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
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_inline.h"
90 #include "invlist_inline.h"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
105 #define MIN(a,b) ((a) < (b) ? (a) : (b))
108 /* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
112 typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
124 /* Certain characters are output as a sequence with the first being a
126 #define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
130 struct RExC_state_t {
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
133 char *precomp; /* uncompiled string. */
134 REGEXP *rx_sv; /* The SV that is the regexp. */
135 regexp *rx; /* perl core regexp structure */
136 regexp_internal *rxi; /* internal data for regexp object
138 char *start; /* Start of input for compile */
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
141 SSize_t whilem_seen; /* number of WHILEM in this expr */
142 regnode *emit_start; /* Start of emitted-code area */
143 regnode *emit_bound; /* First regnode outside of the
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
154 SSize_t size; /* Code size. */
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
158 I32 nestroot; /* root parens we are in - used by
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
172 HV *paren_names; /* Paren names */
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
182 I32 override_recoding;
184 I32 recode_x_to_native;
186 I32 in_multi_char_class;
187 struct reg_code_block *code_blocks; /* positions of literal (?{})
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
191 SSize_t maxlen; /* mininum possible number of chars in string to match */
192 scan_frame *frame_head;
193 scan_frame *frame_last;
196 #ifdef ADD_TO_REGEXEC
197 char *starttry; /* -Dr: where regtry was called. */
198 #define RExC_starttry (pRExC_state->starttry)
200 SV *runtime_code_qr; /* qr with the runtime code blocks */
202 const char *lastparse;
204 AV *paren_name_list; /* idx -> name */
205 U32 study_chunk_recursed_count;
208 #define RExC_lastparse (pRExC_state->lastparse)
209 #define RExC_lastnum (pRExC_state->lastnum)
210 #define RExC_paren_name_list (pRExC_state->paren_name_list)
211 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
212 #define RExC_mysv (pRExC_state->mysv1)
213 #define RExC_mysv1 (pRExC_state->mysv1)
214 #define RExC_mysv2 (pRExC_state->mysv2)
217 bool seen_unfolded_sharp_s;
220 #define RExC_flags (pRExC_state->flags)
221 #define RExC_pm_flags (pRExC_state->pm_flags)
222 #define RExC_precomp (pRExC_state->precomp)
223 #define RExC_rx_sv (pRExC_state->rx_sv)
224 #define RExC_rx (pRExC_state->rx)
225 #define RExC_rxi (pRExC_state->rxi)
226 #define RExC_start (pRExC_state->start)
227 #define RExC_end (pRExC_state->end)
228 #define RExC_parse (pRExC_state->parse)
229 #define RExC_whilem_seen (pRExC_state->whilem_seen)
231 /* Set during the sizing pass when there is a LATIN SMALL LETTER SHARP S in any
232 * EXACTF node, hence was parsed under /di rules. If later in the parse,
233 * something forces the pattern into using /ui rules, the sharp s should be
234 * folded into the sequence 'ss', which takes up more space than previously
235 * calculated. This means that the sizing pass needs to be restarted. (The
236 * node also becomes an EXACTFU_SS.) For all other characters, an EXACTF node
237 * that gets converted to /ui (and EXACTFU) occupies the same amount of space,
238 * so there is no need to resize [perl #125990]. */
239 #define RExC_seen_unfolded_sharp_s (pRExC_state->seen_unfolded_sharp_s)
241 #ifdef RE_TRACK_PATTERN_OFFSETS
242 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
245 #define RExC_emit (pRExC_state->emit)
246 #define RExC_emit_dummy (pRExC_state->emit_dummy)
247 #define RExC_emit_start (pRExC_state->emit_start)
248 #define RExC_emit_bound (pRExC_state->emit_bound)
249 #define RExC_sawback (pRExC_state->sawback)
250 #define RExC_seen (pRExC_state->seen)
251 #define RExC_size (pRExC_state->size)
252 #define RExC_maxlen (pRExC_state->maxlen)
253 #define RExC_npar (pRExC_state->npar)
254 #define RExC_nestroot (pRExC_state->nestroot)
255 #define RExC_extralen (pRExC_state->extralen)
256 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
257 #define RExC_utf8 (pRExC_state->utf8)
258 #define RExC_uni_semantics (pRExC_state->uni_semantics)
259 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
260 #define RExC_open_parens (pRExC_state->open_parens)
261 #define RExC_close_parens (pRExC_state->close_parens)
262 #define RExC_opend (pRExC_state->opend)
263 #define RExC_paren_names (pRExC_state->paren_names)
264 #define RExC_recurse (pRExC_state->recurse)
265 #define RExC_recurse_count (pRExC_state->recurse_count)
266 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
267 #define RExC_study_chunk_recursed_bytes \
268 (pRExC_state->study_chunk_recursed_bytes)
269 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
270 #define RExC_contains_locale (pRExC_state->contains_locale)
271 #define RExC_contains_i (pRExC_state->contains_i)
272 #define RExC_override_recoding (pRExC_state->override_recoding)
274 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
276 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
277 #define RExC_frame_head (pRExC_state->frame_head)
278 #define RExC_frame_last (pRExC_state->frame_last)
279 #define RExC_frame_count (pRExC_state->frame_count)
280 #define RExC_strict (pRExC_state->strict)
282 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
283 * a flag to disable back-off on the fixed/floating substrings - if it's
284 * a high complexity pattern we assume the benefit of avoiding a full match
285 * is worth the cost of checking for the substrings even if they rarely help.
287 #define RExC_naughty (pRExC_state->naughty)
288 #define TOO_NAUGHTY (10)
289 #define MARK_NAUGHTY(add) \
290 if (RExC_naughty < TOO_NAUGHTY) \
291 RExC_naughty += (add)
292 #define MARK_NAUGHTY_EXP(exp, add) \
293 if (RExC_naughty < TOO_NAUGHTY) \
294 RExC_naughty += RExC_naughty / (exp) + (add)
296 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
297 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
298 ((*s) == '{' && regcurly(s)))
301 * Flags to be passed up and down.
303 #define WORST 0 /* Worst case. */
304 #define HASWIDTH 0x01 /* Known to match non-null strings. */
306 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
307 * character. (There needs to be a case: in the switch statement in regexec.c
308 * for any node marked SIMPLE.) Note that this is not the same thing as
311 #define SPSTART 0x04 /* Starts with * or + */
312 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
313 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
314 #define RESTART_PASS1 0x20 /* Need to restart sizing pass */
315 #define NEED_UTF8 0x40 /* In conjunction with RESTART_PASS1, need to
316 calcuate sizes as UTF-8 */
318 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
320 /* whether trie related optimizations are enabled */
321 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
322 #define TRIE_STUDY_OPT
323 #define FULL_TRIE_STUDY
329 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
330 #define PBITVAL(paren) (1 << ((paren) & 7))
331 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
332 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
333 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
335 #define REQUIRE_UTF8(flagp) STMT_START { \
338 *flagp = RESTART_PASS1|NEED_UTF8; \
343 /* Change from /d into /u rules, and restart the parse if we've already seen
344 * something whose size would increase as a result, by setting *flagp and
345 * returning 'restart_retval'. RExC_uni_semantics is a flag that indicates
346 * we've change to /u during the parse. */
347 #define REQUIRE_UNI_RULES(flagp, restart_retval) \
349 if (DEPENDS_SEMANTICS) { \
351 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET); \
352 RExC_uni_semantics = 1; \
353 if (RExC_seen_unfolded_sharp_s) { \
354 *flagp |= RESTART_PASS1; \
355 return restart_retval; \
360 /* This converts the named class defined in regcomp.h to its equivalent class
361 * number defined in handy.h. */
362 #define namedclass_to_classnum(class) ((int) ((class) / 2))
363 #define classnum_to_namedclass(classnum) ((classnum) * 2)
365 #define _invlist_union_complement_2nd(a, b, output) \
366 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
367 #define _invlist_intersection_complement_2nd(a, b, output) \
368 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
370 /* About scan_data_t.
372 During optimisation we recurse through the regexp program performing
373 various inplace (keyhole style) optimisations. In addition study_chunk
374 and scan_commit populate this data structure with information about
375 what strings MUST appear in the pattern. We look for the longest
376 string that must appear at a fixed location, and we look for the
377 longest string that may appear at a floating location. So for instance
382 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
383 strings (because they follow a .* construct). study_chunk will identify
384 both FOO and BAR as being the longest fixed and floating strings respectively.
386 The strings can be composites, for instance
390 will result in a composite fixed substring 'foo'.
392 For each string some basic information is maintained:
394 - offset or min_offset
395 This is the position the string must appear at, or not before.
396 It also implicitly (when combined with minlenp) tells us how many
397 characters must match before the string we are searching for.
398 Likewise when combined with minlenp and the length of the string it
399 tells us how many characters must appear after the string we have
403 Only used for floating strings. This is the rightmost point that
404 the string can appear at. If set to SSize_t_MAX it indicates that the
405 string can occur infinitely far to the right.
408 A pointer to the minimum number of characters of the pattern that the
409 string was found inside. This is important as in the case of positive
410 lookahead or positive lookbehind we can have multiple patterns
415 The minimum length of the pattern overall is 3, the minimum length
416 of the lookahead part is 3, but the minimum length of the part that
417 will actually match is 1. So 'FOO's minimum length is 3, but the
418 minimum length for the F is 1. This is important as the minimum length
419 is used to determine offsets in front of and behind the string being
420 looked for. Since strings can be composites this is the length of the
421 pattern at the time it was committed with a scan_commit. Note that
422 the length is calculated by study_chunk, so that the minimum lengths
423 are not known until the full pattern has been compiled, thus the
424 pointer to the value.
428 In the case of lookbehind the string being searched for can be
429 offset past the start point of the final matching string.
430 If this value was just blithely removed from the min_offset it would
431 invalidate some of the calculations for how many chars must match
432 before or after (as they are derived from min_offset and minlen and
433 the length of the string being searched for).
434 When the final pattern is compiled and the data is moved from the
435 scan_data_t structure into the regexp structure the information
436 about lookbehind is factored in, with the information that would
437 have been lost precalculated in the end_shift field for the
440 The fields pos_min and pos_delta are used to store the minimum offset
441 and the delta to the maximum offset at the current point in the pattern.
445 typedef struct scan_data_t {
446 /*I32 len_min; unused */
447 /*I32 len_delta; unused */
451 SSize_t last_end; /* min value, <0 unless valid. */
452 SSize_t last_start_min;
453 SSize_t last_start_max;
454 SV **longest; /* Either &l_fixed, or &l_float. */
455 SV *longest_fixed; /* longest fixed string found in pattern */
456 SSize_t offset_fixed; /* offset where it starts */
457 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
458 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
459 SV *longest_float; /* longest floating string found in pattern */
460 SSize_t offset_float_min; /* earliest point in string it can appear */
461 SSize_t offset_float_max; /* latest point in string it can appear */
462 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
463 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
466 SSize_t *last_closep;
467 regnode_ssc *start_class;
471 * Forward declarations for pregcomp()'s friends.
474 static const scan_data_t zero_scan_data =
475 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
477 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
478 #define SF_BEFORE_SEOL 0x0001
479 #define SF_BEFORE_MEOL 0x0002
480 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
481 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
483 #define SF_FIX_SHIFT_EOL (+2)
484 #define SF_FL_SHIFT_EOL (+4)
486 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
487 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
489 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
490 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
491 #define SF_IS_INF 0x0040
492 #define SF_HAS_PAR 0x0080
493 #define SF_IN_PAR 0x0100
494 #define SF_HAS_EVAL 0x0200
495 #define SCF_DO_SUBSTR 0x0400
496 #define SCF_DO_STCLASS_AND 0x0800
497 #define SCF_DO_STCLASS_OR 0x1000
498 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
499 #define SCF_WHILEM_VISITED_POS 0x2000
501 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
502 #define SCF_SEEN_ACCEPT 0x8000
503 #define SCF_TRIE_DOING_RESTUDY 0x10000
504 #define SCF_IN_DEFINE 0x20000
509 #define UTF cBOOL(RExC_utf8)
511 /* The enums for all these are ordered so things work out correctly */
512 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
513 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
514 == REGEX_DEPENDS_CHARSET)
515 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
516 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
517 >= REGEX_UNICODE_CHARSET)
518 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
519 == REGEX_ASCII_RESTRICTED_CHARSET)
520 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
521 >= REGEX_ASCII_RESTRICTED_CHARSET)
522 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
523 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
525 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
527 /* For programs that want to be strictly Unicode compatible by dying if any
528 * attempt is made to match a non-Unicode code point against a Unicode
530 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
532 #define OOB_NAMEDCLASS -1
534 /* There is no code point that is out-of-bounds, so this is problematic. But
535 * its only current use is to initialize a variable that is always set before
537 #define OOB_UNICODE 0xDEADBEEF
539 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
540 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
543 /* length of regex to show in messages that don't mark a position within */
544 #define RegexLengthToShowInErrorMessages 127
547 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
548 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
549 * op/pragma/warn/regcomp.
551 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
552 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
554 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
555 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
557 #define REPORT_LOCATION_ARGS(offset) \
558 UTF8fARG(UTF, offset, RExC_precomp), \
559 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
561 /* Used to point after bad bytes for an error message, but avoid skipping
562 * past a nul byte. */
563 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
566 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
567 * arg. Show regex, up to a maximum length. If it's too long, chop and add
570 #define _FAIL(code) STMT_START { \
571 const char *ellipses = ""; \
572 IV len = RExC_end - RExC_precomp; \
575 SAVEFREESV(RExC_rx_sv); \
576 if (len > RegexLengthToShowInErrorMessages) { \
577 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
578 len = RegexLengthToShowInErrorMessages - 10; \
584 #define FAIL(msg) _FAIL( \
585 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
586 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
588 #define FAIL2(msg,arg) _FAIL( \
589 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
590 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
593 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
595 #define Simple_vFAIL(m) STMT_START { \
597 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
598 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
599 m, REPORT_LOCATION_ARGS(offset)); \
603 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
605 #define vFAIL(m) STMT_START { \
607 SAVEFREESV(RExC_rx_sv); \
612 * Like Simple_vFAIL(), but accepts two arguments.
614 #define Simple_vFAIL2(m,a1) STMT_START { \
615 const IV offset = RExC_parse - RExC_precomp; \
616 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
617 REPORT_LOCATION_ARGS(offset)); \
621 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
623 #define vFAIL2(m,a1) STMT_START { \
625 SAVEFREESV(RExC_rx_sv); \
626 Simple_vFAIL2(m, a1); \
631 * Like Simple_vFAIL(), but accepts three arguments.
633 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
634 const IV offset = RExC_parse - RExC_precomp; \
635 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
636 REPORT_LOCATION_ARGS(offset)); \
640 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
642 #define vFAIL3(m,a1,a2) STMT_START { \
644 SAVEFREESV(RExC_rx_sv); \
645 Simple_vFAIL3(m, a1, a2); \
649 * Like Simple_vFAIL(), but accepts four arguments.
651 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
652 const IV offset = RExC_parse - RExC_precomp; \
653 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
654 REPORT_LOCATION_ARGS(offset)); \
657 #define vFAIL4(m,a1,a2,a3) STMT_START { \
659 SAVEFREESV(RExC_rx_sv); \
660 Simple_vFAIL4(m, a1, a2, a3); \
663 /* A specialized version of vFAIL2 that works with UTF8f */
664 #define vFAIL2utf8f(m, a1) STMT_START { \
665 const IV offset = RExC_parse - RExC_precomp; \
667 SAVEFREESV(RExC_rx_sv); \
668 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
669 REPORT_LOCATION_ARGS(offset)); \
672 /* These have asserts in them because of [perl #122671] Many warnings in
673 * regcomp.c can occur twice. If they get output in pass1 and later in that
674 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
675 * would get output again. So they should be output in pass2, and these
676 * asserts make sure new warnings follow that paradigm. */
678 /* m is not necessarily a "literal string", in this macro */
679 #define reg_warn_non_literal_string(loc, m) STMT_START { \
680 const IV offset = loc - RExC_precomp; \
681 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
682 m, REPORT_LOCATION_ARGS(offset)); \
685 #define ckWARNreg(loc,m) STMT_START { \
686 const IV offset = loc - RExC_precomp; \
687 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
688 REPORT_LOCATION_ARGS(offset)); \
691 #define vWARN(loc, m) STMT_START { \
692 const IV offset = loc - RExC_precomp; \
693 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
694 REPORT_LOCATION_ARGS(offset)); \
697 #define vWARN_dep(loc, m) STMT_START { \
698 const IV offset = loc - RExC_precomp; \
699 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
700 REPORT_LOCATION_ARGS(offset)); \
703 #define ckWARNdep(loc,m) STMT_START { \
704 const IV offset = loc - RExC_precomp; \
705 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
707 REPORT_LOCATION_ARGS(offset)); \
710 #define ckWARNregdep(loc,m) STMT_START { \
711 const IV offset = loc - RExC_precomp; \
712 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
714 REPORT_LOCATION_ARGS(offset)); \
717 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
718 const IV offset = loc - RExC_precomp; \
719 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
721 a1, REPORT_LOCATION_ARGS(offset)); \
724 #define ckWARN2reg(loc, m, a1) STMT_START { \
725 const IV offset = loc - RExC_precomp; \
726 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
727 a1, REPORT_LOCATION_ARGS(offset)); \
730 #define vWARN3(loc, m, a1, a2) STMT_START { \
731 const IV offset = loc - RExC_precomp; \
732 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
733 a1, a2, REPORT_LOCATION_ARGS(offset)); \
736 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
737 const IV offset = loc - RExC_precomp; \
738 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
739 a1, a2, REPORT_LOCATION_ARGS(offset)); \
742 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
743 const IV offset = loc - RExC_precomp; \
744 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
745 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
748 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
749 const IV offset = loc - RExC_precomp; \
750 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
751 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
754 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
755 const IV offset = loc - RExC_precomp; \
756 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
757 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
760 /* Macros for recording node offsets. 20001227 mjd@plover.com
761 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
762 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
763 * Element 0 holds the number n.
764 * Position is 1 indexed.
766 #ifndef RE_TRACK_PATTERN_OFFSETS
767 #define Set_Node_Offset_To_R(node,byte)
768 #define Set_Node_Offset(node,byte)
769 #define Set_Cur_Node_Offset
770 #define Set_Node_Length_To_R(node,len)
771 #define Set_Node_Length(node,len)
772 #define Set_Node_Cur_Length(node,start)
773 #define Node_Offset(n)
774 #define Node_Length(n)
775 #define Set_Node_Offset_Length(node,offset,len)
776 #define ProgLen(ri) ri->u.proglen
777 #define SetProgLen(ri,x) ri->u.proglen = x
779 #define ProgLen(ri) ri->u.offsets[0]
780 #define SetProgLen(ri,x) ri->u.offsets[0] = x
781 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
783 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
784 __LINE__, (int)(node), (int)(byte))); \
786 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
789 RExC_offsets[2*(node)-1] = (byte); \
794 #define Set_Node_Offset(node,byte) \
795 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
796 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
798 #define Set_Node_Length_To_R(node,len) STMT_START { \
800 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
801 __LINE__, (int)(node), (int)(len))); \
803 Perl_croak(aTHX_ "value of node is %d in Length macro", \
806 RExC_offsets[2*(node)] = (len); \
811 #define Set_Node_Length(node,len) \
812 Set_Node_Length_To_R((node)-RExC_emit_start, len)
813 #define Set_Node_Cur_Length(node, start) \
814 Set_Node_Length(node, RExC_parse - start)
816 /* Get offsets and lengths */
817 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
818 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
820 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
821 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
822 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
826 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
827 #define EXPERIMENTAL_INPLACESCAN
828 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
830 #define DEBUG_RExC_seen() \
831 DEBUG_OPTIMISE_MORE_r({ \
832 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
834 if (RExC_seen & REG_ZERO_LEN_SEEN) \
835 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
837 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
838 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
840 if (RExC_seen & REG_GPOS_SEEN) \
841 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
843 if (RExC_seen & REG_RECURSE_SEEN) \
844 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
846 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
847 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
849 if (RExC_seen & REG_VERBARG_SEEN) \
850 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
852 if (RExC_seen & REG_CUTGROUP_SEEN) \
853 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
855 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
856 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
858 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
859 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
861 if (RExC_seen & REG_GOSTART_SEEN) \
862 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
864 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
865 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
867 PerlIO_printf(Perl_debug_log,"\n"); \
870 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
871 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
873 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
875 PerlIO_printf(Perl_debug_log, "%s", open_str); \
876 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
877 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
878 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
879 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
880 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
881 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
882 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
883 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
884 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
885 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
886 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
887 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
888 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
889 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
890 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
891 PerlIO_printf(Perl_debug_log, "%s", close_str); \
895 #define DEBUG_STUDYDATA(str,data,depth) \
896 DEBUG_OPTIMISE_MORE_r(if(data){ \
897 PerlIO_printf(Perl_debug_log, \
898 "%*s" str "Pos:%"IVdf"/%"IVdf \
900 (int)(depth)*2, "", \
901 (IV)((data)->pos_min), \
902 (IV)((data)->pos_delta), \
903 (UV)((data)->flags) \
905 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
906 PerlIO_printf(Perl_debug_log, \
907 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
908 (IV)((data)->whilem_c), \
909 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
910 is_inf ? "INF " : "" \
912 if ((data)->last_found) \
913 PerlIO_printf(Perl_debug_log, \
914 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
915 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
916 SvPVX_const((data)->last_found), \
917 (IV)((data)->last_end), \
918 (IV)((data)->last_start_min), \
919 (IV)((data)->last_start_max), \
920 ((data)->longest && \
921 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
922 SvPVX_const((data)->longest_fixed), \
923 (IV)((data)->offset_fixed), \
924 ((data)->longest && \
925 (data)->longest==&((data)->longest_float)) ? "*" : "", \
926 SvPVX_const((data)->longest_float), \
927 (IV)((data)->offset_float_min), \
928 (IV)((data)->offset_float_max) \
930 PerlIO_printf(Perl_debug_log,"\n"); \
933 /* is c a control character for which we have a mnemonic? */
934 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
937 S_cntrl_to_mnemonic(const U8 c)
939 /* Returns the mnemonic string that represents character 'c', if one
940 * exists; NULL otherwise. The only ones that exist for the purposes of
941 * this routine are a few control characters */
944 case '\a': return "\\a";
945 case '\b': return "\\b";
946 case ESC_NATIVE: return "\\e";
947 case '\f': return "\\f";
948 case '\n': return "\\n";
949 case '\r': return "\\r";
950 case '\t': return "\\t";
956 /* Mark that we cannot extend a found fixed substring at this point.
957 Update the longest found anchored substring and the longest found
958 floating substrings if needed. */
961 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
962 SSize_t *minlenp, int is_inf)
964 const STRLEN l = CHR_SVLEN(data->last_found);
965 const STRLEN old_l = CHR_SVLEN(*data->longest);
966 GET_RE_DEBUG_FLAGS_DECL;
968 PERL_ARGS_ASSERT_SCAN_COMMIT;
970 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
971 SvSetMagicSV(*data->longest, data->last_found);
972 if (*data->longest == data->longest_fixed) {
973 data->offset_fixed = l ? data->last_start_min : data->pos_min;
974 if (data->flags & SF_BEFORE_EOL)
976 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
978 data->flags &= ~SF_FIX_BEFORE_EOL;
979 data->minlen_fixed=minlenp;
980 data->lookbehind_fixed=0;
982 else { /* *data->longest == data->longest_float */
983 data->offset_float_min = l ? data->last_start_min : data->pos_min;
984 data->offset_float_max = (l
985 ? data->last_start_max
986 : (data->pos_delta > SSize_t_MAX - data->pos_min
988 : data->pos_min + data->pos_delta));
990 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
991 data->offset_float_max = SSize_t_MAX;
992 if (data->flags & SF_BEFORE_EOL)
994 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
996 data->flags &= ~SF_FL_BEFORE_EOL;
997 data->minlen_float=minlenp;
998 data->lookbehind_float=0;
1001 SvCUR_set(data->last_found, 0);
1003 SV * const sv = data->last_found;
1004 if (SvUTF8(sv) && SvMAGICAL(sv)) {
1005 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
1010 data->last_end = -1;
1011 data->flags &= ~SF_BEFORE_EOL;
1012 DEBUG_STUDYDATA("commit: ",data,0);
1015 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
1016 * list that describes which code points it matches */
1019 S_ssc_anything(pTHX_ regnode_ssc *ssc)
1021 /* Set the SSC 'ssc' to match an empty string or any code point */
1023 PERL_ARGS_ASSERT_SSC_ANYTHING;
1025 assert(is_ANYOF_SYNTHETIC(ssc));
1027 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
1028 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1029 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1033 S_ssc_is_anything(const regnode_ssc *ssc)
1035 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1036 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1037 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1038 * in any way, so there's no point in using it */
1043 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1045 assert(is_ANYOF_SYNTHETIC(ssc));
1047 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1051 /* See if the list consists solely of the range 0 - Infinity */
1052 invlist_iterinit(ssc->invlist);
1053 ret = invlist_iternext(ssc->invlist, &start, &end)
1057 invlist_iterfinish(ssc->invlist);
1063 /* If e.g., both \w and \W are set, matches everything */
1064 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1066 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1067 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1077 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1079 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1080 * string, any code point, or any posix class under locale */
1082 PERL_ARGS_ASSERT_SSC_INIT;
1084 Zero(ssc, 1, regnode_ssc);
1085 set_ANYOF_SYNTHETIC(ssc);
1086 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1089 /* If any portion of the regex is to operate under locale rules that aren't
1090 * fully known at compile time, initialization includes it. The reason
1091 * this isn't done for all regexes is that the optimizer was written under
1092 * the assumption that locale was all-or-nothing. Given the complexity and
1093 * lack of documentation in the optimizer, and that there are inadequate
1094 * test cases for locale, many parts of it may not work properly, it is
1095 * safest to avoid locale unless necessary. */
1096 if (RExC_contains_locale) {
1097 ANYOF_POSIXL_SETALL(ssc);
1100 ANYOF_POSIXL_ZERO(ssc);
1105 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1106 const regnode_ssc *ssc)
1108 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1109 * to the list of code points matched, and locale posix classes; hence does
1110 * not check its flags) */
1115 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1117 assert(is_ANYOF_SYNTHETIC(ssc));
1119 invlist_iterinit(ssc->invlist);
1120 ret = invlist_iternext(ssc->invlist, &start, &end)
1124 invlist_iterfinish(ssc->invlist);
1130 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1138 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1139 const regnode_charclass* const node)
1141 /* Returns a mortal inversion list defining which code points are matched
1142 * by 'node', which is of type ANYOF. Handles complementing the result if
1143 * appropriate. If some code points aren't knowable at this time, the
1144 * returned list must, and will, contain every code point that is a
1147 SV* invlist = sv_2mortal(_new_invlist(0));
1148 SV* only_utf8_locale_invlist = NULL;
1150 const U32 n = ARG(node);
1151 bool new_node_has_latin1 = FALSE;
1153 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1155 /* Look at the data structure created by S_set_ANYOF_arg() */
1156 if (n != ANYOF_ONLY_HAS_BITMAP) {
1157 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1158 AV * const av = MUTABLE_AV(SvRV(rv));
1159 SV **const ary = AvARRAY(av);
1160 assert(RExC_rxi->data->what[n] == 's');
1162 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1163 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1165 else if (ary[0] && ary[0] != &PL_sv_undef) {
1167 /* Here, no compile-time swash, and there are things that won't be
1168 * known until runtime -- we have to assume it could be anything */
1169 return _add_range_to_invlist(invlist, 0, UV_MAX);
1171 else if (ary[3] && ary[3] != &PL_sv_undef) {
1173 /* Here no compile-time swash, and no run-time only data. Use the
1174 * node's inversion list */
1175 invlist = sv_2mortal(invlist_clone(ary[3]));
1178 /* Get the code points valid only under UTF-8 locales */
1179 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1180 && ary[2] && ary[2] != &PL_sv_undef)
1182 only_utf8_locale_invlist = ary[2];
1186 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1187 * code points, and an inversion list for the others, but if there are code
1188 * points that should match only conditionally on the target string being
1189 * UTF-8, those are placed in the inversion list, and not the bitmap.
