 * points) in the ASCII range, so we can't use it here to
18981 * artificially restrict the fold domain, so we check if
18982 * the class does or does not match some EXACTFish node.
18983 * Further, if we aren't under /i, and and the folded-to
18984 * character is part of a multi-character fold, we can't do
18985 * this optimization, as the sequence around it could be
18986 * that multi-character fold, and we don't here know the
18987 * context, so we have to assume it is that multi-char
18988 * fold, to prevent potential bugs.
18990 * To do the general case, we first find the fold of the
18991 * lowest code point (which may be higher than the lowest
18992 * one), then find everything that folds to it. (The data
18993 * structure we have only maps from the folded code points,
18994 * so we have to do the earlier step.) */
18997 U8 foldbuf[UTF8_MAXBYTES_CASE];
18998 UV folded = _to_uni_fold_flags(start[0],
18999 foldbuf, &foldlen, 0);
19000 unsigned int first_fold;
19001 const unsigned int * remaining_folds;
19002 Size_t folds_to_this_cp_count = _inverse_folds(
19006 Size_t folds_count = folds_to_this_cp_count + 1;
19007 SV * fold_list = _new_invlist(folds_count);
19010 /* If there are UTF-8 dependent matches, create a temporary
19011 * list of what this node matches, including them. */
19012 SV * all_cp_list = NULL;
19013 SV ** use_this_list = &cp_list;
19015 if (upper_latin1_only_utf8_matches) {
19016 all_cp_list = _new_invlist(0);
19017 use_this_list = &all_cp_list;
19018 _invlist_union(cp_list,
19019 upper_latin1_only_utf8_matches,
19023 /* Having gotten everything that participates in the fold
19024 * containing the lowest code point, we turn that into an
19025 * inversion list, making sure everything is included. */
19026 fold_list = add_cp_to_invlist(fold_list, start[0]);
19027 fold_list = add_cp_to_invlist(fold_list, folded);
19028 if (folds_to_this_cp_count > 0) {
19029 fold_list = add_cp_to_invlist(fold_list, first_fold);
19030 for (i = 0; i + 1 < folds_to_this_cp_count; i++) {
19031 fold_list = add_cp_to_invlist(fold_list,
19032 remaining_folds[i]);
19036 /* If the fold list is identical to what's in this ANYOF
19037 * node, the node can be represented by an EXACTFish one
19039 if (_invlistEQ(*use_this_list, fold_list,
19040 0 /* Don't complement */ )
19043 /* But, we have to be careful, as mentioned above.
19044 * Just the right sequence of characters could match
19045 * this if it is part of a multi-character fold. That
19046 * IS what we want if we are under /i. But it ISN'T
19047 * what we want if not under /i, as it could match when
19048 * it shouldn't. So, when we aren't under /i and this
19049 * character participates in a multi-char fold, we
19050 * don't optimize into an EXACTFish node. So, for each
19051 * case below we have to check if we are folding
19052 * and if not, if it is not part of a multi-char fold.
19054 if (start[0] > 255) { /* Highish code point */
19055 if (FOLD || ! _invlist_contains_cp(
19056 PL_InMultiCharFold, folded))
19060 : (ASCII_FOLD_RESTRICTED)
19065 } /* Below, the lowest code point < 256 */
19068 && DEPENDS_SEMANTICS)
19069 { /* An EXACTF node containing a single character
19070 's', can be an EXACTFU if it doesn't get
19071 joined with an adjacent 's' */
19072 op = EXACTFU_S_EDGE;
19076 || ! HAS_NONLATIN1_FOLD_CLOSURE(start[0]))
19078 if (upper_latin1_only_utf8_matches) {
19081 /* We can't use the fold, as that only matches
19085 else if ( UNLIKELY(start[0] == MICRO_SIGN)
19087 { /* EXACTFUP is a special node for this
19089 op = (ASCII_FOLD_RESTRICTED)
19092 value = MICRO_SIGN;
19094 else if ( ASCII_FOLD_RESTRICTED
19095 && ! isASCII(start[0]))
19096 { /* For ASCII under /iaa, we can use EXACTFU
19108 SvREFCNT_dec_NN(fold_list);
19109 SvREFCNT_dec(all_cp_list);
19116 /* Here, we have calculated what EXACTish node to use. Have to
19117 * convert to UTF-8 if not already there */
19120 SvREFCNT_dec(cp_list);;
19121 REQUIRE_UTF8(flagp);
19124 /* This is a kludge to the special casing issues with this
19125 * ligature under /aa. FB05 should fold to FB06, but the
19126 * call above to _to_uni_fold_flags() didn't find this, as
19127 * it didn't use the /aa restriction in order to not miss
19128 * other folds that would be affected. This is the only
19129 * instance likely to ever be a problem in all of Unicode.
19130 * So special case it. */
19131 if ( value == LATIN_SMALL_LIGATURE_LONG_S_T
19132 && ASCII_FOLD_RESTRICTED)
19134 value = LATIN_SMALL_LIGATURE_ST;
19138 len = (UTF) ? UVCHR_SKIP(value) : 1;
19140 ret = regnode_guts(pRExC_state, op, len, "exact");
19141 FILL_NODE(ret, op);
19142 RExC_emit += 1 + STR_SZ(len);
19143 setSTR_LEN(REGNODE_p(ret), len);
19145 *STRINGs(REGNODE_p(ret)) = (U8) value;
19148 uvchr_to_utf8((U8 *) STRINGs(REGNODE_p(ret)), value);
19154 if (! has_runtime_dependency) {
19156 /* See if this can be turned into an ANYOFM node. Think about the
19157 * bit patterns in two different bytes. In some positions, the
19158 * bits in each will be 1; and in other positions both will be 0;
19159 * and in some positions the bit will be 1 in one byte, and 0 in
19160 * the other. Let 'n' be the number of positions where the bits
19161 * differ. We create a mask which has exactly 'n' 0 bits, each in
19162 * a position where the two bytes differ. Now take the set of all
19163 * bytes that when ANDed with the mask yield the same result. That
19164 * set has 2**n elements, and is representable by just two 8 bit
19165 * numbers: the result and the mask. Importantly, matching the set
19166 * can be vectorized by creating a word full of the result bytes,
19167 * and a word full of the mask bytes, yielding a significant speed
19168 * up. Here, see if this node matches such a set. As a concrete
19169 * example consider [01], and the byte representing '0' which is
19170 * 0x30 on ASCII machines. It has the bits 0011 0000. Take the
19171 * mask 1111 1110. If we AND 0x31 and 0x30 with that mask we get
19172 * 0x30. Any other bytes ANDed yield something else. So [01],
19173 * which is a common usage, is optimizable into ANYOFM, and can
19174 * benefit from the speed up. We can only do this on UTF-8
19175 * invariant bytes, because they have the same bit patterns under
19177 PERL_UINT_FAST8_T inverted = 0;
19179 const PERL_UINT_FAST8_T max_permissible = 0xFF;
19181 const PERL_UINT_FAST8_T max_permissible = 0x7F;
19183 /* If doesn't fit the criteria for ANYOFM, invert and try again.
19184 * If that works we will instead later generate an NANYOFM, and
19185 * invert back when through */
19186 if (invlist_highest(cp_list) > max_permissible) {
19187 _invlist_invert(cp_list);
19191 if (invlist_highest(cp_list) <= max_permissible) {
19192 UV this_start, this_end;
19193 UV lowest_cp = UV_MAX; /* init'ed to suppress compiler warn */
19194 U8 bits_differing = 0;
19195 Size_t full_cp_count = 0;
19196 bool first_time = TRUE;
19198 /* Go through the bytes and find the bit positions that differ
19200 invlist_iterinit(cp_list);
19201 while (invlist_iternext(cp_list, &this_start, &this_end)) {
19202 unsigned int i = this_start;
19205 if (! UVCHR_IS_INVARIANT(i)) {
19209 first_time = FALSE;
19210 lowest_cp = this_start;
19212 /* We have set up the code point to compare with.
19213 * Don't compare it with itself */
19217 /* Find the bit positions that differ from the lowest code
19218 * point in the node. Keep track of all such positions by
19220 for (; i <= this_end; i++) {
19221 if (! UVCHR_IS_INVARIANT(i)) {
19225 bits_differing |= i ^ lowest_cp;
19228 full_cp_count += this_end - this_start + 1;
19231 /* At the end of the loop, we count how many bits differ from
19232 * the bits in lowest code point, call the count 'd'. If the
19233 * set we found contains 2**d elements, it is the closure of
19234 * all code points that differ only in those bit positions. To
19235 * convince yourself of that, first note that the number in the
19236 * closure must be a power of 2, which we test for. The only
19237 * way we could have that count and it be some differing set,
19238 * is if we got some code points that don't differ from the
19239 * lowest code point in any position, but do differ from each
19240 * other in some other position. That means one code point has
19241 * a 1 in that position, and another has a 0. But that would
19242 * mean that one of them differs from the lowest code point in
19243 * that position, which possibility we've already excluded. */
19244 if ( (inverted || full_cp_count > 1)
19245 && full_cp_count == 1U << PL_bitcount[bits_differing])
19249 op = ANYOFM + inverted;;
19251 /* We need to make the bits that differ be 0's */
19252 ANYOFM_mask = ~ bits_differing; /* This goes into FLAGS */
19254 /* The argument is the lowest code point */
19255 ret = reganode(pRExC_state, op, lowest_cp);
19256 FLAGS(REGNODE_p(ret)) = ANYOFM_mask;
19260 invlist_iterfinish(cp_list);
19264 _invlist_invert(cp_list);
19271 /* XXX We could create an ANYOFR_LOW node here if we saved above if
19272 * all were invariants, it wasn't inverted, and there is a single
19273 * range. This would be faster than some of the posix nodes we
19274 * create below like /\d/a, but would be twice the size. Without
19275 * having actually measured the gain, khw doesn't think the
19276 * tradeoff is really worth it */
19279 if (! (anyof_flags & ANYOF_LOCALE_FLAGS)) {
19280 PERL_UINT_FAST8_T type;
19281 SV * intersection = NULL;
19282 SV* d_invlist = NULL;
19284 /* See if this matches any of the POSIX classes. The POSIXA and
19285 * POSIXD ones are about the same speed as ANYOF ops, but take less
19286 * room; the ones that have above-Latin1 code point matches are
19287 * somewhat faster than ANYOF. */
19289 for (type = POSIXA; type >= POSIXD; type--) {
19292 if (type == POSIXL) { /* But not /l posix classes */
19296 for (posix_class = 0;
19297 posix_class <= _HIGHEST_REGCOMP_DOT_H_SYNC;
19300 SV** our_code_points = &cp_list;
19301 SV** official_code_points;
19304 if (type == POSIXA) {
19305 official_code_points = &PL_Posix_ptrs[posix_class];
19308 official_code_points = &PL_XPosix_ptrs[posix_class];
19311 /* Skip non-existent classes of this type. e.g. \v only
19312 * has an entry in PL_XPosix_ptrs */
19313 if (! *official_code_points) {
19317 /* Try both the regular class, and its inversion */
19318 for (try_inverted = 0; try_inverted < 2; try_inverted++) {
19319 bool this_inverted = invert ^ try_inverted;
19321 if (type != POSIXD) {
19323 /* This class that isn't /d can't match if we have
19324 * /d dependencies */
19325 if (has_runtime_dependency
19326 & HAS_D_RUNTIME_DEPENDENCY)
19331 else /* is /d */ if (! this_inverted) {
19333 /* /d classes don't match anything non-ASCII below
19334 * 256 unconditionally (which cp_list contains) */
19335 _invlist_intersection(cp_list, PL_UpperLatin1,
19337 if (_invlist_len(intersection) != 0) {
19341 SvREFCNT_dec(d_invlist);
19342 d_invlist = invlist_clone(cp_list, NULL);
19344 /* But under UTF-8 it turns into using /u rules.
19345 * Add the things it matches under these conditions
19346 * so that we check below that these are identical
19347 * to what the tested class should match */
19348 if (upper_latin1_only_utf8_matches) {
19351 upper_latin1_only_utf8_matches,
19354 our_code_points = &d_invlist;
19356 else { /* POSIXD, inverted. If this doesn't have this
19357 flag set, it isn't /d. */
19358 if (! (anyof_flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
19362 our_code_points = &cp_list;
19365 /* Here, have weeded out some things. We want to see
19366 * if the list of characters this node contains
19367 * ('*our_code_points') precisely matches those of the
19368 * class we are currently checking against
19369 * ('*official_code_points'). */
19370 if (_invlistEQ(*our_code_points,
19371 *official_code_points,
19374 /* Here, they precisely match. Optimize this ANYOF
19375 * node into its equivalent POSIX one of the
19376 * correct type, possibly inverted */
19377 ret = reg_node(pRExC_state, (try_inverted)
19381 FLAGS(REGNODE_p(ret)) = posix_class;
19382 SvREFCNT_dec(d_invlist);
19383 SvREFCNT_dec(intersection);
19389 SvREFCNT_dec(d_invlist);
19390 SvREFCNT_dec(intersection);
19393 /* If it is a single contiguous range, ANYOFR is an efficient regnode,
19394 * both in size and speed. Currently, a 20 bit range base (smallest
19395 * code point in the range), and a 12 bit maximum delta are packed into
19396 * a 32 bit word. This allows for using it on all of the Unicode code
19397 * points except for the highest plane, which is only for private use
19398 * code points. khw doubts that a bigger delta is likely in real world
19401 && ! has_runtime_dependency
19402 && anyof_flags == 0
19403 && start[0] < (1 << ANYOFR_BASE_BITS)
19404 && end[0] - start[0]
19405 < ((1U << (sizeof(((struct regnode_1 *)NULL)->arg1)
19406 * CHARBITS - ANYOFR_BASE_BITS))))
19409 U8 low_utf8[UTF8_MAXBYTES+1];
19410 U8 high_utf8[UTF8_MAXBYTES+1];
19412 ret = reganode(pRExC_state, ANYOFR,
19413 (start[0] | (end[0] - start[0]) << ANYOFR_BASE_BITS));
19415 /* Place the lowest UTF-8 start byte in the flags field, so as to
19416 * allow efficient ruling out at run time of many possible inputs.
19418 (void) uvchr_to_utf8(low_utf8, start[0]);
19419 (void) uvchr_to_utf8(high_utf8, end[0]);
19421 /* If all code points share the same first byte, this can be an
19422 * ANYOFRb. Otherwise store the lowest UTF-8 start byte which can
19423 * quickly rule out many inputs at run-time without having to
19424 * compute the code point from UTF-8. For EBCDIC, we use I8, as
19425 * not doing that transformation would not rule out nearly so many
19427 if (low_utf8[0] == high_utf8[0]) {
19428 OP(REGNODE_p(ret)) = ANYOFRb;
19429 ANYOF_FLAGS(REGNODE_p(ret)) = low_utf8[0];
19432 ANYOF_FLAGS(REGNODE_p(ret))
19433 = NATIVE_UTF8_TO_I8(low_utf8[0]);
19439 /* If didn't find an optimization and there is no need for a bitmap,
19440 * optimize to indicate that */
19441 if ( start[0] >= NUM_ANYOF_CODE_POINTS
19443 && ! upper_latin1_only_utf8_matches
19444 && anyof_flags == 0)
19446 U8 low_utf8[UTF8_MAXBYTES+1];
19447 UV highest_cp = invlist_highest(cp_list);
19449 /* Currently the maximum allowed code point by the system is
19450 * IV_MAX. Higher ones are reserved for future internal use. This
19451 * particular regnode can be used for higher ones, but we can't
19452 * calculate the code point of those. IV_MAX suffices though, as
19453 * it will be a large first byte */
19454 Size_t low_len = uvchr_to_utf8(low_utf8, MIN(start[0], IV_MAX))
19457 /* We store the lowest possible first byte of the UTF-8
19458 * representation, using the flags field. This allows for quick
19459 * ruling out of some inputs without having to convert from UTF-8
19460 * to code point. For EBCDIC, we use I8, as not doing that
19461 * transformation would not rule out nearly so many things */
19462 anyof_flags = NATIVE_UTF8_TO_I8(low_utf8[0]);
19466 /* If the first UTF-8 start byte for the highest code point in the
19467 * range is suitably small, we may be able to get an upper bound as
19469 if (highest_cp <= IV_MAX) {
19470 U8 high_utf8[UTF8_MAXBYTES+1];
19471 Size_t high_len = uvchr_to_utf8(high_utf8, highest_cp)
19474 /* If the lowest and highest are the same, we can get an exact
19475 * first byte instead of a just minimum or even a sequence of
19476 * exact leading bytes. We signal these with different
19478 if (low_utf8[0] == high_utf8[0]) {
19479 Size_t len = find_first_differing_byte_pos(low_utf8,
19481 MIN(low_len, high_len));
19485 /* No need to convert to I8 for EBCDIC as this is an
19487 anyof_flags = low_utf8[0];
19492 ret = regnode_guts(pRExC_state, op,
19493 regarglen[op] + STR_SZ(len),
19495 FILL_NODE(ret, op);
19496 ((struct regnode_anyofhs *) REGNODE_p(ret))->str_len
19498 Copy(low_utf8, /* Add the common bytes */
19499 ((struct regnode_anyofhs *) REGNODE_p(ret))->string,
19501 RExC_emit += NODE_SZ_STR(REGNODE_p(ret));
19502 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19503 NULL, only_utf8_locale_list);
19507 else if (NATIVE_UTF8_TO_I8(high_utf8[0]) <= MAX_ANYOF_HRx_BYTE)
19510 /* Here, the high byte is not the same as the low, but is
19511 * small enough that its reasonable to have a loose upper
19512 * bound, which is packed in with the strict lower bound.
19513 * See comments at the definition of MAX_ANYOF_HRx_BYTE.
19514 * On EBCDIC platforms, I8 is used. On ASCII platforms I8
19515 * is the same thing as UTF-8 */
19518 U8 max_range_diff = MAX_ANYOF_HRx_BYTE - anyof_flags;
19519 U8 range_diff = NATIVE_UTF8_TO_I8(high_utf8[0])
19522 if (range_diff <= max_range_diff / 8) {
19525 else if (range_diff <= max_range_diff / 4) {
19528 else if (range_diff <= max_range_diff / 2) {
19531 anyof_flags = (anyof_flags - 0xC0) << 2 | bits;
19536 goto done_finding_op;
19538 } /* End of seeing if can optimize it into a different node */
19540 is_anyof: /* It's going to be an ANYOF node. */
19541 op = (has_runtime_dependency & HAS_D_RUNTIME_DEPENDENCY)
19551 ret = regnode_guts(pRExC_state, op, regarglen[op], "anyof");
19552 FILL_NODE(ret, op); /* We set the argument later */
19553 RExC_emit += 1 + regarglen[op];
19554 ANYOF_FLAGS(REGNODE_p(ret)) = anyof_flags;
19556 /* Here, <cp_list> contains all the code points we can determine at
19557 * compile time that match under all conditions. Go through it, and
19558 * for things that belong in the bitmap, put them there, and delete from
19559 * <cp_list>. While we are at it, see if everything above 255 is in the
19560 * list, and if so, set a flag to speed up execution */
19562 populate_ANYOF_from_invlist(REGNODE_p(ret), &cp_list);
19565 ANYOF_POSIXL_SET_TO_BITMAP(REGNODE_p(ret), posixl);
19569 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_INVERT;
19572 /* Here, the bitmap has been populated with all the Latin1 code points that
19573 * always match. Can now add to the overall list those that match only
19574 * when the target string is UTF-8 (<upper_latin1_only_utf8_matches>).