1190 * Since there are circumstances under which they could match, they are
1191 * included in the SSC. But if the ANYOF node is to be inverted, we have
1192 * to exclude them here, so that when we invert below, the end result
1193 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1194 * have to do this here before we add the unconditionally matched code
1196 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1197 _invlist_intersection_complement_2nd(invlist,
1202 /* Add in the points from the bit map */
1203 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1204 if (ANYOF_BITMAP_TEST(node, i)) {
1205 invlist = add_cp_to_invlist(invlist, i);
1206 new_node_has_latin1 = TRUE;
1210 /* If this can match all upper Latin1 code points, have to add them
1212 if (OP(node) == ANYOFD
1213 && (ANYOF_FLAGS(node) & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
1215 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1218 /* Similarly for these */
1219 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1220 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1223 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1224 _invlist_invert(invlist);
1226 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1228 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1229 * locale. We can skip this if there are no 0-255 at all. */
1230 _invlist_union(invlist, PL_Latin1, &invlist);
1233 /* Similarly add the UTF-8 locale possible matches. These have to be
1234 * deferred until after the non-UTF-8 locale ones are taken care of just
1235 * above, or it leads to wrong results under ANYOF_INVERT */
1236 if (only_utf8_locale_invlist) {
1237 _invlist_union_maybe_complement_2nd(invlist,
1238 only_utf8_locale_invlist,
1239 ANYOF_FLAGS(node) & ANYOF_INVERT,
1246 /* These two functions currently do the exact same thing */
1247 #define ssc_init_zero ssc_init
1249 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1250 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1252 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1253 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1254 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1257 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1258 const regnode_charclass *and_with)
1260 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1261 * another SSC or a regular ANYOF class. Can create false positives. */
1266 PERL_ARGS_ASSERT_SSC_AND;
1268 assert(is_ANYOF_SYNTHETIC(ssc));
1270 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1271 * the code point inversion list and just the relevant flags */
1272 if (is_ANYOF_SYNTHETIC(and_with)) {
1273 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1274 anded_flags = ANYOF_FLAGS(and_with);
1276 /* XXX This is a kludge around what appears to be deficiencies in the
1277 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1278 * there are paths through the optimizer where it doesn't get weeded
1279 * out when it should. And if we don't make some extra provision for
1280 * it like the code just below, it doesn't get added when it should.
1281 * This solution is to add it only when AND'ing, which is here, and
1282 * only when what is being AND'ed is the pristine, original node
1283 * matching anything. Thus it is like adding it to ssc_anything() but
1284 * only when the result is to be AND'ed. Probably the same solution
1285 * could be adopted for the same problem we have with /l matching,
1286 * which is solved differently in S_ssc_init(), and that would lead to
1287 * fewer false positives than that solution has. But if this solution
1288 * creates bugs, the consequences are only that a warning isn't raised
1289 * that should be; while the consequences for having /l bugs is
1290 * incorrect matches */
1291 if (ssc_is_anything((regnode_ssc *)and_with)) {
1292 anded_flags |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1296 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1297 if (OP(and_with) == ANYOFD) {
1298 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1301 anded_flags = ANYOF_FLAGS(and_with)
1302 &( ANYOF_COMMON_FLAGS
1303 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER);
1307 ANYOF_FLAGS(ssc) &= anded_flags;
1309 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1310 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1311 * 'and_with' may be inverted. When not inverted, we have the situation of
1313 * (C1 | P1) & (C2 | P2)
1314 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1315 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1316 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1317 * <= ((C1 & C2) | P1 | P2)
1318 * Alternatively, the last few steps could be:
1319 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1320 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1321 * <= (C1 | C2 | (P1 & P2))
1322 * We favor the second approach if either P1 or P2 is non-empty. This is
1323 * because these components are a barrier to doing optimizations, as what
1324 * they match cannot be known until the moment of matching as they are
1325 * dependent on the current locale, 'AND"ing them likely will reduce or
1327 * But we can do better if we know that C1,P1 are in their initial state (a
1328 * frequent occurrence), each matching everything:
1329 * (<everything>) & (C2 | P2) = C2 | P2
1330 * Similarly, if C2,P2 are in their initial state (again a frequent
1331 * occurrence), the result is a no-op
1332 * (C1 | P1) & (<everything>) = C1 | P1
1335 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1336 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1337 * <= (C1 & ~C2) | (P1 & ~P2)
1340 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1341 && ! is_ANYOF_SYNTHETIC(and_with))
1345 ssc_intersection(ssc,
1347 FALSE /* Has already been inverted */
1350 /* If either P1 or P2 is empty, the intersection will be also; can skip
1352 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1353 ANYOF_POSIXL_ZERO(ssc);
1355 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1357 /* Note that the Posix class component P from 'and_with' actually
1359 * P = Pa | Pb | ... | Pn
1360 * where each component is one posix class, such as in [\w\s].
1362 * ~P = ~(Pa | Pb | ... | Pn)
1363 * = ~Pa & ~Pb & ... & ~Pn
1364 * <= ~Pa | ~Pb | ... | ~Pn
1365 * The last is something we can easily calculate, but unfortunately
1366 * is likely to have many false positives. We could do better
1367 * in some (but certainly not all) instances if two classes in
1368 * P have known relationships. For example
1369 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1371 * :lower: & :print: = :lower:
1372 * And similarly for classes that must be disjoint. For example,
1373 * since \s and \w can have no elements in common based on rules in
1374 * the POSIX standard,
1375 * \w & ^\S = nothing
1376 * Unfortunately, some vendor locales do not meet the Posix
1377 * standard, in particular almost everything by Microsoft.
1378 * The loop below just changes e.g., \w into \W and vice versa */
1380 regnode_charclass_posixl temp;
1381 int add = 1; /* To calculate the index of the complement */
1383 ANYOF_POSIXL_ZERO(&temp);
1384 for (i = 0; i < ANYOF_MAX; i++) {
1386 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1387 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1389 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1390 ANYOF_POSIXL_SET(&temp, i + add);
1392 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1394 ANYOF_POSIXL_AND(&temp, ssc);
1396 } /* else ssc already has no posixes */
1397 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1398 in its initial state */
1399 else if (! is_ANYOF_SYNTHETIC(and_with)
1400 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1402 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1403 * copy it over 'ssc' */
1404 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1405 if (is_ANYOF_SYNTHETIC(and_with)) {
1406 StructCopy(and_with, ssc, regnode_ssc);
1409 ssc->invlist = anded_cp_list;
1410 ANYOF_POSIXL_ZERO(ssc);
1411 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1412 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1416 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1417 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1419 /* One or the other of P1, P2 is non-empty. */
1420 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1421 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1423 ssc_union(ssc, anded_cp_list, FALSE);
1425 else { /* P1 = P2 = empty */
1426 ssc_intersection(ssc, anded_cp_list, FALSE);
1432 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1433 const regnode_charclass *or_with)
1435 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1436 * another SSC or a regular ANYOF class. Can create false positives if
1437 * 'or_with' is to be inverted. */
1442 PERL_ARGS_ASSERT_SSC_OR;
1444 assert(is_ANYOF_SYNTHETIC(ssc));
1446 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1447 * the code point inversion list and just the relevant flags */
1448 if (is_ANYOF_SYNTHETIC(or_with)) {
1449 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1450 ored_flags = ANYOF_FLAGS(or_with);
1453 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1454 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1455 if (OP(or_with) != ANYOFD) {
1457 |= ANYOF_FLAGS(or_with)
1458 & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
1462 ANYOF_FLAGS(ssc) |= ored_flags;
1464 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1465 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1466 * 'or_with' may be inverted. When not inverted, we have the simple
1467 * situation of computing:
1468 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1469 * If P1|P2 yields a situation with both a class and its complement are
1470 * set, like having both \w and \W, this matches all code points, and we
1471 * can delete these from the P component of the ssc going forward. XXX We
1472 * might be able to delete all the P components, but I (khw) am not certain
1473 * about this, and it is better to be safe.
1476 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1477 * <= (C1 | P1) | ~C2
1478 * <= (C1 | ~C2) | P1
1479 * (which results in actually simpler code than the non-inverted case)
1482 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1483 && ! is_ANYOF_SYNTHETIC(or_with))
1485 /* We ignore P2, leaving P1 going forward */
1486 } /* else Not inverted */
1487 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1488 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1489 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1491 for (i = 0; i < ANYOF_MAX; i += 2) {
1492 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1494 ssc_match_all_cp(ssc);
1495 ANYOF_POSIXL_CLEAR(ssc, i);
1496 ANYOF_POSIXL_CLEAR(ssc, i+1);
1504 FALSE /* Already has been inverted */
1508 PERL_STATIC_INLINE void
1509 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1511 PERL_ARGS_ASSERT_SSC_UNION;
1513 assert(is_ANYOF_SYNTHETIC(ssc));
1515 _invlist_union_maybe_complement_2nd(ssc->invlist,
1521 PERL_STATIC_INLINE void
1522 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1524 const bool invert2nd)
1526 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1528 assert(is_ANYOF_SYNTHETIC(ssc));
1530 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1536 PERL_STATIC_INLINE void
1537 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1539 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1541 assert(is_ANYOF_SYNTHETIC(ssc));
1543 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1546 PERL_STATIC_INLINE void
1547 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1549 /* AND just the single code point 'cp' into the SSC 'ssc' */
1551 SV* cp_list = _new_invlist(2);
1553 PERL_ARGS_ASSERT_SSC_CP_AND;
1555 assert(is_ANYOF_SYNTHETIC(ssc));
1557 cp_list = add_cp_to_invlist(cp_list, cp);
1558 ssc_intersection(ssc, cp_list,
1559 FALSE /* Not inverted */
1561 SvREFCNT_dec_NN(cp_list);
1564 PERL_STATIC_INLINE void
1565 S_ssc_clear_locale(regnode_ssc *ssc)
1567 /* Set the SSC 'ssc' to not match any locale things */
1568 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1570 assert(is_ANYOF_SYNTHETIC(ssc));
1572 ANYOF_POSIXL_ZERO(ssc);
1573 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1576 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1579 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1581 /* The synthetic start class is used to hopefully quickly winnow down
1582 * places where a pattern could start a match in the target string. If it
1583 * doesn't really narrow things down that much, there isn't much point to
1584 * having the overhead of using it. This function uses some very crude
1585 * heuristics to decide if to use the ssc or not.
1587 * It returns TRUE if 'ssc' rules out more than half what it considers to
1588 * be the "likely" possible matches, but of course it doesn't know what the
1589 * actual things being matched are going to be; these are only guesses
1591 * For /l matches, it assumes that the only likely matches are going to be
1592 * in the 0-255 range, uniformly distributed, so half of that is 127
1593 * For /a and /d matches, it assumes that the likely matches will be just
1594 * the ASCII range, so half of that is 63
1595 * For /u and there isn't anything matching above the Latin1 range, it
1596 * assumes that that is the only range likely to be matched, and uses
1597 * half that as the cut-off: 127. If anything matches above Latin1,
1598 * it assumes that all of Unicode could match (uniformly), except for
1599 * non-Unicode code points and things in the General Category "Other"
1600 * (unassigned, private use, surrogates, controls and formats). This
1601 * is a much large number. */
1603 const U32 max_match = (LOC)
1607 : (invlist_highest(ssc->invlist) < 256)
1609 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1610 U32 count = 0; /* Running total of number of code points matched by
1612 UV start, end; /* Start and end points of current range in inversion
1615 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1617 invlist_iterinit(ssc->invlist);
1618 while (invlist_iternext(ssc->invlist, &start, &end)) {
1620 /* /u is the only thing that we expect to match above 255; so if not /u
1621 * and even if there are matches above 255, ignore them. This catches
1622 * things like \d under /d which does match the digits above 255, but
1623 * since the pattern is /d, it is not likely to be expecting them */
1624 if (! UNI_SEMANTICS) {
1628 end = MIN(end, 255);
1630 count += end - start + 1;
1631 if (count > max_match) {
1632 invlist_iterfinish(ssc->invlist);
1642 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1644 /* The inversion list in the SSC is marked mortal; now we need a more
1645 * permanent copy, which is stored the same way that is done in a regular
1646 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1649 SV* invlist = invlist_clone(ssc->invlist);
1651 PERL_ARGS_ASSERT_SSC_FINALIZE;
1653 assert(is_ANYOF_SYNTHETIC(ssc));
1655 /* The code in this file assumes that all but these flags aren't relevant
1656 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1657 * by the time we reach here */
1658 assert(! (ANYOF_FLAGS(ssc)
1659 & ~( ANYOF_COMMON_FLAGS
1660 |ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER)));
1662 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1664 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1665 NULL, NULL, NULL, FALSE);
1667 /* Make sure is clone-safe */
1668 ssc->invlist = NULL;
1670 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1671 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1674 if (RExC_contains_locale) {
1678 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1681 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1682 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1683 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1684 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1685 ? (TRIE_LIST_CUR( idx ) - 1) \
1691 dump_trie(trie,widecharmap,revcharmap)
1692 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1693 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1695 These routines dump out a trie in a somewhat readable format.
1696 The _interim_ variants are used for debugging the interim
1697 tables that are used to generate the final compressed
1698 representation which is what dump_trie expects.
1700 Part of the reason for their existence is to provide a form
1701 of documentation as to how the different representations function.
1706 Dumps the final compressed table form of the trie to Perl_debug_log.
1707 Used for debugging make_trie().
1711 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1712 AV *revcharmap, U32 depth)
1715 SV *sv=sv_newmortal();
1716 int colwidth= widecharmap ? 6 : 4;
1718 GET_RE_DEBUG_FLAGS_DECL;
1720 PERL_ARGS_ASSERT_DUMP_TRIE;
1722 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1723 (int)depth * 2 + 2,"",
1724 "Match","Base","Ofs" );
1726 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1727 SV ** const tmp = av_fetch( revcharmap, state, 0);
1729 PerlIO_printf( Perl_debug_log, "%*s",
1731 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1732 PL_colors[0], PL_colors[1],
1733 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1734 PERL_PV_ESCAPE_FIRSTCHAR
1739 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1740 (int)depth * 2 + 2,"");
1742 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1743 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1744 PerlIO_printf( Perl_debug_log, "\n");
1746 for( state = 1 ; state < trie->statecount ; state++ ) {
1747 const U32 base = trie->states[ state ].trans.base;
1749 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1750 (int)depth * 2 + 2,"", (UV)state);
1752 if ( trie->states[ state ].wordnum ) {
1753 PerlIO_printf( Perl_debug_log, " W%4X",
1754 trie->states[ state ].wordnum );
1756 PerlIO_printf( Perl_debug_log, "%6s", "" );
1759 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1764 while( ( base + ofs < trie->uniquecharcount ) ||
1765 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1766 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1770 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1772 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1773 if ( ( base + ofs >= trie->uniquecharcount )
1774 && ( base + ofs - trie->uniquecharcount
1776 && trie->trans[ base + ofs
1777 - trie->uniquecharcount ].check == state )
1779 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1781 (UV)trie->trans[ base + ofs
1782 - trie->uniquecharcount ].next );
1784 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1788 PerlIO_printf( Perl_debug_log, "]");
1791 PerlIO_printf( Perl_debug_log, "\n" );
1793 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1795 for (word=1; word <= trie->wordcount; word++) {
1796 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1797 (int)word, (int)(trie->wordinfo[word].prev),
1798 (int)(trie->wordinfo[word].len));
1800 PerlIO_printf(Perl_debug_log, "\n" );
1803 Dumps a fully constructed but uncompressed trie in list form.
1804 List tries normally only are used for construction when the number of
1805 possible chars (trie->uniquecharcount) is very high.
1806 Used for debugging make_trie().
1809 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1810 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1814 SV *sv=sv_newmortal();
1815 int colwidth= widecharmap ? 6 : 4;
1816 GET_RE_DEBUG_FLAGS_DECL;
1818 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1820 /* print out the table precompression. */
1821 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1822 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1823 "------:-----+-----------------\n" );
1825 for( state=1 ; state < next_alloc ; state ++ ) {
1828 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1829 (int)depth * 2 + 2,"", (UV)state );
1830 if ( ! trie->states[ state ].wordnum ) {
1831 PerlIO_printf( Perl_debug_log, "%5s| ","");
1833 PerlIO_printf( Perl_debug_log, "W%4x| ",
1834 trie->states[ state ].wordnum
1837 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1838 SV ** const tmp = av_fetch( revcharmap,
1839 TRIE_LIST_ITEM(state,charid).forid, 0);
1841 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1843 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1845 PL_colors[0], PL_colors[1],
1846 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1847 | PERL_PV_ESCAPE_FIRSTCHAR
1849 TRIE_LIST_ITEM(state,charid).forid,
1850 (UV)TRIE_LIST_ITEM(state,charid).newstate
1853 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1854 (int)((depth * 2) + 14), "");
1857 PerlIO_printf( Perl_debug_log, "\n");
1862 Dumps a fully constructed but uncompressed trie in table form.
1863 This is the normal DFA style state transition table, with a few
1864 twists to facilitate compression later.
1865 Used for debugging make_trie().
1868 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1869 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1874 SV *sv=sv_newmortal();
1875 int colwidth= widecharmap ? 6 : 4;
1876 GET_RE_DEBUG_FLAGS_DECL;
1878 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1881 print out the table precompression so that we can do a visual check
1882 that they are identical.
1885 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1887 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1888 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1890 PerlIO_printf( Perl_debug_log, "%*s",
1892 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1893 PL_colors[0], PL_colors[1],
1894 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1895 PERL_PV_ESCAPE_FIRSTCHAR
1901 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1903 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1904 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1907 PerlIO_printf( Perl_debug_log, "\n" );
1909 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1911 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1912 (int)depth * 2 + 2,"",
1913 (UV)TRIE_NODENUM( state ) );
1915 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1916 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1918 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1920 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1922 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1923 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1924 (UV)trie->trans[ state ].check );
1926 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1927 (UV)trie->trans[ state ].check,
1928 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1936 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1937 startbranch: the first branch in the whole branch sequence
1938 first : start branch of sequence of branch-exact nodes.
1939 May be the same as startbranch
1940 last : Thing following the last branch.
1941 May be the same as tail.
1942 tail : item following the branch sequence
1943 count : words in the sequence
1944 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1945 depth : indent depth
1947 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1949 A trie is an N'ary tree where the branches are determined by digital
1950 decomposition of the key. IE, at the root node you look up the 1st character and
1951 follow that branch repeat until you find the end of the branches. Nodes can be
1952 marked as "accepting" meaning they represent a complete word. Eg:
1956 would convert into the following structure. Numbers represent states, letters
1957 following numbers represent valid transitions on the letter from that state, if
1958 the number is in square brackets it represents an accepting state, otherwise it
1959 will be in parenthesis.
1961 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1965 (1) +-i->(6)-+-s->[7]
1967 +-s->(3)-+-h->(4)-+-e->[5]
1969 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1971 This shows that when matching against the string 'hers' we will begin at state 1
1972 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1973 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1974 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1975 single traverse. We store a mapping from accepting to state to which word was
1976 matched, and then when we have multiple possibilities we try to complete the
1977 rest of the regex in the order in which they occurred in the alternation.
1979 The only prior NFA like behaviour that would be changed by the TRIE support is
1980 the silent ignoring of duplicate alternations which are of the form:
1982 / (DUPE|DUPE) X? (?{ ... }) Y /x
1984 Thus EVAL blocks following a trie may be called a different number of times with
1985 and without the optimisation. With the optimisations dupes will be silently
1986 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1987 the following demonstrates:
1989 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1991 which prints out 'word' three times, but
1993 'words'=~/(word|word|word)(?{ print $1 })S/
1995 which doesnt print it out at all. This is due to other optimisations kicking in.
1997 Example of what happens on a structural level:
1999 The regexp /(ac|ad|ab)+/ will produce the following debug output:
2001 1: CURLYM[1] {1,32767}(18)
2012 This would be optimizable with startbranch=5, first=5, last=16, tail=16
2013 and should turn into:
2015 1: CURLYM[1] {1,32767}(18)
2017 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
2025 Cases where tail != last would be like /(?foo|bar)baz/:
2035 which would be optimizable with startbranch=1, first=1, last=7, tail=8
2036 and would end up looking like:
2039 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
2046 d = uvchr_to_utf8_flags(d, uv, 0);
2048 is the recommended Unicode-aware way of saying
2053 #define TRIE_STORE_REVCHAR(val) \
2056 SV *zlopp = newSV(UTF8_MAXBYTES); \
2057 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2058 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2059 SvCUR_set(zlopp, kapow - flrbbbbb); \
2062 av_push(revcharmap, zlopp); \
2064 char ooooff = (char)val; \
2065 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2069 /* This gets the next character from the input, folding it if not already
2071 #define TRIE_READ_CHAR STMT_START { \
2074 /* if it is UTF then it is either already folded, or does not need \
2076 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2078 else if (folder == PL_fold_latin1) { \
2079 /* This folder implies Unicode rules, which in the range expressible \
2080 * by not UTF is the lower case, with the two exceptions, one of \
2081 * which should have been taken care of before calling this */ \
2082 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2083 uvc = toLOWER_L1(*uc); \
2084 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2087 /* raw data, will be folded later if needed */ \
2095 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2096 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2097 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2098 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2100 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2101 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2102 TRIE_LIST_CUR( state )++; \
2105 #define TRIE_LIST_NEW(state) STMT_START { \
2106 Newxz( trie->states[ state ].trans.list, \
2107 4, reg_trie_trans_le ); \
2108 TRIE_LIST_CUR( state ) = 1; \
2109 TRIE_LIST_LEN( state ) = 4; \
2112 #define TRIE_HANDLE_WORD(state) STMT_START { \
2113 U16 dupe= trie->states[ state ].wordnum; \
2114 regnode * const noper_next = regnext( noper ); \
2117 /* store the word for dumping */ \
2119 if (OP(noper) != NOTHING) \
2120 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2122 tmp = newSVpvn_utf8( "", 0, UTF ); \
2123 av_push( trie_words, tmp ); \
2127 trie->wordinfo[curword].prev = 0; \
2128 trie->wordinfo[curword].len = wordlen; \
2129 trie->wordinfo[curword].accept = state; \
2131 if ( noper_next < tail ) { \
2133 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2135 trie->jump[curword] = (U16)(noper_next - convert); \
2137 jumper = noper_next; \
2139 nextbranch= regnext(cur); \
2143 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2144 /* chain, so that when the bits of chain are later */\
2145 /* linked together, the dups appear in the chain */\
2146 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2147 trie->wordinfo[dupe].prev = curword; \
2149 /* we haven't inserted this word yet. */ \
2150 trie->states[ state ].wordnum = curword; \
2155 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2156 ( ( base + charid >= ucharcount \
2157 && base + charid < ubound \
2158 && state == trie->trans[ base - ucharcount + charid ].check \
2159 && trie->trans[ base - ucharcount + charid ].next ) \
2160 ? trie->trans[ base - ucharcount + charid ].next \
2161 : ( state==1 ? special : 0 ) \
2165 #define MADE_JUMP_TRIE 2
2166 #define MADE_EXACT_TRIE 4
2169 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2170 regnode *first, regnode *last, regnode *tail,
2171 U32 word_count, U32 flags, U32 depth)
2173 /* first pass, loop through and scan words */
2174 reg_trie_data *trie;
2175 HV *widecharmap = NULL;
2176 AV *revcharmap = newAV();
2182 regnode *jumper = NULL;
2183 regnode *nextbranch = NULL;
2184 regnode *convert = NULL;
2185 U32 *prev_states; /* temp array mapping each state to previous one */
2186 /* we just use folder as a flag in utf8 */
2187 const U8 * folder = NULL;
2190 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2191 AV *trie_words = NULL;
2192 /* along with revcharmap, this only used during construction but both are
2193 * useful during debugging so we store them in the struct when debugging.
2196 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2197 STRLEN trie_charcount=0;
2199 SV *re_trie_maxbuff;
2200 GET_RE_DEBUG_FLAGS_DECL;
2202 PERL_ARGS_ASSERT_MAKE_TRIE;
2204 PERL_UNUSED_ARG(depth);
2208 case EXACT: case EXACTL: break;
2212 case EXACTFLU8: folder = PL_fold_latin1; break;
2213 case EXACTF: folder = PL_fold; break;
2214 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2217 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2219 trie->startstate = 1;
2220 trie->wordcount = word_count;
2221 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2222 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2223 if (flags == EXACT || flags == EXACTL)
2224 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2225 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2226 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2229 trie_words = newAV();
2232 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2233 assert(re_trie_maxbuff);
2234 if (!SvIOK(re_trie_maxbuff)) {
2235 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2237 DEBUG_TRIE_COMPILE_r({
2238 PerlIO_printf( Perl_debug_log,
2239 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2240 (int)depth * 2 + 2, "",
2241 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2242 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2245 /* Find the node we are going to overwrite */
2246 if ( first == startbranch && OP( last ) != BRANCH ) {
2247 /* whole branch chain */
2250 /* branch sub-chain */
2251 convert = NEXTOPER( first );
2254 /* -- First loop and Setup --
2256 We first traverse the branches and scan each word to determine if it
2257 contains widechars, and how many unique chars there are, this is
2258 important as we have to build a table with at least as many columns as we
2261 We use an array of integers to represent the character codes 0..255
2262 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2263 the native representation of the character value as the key and IV's for
2266 *TODO* If we keep track of how many times each character is used we can
2267 remap the columns so that the table compression later on is more
2268 efficient in terms of memory by ensuring the most common value is in the
2269 middle and the least common are on the outside. IMO this would be better
2270 than a most to least common mapping as theres a decent chance the most
2271 common letter will share a node with the least common, meaning the node
2272 will not be compressible. With a middle is most common approach the worst
2273 case is when we have the least common nodes twice.
2277 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2278 regnode *noper = NEXTOPER( cur );
2279 const U8 *uc = (U8*)STRING( noper );
2280 const U8 *e = uc + STR_LEN( noper );
2282 U32 wordlen = 0; /* required init */
2283 STRLEN minchars = 0;
2284 STRLEN maxchars = 0;
2285 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2288 if (OP(noper) == NOTHING) {
2289 regnode *noper_next= regnext(noper);
2290 if (noper_next != tail && OP(noper_next) == flags) {
2292 uc= (U8*)STRING(noper);
2293 e= uc + STR_LEN(noper);
2294 trie->minlen= STR_LEN(noper);
2301 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2302 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2303 regardless of encoding */
2304 if (OP( noper ) == EXACTFU_SS) {
2305 /* false positives are ok, so just set this */
2306 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2309 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2311 TRIE_CHARCOUNT(trie)++;
2314 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2315 * is in effect. Under /i, this character can match itself, or
2316 * anything that folds to it. If not under /i, it can match just
2317 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2318 * all fold to k, and all are single characters. But some folds
2319 * expand to more than one character, so for example LATIN SMALL
2320 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2321 * the string beginning at 'uc' is 'ffi', it could be matched by
2322 * three characters, or just by the one ligature character. (It
2323 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2324 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2325 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2326 * match.) The trie needs to know the minimum and maximum number
2327 * of characters that could match so that it can use size alone to
2328 * quickly reject many match attempts. The max is simple: it is
2329 * the number of folded characters in this branch (since a fold is
2330 * never shorter than what folds to it. */
2334 /* And the min is equal to the max if not under /i (indicated by
2335 * 'folder' being NULL), or there are no multi-character folds. If
2336 * there is a multi-character fold, the min is incremented just
2337 * once, for the character that folds to the sequence. Each
2338 * character in the sequence needs to be added to the list below of
2339 * characters in the trie, but we count only the first towards the
2340 * min number of characters needed. This is done through the
2341 * variable 'foldlen', which is returned by the macros that look
2342 * for these sequences as the number of bytes the sequence
2343 * occupies. Each time through the loop, we decrement 'foldlen' by
2344 * how many bytes the current char occupies. Only when it reaches
2345 * 0 do we increment 'minchars' or look for another multi-character
2347 if (folder == NULL) {
2350 else if (foldlen > 0) {
2351 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2356 /* See if *uc is the beginning of a multi-character fold. If
2357 * so, we decrement the length remaining to look at, to account
2358 * for the current character this iteration. (We can use 'uc'
2359 * instead of the fold returned by TRIE_READ_CHAR because for
2360 * non-UTF, the latin1_safe macro is smart enough to account
2361 * for all the unfolded characters, and because for UTF, the
2362 * string will already have been folded earlier in the
2363 * compilation process */
2365 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2366 foldlen -= UTF8SKIP(uc);
2369 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2374 /* The current character (and any potential folds) should be added
2375 * to the possible matching characters for this position in this
2379 U8 folded= folder[ (U8) uvc ];
2380 if ( !trie->charmap[ folded ] ) {
2381 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2382 TRIE_STORE_REVCHAR( folded );
2385 if ( !trie->charmap[ uvc ] ) {
2386 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2387 TRIE_STORE_REVCHAR( uvc );
2390 /* store the codepoint in the bitmap, and its folded
2392 TRIE_BITMAP_SET(trie, uvc);
2394 /* store the folded codepoint */
2395 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2398 /* store first byte of utf8 representation of
2399 variant codepoints */
2400 if (! UVCHR_IS_INVARIANT(uvc)) {
2401 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2404 set_bit = 0; /* We've done our bit :-) */
2408 /* XXX We could come up with the list of code points that fold
2409 * to this using PL_utf8_foldclosures, except not for
2410 * multi-char folds, as there may be multiple combinations
2411 * there that could work, which needs to wait until runtime to
2412 * resolve (The comment about LIGATURE FFI above is such an
2417 widecharmap = newHV();
2419 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2422 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2424 if ( !SvTRUE( *svpp ) ) {
2425 sv_setiv( *svpp, ++trie->uniquecharcount );
2426 TRIE_STORE_REVCHAR(uvc);
2429 } /* end loop through characters in this branch of the trie */
2431 /* We take the min and max for this branch and combine to find the min
2432 * and max for all branches processed so far */
2433 if( cur == first ) {
2434 trie->minlen = minchars;
2435 trie->maxlen = maxchars;
2436 } else if (minchars < trie->minlen) {
2437 trie->minlen = minchars;
2438 } else if (maxchars > trie->maxlen) {
2439 trie->maxlen = maxchars;
2441 } /* end first pass */
2442 DEBUG_TRIE_COMPILE_r(
2443 PerlIO_printf( Perl_debug_log,
2444 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2445 (int)depth * 2 + 2,"",
2446 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2447 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2448 (int)trie->minlen, (int)trie->maxlen )
2452 We now know what we are dealing with in terms of unique chars and
2453 string sizes so we can calculate how much memory a naive
2454 representation using a flat table will take. If it's over a reasonable
2455 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2456 conservative but potentially much slower representation using an array
2459 At the end we convert both representations into the same compressed
2460 form that will be used in regexec.c for matching with. The latter
2461 is a form that cannot be used to construct with but has memory
2462 properties similar to the list form and access properties similar
2463 to the table form making it both suitable for fast searches and
2464 small enough that its feasable to store for the duration of a program.
2466 See the comment in the code where the compressed table is produced
2467 inplace from the flat tabe representation for an explanation of how
2468 the compression works.
2473 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2476 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2477 > SvIV(re_trie_maxbuff) )
2480 Second Pass -- Array Of Lists Representation
2482 Each state will be represented by a list of charid:state records
2483 (reg_trie_trans_le) the first such element holds the CUR and LEN
2484 points of the allocated array. (See defines above).
2486 We build the initial structure using the lists, and then convert
2487 it into the compressed table form which allows faster lookups
2488 (but cant be modified once converted).
2491 STRLEN transcount = 1;
2493 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2494 "%*sCompiling trie using list compiler\n",
2495 (int)depth * 2 + 2, ""));
2497 trie->states = (reg_trie_state *)
2498 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2499 sizeof(reg_trie_state) );
2503 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2505 regnode *noper = NEXTOPER( cur );
2506 U8 *uc = (U8*)STRING( noper );
2507 const U8 *e = uc + STR_LEN( noper );
2508 U32 state = 1; /* required init */
2509 U16 charid = 0; /* sanity init */
2510 U32 wordlen = 0; /* required init */
2512 if (OP(noper) == NOTHING) {
2513 regnode *noper_next= regnext(noper);
2514 if (noper_next != tail && OP(noper_next) == flags) {
2516 uc= (U8*)STRING(noper);
2517 e= uc + STR_LEN(noper);
2521 if (OP(noper) != NOTHING) {
2522 for ( ; uc < e ; uc += len ) {
2527 charid = trie->charmap[ uvc ];
2529 SV** const svpp = hv_fetch( widecharmap,
2536 charid=(U16)SvIV( *svpp );
2539 /* charid is now 0 if we dont know the char read, or
2540 * nonzero if we do */
2547 if ( !trie->states[ state ].trans.list ) {
2548 TRIE_LIST_NEW( state );
2551 check <= TRIE_LIST_USED( state );
2554 if ( TRIE_LIST_ITEM( state, check ).forid
2557 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2562 newstate = next_alloc++;
2563 prev_states[newstate] = state;
2564 TRIE_LIST_PUSH( state, charid, newstate );
2569 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2573 TRIE_HANDLE_WORD(state);
2575 } /* end second pass */
2577 /* next alloc is the NEXT state to be allocated */
2578 trie->statecount = next_alloc;
2579 trie->states = (reg_trie_state *)
2580 PerlMemShared_realloc( trie->states,
2582 * sizeof(reg_trie_state) );
2584 /* and now dump it out before we compress it */
2585 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2586 revcharmap, next_alloc,
2590 trie->trans = (reg_trie_trans *)
2591 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2598 for( state=1 ; state < next_alloc ; state ++ ) {
2602 DEBUG_TRIE_COMPILE_MORE_r(
2603 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2607 if (trie->states[state].trans.list) {
2608 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2612 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2613 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2614 if ( forid < minid ) {
2616 } else if ( forid > maxid ) {
2620 if ( transcount < tp + maxid - minid + 1) {
2622 trie->trans = (reg_trie_trans *)
2623 PerlMemShared_realloc( trie->trans,
2625 * sizeof(reg_trie_trans) );
2626 Zero( trie->trans + (transcount / 2),
2630 base = trie->uniquecharcount + tp - minid;
2631 if ( maxid == minid ) {
2633 for ( ; zp < tp ; zp++ ) {
2634 if ( ! trie->trans[ zp ].next ) {
2635 base = trie->uniquecharcount + zp - minid;
2636 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2638 trie->trans[ zp ].check = state;
2644 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2646 trie->trans[ tp ].check = state;
2651 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2652 const U32 tid = base
2653 - trie->uniquecharcount
2654 + TRIE_LIST_ITEM( state, idx ).forid;
2655 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2657 trie->trans[ tid ].check = state;
2659 tp += ( maxid - minid + 1 );
2661 Safefree(trie->states[ state ].trans.list);
2664 DEBUG_TRIE_COMPILE_MORE_r(
2665 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2668 trie->states[ state ].trans.base=base;
2670 trie->lasttrans = tp + 1;
2674 Second Pass -- Flat Table Representation.
2676 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2677 each. We know that we will need Charcount+1 trans at most to store
2678 the data (one row per char at worst case) So we preallocate both
2679 structures assuming worst case.
2681 We then construct the trie using only the .next slots of the entry
2684 We use the .check field of the first entry of the node temporarily
2685 to make compression both faster and easier by keeping track of how
2686 many non zero fields are in the node.
2688 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2691 There are two terms at use here: state as a TRIE_NODEIDX() which is
2692 a number representing the first entry of the node, and state as a
2693 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2694 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2695 if there are 2 entrys per node. eg:
2703 The table is internally in the right hand, idx form. However as we
2704 also have to deal with the states array which is indexed by nodenum
2705 we have to use TRIE_NODENUM() to convert.
2708 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2709 "%*sCompiling trie using table compiler\n",
2710 (int)depth * 2 + 2, ""));
2712 trie->trans = (reg_trie_trans *)
2713 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2714 * trie->uniquecharcount + 1,
2715 sizeof(reg_trie_trans) );
2716 trie->states = (reg_trie_state *)
2717 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2718 sizeof(reg_trie_state) );
2719 next_alloc = trie->uniquecharcount + 1;
2722 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2724 regnode *noper = NEXTOPER( cur );
2725 const U8 *uc = (U8*)STRING( noper );
2726 const U8 *e = uc + STR_LEN( noper );
2728 U32 state = 1; /* required init */
2730 U16 charid = 0; /* sanity init */
2731 U32 accept_state = 0; /* sanity init */
2733 U32 wordlen = 0; /* required init */
2735 if (OP(noper) == NOTHING) {
2736 regnode *noper_next= regnext(noper);
2737 if (noper_next != tail && OP(noper_next) == flags) {
2739 uc= (U8*)STRING(noper);
2740 e= uc + STR_LEN(noper);
2744 if ( OP(noper) != NOTHING ) {
2745 for ( ; uc < e ; uc += len ) {
2750 charid = trie->charmap[ uvc ];
2752 SV* const * const svpp = hv_fetch( widecharmap,
2756 charid = svpp ? (U16)SvIV(*svpp) : 0;
2760 if ( !trie->trans[ state + charid ].next ) {
2761 trie->trans[ state + charid ].next = next_alloc;
2762 trie->trans[ state ].check++;
2763 prev_states[TRIE_NODENUM(next_alloc)]
2764 = TRIE_NODENUM(state);
2765 next_alloc += trie->uniquecharcount;
2767 state = trie->trans[ state + charid ].next;
2769 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2771 /* charid is now 0 if we dont know the char read, or
2772 * nonzero if we do */
2775 accept_state = TRIE_NODENUM( state );
2776 TRIE_HANDLE_WORD(accept_state);
2778 } /* end second pass */
2780 /* and now dump it out before we compress it */
2781 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2783 next_alloc, depth+1));
2787 * Inplace compress the table.*
2789 For sparse data sets the table constructed by the trie algorithm will
2790 be mostly 0/FAIL transitions or to put it another way mostly empty.
2791 (Note that leaf nodes will not contain any transitions.)
2793 This algorithm compresses the tables by eliminating most such
2794 transitions, at the cost of a modest bit of extra work during lookup:
2796 - Each states[] entry contains a .base field which indicates the
2797 index in the state[] array wheres its transition data is stored.
2799 - If .base is 0 there are no valid transitions from that node.
2801 - If .base is nonzero then charid is added to it to find an entry in
2804 -If trans[states[state].base+charid].check!=state then the
2805 transition is taken to be a 0/Fail transition. Thus if there are fail
2806 transitions at the front of the node then the .base offset will point
2807 somewhere inside the previous nodes data (or maybe even into a node
2808 even earlier), but the .check field determines if the transition is
2812 The following process inplace converts the table to the compressed
2813 table: We first do not compress the root node 1,and mark all its
2814 .check pointers as 1 and set its .base pointer as 1 as well. This
2815 allows us to do a DFA construction from the compressed table later,
2816 and ensures that any .base pointers we calculate later are greater
2819 - We set 'pos' to indicate the first entry of the second node.
2821 - We then iterate over the columns of the node, finding the first and
2822 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2823 and set the .check pointers accordingly, and advance pos
2824 appropriately and repreat for the next node. Note that when we copy
2825 the next pointers we have to convert them from the original
2826 NODEIDX form to NODENUM form as the former is not valid post
2829 - If a node has no transitions used we mark its base as 0 and do not
2830 advance the pos pointer.
2832 - If a node only has one transition we use a second pointer into the
2833 structure to fill in allocated fail transitions from other states.
2834 This pointer is independent of the main pointer and scans forward
2835 looking for null transitions that are allocated to a state. When it
2836 finds one it writes the single transition into the "hole". If the
2837 pointer doesnt find one the single transition is appended as normal.
2839 - Once compressed we can Renew/realloc the structures to release the
2842 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2843 specifically Fig 3.47 and the associated pseudocode.
2847 const U32 laststate = TRIE_NODENUM( next_alloc );
2850 trie->statecount = laststate;
2852 for ( state = 1 ; state < laststate ; state++ ) {
2854 const U32 stateidx = TRIE_NODEIDX( state );
2855 const U32 o_used = trie->trans[ stateidx ].check;
2856 U32 used = trie->trans[ stateidx ].check;
2857 trie->trans[ stateidx ].check = 0;
2860 used && charid < trie->uniquecharcount;
2863 if ( flag || trie->trans[ stateidx + charid ].next ) {
2864 if ( trie->trans[ stateidx + charid ].next ) {
2866 for ( ; zp < pos ; zp++ ) {
2867 if ( ! trie->trans[ zp ].next ) {
2871 trie->states[ state ].trans.base
2873 + trie->uniquecharcount
2875 trie->trans[ zp ].next
2876 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2878 trie->trans[ zp ].check = state;
2879 if ( ++zp > pos ) pos = zp;
2886 trie->states[ state ].trans.base
2887 = pos + trie->uniquecharcount - charid ;
2889 trie->trans[ pos ].next
2890 = SAFE_TRIE_NODENUM(
2891 trie->trans[ stateidx + charid ].next );
2892 trie->trans[ pos ].check = state;
2897 trie->lasttrans = pos + 1;
2898 trie->states = (reg_trie_state *)
2899 PerlMemShared_realloc( trie->states, laststate
2900 * sizeof(reg_trie_state) );
2901 DEBUG_TRIE_COMPILE_MORE_r(
2902 PerlIO_printf( Perl_debug_log,
2903 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2904 (int)depth * 2 + 2,"",
2905 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2909 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2912 } /* end table compress */
2914 DEBUG_TRIE_COMPILE_MORE_r(
2915 PerlIO_printf(Perl_debug_log,
2916 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2917 (int)depth * 2 + 2, "",
2918 (UV)trie->statecount,
2919 (UV)trie->lasttrans)
2921 /* resize the trans array to remove unused space */
2922 trie->trans = (reg_trie_trans *)
2923 PerlMemShared_realloc( trie->trans, trie->lasttrans
2924 * sizeof(reg_trie_trans) );
2926 { /* Modify the program and insert the new TRIE node */
2927 U8 nodetype =(U8)(flags & 0xFF);
2931 regnode *optimize = NULL;
2932 #ifdef RE_TRACK_PATTERN_OFFSETS
2935 U32 mjd_nodelen = 0;
2936 #endif /* RE_TRACK_PATTERN_OFFSETS */
2937 #endif /* DEBUGGING */
2939 This means we convert either the first branch or the first Exact,
2940 depending on whether the thing following (in 'last') is a branch
2941 or not and whther first is the startbranch (ie is it a sub part of
2942 the alternation or is it the whole thing.)
2943 Assuming its a sub part we convert the EXACT otherwise we convert
2944 the whole branch sequence, including the first.
2946 /* Find the node we are going to overwrite */
2947 if ( first != startbranch || OP( last ) == BRANCH ) {
2948 /* branch sub-chain */
2949 NEXT_OFF( first ) = (U16)(last - first);
2950 #ifdef RE_TRACK_PATTERN_OFFSETS
2952 mjd_offset= Node_Offset((convert));
2953 mjd_nodelen= Node_Length((convert));
2956 /* whole branch chain */
2958 #ifdef RE_TRACK_PATTERN_OFFSETS
2961 const regnode *nop = NEXTOPER( convert );
2962 mjd_offset= Node_Offset((nop));
2963 mjd_nodelen= Node_Length((nop));
2967 PerlIO_printf(Perl_debug_log,
2968 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2969 (int)depth * 2 + 2, "",
2970 (UV)mjd_offset, (UV)mjd_nodelen)
2973 /* But first we check to see if there is a common prefix we can
2974 split out as an EXACT and put in front of the TRIE node. */
2975 trie->startstate= 1;
2976 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2978 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2982 const U32 base = trie->states[ state ].trans.base;
2984 if ( trie->states[state].wordnum )
2987 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2988 if ( ( base + ofs >= trie->uniquecharcount ) &&
2989 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2990 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2992 if ( ++count > 1 ) {
2993 SV **tmp = av_fetch( revcharmap, ofs, 0);
2994 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2995 if ( state == 1 ) break;
2997 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2999 PerlIO_printf(Perl_debug_log,
3000 "%*sNew Start State=%"UVuf" Class: [",
3001 (int)depth * 2 + 2, "",
3004 SV ** const tmp = av_fetch( revcharmap, idx, 0);
3005 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
3007 TRIE_BITMAP_SET(trie,*ch);
3009 TRIE_BITMAP_SET(trie, folder[ *ch ]);
3011 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
3015 TRIE_BITMAP_SET(trie,*ch);
3017 TRIE_BITMAP_SET(trie,folder[ *ch ]);
3018 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
3024 SV **tmp = av_fetch( revcharmap, idx, 0);
3026 char *ch = SvPV( *tmp, len );
3028 SV *sv=sv_newmortal();
3029 PerlIO_printf( Perl_debug_log,
3030 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
3031 (int)depth * 2 + 2, "",
3033 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
3034 PL_colors[0], PL_colors[1],
3035 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
3036 PERL_PV_ESCAPE_FIRSTCHAR
3041 OP( convert ) = nodetype;
3042 str=STRING(convert);
3045 STR_LEN(convert) += len;
3051 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3056 trie->prefixlen = (state-1);
3058 regnode *n = convert+NODE_SZ_STR(convert);
3059 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3060 trie->startstate = state;
3061 trie->minlen -= (state - 1);
3062 trie->maxlen -= (state - 1);
3064 /* At least the UNICOS C compiler choked on this
3065 * being argument to DEBUG_r(), so let's just have
3068 #ifdef PERL_EXT_RE_BUILD
3074 regnode *fix = convert;
3075 U32 word = trie->wordcount;
3077 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3078 while( ++fix < n ) {
3079 Set_Node_Offset_Length(fix, 0, 0);
3082 SV ** const tmp = av_fetch( trie_words, word, 0 );
3084 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3085 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3087 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3095 NEXT_OFF(convert) = (U16)(tail - convert);
3096 DEBUG_r(optimize= n);
3102 if ( trie->maxlen ) {
3103 NEXT_OFF( convert ) = (U16)(tail - convert);
3104 ARG_SET( convert, data_slot );
3105 /* Store the offset to the first unabsorbed branch in
3106 jump[0], which is otherwise unused by the jump logic.
3107 We use this when dumping a trie and during optimisation. */
3109 trie->jump[0] = (U16)(nextbranch - convert);
3111 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3112 * and there is a bitmap
3113 * and the first "jump target" node we found leaves enough room
3114 * then convert the TRIE node into a TRIEC node, with the bitmap
3115 * embedded inline in the opcode - this is hypothetically faster.
3117 if ( !trie->states[trie->startstate].wordnum
3119 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3121 OP( convert ) = TRIEC;
3122 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3123 PerlMemShared_free(trie->bitmap);
3126 OP( convert ) = TRIE;
3128 /* store the type in the flags */
3129 convert->flags = nodetype;
3133 + regarglen[ OP( convert ) ];
3135 /* XXX We really should free up the resource in trie now,
3136 as we won't use them - (which resources?) dmq */
3138 /* needed for dumping*/
3139 DEBUG_r(if (optimize) {
3140 regnode *opt = convert;
3142 while ( ++opt < optimize) {
3143 Set_Node_Offset_Length(opt,0,0);
3146 Try to clean up some of the debris left after the
3149 while( optimize < jumper ) {
3150 mjd_nodelen += Node_Length((optimize));
3151 OP( optimize ) = OPTIMIZED;
3152 Set_Node_Offset_Length(optimize,0,0);
3155 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3157 } /* end node insert */
3159 /* Finish populating the prev field of the wordinfo array. Walk back
3160 * from each accept state until we find another accept state, and if
3161 * so, point the first word's .prev field at the second word. If the
3162 * second already has a .prev field set, stop now. This will be the
3163 * case either if we've already processed that word's accept state,
3164 * or that state had multiple words, and the overspill words were
3165 * already linked up earlier.
3172 for (word=1; word <= trie->wordcount; word++) {
3174 if (trie->wordinfo[word].prev)
3176 state = trie->wordinfo[word].accept;
3178 state = prev_states[state];
3181 prev = trie->states[state].wordnum;
3185 trie->wordinfo[word].prev = prev;
3187 Safefree(prev_states);
3191 /* and now dump out the compressed format */
3192 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3194 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3196 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3197 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3199 SvREFCNT_dec_NN(revcharmap);
3203 : trie->startstate>1
3209 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3211 /* The Trie is constructed and compressed now so we can build a fail array if
3214 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3216 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3220 We find the fail state for each state in the trie, this state is the longest
3221 proper suffix of the current state's 'word' that is also a proper prefix of
3222 another word in our trie. State 1 represents the word '' and is thus the
3223 default fail state. This allows the DFA not to have to restart after its
3224 tried and failed a word at a given point, it simply continues as though it
3225 had been matching the other word in the first place.
3227 'abcdgu'=~/abcdefg|cdgu/
3228 When we get to 'd' we are still matching the first word, we would encounter
3229 'g' which would fail, which would bring us to the state representing 'd' in
3230 the second word where we would try 'g' and succeed, proceeding to match
3233 /* add a fail transition */
3234 const U32 trie_offset = ARG(source);
3235 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3237 const U32 ucharcount = trie->uniquecharcount;
3238 const U32 numstates = trie->statecount;
3239 const U32 ubound = trie->lasttrans + ucharcount;
3243 U32 base = trie->states[ 1 ].trans.base;
3246 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3248 GET_RE_DEBUG_FLAGS_DECL;
3250 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3251 PERL_UNUSED_CONTEXT;
3253 PERL_UNUSED_ARG(depth);
3256 if ( OP(source) == TRIE ) {
3257 struct regnode_1 *op = (struct regnode_1 *)
3258 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3259 StructCopy(source,op,struct regnode_1);
3260 stclass = (regnode *)op;
3262 struct regnode_charclass *op = (struct regnode_charclass *)
3263 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3264 StructCopy(source,op,struct regnode_charclass);
3265 stclass = (regnode *)op;
3267 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3269 ARG_SET( stclass, data_slot );
3270 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3271 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3272 aho->trie=trie_offset;
3273 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3274 Copy( trie->states, aho->states, numstates, reg_trie_state );
3275 Newxz( q, numstates, U32);
3276 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3279 /* initialize fail[0..1] to be 1 so that we always have
3280 a valid final fail state */
3281 fail[ 0 ] = fail[ 1 ] = 1;
3283 for ( charid = 0; charid < ucharcount ; charid++ ) {
3284 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3286 q[ q_write ] = newstate;
3287 /* set to point at the root */
3288 fail[ q[ q_write++ ] ]=1;
3291 while ( q_read < q_write) {
3292 const U32 cur = q[ q_read++ % numstates ];
3293 base = trie->states[ cur ].trans.base;
3295 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3296 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3298 U32 fail_state = cur;
3301 fail_state = fail[ fail_state ];
3302 fail_base = aho->states[ fail_state ].trans.base;
3303 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3305 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3306 fail[ ch_state ] = fail_state;
3307 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3309 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3311 q[ q_write++ % numstates] = ch_state;
3315 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3316 when we fail in state 1, this allows us to use the
3317 charclass scan to find a valid start char. This is based on the principle
3318 that theres a good chance the string being searched contains lots of stuff
3319 that cant be a start char.
3321 fail[ 0 ] = fail[ 1 ] = 0;
3322 DEBUG_TRIE_COMPILE_r({
3323 PerlIO_printf(Perl_debug_log,
3324 "%*sStclass Failtable (%"UVuf" states): 0",
3325 (int)(depth * 2), "", (UV)numstates
3327 for( q_read=1; q_read<numstates; q_read++ ) {
3328 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3330 PerlIO_printf(Perl_debug_log, "\n");
3333 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3338 #define DEBUG_PEEP(str,scan,depth) \
3339 DEBUG_OPTIMISE_r({if (scan){ \
3340 regnode *Next = regnext(scan); \
3341 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3342 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3343 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3344 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3345 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3346 PerlIO_printf(Perl_debug_log, "\n"); \
3349 /* The below joins as many adjacent EXACTish nodes as possible into a single
3350 * one. The regop may be changed if the node(s) contain certain sequences that
3351 * require special handling. The joining is only done if:
3352 * 1) there is room in the current conglomerated node to entirely contain the
3354 * 2) they are the exact same node type
3356 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3357 * these get optimized out
3359 * If a node is to match under /i (folded), the number of characters it matches
3360 * can be different than its character length if it contains a multi-character
3361 * fold. *min_subtract is set to the total delta number of characters of the
3364 * And *unfolded_multi_char is set to indicate whether or not the node contains
3365 * an unfolded multi-char fold. This happens when whether the fold is valid or
3366 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3367 * SMALL LETTER SHARP S, as only if the target string being matched against
3368 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3369 * folding rules depend on the locale in force at runtime. (Multi-char folds
3370 * whose components are all above the Latin1 range are not run-time locale
3371 * dependent, and have already been folded by the time this function is
3374 * This is as good a place as any to discuss the design of handling these
3375 * multi-character fold sequences. It's been wrong in Perl for a very long
3376 * time. There are three code points in Unicode whose multi-character folds
3377 * were long ago discovered to mess things up. The previous designs for
3378 * dealing with these involved assigning a special node for them. This
3379 * approach doesn't always work, as evidenced by this example:
3380 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3381 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3382 * would match just the \xDF, it won't be able to handle the case where a
3383 * successful match would have to cross the node's boundary. The new approach
3384 * that hopefully generally solves the problem generates an EXACTFU_SS node
3385 * that is "sss" in this case.
3387 * It turns out that there are problems with all multi-character folds, and not
3388 * just these three. Now the code is general, for all such cases. The
3389 * approach taken is:
3390 * 1) This routine examines each EXACTFish node that could contain multi-
3391 * character folded sequences. Since a single character can fold into
3392 * such a sequence, the minimum match length for this node is less than
3393 * the number of characters in the node. This routine returns in
3394 * *min_subtract how many characters to subtract from the the actual
3395 * length of the string to get a real minimum match length; it is 0 if
3396 * there are no multi-char foldeds. This delta is used by the caller to
3397 * adjust the min length of the match, and the delta between min and max,
3398 * so that the optimizer doesn't reject these possibilities based on size
3400 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3401 * is used for an EXACTFU node that contains at least one "ss" sequence in
3402 * it. For non-UTF-8 patterns and strings, this is the only case where
3403 * there is a possible fold length change. That means that a regular
3404 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3405 * with length changes, and so can be processed faster. regexec.c takes
3406 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3407 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3408 * known until runtime). This saves effort in regex matching. However,
3409 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3410 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3411 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3412 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3413 * possibilities for the non-UTF8 patterns are quite simple, except for
3414 * the sharp s. All the ones that don't involve a UTF-8 target string are
3415 * members of a fold-pair, and arrays are set up for all of them so that
3416 * the other member of the pair can be found quickly. Code elsewhere in
3417 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3418 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3419 * described in the next item.