19576 if (upper_latin1_only_utf8_matches) {
19578 _invlist_union(cp_list,
19579 upper_latin1_only_utf8_matches,
19581 SvREFCNT_dec_NN(upper_latin1_only_utf8_matches);
19584 cp_list = upper_latin1_only_utf8_matches;
19586 ANYOF_FLAGS(REGNODE_p(ret)) |= ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP;
19589 set_ANYOF_arg(pRExC_state, REGNODE_p(ret), cp_list,
19590 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
19593 only_utf8_locale_list);
19594 SvREFCNT_dec(cp_list);;
19595 SvREFCNT_dec(only_utf8_locale_list);
19600 /* Here, the node is getting optimized into something that's not an ANYOF
19601 * one. Finish up. */
19603 Set_Node_Offset_Length(REGNODE_p(ret), orig_parse - RExC_start,
19604 RExC_parse - orig_parse);;
19605 SvREFCNT_dec(cp_list);;
19606 SvREFCNT_dec(only_utf8_locale_list);
19610 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
19613 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
19614 regnode* const node,
19616 SV* const runtime_defns,
19617 SV* const only_utf8_locale_list)
19619 /* Sets the arg field of an ANYOF-type node 'node', using information about
19620 * the node passed-in. If there is nothing outside the node's bitmap, the
19621 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
19622 * the count returned by add_data(), having allocated and stored an array,
19625 * av[0] stores the inversion list defining this class as far as known at
19626 * this time, or PL_sv_undef if nothing definite is now known.
19627 * av[1] stores the inversion list of code points that match only if the
19628 * current locale is UTF-8, or if none, PL_sv_undef if there is an
19629 * av[2], or no entry otherwise.
19630 * av[2] stores the list of user-defined properties whose subroutine
19631 * definitions aren't known at this time, or no entry if none. */
19635 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
19637 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
19638 assert(! (ANYOF_FLAGS(node)
19639 & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP));
19640 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
19643 AV * const av = newAV();
19647 av_store(av, INVLIST_INDEX, SvREFCNT_inc(cp_list));
19650 if (only_utf8_locale_list) {
19651 av_store(av, ONLY_LOCALE_MATCHES_INDEX,
19652 SvREFCNT_inc(only_utf8_locale_list));
19655 if (runtime_defns) {
19656 av_store(av, DEFERRED_USER_DEFINED_INDEX, SvREFCNT_inc(runtime_defns));
19659 rv = newRV_noinc(MUTABLE_SV(av));
19660 n = add_data(pRExC_state, STR_WITH_LEN("s"));
19661 RExC_rxi->data->data[n] = (void*)rv;
19666 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
19668 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
19669 const regnode* node,
19672 SV** only_utf8_locale_ptr,
19673 SV** output_invlist)
19676 /* For internal core use only.
19677 * Returns the inversion list for the input 'node' in the regex 'prog'.
19678 * If <doinit> is 'true', will attempt to create the inversion list if not
19680 * If <listsvp> is non-null, will return the printable contents of the
19681 * property definition. This can be used to get debugging information
19682 * even before the inversion list exists, by calling this function with
19683 * 'doinit' set to false, in which case the components that will be used
19684 * to eventually create the inversion list are returned (in a printable
19686 * If <only_utf8_locale_ptr> is not NULL, it is where this routine is to
19687 * store an inversion list of code points that should match only if the
19688 * execution-time locale is a UTF-8 one.
19689 * If <output_invlist> is not NULL, it is where this routine is to store an
19690 * inversion list of the code points that would be instead returned in
19691 * <listsvp> if this were NULL. Thus, what gets output in <listsvp>
19692 * when this parameter is used, is just the non-code point data that
19693 * will go into creating the inversion list. This currently should be just
19694 * user-defined properties whose definitions were not known at compile
19695 * time. Using this parameter allows for easier manipulation of the
19696 * inversion list's data by the caller. It is illegal to call this
19697 * function with this parameter set, but not <listsvp>
19699 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
19700 * that, in spite of this function's name, the inversion list it returns
19701 * may include the bitmap data as well */
19703 SV *si = NULL; /* Input initialization string */
19704 SV* invlist = NULL;
19706 RXi_GET_DECL(prog, progi);
19707 const struct reg_data * const data = prog ? progi->data : NULL;
19709 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
19710 assert(! output_invlist || listsvp);
19712 if (data && data->count) {
19713 const U32 n = ARG(node);
19715 if (data->what[n] == 's') {
19716 SV * const rv = MUTABLE_SV(data->data[n]);
19717 AV * const av = MUTABLE_AV(SvRV(rv));
19718 SV **const ary = AvARRAY(av);
19720 invlist = ary[INVLIST_INDEX];
19722 if (av_tindex_skip_len_mg(av) >= ONLY_LOCALE_MATCHES_INDEX) {
19723 *only_utf8_locale_ptr = ary[ONLY_LOCALE_MATCHES_INDEX];
19726 if (av_tindex_skip_len_mg(av) >= DEFERRED_USER_DEFINED_INDEX) {
19727 si = ary[DEFERRED_USER_DEFINED_INDEX];
19730 if (doinit && (si || invlist)) {
19733 SV * msg = newSVpvs_flags("", SVs_TEMP);
19735 SV * prop_definition = handle_user_defined_property(
19736 "", 0, FALSE, /* There is no \p{}, \P{} */
19737 SvPVX_const(si)[1] - '0', /* /i or not has been
19738 stored here for just
19740 TRUE, /* run time */
19741 FALSE, /* This call must find the defn */
19742 si, /* The property definition */
19745 0 /* base level call */
19749 assert(prop_definition == NULL);
19751 Perl_croak(aTHX_ "%" UTF8f,
19752 UTF8fARG(SvUTF8(msg), SvCUR(msg), SvPVX(msg)));
19756 _invlist_union(invlist, prop_definition, &invlist);
19757 SvREFCNT_dec_NN(prop_definition);
19760 invlist = prop_definition;
19763 STATIC_ASSERT_STMT(ONLY_LOCALE_MATCHES_INDEX == 1 + INVLIST_INDEX);
19764 STATIC_ASSERT_STMT(DEFERRED_USER_DEFINED_INDEX == 1 + ONLY_LOCALE_MATCHES_INDEX);
19766 ary[INVLIST_INDEX] = invlist;
19767 av_fill(av, (ary[ONLY_LOCALE_MATCHES_INDEX])
19768 ? ONLY_LOCALE_MATCHES_INDEX
19776 /* If requested, return a printable version of what this ANYOF node matches
19779 SV* matches_string = NULL;
19781 /* This function can be called at compile-time, before everything gets
19782 * resolved, in which case we return the currently best available
19783 * information, which is the string that will eventually be used to do
19784 * that resolving, 'si' */
19786 /* Here, we only have 'si' (and possibly some passed-in data in
19787 * 'invlist', which is handled below) If the caller only wants
19788 * 'si', use that. */
19789 if (! output_invlist) {
19790 matches_string = newSVsv(si);
19793 /* But if the caller wants an inversion list of the node, we
19794 * need to parse 'si' and place as much as possible in the
19795 * desired output inversion list, making 'matches_string' only
19796 * contain the currently unresolvable things */
19797 const char *si_string = SvPVX(si);
19798 STRLEN remaining = SvCUR(si);
19802 /* Ignore everything before the first new-line */
19803 while (*si_string != '\n' && remaining > 0) {
19807 assert(remaining > 0);
19812 while (remaining > 0) {
19814 /* The data consists of just strings defining user-defined
19815 * property names, but in prior incarnations, and perhaps
19816 * somehow from pluggable regex engines, it could still
19817 * hold hex code point definitions. Each component of a
19818 * range would be separated by a tab, and each range by a
19819 * new-line. If these are found, instead add them to the
19820 * inversion list */
19821 I32 grok_flags = PERL_SCAN_SILENT_ILLDIGIT
19822 |PERL_SCAN_SILENT_NON_PORTABLE;
19823 STRLEN len = remaining;
19824 UV cp = grok_hex(si_string, &len, &grok_flags, NULL);
19826 /* If the hex decode routine found something, it should go
19827 * up to the next \n */
19828 if ( *(si_string + len) == '\n') {
19829 if (count) { /* 2nd code point on line */
19830 *output_invlist = _add_range_to_invlist(*output_invlist, prev_cp, cp);
19833 *output_invlist = add_cp_to_invlist(*output_invlist, cp);
19836 goto prepare_for_next_iteration;
19839 /* If the hex decode was instead for the lower range limit,
19840 * save it, and go parse the upper range limit */
19841 if (*(si_string + len) == '\t') {
19842 assert(count == 0);
19846 prepare_for_next_iteration:
19847 si_string += len + 1;
19848 remaining -= len + 1;
19852 /* Here, didn't find a legal hex number. Just add it from
19853 * here to the next \n */
19856 while (*(si_string + len) != '\n' && remaining > 0) {
19860 if (*(si_string + len) == '\n') {
19864 if (matches_string) {
19865 sv_catpvn(matches_string, si_string, len - 1);
19868 matches_string = newSVpvn(si_string, len - 1);
19871 sv_catpvs(matches_string, " ");
19872 } /* end of loop through the text */
19874 assert(matches_string);
19875 if (SvCUR(matches_string)) { /* Get rid of trailing blank */
19876 SvCUR_set(matches_string, SvCUR(matches_string) - 1);
19878 } /* end of has an 'si' */
19881 /* Add the stuff that's already known */
19884 /* Again, if the caller doesn't want the output inversion list, put
19885 * everything in 'matches-string' */
19886 if (! output_invlist) {
19887 if ( ! matches_string) {
19888 matches_string = newSVpvs("\n");
19890 sv_catsv(matches_string, invlist_contents(invlist,
19891 TRUE /* traditional style */
19894 else if (! *output_invlist) {
19895 *output_invlist = invlist_clone(invlist, NULL);
19898 _invlist_union(*output_invlist, invlist, output_invlist);
19902 *listsvp = matches_string;
19907 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
19909 /* reg_skipcomment()
19911 Absorbs an /x style # comment from the input stream,
19912 returning a pointer to the first character beyond the comment, or if the
19913 comment terminates the pattern without anything following it, this returns
19914 one past the final character of the pattern (in other words, RExC_end) and
19915 sets the REG_RUN_ON_COMMENT_SEEN flag.
19917 Note it's the callers responsibility to ensure that we are
19918 actually in /x mode
19922 PERL_STATIC_INLINE char*
19923 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
19925 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
19929 while (p < RExC_end) {
19930 if (*(++p) == '\n') {
19935 /* we ran off the end of the pattern without ending the comment, so we have
19936 * to add an \n when wrapping */
19937 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
19942 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
19944 const bool force_to_xmod
19947 /* If the text at the current parse position '*p' is a '(?#...)' comment,
19948 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
19949 * is /x whitespace, advance '*p' so that on exit it points to the first
19950 * byte past all such white space and comments */
19952 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
19954 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
19956 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
19959 if (RExC_end - (*p) >= 3
19961 && *(*p + 1) == '?'
19962 && *(*p + 2) == '#')
19964 while (*(*p) != ')') {
19965 if ((*p) == RExC_end)
19966 FAIL("Sequence (?#... not terminated");
19974 const char * save_p = *p;
19975 while ((*p) < RExC_end) {
19977 if ((len = is_PATWS_safe((*p), RExC_end, UTF))) {
19980 else if (*(*p) == '#') {
19981 (*p) = reg_skipcomment(pRExC_state, (*p));
19987 if (*p != save_p) {
20000 Advances the parse position by one byte, unless that byte is the beginning
20001 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
20002 those two cases, the parse position is advanced beyond all such comments and
20005 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
20009 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
20011 PERL_ARGS_ASSERT_NEXTCHAR;
20013 if (RExC_parse < RExC_end) {
20015 || UTF8_IS_INVARIANT(*RExC_parse)
20016 || UTF8_IS_START(*RExC_parse));
20018 RExC_parse += (UTF)
20019 ? UTF8_SAFE_SKIP(RExC_parse, RExC_end)
20022 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
20023 FALSE /* Don't force /x */ );
20028 S_change_engine_size(pTHX_ RExC_state_t *pRExC_state, const Ptrdiff_t size)
20030 /* 'size' is the delta number of smallest regnode equivalents to add or
20031 * subtract from the current memory allocated to the regex engine being
20034 PERL_ARGS_ASSERT_CHANGE_ENGINE_SIZE;
20039 sizeof(regexp_internal) + (RExC_size + 1) * sizeof(regnode),
20040 /* +1 for REG_MAGIC */
20043 if ( RExC_rxi == NULL )
20044 FAIL("Regexp out of space");
20045 RXi_SET(RExC_rx, RExC_rxi);
20047 RExC_emit_start = RExC_rxi->program;
20049 Zero(REGNODE_p(RExC_emit), size, regnode);
20052 #ifdef RE_TRACK_PATTERN_OFFSETS
20053 Renew(RExC_offsets, 2*RExC_size+1, U32);
20055 Zero(RExC_offsets + 2*(RExC_size - size) + 1, 2 * size, U32);
20057 RExC_offsets[0] = RExC_size;
20061 STATIC regnode_offset
20062 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
20064 /* Allocate a regnode for 'op', with 'extra_size' extra (smallest) regnode
20065 * equivalents space. It aligns and increments RExC_size
20067 * It returns the regnode's offset into the regex engine program */
20069 const regnode_offset ret = RExC_emit;
20071 GET_RE_DEBUG_FLAGS_DECL;
20073 PERL_ARGS_ASSERT_REGNODE_GUTS;
20075 SIZE_ALIGN(RExC_size);
20076 change_engine_size(pRExC_state, (Ptrdiff_t) 1 + extra_size);
20077 NODE_ALIGN_FILL(REGNODE_p(ret));
20078 #ifndef RE_TRACK_PATTERN_OFFSETS
20079 PERL_UNUSED_ARG(name);
20080 PERL_UNUSED_ARG(op);
20082 assert(extra_size >= regarglen[op] || PL_regkind[op] == ANYOF);
20084 if (RExC_offsets) { /* MJD */
20086 ("%s:%d: (op %s) %s %" UVuf " (len %" UVuf ") (max %" UVuf ").\n",
20089 (UV)(RExC_emit) > RExC_offsets[0]
20090 ? "Overwriting end of array!\n" : "OK",
20092 (UV)(RExC_parse - RExC_start),
20093 (UV)RExC_offsets[0]));
20094 Set_Node_Offset(REGNODE_p(RExC_emit), RExC_parse + (op == END));
20101 - reg_node - emit a node
20103 STATIC regnode_offset /* Location. */
20104 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
20106 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
20107 regnode_offset ptr = ret;
20109 PERL_ARGS_ASSERT_REG_NODE;
20111 assert(regarglen[op] == 0);
20113 FILL_ADVANCE_NODE(ptr, op);
20119 - reganode - emit a node with an argument
20121 STATIC regnode_offset /* Location. */
20122 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
20124 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
20125 regnode_offset ptr = ret;
20127 PERL_ARGS_ASSERT_REGANODE;
20129 /* ANYOF are special cased to allow non-length 1 args */
20130 assert(regarglen[op] == 1);
20132 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
20137 STATIC regnode_offset
20138 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
20140 /* emit a node with U32 and I32 arguments */
20142 const regnode_offset ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
20143 regnode_offset ptr = ret;
20145 PERL_ARGS_ASSERT_REG2LANODE;
20147 assert(regarglen[op] == 2);
20149 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
20155 - reginsert - insert an operator in front of already-emitted operand
20157 * That means that on exit 'operand' is the offset of the newly inserted
20158 * operator, and the original operand has been relocated.
20160 * IMPORTANT NOTE - it is the *callers* responsibility to correctly
20161 * set up NEXT_OFF() of the inserted node if needed. Something like this:
20163 * reginsert(pRExC, OPFAIL, orig_emit, depth+1);
20164 * NEXT_OFF(orig_emit) = regarglen[OPFAIL] + NODE_STEP_REGNODE;
20166 * ALSO NOTE - FLAGS(newly-inserted-operator) will be set to 0 as well.
20169 S_reginsert(pTHX_ RExC_state_t *pRExC_state, const U8 op,
20170 const regnode_offset operand, const U32 depth)
20175 const int offset = regarglen[(U8)op];
20176 const int size = NODE_STEP_REGNODE + offset;
20177 GET_RE_DEBUG_FLAGS_DECL;
20179 PERL_ARGS_ASSERT_REGINSERT;
20180 PERL_UNUSED_CONTEXT;
20181 PERL_UNUSED_ARG(depth);
20182 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
20183 DEBUG_PARSE_FMT("inst"," - %s", PL_reg_name[op]);
20184 assert(!RExC_study_started); /* I believe we should never use reginsert once we have started
20185 studying. If this is wrong then we need to adjust RExC_recurse
20186 below like we do with RExC_open_parens/RExC_close_parens. */
20187 change_engine_size(pRExC_state, (Ptrdiff_t) size);
20188 src = REGNODE_p(RExC_emit);
20190 dst = REGNODE_p(RExC_emit);
20192 /* If we are in a "count the parentheses" pass, the numbers are unreliable,
20193 * and [perl #133871] shows this can lead to problems, so skip this
20194 * realignment of parens until a later pass when they are reliable */
20195 if (! IN_PARENS_PASS && RExC_open_parens) {
20197 /*DEBUG_PARSE_FMT("inst"," - %" IVdf, (IV)RExC_npar);*/
20198 /* remember that RExC_npar is rex->nparens + 1,
20199 * iow it is 1 more than the number of parens seen in
20200 * the pattern so far. */
20201 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
20202 /* note, RExC_open_parens[0] is the start of the
20203 * regex, it can't move. RExC_close_parens[0] is the end
20204 * of the regex, it *can* move. */
20205 if ( paren && RExC_open_parens[paren] >= operand ) {
20206 /*DEBUG_PARSE_FMT("open"," - %d", size);*/
20207 RExC_open_parens[paren] += size;
20209 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
20211 if ( RExC_close_parens[paren] >= operand ) {
20212 /*DEBUG_PARSE_FMT("close"," - %d", size);*/
20213 RExC_close_parens[paren] += size;
20215 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
20220 RExC_end_op += size;
20222 while (src > REGNODE_p(operand)) {
20223 StructCopy(--src, --dst, regnode);
20224 #ifdef RE_TRACK_PATTERN_OFFSETS
20225 if (RExC_offsets) { /* MJD 20010112 */
20227 ("%s(%d): (op %s) %s copy %" UVuf " -> %" UVuf " (max %" UVuf ").\n",
20231 (UV)(REGNODE_OFFSET(dst)) > RExC_offsets[0]
20232 ? "Overwriting end of array!\n" : "OK",
20233 (UV)REGNODE_OFFSET(src),
20234 (UV)REGNODE_OFFSET(dst),
20235 (UV)RExC_offsets[0]));
20236 Set_Node_Offset_To_R(REGNODE_OFFSET(dst), Node_Offset(src));
20237 Set_Node_Length_To_R(REGNODE_OFFSET(dst), Node_Length(src));
20242 place = REGNODE_p(operand); /* Op node, where operand used to be. */
20243 #ifdef RE_TRACK_PATTERN_OFFSETS
20244 if (RExC_offsets) { /* MJD */
20246 ("%s(%d): (op %s) %s %" UVuf " <- %" UVuf " (max %" UVuf ").\n",
20250 (UV)REGNODE_OFFSET(place) > RExC_offsets[0]
20251 ? "Overwriting end of array!\n" : "OK",
20252 (UV)REGNODE_OFFSET(place),
20253 (UV)(RExC_parse - RExC_start),
20254 (UV)RExC_offsets[0]));
20255 Set_Node_Offset(place, RExC_parse);
20256 Set_Node_Length(place, 1);
20259 src = NEXTOPER(place);
20261 FILL_NODE(operand, op);
20263 /* Zero out any arguments in the new node */
20264 Zero(src, offset, regnode);
20268 - regtail - set the next-pointer at the end of a node chain of p to val. If
20269 that value won't fit in the space available, instead returns FALSE.
20270 (Except asserts if we can't fit in the largest space the regex
20271 engine is designed for.)