3420 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3421 * validity of the fold won't be known until runtime, and so must remain
3422 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3423 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3424 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3425 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3426 * The reason this is a problem is that the optimizer part of regexec.c
3427 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3428 * that a character in the pattern corresponds to at most a single
3429 * character in the target string. (And I do mean character, and not byte
3430 * here, unlike other parts of the documentation that have never been
3431 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3432 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3433 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3434 * nodes, violate the assumption, and they are the only instances where it
3435 * is violated. I'm reluctant to try to change the assumption, as the
3436 * code involved is impenetrable to me (khw), so instead the code here
3437 * punts. This routine examines EXACTFL nodes, and (when the pattern
3438 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3439 * boolean indicating whether or not the node contains such a fold. When
3440 * it is true, the caller sets a flag that later causes the optimizer in
3441 * this file to not set values for the floating and fixed string lengths,
3442 * and thus avoids the optimizer code in regexec.c that makes the invalid
3443 * assumption. Thus, there is no optimization based on string lengths for
3444 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3445 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3446 * assumption is wrong only in these cases is that all other non-UTF-8
3447 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3448 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3449 * EXACTF nodes because we don't know at compile time if it actually
3450 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3451 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3452 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3453 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3454 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3455 * string would require the pattern to be forced into UTF-8, the overhead
3456 * of which we want to avoid. Similarly the unfolded multi-char folds in
3457 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3460 * Similarly, the code that generates tries doesn't currently handle
3461 * not-already-folded multi-char folds, and it looks like a pain to change
3462 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3463 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3464 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3465 * using /iaa matching will be doing so almost entirely with ASCII
3466 * strings, so this should rarely be encountered in practice */
3468 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3469 if (PL_regkind[OP(scan)] == EXACT) \
3470 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3473 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3474 UV *min_subtract, bool *unfolded_multi_char,
3475 U32 flags,regnode *val, U32 depth)
3477 /* Merge several consecutive EXACTish nodes into one. */
3478 regnode *n = regnext(scan);
3480 regnode *next = scan + NODE_SZ_STR(scan);
3484 regnode *stop = scan;
3485 GET_RE_DEBUG_FLAGS_DECL;
3487 PERL_UNUSED_ARG(depth);
3490 PERL_ARGS_ASSERT_JOIN_EXACT;
3491 #ifndef EXPERIMENTAL_INPLACESCAN
3492 PERL_UNUSED_ARG(flags);
3493 PERL_UNUSED_ARG(val);
3495 DEBUG_PEEP("join",scan,depth);
3497 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3498 * EXACT ones that are mergeable to the current one. */
3500 && (PL_regkind[OP(n)] == NOTHING
3501 || (stringok && OP(n) == OP(scan)))
3503 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3506 if (OP(n) == TAIL || n > next)
3508 if (PL_regkind[OP(n)] == NOTHING) {
3509 DEBUG_PEEP("skip:",n,depth);
3510 NEXT_OFF(scan) += NEXT_OFF(n);
3511 next = n + NODE_STEP_REGNODE;
3518 else if (stringok) {
3519 const unsigned int oldl = STR_LEN(scan);
3520 regnode * const nnext = regnext(n);
3522 /* XXX I (khw) kind of doubt that this works on platforms (should
3523 * Perl ever run on one) where U8_MAX is above 255 because of lots
3524 * of other assumptions */
3525 /* Don't join if the sum can't fit into a single node */
3526 if (oldl + STR_LEN(n) > U8_MAX)
3529 DEBUG_PEEP("merg",n,depth);
3532 NEXT_OFF(scan) += NEXT_OFF(n);
3533 STR_LEN(scan) += STR_LEN(n);
3534 next = n + NODE_SZ_STR(n);
3535 /* Now we can overwrite *n : */
3536 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3544 #ifdef EXPERIMENTAL_INPLACESCAN
3545 if (flags && !NEXT_OFF(n)) {
3546 DEBUG_PEEP("atch", val, depth);
3547 if (reg_off_by_arg[OP(n)]) {
3548 ARG_SET(n, val - n);
3551 NEXT_OFF(n) = val - n;
3559 *unfolded_multi_char = FALSE;
3561 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3562 * can now analyze for sequences of problematic code points. (Prior to
3563 * this final joining, sequences could have been split over boundaries, and
3564 * hence missed). The sequences only happen in folding, hence for any
3565 * non-EXACT EXACTish node */
3566 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3567 U8* s0 = (U8*) STRING(scan);
3569 U8* s_end = s0 + STR_LEN(scan);
3571 int total_count_delta = 0; /* Total delta number of characters that
3572 multi-char folds expand to */
3574 /* One pass is made over the node's string looking for all the
3575 * possibilities. To avoid some tests in the loop, there are two main
3576 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3581 if (OP(scan) == EXACTFL) {
3584 /* An EXACTFL node would already have been changed to another
3585 * node type unless there is at least one character in it that
3586 * is problematic; likely a character whose fold definition
3587 * won't be known until runtime, and so has yet to be folded.
3588 * For all but the UTF-8 locale, folds are 1-1 in length, but
3589 * to handle the UTF-8 case, we need to create a temporary
3590 * folded copy using UTF-8 locale rules in order to analyze it.
3591 * This is because our macros that look to see if a sequence is
3592 * a multi-char fold assume everything is folded (otherwise the
3593 * tests in those macros would be too complicated and slow).
3594 * Note that here, the non-problematic folds will have already
3595 * been done, so we can just copy such characters. We actually
3596 * don't completely fold the EXACTFL string. We skip the
3597 * unfolded multi-char folds, as that would just create work
3598 * below to figure out the size they already are */
3600 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3603 STRLEN s_len = UTF8SKIP(s);
3604 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3605 Copy(s, d, s_len, U8);
3608 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3609 *unfolded_multi_char = TRUE;
3610 Copy(s, d, s_len, U8);
3613 else if (isASCII(*s)) {
3614 *(d++) = toFOLD(*s);
3618 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3624 /* Point the remainder of the routine to look at our temporary
3628 } /* End of creating folded copy of EXACTFL string */
3630 /* Examine the string for a multi-character fold sequence. UTF-8
3631 * patterns have all characters pre-folded by the time this code is
3633 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3634 length sequence we are looking for is 2 */
3636 int count = 0; /* How many characters in a multi-char fold */
3637 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3638 if (! len) { /* Not a multi-char fold: get next char */
3643 /* Nodes with 'ss' require special handling, except for
3644 * EXACTFA-ish for which there is no multi-char fold to this */
3645 if (len == 2 && *s == 's' && *(s+1) == 's'
3646 && OP(scan) != EXACTFA
3647 && OP(scan) != EXACTFA_NO_TRIE)
3650 if (OP(scan) != EXACTFL) {
3651 OP(scan) = EXACTFU_SS;
3655 else { /* Here is a generic multi-char fold. */
3656 U8* multi_end = s + len;
3658 /* Count how many characters are in it. In the case of
3659 * /aa, no folds which contain ASCII code points are
3660 * allowed, so check for those, and skip if found. */
3661 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3662 count = utf8_length(s, multi_end);
3666 while (s < multi_end) {
3669 goto next_iteration;
3679 /* The delta is how long the sequence is minus 1 (1 is how long
3680 * the character that folds to the sequence is) */
3681 total_count_delta += count - 1;
3685 /* We created a temporary folded copy of the string in EXACTFL
3686 * nodes. Therefore we need to be sure it doesn't go below zero,
3687 * as the real string could be shorter */
3688 if (OP(scan) == EXACTFL) {
3689 int total_chars = utf8_length((U8*) STRING(scan),
3690 (U8*) STRING(scan) + STR_LEN(scan));
3691 if (total_count_delta > total_chars) {
3692 total_count_delta = total_chars;
3696 *min_subtract += total_count_delta;
3699 else if (OP(scan) == EXACTFA) {
3701 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3702 * fold to the ASCII range (and there are no existing ones in the
3703 * upper latin1 range). But, as outlined in the comments preceding
3704 * this function, we need to flag any occurrences of the sharp s.
3705 * This character forbids trie formation (because of added
3707 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
3708 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
3709 || UNICODE_DOT_DOT_VERSION > 0)
3711 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3712 OP(scan) = EXACTFA_NO_TRIE;
3713 *unfolded_multi_char = TRUE;
3721 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3722 * folds that are all Latin1. As explained in the comments
3723 * preceding this function, we look also for the sharp s in EXACTF
3724 * and EXACTFL nodes; it can be in the final position. Otherwise
3725 * we can stop looking 1 byte earlier because have to find at least
3726 * two characters for a multi-fold */
3727 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3732 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3733 if (! len) { /* Not a multi-char fold. */
3734 if (*s == LATIN_SMALL_LETTER_SHARP_S
3735 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3737 *unfolded_multi_char = TRUE;
3744 && isALPHA_FOLD_EQ(*s, 's')
3745 && isALPHA_FOLD_EQ(*(s+1), 's'))
3748 /* EXACTF nodes need to know that the minimum length
3749 * changed so that a sharp s in the string can match this
3750 * ss in the pattern, but they remain EXACTF nodes, as they
3751 * won't match this unless the target string is is UTF-8,
3752 * which we don't know until runtime. EXACTFL nodes can't
3753 * transform into EXACTFU nodes */
3754 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3755 OP(scan) = EXACTFU_SS;
3759 *min_subtract += len - 1;
3767 /* Allow dumping but overwriting the collection of skipped
3768 * ops and/or strings with fake optimized ops */
3769 n = scan + NODE_SZ_STR(scan);
3777 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3781 /* REx optimizer. Converts nodes into quicker variants "in place".
3782 Finds fixed substrings. */
3784 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3785 to the position after last scanned or to NULL. */
3787 #define INIT_AND_WITHP \
3788 assert(!and_withp); \
3789 Newx(and_withp,1, regnode_ssc); \
3790 SAVEFREEPV(and_withp)
3794 S_unwind_scan_frames(pTHX_ const void *p)
3796 scan_frame *f= (scan_frame *)p;
3798 scan_frame *n= f->next_frame;
3806 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3807 SSize_t *minlenp, SSize_t *deltap,
3812 regnode_ssc *and_withp,
3813 U32 flags, U32 depth)
3814 /* scanp: Start here (read-write). */
3815 /* deltap: Write maxlen-minlen here. */
3816 /* last: Stop before this one. */
3817 /* data: string data about the pattern */
3818 /* stopparen: treat close N as END */
3819 /* recursed: which subroutines have we recursed into */
3820 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3822 /* There must be at least this number of characters to match */
3825 regnode *scan = *scanp, *next;
3827 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3828 int is_inf_internal = 0; /* The studied chunk is infinite */
3829 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3830 scan_data_t data_fake;
3831 SV *re_trie_maxbuff = NULL;
3832 regnode *first_non_open = scan;
3833 SSize_t stopmin = SSize_t_MAX;
3834 scan_frame *frame = NULL;
3835 GET_RE_DEBUG_FLAGS_DECL;
3837 PERL_ARGS_ASSERT_STUDY_CHUNK;
3841 while (first_non_open && OP(first_non_open) == OPEN)
3842 first_non_open=regnext(first_non_open);
3848 RExC_study_chunk_recursed_count++;
3850 DEBUG_OPTIMISE_MORE_r(
3852 PerlIO_printf(Perl_debug_log,
3853 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3854 (int)(depth*2), "", (long)stopparen,
3855 (unsigned long)RExC_study_chunk_recursed_count,
3856 (unsigned long)depth, (unsigned long)recursed_depth,
3859 if (recursed_depth) {
3862 for ( j = 0 ; j < recursed_depth ; j++ ) {
3863 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3865 PAREN_TEST(RExC_study_chunk_recursed +
3866 ( j * RExC_study_chunk_recursed_bytes), i )
3869 !PAREN_TEST(RExC_study_chunk_recursed +
3870 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3873 PerlIO_printf(Perl_debug_log," %d",(int)i);
3877 if ( j + 1 < recursed_depth ) {
3878 PerlIO_printf(Perl_debug_log, ",");
3882 PerlIO_printf(Perl_debug_log,"\n");
3885 while ( scan && OP(scan) != END && scan < last ){
3886 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3887 node length to get a real minimum (because
3888 the folded version may be shorter) */
3889 bool unfolded_multi_char = FALSE;
3890 /* Peephole optimizer: */
3891 DEBUG_STUDYDATA("Peep:", data, depth);
3892 DEBUG_PEEP("Peep", scan, depth);
3895 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3896 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3897 * by a different invocation of reg() -- Yves
3899 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3901 /* Follow the next-chain of the current node and optimize
3902 away all the NOTHINGs from it. */
3903 if (OP(scan) != CURLYX) {
3904 const int max = (reg_off_by_arg[OP(scan)]
3906 /* I32 may be smaller than U16 on CRAYs! */
3907 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3908 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3912 /* Skip NOTHING and LONGJMP. */
3913 while ((n = regnext(n))
3914 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3915 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3916 && off + noff < max)
3918 if (reg_off_by_arg[OP(scan)])
3921 NEXT_OFF(scan) = off;
3924 /* The principal pseudo-switch. Cannot be a switch, since we
3925 look into several different things. */
3926 if ( OP(scan) == DEFINEP ) {
3928 SSize_t deltanext = 0;
3929 SSize_t fake_last_close = 0;
3930 I32 f = SCF_IN_DEFINE;
3932 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3933 scan = regnext(scan);
3934 assert( OP(scan) == IFTHEN );
3935 DEBUG_PEEP("expect IFTHEN", scan, depth);
3937 data_fake.last_closep= &fake_last_close;
3939 next = regnext(scan);
3940 scan = NEXTOPER(NEXTOPER(scan));
3941 DEBUG_PEEP("scan", scan, depth);
3942 DEBUG_PEEP("next", next, depth);
3944 /* we suppose the run is continuous, last=next...
3945 * NOTE we dont use the return here! */
3946 (void)study_chunk(pRExC_state, &scan, &minlen,
3947 &deltanext, next, &data_fake, stopparen,
3948 recursed_depth, NULL, f, depth+1);
3953 OP(scan) == BRANCH ||
3954 OP(scan) == BRANCHJ ||
3957 next = regnext(scan);
3960 /* The op(next)==code check below is to see if we
3961 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3962 * IFTHEN is special as it might not appear in pairs.
3963 * Not sure whether BRANCH-BRANCHJ is possible, regardless
3964 * we dont handle it cleanly. */
3965 if (OP(next) == code || code == IFTHEN) {
3966 /* NOTE - There is similar code to this block below for
3967 * handling TRIE nodes on a re-study. If you change stuff here
3968 * check there too. */
3969 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3971 regnode * const startbranch=scan;
3973 if (flags & SCF_DO_SUBSTR) {
3974 /* Cannot merge strings after this. */
3975 scan_commit(pRExC_state, data, minlenp, is_inf);
3978 if (flags & SCF_DO_STCLASS)
3979 ssc_init_zero(pRExC_state, &accum);
3981 while (OP(scan) == code) {
3982 SSize_t deltanext, minnext, fake;
3984 regnode_ssc this_class;
3986 DEBUG_PEEP("Branch", scan, depth);
3989 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3991 data_fake.whilem_c = data->whilem_c;
3992 data_fake.last_closep = data->last_closep;
3995 data_fake.last_closep = &fake;
3997 data_fake.pos_delta = delta;
3998 next = regnext(scan);
4000 scan = NEXTOPER(scan); /* everything */
4001 if (code != BRANCH) /* everything but BRANCH */
4002 scan = NEXTOPER(scan);
4004 if (flags & SCF_DO_STCLASS) {
4005 ssc_init(pRExC_state, &this_class);
4006 data_fake.start_class = &this_class;
4007 f = SCF_DO_STCLASS_AND;
4009 if (flags & SCF_WHILEM_VISITED_POS)
4010 f |= SCF_WHILEM_VISITED_POS;
4012 /* we suppose the run is continuous, last=next...*/
4013 minnext = study_chunk(pRExC_state, &scan, minlenp,
4014 &deltanext, next, &data_fake, stopparen,
4015 recursed_depth, NULL, f,depth+1);
4019 if (deltanext == SSize_t_MAX) {
4020 is_inf = is_inf_internal = 1;
4022 } else if (max1 < minnext + deltanext)
4023 max1 = minnext + deltanext;
4025 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4027 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4028 if ( stopmin > minnext)
4029 stopmin = min + min1;
4030 flags &= ~SCF_DO_SUBSTR;
4032 data->flags |= SCF_SEEN_ACCEPT;
4035 if (data_fake.flags & SF_HAS_EVAL)
4036 data->flags |= SF_HAS_EVAL;
4037 data->whilem_c = data_fake.whilem_c;
4039 if (flags & SCF_DO_STCLASS)
4040 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
4042 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
4044 if (flags & SCF_DO_SUBSTR) {
4045 data->pos_min += min1;
4046 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
4047 data->pos_delta = SSize_t_MAX;
4049 data->pos_delta += max1 - min1;
4050 if (max1 != min1 || is_inf)
4051 data->longest = &(data->longest_float);
4054 if (delta == SSize_t_MAX
4055 || SSize_t_MAX - delta - (max1 - min1) < 0)
4056 delta = SSize_t_MAX;
4058 delta += max1 - min1;
4059 if (flags & SCF_DO_STCLASS_OR) {
4060 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4062 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4063 flags &= ~SCF_DO_STCLASS;
4066 else if (flags & SCF_DO_STCLASS_AND) {
4068 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4069 flags &= ~SCF_DO_STCLASS;
4072 /* Switch to OR mode: cache the old value of
4073 * data->start_class */
4075 StructCopy(data->start_class, and_withp, regnode_ssc);
4076 flags &= ~SCF_DO_STCLASS_AND;
4077 StructCopy(&accum, data->start_class, regnode_ssc);
4078 flags |= SCF_DO_STCLASS_OR;
4082 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4083 OP( startbranch ) == BRANCH )
4087 Assuming this was/is a branch we are dealing with: 'scan'
4088 now points at the item that follows the branch sequence,
4089 whatever it is. We now start at the beginning of the
4090 sequence and look for subsequences of
4096 which would be constructed from a pattern like
4099 If we can find such a subsequence we need to turn the first
4100 element into a trie and then add the subsequent branch exact
4101 strings to the trie.
4105 1. patterns where the whole set of branches can be
4108 2. patterns where only a subset can be converted.
4110 In case 1 we can replace the whole set with a single regop
4111 for the trie. In case 2 we need to keep the start and end
4114 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4115 becomes BRANCH TRIE; BRANCH X;
4117 There is an additional case, that being where there is a
4118 common prefix, which gets split out into an EXACT like node
4119 preceding the TRIE node.
4121 If x(1..n)==tail then we can do a simple trie, if not we make
4122 a "jump" trie, such that when we match the appropriate word
4123 we "jump" to the appropriate tail node. Essentially we turn
4124 a nested if into a case structure of sorts.
4129 if (!re_trie_maxbuff) {
4130 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4131 if (!SvIOK(re_trie_maxbuff))
4132 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4134 if ( SvIV(re_trie_maxbuff)>=0 ) {
4136 regnode *first = (regnode *)NULL;
4137 regnode *last = (regnode *)NULL;
4138 regnode *tail = scan;
4142 /* var tail is used because there may be a TAIL
4143 regop in the way. Ie, the exacts will point to the
4144 thing following the TAIL, but the last branch will
4145 point at the TAIL. So we advance tail. If we
4146 have nested (?:) we may have to move through several
4150 while ( OP( tail ) == TAIL ) {
4151 /* this is the TAIL generated by (?:) */
4152 tail = regnext( tail );
4156 DEBUG_TRIE_COMPILE_r({
4157 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4158 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
4159 (int)depth * 2 + 2, "",
4160 "Looking for TRIE'able sequences. Tail node is: ",
4161 SvPV_nolen_const( RExC_mysv )
4167 Step through the branches
4168 cur represents each branch,
4169 noper is the first thing to be matched as part
4171 noper_next is the regnext() of that node.
4173 We normally handle a case like this
4174 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4175 support building with NOJUMPTRIE, which restricts
4176 the trie logic to structures like /FOO|BAR/.
4178 If noper is a trieable nodetype then the branch is
4179 a possible optimization target. If we are building
4180 under NOJUMPTRIE then we require that noper_next is
4181 the same as scan (our current position in the regex
4184 Once we have two or more consecutive such branches
4185 we can create a trie of the EXACT's contents and
4186 stitch it in place into the program.
4188 If the sequence represents all of the branches in
4189 the alternation we replace the entire thing with a
4192 Otherwise when it is a subsequence we need to
4193 stitch it in place and replace only the relevant
4194 branches. This means the first branch has to remain
4195 as it is used by the alternation logic, and its
4196 next pointer, and needs to be repointed at the item
4197 on the branch chain following the last branch we
4198 have optimized away.
4200 This could be either a BRANCH, in which case the
4201 subsequence is internal, or it could be the item
4202 following the branch sequence in which case the
4203 subsequence is at the end (which does not
4204 necessarily mean the first node is the start of the
4207 TRIE_TYPE(X) is a define which maps the optype to a
4211 ----------------+-----------
4215 EXACTFU_SS | EXACTFU
4218 EXACTFLU8 | EXACTFLU8
4222 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4224 : ( EXACT == (X) ) \
4226 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4228 : ( EXACTFA == (X) ) \
4230 : ( EXACTL == (X) ) \
4232 : ( EXACTFLU8 == (X) ) \
4236 /* dont use tail as the end marker for this traverse */
4237 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4238 regnode * const noper = NEXTOPER( cur );
4239 U8 noper_type = OP( noper );
4240 U8 noper_trietype = TRIE_TYPE( noper_type );
4241 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4242 regnode * const noper_next = regnext( noper );
4243 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
4244 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
4247 DEBUG_TRIE_COMPILE_r({
4248 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4249 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
4250 (int)depth * 2 + 2,"", SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4252 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4253 PerlIO_printf( Perl_debug_log, " -> %s",
4254 SvPV_nolen_const(RExC_mysv));
4257 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4258 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
4259 SvPV_nolen_const(RExC_mysv));
4261 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
4262 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4263 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4267 /* Is noper a trieable nodetype that can be merged
4268 * with the current trie (if there is one)? */
4272 ( noper_trietype == NOTHING)
4273 || ( trietype == NOTHING )
4274 || ( trietype == noper_trietype )
4277 && noper_next == tail
4281 /* Handle mergable triable node Either we are
4282 * the first node in a new trieable sequence,
4283 * in which case we do some bookkeeping,
4284 * otherwise we update the end pointer. */
4287 if ( noper_trietype == NOTHING ) {
4288 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4289 regnode * const noper_next = regnext( noper );
4290 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4291 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4294 if ( noper_next_trietype ) {
4295 trietype = noper_next_trietype;
4296 } else if (noper_next_type) {
4297 /* a NOTHING regop is 1 regop wide.
4298 * We need at least two for a trie
4299 * so we can't merge this in */
4303 trietype = noper_trietype;
4306 if ( trietype == NOTHING )
4307 trietype = noper_trietype;
4312 } /* end handle mergable triable node */
4314 /* handle unmergable node -
4315 * noper may either be a triable node which can
4316 * not be tried together with the current trie,
4317 * or a non triable node */
4319 /* If last is set and trietype is not
4320 * NOTHING then we have found at least two
4321 * triable branch sequences in a row of a
4322 * similar trietype so we can turn them
4323 * into a trie. If/when we allow NOTHING to
4324 * start a trie sequence this condition
4325 * will be required, and it isn't expensive
4326 * so we leave it in for now. */
4327 if ( trietype && trietype != NOTHING )
4328 make_trie( pRExC_state,
4329 startbranch, first, cur, tail,
4330 count, trietype, depth+1 );
4331 last = NULL; /* note: we clear/update
4332 first, trietype etc below,
4333 so we dont do it here */
4337 && noper_next == tail
4340 /* noper is triable, so we can start a new
4344 trietype = noper_trietype;
4346 /* if we already saw a first but the
4347 * current node is not triable then we have
4348 * to reset the first information. */
4353 } /* end handle unmergable node */
4354 } /* loop over branches */
4355 DEBUG_TRIE_COMPILE_r({
4356 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4357 PerlIO_printf( Perl_debug_log,
4358 "%*s- %s (%d) <SCAN FINISHED>\n",
4360 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4363 if ( last && trietype ) {
4364 if ( trietype != NOTHING ) {
4365 /* the last branch of the sequence was part of
4366 * a trie, so we have to construct it here
4367 * outside of the loop */
4368 made= make_trie( pRExC_state, startbranch,
4369 first, scan, tail, count,
4370 trietype, depth+1 );
4371 #ifdef TRIE_STUDY_OPT
4372 if ( ((made == MADE_EXACT_TRIE &&
4373 startbranch == first)
4374 || ( first_non_open == first )) &&
4376 flags |= SCF_TRIE_RESTUDY;
4377 if ( startbranch == first
4380 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4385 /* at this point we know whatever we have is a
4386 * NOTHING sequence/branch AND if 'startbranch'
4387 * is 'first' then we can turn the whole thing
4390 if ( startbranch == first ) {
4392 /* the entire thing is a NOTHING sequence,
4393 * something like this: (?:|) So we can
4394 * turn it into a plain NOTHING op. */
4395 DEBUG_TRIE_COMPILE_r({
4396 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4397 PerlIO_printf( Perl_debug_log,
4398 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4399 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4402 OP(startbranch)= NOTHING;
4403 NEXT_OFF(startbranch)= tail - startbranch;
4404 for ( opt= startbranch + 1; opt < tail ; opt++ )
4408 } /* end if ( last) */
4409 } /* TRIE_MAXBUF is non zero */
4414 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4415 scan = NEXTOPER(NEXTOPER(scan));
4416 } else /* single branch is optimized. */
4417 scan = NEXTOPER(scan);
4419 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4421 regnode *start = NULL;
4422 regnode *end = NULL;
4423 U32 my_recursed_depth= recursed_depth;
4426 if (OP(scan) != SUSPEND) { /* GOSUB/GOSTART */
4427 /* Do setup, note this code has side effects beyond
4428 * the rest of this block. Specifically setting
4429 * RExC_recurse[] must happen at least once during
4431 if (OP(scan) == GOSUB) {
4433 RExC_recurse[ARG2L(scan)] = scan;
4434 start = RExC_open_parens[paren-1];
4435 end = RExC_close_parens[paren-1];
4437 start = RExC_rxi->program + 1;
4440 /* NOTE we MUST always execute the above code, even
4441 * if we do nothing with a GOSUB/GOSTART */
4443 ( flags & SCF_IN_DEFINE )
4446 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4448 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4451 /* no need to do anything here if we are in a define. */
4452 /* or we are after some kind of infinite construct
4453 * so we can skip recursing into this item.
4454 * Since it is infinite we will not change the maxlen
4455 * or delta, and if we miss something that might raise
4456 * the minlen it will merely pessimise a little.
4458 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4459 * might result in a minlen of 1 and not of 4,
4460 * but this doesn't make us mismatch, just try a bit
4461 * harder than we should.
4463 scan= regnext(scan);
4470 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4472 /* it is quite possible that there are more efficient ways
4473 * to do this. We maintain a bitmap per level of recursion
4474 * of which patterns we have entered so we can detect if a
4475 * pattern creates a possible infinite loop. When we
4476 * recurse down a level we copy the previous levels bitmap
4477 * down. When we are at recursion level 0 we zero the top
4478 * level bitmap. It would be nice to implement a different
4479 * more efficient way of doing this. In particular the top
4480 * level bitmap may be unnecessary.
4482 if (!recursed_depth) {
4483 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4485 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4486 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4487 RExC_study_chunk_recursed_bytes, U8);
4489 /* we havent recursed into this paren yet, so recurse into it */
4490 DEBUG_STUDYDATA("set:", data,depth);
4491 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4492 my_recursed_depth= recursed_depth + 1;
4494 DEBUG_STUDYDATA("inf:", data,depth);
4495 /* some form of infinite recursion, assume infinite length
4497 if (flags & SCF_DO_SUBSTR) {
4498 scan_commit(pRExC_state, data, minlenp, is_inf);
4499 data->longest = &(data->longest_float);
4501 is_inf = is_inf_internal = 1;
4502 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4503 ssc_anything(data->start_class);
4504 flags &= ~SCF_DO_STCLASS;
4506 start= NULL; /* reset start so we dont recurse later on. */
4511 end = regnext(scan);
4514 scan_frame *newframe;
4516 if (!RExC_frame_last) {
4517 Newxz(newframe, 1, scan_frame);
4518 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4519 RExC_frame_head= newframe;
4521 } else if (!RExC_frame_last->next_frame) {
4522 Newxz(newframe,1,scan_frame);
4523 RExC_frame_last->next_frame= newframe;
4524 newframe->prev_frame= RExC_frame_last;
4527 newframe= RExC_frame_last->next_frame;
4529 RExC_frame_last= newframe;
4531 newframe->next_regnode = regnext(scan);
4532 newframe->last_regnode = last;
4533 newframe->stopparen = stopparen;
4534 newframe->prev_recursed_depth = recursed_depth;
4535 newframe->this_prev_frame= frame;
4537 DEBUG_STUDYDATA("frame-new:",data,depth);
4538 DEBUG_PEEP("fnew", scan, depth);
4545 recursed_depth= my_recursed_depth;
4550 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4551 SSize_t l = STR_LEN(scan);
4554 const U8 * const s = (U8*)STRING(scan);
4555 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4556 l = utf8_length(s, s + l);
4558 uc = *((U8*)STRING(scan));
4561 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4562 /* The code below prefers earlier match for fixed
4563 offset, later match for variable offset. */
4564 if (data->last_end == -1) { /* Update the start info. */
4565 data->last_start_min = data->pos_min;
4566 data->last_start_max = is_inf
4567 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4569 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4571 SvUTF8_on(data->last_found);
4573 SV * const sv = data->last_found;
4574 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4575 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4576 if (mg && mg->mg_len >= 0)
4577 mg->mg_len += utf8_length((U8*)STRING(scan),
4578 (U8*)STRING(scan)+STR_LEN(scan));
4580 data->last_end = data->pos_min + l;
4581 data->pos_min += l; /* As in the first entry. */
4582 data->flags &= ~SF_BEFORE_EOL;
4585 /* ANDing the code point leaves at most it, and not in locale, and
4586 * can't match null string */
4587 if (flags & SCF_DO_STCLASS_AND) {
4588 ssc_cp_and(data->start_class, uc);
4589 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4590 ssc_clear_locale(data->start_class);
4592 else if (flags & SCF_DO_STCLASS_OR) {
4593 ssc_add_cp(data->start_class, uc);
4594 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4596 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4597 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4599 flags &= ~SCF_DO_STCLASS;
4601 else if (PL_regkind[OP(scan)] == EXACT) {
4602 /* But OP != EXACT!, so is EXACTFish */
4603 SSize_t l = STR_LEN(scan);
4604 const U8 * s = (U8*)STRING(scan);
4606 /* Search for fixed substrings supports EXACT only. */
4607 if (flags & SCF_DO_SUBSTR) {
4609 scan_commit(pRExC_state, data, minlenp, is_inf);
4612 l = utf8_length(s, s + l);
4614 if (unfolded_multi_char) {
4615 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4617 min += l - min_subtract;
4619 delta += min_subtract;
4620 if (flags & SCF_DO_SUBSTR) {
4621 data->pos_min += l - min_subtract;
4622 if (data->pos_min < 0) {
4625 data->pos_delta += min_subtract;
4627 data->longest = &(data->longest_float);
4631 if (flags & SCF_DO_STCLASS) {
4632 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4634 assert(EXACTF_invlist);
4635 if (flags & SCF_DO_STCLASS_AND) {
4636 if (OP(scan) != EXACTFL)
4637 ssc_clear_locale(data->start_class);
4638 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4639 ANYOF_POSIXL_ZERO(data->start_class);
4640 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4642 else { /* SCF_DO_STCLASS_OR */
4643 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4644 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4646 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4647 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4649 flags &= ~SCF_DO_STCLASS;
4650 SvREFCNT_dec(EXACTF_invlist);
4653 else if (REGNODE_VARIES(OP(scan))) {
4654 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4655 I32 fl = 0, f = flags;
4656 regnode * const oscan = scan;
4657 regnode_ssc this_class;
4658 regnode_ssc *oclass = NULL;
4659 I32 next_is_eval = 0;
4661 switch (PL_regkind[OP(scan)]) {
4662 case WHILEM: /* End of (?:...)* . */
4663 scan = NEXTOPER(scan);
4666 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4667 next = NEXTOPER(scan);
4668 if (OP(next) == EXACT
4669 || OP(next) == EXACTL
4670 || (flags & SCF_DO_STCLASS))
4673 maxcount = REG_INFTY;
4674 next = regnext(scan);
4675 scan = NEXTOPER(scan);
4679 if (flags & SCF_DO_SUBSTR)
4684 if (flags & SCF_DO_STCLASS) {
4686 maxcount = REG_INFTY;
4687 next = regnext(scan);
4688 scan = NEXTOPER(scan);
4691 if (flags & SCF_DO_SUBSTR) {
4692 scan_commit(pRExC_state, data, minlenp, is_inf);
4693 /* Cannot extend fixed substrings */
4694 data->longest = &(data->longest_float);
4696 is_inf = is_inf_internal = 1;
4697 scan = regnext(scan);
4698 goto optimize_curly_tail;
4700 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4701 && (scan->flags == stopparen))
4706 mincount = ARG1(scan);
4707 maxcount = ARG2(scan);
4709 next = regnext(scan);
4710 if (OP(scan) == CURLYX) {
4711 I32 lp = (data ? *(data->last_closep) : 0);
4712 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4714 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4715 next_is_eval = (OP(scan) == EVAL);
4717 if (flags & SCF_DO_SUBSTR) {
4719 scan_commit(pRExC_state, data, minlenp, is_inf);
4720 /* Cannot extend fixed substrings */
4721 pos_before = data->pos_min;
4725 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4727 data->flags |= SF_IS_INF;
4729 if (flags & SCF_DO_STCLASS) {
4730 ssc_init(pRExC_state, &this_class);
4731 oclass = data->start_class;
4732 data->start_class = &this_class;
4733 f |= SCF_DO_STCLASS_AND;
4734 f &= ~SCF_DO_STCLASS_OR;
4736 /* Exclude from super-linear cache processing any {n,m}
4737 regops for which the combination of input pos and regex
4738 pos is not enough information to determine if a match
4741 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4742 regex pos at the \s*, the prospects for a match depend not
4743 only on the input position but also on how many (bar\s*)
4744 repeats into the {4,8} we are. */
4745 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4746 f &= ~SCF_WHILEM_VISITED_POS;
4748 /* This will finish on WHILEM, setting scan, or on NULL: */
4749 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4750 last, data, stopparen, recursed_depth, NULL,
4752 ? (f & ~SCF_DO_SUBSTR)
4756 if (flags & SCF_DO_STCLASS)
4757 data->start_class = oclass;
4758 if (mincount == 0 || minnext == 0) {
4759 if (flags & SCF_DO_STCLASS_OR) {
4760 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4762 else if (flags & SCF_DO_STCLASS_AND) {
4763 /* Switch to OR mode: cache the old value of
4764 * data->start_class */
4766 StructCopy(data->start_class, and_withp, regnode_ssc);
4767 flags &= ~SCF_DO_STCLASS_AND;
4768 StructCopy(&this_class, data->start_class, regnode_ssc);
4769 flags |= SCF_DO_STCLASS_OR;
4770 ANYOF_FLAGS(data->start_class)
4771 |= SSC_MATCHES_EMPTY_STRING;
4773 } else { /* Non-zero len */
4774 if (flags & SCF_DO_STCLASS_OR) {
4775 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4776 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4778 else if (flags & SCF_DO_STCLASS_AND)
4779 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4780 flags &= ~SCF_DO_STCLASS;
4782 if (!scan) /* It was not CURLYX, but CURLY. */
4784 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4785 /* ? quantifier ok, except for (?{ ... }) */
4786 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4787 && (minnext == 0) && (deltanext == 0)
4788 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4789 && maxcount <= REG_INFTY/3) /* Complement check for big
4792 /* Fatal warnings may leak the regexp without this: */
4793 SAVEFREESV(RExC_rx_sv);
4794 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
4795 "Quantifier unexpected on zero-length expression "
4796 "in regex m/%"UTF8f"/",
4797 UTF8fARG(UTF, RExC_end - RExC_precomp,
4799 (void)ReREFCNT_inc(RExC_rx_sv);
4802 min += minnext * mincount;
4803 is_inf_internal |= deltanext == SSize_t_MAX
4804 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4805 is_inf |= is_inf_internal;
4807 delta = SSize_t_MAX;
4809 delta += (minnext + deltanext) * maxcount
4810 - minnext * mincount;
4812 /* Try powerful optimization CURLYX => CURLYN. */
4813 if ( OP(oscan) == CURLYX && data
4814 && data->flags & SF_IN_PAR
4815 && !(data->flags & SF_HAS_EVAL)
4816 && !deltanext && minnext == 1 ) {
4817 /* Try to optimize to CURLYN. */
4818 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4819 regnode * const nxt1 = nxt;
4826 if (!REGNODE_SIMPLE(OP(nxt))
4827 && !(PL_regkind[OP(nxt)] == EXACT
4828 && STR_LEN(nxt) == 1))
4834 if (OP(nxt) != CLOSE)
4836 if (RExC_open_parens) {
4837 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4838 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4840 /* Now we know that nxt2 is the only contents: */
4841 oscan->flags = (U8)ARG(nxt);
4843 OP(nxt1) = NOTHING; /* was OPEN. */
4846 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4847 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4848 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4849 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4850 OP(nxt + 1) = OPTIMIZED; /* was count. */
4851 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4856 /* Try optimization CURLYX => CURLYM. */
4857 if ( OP(oscan) == CURLYX && data
4858 && !(data->flags & SF_HAS_PAR)
4859 && !(data->flags & SF_HAS_EVAL)
4860 && !deltanext /* atom is fixed width */
4861 && minnext != 0 /* CURLYM can't handle zero width */
4863 /* Nor characters whose fold at run-time may be
4864 * multi-character */
4865 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4867 /* XXXX How to optimize if data == 0? */
4868 /* Optimize to a simpler form. */
4869 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4873 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4874 && (OP(nxt2) != WHILEM))
4876 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4877 /* Need to optimize away parenths. */
4878 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4879 /* Set the parenth number. */
4880 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4882 oscan->flags = (U8)ARG(nxt);
4883 if (RExC_open_parens) {
4884 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4885 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4887 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4888 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4891 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4892 OP(nxt + 1) = OPTIMIZED; /* was count. */
4893 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4894 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4897 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4898 regnode *nnxt = regnext(nxt1);
4900 if (reg_off_by_arg[OP(nxt1)])
4901 ARG_SET(nxt1, nxt2 - nxt1);
4902 else if (nxt2 - nxt1 < U16_MAX)
4903 NEXT_OFF(nxt1) = nxt2 - nxt1;
4905 OP(nxt) = NOTHING; /* Cannot beautify */
4910 /* Optimize again: */
4911 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4912 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4917 else if ((OP(oscan) == CURLYX)
4918 && (flags & SCF_WHILEM_VISITED_POS)
4919 /* See the comment on a similar expression above.
4920 However, this time it's not a subexpression
4921 we care about, but the expression itself. */
4922 && (maxcount == REG_INFTY)
4923 && data && ++data->whilem_c < 16) {
4924 /* This stays as CURLYX, we can put the count/of pair. */
4925 /* Find WHILEM (as in regexec.c) */
4926 regnode *nxt = oscan + NEXT_OFF(oscan);
4928 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4930 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4931 | (RExC_whilem_seen << 4)); /* On WHILEM */
4933 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4935 if (flags & SCF_DO_SUBSTR) {
4936 SV *last_str = NULL;
4937 STRLEN last_chrs = 0;
4938 int counted = mincount != 0;
4940 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4942 SSize_t b = pos_before >= data->last_start_min
4943 ? pos_before : data->last_start_min;
4945 const char * const s = SvPV_const(data->last_found, l);
4946 SSize_t old = b - data->last_start_min;
4949 old = utf8_hop((U8*)s, old) - (U8*)s;
4951 /* Get the added string: */
4952 last_str = newSVpvn_utf8(s + old, l, UTF);
4953 last_chrs = UTF ? utf8_length((U8*)(s + old),
4954 (U8*)(s + old + l)) : l;
4955 if (deltanext == 0 && pos_before == b) {
4956 /* What was added is a constant string */
4959 SvGROW(last_str, (mincount * l) + 1);
4960 repeatcpy(SvPVX(last_str) + l,
4961 SvPVX_const(last_str), l,
4963 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4964 /* Add additional parts. */
4965 SvCUR_set(data->last_found,
4966 SvCUR(data->last_found) - l);
4967 sv_catsv(data->last_found, last_str);
4969 SV * sv = data->last_found;
4971 SvUTF8(sv) && SvMAGICAL(sv) ?
4972 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4973 if (mg && mg->mg_len >= 0)
4974 mg->mg_len += last_chrs * (mincount-1);
4976 last_chrs *= mincount;
4977 data->last_end += l * (mincount - 1);
4980 /* start offset must point into the last copy */
4981 data->last_start_min += minnext * (mincount - 1);
4982 data->last_start_max =
4985 : data->last_start_max +
4986 (maxcount - 1) * (minnext + data->pos_delta);
4989 /* It is counted once already... */
4990 data->pos_min += minnext * (mincount - counted);
4992 PerlIO_printf(Perl_debug_log, "counted=%"UVuf" deltanext=%"UVuf
4993 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
4994 " maxcount=%"UVuf" mincount=%"UVuf"\n",
4995 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4997 if (deltanext != SSize_t_MAX)
4998 PerlIO_printf(Perl_debug_log, "LHS=%"UVuf" RHS=%"UVuf"\n",
4999 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
5000 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
5002 if (deltanext == SSize_t_MAX
5003 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
5004 data->pos_delta = SSize_t_MAX;
5006 data->pos_delta += - counted * deltanext +
5007 (minnext + deltanext) * maxcount - minnext * mincount;
5008 if (mincount != maxcount) {
5009 /* Cannot extend fixed substrings found inside
5011 scan_commit(pRExC_state, data, minlenp, is_inf);
5012 if (mincount && last_str) {
5013 SV * const sv = data->last_found;
5014 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
5015 mg_find(sv, PERL_MAGIC_utf8) : NULL;
5019 sv_setsv(sv, last_str);
5020 data->last_end = data->pos_min;
5021 data->last_start_min = data->pos_min - last_chrs;
5022 data->last_start_max = is_inf
5024 : data->pos_min + data->pos_delta - last_chrs;
5026 data->longest = &(data->longest_float);
5028 SvREFCNT_dec(last_str);
5030 if (data && (fl & SF_HAS_EVAL))
5031 data->flags |= SF_HAS_EVAL;
5032 optimize_curly_tail:
5033 if (OP(oscan) != CURLYX) {
5034 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
5036 NEXT_OFF(oscan) += NEXT_OFF(next);
5042 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
5047 if (flags & SCF_DO_SUBSTR) {
5048 /* Cannot expect anything... */
5049 scan_commit(pRExC_state, data, minlenp, is_inf);
5050 data->longest = &(data->longest_float);
5052 is_inf = is_inf_internal = 1;
5053 if (flags & SCF_DO_STCLASS_OR) {
5054 if (OP(scan) == CLUMP) {
5055 /* Actually is any start char, but very few code points
5056 * aren't start characters */
5057 ssc_match_all_cp(data->start_class);
5060 ssc_anything(data->start_class);
5063 flags &= ~SCF_DO_STCLASS;
5067 else if (OP(scan) == LNBREAK) {
5068 if (flags & SCF_DO_STCLASS) {
5069 if (flags & SCF_DO_STCLASS_AND) {
5070 ssc_intersection(data->start_class,
5071 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5072 ssc_clear_locale(data->start_class);
5073 ANYOF_FLAGS(data->start_class)
5074 &= ~SSC_MATCHES_EMPTY_STRING;
5076 else if (flags & SCF_DO_STCLASS_OR) {
5077 ssc_union(data->start_class,
5078 PL_XPosix_ptrs[_CC_VERTSPACE],
5080 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5082 /* See commit msg for
5083 * 749e076fceedeb708a624933726e7989f2302f6a */
5084 ANYOF_FLAGS(data->start_class)
5085 &= ~SSC_MATCHES_EMPTY_STRING;
5087 flags &= ~SCF_DO_STCLASS;
5090 if (delta != SSize_t_MAX)
5091 delta++; /* Because of the 2 char string cr-lf */
5092 if (flags & SCF_DO_SUBSTR) {
5093 /* Cannot expect anything... */
5094 scan_commit(pRExC_state, data, minlenp, is_inf);
5096 data->pos_delta += 1;
5097 data->longest = &(data->longest_float);
5100 else if (REGNODE_SIMPLE(OP(scan))) {
5102 if (flags & SCF_DO_SUBSTR) {
5103 scan_commit(pRExC_state, data, minlenp, is_inf);
5107 if (flags & SCF_DO_STCLASS) {
5109 SV* my_invlist = NULL;
5112 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5113 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5115 /* Some of the logic below assumes that switching
5116 locale on will only add false positives. */
5121 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5125 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5126 ssc_match_all_cp(data->start_class);
5131 SV* REG_ANY_invlist = _new_invlist(2);
5132 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5134 if (flags & SCF_DO_STCLASS_OR) {
5135 ssc_union(data->start_class,
5137 TRUE /* TRUE => invert, hence all but \n
5141 else if (flags & SCF_DO_STCLASS_AND) {
5142 ssc_intersection(data->start_class,
5144 TRUE /* TRUE => invert */
5146 ssc_clear_locale(data->start_class);
5148 SvREFCNT_dec_NN(REG_ANY_invlist);
5155 if (flags & SCF_DO_STCLASS_AND)
5156 ssc_and(pRExC_state, data->start_class,
5157 (regnode_charclass *) scan);
5159 ssc_or(pRExC_state, data->start_class,
5160 (regnode_charclass *) scan);
5168 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5169 if (flags & SCF_DO_STCLASS_AND) {
5170 bool was_there = cBOOL(
5171 ANYOF_POSIXL_TEST(data->start_class,
5173 ANYOF_POSIXL_ZERO(data->start_class);
5174 if (was_there) { /* Do an AND */
5175 ANYOF_POSIXL_SET(data->start_class, namedclass);
5177 /* No individual code points can now match */
5178 data->start_class->invlist
5179 = sv_2mortal(_new_invlist(0));
5182 int complement = namedclass + ((invert) ? -1 : 1);
5184 assert(flags & SCF_DO_STCLASS_OR);
5186 /* If the complement of this class was already there,
5187 * the result is that they match all code points,
5188 * (\d + \D == everything). Remove the classes from
5189 * future consideration. Locale is not relevant in
5191 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5192 ssc_match_all_cp(data->start_class);
5193 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5194 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5196 else { /* The usual case; just add this class to the
5198 ANYOF_POSIXL_SET(data->start_class, namedclass);
5203 case NPOSIXA: /* For these, we always know the exact set of
5208 if (FLAGS(scan) == _CC_ASCII) {
5209 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5212 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5213 PL_XPosix_ptrs[_CC_ASCII],
5224 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5226 /* NPOSIXD matches all upper Latin1 code points unless the
5227 * target string being matched is UTF-8, which is
5228 * unknowable until match time. Since we are going to
5229 * invert, we want to get rid of all of them so that the
5230 * inversion will match all */
5231 if (OP(scan) == NPOSIXD) {
5232 _invlist_subtract(my_invlist, PL_UpperLatin1,
5238 if (flags & SCF_DO_STCLASS_AND) {
5239 ssc_intersection(data->start_class, my_invlist, invert);
5240 ssc_clear_locale(data->start_class);
5243 assert(flags & SCF_DO_STCLASS_OR);
5244 ssc_union(data->start_class, my_invlist, invert);
5246 SvREFCNT_dec(my_invlist);
5248 if (flags & SCF_DO_STCLASS_OR)
5249 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5250 flags &= ~SCF_DO_STCLASS;
5253 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5254 data->flags |= (OP(scan) == MEOL
5257 scan_commit(pRExC_state, data, minlenp, is_inf);
5260 else if ( PL_regkind[OP(scan)] == BRANCHJ
5261 /* Lookbehind, or need to calculate parens/evals/stclass: */
5262 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5263 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5265 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5266 || OP(scan) == UNLESSM )
5268 /* Negative Lookahead/lookbehind
5269 In this case we can't do fixed string optimisation.
5272 SSize_t deltanext, minnext, fake = 0;
5277 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5279 data_fake.whilem_c = data->whilem_c;
5280 data_fake.last_closep = data->last_closep;
5283 data_fake.last_closep = &fake;
5284 data_fake.pos_delta = delta;
5285 if ( flags & SCF_DO_STCLASS && !scan->flags
5286 && OP(scan) == IFMATCH ) { /* Lookahead */
5287 ssc_init(pRExC_state, &intrnl);
5288 data_fake.start_class = &intrnl;
5289 f |= SCF_DO_STCLASS_AND;
5291 if (flags & SCF_WHILEM_VISITED_POS)
5292 f |= SCF_WHILEM_VISITED_POS;
5293 next = regnext(scan);
5294 nscan = NEXTOPER(NEXTOPER(scan));
5295 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5296 last, &data_fake, stopparen,
5297 recursed_depth, NULL, f, depth+1);
5300 FAIL("Variable length lookbehind not implemented");
5302 else if (minnext > (I32)U8_MAX) {
5303 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5306 scan->flags = (U8)minnext;
5309 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5311 if (data_fake.flags & SF_HAS_EVAL)
5312 data->flags |= SF_HAS_EVAL;
5313 data->whilem_c = data_fake.whilem_c;
5315 if (f & SCF_DO_STCLASS_AND) {
5316 if (flags & SCF_DO_STCLASS_OR) {
5317 /* OR before, AND after: ideally we would recurse with
5318 * data_fake to get the AND applied by study of the
5319 * remainder of the pattern, and then derecurse;
5320 * *** HACK *** for now just treat as "no information".
5321 * See [perl #56690].
5323 ssc_init(pRExC_state, data->start_class);
5325 /* AND before and after: combine and continue. These
5326 * assertions are zero-length, so can match an EMPTY
5328 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5329 ANYOF_FLAGS(data->start_class)
5330 |= SSC_MATCHES_EMPTY_STRING;
5334 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5336 /* Positive Lookahead/lookbehind
5337 In this case we can do fixed string optimisation,
5338 but we must be careful about it. Note in the case of
5339 lookbehind the positions will be offset by the minimum
5340 length of the pattern, something we won't know about
5341 until after the recurse.
5343 SSize_t deltanext, fake = 0;
5347 /* We use SAVEFREEPV so that when the full compile
5348 is finished perl will clean up the allocated
5349 minlens when it's all done. This way we don't
5350 have to worry about freeing them when we know
5351 they wont be used, which would be a pain.
5354 Newx( minnextp, 1, SSize_t );
5355 SAVEFREEPV(minnextp);
5358 StructCopy(data, &data_fake, scan_data_t);
5359 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5362 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5363 data_fake.last_found=newSVsv(data->last_found);
5367 data_fake.last_closep = &fake;
5368 data_fake.flags = 0;
5369 data_fake.pos_delta = delta;
5371 data_fake.flags |= SF_IS_INF;
5372 if ( flags & SCF_DO_STCLASS && !scan->flags
5373 && OP(scan) == IFMATCH ) { /* Lookahead */
5374 ssc_init(pRExC_state, &intrnl);
5375 data_fake.start_class = &intrnl;
5376 f |= SCF_DO_STCLASS_AND;
5378 if (flags & SCF_WHILEM_VISITED_POS)
5379 f |= SCF_WHILEM_VISITED_POS;
5380 next = regnext(scan);
5381 nscan = NEXTOPER(NEXTOPER(scan));
5383 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5384 &deltanext, last, &data_fake,
5385 stopparen, recursed_depth, NULL,
5389 FAIL("Variable length lookbehind not implemented");
5391 else if (*minnextp > (I32)U8_MAX) {
5392 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5395 scan->flags = (U8)*minnextp;
5400 if (f & SCF_DO_STCLASS_AND) {
5401 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5402 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5405 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5407 if (data_fake.flags & SF_HAS_EVAL)
5408 data->flags |= SF_HAS_EVAL;
5409 data->whilem_c = data_fake.whilem_c;
5410 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5411 if (RExC_rx->minlen<*minnextp)
5412 RExC_rx->minlen=*minnextp;
5413 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5414 SvREFCNT_dec_NN(data_fake.last_found);
5416 if ( data_fake.minlen_fixed != minlenp )
5418 data->offset_fixed= data_fake.offset_fixed;
5419 data->minlen_fixed= data_fake.minlen_fixed;
5420 data->lookbehind_fixed+= scan->flags;
5422 if ( data_fake.minlen_float != minlenp )
5424 data->minlen_float= data_fake.minlen_float;
5425 data->offset_float_min=data_fake.offset_float_min;
5426 data->offset_float_max=data_fake.offset_float_max;
5427 data->lookbehind_float+= scan->flags;
5434 else if (OP(scan) == OPEN) {
5435 if (stopparen != (I32)ARG(scan))
5438 else if (OP(scan) == CLOSE) {
5439 if (stopparen == (I32)ARG(scan)) {
5442 if ((I32)ARG(scan) == is_par) {
5443 next = regnext(scan);
5445 if ( next && (OP(next) != WHILEM) && next < last)
5446 is_par = 0; /* Disable optimization */
5449 *(data->last_closep) = ARG(scan);
5451 else if (OP(scan) == EVAL) {
5453 data->flags |= SF_HAS_EVAL;
5455 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5456 if (flags & SCF_DO_SUBSTR) {
5457 scan_commit(pRExC_state, data, minlenp, is_inf);
5458 flags &= ~SCF_DO_SUBSTR;
5460 if (data && OP(scan)==ACCEPT) {
5461 data->flags |= SCF_SEEN_ACCEPT;
5466 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5468 if (flags & SCF_DO_SUBSTR) {
5469 scan_commit(pRExC_state, data, minlenp, is_inf);
5470 data->longest = &(data->longest_float);
5472 is_inf = is_inf_internal = 1;
5473 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5474 ssc_anything(data->start_class);
5475 flags &= ~SCF_DO_STCLASS;
5477 else if (OP(scan) == GPOS) {
5478 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5479 !(delta || is_inf || (data && data->pos_delta)))
5481 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5482 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5483 if (RExC_rx->gofs < (STRLEN)min)
5484 RExC_rx->gofs = min;
5486 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5490 #ifdef TRIE_STUDY_OPT
5491 #ifdef FULL_TRIE_STUDY
5492 else if (PL_regkind[OP(scan)] == TRIE) {
5493 /* NOTE - There is similar code to this block above for handling
5494 BRANCH nodes on the initial study. If you change stuff here
5496 regnode *trie_node= scan;
5497 regnode *tail= regnext(scan);
5498 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5499 SSize_t max1 = 0, min1 = SSize_t_MAX;
5502 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5503 /* Cannot merge strings after this. */
5504 scan_commit(pRExC_state, data, minlenp, is_inf);
5506 if (flags & SCF_DO_STCLASS)
5507 ssc_init_zero(pRExC_state, &accum);
5513 const regnode *nextbranch= NULL;
5516 for ( word=1 ; word <= trie->wordcount ; word++)
5518 SSize_t deltanext=0, minnext=0, f = 0, fake;
5519 regnode_ssc this_class;
5521 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5523 data_fake.whilem_c = data->whilem_c;
5524 data_fake.last_closep = data->last_closep;
5527 data_fake.last_closep = &fake;
5528 data_fake.pos_delta = delta;
5529 if (flags & SCF_DO_STCLASS) {
5530 ssc_init(pRExC_state, &this_class);
5531 data_fake.start_class = &this_class;
5532 f = SCF_DO_STCLASS_AND;
5534 if (flags & SCF_WHILEM_VISITED_POS)
5535 f |= SCF_WHILEM_VISITED_POS;
5537 if (trie->jump[word]) {
5539 nextbranch = trie_node + trie->jump[0];
5540 scan= trie_node + trie->jump[word];
5541 /* We go from the jump point to the branch that follows
5542 it. Note this means we need the vestigal unused
5543 branches even though they arent otherwise used. */
5544 minnext = study_chunk(pRExC_state, &scan, minlenp,
5545 &deltanext, (regnode *)nextbranch, &data_fake,
5546 stopparen, recursed_depth, NULL, f,depth+1);
5548 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5549 nextbranch= regnext((regnode*)nextbranch);
5551 if (min1 > (SSize_t)(minnext + trie->minlen))
5552 min1 = minnext + trie->minlen;
5553 if (deltanext == SSize_t_MAX) {
5554 is_inf = is_inf_internal = 1;
5556 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5557 max1 = minnext + deltanext + trie->maxlen;
5559 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5561 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5562 if ( stopmin > min + min1)
5563 stopmin = min + min1;
5564 flags &= ~SCF_DO_SUBSTR;
5566 data->flags |= SCF_SEEN_ACCEPT;
5569 if (data_fake.flags & SF_HAS_EVAL)
5570 data->flags |= SF_HAS_EVAL;
5571 data->whilem_c = data_fake.whilem_c;
5573 if (flags & SCF_DO_STCLASS)
5574 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5577 if (flags & SCF_DO_SUBSTR) {
5578 data->pos_min += min1;
5579 data->pos_delta += max1 - min1;
5580 if (max1 != min1 || is_inf)
5581 data->longest = &(data->longest_float);
5584 if (delta != SSize_t_MAX)
5585 delta += max1 - min1;
5586 if (flags & SCF_DO_STCLASS_OR) {
5587 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5589 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5590 flags &= ~SCF_DO_STCLASS;
5593 else if (flags & SCF_DO_STCLASS_AND) {
5595 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5596 flags &= ~SCF_DO_STCLASS;
5599 /* Switch to OR mode: cache the old value of
5600 * data->start_class */
5602 StructCopy(data->start_class, and_withp, regnode_ssc);
5603 flags &= ~SCF_DO_STCLASS_AND;
5604 StructCopy(&accum, data->start_class, regnode_ssc);
5605 flags |= SCF_DO_STCLASS_OR;
5612 else if (PL_regkind[OP(scan)] == TRIE) {
5613 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5616 min += trie->minlen;
5617 delta += (trie->maxlen - trie->minlen);
5618 flags &= ~SCF_DO_STCLASS; /* xxx */
5619 if (flags & SCF_DO_SUBSTR) {
5620 /* Cannot expect anything... */
5621 scan_commit(pRExC_state, data, minlenp, is_inf);
5622 data->pos_min += trie->minlen;
5623 data->pos_delta += (trie->maxlen - trie->minlen);
5624 if (trie->maxlen != trie->minlen)
5625 data->longest = &(data->longest_float);
5627 if (trie->jump) /* no more substrings -- for now /grr*/
5628 flags &= ~SCF_DO_SUBSTR;
5630 #endif /* old or new */
5631 #endif /* TRIE_STUDY_OPT */
5633 /* Else: zero-length, ignore. */
5634 scan = regnext(scan);
5636 /* If we are exiting a recursion we can unset its recursed bit
5637 * and allow ourselves to enter it again - no danger of an
5638 * infinite loop there.