20272 - SEE ALSO: regtail_study
20275 S_regtail(pTHX_ RExC_state_t * pRExC_state,
20276 const regnode_offset p,
20277 const regnode_offset val,
20280 regnode_offset scan;
20281 GET_RE_DEBUG_FLAGS_DECL;
20283 PERL_ARGS_ASSERT_REGTAIL;
20285 PERL_UNUSED_ARG(depth);
20288 /* Find last node. */
20289 scan = (regnode_offset) p;
20291 regnode * const temp = regnext(REGNODE_p(scan));
20293 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
20294 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20295 Perl_re_printf( aTHX_ "~ %s (%d) %s %s\n",
20296 SvPV_nolen_const(RExC_mysv), scan,
20297 (temp == NULL ? "->" : ""),
20298 (temp == NULL ? PL_reg_name[OP(REGNODE_p(val))] : "")
20303 scan = REGNODE_OFFSET(temp);
20306 assert(val >= scan);
20307 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20308 assert((UV) (val - scan) <= U32_MAX);
20309 ARG_SET(REGNODE_p(scan), val - scan);
20312 if (val - scan > U16_MAX) {
20313 /* Populate this with something that won't loop and will likely
20314 * lead to a crash if the caller ignores the failure return, and
20315 * execution continues */
20316 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20319 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20327 - regtail_study - set the next-pointer at the end of a node chain of p to val.
20328 - Look for optimizable sequences at the same time.
20329 - currently only looks for EXACT chains.
20331 This is experimental code. The idea is to use this routine to perform
20332 in place optimizations on branches and groups as they are constructed,
20333 with the long term intention of removing optimization from study_chunk so
20334 that it is purely analytical.
20336 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
20337 to control which is which.
20339 This used to return a value that was ignored. It was a problem that it is
20340 #ifdef'd to be another function that didn't return a value. khw has changed it
20341 so both currently return a pass/fail return.
20344 /* TODO: All four parms should be const */
20347 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode_offset p,
20348 const regnode_offset val, U32 depth)
20350 regnode_offset scan;
20352 #ifdef EXPERIMENTAL_INPLACESCAN
20355 GET_RE_DEBUG_FLAGS_DECL;
20357 PERL_ARGS_ASSERT_REGTAIL_STUDY;
20360 /* Find last node. */
20364 regnode * const temp = regnext(REGNODE_p(scan));
20365 #ifdef EXPERIMENTAL_INPLACESCAN
20366 if (PL_regkind[OP(REGNODE_p(scan))] == EXACT) {
20367 bool unfolded_multi_char; /* Unexamined in this routine */
20368 if (join_exact(pRExC_state, scan, &min,
20369 &unfolded_multi_char, 1, REGNODE_p(val), depth+1))
20370 return TRUE; /* Was return EXACT */
20374 switch (OP(REGNODE_p(scan))) {
20381 case EXACTFU_S_EDGE:
20382 case EXACTFAA_NO_TRIE:
20389 if( exact == PSEUDO )
20390 exact= OP(REGNODE_p(scan));
20391 else if ( exact != OP(REGNODE_p(scan)) )
20400 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
20401 regprop(RExC_rx, RExC_mysv, REGNODE_p(scan), NULL, pRExC_state);
20402 Perl_re_printf( aTHX_ "~ %s (%d) -> %s\n",
20403 SvPV_nolen_const(RExC_mysv),
20405 PL_reg_name[exact]);
20409 scan = REGNODE_OFFSET(temp);
20412 DEBUG_PARSE_MSG("");
20413 regprop(RExC_rx, RExC_mysv, REGNODE_p(val), NULL, pRExC_state);
20414 Perl_re_printf( aTHX_
20415 "~ attach to %s (%" IVdf ") offset to %" IVdf "\n",
20416 SvPV_nolen_const(RExC_mysv),
20421 if (reg_off_by_arg[OP(REGNODE_p(scan))]) {
20422 assert((UV) (val - scan) <= U32_MAX);
20423 ARG_SET(REGNODE_p(scan), val - scan);
20426 if (val - scan > U16_MAX) {
20427 /* Populate this with something that won't loop and will likely
20428 * lead to a crash if the caller ignores the failure return, and
20429 * execution continues */
20430 NEXT_OFF(REGNODE_p(scan)) = U16_MAX;
20433 NEXT_OFF(REGNODE_p(scan)) = val - scan;
20436 return TRUE; /* Was 'return exact' */
20441 S_get_ANYOFM_contents(pTHX_ const regnode * n) {
20443 /* Returns an inversion list of all the code points matched by the
20444 * ANYOFM/NANYOFM node 'n' */
20446 SV * cp_list = _new_invlist(-1);
20447 const U8 lowest = (U8) ARG(n);
20450 U8 needed = 1U << PL_bitcount[ (U8) ~ FLAGS(n)];
20452 PERL_ARGS_ASSERT_GET_ANYOFM_CONTENTS;
20454 /* Starting with the lowest code point, any code point that ANDed with the
20455 * mask yields the lowest code point is in the set */
20456 for (i = lowest; i <= 0xFF; i++) {
20457 if ((i & FLAGS(n)) == ARG(n)) {
20458 cp_list = add_cp_to_invlist(cp_list, i);
20461 /* We know how many code points (a power of two) that are in the
20462 * set. No use looking once we've got that number */
20463 if (count >= needed) break;
20467 if (OP(n) == NANYOFM) {
20468 _invlist_invert(cp_list);
20474 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
20479 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
20484 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20486 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
20487 if (flags & (1<<bit)) {
20488 if (!set++ && lead)
20489 Perl_re_printf( aTHX_ "%s", lead);
20490 Perl_re_printf( aTHX_ "%s ", PL_reg_intflags_name[bit]);
20495 Perl_re_printf( aTHX_ "\n");
20497 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20502 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
20508 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
20510 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
20511 if (flags & (1<<bit)) {
20512 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
20515 if (!set++ && lead)
20516 Perl_re_printf( aTHX_ "%s", lead);
20517 Perl_re_printf( aTHX_ "%s ", PL_reg_extflags_name[bit]);
20520 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
20521 if (!set++ && lead) {
20522 Perl_re_printf( aTHX_ "%s", lead);
20525 case REGEX_UNICODE_CHARSET:
20526 Perl_re_printf( aTHX_ "UNICODE");
20528 case REGEX_LOCALE_CHARSET:
20529 Perl_re_printf( aTHX_ "LOCALE");
20531 case REGEX_ASCII_RESTRICTED_CHARSET:
20532 Perl_re_printf( aTHX_ "ASCII-RESTRICTED");
20534 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
20535 Perl_re_printf( aTHX_ "ASCII-MORE_RESTRICTED");
20538 Perl_re_printf( aTHX_ "UNKNOWN CHARACTER SET");
20544 Perl_re_printf( aTHX_ "\n");
20546 Perl_re_printf( aTHX_ "%s[none-set]\n", lead);
20552 Perl_regdump(pTHX_ const regexp *r)
20556 SV * const sv = sv_newmortal();
20557 SV *dsv= sv_newmortal();
20558 RXi_GET_DECL(r, ri);
20559 GET_RE_DEBUG_FLAGS_DECL;
20561 PERL_ARGS_ASSERT_REGDUMP;
20563 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
20565 /* Header fields of interest. */
20566 for (i = 0; i < 2; i++) {
20567 if (r->substrs->data[i].substr) {
20568 RE_PV_QUOTED_DECL(s, 0, dsv,
20569 SvPVX_const(r->substrs->data[i].substr),
20570 RE_SV_DUMPLEN(r->substrs->data[i].substr),
20571 PL_dump_re_max_len);
20572 Perl_re_printf( aTHX_
20573 "%s %s%s at %" IVdf "..%" UVuf " ",
20574 i ? "floating" : "anchored",
20576 RE_SV_TAIL(r->substrs->data[i].substr),
20577 (IV)r->substrs->data[i].min_offset,
20578 (UV)r->substrs->data[i].max_offset);
20580 else if (r->substrs->data[i].utf8_substr) {
20581 RE_PV_QUOTED_DECL(s, 1, dsv,
20582 SvPVX_const(r->substrs->data[i].utf8_substr),
20583 RE_SV_DUMPLEN(r->substrs->data[i].utf8_substr),
20585 Perl_re_printf( aTHX_
20586 "%s utf8 %s%s at %" IVdf "..%" UVuf " ",
20587 i ? "floating" : "anchored",
20589 RE_SV_TAIL(r->substrs->data[i].utf8_substr),
20590 (IV)r->substrs->data[i].min_offset,
20591 (UV)r->substrs->data[i].max_offset);
20595 if (r->check_substr || r->check_utf8)
20596 Perl_re_printf( aTHX_
20598 ( r->check_substr == r->substrs->data[1].substr
20599 && r->check_utf8 == r->substrs->data[1].utf8_substr
20600 ? "(checking floating" : "(checking anchored"));
20601 if (r->intflags & PREGf_NOSCAN)
20602 Perl_re_printf( aTHX_ " noscan");
20603 if (r->extflags & RXf_CHECK_ALL)
20604 Perl_re_printf( aTHX_ " isall");
20605 if (r->check_substr || r->check_utf8)
20606 Perl_re_printf( aTHX_ ") ");
20608 if (ri->regstclass) {
20609 regprop(r, sv, ri->regstclass, NULL, NULL);
20610 Perl_re_printf( aTHX_ "stclass %s ", SvPVX_const(sv));
20612 if (r->intflags & PREGf_ANCH) {
20613 Perl_re_printf( aTHX_ "anchored");
20614 if (r->intflags & PREGf_ANCH_MBOL)
20615 Perl_re_printf( aTHX_ "(MBOL)");
20616 if (r->intflags & PREGf_ANCH_SBOL)
20617 Perl_re_printf( aTHX_ "(SBOL)");
20618 if (r->intflags & PREGf_ANCH_GPOS)
20619 Perl_re_printf( aTHX_ "(GPOS)");
20620 Perl_re_printf( aTHX_ " ");
20622 if (r->intflags & PREGf_GPOS_SEEN)
20623 Perl_re_printf( aTHX_ "GPOS:%" UVuf " ", (UV)r->gofs);
20624 if (r->intflags & PREGf_SKIP)
20625 Perl_re_printf( aTHX_ "plus ");
20626 if (r->intflags & PREGf_IMPLICIT)
20627 Perl_re_printf( aTHX_ "implicit ");
20628 Perl_re_printf( aTHX_ "minlen %" IVdf " ", (IV)r->minlen);
20629 if (r->extflags & RXf_EVAL_SEEN)
20630 Perl_re_printf( aTHX_ "with eval ");
20631 Perl_re_printf( aTHX_ "\n");
20633 regdump_extflags("r->extflags: ", r->extflags);
20634 regdump_intflags("r->intflags: ", r->intflags);
20637 PERL_ARGS_ASSERT_REGDUMP;
20638 PERL_UNUSED_CONTEXT;
20639 PERL_UNUSED_ARG(r);
20640 #endif /* DEBUGGING */
20643 /* Should be synchronized with ANYOF_ #defines in regcomp.h */
20646 # if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 \
20647 || _CC_LOWER != 3 || _CC_UPPER != 4 || _CC_PUNCT != 5 \
20648 || _CC_PRINT != 6 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 \
20649 || _CC_CASED != 9 || _CC_SPACE != 10 || _CC_BLANK != 11 \
20650 || _CC_XDIGIT != 12 || _CC_CNTRL != 13 || _CC_ASCII != 14 \
20651 || _CC_VERTSPACE != 15
20652 # error Need to adjust order of anyofs[]
20654 static const char * const anyofs[] = {
20691 - regprop - printable representation of opcode, with run time support
20695 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
20700 RXi_GET_DECL(prog, progi);
20701 GET_RE_DEBUG_FLAGS_DECL;
20703 PERL_ARGS_ASSERT_REGPROP;
20707 if (OP(o) > REGNODE_MAX) { /* regnode.type is unsigned */
20708 if (pRExC_state) { /* This gives more info, if we have it */
20709 FAIL3("panic: corrupted regexp opcode %d > %d",
20710 (int)OP(o), (int)REGNODE_MAX);
20713 Perl_croak(aTHX_ "panic: corrupted regexp opcode %d > %d",
20714 (int)OP(o), (int)REGNODE_MAX);
20717 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
20719 k = PL_regkind[OP(o)];
20722 sv_catpvs(sv, " ");
20723 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
20724 * is a crude hack but it may be the best for now since
20725 * we have no flag "this EXACTish node was UTF-8"
20727 pv_pretty(sv, STRING(o), STR_LEN(o), PL_dump_re_max_len,
20728 PL_colors[0], PL_colors[1],
20729 PERL_PV_ESCAPE_UNI_DETECT |
20730 PERL_PV_ESCAPE_NONASCII |
20731 PERL_PV_PRETTY_ELLIPSES |
20732 PERL_PV_PRETTY_LTGT |
20733 PERL_PV_PRETTY_NOCLEAR
20735 } else if (k == TRIE) {
20736 /* print the details of the trie in dumpuntil instead, as
20737 * progi->data isn't available here */
20738 const char op = OP(o);
20739 const U32 n = ARG(o);
20740 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
20741 (reg_ac_data *)progi->data->data[n] :
20743 const reg_trie_data * const trie
20744 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
20746 Perl_sv_catpvf(aTHX_ sv, "-%s", PL_reg_name[o->flags]);
20747 DEBUG_TRIE_COMPILE_r({
20749 sv_catpvs(sv, "(JUMP)");
20750 Perl_sv_catpvf(aTHX_ sv,
20751 "<S:%" UVuf "/%" IVdf " W:%" UVuf " L:%" UVuf "/%" UVuf " C:%" UVuf "/%" UVuf ">",
20752 (UV)trie->startstate,
20753 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
20754 (UV)trie->wordcount,
20757 (UV)TRIE_CHARCOUNT(trie),
20758 (UV)trie->uniquecharcount
20761 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
20762 sv_catpvs(sv, "[");
20763 (void) put_charclass_bitmap_innards(sv,
20764 ((IS_ANYOF_TRIE(op))
20766 : TRIE_BITMAP(trie)),
20773 sv_catpvs(sv, "]");
20775 } else if (k == CURLY) {
20776 U32 lo = ARG1(o), hi = ARG2(o);
20777 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
20778 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
20779 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
20780 if (hi == REG_INFTY)
20781 sv_catpvs(sv, "INFTY");
20783 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
20784 sv_catpvs(sv, "}");
20786 else if (k == WHILEM && o->flags) /* Ordinal/of */
20787 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
20788 else if (k == REF || k == OPEN || k == CLOSE
20789 || k == GROUPP || OP(o)==ACCEPT)
20791 AV *name_list= NULL;
20792 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
20793 Perl_sv_catpvf(aTHX_ sv, "%" UVuf, (UV)parno); /* Parenth number */
20794 if ( RXp_PAREN_NAMES(prog) ) {
20795 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20796 } else if ( pRExC_state ) {
20797 name_list= RExC_paren_name_list;
20800 if ( k != REF || (OP(o) < REFN)) {
20801 SV **name= av_fetch(name_list, parno, 0 );
20803 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20806 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
20807 I32 *nums=(I32*)SvPVX(sv_dat);
20808 SV **name= av_fetch(name_list, nums[0], 0 );
20811 for ( n=0; n<SvIVX(sv_dat); n++ ) {
20812 Perl_sv_catpvf(aTHX_ sv, "%s%" IVdf,
20813 (n ? "," : ""), (IV)nums[n]);
20815 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20819 if ( k == REF && reginfo) {
20820 U32 n = ARG(o); /* which paren pair */
20821 I32 ln = prog->offs[n].start;
20822 if (prog->lastparen < n || ln == -1 || prog->offs[n].end == -1)
20823 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
20824 else if (ln == prog->offs[n].end)
20825 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
20827 const char *s = reginfo->strbeg + ln;
20828 Perl_sv_catpvf(aTHX_ sv, ": ");
20829 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
20830 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
20833 } else if (k == GOSUB) {
20834 AV *name_list= NULL;
20835 if ( RXp_PAREN_NAMES(prog) ) {
20836 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
20837 } else if ( pRExC_state ) {
20838 name_list= RExC_paren_name_list;
20841 /* Paren and offset */
20842 Perl_sv_catpvf(aTHX_ sv, "%d[%+d:%d]", (int)ARG(o),(int)ARG2L(o),
20843 (int)((o + (int)ARG2L(o)) - progi->program) );
20845 SV **name= av_fetch(name_list, ARG(o), 0 );
20847 Perl_sv_catpvf(aTHX_ sv, " '%" SVf "'", SVfARG(*name));
20850 else if (k == LOGICAL)
20851 /* 2: embedded, otherwise 1 */
20852 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
20853 else if (k == ANYOF || k == ANYOFR) {
20857 bool do_sep = FALSE; /* Do we need to separate various components of
20859 /* Set if there is still an unresolved user-defined property */
20860 SV *unresolved = NULL;
20862 /* Things that are ignored except when the runtime locale is UTF-8 */
20863 SV *only_utf8_locale_invlist = NULL;
20865 /* Code points that don't fit in the bitmap */
20866 SV *nonbitmap_invlist = NULL;
20868 /* And things that aren't in the bitmap, but are small enough to be */
20869 SV* bitmap_range_not_in_bitmap = NULL;
20873 if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
20879 flags = ANYOF_FLAGS(o);
20880 bitmap = ANYOF_BITMAP(o);
20884 if (OP(o) == ANYOFL || OP(o) == ANYOFPOSIXL) {
20885 if (ANYOFL_UTF8_LOCALE_REQD(flags)) {
20886 sv_catpvs(sv, "{utf8-locale-reqd}");
20888 if (flags & ANYOFL_FOLD) {
20889 sv_catpvs(sv, "{i}");
20893 inverted = flags & ANYOF_INVERT;
20895 /* If there is stuff outside the bitmap, get it */
20896 if (arg != ANYOF_ONLY_HAS_BITMAP) {
20897 if (inRANGE(OP(o), ANYOFR, ANYOFRb)) {
20898 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
20900 ANYOFRbase(o) + ANYOFRdelta(o));
20903 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
20905 &only_utf8_locale_invlist,
20906 &nonbitmap_invlist);
20909 /* The non-bitmap data may contain stuff that could fit in the
20910 * bitmap. This could come from a user-defined property being
20911 * finally resolved when this call was done; or much more likely
20912 * because there are matches that require UTF-8 to be valid, and so
20913 * aren't in the bitmap (or ANYOFR). This is teased apart later */
20914 _invlist_intersection(nonbitmap_invlist,
20916 &bitmap_range_not_in_bitmap);
20917 /* Leave just the things that don't fit into the bitmap */
20918 _invlist_subtract(nonbitmap_invlist,
20920 &nonbitmap_invlist);
20923 /* Obey this flag to add all above-the-bitmap code points */
20924 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
20925 nonbitmap_invlist = _add_range_to_invlist(nonbitmap_invlist,
20926 NUM_ANYOF_CODE_POINTS,
20930 /* Ready to start outputting. First, the initial left bracket */
20931 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
20933 /* ANYOFH by definition doesn't have anything that will fit inside the
20934 * bitmap; ANYOFR may or may not. */
20935 if ( ! inRANGE(OP(o), ANYOFH, ANYOFHr)
20936 && ( ! inRANGE(OP(o), ANYOFR, ANYOFRb)
20937 || ANYOFRbase(o) < NUM_ANYOF_CODE_POINTS))
20939 /* Then all the things that could fit in the bitmap */
20940 do_sep = put_charclass_bitmap_innards(sv,
20942 bitmap_range_not_in_bitmap,
20943 only_utf8_locale_invlist,
20947 /* Can't try inverting for a
20948 * better display if there
20949 * are things that haven't
20952 || inRANGE(OP(o), ANYOFR, ANYOFRb));
20953 SvREFCNT_dec(bitmap_range_not_in_bitmap);
20955 /* If there are user-defined properties which haven't been defined
20956 * yet, output them. If the result is not to be inverted, it is
20957 * clearest to output them in a separate [] from the bitmap range
20958 * stuff. If the result is to be complemented, we have to show
20959 * everything in one [], as the inversion applies to the whole
20960 * thing. Use {braces} to separate them from anything in the
20961 * bitmap and anything above the bitmap. */
20964 if (! do_sep) { /* If didn't output anything in the bitmap
20966 sv_catpvs(sv, "^");
20968 sv_catpvs(sv, "{");
20971 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1],
20974 sv_catsv(sv, unresolved);
20976 sv_catpvs(sv, "}");
20978 do_sep = ! inverted;
20982 /* And, finally, add the above-the-bitmap stuff */
20983 if (nonbitmap_invlist && _invlist_len(nonbitmap_invlist)) {
20986 /* See if truncation size is overridden */
20987 const STRLEN dump_len = (PL_dump_re_max_len > 256)
20988 ? PL_dump_re_max_len
20991 /* This is output in a separate [] */
20993 Perl_sv_catpvf(aTHX_ sv,"%s][%s", PL_colors[1], PL_colors[0]);
20996 /* And, for easy of understanding, it is shown in the
20997 * uncomplemented form if possible. The one exception being if
20998 * there are unresolved items, where the inversion has to be
20999 * delayed until runtime */
21000 if (inverted && ! unresolved) {
21001 _invlist_invert(nonbitmap_invlist);
21002 _invlist_subtract(nonbitmap_invlist, PL_InBitmap, &nonbitmap_invlist);
21005 contents = invlist_contents(nonbitmap_invlist,
21006 FALSE /* output suitable for catsv */
21009 /* If the output is shorter than the permissible maximum, just do it. */
21010 if (SvCUR(contents) <= dump_len) {
21011 sv_catsv(sv, contents);
21014 const char * contents_string = SvPVX(contents);
21015 STRLEN i = dump_len;
21017 /* Otherwise, start at the permissible max and work back to the
21018 * first break possibility */
21019 while (i > 0 && contents_string[i] != ' ') {
21022 if (i == 0) { /* Fail-safe. Use the max if we couldn't
21023 find a legal break */
21027 sv_catpvn(sv, contents_string, i);
21028 sv_catpvs(sv, "...");
21031 SvREFCNT_dec_NN(contents);
21032 SvREFCNT_dec_NN(nonbitmap_invlist);
21035 /* And finally the matching, closing ']' */
21036 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21038 if (OP(o) == ANYOFHs) {
21039 Perl_sv_catpvf(aTHX_ sv, " (Leading UTF-8 bytes=%s", _byte_dump_string((U8 *) ((struct regnode_anyofhs *) o)->string, FLAGS(o), 1));
21041 else if (inRANGE(OP(o), ANYOFH, ANYOFRb)) {
21042 U8 lowest = (OP(o) != ANYOFHr)
21044 : LOWEST_ANYOF_HRx_BYTE(FLAGS(o));
21045 U8 highest = (OP(o) == ANYOFHr)
21046 ? HIGHEST_ANYOF_HRx_BYTE(FLAGS(o))
21047 : (OP(o) == ANYOFH || OP(o) == ANYOFR)
21050 Perl_sv_catpvf(aTHX_ sv, " (First UTF-8 byte=%02X", lowest);
21051 if (lowest != highest) {
21052 Perl_sv_catpvf(aTHX_ sv, "-%02X", highest);
21054 Perl_sv_catpvf(aTHX_ sv, ")");
21057 SvREFCNT_dec(unresolved);
21059 else if (k == ANYOFM) {
21060 SV * cp_list = get_ANYOFM_contents(o);
21062 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
21063 if (OP(o) == NANYOFM) {
21064 _invlist_invert(cp_list);
21067 put_charclass_bitmap_innards(sv, NULL, cp_list, NULL, NULL, 0, TRUE);
21068 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
21070 SvREFCNT_dec(cp_list);
21072 else if (k == POSIXD || k == NPOSIXD) {
21073 U8 index = FLAGS(o) * 2;
21074 if (index < C_ARRAY_LENGTH(anyofs)) {
21075 if (*anyofs[index] != '[') {
21076 sv_catpvs(sv, "[");
21078 sv_catpv(sv, anyofs[index]);
21079 if (*anyofs[index] != '[') {
21080 sv_catpvs(sv, "]");
21084 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
21087 else if (k == BOUND || k == NBOUND) {
21088 /* Must be synced with order of 'bound_type' in regcomp.h */
21089 const char * const bounds[] = {
21090 "", /* Traditional */
21096 assert(FLAGS(o) < C_ARRAY_LENGTH(bounds));
21097 sv_catpv(sv, bounds[FLAGS(o)]);
21099 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH)) {
21100 Perl_sv_catpvf(aTHX_ sv, "[%d", -(o->flags));
21102 Perl_sv_catpvf(aTHX_ sv, "..-%d", o->flags - o->next_off);
21104 Perl_sv_catpvf(aTHX_ sv, "]");
21106 else if (OP(o) == SBOL)
21107 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
21109 /* add on the verb argument if there is one */
21110 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
21112 Perl_sv_catpvf(aTHX_ sv, ":%" SVf,
21113 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
21115 sv_catpvs(sv, ":NULL");
21118 PERL_UNUSED_CONTEXT;
21119 PERL_UNUSED_ARG(sv);
21120 PERL_UNUSED_ARG(o);
21121 PERL_UNUSED_ARG(prog);
21122 PERL_UNUSED_ARG(reginfo);
21123 PERL_UNUSED_ARG(pRExC_state);
21124 #endif /* DEBUGGING */
21130 Perl_re_intuit_string(pTHX_ REGEXP * const r)
21131 { /* Assume that RE_INTUIT is set */
21132 struct regexp *const prog = ReANY(r);
21133 GET_RE_DEBUG_FLAGS_DECL;
21135 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
21136 PERL_UNUSED_CONTEXT;
21140 const char * const s = SvPV_nolen_const(RX_UTF8(r)
21141 ? prog->check_utf8 : prog->check_substr);
21143 if (!PL_colorset) reginitcolors();
21144 Perl_re_printf( aTHX_
21145 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
21147 RX_UTF8(r) ? "utf8 " : "",
21148 PL_colors[5], PL_colors[0],
21151 (strlen(s) > PL_dump_re_max_len ? "..." : ""));
21154 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
21155 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
21161 handles refcounting and freeing the perl core regexp structure. When
21162 it is necessary to actually free the structure the first thing it
21163 does is call the 'free' method of the regexp_engine associated to
21164 the regexp, allowing the handling of the void *pprivate; member
21165 first. (This routine is not overridable by extensions, which is why
21166 the extensions free is called first.)