5639 if (stopparen > -1 && recursed) {
5640 DEBUG_STUDYDATA("unset:", data,depth);
5641 PAREN_UNSET( recursed, stopparen);
5647 DEBUG_STUDYDATA("frame-end:",data,depth);
5648 DEBUG_PEEP("fend", scan, depth);
5650 /* restore previous context */
5651 last = frame->last_regnode;
5652 scan = frame->next_regnode;
5653 stopparen = frame->stopparen;
5654 recursed_depth = frame->prev_recursed_depth;
5656 RExC_frame_last = frame->prev_frame;
5657 frame = frame->this_prev_frame;
5658 goto fake_study_recurse;
5663 DEBUG_STUDYDATA("pre-fin:",data,depth);
5666 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5668 if (flags & SCF_DO_SUBSTR && is_inf)
5669 data->pos_delta = SSize_t_MAX - data->pos_min;
5670 if (is_par > (I32)U8_MAX)
5672 if (is_par && pars==1 && data) {
5673 data->flags |= SF_IN_PAR;
5674 data->flags &= ~SF_HAS_PAR;
5676 else if (pars && data) {
5677 data->flags |= SF_HAS_PAR;
5678 data->flags &= ~SF_IN_PAR;
5680 if (flags & SCF_DO_STCLASS_OR)
5681 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5682 if (flags & SCF_TRIE_RESTUDY)
5683 data->flags |= SCF_TRIE_RESTUDY;
5685 DEBUG_STUDYDATA("post-fin:",data,depth);
5688 SSize_t final_minlen= min < stopmin ? min : stopmin;
5690 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5691 if (final_minlen > SSize_t_MAX - delta)
5692 RExC_maxlen = SSize_t_MAX;
5693 else if (RExC_maxlen < final_minlen + delta)
5694 RExC_maxlen = final_minlen + delta;
5696 return final_minlen;
5698 NOT_REACHED; /* NOTREACHED */
5702 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5704 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5706 PERL_ARGS_ASSERT_ADD_DATA;
5708 Renewc(RExC_rxi->data,
5709 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5710 char, struct reg_data);
5712 Renew(RExC_rxi->data->what, count + n, U8);
5714 Newx(RExC_rxi->data->what, n, U8);
5715 RExC_rxi->data->count = count + n;
5716 Copy(s, RExC_rxi->data->what + count, n, U8);
5720 /*XXX: todo make this not included in a non debugging perl, but appears to be
5721 * used anyway there, in 'use re' */
5722 #ifndef PERL_IN_XSUB_RE
5724 Perl_reginitcolors(pTHX)
5726 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5728 char *t = savepv(s);
5732 t = strchr(t, '\t');
5738 PL_colors[i] = t = (char *)"";
5743 PL_colors[i++] = (char *)"";
5750 #ifdef TRIE_STUDY_OPT
5751 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5754 (data.flags & SCF_TRIE_RESTUDY) \
5762 #define CHECK_RESTUDY_GOTO_butfirst
5766 * pregcomp - compile a regular expression into internal code
5768 * Decides which engine's compiler to call based on the hint currently in
5772 #ifndef PERL_IN_XSUB_RE
5774 /* return the currently in-scope regex engine (or the default if none) */
5776 regexp_engine const *
5777 Perl_current_re_engine(pTHX)
5779 if (IN_PERL_COMPILETIME) {
5780 HV * const table = GvHV(PL_hintgv);
5783 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5784 return &PL_core_reg_engine;
5785 ptr = hv_fetchs(table, "regcomp", FALSE);
5786 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5787 return &PL_core_reg_engine;
5788 return INT2PTR(regexp_engine*,SvIV(*ptr));
5792 if (!PL_curcop->cop_hints_hash)
5793 return &PL_core_reg_engine;
5794 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5795 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5796 return &PL_core_reg_engine;
5797 return INT2PTR(regexp_engine*,SvIV(ptr));
5803 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5805 regexp_engine const *eng = current_re_engine();
5806 GET_RE_DEBUG_FLAGS_DECL;
5808 PERL_ARGS_ASSERT_PREGCOMP;
5810 /* Dispatch a request to compile a regexp to correct regexp engine. */
5812 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5815 return CALLREGCOMP_ENG(eng, pattern, flags);
5819 /* public(ish) entry point for the perl core's own regex compiling code.
5820 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5821 * pattern rather than a list of OPs, and uses the internal engine rather
5822 * than the current one */
5825 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5827 SV *pat = pattern; /* defeat constness! */
5828 PERL_ARGS_ASSERT_RE_COMPILE;
5829 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5830 #ifdef PERL_IN_XSUB_RE
5833 &PL_core_reg_engine,
5835 NULL, NULL, rx_flags, 0);
5839 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5840 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5841 * point to the realloced string and length.
5843 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5847 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5848 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5850 U8 *const src = (U8*)*pat_p;
5855 GET_RE_DEBUG_FLAGS_DECL;
5857 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5858 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5860 Newx(dst, *plen_p * 2 + 1, U8);
5863 while (s < *plen_p) {
5864 append_utf8_from_native_byte(src[s], &d);
5865 if (n < num_code_blocks) {
5866 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5867 pRExC_state->code_blocks[n].start = d - dst - 1;
5868 assert(*(d - 1) == '(');
5871 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5872 pRExC_state->code_blocks[n].end = d - dst - 1;
5873 assert(*(d - 1) == ')');
5882 *pat_p = (char*) dst;
5884 RExC_orig_utf8 = RExC_utf8 = 1;
5889 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5890 * while recording any code block indices, and handling overloading,
5891 * nested qr// objects etc. If pat is null, it will allocate a new
5892 * string, or just return the first arg, if there's only one.
5894 * Returns the malloced/updated pat.
5895 * patternp and pat_count is the array of SVs to be concatted;
5896 * oplist is the optional list of ops that generated the SVs;
5897 * recompile_p is a pointer to a boolean that will be set if
5898 * the regex will need to be recompiled.
5899 * delim, if non-null is an SV that will be inserted between each element
5903 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5904 SV *pat, SV ** const patternp, int pat_count,
5905 OP *oplist, bool *recompile_p, SV *delim)
5909 bool use_delim = FALSE;
5910 bool alloced = FALSE;
5912 /* if we know we have at least two args, create an empty string,
5913 * then concatenate args to that. For no args, return an empty string */
5914 if (!pat && pat_count != 1) {
5920 for (svp = patternp; svp < patternp + pat_count; svp++) {
5923 STRLEN orig_patlen = 0;
5925 SV *msv = use_delim ? delim : *svp;
5926 if (!msv) msv = &PL_sv_undef;
5928 /* if we've got a delimiter, we go round the loop twice for each
5929 * svp slot (except the last), using the delimiter the second
5938 if (SvTYPE(msv) == SVt_PVAV) {
5939 /* we've encountered an interpolated array within
5940 * the pattern, e.g. /...@a..../. Expand the list of elements,
5941 * then recursively append elements.
5942 * The code in this block is based on S_pushav() */
5944 AV *const av = (AV*)msv;
5945 const SSize_t maxarg = AvFILL(av) + 1;
5949 assert(oplist->op_type == OP_PADAV
5950 || oplist->op_type == OP_RV2AV);
5951 oplist = OpSIBLING(oplist);
5954 if (SvRMAGICAL(av)) {
5957 Newx(array, maxarg, SV*);
5959 for (i=0; i < maxarg; i++) {
5960 SV ** const svp = av_fetch(av, i, FALSE);
5961 array[i] = svp ? *svp : &PL_sv_undef;
5965 array = AvARRAY(av);
5967 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5968 array, maxarg, NULL, recompile_p,
5970 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5976 /* we make the assumption here that each op in the list of
5977 * op_siblings maps to one SV pushed onto the stack,
5978 * except for code blocks, with have both an OP_NULL and
5980 * This allows us to match up the list of SVs against the
5981 * list of OPs to find the next code block.
5983 * Note that PUSHMARK PADSV PADSV ..
5985 * PADRANGE PADSV PADSV ..
5986 * so the alignment still works. */
5989 if (oplist->op_type == OP_NULL
5990 && (oplist->op_flags & OPf_SPECIAL))
5992 assert(n < pRExC_state->num_code_blocks);
5993 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5994 pRExC_state->code_blocks[n].block = oplist;
5995 pRExC_state->code_blocks[n].src_regex = NULL;
5998 oplist = OpSIBLING(oplist); /* skip CONST */
6001 oplist = OpSIBLING(oplist);;
6004 /* apply magic and QR overloading to arg */
6007 if (SvROK(msv) && SvAMAGIC(msv)) {
6008 SV *sv = AMG_CALLunary(msv, regexp_amg);
6012 if (SvTYPE(sv) != SVt_REGEXP)
6013 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
6018 /* try concatenation overload ... */
6019 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
6020 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
6023 /* overloading involved: all bets are off over literal
6024 * code. Pretend we haven't seen it */
6025 pRExC_state->num_code_blocks -= n;
6029 /* ... or failing that, try "" overload */
6030 while (SvAMAGIC(msv)
6031 && (sv = AMG_CALLunary(msv, string_amg))
6035 && SvRV(msv) == SvRV(sv))
6040 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
6044 /* this is a partially unrolled
6045 * sv_catsv_nomg(pat, msv);
6046 * that allows us to adjust code block indices if
6049 char *dst = SvPV_force_nomg(pat, dlen);
6051 if (SvUTF8(msv) && !SvUTF8(pat)) {
6052 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6053 sv_setpvn(pat, dst, dlen);
6056 sv_catsv_nomg(pat, msv);
6063 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6066 /* extract any code blocks within any embedded qr//'s */
6067 if (rx && SvTYPE(rx) == SVt_REGEXP
6068 && RX_ENGINE((REGEXP*)rx)->op_comp)
6071 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6072 if (ri->num_code_blocks) {
6074 /* the presence of an embedded qr// with code means
6075 * we should always recompile: the text of the
6076 * qr// may not have changed, but it may be a
6077 * different closure than last time */
6079 Renew(pRExC_state->code_blocks,
6080 pRExC_state->num_code_blocks + ri->num_code_blocks,
6081 struct reg_code_block);
6082 pRExC_state->num_code_blocks += ri->num_code_blocks;
6084 for (i=0; i < ri->num_code_blocks; i++) {
6085 struct reg_code_block *src, *dst;
6086 STRLEN offset = orig_patlen
6087 + ReANY((REGEXP *)rx)->pre_prefix;
6088 assert(n < pRExC_state->num_code_blocks);
6089 src = &ri->code_blocks[i];
6090 dst = &pRExC_state->code_blocks[n];
6091 dst->start = src->start + offset;
6092 dst->end = src->end + offset;
6093 dst->block = src->block;
6094 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6103 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6112 /* see if there are any run-time code blocks in the pattern.
6113 * False positives are allowed */
6116 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6117 char *pat, STRLEN plen)
6122 PERL_UNUSED_CONTEXT;
6124 for (s = 0; s < plen; s++) {
6125 if (n < pRExC_state->num_code_blocks
6126 && s == pRExC_state->code_blocks[n].start)
6128 s = pRExC_state->code_blocks[n].end;
6132 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6134 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6136 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6143 /* Handle run-time code blocks. We will already have compiled any direct
6144 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6145 * copy of it, but with any literal code blocks blanked out and
6146 * appropriate chars escaped; then feed it into
6148 * eval "qr'modified_pattern'"
6152 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6156 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6158 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6159 * and merge them with any code blocks of the original regexp.
6161 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6162 * instead, just save the qr and return FALSE; this tells our caller that
6163 * the original pattern needs upgrading to utf8.
6167 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6168 char *pat, STRLEN plen)
6172 GET_RE_DEBUG_FLAGS_DECL;
6174 if (pRExC_state->runtime_code_qr) {
6175 /* this is the second time we've been called; this should
6176 * only happen if the main pattern got upgraded to utf8
6177 * during compilation; re-use the qr we compiled first time
6178 * round (which should be utf8 too)
6180 qr = pRExC_state->runtime_code_qr;
6181 pRExC_state->runtime_code_qr = NULL;
6182 assert(RExC_utf8 && SvUTF8(qr));
6188 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6192 /* determine how many extra chars we need for ' and \ escaping */
6193 for (s = 0; s < plen; s++) {
6194 if (pat[s] == '\'' || pat[s] == '\\')
6198 Newx(newpat, newlen, char);
6200 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6202 for (s = 0; s < plen; s++) {
6203 if (n < pRExC_state->num_code_blocks
6204 && s == pRExC_state->code_blocks[n].start)
6206 /* blank out literal code block */
6207 assert(pat[s] == '(');
6208 while (s <= pRExC_state->code_blocks[n].end) {
6216 if (pat[s] == '\'' || pat[s] == '\\')
6221 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6225 PerlIO_printf(Perl_debug_log,
6226 "%sre-parsing pattern for runtime code:%s %s\n",
6227 PL_colors[4],PL_colors[5],newpat);
6230 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6236 PUSHSTACKi(PERLSI_REQUIRE);
6237 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6238 * parsing qr''; normally only q'' does this. It also alters
6240 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6241 SvREFCNT_dec_NN(sv);
6246 SV * const errsv = ERRSV;
6247 if (SvTRUE_NN(errsv))
6249 Safefree(pRExC_state->code_blocks);
6250 /* use croak_sv ? */
6251 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6254 assert(SvROK(qr_ref));
6256 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6257 /* the leaving below frees the tmp qr_ref.
6258 * Give qr a life of its own */
6266 if (!RExC_utf8 && SvUTF8(qr)) {
6267 /* first time through; the pattern got upgraded; save the
6268 * qr for the next time through */
6269 assert(!pRExC_state->runtime_code_qr);
6270 pRExC_state->runtime_code_qr = qr;
6275 /* extract any code blocks within the returned qr// */
6278 /* merge the main (r1) and run-time (r2) code blocks into one */
6280 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6281 struct reg_code_block *new_block, *dst;
6282 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6285 if (!r2->num_code_blocks) /* we guessed wrong */
6287 SvREFCNT_dec_NN(qr);
6292 r1->num_code_blocks + r2->num_code_blocks,
6293 struct reg_code_block);
6296 while ( i1 < r1->num_code_blocks
6297 || i2 < r2->num_code_blocks)
6299 struct reg_code_block *src;
6302 if (i1 == r1->num_code_blocks) {
6303 src = &r2->code_blocks[i2++];
6306 else if (i2 == r2->num_code_blocks)
6307 src = &r1->code_blocks[i1++];
6308 else if ( r1->code_blocks[i1].start
6309 < r2->code_blocks[i2].start)
6311 src = &r1->code_blocks[i1++];
6312 assert(src->end < r2->code_blocks[i2].start);
6315 assert( r1->code_blocks[i1].start
6316 > r2->code_blocks[i2].start);
6317 src = &r2->code_blocks[i2++];
6319 assert(src->end < r1->code_blocks[i1].start);
6322 assert(pat[src->start] == '(');
6323 assert(pat[src->end] == ')');
6324 dst->start = src->start;
6325 dst->end = src->end;
6326 dst->block = src->block;
6327 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6331 r1->num_code_blocks += r2->num_code_blocks;
6332 Safefree(r1->code_blocks);
6333 r1->code_blocks = new_block;
6336 SvREFCNT_dec_NN(qr);
6342 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6343 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6344 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6345 STRLEN longest_length, bool eol, bool meol)
6347 /* This is the common code for setting up the floating and fixed length
6348 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6349 * as to whether succeeded or not */
6354 if (! (longest_length
6355 || (eol /* Can't have SEOL and MULTI */
6356 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6358 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6359 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6364 /* copy the information about the longest from the reg_scan_data
6365 over to the program. */
6366 if (SvUTF8(sv_longest)) {
6367 *rx_utf8 = sv_longest;
6370 *rx_substr = sv_longest;
6373 /* end_shift is how many chars that must be matched that
6374 follow this item. We calculate it ahead of time as once the
6375 lookbehind offset is added in we lose the ability to correctly
6377 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6378 *rx_end_shift = ml - offset
6379 - longest_length + (SvTAIL(sv_longest) != 0)
6382 t = (eol/* Can't have SEOL and MULTI */
6383 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6384 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6390 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6391 * regular expression into internal code.
6392 * The pattern may be passed either as:
6393 * a list of SVs (patternp plus pat_count)
6394 * a list of OPs (expr)
6395 * If both are passed, the SV list is used, but the OP list indicates
6396 * which SVs are actually pre-compiled code blocks
6398 * The SVs in the list have magic and qr overloading applied to them (and
6399 * the list may be modified in-place with replacement SVs in the latter
6402 * If the pattern hasn't changed from old_re, then old_re will be
6405 * eng is the current engine. If that engine has an op_comp method, then
6406 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6407 * do the initial concatenation of arguments and pass on to the external
6410 * If is_bare_re is not null, set it to a boolean indicating whether the
6411 * arg list reduced (after overloading) to a single bare regex which has
6412 * been returned (i.e. /$qr/).
6414 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6416 * pm_flags contains the PMf_* flags, typically based on those from the
6417 * pm_flags field of the related PMOP. Currently we're only interested in
6418 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6420 * We can't allocate space until we know how big the compiled form will be,
6421 * but we can't compile it (and thus know how big it is) until we've got a
6422 * place to put the code. So we cheat: we compile it twice, once with code
6423 * generation turned off and size counting turned on, and once "for real".
6424 * This also means that we don't allocate space until we are sure that the
6425 * thing really will compile successfully, and we never have to move the
6426 * code and thus invalidate pointers into it. (Note that it has to be in
6427 * one piece because free() must be able to free it all.) [NB: not true in perl]
6429 * Beware that the optimization-preparation code in here knows about some
6430 * of the structure of the compiled regexp. [I'll say.]
6434 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6435 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6436 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6440 regexp_internal *ri;
6448 SV *code_blocksv = NULL;
6449 SV** new_patternp = patternp;
6451 /* these are all flags - maybe they should be turned
6452 * into a single int with different bit masks */
6453 I32 sawlookahead = 0;
6458 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6460 bool runtime_code = 0;
6462 RExC_state_t RExC_state;
6463 RExC_state_t * const pRExC_state = &RExC_state;
6464 #ifdef TRIE_STUDY_OPT
6466 RExC_state_t copyRExC_state;
6468 GET_RE_DEBUG_FLAGS_DECL;
6470 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6472 DEBUG_r(if (!PL_colorset) reginitcolors());
6474 /* Initialize these here instead of as-needed, as is quick and avoids
6475 * having to test them each time otherwise */
6476 if (! PL_AboveLatin1) {
6477 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6478 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6479 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6480 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6481 PL_HasMultiCharFold =
6482 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6484 /* This is calculated here, because the Perl program that generates the
6485 * static global ones doesn't currently have access to
6486 * NUM_ANYOF_CODE_POINTS */
6487 PL_InBitmap = _new_invlist(2);
6488 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6489 NUM_ANYOF_CODE_POINTS - 1);
6492 pRExC_state->code_blocks = NULL;
6493 pRExC_state->num_code_blocks = 0;
6496 *is_bare_re = FALSE;
6498 if (expr && (expr->op_type == OP_LIST ||
6499 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6500 /* allocate code_blocks if needed */
6504 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6505 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6506 ncode++; /* count of DO blocks */
6508 pRExC_state->num_code_blocks = ncode;
6509 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6514 /* compile-time pattern with just OP_CONSTs and DO blocks */
6519 /* find how many CONSTs there are */
6522 if (expr->op_type == OP_CONST)
6525 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6526 if (o->op_type == OP_CONST)
6530 /* fake up an SV array */
6532 assert(!new_patternp);
6533 Newx(new_patternp, n, SV*);
6534 SAVEFREEPV(new_patternp);
6538 if (expr->op_type == OP_CONST)
6539 new_patternp[n] = cSVOPx_sv(expr);
6541 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6542 if (o->op_type == OP_CONST)
6543 new_patternp[n++] = cSVOPo_sv;
6548 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6549 "Assembling pattern from %d elements%s\n", pat_count,
6550 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6552 /* set expr to the first arg op */
6554 if (pRExC_state->num_code_blocks
6555 && expr->op_type != OP_CONST)
6557 expr = cLISTOPx(expr)->op_first;
6558 assert( expr->op_type == OP_PUSHMARK
6559 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6560 || expr->op_type == OP_PADRANGE);
6561 expr = OpSIBLING(expr);
6564 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6565 expr, &recompile, NULL);
6567 /* handle bare (possibly after overloading) regex: foo =~ $re */
6572 if (SvTYPE(re) == SVt_REGEXP) {
6576 Safefree(pRExC_state->code_blocks);
6577 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6578 "Precompiled pattern%s\n",
6579 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6585 exp = SvPV_nomg(pat, plen);
6587 if (!eng->op_comp) {
6588 if ((SvUTF8(pat) && IN_BYTES)
6589 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6591 /* make a temporary copy; either to convert to bytes,
6592 * or to avoid repeating get-magic / overloaded stringify */
6593 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6594 (IN_BYTES ? 0 : SvUTF8(pat)));
6596 Safefree(pRExC_state->code_blocks);
6597 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6600 /* ignore the utf8ness if the pattern is 0 length */
6601 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6603 RExC_uni_semantics = 0;
6604 RExC_seen_unfolded_sharp_s = 0;
6605 RExC_contains_locale = 0;
6606 RExC_contains_i = 0;
6607 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6608 pRExC_state->runtime_code_qr = NULL;
6609 RExC_frame_head= NULL;
6610 RExC_frame_last= NULL;
6611 RExC_frame_count= 0;
6614 RExC_mysv1= sv_newmortal();
6615 RExC_mysv2= sv_newmortal();
6618 SV *dsv= sv_newmortal();
6619 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6620 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6621 PL_colors[4],PL_colors[5],s);
6625 /* we jump here if we have to recompile, e.g., from upgrading the pattern
6628 if ((pm_flags & PMf_USE_RE_EVAL)
6629 /* this second condition covers the non-regex literal case,
6630 * i.e. $foo =~ '(?{})'. */
6631 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6633 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6635 /* return old regex if pattern hasn't changed */
6636 /* XXX: note in the below we have to check the flags as well as the
6639 * Things get a touch tricky as we have to compare the utf8 flag
6640 * independently from the compile flags. */
6644 && !!RX_UTF8(old_re) == !!RExC_utf8
6645 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6646 && RX_PRECOMP(old_re)
6647 && RX_PRELEN(old_re) == plen
6648 && memEQ(RX_PRECOMP(old_re), exp, plen)
6649 && !runtime_code /* with runtime code, always recompile */ )
6651 Safefree(pRExC_state->code_blocks);
6655 rx_flags = orig_rx_flags;
6657 if (rx_flags & PMf_FOLD) {
6658 RExC_contains_i = 1;
6660 if ( initial_charset == REGEX_DEPENDS_CHARSET
6661 && (RExC_utf8 ||RExC_uni_semantics))
6664 /* Set to use unicode semantics if the pattern is in utf8 and has the
6665 * 'depends' charset specified, as it means unicode when utf8 */
6666 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6670 RExC_flags = rx_flags;
6671 RExC_pm_flags = pm_flags;
6674 if (TAINTING_get && TAINT_get)
6675 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6677 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6678 /* whoops, we have a non-utf8 pattern, whilst run-time code
6679 * got compiled as utf8. Try again with a utf8 pattern */
6680 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6681 pRExC_state->num_code_blocks);
6682 goto redo_first_pass;
6685 assert(!pRExC_state->runtime_code_qr);
6691 RExC_in_lookbehind = 0;
6692 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6694 RExC_override_recoding = 0;
6696 RExC_recode_x_to_native = 0;
6698 RExC_in_multi_char_class = 0;
6700 /* First pass: determine size, legality. */
6703 RExC_end = exp + plen;
6708 RExC_emit = (regnode *) &RExC_emit_dummy;
6709 RExC_whilem_seen = 0;
6710 RExC_open_parens = NULL;
6711 RExC_close_parens = NULL;
6713 RExC_paren_names = NULL;
6715 RExC_paren_name_list = NULL;
6717 RExC_recurse = NULL;
6718 RExC_study_chunk_recursed = NULL;
6719 RExC_study_chunk_recursed_bytes= 0;
6720 RExC_recurse_count = 0;
6721 pRExC_state->code_index = 0;
6724 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6726 RExC_lastparse=NULL;
6728 /* reg may croak on us, not giving us a chance to free
6729 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6730 need it to survive as long as the regexp (qr/(?{})/).