21168 See regdupe and regdupe_internal if you change anything here.
21170 #ifndef PERL_IN_XSUB_RE
21172 Perl_pregfree(pTHX_ REGEXP *r)
21178 Perl_pregfree2(pTHX_ REGEXP *rx)
21180 struct regexp *const r = ReANY(rx);
21181 GET_RE_DEBUG_FLAGS_DECL;
21183 PERL_ARGS_ASSERT_PREGFREE2;
21188 if (r->mother_re) {
21189 ReREFCNT_dec(r->mother_re);
21191 CALLREGFREE_PVT(rx); /* free the private data */
21192 SvREFCNT_dec(RXp_PAREN_NAMES(r));
21196 for (i = 0; i < 2; i++) {
21197 SvREFCNT_dec(r->substrs->data[i].substr);
21198 SvREFCNT_dec(r->substrs->data[i].utf8_substr);
21200 Safefree(r->substrs);
21202 RX_MATCH_COPY_FREE(rx);
21203 #ifdef PERL_ANY_COW
21204 SvREFCNT_dec(r->saved_copy);
21207 SvREFCNT_dec(r->qr_anoncv);
21208 if (r->recurse_locinput)
21209 Safefree(r->recurse_locinput);
21215 Copy ssv to dsv, both of which should of type SVt_REGEXP or SVt_PVLV,
21216 except that dsv will be created if NULL.
21218 This function is used in two main ways. First to implement
21219 $r = qr/....; $s = $$r;
21221 Secondly, it is used as a hacky workaround to the structural issue of
21223 being stored in the regexp structure which is in turn stored in
21224 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
21225 could be PL_curpm in multiple contexts, and could require multiple
21226 result sets being associated with the pattern simultaneously, such
21227 as when doing a recursive match with (??{$qr})
21229 The solution is to make a lightweight copy of the regexp structure
21230 when a qr// is returned from the code executed by (??{$qr}) this
21231 lightweight copy doesn't actually own any of its data except for
21232 the starp/end and the actual regexp structure itself.
21238 Perl_reg_temp_copy(pTHX_ REGEXP *dsv, REGEXP *ssv)
21240 struct regexp *drx;
21241 struct regexp *const srx = ReANY(ssv);
21242 const bool islv = dsv && SvTYPE(dsv) == SVt_PVLV;
21244 PERL_ARGS_ASSERT_REG_TEMP_COPY;
21247 dsv = (REGEXP*) newSV_type(SVt_REGEXP);
21249 assert(SvTYPE(dsv) == SVt_REGEXP || (SvTYPE(dsv) == SVt_PVLV));
21251 /* our only valid caller, sv_setsv_flags(), should have done
21252 * a SV_CHECK_THINKFIRST_COW_DROP() by now */
21253 assert(!SvOOK(dsv));
21254 assert(!SvIsCOW(dsv));
21255 assert(!SvROK(dsv));
21257 if (SvPVX_const(dsv)) {
21259 Safefree(SvPVX(dsv));
21264 SvOK_off((SV *)dsv);
21267 /* For PVLVs, the head (sv_any) points to an XPVLV, while
21268 * the LV's xpvlenu_rx will point to a regexp body, which
21269 * we allocate here */
21270 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
21271 assert(!SvPVX(dsv));
21272 ((XPV*)SvANY(dsv))->xpv_len_u.xpvlenu_rx = temp->sv_any;
21273 temp->sv_any = NULL;
21274 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
21275 SvREFCNT_dec_NN(temp);
21276 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
21277 ing below will not set it. */
21278 SvCUR_set(dsv, SvCUR(ssv));
21281 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
21282 sv_force_normal(sv) is called. */
21286 SvFLAGS(dsv) |= SvFLAGS(ssv) & (SVf_POK|SVp_POK|SVf_UTF8);
21287 SvPV_set(dsv, RX_WRAPPED(ssv));
21288 /* We share the same string buffer as the original regexp, on which we
21289 hold a reference count, incremented when mother_re is set below.
21290 The string pointer is copied here, being part of the regexp struct.
21292 memcpy(&(drx->xpv_cur), &(srx->xpv_cur),
21293 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
21297 const I32 npar = srx->nparens+1;
21298 Newx(drx->offs, npar, regexp_paren_pair);
21299 Copy(srx->offs, drx->offs, npar, regexp_paren_pair);
21301 if (srx->substrs) {
21303 Newx(drx->substrs, 1, struct reg_substr_data);
21304 StructCopy(srx->substrs, drx->substrs, struct reg_substr_data);
21306 for (i = 0; i < 2; i++) {
21307 SvREFCNT_inc_void(drx->substrs->data[i].substr);
21308 SvREFCNT_inc_void(drx->substrs->data[i].utf8_substr);
21311 /* check_substr and check_utf8, if non-NULL, point to either their
21312 anchored or float namesakes, and don't hold a second reference. */
21314 RX_MATCH_COPIED_off(dsv);
21315 #ifdef PERL_ANY_COW
21316 drx->saved_copy = NULL;
21318 drx->mother_re = ReREFCNT_inc(srx->mother_re ? srx->mother_re : ssv);
21319 SvREFCNT_inc_void(drx->qr_anoncv);
21320 if (srx->recurse_locinput)
21321 Newx(drx->recurse_locinput, srx->nparens + 1, char *);
21328 /* regfree_internal()
21330 Free the private data in a regexp. This is overloadable by
21331 extensions. Perl takes care of the regexp structure in pregfree(),
21332 this covers the *pprivate pointer which technically perl doesn't
21333 know about, however of course we have to handle the
21334 regexp_internal structure when no extension is in use.
21336 Note this is called before freeing anything in the regexp
21341 Perl_regfree_internal(pTHX_ REGEXP * const rx)
21343 struct regexp *const r = ReANY(rx);
21344 RXi_GET_DECL(r, ri);
21345 GET_RE_DEBUG_FLAGS_DECL;
21347 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
21357 SV *dsv= sv_newmortal();
21358 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
21359 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), PL_dump_re_max_len);
21360 Perl_re_printf( aTHX_ "%sFreeing REx:%s %s\n",
21361 PL_colors[4], PL_colors[5], s);
21365 #ifdef RE_TRACK_PATTERN_OFFSETS
21367 Safefree(ri->u.offsets); /* 20010421 MJD */
21369 if (ri->code_blocks)
21370 S_free_codeblocks(aTHX_ ri->code_blocks);
21373 int n = ri->data->count;
21376 /* If you add a ->what type here, update the comment in regcomp.h */
21377 switch (ri->data->what[n]) {
21383 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
21386 Safefree(ri->data->data[n]);
21392 { /* Aho Corasick add-on structure for a trie node.
21393 Used in stclass optimization only */
21395 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
21396 #ifdef USE_ITHREADS
21400 refcount = --aho->refcount;
21403 PerlMemShared_free(aho->states);
21404 PerlMemShared_free(aho->fail);
21405 /* do this last!!!! */
21406 PerlMemShared_free(ri->data->data[n]);
21407 /* we should only ever get called once, so
21408 * assert as much, and also guard the free
21409 * which /might/ happen twice. At the least
21410 * it will make code anlyzers happy and it
21411 * doesn't cost much. - Yves */
21412 assert(ri->regstclass);
21413 if (ri->regstclass) {
21414 PerlMemShared_free(ri->regstclass);
21415 ri->regstclass = 0;
21422 /* trie structure. */
21424 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
21425 #ifdef USE_ITHREADS
21429 refcount = --trie->refcount;
21432 PerlMemShared_free(trie->charmap);
21433 PerlMemShared_free(trie->states);
21434 PerlMemShared_free(trie->trans);
21436 PerlMemShared_free(trie->bitmap);
21438 PerlMemShared_free(trie->jump);
21439 PerlMemShared_free(trie->wordinfo);
21440 /* do this last!!!! */
21441 PerlMemShared_free(ri->data->data[n]);
21446 Perl_croak(aTHX_ "panic: regfree data code '%c'",
21447 ri->data->what[n]);
21450 Safefree(ri->data->what);
21451 Safefree(ri->data);
21457 #define av_dup_inc(s, t) MUTABLE_AV(sv_dup_inc((const SV *)s, t))
21458 #define hv_dup_inc(s, t) MUTABLE_HV(sv_dup_inc((const SV *)s, t))
21459 #define SAVEPVN(p, n) ((p) ? savepvn(p, n) : NULL)
21462 re_dup_guts - duplicate a regexp.
21464 This routine is expected to clone a given regexp structure. It is only
21465 compiled under USE_ITHREADS.
21467 After all of the core data stored in struct regexp is duplicated
21468 the regexp_engine.dupe method is used to copy any private data
21469 stored in the *pprivate pointer. This allows extensions to handle
21470 any duplication it needs to do.
21472 See pregfree() and regfree_internal() if you change anything here.
21474 #if defined(USE_ITHREADS)
21475 #ifndef PERL_IN_XSUB_RE
21477 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
21481 const struct regexp *r = ReANY(sstr);
21482 struct regexp *ret = ReANY(dstr);
21484 PERL_ARGS_ASSERT_RE_DUP_GUTS;
21486 npar = r->nparens+1;
21487 Newx(ret->offs, npar, regexp_paren_pair);
21488 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
21490 if (ret->substrs) {
21491 /* Do it this way to avoid reading from *r after the StructCopy().
21492 That way, if any of the sv_dup_inc()s dislodge *r from the L1
21493 cache, it doesn't matter. */
21495 const bool anchored = r->check_substr
21496 ? r->check_substr == r->substrs->data[0].substr
21497 : r->check_utf8 == r->substrs->data[0].utf8_substr;
21498 Newx(ret->substrs, 1, struct reg_substr_data);
21499 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
21501 for (i = 0; i < 2; i++) {
21502 ret->substrs->data[i].substr =
21503 sv_dup_inc(ret->substrs->data[i].substr, param);
21504 ret->substrs->data[i].utf8_substr =
21505 sv_dup_inc(ret->substrs->data[i].utf8_substr, param);
21508 /* check_substr and check_utf8, if non-NULL, point to either their
21509 anchored or float namesakes, and don't hold a second reference. */
21511 if (ret->check_substr) {
21513 assert(r->check_utf8 == r->substrs->data[0].utf8_substr);
21515 ret->check_substr = ret->substrs->data[0].substr;
21516 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21518 assert(r->check_substr == r->substrs->data[1].substr);
21519 assert(r->check_utf8 == r->substrs->data[1].utf8_substr);
21521 ret->check_substr = ret->substrs->data[1].substr;
21522 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21524 } else if (ret->check_utf8) {
21526 ret->check_utf8 = ret->substrs->data[0].utf8_substr;
21528 ret->check_utf8 = ret->substrs->data[1].utf8_substr;
21533 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
21534 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
21535 if (r->recurse_locinput)
21536 Newx(ret->recurse_locinput, r->nparens + 1, char *);
21539 RXi_SET(ret, CALLREGDUPE_PVT(dstr, param));
21541 if (RX_MATCH_COPIED(dstr))
21542 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
21544 ret->subbeg = NULL;
21545 #ifdef PERL_ANY_COW
21546 ret->saved_copy = NULL;
21549 /* Whether mother_re be set or no, we need to copy the string. We
21550 cannot refrain from copying it when the storage points directly to
21551 our mother regexp, because that's
21552 1: a buffer in a different thread
21553 2: something we no longer hold a reference on
21554 so we need to copy it locally. */
21555 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED_const(sstr), SvCUR(sstr)+1);
21556 /* set malloced length to a non-zero value so it will be freed
21557 * (otherwise in combination with SVf_FAKE it looks like an alien
21558 * buffer). It doesn't have to be the actual malloced size, since it
21559 * should never be grown */
21560 SvLEN_set(dstr, SvCUR(sstr)+1);
21561 ret->mother_re = NULL;
21563 #endif /* PERL_IN_XSUB_RE */
21568 This is the internal complement to regdupe() which is used to copy
21569 the structure pointed to by the *pprivate pointer in the regexp.
21570 This is the core version of the extension overridable cloning hook.
21571 The regexp structure being duplicated will be copied by perl prior
21572 to this and will be provided as the regexp *r argument, however
21573 with the /old/ structures pprivate pointer value. Thus this routine
21574 may override any copying normally done by perl.
21576 It returns a pointer to the new regexp_internal structure.
21580 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
21583 struct regexp *const r = ReANY(rx);
21584 regexp_internal *reti;
21586 RXi_GET_DECL(r, ri);
21588 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
21592 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
21593 char, regexp_internal);
21594 Copy(ri->program, reti->program, len+1, regnode);
21597 if (ri->code_blocks) {
21599 Newx(reti->code_blocks, 1, struct reg_code_blocks);
21600 Newx(reti->code_blocks->cb, ri->code_blocks->count,
21601 struct reg_code_block);
21602 Copy(ri->code_blocks->cb, reti->code_blocks->cb,
21603 ri->code_blocks->count, struct reg_code_block);
21604 for (n = 0; n < ri->code_blocks->count; n++)
21605 reti->code_blocks->cb[n].src_regex = (REGEXP*)
21606 sv_dup_inc((SV*)(ri->code_blocks->cb[n].src_regex), param);
21607 reti->code_blocks->count = ri->code_blocks->count;
21608 reti->code_blocks->refcnt = 1;
21611 reti->code_blocks = NULL;
21613 reti->regstclass = NULL;
21616 struct reg_data *d;
21617 const int count = ri->data->count;
21620 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
21621 char, struct reg_data);
21622 Newx(d->what, count, U8);
21625 for (i = 0; i < count; i++) {
21626 d->what[i] = ri->data->what[i];
21627 switch (d->what[i]) {
21628 /* see also regcomp.h and regfree_internal() */
21629 case 'a': /* actually an AV, but the dup function is identical.
21630 values seem to be "plain sv's" generally. */
21631 case 'r': /* a compiled regex (but still just another SV) */
21632 case 's': /* an RV (currently only used for an RV to an AV by the ANYOF code)
21633 this use case should go away, the code could have used
21634 'a' instead - see S_set_ANYOF_arg() for array contents. */
21635 case 'S': /* actually an SV, but the dup function is identical. */
21636 case 'u': /* actually an HV, but the dup function is identical.
21637 values are "plain sv's" */
21638 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
21641 /* Synthetic Start Class - "Fake" charclass we generate to optimize
21642 * patterns which could start with several different things. Pre-TRIE
21643 * this was more important than it is now, however this still helps
21644 * in some places, for instance /x?a+/ might produce a SSC equivalent
21645 * to [xa]. This is used by Perl_re_intuit_start() and S_find_byclass()
21648 /* This is cheating. */
21649 Newx(d->data[i], 1, regnode_ssc);
21650 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
21651 reti->regstclass = (regnode*)d->data[i];
21654 /* AHO-CORASICK fail table */
21655 /* Trie stclasses are readonly and can thus be shared
21656 * without duplication. We free the stclass in pregfree
21657 * when the corresponding reg_ac_data struct is freed.
21659 reti->regstclass= ri->regstclass;
21662 /* TRIE transition table */
21664 ((reg_trie_data*)ri->data->data[i])->refcount++;
21667 case 'l': /* (?{...}) or (??{ ... }) code (cb->block) */
21668 case 'L': /* same when RExC_pm_flags & PMf_HAS_CV and code
21669 is not from another regexp */
21670 d->data[i] = ri->data->data[i];
21673 Perl_croak(aTHX_ "panic: re_dup_guts unknown data code '%c'",
21674 ri->data->what[i]);
21683 reti->name_list_idx = ri->name_list_idx;
21685 #ifdef RE_TRACK_PATTERN_OFFSETS
21686 if (ri->u.offsets) {
21687 Newx(reti->u.offsets, 2*len+1, U32);
21688 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
21691 SetProgLen(reti, len);
21694 return (void*)reti;
21697 #endif /* USE_ITHREADS */
21699 #ifndef PERL_IN_XSUB_RE
21702 - regnext - dig the "next" pointer out of a node
21705 Perl_regnext(pTHX_ regnode *p)
21712 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
21713 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
21714 (int)OP(p), (int)REGNODE_MAX);
21717 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
21727 S_re_croak2(pTHX_ bool utf8, const char* pat1, const char* pat2,...)
21730 STRLEN l1 = strlen(pat1);
21731 STRLEN l2 = strlen(pat2);
21734 const char *message;
21736 PERL_ARGS_ASSERT_RE_CROAK2;
21742 Copy(pat1, buf, l1 , char);
21743 Copy(pat2, buf + l1, l2 , char);
21744 buf[l1 + l2] = '\n';
21745 buf[l1 + l2 + 1] = '\0';
21746 va_start(args, pat2);
21747 msv = vmess(buf, &args);
21749 message = SvPV_const(msv, l1);
21752 Copy(message, buf, l1 , char);
21753 /* l1-1 to avoid \n */
21754 Perl_croak(aTHX_ "%" UTF8f, UTF8fARG(utf8, l1-1, buf));
21757 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
21759 #ifndef PERL_IN_XSUB_RE
21761 Perl_save_re_context(pTHX)
21766 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
21769 const REGEXP * const rx = PM_GETRE(PL_curpm);
21771 nparens = RX_NPARENS(rx);
21774 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
21775 * that PL_curpm will be null, but that utf8.pm and the modules it
21776 * loads will only use $1..$3.
21777 * The t/porting/re_context.t test file checks this assumption.
21782 for (i = 1; i <= nparens; i++) {
21783 char digits[TYPE_CHARS(long)];
21784 const STRLEN len = my_snprintf(digits, sizeof(digits),
21786 GV *const *const gvp
21787 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
21790 GV * const gv = *gvp;
21791 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
21801 S_put_code_point(pTHX_ SV *sv, UV c)
21803 PERL_ARGS_ASSERT_PUT_CODE_POINT;
21806 Perl_sv_catpvf(aTHX_ sv, "\\x{%04" UVXf "}", c);
21808 else if (isPRINT(c)) {
21809 const char string = (char) c;
21811 /* We use {phrase} as metanotation in the class, so also escape literal
21813 if (isBACKSLASHED_PUNCT(c) || c == '{' || c == '}')
21814 sv_catpvs(sv, "\\");
21815 sv_catpvn(sv, &string, 1);
21817 else if (isMNEMONIC_CNTRL(c)) {
21818 Perl_sv_catpvf(aTHX_ sv, "%s", cntrl_to_mnemonic((U8) c));
21821 Perl_sv_catpvf(aTHX_ sv, "\\x%02X", (U8) c);
21825 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
21828 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
21830 /* Appends to 'sv' a displayable version of the range of code points from
21831 * 'start' to 'end'. Mnemonics (like '\r') are used for the few controls
21832 * that have them, when they occur at the beginning or end of the range.
21833 * It uses hex to output the remaining code points, unless 'allow_literals'
21834 * is true, in which case the printable ASCII ones are output as-is (though
21835 * some of these will be escaped by put_code_point()).
21837 * NOTE: This is designed only for printing ranges of code points that fit
21838 * inside an ANYOF bitmap. Higher code points are simply suppressed
21841 const unsigned int min_range_count = 3;
21843 assert(start <= end);
21845 PERL_ARGS_ASSERT_PUT_RANGE;
21847 while (start <= end) {
21849 const char * format;
21851 if (end - start < min_range_count) {
21853 /* Output chars individually when they occur in short ranges */
21854 for (; start <= end; start++) {
21855 put_code_point(sv, start);
21860 /* If permitted by the input options, and there is a possibility that
21861 * this range contains a printable literal, look to see if there is
21863 if (allow_literals && start <= MAX_PRINT_A) {
21865 /* If the character at the beginning of the range isn't an ASCII
21866 * printable, effectively split the range into two parts:
21867 * 1) the portion before the first such printable,
21869 * and output them separately. */
21870 if (! isPRINT_A(start)) {
21871 UV temp_end = start + 1;
21873 /* There is no point looking beyond the final possible
21874 * printable, in MAX_PRINT_A */
21875 UV max = MIN(end, MAX_PRINT_A);
21877 while (temp_end <= max && ! isPRINT_A(temp_end)) {
21881 /* Here, temp_end points to one beyond the first printable if
21882 * found, or to one beyond 'max' if not. If none found, make
21883 * sure that we use the entire range */
21884 if (temp_end > MAX_PRINT_A) {
21885 temp_end = end + 1;
21888 /* Output the first part of the split range: the part that
21889 * doesn't have printables, with the parameter set to not look
21890 * for literals (otherwise we would infinitely recurse) */
21891 put_range(sv, start, temp_end - 1, FALSE);
21893 /* The 2nd part of the range (if any) starts here. */
21896 /* We do a continue, instead of dropping down, because even if
21897 * the 2nd part is non-empty, it could be so short that we want
21898 * to output it as individual characters, as tested for at the
21899 * top of this loop. */
21903 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
21904 * output a sub-range of just the digits or letters, then process
21905 * the remaining portion as usual. */
21906 if (isALPHANUMERIC_A(start)) {
21907 UV mask = (isDIGIT_A(start))
21912 UV temp_end = start + 1;
21914 /* Find the end of the sub-range that includes just the
21915 * characters in the same class as the first character in it */
21916 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
21921 /* For short ranges, don't duplicate the code above to output
21922 * them; just call recursively */
21923 if (temp_end - start < min_range_count) {
21924 put_range(sv, start, temp_end, FALSE);
21926 else { /* Output as a range */
21927 put_code_point(sv, start);
21928 sv_catpvs(sv, "-");
21929 put_code_point(sv, temp_end);
21931 start = temp_end + 1;
21935 /* We output any other printables as individual characters */
21936 if (isPUNCT_A(start) || isSPACE_A(start)) {
21937 while (start <= end && (isPUNCT_A(start)
21938 || isSPACE_A(start)))
21940 put_code_point(sv, start);
21945 } /* End of looking for literals */
21947 /* Here is not to output as a literal. Some control characters have
21948 * mnemonic names. Split off any of those at the beginning and end of
21949 * the range to print mnemonically. It isn't possible for many of
21950 * these to be in a row, so this won't overwhelm with output */
21952 && (isMNEMONIC_CNTRL(start) || isMNEMONIC_CNTRL(end)))
21954 while (isMNEMONIC_CNTRL(start) && start <= end) {
21955 put_code_point(sv, start);
21959 /* If this didn't take care of the whole range ... */
21960 if (start <= end) {
21962 /* Look backwards from the end to find the final non-mnemonic
21965 while (isMNEMONIC_CNTRL(temp_end)) {
21969 /* And separately output the interior range that doesn't start
21970 * or end with mnemonics */
21971 put_range(sv, start, temp_end, FALSE);
21973 /* Then output the mnemonic trailing controls */
21974 start = temp_end + 1;
21975 while (start <= end) {
21976 put_code_point(sv, start);
21983 /* As a final resort, output the range or subrange as hex. */
21985 if (start >= NUM_ANYOF_CODE_POINTS) {
21988 else { /* Have to split range at the bitmap boundary */
21989 this_end = (end < NUM_ANYOF_CODE_POINTS)
21991 : NUM_ANYOF_CODE_POINTS - 1;
21993 #if NUM_ANYOF_CODE_POINTS > 256
21994 format = (this_end < 256)
21995 ? "\\x%02" UVXf "-\\x%02" UVXf
21996 : "\\x{%04" UVXf "}-\\x{%04" UVXf "}";
21998 format = "\\x%02" UVXf "-\\x%02" UVXf;
22000 GCC_DIAG_IGNORE_STMT(-Wformat-nonliteral);
22001 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
22002 GCC_DIAG_RESTORE_STMT;
22008 S_put_charclass_bitmap_innards_invlist(pTHX_ SV *sv, SV* invlist)
22010 /* Concatenate onto the PV in 'sv' a displayable form of the inversion list
22014 bool allow_literals = TRUE;
22016 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_INVLIST;
22018 /* Generally, it is more readable if printable characters are output as
22019 * literals, but if a range (nearly) spans all of them, it's best to output
22020 * it as a single range. This code will use a single range if all but 2
22021 * ASCII printables are in it */
22022 invlist_iterinit(invlist);
22023 while (invlist_iternext(invlist, &start, &end)) {
22025 /* If the range starts beyond the final printable, it doesn't have any
22027 if (start > MAX_PRINT_A) {
22031 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
22032 * all but two, the range must start and end no later than 2 from
22034 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
22035 if (end > MAX_PRINT_A) {
22041 if (end - start >= MAX_PRINT_A - ' ' - 2) {
22042 allow_literals = FALSE;
22047 invlist_iterfinish(invlist);
22049 /* Here we have figured things out. Output each range */
22050 invlist_iterinit(invlist);
22051 while (invlist_iternext(invlist, &start, &end)) {
22052 if (start >= NUM_ANYOF_CODE_POINTS) {
22055 put_range(sv, start, end, allow_literals);
22057 invlist_iterfinish(invlist);
22063 S_put_charclass_bitmap_innards_common(pTHX_
22064 SV* invlist, /* The bitmap */
22065 SV* posixes, /* Under /l, things like [:word:], \S */
22066 SV* only_utf8, /* Under /d, matches iff the target is UTF-8 */
22067 SV* not_utf8, /* /d, matches iff the target isn't UTF-8 */
22068 SV* only_utf8_locale, /* Under /l, matches if the locale is UTF-8 */
22069 const bool invert /* Is the result to be inverted? */
22072 /* Create and return an SV containing a displayable version of the bitmap
22073 * and associated information determined by the input parameters. If the
22074 * output would have been only the inversion indicator '^', NULL is instead
22080 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS_COMMON;
22083 output = newSVpvs("^");
22086 output = newSVpvs("");
22089 /* First, the code points in the bitmap that are unconditionally there */
22090 put_charclass_bitmap_innards_invlist(output, invlist);
22092 /* Traditionally, these have been placed after the main code points */
22094 sv_catsv(output, posixes);
22097 if (only_utf8 && _invlist_len(only_utf8)) {
22098 Perl_sv_catpvf(aTHX_ output, "%s{utf8}%s", PL_colors[1], PL_colors[0]);
22099 put_charclass_bitmap_innards_invlist(output, only_utf8);
22102 if (not_utf8 && _invlist_len(not_utf8)) {
22103 Perl_sv_catpvf(aTHX_ output, "%s{not utf8}%s", PL_colors[1], PL_colors[0]);
22104 put_charclass_bitmap_innards_invlist(output, not_utf8);
22107 if (only_utf8_locale && _invlist_len(only_utf8_locale)) {
22108 Perl_sv_catpvf(aTHX_ output, "%s{utf8 locale}%s", PL_colors[1], PL_colors[0]);
22109 put_charclass_bitmap_innards_invlist(output, only_utf8_locale);
22111 /* This is the only list in this routine that can legally contain code
22112 * points outside the bitmap range. The call just above to
22113 * 'put_charclass_bitmap_innards_invlist' will simply suppress them, so
22114 * output them here. There's about a half-dozen possible, and none in
22115 * contiguous ranges longer than 2 */
22116 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22118 SV* above_bitmap = NULL;
22120 _invlist_subtract(only_utf8_locale, PL_InBitmap, &above_bitmap);
22122 invlist_iterinit(above_bitmap);
22123 while (invlist_iternext(above_bitmap, &start, &end)) {
22126 for (i = start; i <= end; i++) {
22127 put_code_point(output, i);
22130 invlist_iterfinish(above_bitmap);
22131 SvREFCNT_dec_NN(above_bitmap);
22135 if (invert && SvCUR(output) == 1) {
22143 S_put_charclass_bitmap_innards(pTHX_ SV *sv,
22145 SV *nonbitmap_invlist,
22146 SV *only_utf8_locale_invlist,
22147 const regnode * const node,
22149 const bool force_as_is_display)
22151 /* Appends to 'sv' a displayable version of the innards of the bracketed
22152 * character class defined by the other arguments:
22153 * 'bitmap' points to the bitmap, or NULL if to ignore that.
22154 * 'nonbitmap_invlist' is an inversion list of the code points that are in
22155 * the bitmap range, but for some reason aren't in the bitmap; NULL if
22156 * none. The reasons for this could be that they require some
22157 * condition such as the target string being or not being in UTF-8
22158 * (under /d), or because they came from a user-defined property that
22159 * was not resolved at the time of the regex compilation (under /u)
22160 * 'only_utf8_locale_invlist' is an inversion list of the code points that
22161 * are valid only if the runtime locale is a UTF-8 one; NULL if none
22162 * 'node' is the regex pattern ANYOF node. It is needed only when the
22163 * above two parameters are not null, and is passed so that this
22164 * routine can tease apart the various reasons for them.
22165 * 'flags' is the flags field of 'node'
22166 * 'force_as_is_display' is TRUE if this routine should definitely NOT try
22167 * to invert things to see if that leads to a cleaner display. If
22168 * FALSE, this routine is free to use its judgment about doing this.
22170 * It returns TRUE if there was actually something output. (It may be that
22171 * the bitmap, etc is empty.)
22173 * When called for outputting the bitmap of a non-ANYOF node, just pass the
22174 * bitmap, with the succeeding parameters set to NULL, and the final one to
22178 /* In general, it tries to display the 'cleanest' representation of the
22179 * innards, choosing whether to display them inverted or not, regardless of
22180 * whether the class itself is to be inverted. However, there are some
22181 * cases where it can't try inverting, as what actually matches isn't known
22182 * until runtime, and hence the inversion isn't either. */
22185 bool inverting_allowed = ! force_as_is_display;
22188 STRLEN orig_sv_cur = SvCUR(sv);
22190 SV* invlist; /* Inversion list we accumulate of code points that
22191 are unconditionally matched */
22192 SV* only_utf8 = NULL; /* Under /d, list of matches iff the target is
22194 SV* not_utf8 = NULL; /* /d, list of matches iff the target isn't UTF-8
22196 SV* posixes = NULL; /* Under /l, string of things like [:word:], \D */
22197 SV* only_utf8_locale = NULL; /* Under /l, list of matches if the locale
22200 SV* as_is_display; /* The output string when we take the inputs
22202 SV* inverted_display; /* The output string when we invert the inputs */
22204 bool invert = cBOOL(flags & ANYOF_INVERT); /* Is the input to be inverted
22206 /* We are biased in favor of displaying things without them being inverted,
22207 * as that is generally easier to understand */
22208 const int bias = 5;
22210 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
22212 /* Start off with whatever code points are passed in. (We clone, so we
22213 * don't change the caller's list) */
22214 if (nonbitmap_invlist) {
22215 assert(invlist_highest(nonbitmap_invlist) < NUM_ANYOF_CODE_POINTS);
22216 invlist = invlist_clone(nonbitmap_invlist, NULL);
22218 else { /* Worst case size is every other code point is matched */
22219 invlist = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
22223 if (OP(node) == ANYOFD) {
22225 /* This flag indicates that the code points below 0x100 in the
22226 * nonbitmap list are precisely the ones that match only when the
22227 * target is UTF-8 (they should all be non-ASCII). */
22228 if (flags & ANYOF_SHARED_d_UPPER_LATIN1_UTF8_STRING_MATCHES_non_d_RUNTIME_USER_PROP)
22230 _invlist_intersection(invlist, PL_UpperLatin1, &only_utf8);
22231 _invlist_subtract(invlist, only_utf8, &invlist);
22234 /* And this flag for matching all non-ASCII 0xFF and below */
22235 if (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)
22237 not_utf8 = invlist_clone(PL_UpperLatin1, NULL);
22240 else if (OP(node) == ANYOFL || OP(node) == ANYOFPOSIXL) {
22242 /* If either of these flags are set, what matches isn't
22243 * determinable except during execution, so don't know enough here
22245 if (flags & (ANYOFL_FOLD|ANYOF_MATCHES_POSIXL)) {
22246 inverting_allowed = FALSE;
22249 /* What the posix classes match also varies at runtime, so these
22250 * will be output symbolically. */
22251 if (ANYOF_POSIXL_TEST_ANY_SET(node)) {
22254 posixes = newSVpvs("");
22255 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
22256 if (ANYOF_POSIXL_TEST(node, i)) {
22257 sv_catpv(posixes, anyofs[i]);
22264 /* Accumulate the bit map into the unconditional match list */
22266 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
22267 if (BITMAP_TEST(bitmap, i)) {
22270 i < NUM_ANYOF_CODE_POINTS && BITMAP_TEST(bitmap, i);
22273 invlist = _add_range_to_invlist(invlist, start, i-1);
22278 /* Make sure that the conditional match lists don't have anything in them
22279 * that match unconditionally; otherwise the output is quite confusing.