6731 We must check that code_blocksv is not already set, because we may
6732 have jumped back to restart the sizing pass. */
6733 if (pRExC_state->code_blocks && !code_blocksv) {
6734 code_blocksv = newSV_type(SVt_PV);
6735 SAVEFREESV(code_blocksv);
6736 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6737 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6739 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6740 /* It's possible to write a regexp in ascii that represents Unicode
6741 codepoints outside of the byte range, such as via \x{100}. If we
6742 detect such a sequence we have to convert the entire pattern to utf8
6743 and then recompile, as our sizing calculation will have been based
6744 on 1 byte == 1 character, but we will need to use utf8 to encode
6745 at least some part of the pattern, and therefore must convert the whole
6748 if (flags & RESTART_PASS1) {
6749 if (flags & NEED_UTF8) {
6750 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6751 pRExC_state->num_code_blocks);
6754 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6755 "Need to redo pass 1\n"));
6758 goto redo_first_pass;
6760 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6763 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6766 PerlIO_printf(Perl_debug_log,
6767 "Required size %"IVdf" nodes\n"
6768 "Starting second pass (creation)\n",
6771 RExC_lastparse=NULL;
6774 /* The first pass could have found things that force Unicode semantics */
6775 if ((RExC_utf8 || RExC_uni_semantics)
6776 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6778 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6781 /* Small enough for pointer-storage convention?
6782 If extralen==0, this means that we will not need long jumps. */
6783 if (RExC_size >= 0x10000L && RExC_extralen)
6784 RExC_size += RExC_extralen;
6787 if (RExC_whilem_seen > 15)
6788 RExC_whilem_seen = 15;
6790 /* Allocate space and zero-initialize. Note, the two step process
6791 of zeroing when in debug mode, thus anything assigned has to
6792 happen after that */
6793 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6795 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6796 char, regexp_internal);
6797 if ( r == NULL || ri == NULL )
6798 FAIL("Regexp out of space");
6800 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6801 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6804 /* bulk initialize base fields with 0. */
6805 Zero(ri, sizeof(regexp_internal), char);
6808 /* non-zero initialization begins here */
6811 r->extflags = rx_flags;
6812 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6814 if (pm_flags & PMf_IS_QR) {
6815 ri->code_blocks = pRExC_state->code_blocks;
6816 ri->num_code_blocks = pRExC_state->num_code_blocks;
6821 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6822 if (pRExC_state->code_blocks[n].src_regex)
6823 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6824 if(pRExC_state->code_blocks)
6825 SAVEFREEPV(pRExC_state->code_blocks); /* often null */
6829 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6830 bool has_charset = (get_regex_charset(r->extflags)
6831 != REGEX_DEPENDS_CHARSET);
6833 /* The caret is output if there are any defaults: if not all the STD
6834 * flags are set, or if no character set specifier is needed */
6836 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6838 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6839 == REG_RUN_ON_COMMENT_SEEN);
6840 U8 reganch = (U8)((r->extflags & RXf_PMf_STD_PMMOD)
6841 >> RXf_PMf_STD_PMMOD_SHIFT);
6842 const char *fptr = STD_PAT_MODS; /*"msixn"*/
6845 /* We output all the necessary flags; we never output a minus, as all
6846 * those are defaults, so are
6847 * covered by the caret */
6848 const STRLEN wraplen = plen + has_p + has_runon
6849 + has_default /* If needs a caret */
6850 + PL_bitcount[reganch] /* 1 char for each set standard flag */
6852 /* If needs a character set specifier */
6853 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6854 + (sizeof("(?:)") - 1);
6856 /* make sure PL_bitcount bounds not exceeded */
6857 assert(sizeof(STD_PAT_MODS) <= 8);
6859 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6860 r->xpv_len_u.xpvlenu_pv = p;
6862 SvFLAGS(rx) |= SVf_UTF8;
6865 /* If a default, cover it using the caret */
6867 *p++= DEFAULT_PAT_MOD;
6871 const char* const name = get_regex_charset_name(r->extflags, &len);
6872 Copy(name, p, len, char);
6876 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6879 while((ch = *fptr++)) {
6887 Copy(RExC_precomp, p, plen, char);
6888 assert ((RX_WRAPPED(rx) - p) < 16);
6889 r->pre_prefix = p - RX_WRAPPED(rx);
6895 SvCUR_set(rx, p - RX_WRAPPED(rx));
6899 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6901 /* setup various meta data about recursion, this all requires
6902 * RExC_npar to be correctly set, and a bit later on we clear it */
6903 if (RExC_seen & REG_RECURSE_SEEN) {
6904 Newxz(RExC_open_parens, RExC_npar,regnode *);
6905 SAVEFREEPV(RExC_open_parens);
6906 Newxz(RExC_close_parens,RExC_npar,regnode *);
6907 SAVEFREEPV(RExC_close_parens);
6909 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6910 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6911 * So its 1 if there are no parens. */
6912 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6913 ((RExC_npar & 0x07) != 0);
6914 Newx(RExC_study_chunk_recursed,
6915 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6916 SAVEFREEPV(RExC_study_chunk_recursed);
6919 /* Useful during FAIL. */
6920 #ifdef RE_TRACK_PATTERN_OFFSETS
6921 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6922 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6923 "%s %"UVuf" bytes for offset annotations.\n",
6924 ri->u.offsets ? "Got" : "Couldn't get",
6925 (UV)((2*RExC_size+1) * sizeof(U32))));
6927 SetProgLen(ri,RExC_size);
6932 /* Second pass: emit code. */
6933 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6934 RExC_pm_flags = pm_flags;
6936 RExC_end = exp + plen;
6939 RExC_emit_start = ri->program;
6940 RExC_emit = ri->program;
6941 RExC_emit_bound = ri->program + RExC_size + 1;
6942 pRExC_state->code_index = 0;
6944 *((char*) RExC_emit++) = (char) REG_MAGIC;
6945 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6947 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6949 /* XXXX To minimize changes to RE engine we always allocate
6950 3-units-long substrs field. */
6951 Newx(r->substrs, 1, struct reg_substr_data);
6952 if (RExC_recurse_count) {
6953 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6954 SAVEFREEPV(RExC_recurse);
6958 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6960 RExC_study_chunk_recursed_count= 0;
6962 Zero(r->substrs, 1, struct reg_substr_data);
6963 if (RExC_study_chunk_recursed) {
6964 Zero(RExC_study_chunk_recursed,
6965 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6969 #ifdef TRIE_STUDY_OPT
6971 StructCopy(&zero_scan_data, &data, scan_data_t);
6972 copyRExC_state = RExC_state;
6975 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6977 RExC_state = copyRExC_state;
6978 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6979 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6981 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6982 StructCopy(&zero_scan_data, &data, scan_data_t);
6985 StructCopy(&zero_scan_data, &data, scan_data_t);
6988 /* Dig out information for optimizations. */
6989 r->extflags = RExC_flags; /* was pm_op */
6990 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6993 SvUTF8_on(rx); /* Unicode in it? */
6994 ri->regstclass = NULL;
6995 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
6996 r->intflags |= PREGf_NAUGHTY;
6997 scan = ri->program + 1; /* First BRANCH. */
6999 /* testing for BRANCH here tells us whether there is "must appear"
7000 data in the pattern. If there is then we can use it for optimisations */
7001 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
7004 STRLEN longest_float_length, longest_fixed_length;
7005 regnode_ssc ch_class; /* pointed to by data */
7007 SSize_t last_close = 0; /* pointed to by data */
7008 regnode *first= scan;
7009 regnode *first_next= regnext(first);
7011 * Skip introductions and multiplicators >= 1
7012 * so that we can extract the 'meat' of the pattern that must
7013 * match in the large if() sequence following.
7014 * NOTE that EXACT is NOT covered here, as it is normally
7015 * picked up by the optimiser separately.
7017 * This is unfortunate as the optimiser isnt handling lookahead
7018 * properly currently.
7021 while ((OP(first) == OPEN && (sawopen = 1)) ||
7022 /* An OR of *one* alternative - should not happen now. */
7023 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
7024 /* for now we can't handle lookbehind IFMATCH*/
7025 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
7026 (OP(first) == PLUS) ||
7027 (OP(first) == MINMOD) ||
7028 /* An {n,m} with n>0 */
7029 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
7030 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
7033 * the only op that could be a regnode is PLUS, all the rest
7034 * will be regnode_1 or regnode_2.
7036 * (yves doesn't think this is true)
7038 if (OP(first) == PLUS)
7041 if (OP(first) == MINMOD)
7043 first += regarglen[OP(first)];
7045 first = NEXTOPER(first);
7046 first_next= regnext(first);
7049 /* Starting-point info. */
7051 DEBUG_PEEP("first:",first,0);
7052 /* Ignore EXACT as we deal with it later. */
7053 if (PL_regkind[OP(first)] == EXACT) {
7054 if (OP(first) == EXACT || OP(first) == EXACTL)
7055 NOOP; /* Empty, get anchored substr later. */
7057 ri->regstclass = first;
7060 else if (PL_regkind[OP(first)] == TRIE &&
7061 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7063 /* this can happen only on restudy */
7064 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7067 else if (REGNODE_SIMPLE(OP(first)))
7068 ri->regstclass = first;
7069 else if (PL_regkind[OP(first)] == BOUND ||
7070 PL_regkind[OP(first)] == NBOUND)
7071 ri->regstclass = first;
7072 else if (PL_regkind[OP(first)] == BOL) {
7073 r->intflags |= (OP(first) == MBOL
7076 first = NEXTOPER(first);
7079 else if (OP(first) == GPOS) {
7080 r->intflags |= PREGf_ANCH_GPOS;
7081 first = NEXTOPER(first);
7084 else if ((!sawopen || !RExC_sawback) &&
7086 (OP(first) == STAR &&
7087 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7088 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7090 /* turn .* into ^.* with an implied $*=1 */
7092 (OP(NEXTOPER(first)) == REG_ANY)
7095 r->intflags |= (type | PREGf_IMPLICIT);
7096 first = NEXTOPER(first);
7099 if (sawplus && !sawminmod && !sawlookahead
7100 && (!sawopen || !RExC_sawback)
7101 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7102 /* x+ must match at the 1st pos of run of x's */
7103 r->intflags |= PREGf_SKIP;
7105 /* Scan is after the zeroth branch, first is atomic matcher. */
7106 #ifdef TRIE_STUDY_OPT
7109 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7110 (IV)(first - scan + 1))
7114 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7115 (IV)(first - scan + 1))
7121 * If there's something expensive in the r.e., find the
7122 * longest literal string that must appear and make it the
7123 * regmust. Resolve ties in favor of later strings, since
7124 * the regstart check works with the beginning of the r.e.
7125 * and avoiding duplication strengthens checking. Not a
7126 * strong reason, but sufficient in the absence of others.
7127 * [Now we resolve ties in favor of the earlier string if
7128 * it happens that c_offset_min has been invalidated, since the
7129 * earlier string may buy us something the later one won't.]
7132 data.longest_fixed = newSVpvs("");
7133 data.longest_float = newSVpvs("");
7134 data.last_found = newSVpvs("");
7135 data.longest = &(data.longest_fixed);
7136 ENTER_with_name("study_chunk");
7137 SAVEFREESV(data.longest_fixed);
7138 SAVEFREESV(data.longest_float);
7139 SAVEFREESV(data.last_found);
7141 if (!ri->regstclass) {
7142 ssc_init(pRExC_state, &ch_class);
7143 data.start_class = &ch_class;
7144 stclass_flag = SCF_DO_STCLASS_AND;
7145 } else /* XXXX Check for BOUND? */
7147 data.last_closep = &last_close;
7150 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7151 scan + RExC_size, /* Up to end */
7153 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7154 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7158 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7161 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7162 && data.last_start_min == 0 && data.last_end > 0
7163 && !RExC_seen_zerolen
7164 && !(RExC_seen & REG_VERBARG_SEEN)
7165 && !(RExC_seen & REG_GPOS_SEEN)
7167 r->extflags |= RXf_CHECK_ALL;
7169 scan_commit(pRExC_state, &data,&minlen,0);
7171 longest_float_length = CHR_SVLEN(data.longest_float);
7173 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7174 && data.offset_fixed == data.offset_float_min
7175 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7176 && S_setup_longest (aTHX_ pRExC_state,
7180 &(r->float_end_shift),
7181 data.lookbehind_float,
7182 data.offset_float_min,
7184 longest_float_length,
7185 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7186 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7188 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7189 r->float_max_offset = data.offset_float_max;
7190 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7191 r->float_max_offset -= data.lookbehind_float;
7192 SvREFCNT_inc_simple_void_NN(data.longest_float);
7195 r->float_substr = r->float_utf8 = NULL;
7196 longest_float_length = 0;
7199 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7201 if (S_setup_longest (aTHX_ pRExC_state,
7203 &(r->anchored_utf8),
7204 &(r->anchored_substr),
7205 &(r->anchored_end_shift),
7206 data.lookbehind_fixed,
7209 longest_fixed_length,
7210 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7211 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7213 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7214 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7217 r->anchored_substr = r->anchored_utf8 = NULL;
7218 longest_fixed_length = 0;
7220 LEAVE_with_name("study_chunk");
7223 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7224 ri->regstclass = NULL;
7226 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7228 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7229 && is_ssc_worth_it(pRExC_state, data.start_class))
7231 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7233 ssc_finalize(pRExC_state, data.start_class);
7235 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7236 StructCopy(data.start_class,
7237 (regnode_ssc*)RExC_rxi->data->data[n],
7239 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7240 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7241 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7242 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7243 PerlIO_printf(Perl_debug_log,
7244 "synthetic stclass \"%s\".\n",
7245 SvPVX_const(sv));});
7246 data.start_class = NULL;
7249 /* A temporary algorithm prefers floated substr to fixed one to dig
7251 if (longest_fixed_length > longest_float_length) {
7252 r->substrs->check_ix = 0;
7253 r->check_end_shift = r->anchored_end_shift;
7254 r->check_substr = r->anchored_substr;
7255 r->check_utf8 = r->anchored_utf8;
7256 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7257 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7258 r->intflags |= PREGf_NOSCAN;
7261 r->substrs->check_ix = 1;
7262 r->check_end_shift = r->float_end_shift;
7263 r->check_substr = r->float_substr;
7264 r->check_utf8 = r->float_utf8;
7265 r->check_offset_min = r->float_min_offset;
7266 r->check_offset_max = r->float_max_offset;
7268 if ((r->check_substr || r->check_utf8) ) {
7269 r->extflags |= RXf_USE_INTUIT;
7270 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7271 r->extflags |= RXf_INTUIT_TAIL;
7273 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7275 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7276 if ( (STRLEN)minlen < longest_float_length )
7277 minlen= longest_float_length;
7278 if ( (STRLEN)minlen < longest_fixed_length )
7279 minlen= longest_fixed_length;
7283 /* Several toplevels. Best we can is to set minlen. */
7285 regnode_ssc ch_class;
7286 SSize_t last_close = 0;
7288 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7290 scan = ri->program + 1;
7291 ssc_init(pRExC_state, &ch_class);
7292 data.start_class = &ch_class;
7293 data.last_closep = &last_close;
7296 minlen = study_chunk(pRExC_state,
7297 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7298 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7299 ? SCF_TRIE_DOING_RESTUDY
7303 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7305 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7306 = r->float_substr = r->float_utf8 = NULL;
7308 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7309 && is_ssc_worth_it(pRExC_state, data.start_class))
7311 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7313 ssc_finalize(pRExC_state, data.start_class);
7315 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7316 StructCopy(data.start_class,
7317 (regnode_ssc*)RExC_rxi->data->data[n],
7319 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7320 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7321 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7322 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7323 PerlIO_printf(Perl_debug_log,
7324 "synthetic stclass \"%s\".\n",
7325 SvPVX_const(sv));});
7326 data.start_class = NULL;
7330 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7331 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7332 r->maxlen = REG_INFTY;
7335 r->maxlen = RExC_maxlen;
7338 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7339 the "real" pattern. */
7341 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%"IVdf"\n",
7342 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7344 r->minlenret = minlen;
7345 if (r->minlen < minlen)
7348 if (RExC_seen & REG_GPOS_SEEN)
7349 r->intflags |= PREGf_GPOS_SEEN;
7350 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7351 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7353 if (pRExC_state->num_code_blocks)
7354 r->extflags |= RXf_EVAL_SEEN;
7355 if (RExC_seen & REG_VERBARG_SEEN)
7357 r->intflags |= PREGf_VERBARG_SEEN;
7358 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7360 if (RExC_seen & REG_CUTGROUP_SEEN)
7361 r->intflags |= PREGf_CUTGROUP_SEEN;
7362 if (pm_flags & PMf_USE_RE_EVAL)
7363 r->intflags |= PREGf_USE_RE_EVAL;
7364 if (RExC_paren_names)
7365 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7367 RXp_PAREN_NAMES(r) = NULL;
7369 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7370 * so it can be used in pp.c */
7371 if (r->intflags & PREGf_ANCH)
7372 r->extflags |= RXf_IS_ANCHORED;
7376 /* this is used to identify "special" patterns that might result
7377 * in Perl NOT calling the regex engine and instead doing the match "itself",
7378 * particularly special cases in split//. By having the regex compiler
7379 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7380 * we avoid weird issues with equivalent patterns resulting in different behavior,
7381 * AND we allow non Perl engines to get the same optimizations by the setting the
7382 * flags appropriately - Yves */
7383 regnode *first = ri->program + 1;
7385 regnode *next = regnext(first);
7388 if (PL_regkind[fop] == NOTHING && nop == END)
7389 r->extflags |= RXf_NULL;
7390 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7391 /* when fop is SBOL first->flags will be true only when it was
7392 * produced by parsing /\A/, and not when parsing /^/. This is
7393 * very important for the split code as there we want to
7394 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7395 * See rt #122761 for more details. -- Yves */
7396 r->extflags |= RXf_START_ONLY;
7397 else if (fop == PLUS
7398 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7400 r->extflags |= RXf_WHITE;
7401 else if ( r->extflags & RXf_SPLIT
7402 && (fop == EXACT || fop == EXACTL)
7403 && STR_LEN(first) == 1
7404 && *(STRING(first)) == ' '
7406 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7410 if (RExC_contains_locale) {
7411 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7415 if (RExC_paren_names) {
7416 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7417 ri->data->data[ri->name_list_idx]
7418 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7421 ri->name_list_idx = 0;
7423 if (RExC_recurse_count) {
7424 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7425 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7426 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7429 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7430 /* assume we don't need to swap parens around before we match */
7432 PerlIO_printf(Perl_debug_log,"study_chunk_recursed_count: %lu\n",
7433 (unsigned long)RExC_study_chunk_recursed_count);
7437 PerlIO_printf(Perl_debug_log,"Final program:\n");
7440 #ifdef RE_TRACK_PATTERN_OFFSETS
7441 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7442 const STRLEN len = ri->u.offsets[0];
7444 GET_RE_DEBUG_FLAGS_DECL;
7445 PerlIO_printf(Perl_debug_log,
7446 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7447 for (i = 1; i <= len; i++) {
7448 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7449 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7450 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7452 PerlIO_printf(Perl_debug_log, "\n");
7457 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7458 * by setting the regexp SV to readonly-only instead. If the
7459 * pattern's been recompiled, the USEDness should remain. */
7460 if (old_re && SvREADONLY(old_re))
7468 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7471 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7473 PERL_UNUSED_ARG(value);
7475 if (flags & RXapif_FETCH) {
7476 return reg_named_buff_fetch(rx, key, flags);
7477 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7478 Perl_croak_no_modify();
7480 } else if (flags & RXapif_EXISTS) {
7481 return reg_named_buff_exists(rx, key, flags)
7484 } else if (flags & RXapif_REGNAMES) {
7485 return reg_named_buff_all(rx, flags);
7486 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7487 return reg_named_buff_scalar(rx, flags);
7489 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7495 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7498 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7499 PERL_UNUSED_ARG(lastkey);
7501 if (flags & RXapif_FIRSTKEY)
7502 return reg_named_buff_firstkey(rx, flags);
7503 else if (flags & RXapif_NEXTKEY)
7504 return reg_named_buff_nextkey(rx, flags);
7506 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7513 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7516 AV *retarray = NULL;
7518 struct regexp *const rx = ReANY(r);
7520 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7522 if (flags & RXapif_ALL)
7525 if (rx && RXp_PAREN_NAMES(rx)) {
7526 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7529 SV* sv_dat=HeVAL(he_str);
7530 I32 *nums=(I32*)SvPVX(sv_dat);
7531 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7532 if ((I32)(rx->nparens) >= nums[i]
7533 && rx->offs[nums[i]].start != -1
7534 && rx->offs[nums[i]].end != -1)
7537 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7542 ret = newSVsv(&PL_sv_undef);
7545 av_push(retarray, ret);
7548 return newRV_noinc(MUTABLE_SV(retarray));
7555 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7558 struct regexp *const rx = ReANY(r);
7560 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7562 if (rx && RXp_PAREN_NAMES(rx)) {
7563 if (flags & RXapif_ALL) {
7564 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7566 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7568 SvREFCNT_dec_NN(sv);
7580 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7582 struct regexp *const rx = ReANY(r);
7584 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7586 if ( rx && RXp_PAREN_NAMES(rx) ) {
7587 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7589 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7596 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7598 struct regexp *const rx = ReANY(r);
7599 GET_RE_DEBUG_FLAGS_DECL;
7601 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7603 if (rx && RXp_PAREN_NAMES(rx)) {
7604 HV *hv = RXp_PAREN_NAMES(rx);
7606 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7609 SV* sv_dat = HeVAL(temphe);
7610 I32 *nums = (I32*)SvPVX(sv_dat);
7611 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7612 if ((I32)(rx->lastparen) >= nums[i] &&
7613 rx->offs[nums[i]].start != -1 &&
7614 rx->offs[nums[i]].end != -1)
7620 if (parno || flags & RXapif_ALL) {
7621 return newSVhek(HeKEY_hek(temphe));
7629 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7634 struct regexp *const rx = ReANY(r);
7636 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7638 if (rx && RXp_PAREN_NAMES(rx)) {
7639 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7640 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7641 } else if (flags & RXapif_ONE) {
7642 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7643 av = MUTABLE_AV(SvRV(ret));
7644 length = av_tindex(av);
7645 SvREFCNT_dec_NN(ret);
7646 return newSViv(length + 1);
7648 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7653 return &PL_sv_undef;
7657 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7659 struct regexp *const rx = ReANY(r);
7662 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7664 if (rx && RXp_PAREN_NAMES(rx)) {
7665 HV *hv= RXp_PAREN_NAMES(rx);
7667 (void)hv_iterinit(hv);
7668 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7671 SV* sv_dat = HeVAL(temphe);
7672 I32 *nums = (I32*)SvPVX(sv_dat);
7673 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7674 if ((I32)(rx->lastparen) >= nums[i] &&
7675 rx->offs[nums[i]].start != -1 &&
7676 rx->offs[nums[i]].end != -1)
7682 if (parno || flags & RXapif_ALL) {
7683 av_push(av, newSVhek(HeKEY_hek(temphe)));
7688 return newRV_noinc(MUTABLE_SV(av));
7692 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7695 struct regexp *const rx = ReANY(r);
7701 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7703 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7704 || n == RX_BUFF_IDX_CARET_FULLMATCH
7705 || n == RX_BUFF_IDX_CARET_POSTMATCH
7708 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7710 /* on something like
7713 * the KEEPCOPY is set on the PMOP rather than the regex */
7714 if (PL_curpm && r == PM_GETRE(PL_curpm))
7715 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7724 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7725 /* no need to distinguish between them any more */
7726 n = RX_BUFF_IDX_FULLMATCH;
7728 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7729 && rx->offs[0].start != -1)
7731 /* $`, ${^PREMATCH} */
7732 i = rx->offs[0].start;
7736 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7737 && rx->offs[0].end != -1)
7739 /* $', ${^POSTMATCH} */
7740 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7741 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7744 if ( 0 <= n && n <= (I32)rx->nparens &&
7745 (s1 = rx->offs[n].start) != -1 &&
7746 (t1 = rx->offs[n].end) != -1)
7748 /* $&, ${^MATCH}, $1 ... */
7750 s = rx->subbeg + s1 - rx->suboffset;
7755 assert(s >= rx->subbeg);
7756 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7758 #ifdef NO_TAINT_SUPPORT
7759 sv_setpvn(sv, s, i);
7761 const int oldtainted = TAINT_get;
7763 sv_setpvn(sv, s, i);
7764 TAINT_set(oldtainted);
7766 if (RXp_MATCH_UTF8(rx))
7771 if (RXp_MATCH_TAINTED(rx)) {
7772 if (SvTYPE(sv) >= SVt_PVMG) {
7773 MAGIC* const mg = SvMAGIC(sv);
7776 SvMAGIC_set(sv, mg->mg_moremagic);
7778 if ((mgt = SvMAGIC(sv))) {
7779 mg->mg_moremagic = mgt;
7780 SvMAGIC_set(sv, mg);
7791 sv_setsv(sv,&PL_sv_undef);
7797 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7798 SV const * const value)
7800 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7802 PERL_UNUSED_ARG(rx);
7803 PERL_UNUSED_ARG(paren);
7804 PERL_UNUSED_ARG(value);
7807 Perl_croak_no_modify();
7811 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7814 struct regexp *const rx = ReANY(r);
7818 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7820 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7821 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7822 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7825 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7827 /* on something like
7830 * the KEEPCOPY is set on the PMOP rather than the regex */
7831 if (PL_curpm && r == PM_GETRE(PL_curpm))
7832 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7838 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7840 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7841 case RX_BUFF_IDX_PREMATCH: /* $` */
7842 if (rx->offs[0].start != -1) {
7843 i = rx->offs[0].start;
7852 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7853 case RX_BUFF_IDX_POSTMATCH: /* $' */
7854 if (rx->offs[0].end != -1) {
7855 i = rx->sublen - rx->offs[0].end;
7857 s1 = rx->offs[0].end;
7864 default: /* $& / ${^MATCH}, $1, $2, ... */
7865 if (paren <= (I32)rx->nparens &&
7866 (s1 = rx->offs[paren].start) != -1 &&
7867 (t1 = rx->offs[paren].end) != -1)
7873 if (ckWARN(WARN_UNINITIALIZED))
7874 report_uninit((const SV *)sv);
7879 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7880 const char * const s = rx->subbeg - rx->suboffset + s1;
7885 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7892 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7894 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7895 PERL_UNUSED_ARG(rx);
7899 return newSVpvs("Regexp");
7902 /* Scans the name of a named buffer from the pattern.
7903 * If flags is REG_RSN_RETURN_NULL returns null.
7904 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7905 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7906 * to the parsed name as looked up in the RExC_paren_names hash.
7907 * If there is an error throws a vFAIL().. type exception.