22280 * This could happen if the code that populates these misses some
22283 _invlist_subtract(only_utf8, invlist, &only_utf8);
22286 _invlist_subtract(not_utf8, invlist, ¬_utf8);
22289 if (only_utf8_locale_invlist) {
22291 /* Since this list is passed in, we have to make a copy before
22293 only_utf8_locale = invlist_clone(only_utf8_locale_invlist, NULL);
22295 _invlist_subtract(only_utf8_locale, invlist, &only_utf8_locale);
22297 /* And, it can get really weird for us to try outputting an inverted
22298 * form of this list when it has things above the bitmap, so don't even
22300 if (invlist_highest(only_utf8_locale) >= NUM_ANYOF_CODE_POINTS) {
22301 inverting_allowed = FALSE;
22305 /* Calculate what the output would be if we take the input as-is */
22306 as_is_display = put_charclass_bitmap_innards_common(invlist,
22313 /* If have to take the output as-is, just do that */
22314 if (! inverting_allowed) {
22315 if (as_is_display) {
22316 sv_catsv(sv, as_is_display);
22317 SvREFCNT_dec_NN(as_is_display);
22320 else { /* But otherwise, create the output again on the inverted input, and
22321 use whichever version is shorter */
22323 int inverted_bias, as_is_bias;
22325 /* We will apply our bias to whichever of the the results doesn't have
22335 inverted_bias = bias;
22338 /* Now invert each of the lists that contribute to the output,
22339 * excluding from the result things outside the possible range */
22341 /* For the unconditional inversion list, we have to add in all the
22342 * conditional code points, so that when inverted, they will be gone
22344 _invlist_union(only_utf8, invlist, &invlist);
22345 _invlist_union(not_utf8, invlist, &invlist);
22346 _invlist_union(only_utf8_locale, invlist, &invlist);
22347 _invlist_invert(invlist);
22348 _invlist_intersection(invlist, PL_InBitmap, &invlist);
22351 _invlist_invert(only_utf8);
22352 _invlist_intersection(only_utf8, PL_UpperLatin1, &only_utf8);
22354 else if (not_utf8) {
22356 /* If a code point matches iff the target string is not in UTF-8,
22357 * then complementing the result has it not match iff not in UTF-8,
22358 * which is the same thing as matching iff it is UTF-8. */
22359 only_utf8 = not_utf8;
22363 if (only_utf8_locale) {
22364 _invlist_invert(only_utf8_locale);
22365 _invlist_intersection(only_utf8_locale,
22367 &only_utf8_locale);
22370 inverted_display = put_charclass_bitmap_innards_common(
22375 only_utf8_locale, invert);
22377 /* Use the shortest representation, taking into account our bias
22378 * against showing it inverted */
22379 if ( inverted_display
22380 && ( ! as_is_display
22381 || ( SvCUR(inverted_display) + inverted_bias
22382 < SvCUR(as_is_display) + as_is_bias)))
22384 sv_catsv(sv, inverted_display);
22386 else if (as_is_display) {
22387 sv_catsv(sv, as_is_display);
22390 SvREFCNT_dec(as_is_display);
22391 SvREFCNT_dec(inverted_display);
22394 SvREFCNT_dec_NN(invlist);
22395 SvREFCNT_dec(only_utf8);
22396 SvREFCNT_dec(not_utf8);
22397 SvREFCNT_dec(posixes);
22398 SvREFCNT_dec(only_utf8_locale);
22400 return SvCUR(sv) > orig_sv_cur;
22403 #define CLEAR_OPTSTART \
22404 if (optstart) STMT_START { \
22405 DEBUG_OPTIMISE_r(Perl_re_printf( aTHX_ \
22406 " (%" IVdf " nodes)\n", (IV)(node - optstart))); \
22410 #define DUMPUNTIL(b,e) \
22412 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
22414 STATIC const regnode *
22415 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
22416 const regnode *last, const regnode *plast,
22417 SV* sv, I32 indent, U32 depth)
22419 U8 op = PSEUDO; /* Arbitrary non-END op. */
22420 const regnode *next;
22421 const regnode *optstart= NULL;
22423 RXi_GET_DECL(r, ri);
22424 GET_RE_DEBUG_FLAGS_DECL;
22426 PERL_ARGS_ASSERT_DUMPUNTIL;
22428 #ifdef DEBUG_DUMPUNTIL
22429 Perl_re_printf( aTHX_ "--- %d : %d - %d - %d\n", indent, node-start,
22430 last ? last-start : 0, plast ? plast-start : 0);
22433 if (plast && plast < last)
22436 while (PL_regkind[op] != END && (!last || node < last)) {
22438 /* While that wasn't END last time... */
22441 if (op == CLOSE || op == SRCLOSE || op == WHILEM)
22443 next = regnext((regnode *)node);
22446 if (OP(node) == OPTIMIZED) {
22447 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
22454 regprop(r, sv, node, NULL, NULL);
22455 Perl_re_printf( aTHX_ "%4" IVdf ":%*s%s", (IV)(node - start),
22456 (int)(2*indent + 1), "", SvPVX_const(sv));
22458 if (OP(node) != OPTIMIZED) {
22459 if (next == NULL) /* Next ptr. */
22460 Perl_re_printf( aTHX_ " (0)");
22461 else if (PL_regkind[(U8)op] == BRANCH
22462 && PL_regkind[OP(next)] != BRANCH )
22463 Perl_re_printf( aTHX_ " (FAIL)");
22465 Perl_re_printf( aTHX_ " (%" IVdf ")", (IV)(next - start));
22466 Perl_re_printf( aTHX_ "\n");
22470 if (PL_regkind[(U8)op] == BRANCHJ) {
22473 const regnode *nnode = (OP(next) == LONGJMP
22474 ? regnext((regnode *)next)
22476 if (last && nnode > last)
22478 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
22481 else if (PL_regkind[(U8)op] == BRANCH) {
22483 DUMPUNTIL(NEXTOPER(node), next);
22485 else if ( PL_regkind[(U8)op] == TRIE ) {
22486 const regnode *this_trie = node;
22487 const char op = OP(node);
22488 const U32 n = ARG(node);
22489 const reg_ac_data * const ac = op>=AHOCORASICK ?
22490 (reg_ac_data *)ri->data->data[n] :
22492 const reg_trie_data * const trie =
22493 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
22495 AV *const trie_words
22496 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
22498 const regnode *nextbranch= NULL;
22501 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
22502 SV ** const elem_ptr = av_fetch(trie_words, word_idx, 0);
22504 Perl_re_indentf( aTHX_ "%s ",
22507 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
22508 SvCUR(*elem_ptr), PL_dump_re_max_len,
22509 PL_colors[0], PL_colors[1],
22511 ? PERL_PV_ESCAPE_UNI
22513 | PERL_PV_PRETTY_ELLIPSES
22514 | PERL_PV_PRETTY_LTGT
22519 U16 dist= trie->jump[word_idx+1];
22520 Perl_re_printf( aTHX_ "(%" UVuf ")\n",
22521 (UV)((dist ? this_trie + dist : next) - start));
22524 nextbranch= this_trie + trie->jump[0];
22525 DUMPUNTIL(this_trie + dist, nextbranch);
22527 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
22528 nextbranch= regnext((regnode *)nextbranch);
22530 Perl_re_printf( aTHX_ "\n");
22533 if (last && next > last)
22538 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
22539 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
22540 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
22542 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
22544 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
22546 else if ( op == PLUS || op == STAR) {
22547 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
22549 else if (PL_regkind[(U8)op] == EXACT || op == ANYOFHs) {
22550 /* Literal string, where present. */
22551 node += NODE_SZ_STR(node) - 1;
22552 node = NEXTOPER(node);
22555 node = NEXTOPER(node);
22556 node += regarglen[(U8)op];
22558 if (op == CURLYX || op == OPEN || op == SROPEN)
22562 #ifdef DEBUG_DUMPUNTIL
22563 Perl_re_printf( aTHX_ "--- %d\n", (int)indent);
22568 #endif /* DEBUGGING */
22570 #ifndef PERL_IN_XSUB_RE
22572 #include "uni_keywords.h"
22575 Perl_init_uniprops(pTHX)
22580 char * dump_len_string;
22582 dump_len_string = PerlEnv_getenv("PERL_DUMP_RE_MAX_LEN");
22583 if ( ! dump_len_string
22584 || ! grok_atoUV(dump_len_string, (UV *)&PL_dump_re_max_len, NULL))
22586 PL_dump_re_max_len = 60; /* A reasonable default */
22590 PL_user_def_props = newHV();
22592 #ifdef USE_ITHREADS
22594 HvSHAREKEYS_off(PL_user_def_props);
22595 PL_user_def_props_aTHX = aTHX;
22599 /* Set up the inversion list interpreter-level variables */
22601 PL_XPosix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22602 PL_XPosix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALNUM]);
22603 PL_XPosix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXALPHA]);
22604 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXBLANK]);
22605 PL_XPosix_ptrs[_CC_CASED] = _new_invlist_C_array(uni_prop_ptrs[UNI_CASED]);
22606 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXCNTRL]);
22607 PL_XPosix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXDIGIT]);
22608 PL_XPosix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXGRAPH]);
22609 PL_XPosix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXLOWER]);
22610 PL_XPosix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPRINT]);
22611 PL_XPosix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXPUNCT]);
22612 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXSPACE]);
22613 PL_XPosix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXUPPER]);
22614 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_VERTSPACE]);
22615 PL_XPosix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXWORD]);
22616 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_XPOSIXXDIGIT]);
22618 PL_Posix_ptrs[_CC_ASCII] = _new_invlist_C_array(uni_prop_ptrs[UNI_ASCII]);
22619 PL_Posix_ptrs[_CC_ALPHANUMERIC] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALNUM]);
22620 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXALPHA]);
22621 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXBLANK]);
22622 PL_Posix_ptrs[_CC_CASED] = PL_Posix_ptrs[_CC_ALPHA];
22623 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXCNTRL]);
22624 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXDIGIT]);
22625 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXGRAPH]);
22626 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXLOWER]);
22627 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPRINT]);
22628 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXPUNCT]);
22629 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXSPACE]);
22630 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXUPPER]);
22631 PL_Posix_ptrs[_CC_VERTSPACE] = NULL;
22632 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXWORD]);
22633 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(uni_prop_ptrs[UNI_POSIXXDIGIT]);
22635 PL_GCB_invlist = _new_invlist_C_array(_Perl_GCB_invlist);
22636 PL_SB_invlist = _new_invlist_C_array(_Perl_SB_invlist);
22637 PL_WB_invlist = _new_invlist_C_array(_Perl_WB_invlist);
22638 PL_LB_invlist = _new_invlist_C_array(_Perl_LB_invlist);
22639 PL_SCX_invlist = _new_invlist_C_array(_Perl_SCX_invlist);
22641 PL_InBitmap = _new_invlist_C_array(_Perl_InBitmap_invlist);
22642 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
22643 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
22644 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
22646 PL_Assigned_invlist = _new_invlist_C_array(uni_prop_ptrs[UNI_ASSIGNED]);
22648 PL_utf8_perl_idstart = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDSTART]);
22649 PL_utf8_perl_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_IDCONT]);
22651 PL_utf8_charname_begin = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_BEGIN]);
22652 PL_utf8_charname_continue = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_CHARNAME_CONTINUE]);
22654 PL_in_some_fold = _new_invlist_C_array(uni_prop_ptrs[UNI__PERL_ANY_FOLDS]);
22655 PL_HasMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22656 UNI__PERL_FOLDS_TO_MULTI_CHAR]);
22657 PL_InMultiCharFold = _new_invlist_C_array(uni_prop_ptrs[
22658 UNI__PERL_IS_IN_MULTI_CHAR_FOLD]);
22659 PL_utf8_toupper = _new_invlist_C_array(Uppercase_Mapping_invlist);
22660 PL_utf8_tolower = _new_invlist_C_array(Lowercase_Mapping_invlist);
22661 PL_utf8_totitle = _new_invlist_C_array(Titlecase_Mapping_invlist);
22662 PL_utf8_tofold = _new_invlist_C_array(Case_Folding_invlist);
22663 PL_utf8_tosimplefold = _new_invlist_C_array(Simple_Case_Folding_invlist);
22664 PL_utf8_foldclosures = _new_invlist_C_array(_Perl_IVCF_invlist);
22665 PL_utf8_mark = _new_invlist_C_array(uni_prop_ptrs[UNI_M]);
22666 PL_CCC_non0_non230 = _new_invlist_C_array(_Perl_CCC_non0_non230_invlist);
22667 PL_Private_Use = _new_invlist_C_array(uni_prop_ptrs[UNI_CO]);
22670 /* The below are used only by deprecated functions. They could be removed */
22671 PL_utf8_xidcont = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDC]);
22672 PL_utf8_idcont = _new_invlist_C_array(uni_prop_ptrs[UNI_IDC]);
22673 PL_utf8_xidstart = _new_invlist_C_array(uni_prop_ptrs[UNI_XIDS]);
22679 This code was mainly added for backcompat to give a warning for non-portable
22680 code points in user-defined properties. But experiments showed that the
22681 warning in earlier perls were only omitted on overflow, which should be an
22682 error, so there really isnt a backcompat issue, and actually adding the
22683 warning when none was present before might cause breakage, for little gain. So
22684 khw left this code in, but not enabled. Tests were never added.
22687 Ei |const char *|get_extended_utf8_msg|const UV cp
22689 PERL_STATIC_INLINE const char *
22690 S_get_extended_utf8_msg(pTHX_ const UV cp)
22692 U8 dummy[UTF8_MAXBYTES + 1];
22696 uvchr_to_utf8_flags_msgs(dummy, cp, UNICODE_WARN_PERL_EXTENDED,
22699 msg = hv_fetchs(msgs, "text", 0);
22702 (void) sv_2mortal((SV *) msgs);
22704 return SvPVX(*msg);
22710 Perl_handle_user_defined_property(pTHX_
22712 /* Parses the contents of a user-defined property definition; returning the
22713 * expanded definition if possible. If so, the return is an inversion
22716 * If there are subroutines that are part of the expansion and which aren't
22717 * known at the time of the call to this function, this returns what
22718 * parse_uniprop_string() returned for the first one encountered.
22720 * If an error was found, NULL is returned, and 'msg' gets a suitable
22721 * message appended to it. (Appending allows the back trace of how we got
22722 * to the faulty definition to be displayed through nested calls of
22723 * user-defined subs.)
22725 * The caller IS responsible for freeing any returned SV.
22727 * The syntax of the contents is pretty much described in perlunicode.pod,
22728 * but we also allow comments on each line */
22730 const char * name, /* Name of property */
22731 const STRLEN name_len, /* The name's length in bytes */
22732 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
22733 const bool to_fold, /* ? Is this under /i */
22734 const bool runtime, /* ? Are we in compile- or run-time */
22735 const bool deferrable, /* Is it ok for this property's full definition
22736 to be deferred until later? */
22737 SV* contents, /* The property's definition */
22738 bool *user_defined_ptr, /* This will be set TRUE as we wouldn't be
22739 getting called unless this is thought to be
22740 a user-defined property */
22741 SV * msg, /* Any error or warning msg(s) are appended to
22743 const STRLEN level) /* Recursion level of this call */
22746 const char * string = SvPV_const(contents, len);
22747 const char * const e = string + len;
22748 const bool is_contents_utf8 = cBOOL(SvUTF8(contents));
22749 const STRLEN msgs_length_on_entry = SvCUR(msg);
22751 const char * s0 = string; /* Points to first byte in the current line
22752 being parsed in 'string' */
22753 const char overflow_msg[] = "Code point too large in \"";
22754 SV* running_definition = NULL;
22756 PERL_ARGS_ASSERT_HANDLE_USER_DEFINED_PROPERTY;
22758 *user_defined_ptr = TRUE;
22760 /* Look at each line */
22762 const char * s; /* Current byte */
22763 char op = '+'; /* Default operation is 'union' */
22764 IV min = 0; /* range begin code point */
22765 IV max = -1; /* and range end */
22766 SV* this_definition;
22768 /* Skip comment lines */
22770 s0 = strchr(s0, '\n');
22778 /* For backcompat, allow an empty first line */
22784 /* First character in the line may optionally be the operation */
22793 /* If the line is one or two hex digits separated by blank space, its
22794 * a range; otherwise it is either another user-defined property or an
22799 if (! isXDIGIT(*s)) {
22800 goto check_if_property;
22803 do { /* Each new hex digit will add 4 bits. */
22804 if (min > ( (IV) MAX_LEGAL_CP >> 4)) {
22805 s = strchr(s, '\n');
22809 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22810 sv_catpv(msg, overflow_msg);
22811 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22812 UTF8fARG(is_contents_utf8, s - s0, s0));
22813 sv_catpvs(msg, "\"");
22814 goto return_failure;
22817 /* Accumulate this digit into the value */
22818 min = (min << 4) + READ_XDIGIT(s);
22819 } while (isXDIGIT(*s));
22821 while (isBLANK(*s)) { s++; }
22823 /* We allow comments at the end of the line */
22825 s = strchr(s, '\n');
22831 else if (s < e && *s != '\n') {
22832 if (! isXDIGIT(*s)) {
22833 goto check_if_property;
22836 /* Look for the high point of the range */
22839 if (max > ( (IV) MAX_LEGAL_CP >> 4)) {
22840 s = strchr(s, '\n');
22844 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22845 sv_catpv(msg, overflow_msg);
22846 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22847 UTF8fARG(is_contents_utf8, s - s0, s0));
22848 sv_catpvs(msg, "\"");
22849 goto return_failure;
22852 max = (max << 4) + READ_XDIGIT(s);
22853 } while (isXDIGIT(*s));
22855 while (isBLANK(*s)) { s++; }
22858 s = strchr(s, '\n');
22863 else if (s < e && *s != '\n') {
22864 goto check_if_property;
22868 if (max == -1) { /* The line only had one entry */
22871 else if (max < min) {
22872 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22873 sv_catpvs(msg, "Illegal range in \"");
22874 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22875 UTF8fARG(is_contents_utf8, s - s0, s0));
22876 sv_catpvs(msg, "\"");
22877 goto return_failure;
22880 #if 0 /* See explanation at definition above of get_extended_utf8_msg() */
22882 if ( UNICODE_IS_PERL_EXTENDED(min)
22883 || UNICODE_IS_PERL_EXTENDED(max))
22885 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
22887 /* If both code points are non-portable, warn only on the lower
22889 sv_catpv(msg, get_extended_utf8_msg(
22890 (UNICODE_IS_PERL_EXTENDED(min))
22892 sv_catpvs(msg, " in \"");
22893 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f,
22894 UTF8fARG(is_contents_utf8, s - s0, s0));
22895 sv_catpvs(msg, "\"");
22900 /* Here, this line contains a legal range */
22901 this_definition = sv_2mortal(_new_invlist(2));
22902 this_definition = _add_range_to_invlist(this_definition, min, max);
22907 /* Here it isn't a legal range line. See if it is a legal property
22908 * line. First find the end of the meat of the line */
22909 s = strpbrk(s, "#\n");
22914 /* Ignore trailing blanks in keeping with the requirements of
22915 * parse_uniprop_string() */
22917 while (s > s0 && isBLANK_A(*s)) {
22922 this_definition = parse_uniprop_string(s0, s - s0,
22923 is_utf8, to_fold, runtime,
22925 user_defined_ptr, msg,
22927 ? level /* Don't increase level
22928 if input is empty */
22931 if (this_definition == NULL) {
22932 goto return_failure; /* 'msg' should have had the reason
22933 appended to it by the above call */
22936 if (! is_invlist(this_definition)) { /* Unknown at this time */
22937 return newSVsv(this_definition);
22941 s = strchr(s, '\n');
22951 _invlist_union(running_definition, this_definition,
22952 &running_definition);
22955 _invlist_subtract(running_definition, this_definition,
22956 &running_definition);
22959 _invlist_intersection(running_definition, this_definition,
22960 &running_definition);
22963 _invlist_union_complement_2nd(running_definition,
22964 this_definition, &running_definition);
22967 Perl_croak(aTHX_ "panic: %s: %d: Unexpected operation %d",
22968 __FILE__, __LINE__, op);
22972 /* Position past the '\n' */
22974 } /* End of loop through the lines of 'contents' */
22976 /* Here, we processed all the lines in 'contents' without error. If we
22977 * didn't add any warnings, simply return success */
22978 if (msgs_length_on_entry == SvCUR(msg)) {
22980 /* If the expansion was empty, the answer isn't nothing: its an empty
22981 * inversion list */
22982 if (running_definition == NULL) {
22983 running_definition = _new_invlist(1);
22986 return running_definition;
22989 /* Otherwise, add some explanatory text, but we will return success */
22993 running_definition = NULL;
22997 if (name_len > 0) {
22998 sv_catpvs(msg, " in expansion of ");
22999 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
23002 return running_definition;
23005 /* As explained below, certain operations need to take place in the first
23006 * thread created. These macros switch contexts */
23007 #ifdef USE_ITHREADS
23008 # define DECLARATION_FOR_GLOBAL_CONTEXT \
23009 PerlInterpreter * save_aTHX = aTHX;
23010 # define SWITCH_TO_GLOBAL_CONTEXT \
23011 PERL_SET_CONTEXT((aTHX = PL_user_def_props_aTHX))
23012 # define RESTORE_CONTEXT PERL_SET_CONTEXT((aTHX = save_aTHX));
23013 # define CUR_CONTEXT aTHX
23014 # define ORIGINAL_CONTEXT save_aTHX
23016 # define DECLARATION_FOR_GLOBAL_CONTEXT
23017 # define SWITCH_TO_GLOBAL_CONTEXT NOOP
23018 # define RESTORE_CONTEXT NOOP
23019 # define CUR_CONTEXT NULL
23020 # define ORIGINAL_CONTEXT NULL
23024 S_delete_recursion_entry(pTHX_ void *key)
23026 /* Deletes the entry used to detect recursion when expanding user-defined
23027 * properties. This is a function so it can be set up to be called even if
23028 * the program unexpectedly quits */
23031 SV ** current_entry;
23032 const STRLEN key_len = strlen((const char *) key);
23033 DECLARATION_FOR_GLOBAL_CONTEXT;
23035 SWITCH_TO_GLOBAL_CONTEXT;
23037 /* If the entry is one of these types, it is a permanent entry, and not the
23038 * one used to detect recursions. This function should delete only the
23039 * recursion entry */
23040 current_entry = hv_fetch(PL_user_def_props, (const char *) key, key_len, 0);
23042 && ! is_invlist(*current_entry)
23043 && ! SvPOK(*current_entry))
23045 (void) hv_delete(PL_user_def_props, (const char *) key, key_len,
23053 S_get_fq_name(pTHX_
23054 const char * const name, /* The first non-blank in the \p{}, \P{} */
23055 const Size_t name_len, /* Its length in bytes, not including any trailing space */
23056 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23057 const bool has_colon_colon
23060 /* Returns a mortal SV containing the fully qualified version of the input
23065 fq_name = newSVpvs_flags("", SVs_TEMP);
23067 /* Use the current package if it wasn't included in our input */
23068 if (! has_colon_colon) {
23069 const HV * pkg = (IN_PERL_COMPILETIME)
23071 : CopSTASH(PL_curcop);
23072 const char* pkgname = HvNAME(pkg);
23074 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23075 UTF8fARG(is_utf8, strlen(pkgname), pkgname));
23076 sv_catpvs(fq_name, "::");
23079 Perl_sv_catpvf(aTHX_ fq_name, "%" UTF8f,
23080 UTF8fARG(is_utf8, name_len, name));
23085 Perl_parse_uniprop_string(pTHX_
23087 /* Parse the interior of a \p{}, \P{}. Returns its definition if knowable
23088 * now. If so, the return is an inversion list.