7910 #define REG_RSN_RETURN_NULL 0
7911 #define REG_RSN_RETURN_NAME 1
7912 #define REG_RSN_RETURN_DATA 2
7915 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7917 char *name_start = RExC_parse;
7919 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7921 assert (RExC_parse <= RExC_end);
7922 if (RExC_parse == RExC_end) NOOP;
7923 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7924 /* skip IDFIRST by using do...while */
7927 RExC_parse += UTF8SKIP(RExC_parse);
7928 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7932 } while (isWORDCHAR(*RExC_parse));
7934 RExC_parse++; /* so the <- from the vFAIL is after the offending
7936 vFAIL("Group name must start with a non-digit word character");
7940 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7941 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7942 if ( flags == REG_RSN_RETURN_NAME)
7944 else if (flags==REG_RSN_RETURN_DATA) {
7947 if ( ! sv_name ) /* should not happen*/
7948 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7949 if (RExC_paren_names)
7950 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7952 sv_dat = HeVAL(he_str);
7954 vFAIL("Reference to nonexistent named group");
7958 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7959 (unsigned long) flags);
7961 NOT_REACHED; /* NOTREACHED */
7966 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7968 if (RExC_lastparse!=RExC_parse) { \
7969 PerlIO_printf(Perl_debug_log, "%s", \
7970 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
7971 RExC_end - RExC_parse, 16, \
7973 PERL_PV_ESCAPE_UNI_DETECT | \
7974 PERL_PV_PRETTY_ELLIPSES | \
7975 PERL_PV_PRETTY_LTGT | \
7976 PERL_PV_ESCAPE_RE | \
7977 PERL_PV_PRETTY_EXACTSIZE \
7981 PerlIO_printf(Perl_debug_log,"%16s",""); \
7984 num = RExC_size + 1; \
7986 num=REG_NODE_NUM(RExC_emit); \
7987 if (RExC_lastnum!=num) \
7988 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7990 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7991 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7992 (int)((depth*2)), "", \
7996 RExC_lastparse=RExC_parse; \
8001 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
8002 DEBUG_PARSE_MSG((funcname)); \
8003 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
8005 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
8006 DEBUG_PARSE_MSG((funcname)); \
8007 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
8010 /* This section of code defines the inversion list object and its methods. The
8011 * interfaces are highly subject to change, so as much as possible is static to
8012 * this file. An inversion list is here implemented as a malloc'd C UV array
8013 * as an SVt_INVLIST scalar.
8015 * An inversion list for Unicode is an array of code points, sorted by ordinal
8016 * number. The zeroth element is the first code point in the list. The 1th
8017 * element is the first element beyond that not in the list. In other words,
8018 * the first range is
8019 * invlist[0]..(invlist[1]-1)
8020 * The other ranges follow. Thus every element whose index is divisible by two
8021 * marks the beginning of a range that is in the list, and every element not
8022 * divisible by two marks the beginning of a range not in the list. A single
8023 * element inversion list that contains the single code point N generally
8024 * consists of two elements
8027 * (The exception is when N is the highest representable value on the
8028 * machine, in which case the list containing just it would be a single
8029 * element, itself. By extension, if the last range in the list extends to
8030 * infinity, then the first element of that range will be in the inversion list
8031 * at a position that is divisible by two, and is the final element in the
8033 * Taking the complement (inverting) an inversion list is quite simple, if the
8034 * first element is 0, remove it; otherwise add a 0 element at the beginning.
8035 * This implementation reserves an element at the beginning of each inversion
8036 * list to always contain 0; there is an additional flag in the header which
8037 * indicates if the list begins at the 0, or is offset to begin at the next
8040 * More about inversion lists can be found in "Unicode Demystified"
8041 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
8042 * More will be coming when functionality is added later.
8044 * The inversion list data structure is currently implemented as an SV pointing
8045 * to an array of UVs that the SV thinks are bytes. This allows us to have an
8046 * array of UV whose memory management is automatically handled by the existing
8047 * facilities for SV's.
8049 * Some of the methods should always be private to the implementation, and some
8050 * should eventually be made public */
8052 /* The header definitions are in F<invlist_inline.h> */
8054 PERL_STATIC_INLINE UV*
8055 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8057 /* Returns a pointer to the first element in the inversion list's array.
8058 * This is called upon initialization of an inversion list. Where the
8059 * array begins depends on whether the list has the code point U+0000 in it
8060 * or not. The other parameter tells it whether the code that follows this
8061 * call is about to put a 0 in the inversion list or not. The first
8062 * element is either the element reserved for 0, if TRUE, or the element
8063 * after it, if FALSE */
8065 bool* offset = get_invlist_offset_addr(invlist);
8066 UV* zero_addr = (UV *) SvPVX(invlist);
8068 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8071 assert(! _invlist_len(invlist));
8075 /* 1^1 = 0; 1^0 = 1 */
8076 *offset = 1 ^ will_have_0;
8077 return zero_addr + *offset;
8080 PERL_STATIC_INLINE void
8081 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8083 /* Sets the current number of elements stored in the inversion list.
8084 * Updates SvCUR correspondingly */
8085 PERL_UNUSED_CONTEXT;
8086 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8088 assert(SvTYPE(invlist) == SVt_INVLIST);
8093 : TO_INTERNAL_SIZE(len + offset));
8094 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8097 #ifndef PERL_IN_XSUB_RE
8099 PERL_STATIC_INLINE IV*
8100 S_get_invlist_previous_index_addr(SV* invlist)
8102 /* Return the address of the IV that is reserved to hold the cached index
8104 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8106 assert(SvTYPE(invlist) == SVt_INVLIST);
8108 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8111 PERL_STATIC_INLINE IV
8112 S_invlist_previous_index(SV* const invlist)
8114 /* Returns cached index of previous search */
8116 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8118 return *get_invlist_previous_index_addr(invlist);
8121 PERL_STATIC_INLINE void
8122 S_invlist_set_previous_index(SV* const invlist, const IV index)
8124 /* Caches <index> for later retrieval */
8126 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8128 assert(index == 0 || index < (int) _invlist_len(invlist));
8130 *get_invlist_previous_index_addr(invlist) = index;
8133 PERL_STATIC_INLINE void
8134 S_invlist_trim(SV* const invlist)
8136 PERL_ARGS_ASSERT_INVLIST_TRIM;
8138 assert(SvTYPE(invlist) == SVt_INVLIST);
8140 /* Change the length of the inversion list to how many entries it currently
8142 SvPV_shrink_to_cur((SV *) invlist);
8145 PERL_STATIC_INLINE bool
8146 S_invlist_is_iterating(SV* const invlist)
8148 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8150 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8153 #endif /* ifndef PERL_IN_XSUB_RE */
8155 PERL_STATIC_INLINE UV
8156 S_invlist_max(SV* const invlist)
8158 /* Returns the maximum number of elements storable in the inversion list's
8159 * array, without having to realloc() */
8161 PERL_ARGS_ASSERT_INVLIST_MAX;
8163 assert(SvTYPE(invlist) == SVt_INVLIST);
8165 /* Assumes worst case, in which the 0 element is not counted in the
8166 * inversion list, so subtracts 1 for that */
8167 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8168 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8169 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8172 #ifndef PERL_IN_XSUB_RE
8174 Perl__new_invlist(pTHX_ IV initial_size)
8177 /* Return a pointer to a newly constructed inversion list, with enough
8178 * space to store 'initial_size' elements. If that number is negative, a
8179 * system default is used instead */
8183 if (initial_size < 0) {
8187 /* Allocate the initial space */
8188 new_list = newSV_type(SVt_INVLIST);
8190 /* First 1 is in case the zero element isn't in the list; second 1 is for
8192 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8193 invlist_set_len(new_list, 0, 0);
8195 /* Force iterinit() to be used to get iteration to work */
8196 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8198 *get_invlist_previous_index_addr(new_list) = 0;
8204 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8206 /* Return a pointer to a newly constructed inversion list, initialized to
8207 * point to <list>, which has to be in the exact correct inversion list
8208 * form, including internal fields. Thus this is a dangerous routine that
8209 * should not be used in the wrong hands. The passed in 'list' contains
8210 * several header fields at the beginning that are not part of the
8211 * inversion list body proper */
8213 const STRLEN length = (STRLEN) list[0];
8214 const UV version_id = list[1];
8215 const bool offset = cBOOL(list[2]);
8216 #define HEADER_LENGTH 3
8217 /* If any of the above changes in any way, you must change HEADER_LENGTH
8218 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8219 * perl -E 'say int(rand 2**31-1)'
8221 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8222 data structure type, so that one being
8223 passed in can be validated to be an
8224 inversion list of the correct vintage.
8227 SV* invlist = newSV_type(SVt_INVLIST);
8229 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8231 if (version_id != INVLIST_VERSION_ID) {
8232 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8235 /* The generated array passed in includes header elements that aren't part
8236 * of the list proper, so start it just after them */
8237 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8239 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8240 shouldn't touch it */
8242 *(get_invlist_offset_addr(invlist)) = offset;
8244 /* The 'length' passed to us is the physical number of elements in the
8245 * inversion list. But if there is an offset the logical number is one
8247 invlist_set_len(invlist, length - offset, offset);
8249 invlist_set_previous_index(invlist, 0);
8251 /* Initialize the iteration pointer. */
8252 invlist_iterfinish(invlist);
8254 SvREADONLY_on(invlist);
8258 #endif /* ifndef PERL_IN_XSUB_RE */
8261 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8263 /* Grow the maximum size of an inversion list */
8265 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8267 assert(SvTYPE(invlist) == SVt_INVLIST);
8269 /* Add one to account for the zero element at the beginning which may not
8270 * be counted by the calling parameters */
8271 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8275 S__append_range_to_invlist(pTHX_ SV* const invlist,
8276 const UV start, const UV end)
8278 /* Subject to change or removal. Append the range from 'start' to 'end' at
8279 * the end of the inversion list. The range must be above any existing
8283 UV max = invlist_max(invlist);
8284 UV len = _invlist_len(invlist);
8287 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8289 if (len == 0) { /* Empty lists must be initialized */
8290 offset = start != 0;
8291 array = _invlist_array_init(invlist, ! offset);
8294 /* Here, the existing list is non-empty. The current max entry in the
8295 * list is generally the first value not in the set, except when the
8296 * set extends to the end of permissible values, in which case it is
8297 * the first entry in that final set, and so this call is an attempt to
8298 * append out-of-order */
8300 UV final_element = len - 1;
8301 array = invlist_array(invlist);
8302 if (array[final_element] > start
8303 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8305 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",
8306 array[final_element], start,
8307 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8310 /* Here, it is a legal append. If the new range begins with the first
8311 * value not in the set, it is extending the set, so the new first
8312 * value not in the set is one greater than the newly extended range.
8314 offset = *get_invlist_offset_addr(invlist);
8315 if (array[final_element] == start) {
8316 if (end != UV_MAX) {
8317 array[final_element] = end + 1;
8320 /* But if the end is the maximum representable on the machine,
8321 * just let the range that this would extend to have no end */
8322 invlist_set_len(invlist, len - 1, offset);
8328 /* Here the new range doesn't extend any existing set. Add it */
8330 len += 2; /* Includes an element each for the start and end of range */
8332 /* If wll overflow the existing space, extend, which may cause the array to
8335 invlist_extend(invlist, len);
8337 /* Have to set len here to avoid assert failure in invlist_array() */
8338 invlist_set_len(invlist, len, offset);
8340 array = invlist_array(invlist);
8343 invlist_set_len(invlist, len, offset);
8346 /* The next item on the list starts the range, the one after that is
8347 * one past the new range. */
8348 array[len - 2] = start;
8349 if (end != UV_MAX) {
8350 array[len - 1] = end + 1;
8353 /* But if the end is the maximum representable on the machine, just let
8354 * the range have no end */
8355 invlist_set_len(invlist, len - 1, offset);
8359 #ifndef PERL_IN_XSUB_RE
8362 Perl__invlist_search(SV* const invlist, const UV cp)
8364 /* Searches the inversion list for the entry that contains the input code
8365 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8366 * return value is the index into the list's array of the range that
8371 IV high = _invlist_len(invlist);
8372 const IV highest_element = high - 1;
8375 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8377 /* If list is empty, return failure. */
8382 /* (We can't get the array unless we know the list is non-empty) */
8383 array = invlist_array(invlist);
8385 mid = invlist_previous_index(invlist);
8386 assert(mid >=0 && mid <= highest_element);
8388 /* <mid> contains the cache of the result of the previous call to this
8389 * function (0 the first time). See if this call is for the same result,
8390 * or if it is for mid-1. This is under the theory that calls to this
8391 * function will often be for related code points that are near each other.
8392 * And benchmarks show that caching gives better results. We also test
8393 * here if the code point is within the bounds of the list. These tests
8394 * replace others that would have had to be made anyway to make sure that
8395 * the array bounds were not exceeded, and these give us extra information
8396 * at the same time */
8397 if (cp >= array[mid]) {
8398 if (cp >= array[highest_element]) {
8399 return highest_element;
8402 /* Here, array[mid] <= cp < array[highest_element]. This means that
8403 * the final element is not the answer, so can exclude it; it also
8404 * means that <mid> is not the final element, so can refer to 'mid + 1'
8406 if (cp < array[mid + 1]) {
8412 else { /* cp < aray[mid] */
8413 if (cp < array[0]) { /* Fail if outside the array */
8417 if (cp >= array[mid - 1]) {
8422 /* Binary search. What we are looking for is <i> such that
8423 * array[i] <= cp < array[i+1]
8424 * The loop below converges on the i+1. Note that there may not be an
8425 * (i+1)th element in the array, and things work nonetheless */
8426 while (low < high) {
8427 mid = (low + high) / 2;
8428 assert(mid <= highest_element);
8429 if (array[mid] <= cp) { /* cp >= array[mid] */
8432 /* We could do this extra test to exit the loop early.
8433 if (cp < array[low]) {
8438 else { /* cp < array[mid] */
8445 invlist_set_previous_index(invlist, high);
8450 Perl__invlist_populate_swatch(SV* const invlist,
8451 const UV start, const UV end, U8* swatch)
8453 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8454 * but is used when the swash has an inversion list. This makes this much
8455 * faster, as it uses a binary search instead of a linear one. This is
8456 * intimately tied to that function, and perhaps should be in utf8.c,
8457 * except it is intimately tied to inversion lists as well. It assumes
8458 * that <swatch> is all 0's on input */
8461 const IV len = _invlist_len(invlist);
8465 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8467 if (len == 0) { /* Empty inversion list */
8471 array = invlist_array(invlist);
8473 /* Find which element it is */
8474 i = _invlist_search(invlist, start);
8476 /* We populate from <start> to <end> */
8477 while (current < end) {
8480 /* The inversion list gives the results for every possible code point
8481 * after the first one in the list. Only those ranges whose index is
8482 * even are ones that the inversion list matches. For the odd ones,
8483 * and if the initial code point is not in the list, we have to skip
8484 * forward to the next element */
8485 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8487 if (i >= len) { /* Finished if beyond the end of the array */
8491 if (current >= end) { /* Finished if beyond the end of what we
8493 if (LIKELY(end < UV_MAX)) {
8497 /* We get here when the upper bound is the maximum
8498 * representable on the machine, and we are looking for just
8499 * that code point. Have to special case it */
8501 goto join_end_of_list;
8504 assert(current >= start);
8506 /* The current range ends one below the next one, except don't go past
8509 upper = (i < len && array[i] < end) ? array[i] : end;
8511 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8512 * for each code point in it */
8513 for (; current < upper; current++) {
8514 const STRLEN offset = (STRLEN)(current - start);
8515 swatch[offset >> 3] |= 1 << (offset & 7);
8520 /* Quit if at the end of the list */
8523 /* But first, have to deal with the highest possible code point on
8524 * the platform. The previous code assumes that <end> is one
8525 * beyond where we want to populate, but that is impossible at the
8526 * platform's infinity, so have to handle it specially */
8527 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8529 const STRLEN offset = (STRLEN)(end - start);
8530 swatch[offset >> 3] |= 1 << (offset & 7);
8535 /* Advance to the next range, which will be for code points not in the
8544 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8545 const bool complement_b, SV** output)
8547 /* Take the union of two inversion lists and point <output> to it. *output
8548 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8549 * the reference count to that list will be decremented if not already a
8550 * temporary (mortal); otherwise *output will be made correspondingly
8551 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8552 * second list is returned. If <complement_b> is TRUE, the union is taken
8553 * of the complement (inversion) of <b> instead of b itself.
8555 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8556 * Richard Gillam, published by Addison-Wesley, and explained at some
8557 * length there. The preface says to incorporate its examples into your
8558 * code at your own risk.
8560 * The algorithm is like a merge sort.
8562 * XXX A potential performance improvement is to keep track as we go along
8563 * if only one of the inputs contributes to the result, meaning the other
8564 * is a subset of that one. In that case, we can skip the final copy and
8565 * return the larger of the input lists, but then outside code might need
8566 * to keep track of whether to free the input list or not */
8568 const UV* array_a; /* a's array */
8570 UV len_a; /* length of a's array */
8573 SV* u; /* the resulting union */
8577 UV i_a = 0; /* current index into a's array */
8581 /* running count, as explained in the algorithm source book; items are
8582 * stopped accumulating and are output when the count changes to/from 0.
8583 * The count is incremented when we start a range that's in the set, and
8584 * decremented when we start a range that's not in the set. So its range
8585 * is 0 to 2. Only when the count is zero is something not in the set.
8589 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8592 /* If either one is empty, the union is the other one */
8593 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8594 bool make_temp = FALSE; /* Should we mortalize the result? */
8598 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8604 *output = invlist_clone(b);
8606 _invlist_invert(*output);
8608 } /* else *output already = b; */
8611 sv_2mortal(*output);
8615 else if ((len_b = _invlist_len(b)) == 0) {
8616 bool make_temp = FALSE;
8618 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8623 /* The complement of an empty list is a list that has everything in it,
8624 * so the union with <a> includes everything too */
8627 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8631 *output = _new_invlist(1);
8632 _append_range_to_invlist(*output, 0, UV_MAX);
8634 else if (*output != a) {
8635 *output = invlist_clone(a);
8637 /* else *output already = a; */
8640 sv_2mortal(*output);
8645 /* Here both lists exist and are non-empty */
8646 array_a = invlist_array(a);
8647 array_b = invlist_array(b);
8649 /* If are to take the union of 'a' with the complement of b, set it
8650 * up so are looking at b's complement. */
8653 /* To complement, we invert: if the first element is 0, remove it. To
8654 * do this, we just pretend the array starts one later */
8655 if (array_b[0] == 0) {
8661 /* But if the first element is not zero, we pretend the list starts
8662 * at the 0 that is always stored immediately before the array. */
8668 /* Size the union for the worst case: that the sets are completely
8670 u = _new_invlist(len_a + len_b);
8672 /* Will contain U+0000 if either component does */
8673 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8674 || (len_b > 0 && array_b[0] == 0));
8676 /* Go through each list item by item, stopping when exhausted one of
8678 while (i_a < len_a && i_b < len_b) {
8679 UV cp; /* The element to potentially add to the union's array */
8680 bool cp_in_set; /* is it in the the input list's set or not */
8682 /* We need to take one or the other of the two inputs for the union.
8683 * Since we are merging two sorted lists, we take the smaller of the
8684 * next items. In case of a tie, we take the one that is in its set
8685 * first. If we took one not in the set first, it would decrement the
8686 * count, possibly to 0 which would cause it to be output as ending the
8687 * range, and the next time through we would take the same number, and
8688 * output it again as beginning the next range. By doing it the
8689 * opposite way, there is no possibility that the count will be
8690 * momentarily decremented to 0, and thus the two adjoining ranges will
8691 * be seamlessly merged. (In a tie and both are in the set or both not
8692 * in the set, it doesn't matter which we take first.) */
8693 if (array_a[i_a] < array_b[i_b]
8694 || (array_a[i_a] == array_b[i_b]
8695 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8697 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8701 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8702 cp = array_b[i_b++];
8705 /* Here, have chosen which of the two inputs to look at. Only output
8706 * if the running count changes to/from 0, which marks the
8707 * beginning/end of a range in that's in the set */
8710 array_u[i_u++] = cp;
8717 array_u[i_u++] = cp;
8722 /* Here, we are finished going through at least one of the lists, which
8723 * means there is something remaining in at most one. We check if the list
8724 * that hasn't been exhausted is positioned such that we are in the middle
8725 * of a range in its set or not. (i_a and i_b point to the element beyond
8726 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8727 * is potentially more to output.
8728 * There are four cases:
8729 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8730 * in the union is entirely from the non-exhausted set.
8731 * 2) Both were in their sets, count is 2. Nothing further should
8732 * be output, as everything that remains will be in the exhausted
8733 * list's set, hence in the union; decrementing to 1 but not 0 insures
8735 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8736 * Nothing further should be output because the union includes
8737 * everything from the exhausted set. Not decrementing ensures that.
8738 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8739 * decrementing to 0 insures that we look at the remainder of the
8740 * non-exhausted set */
8741 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8742 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8747 /* The final length is what we've output so far, plus what else is about to
8748 * be output. (If 'count' is non-zero, then the input list we exhausted
8749 * has everything remaining up to the machine's limit in its set, and hence
8750 * in the union, so there will be no further output. */
8753 /* At most one of the subexpressions will be non-zero */
8754 len_u += (len_a - i_a) + (len_b - i_b);
8757 /* Set result to final length, which can change the pointer to array_u, so
8759 if (len_u != _invlist_len(u)) {
8760 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8762 array_u = invlist_array(u);
8765 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8766 * the other) ended with everything above it not in its set. That means
8767 * that the remaining part of the union is precisely the same as the
8768 * non-exhausted list, so can just copy it unchanged. (If both list were
8769 * exhausted at the same time, then the operations below will be both 0.)
8772 IV copy_count; /* At most one will have a non-zero copy count */
8773 if ((copy_count = len_a - i_a) > 0) {
8774 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8776 else if ((copy_count = len_b - i_b) > 0) {
8777 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8781 /* We may be removing a reference to one of the inputs. If so, the output
8782 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8783 * count decremented) */
8784 if (a == *output || b == *output) {
8785 assert(! invlist_is_iterating(*output));
8786 if ((SvTEMP(*output))) {
8790 SvREFCNT_dec_NN(*output);
8800 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8801 const bool complement_b, SV** i)
8803 /* Take the intersection of two inversion lists and point <i> to it. *i
8804 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8805 * the reference count to that list will be decremented if not already a
8806 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8807 * The first list, <a>, may be NULL, in which case an empty list is
8808 * returned. If <complement_b> is TRUE, the result will be the
8809 * intersection of <a> and the complement (or inversion) of <b> instead of
8812 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8813 * Richard Gillam, published by Addison-Wesley, and explained at some
8814 * length there. The preface says to incorporate its examples into your
8815 * code at your own risk. In fact, it had bugs
8817 * The algorithm is like a merge sort, and is essentially the same as the
8821 const UV* array_a; /* a's array */
8823 UV len_a; /* length of a's array */
8826 SV* r; /* the resulting intersection */
8830 UV i_a = 0; /* current index into a's array */
8834 /* running count, as explained in the algorithm source book; items are
8835 * stopped accumulating and are output when the count changes to/from 2.
8836 * The count is incremented when we start a range that's in the set, and
8837 * decremented when we start a range that's not in the set. So its range
8838 * is 0 to 2. Only when the count is 2 is something in the intersection.
8842 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8845 /* Special case if either one is empty */
8846 len_a = (a == NULL) ? 0 : _invlist_len(a);
8847 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8848 bool make_temp = FALSE;
8850 if (len_a != 0 && complement_b) {
8852 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8853 * be empty. Here, also we are using 'b's complement, which hence
8854 * must be every possible code point. Thus the intersection is
8858 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8863 *i = invlist_clone(a);
8865 /* else *i is already 'a' */
8873 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8874 * intersection must be empty */
8876 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8881 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8885 *i = _new_invlist(0);
8893 /* Here both lists exist and are non-empty */
8894 array_a = invlist_array(a);
8895 array_b = invlist_array(b);
8897 /* If are to take the intersection of 'a' with the complement of b, set it
8898 * up so are looking at b's complement. */
8901 /* To complement, we invert: if the first element is 0, remove it. To
8902 * do this, we just pretend the array starts one later */
8903 if (array_b[0] == 0) {
8909 /* But if the first element is not zero, we pretend the list starts
8910 * at the 0 that is always stored immediately before the array. */
8916 /* Size the intersection for the worst case: that the intersection ends up
8917 * fragmenting everything to be completely disjoint */
8918 r= _new_invlist(len_a + len_b);
8920 /* Will contain U+0000 iff both components do */
8921 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8922 && len_b > 0 && array_b[0] == 0);
8924 /* Go through each list item by item, stopping when exhausted one of
8926 while (i_a < len_a && i_b < len_b) {
8927 UV cp; /* The element to potentially add to the intersection's
8929 bool cp_in_set; /* Is it in the input list's set or not */
8931 /* We need to take one or the other of the two inputs for the
8932 * intersection. Since we are merging two sorted lists, we take the
8933 * smaller of the next items. In case of a tie, we take the one that
8934 * is not in its set first (a difference from the union algorithm). If
8935 * we took one in the set first, it would increment the count, possibly
8936 * to 2 which would cause it to be output as starting a range in the
8937 * intersection, and the next time through we would take that same
8938 * number, and output it again as ending the set. By doing it the
8939 * opposite of this, there is no possibility that the count will be
8940 * momentarily incremented to 2. (In a tie and both are in the set or
8941 * both not in the set, it doesn't matter which we take first.) */
8942 if (array_a[i_a] < array_b[i_b]
8943 || (array_a[i_a] == array_b[i_b]
8944 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8946 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8950 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8954 /* Here, have chosen which of the two inputs to look at. Only output
8955 * if the running count changes to/from 2, which marks the
8956 * beginning/end of a range that's in the intersection */
8960 array_r[i_r++] = cp;
8965 array_r[i_r++] = cp;
8971 /* Here, we are finished going through at least one of the lists, which
8972 * means there is something remaining in at most one. We check if the list
8973 * that has been exhausted is positioned such that we are in the middle
8974 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8975 * the ones we care about.) There are four cases:
8976 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8977 * nothing left in the intersection.
8978 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8979 * above 2. What should be output is exactly that which is in the
8980 * non-exhausted set, as everything it has is also in the intersection
8981 * set, and everything it doesn't have can't be in the intersection
8982 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8983 * gets incremented to 2. Like the previous case, the intersection is
8984 * everything that remains in the non-exhausted set.
8985 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8986 * remains 1. And the intersection has nothing more. */
8987 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8988 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8993 /* The final length is what we've output so far plus what else is in the
8994 * intersection. At most one of the subexpressions below will be non-zero
8998 len_r += (len_a - i_a) + (len_b - i_b);
9001 /* Set result to final length, which can change the pointer to array_r, so
9003 if (len_r != _invlist_len(r)) {
9004 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
9006 array_r = invlist_array(r);
9009 /* Finish outputting any remaining */
9010 if (count >= 2) { /* At most one will have a non-zero copy count */
9012 if ((copy_count = len_a - i_a) > 0) {
9013 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
9015 else if ((copy_count = len_b - i_b) > 0) {
9016 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
9020 /* We may be removing a reference to one of the inputs. If so, the output
9021 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
9022 * count decremented) */
9023 if (a == *i || b == *i) {
9024 assert(! invlist_is_iterating(*i));
9029 SvREFCNT_dec_NN(*i);
9039 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
9041 /* Add the range from 'start' to 'end' inclusive to the inversion list's
9042 * set. A pointer to the inversion list is returned. This may actually be
9043 * a new list, in which case the passed in one has been destroyed. The
9044 * passed-in inversion list can be NULL, in which case a new one is created
9045 * with just the one range in it */
9050 if (invlist == NULL) {
9051 invlist = _new_invlist(2);
9055 len = _invlist_len(invlist);
9058 /* If comes after the final entry actually in the list, can just append it
9061 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
9062 && start >= invlist_array(invlist)[len - 1]))
9064 _append_range_to_invlist(invlist, start, end);
9068 /* Here, can't just append things, create and return a new inversion list
9069 * which is the union of this range and the existing inversion list. (If
9070 * the new range is well-behaved wrt to the old one, we could just insert
9071 * it, doing a Move() down on the tail of the old one (potentially growing
9072 * it first). But to determine that means we would have the extra
9073 * (possibly throw-away) work of first finding where the new one goes and
9074 * whether it disrupts (splits) an existing range, so it doesn't appear to
9075 * me (khw) that it's worth it) */
9076 range_invlist = _new_invlist(2);
9077 _append_range_to_invlist(range_invlist, start, end);
9079 _invlist_union(invlist, range_invlist, &invlist);
9081 /* The temporary can be freed */
9082 SvREFCNT_dec_NN(range_invlist);
9088 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9089 UV** other_elements_ptr)
9091 /* Create and return an inversion list whose contents are to be populated
9092 * by the caller. The caller gives the number of elements (in 'size