23090 * If the property is user-defined, it is a subroutine, which in turn
23091 * may call other subroutines. This function will call the whole nest of
23092 * them to get the definition they return; if some aren't known at the time
23093 * of the call to this function, the fully qualified name of the highest
23094 * level sub is returned. It is an error to call this function at runtime
23095 * without every sub defined.
23097 * If an error was found, NULL is returned, and 'msg' gets a suitable
23098 * message appended to it. (Appending allows the back trace of how we got
23099 * to the faulty definition to be displayed through nested calls of
23100 * user-defined subs.)
23102 * The caller should NOT try to free any returned inversion list.
23104 * Other parameters will be set on return as described below */
23106 const char * const name, /* The first non-blank in the \p{}, \P{} */
23107 const Size_t name_len, /* Its length in bytes, not including any
23109 const bool is_utf8, /* ? Is 'name' encoded in UTF-8 */
23110 const bool to_fold, /* ? Is this under /i */
23111 const bool runtime, /* TRUE if this is being called at run time */
23112 const bool deferrable, /* TRUE if it's ok for the definition to not be
23113 known at this call */
23114 bool *user_defined_ptr, /* Upon return from this function it will be
23115 set to TRUE if any component is a
23116 user-defined property */
23117 SV * msg, /* Any error or warning msg(s) are appended to
23119 const STRLEN level) /* Recursion level of this call */
23122 char* lookup_name; /* normalized name for lookup in our tables */
23123 unsigned lookup_len; /* Its length */
23124 bool stricter = FALSE; /* Some properties have stricter name
23125 normalization rules, which we decide upon
23126 based on parsing */
23128 /* nv= or numeric_value=, or possibly one of the cjk numeric properties
23129 * (though it requires extra effort to download them from Unicode and
23130 * compile perl to know about them) */
23131 bool is_nv_type = FALSE;
23133 unsigned int i, j = 0;
23134 int equals_pos = -1; /* Where the '=' is found, or negative if none */
23135 int slash_pos = -1; /* Where the '/' is found, or negative if none */
23136 int table_index = 0; /* The entry number for this property in the table
23137 of all Unicode property names */
23138 bool starts_with_Is = FALSE; /* ? Does the name start with 'Is' */
23139 Size_t lookup_offset = 0; /* Used to ignore the first few characters of
23140 the normalized name in certain situations */
23141 Size_t non_pkg_begin = 0; /* Offset of first byte in 'name' that isn't
23142 part of a package name */
23143 bool could_be_user_defined = TRUE; /* ? Could this be a user-defined
23144 property rather than a Unicode
23146 SV * prop_definition = NULL; /* The returned definition of 'name' or NULL
23147 if an error. If it is an inversion list,
23148 it is the definition. Otherwise it is a
23149 string containing the fully qualified sub
23151 SV * fq_name = NULL; /* For user-defined properties, the fully
23153 bool invert_return = FALSE; /* ? Do we need to complement the result before
23156 PERL_ARGS_ASSERT_PARSE_UNIPROP_STRING;
23158 /* The input will be normalized into 'lookup_name' */
23159 Newx(lookup_name, name_len, char);
23160 SAVEFREEPV(lookup_name);
23162 /* Parse the input. */
23163 for (i = 0; i < name_len; i++) {
23164 char cur = name[i];
23166 /* Most of the characters in the input will be of this ilk, being parts
23168 if (isIDCONT_A(cur)) {
23170 /* Case differences are ignored. Our lookup routine assumes
23171 * everything is lowercase, so normalize to that */
23172 if (isUPPER_A(cur)) {
23173 lookup_name[j++] = toLOWER_A(cur);
23177 if (cur == '_') { /* Don't include these in the normalized name */
23181 lookup_name[j++] = cur;
23183 /* The first character in a user-defined name must be of this type.
23185 if (i - non_pkg_begin == 0 && ! isIDFIRST_A(cur)) {
23186 could_be_user_defined = FALSE;
23192 /* Here, the character is not something typically in a name, But these
23193 * two types of characters (and the '_' above) can be freely ignored in
23194 * most situations. Later it may turn out we shouldn't have ignored
23195 * them, and we have to reparse, but we don't have enough information
23196 * yet to make that decision */
23197 if (cur == '-' || isSPACE_A(cur)) {
23198 could_be_user_defined = FALSE;
23202 /* An equals sign or single colon mark the end of the first part of
23203 * the property name */
23205 || (cur == ':' && (i >= name_len - 1 || name[i+1] != ':')))
23207 lookup_name[j++] = '='; /* Treat the colon as an '=' */
23208 equals_pos = j; /* Note where it occurred in the input */
23209 could_be_user_defined = FALSE;
23213 /* Otherwise, this character is part of the name. */
23214 lookup_name[j++] = cur;
23216 /* Here it isn't a single colon, so if it is a colon, it must be a
23220 /* A double colon should be a package qualifier. We note its
23221 * position and continue. Note that one could have
23222 * pkg1::pkg2::...::foo
23223 * so that the position at the end of the loop will be just after
23224 * the final qualifier */
23227 non_pkg_begin = i + 1;
23228 lookup_name[j++] = ':';
23230 else { /* Only word chars (and '::') can be in a user-defined name */
23231 could_be_user_defined = FALSE;
23233 } /* End of parsing through the lhs of the property name (or all of it if
23236 #define STRLENs(s) (sizeof("" s "") - 1)
23238 /* If there is a single package name 'utf8::', it is ambiguous. It could
23239 * be for a user-defined property, or it could be a Unicode property, as
23240 * all of them are considered to be for that package. For the purposes of
23241 * parsing the rest of the property, strip it off */
23242 if (non_pkg_begin == STRLENs("utf8::") && memBEGINPs(name, name_len, "utf8::")) {
23243 lookup_name += STRLENs("utf8::");
23244 j -= STRLENs("utf8::");
23245 equals_pos -= STRLENs("utf8::");
23248 /* Here, we are either done with the whole property name, if it was simple;
23249 * or are positioned just after the '=' if it is compound. */
23251 if (equals_pos >= 0) {
23252 assert(! stricter); /* We shouldn't have set this yet */
23254 /* Space immediately after the '=' is ignored */
23256 for (; i < name_len; i++) {
23257 if (! isSPACE_A(name[i])) {
23262 /* Most punctuation after the equals indicates a subpattern, like
23264 if ( isPUNCT_A(name[i])
23270 /* Find the property. The table includes the equals sign, so we
23272 table_index = match_uniprop((U8 *) lookup_name, j);
23274 const char * const * prop_values
23275 = UNI_prop_value_ptrs[table_index];
23277 Size_t subpattern_len;
23278 REGEXP * subpattern_re;
23279 char open = name[i++];
23281 const char * pos_in_brackets;
23284 /* A backslash means the real delimitter is the next character.
23286 if (open == '\\') {
23291 /* This data structure is constructed so that the matching
23292 * closing bracket is 3 past its matching opening. The second
23293 * set of closing is so that if the opening is something like
23294 * ']', the closing will be that as well. Something similar is
23295 * done in toke.c */
23296 pos_in_brackets = strchr("([<)]>)]>", open);
23297 close = (pos_in_brackets) ? pos_in_brackets[3] : open;
23300 || name[name_len-1] != close
23301 || (escaped && name[name_len-2] != '\\'))
23303 sv_catpvs(msg, "Unicode property wildcard not terminated");
23304 goto append_name_to_msg;
23307 Perl_ck_warner_d(aTHX_
23308 packWARN(WARN_EXPERIMENTAL__UNIPROP_WILDCARDS),
23309 "The Unicode property wildcards feature is experimental");
23311 /* Now create and compile the wildcard subpattern. Use /iaa
23312 * because nothing outside of ASCII will match, and it the
23313 * property values should all match /i. Note that when the
23314 * pattern fails to compile, our added text to the user's
23315 * pattern will be displayed to the user, which is not so
23317 subpattern_len = name_len - i - 1 - escaped;
23318 subpattern = Perl_newSVpvf(aTHX_ "(?iaa:%.*s)",
23319 (unsigned) subpattern_len,
23321 subpattern = sv_2mortal(subpattern);
23322 subpattern_re = re_compile(subpattern, 0);
23323 assert(subpattern_re); /* Should have died if didn't compile
23326 /* For each legal property value, see if the supplied pattern
23328 while (*prop_values) {
23329 const char * const entry = *prop_values;
23330 const Size_t len = strlen(entry);
23331 SV* entry_sv = newSVpvn_flags(entry, len, SVs_TEMP);
23333 if (pregexec(subpattern_re,
23335 (char *) entry + len,
23339 { /* Here, matched. Add to the returned list */
23340 Size_t total_len = j + len;
23341 SV * sub_invlist = NULL;
23342 char * this_string;
23344 /* We know this is a legal \p{property=value}. Call
23345 * the function to return the list of code points that
23347 Newxz(this_string, total_len + 1, char);
23348 Copy(lookup_name, this_string, j, char);
23349 my_strlcat(this_string, entry, total_len + 1);
23350 SAVEFREEPV(this_string);
23351 sub_invlist = parse_uniprop_string(this_string,
23360 _invlist_union(prop_definition, sub_invlist,
23364 prop_values++; /* Next iteration, look at next propvalue */
23365 } /* End of looking through property values; (the data
23366 structure is terminated by a NULL ptr) */
23368 SvREFCNT_dec_NN(subpattern_re);
23370 if (prop_definition) {
23371 return prop_definition;
23374 sv_catpvs(msg, "No Unicode property value wildcard matches:");
23375 goto append_name_to_msg;
23378 /* Here's how khw thinks we should proceed to handle the properties
23379 * not yet done: Bidi Mirroring Glyph
23380 Bidi Paired Bracket
23381 Case Folding (both full and simple)
23382 Decomposition Mapping
23383 Equivalent Unified Ideograph
23386 Lowercase Mapping (both full and simple)
23388 Titlecase Mapping (both full and simple)
23389 Uppercase Mapping (both full and simple)
23390 * Move the part that looks at the property values into a perl
23391 * script, like utf8_heavy.pl was done. This makes things somewhat
23392 * easier, but most importantly, it avoids always adding all these
23393 * strings to the memory usage when the feature is little-used.
23395 * The property values would all be concatenated into a single
23396 * string per property with each value on a separate line, and the
23397 * code point it's for on alternating lines. Then we match the
23398 * user's input pattern m//mg, without having to worry about their
23399 * uses of '^' and '$'. Only the values that aren't the default
23400 * would be in the strings. Code points would be in UTF-8. The
23401 * search pattern that we would construct would look like
23402 * (?: \n (code-point_re) \n (?aam: user-re ) \n )
23403 * And so $1 would contain the code point that matched the user-re.
23404 * For properties where the default is the code point itself, such
23405 * as any of the case changing mappings, the string would otherwise
23406 * consist of all Unicode code points in UTF-8 strung together.
23407 * This would be impractical. So instead, examine their compiled
23408 * pattern, looking at the ssc. If none, reject the pattern as an
23409 * error. Otherwise run the pattern against every code point in
23410 * the ssc. The ssc is kind of like tr18's 3.9 Possible Match Sets
23411 * And it might be good to create an API to return the ssc.
23413 * For the name properties, a new function could be created in
23414 * charnames which essentially does the same thing as above,
23415 * sharing Name.pl with the other charname functions. Don't know
23416 * about loose name matching, or algorithmically determined names.
23417 * Decomposition.pl similarly.
23419 * It might be that a new pattern modifier would have to be
23420 * created, like /t for resTricTed, which changed the behavior of
23421 * some constructs in their subpattern, like \A. */
23422 } /* End of is a wildcard subppattern */
23425 /* Certain properties whose values are numeric need special handling.
23426 * They may optionally be prefixed by 'is'. Ignore that prefix for the
23427 * purposes of checking if this is one of those properties */
23428 if (memBEGINPs(lookup_name, j, "is")) {
23432 /* Then check if it is one of these specially-handled properties. The
23433 * possibilities are hard-coded because easier this way, and the list
23434 * is unlikely to change.
23436 * All numeric value type properties are of this ilk, and are also
23437 * special in a different way later on. So find those first. There
23438 * are several numeric value type properties in the Unihan DB (which is
23439 * unlikely to be compiled with perl, but we handle it here in case it
23440 * does get compiled). They all end with 'numeric'. The interiors
23441 * aren't checked for the precise property. This would stop working if
23442 * a cjk property were to be created that ended with 'numeric' and
23443 * wasn't a numeric type */
23444 is_nv_type = memEQs(lookup_name + lookup_offset,
23445 j - 1 - lookup_offset, "numericvalue")
23446 || memEQs(lookup_name + lookup_offset,
23447 j - 1 - lookup_offset, "nv")
23448 || ( memENDPs(lookup_name + lookup_offset,
23449 j - 1 - lookup_offset, "numeric")
23450 && ( memBEGINPs(lookup_name + lookup_offset,
23451 j - 1 - lookup_offset, "cjk")
23452 || memBEGINPs(lookup_name + lookup_offset,
23453 j - 1 - lookup_offset, "k")));
23455 || memEQs(lookup_name + lookup_offset,
23456 j - 1 - lookup_offset, "canonicalcombiningclass")
23457 || memEQs(lookup_name + lookup_offset,
23458 j - 1 - lookup_offset, "ccc")
23459 || memEQs(lookup_name + lookup_offset,
23460 j - 1 - lookup_offset, "age")
23461 || memEQs(lookup_name + lookup_offset,
23462 j - 1 - lookup_offset, "in")
23463 || memEQs(lookup_name + lookup_offset,
23464 j - 1 - lookup_offset, "presentin"))
23468 /* Since the stuff after the '=' is a number, we can't throw away
23469 * '-' willy-nilly, as those could be a minus sign. Other stricter
23470 * rules also apply. However, these properties all can have the
23471 * rhs not be a number, in which case they contain at least one
23472 * alphabetic. In those cases, the stricter rules don't apply.
23473 * But the numeric type properties can have the alphas [Ee] to
23474 * signify an exponent, and it is still a number with stricter
23475 * rules. So look for an alpha that signifies not-strict */
23477 for (k = i; k < name_len; k++) {
23478 if ( isALPHA_A(name[k])
23479 && (! is_nv_type || ! isALPHA_FOLD_EQ(name[k], 'E')))
23489 /* A number may have a leading '+' or '-'. The latter is retained
23491 if (name[i] == '+') {
23494 else if (name[i] == '-') {
23495 lookup_name[j++] = '-';
23499 /* Skip leading zeros including single underscores separating the
23500 * zeros, or between the final leading zero and the first other
23502 for (; i < name_len - 1; i++) {
23503 if ( name[i] != '0'
23504 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23511 else { /* No '=' */
23513 /* Only a few properties without an '=' should be parsed with stricter
23514 * rules. The list is unlikely to change. */
23515 if ( memBEGINPs(lookup_name, j, "perl")
23516 && memNEs(lookup_name + 4, j - 4, "space")
23517 && memNEs(lookup_name + 4, j - 4, "word"))
23521 /* We set the inputs back to 0 and the code below will reparse,
23527 /* Here, we have either finished the property, or are positioned to parse
23528 * the remainder, and we know if stricter rules apply. Finish out, if not
23530 for (; i < name_len; i++) {
23531 char cur = name[i];
23533 /* In all instances, case differences are ignored, and we normalize to
23535 if (isUPPER_A(cur)) {
23536 lookup_name[j++] = toLOWER(cur);
23540 /* An underscore is skipped, but not under strict rules unless it
23541 * separates two digits */
23544 && ( i == 0 || (int) i == equals_pos || i == name_len- 1
23545 || ! isDIGIT_A(name[i-1]) || ! isDIGIT_A(name[i+1])))
23547 lookup_name[j++] = '_';
23552 /* Hyphens are skipped except under strict */
23553 if (cur == '-' && ! stricter) {
23557 /* XXX Bug in documentation. It says white space skipped adjacent to
23558 * non-word char. Maybe we should, but shouldn't skip it next to a dot
23560 if (isSPACE_A(cur) && ! stricter) {
23564 lookup_name[j++] = cur;
23566 /* Unless this is a non-trailing slash, we are done with it */
23567 if (i >= name_len - 1 || cur != '/') {
23573 /* A slash in the 'numeric value' property indicates that what follows
23574 * is a denominator. It can have a leading '+' and '0's that should be
23575 * skipped. But we have never allowed a negative denominator, so treat
23576 * a minus like every other character. (No need to rule out a second
23577 * '/', as that won't match anything anyway */
23580 if (i < name_len && name[i] == '+') {
23584 /* Skip leading zeros including underscores separating digits */
23585 for (; i < name_len - 1; i++) {
23586 if ( name[i] != '0'
23587 && (name[i] != '_' || ! isDIGIT_A(name[i+1])))
23593 /* Store the first real character in the denominator */
23594 if (i < name_len) {
23595 lookup_name[j++] = name[i];
23600 /* Here are completely done parsing the input 'name', and 'lookup_name'
23601 * contains a copy, normalized.
23603 * This special case is grandfathered in: 'L_' and 'GC=L_' are accepted and
23604 * different from without the underscores. */
23605 if ( ( UNLIKELY(memEQs(lookup_name, j, "l"))
23606 || UNLIKELY(memEQs(lookup_name, j, "gc=l")))
23607 && UNLIKELY(name[name_len-1] == '_'))
23609 lookup_name[j++] = '&';
23612 /* If the original input began with 'In' or 'Is', it could be a subroutine
23613 * call to a user-defined property instead of a Unicode property name. */
23614 if ( name_len - non_pkg_begin > 2
23615 && name[non_pkg_begin+0] == 'I'
23616 && (name[non_pkg_begin+1] == 'n' || name[non_pkg_begin+1] == 's'))
23618 /* Names that start with In have different characterstics than those
23619 * that start with Is */
23620 if (name[non_pkg_begin+1] == 's') {
23621 starts_with_Is = TRUE;
23625 could_be_user_defined = FALSE;
23628 if (could_be_user_defined) {
23631 /* If the user defined property returns the empty string, it could
23632 * easily be because the pattern is being compiled before the data it
23633 * actually needs to compile is available. This could be argued to be
23634 * a bug in the perl code, but this is a change of behavior for Perl,
23635 * so we handle it. This means that intentionally returning nothing
23636 * will not be resolved until runtime */
23637 bool empty_return = FALSE;
23639 /* Here, the name could be for a user defined property, which are
23640 * implemented as subs. */
23641 user_sub = get_cvn_flags(name, name_len, 0);
23643 const char insecure[] = "Insecure user-defined property";
23645 /* Here, there is a sub by the correct name. Normally we call it
23646 * to get the property definition */
23648 SV * user_sub_sv = MUTABLE_SV(user_sub);
23649 SV * error; /* Any error returned by calling 'user_sub' */
23650 SV * key; /* The key into the hash of user defined sub names
23653 SV ** saved_user_prop_ptr; /* Hash entry for this property */
23655 /* How many times to retry when another thread is in the middle of
23656 * expanding the same definition we want */
23657 PERL_INT_FAST8_T retry_countdown = 10;
23659 DECLARATION_FOR_GLOBAL_CONTEXT;
23661 /* If we get here, we know this property is user-defined */
23662 *user_defined_ptr = TRUE;
23664 /* We refuse to call a potentially tainted subroutine; returning an
23667 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23668 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23669 goto append_name_to_msg;
23672 /* In principal, we only call each subroutine property definition
23673 * once during the life of the program. This guarantees that the
23674 * property definition never changes. The results of the single
23675 * sub call are stored in a hash, which is used instead for future
23676 * references to this property. The property definition is thus
23677 * immutable. But, to allow the user to have a /i-dependent
23678 * definition, we call the sub once for non-/i, and once for /i,
23679 * should the need arise, passing the /i status as a parameter.
23681 * We start by constructing the hash key name, consisting of the
23682 * fully qualified subroutine name, preceded by the /i status, so
23683 * that there is a key for /i and a different key for non-/i */
23684 key = newSVpvn(((to_fold) ? "1" : "0"), 1);
23685 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
23686 non_pkg_begin != 0);
23687 sv_catsv(key, fq_name);
23690 /* We only call the sub once throughout the life of the program
23691 * (with the /i, non-/i exception noted above). That means the
23692 * hash must be global and accessible to all threads. It is
23693 * created at program start-up, before any threads are created, so
23694 * is accessible to all children. But this creates some
23697 * 1) The keys can't be shared, or else problems arise; sharing is
23698 * turned off at hash creation time
23699 * 2) All SVs in it are there for the remainder of the life of the
23700 * program, and must be created in the same interpreter context
23701 * as the hash, or else they will be freed from the wrong pool
23702 * at global destruction time. This is handled by switching to
23703 * the hash's context to create each SV going into it, and then
23704 * immediately switching back
23705 * 3) All accesses to the hash must be controlled by a mutex, to
23706 * prevent two threads from getting an unstable state should
23707 * they simultaneously be accessing it. The code below is
23708 * crafted so that the mutex is locked whenever there is an
23709 * access and unlocked only when the next stable state is
23712 * The hash stores either the definition of the property if it was
23713 * valid, or, if invalid, the error message that was raised. We
23714 * use the type of SV to distinguish.
23716 * There's also the need to guard against the definition expansion
23717 * from infinitely recursing. This is handled by storing the aTHX
23718 * of the expanding thread during the expansion. Again the SV type
23719 * is used to distinguish this from the other two cases. If we
23720 * come to here and the hash entry for this property is our aTHX,
23721 * it means we have recursed, and the code assumes that we would
23722 * infinitely recurse, so instead stops and raises an error.
23723 * (Any recursion has always been treated as infinite recursion in
23726 * If instead, the entry is for a different aTHX, it means that
23727 * that thread has gotten here first, and hasn't finished expanding
23728 * the definition yet. We just have to wait until it is done. We
23729 * sleep and retry a few times, returning an error if the other
23730 * thread doesn't complete. */
23733 USER_PROP_MUTEX_LOCK;
23735 /* If we have an entry for this key, the subroutine has already
23736 * been called once with this /i status. */
23737 saved_user_prop_ptr = hv_fetch(PL_user_def_props,
23738 SvPVX(key), SvCUR(key), 0);
23739 if (saved_user_prop_ptr) {
23741 /* If the saved result is an inversion list, it is the valid
23742 * definition of this property */
23743 if (is_invlist(*saved_user_prop_ptr)) {
23744 prop_definition = *saved_user_prop_ptr;
23746 /* The SV in the hash won't be removed until global
23747 * destruction, so it is stable and we can unlock */
23748 USER_PROP_MUTEX_UNLOCK;
23750 /* The caller shouldn't try to free this SV */
23751 return prop_definition;
23754 /* Otherwise, if it is a string, it is the error message
23755 * that was returned when we first tried to evaluate this
23756 * property. Fail, and append the message */
23757 if (SvPOK(*saved_user_prop_ptr)) {
23758 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23759 sv_catsv(msg, *saved_user_prop_ptr);
23761 /* The SV in the hash won't be removed until global
23762 * destruction, so it is stable and we can unlock */
23763 USER_PROP_MUTEX_UNLOCK;
23768 assert(SvIOK(*saved_user_prop_ptr));
23770 /* Here, we have an unstable entry in the hash. Either another
23771 * thread is in the middle of expanding the property's
23772 * definition, or we are ourselves recursing. We use the aTHX
23773 * in it to distinguish */
23774 if (SvIV(*saved_user_prop_ptr) != PTR2IV(CUR_CONTEXT)) {
23776 /* Here, it's another thread doing the expanding. We've
23777 * looked as much as we are going to at the contents of the
23778 * hash entry. It's safe to unlock. */
23779 USER_PROP_MUTEX_UNLOCK;
23781 /* Retry a few times */
23782 if (retry_countdown-- > 0) {
23787 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23788 sv_catpvs(msg, "Timeout waiting for another thread to "
23790 goto append_name_to_msg;
23793 /* Here, we are recursing; don't dig any deeper */
23794 USER_PROP_MUTEX_UNLOCK;
23796 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23798 "Infinite recursion in user-defined property");
23799 goto append_name_to_msg;
23802 /* Here, this thread has exclusive control, and there is no entry
23803 * for this property in the hash. So we have the go ahead to
23804 * expand the definition ourselves. */
23806 PUSHSTACKi(PERLSI_MAGIC);
23809 /* Create a temporary placeholder in the hash to detect recursion
23811 SWITCH_TO_GLOBAL_CONTEXT;
23812 placeholder= newSVuv(PTR2IV(ORIGINAL_CONTEXT));
23813 (void) hv_store_ent(PL_user_def_props, key, placeholder, 0);
23816 /* Now that we have a placeholder, we can let other threads
23818 USER_PROP_MUTEX_UNLOCK;
23820 /* Make sure the placeholder always gets destroyed */
23821 SAVEDESTRUCTOR_X(S_delete_recursion_entry, SvPVX(key));
23826 /* Call the user's function, with the /i status as a parameter.
23827 * Note that we have gone to a lot of trouble to keep this call
23828 * from being within the locked mutex region. */
23829 XPUSHs(boolSV(to_fold));
23832 /* The following block was taken from swash_init(). Presumably
23833 * they apply to here as well, though we no longer use a swash --
23837 /* We might get here via a subroutine signature which uses a utf8
23838 * parameter name, at which point PL_subname will have been set
23839 * but not yet used. */
23840 save_item(PL_subname);
23842 (void) call_sv(user_sub_sv, G_EVAL|G_SCALAR);
23847 if (TAINT_get || SvTRUE(error)) {
23848 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23849 if (SvTRUE(error)) {
23850 sv_catpvs(msg, "Error \"");
23851 sv_catsv(msg, error);
23852 sv_catpvs(msg, "\"");
23855 if (SvTRUE(error)) sv_catpvs(msg, "; ");
23856 sv_catpvn(msg, insecure, sizeof(insecure) - 1);
23859 if (name_len > 0) {
23860 sv_catpvs(msg, " in expansion of ");
23861 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8,
23867 prop_definition = NULL;
23869 else { /* G_SCALAR guarantees a single return value */
23870 SV * contents = POPs;
23872 /* The contents is supposed to be the expansion of the property
23873 * definition. If the definition is deferrable, and we got an
23874 * empty string back, set a flag to later defer it (after clean
23877 && (! SvPOK(contents) || SvCUR(contents) == 0))
23879 empty_return = TRUE;
23881 else { /* Otherwise, call a function to check for valid syntax,
23884 prop_definition = handle_user_defined_property(
23886 is_utf8, to_fold, runtime,
23888 contents, user_defined_ptr,
23894 /* Here, we have the results of the expansion. Delete the
23895 * placeholder, and if the definition is now known, replace it with
23896 * that definition. We need exclusive access to the hash, and we
23897 * can't let anyone else in, between when we delete the placeholder
23898 * and add the permanent entry */
23899 USER_PROP_MUTEX_LOCK;
23901 S_delete_recursion_entry(aTHX_ SvPVX(key));
23903 if ( ! empty_return
23904 && (! prop_definition || is_invlist(prop_definition)))
23906 /* If we got success we use the inversion list defining the
23907 * property; otherwise use the error message */
23908 SWITCH_TO_GLOBAL_CONTEXT;
23909 (void) hv_store_ent(PL_user_def_props,
23912 ? newSVsv(prop_definition)
23918 /* All done, and the hash now has a permanent entry for this
23919 * property. Give up exclusive control */
23920 USER_PROP_MUTEX_UNLOCK;
23926 if (empty_return) {
23927 goto definition_deferred;
23930 if (prop_definition) {
23932 /* If the definition is for something not known at this time,
23933 * we toss it, and go return the main property name, as that's
23934 * the one the user will be aware of */
23935 if (! is_invlist(prop_definition)) {
23936 SvREFCNT_dec_NN(prop_definition);
23937 goto definition_deferred;
23940 sv_2mortal(prop_definition);
23944 return prop_definition;
23946 } /* End of calling the subroutine for the user-defined property */
23947 } /* End of it could be a user-defined property */
23949 /* Here it wasn't a user-defined property that is known at this time. See
23950 * if it is a Unicode property */
23952 lookup_len = j; /* This is a more mnemonic name than 'j' */
23954 /* Get the index into our pointer table of the inversion list corresponding
23955 * to the property */
23956 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23958 /* If it didn't find the property ... */
23959 if (table_index == 0) {
23961 /* Try again stripping off any initial 'Is'. This is because we
23962 * promise that an initial Is is optional. The same isn't true of
23963 * names that start with 'In'. Those can match only blocks, and the
23964 * lookup table already has those accounted for. */
23965 if (starts_with_Is) {
23971 table_index = match_uniprop((U8 *) lookup_name, lookup_len);
23974 if (table_index == 0) {
23977 /* Here, we didn't find it. If not a numeric type property, and
23978 * can't be a user-defined one, it isn't a legal property */
23979 if (! is_nv_type) {
23980 if (! could_be_user_defined) {
23984 /* Here, the property name is legal as a user-defined one. At
23985 * compile time, it might just be that the subroutine for that
23986 * property hasn't been encountered yet, but at runtime, it's
23987 * an error to try to use an undefined one */
23988 if (! deferrable) {
23989 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
23990 sv_catpvs(msg, "Unknown user-defined property name");
23991 goto append_name_to_msg;
23994 goto definition_deferred;
23995 } /* End of isn't a numeric type property */
23997 /* The numeric type properties need more work to decide. What we
23998 * do is make sure we have the number in canonical form and look
24001 if (slash_pos < 0) { /* No slash */
24003 /* When it isn't a rational, take the input, convert it to a
24004 * NV, then create a canonical string representation of that
24008 SSize_t value_len = lookup_len - equals_pos;
24010 /* Get the value */
24011 if ( value_len <= 0
24012 || my_atof3(lookup_name + equals_pos, &value,
24014 != lookup_name + lookup_len)
24019 /* If the value is an integer, the canonical value is integral
24021 if (Perl_ceil(value) == value) {
24022 canonical = Perl_form(aTHX_ "%.*s%.0" NVff,
24023 equals_pos, lookup_name, value);
24025 else { /* Otherwise, it is %e with a known precision */
24028 canonical = Perl_form(aTHX_ "%.*s%.*" NVef,
24029 equals_pos, lookup_name,
24030 PL_E_FORMAT_PRECISION, value);
24032 /* The exponent generated is expecting two digits, whereas
24033 * %e on some systems will generate three. Remove leading
24034 * zeros in excess of 2 from the exponent. We start
24035 * looking for them after the '=' */
24036 exp_ptr = strchr(canonical + equals_pos, 'e');
24038 char * cur_ptr = exp_ptr + 2; /* past the 'e[+-]' */
24039 SSize_t excess_exponent_len = strlen(cur_ptr) - 2;
24041 assert(*(cur_ptr - 1) == '-' || *(cur_ptr - 1) == '+');
24043 if (excess_exponent_len > 0) {
24044 SSize_t leading_zeros = strspn(cur_ptr, "0");
24045 SSize_t excess_leading_zeros
24046 = MIN(leading_zeros, excess_exponent_len);
24047 if (excess_leading_zeros > 0) {
24048 Move(cur_ptr + excess_leading_zeros,
24050 strlen(cur_ptr) - excess_leading_zeros
24051 + 1, /* Copy the NUL as well */
24058 else { /* Has a slash. Create a rational in canonical form */
24059 UV numerator, denominator, gcd, trial;
24060 const char * end_ptr;
24061 const char * sign = "";
24063 /* We can't just find the numerator, denominator, and do the
24064 * division, then use the method above, because that is
24065 * inexact. And the input could be a rational that is within
24066 * epsilon (given our precision) of a valid rational, and would
24067 * then incorrectly compare valid.
24069 * We're only interested in the part after the '=' */
24070 const char * this_lookup_name = lookup_name + equals_pos;
24071 lookup_len -= equals_pos;
24072 slash_pos -= equals_pos;
24074 /* Handle any leading minus */
24075 if (this_lookup_name[0] == '-') {
24077 this_lookup_name++;
24082 /* Convert the numerator to numeric */
24083 end_ptr = this_lookup_name + slash_pos;
24084 if (! grok_atoUV(this_lookup_name, &numerator, &end_ptr)) {
24088 /* It better have included all characters before the slash */
24089 if (*end_ptr != '/') {
24093 /* Set to look at just the denominator */
24094 this_lookup_name += slash_pos;
24095 lookup_len -= slash_pos;
24096 end_ptr = this_lookup_name + lookup_len;
24098 /* Convert the denominator to numeric */
24099 if (! grok_atoUV(this_lookup_name, &denominator, &end_ptr)) {
24103 /* It better be the rest of the characters, and don't divide by
24105 if ( end_ptr != this_lookup_name + lookup_len
24106 || denominator == 0)
24111 /* Get the greatest common denominator using
24112 http://en.wikipedia.org/wiki/Euclidean_algorithm */
24114 trial = denominator;
24115 while (trial != 0) {
24117 trial = gcd % trial;
24121 /* If already in lowest possible terms, we have already tried
24122 * looking this up */
24127 /* Reduce the rational, which should put it in canonical form
24130 denominator /= gcd;
24132 canonical = Perl_form(aTHX_ "%.*s%s%" UVuf "/%" UVuf,
24133 equals_pos, lookup_name, sign, numerator, denominator);
24136 /* Here, we have the number in canonical form. Try that */
24137 table_index = match_uniprop((U8 *) canonical, strlen(canonical));
24138 if (table_index == 0) {
24141 } /* End of still didn't find the property in our table */
24142 } /* End of didn't find the property in our table */
24144 /* Here, we have a non-zero return, which is an index into a table of ptrs.
24145 * A negative return signifies that the real index is the absolute value,
24146 * but the result needs to be inverted */
24147 if (table_index < 0) {
24148 invert_return = TRUE;
24149 table_index = -table_index;
24152 /* Out-of band indices indicate a deprecated property. The proper index is
24153 * modulo it with the table size. And dividing by the table size yields
24154 * an offset into a table constructed by regen/mk_invlists.pl to contain
24155 * the corresponding warning message */
24156 if (table_index > MAX_UNI_KEYWORD_INDEX) {
24157 Size_t warning_offset = table_index / MAX_UNI_KEYWORD_INDEX;
24158 table_index %= MAX_UNI_KEYWORD_INDEX;
24159 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED),
24160 "Use of '%.*s' in \\p{} or \\P{} is deprecated because: %s",
24161 (int) name_len, name, deprecated_property_msgs[warning_offset]);
24164 /* In a few properties, a different property is used under /i. These are
24165 * unlikely to change, so are hard-coded here. */
24167 if ( table_index == UNI_XPOSIXUPPER
24168 || table_index == UNI_XPOSIXLOWER
24169 || table_index == UNI_TITLE)
24171 table_index = UNI_CASED;
24173 else if ( table_index == UNI_UPPERCASELETTER
24174 || table_index == UNI_LOWERCASELETTER
24175 # ifdef UNI_TITLECASELETTER /* Missing from early Unicodes */
24176 || table_index == UNI_TITLECASELETTER
24179 table_index = UNI_CASEDLETTER;
24181 else if ( table_index == UNI_POSIXUPPER
24182 || table_index == UNI_POSIXLOWER)
24184 table_index = UNI_POSIXALPHA;
24188 /* Create and return the inversion list */
24189 prop_definition =_new_invlist_C_array(uni_prop_ptrs[table_index]);
24190 sv_2mortal(prop_definition);
24193 /* See if there is a private use override to add to this definition */
24195 COPHH * hinthash = (IN_PERL_COMPILETIME)
24196 ? CopHINTHASH_get(&PL_compiling)
24197 : CopHINTHASH_get(PL_curcop);
24198 SV * pu_overrides = cophh_fetch_pv(hinthash, "private_use", 0, 0);
24200 if (UNLIKELY(pu_overrides && SvPOK(pu_overrides))) {
24202 /* See if there is an element in the hints hash for this table */
24203 SV * pu_lookup = Perl_newSVpvf(aTHX_ "%d=", table_index);
24204 const char * pos = strstr(SvPVX(pu_overrides), SvPVX(pu_lookup));
24208 SV * pu_definition;
24210 SV * expanded_prop_definition =
24211 sv_2mortal(invlist_clone(prop_definition, NULL));
24213 /* If so, it's definition is the string from here to the next
24214 * \a character. And its format is the same as a user-defined
24216 pos += SvCUR(pu_lookup);
24217 pu_definition = newSVpvn(pos, strchr(pos, '\a') - pos);
24218 pu_invlist = handle_user_defined_property(lookup_name,
24221 0, /* Not folded */
24229 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24230 sv_catpvs(msg, "Insecure private-use override");
24231 goto append_name_to_msg;
24234 /* For now, as a safety measure, make sure that it doesn't
24235 * override non-private use code points */
24236 _invlist_intersection(pu_invlist, PL_Private_Use, &pu_invlist);
24238 /* Add it to the list to be returned */
24239 _invlist_union(prop_definition, pu_invlist,
24240 &expanded_prop_definition);
24241 prop_definition = expanded_prop_definition;
24242 Perl_ck_warner_d(aTHX_ packWARN(WARN_EXPERIMENTAL__PRIVATE_USE), "The private_use feature is experimental");
24247 if (invert_return) {
24248 _invlist_invert(prop_definition);
24250 return prop_definition;
24254 if (non_pkg_begin != 0) {
24255 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24256 sv_catpvs(msg, "Illegal user-defined property name");
24259 if (SvCUR(msg) > 0) sv_catpvs(msg, "; ");
24260 sv_catpvs(msg, "Can't find Unicode property definition");
24264 append_name_to_msg:
24266 const char * prefix = (runtime && level == 0) ? " \\p{" : " \"";
24267 const char * suffix = (runtime && level == 0) ? "}" : "\"";
24269 sv_catpv(msg, prefix);
24270 Perl_sv_catpvf(aTHX_ msg, "%" UTF8f, UTF8fARG(is_utf8, name_len, name));
24271 sv_catpv(msg, suffix);
24276 definition_deferred:
24278 /* Here it could yet to be defined, so defer evaluation of this
24279 * until its needed at runtime. We need the fully qualified property name
24280 * to avoid ambiguity, and a trailing newline */
24282 fq_name = S_get_fq_name(aTHX_ name, name_len, is_utf8,
24283 non_pkg_begin != 0 /* If has "::" */
24286 sv_catpvs(fq_name, "\n");
24288 *user_defined_ptr = TRUE;
24295 * ex: set ts=8 sts=4 sw=4 et: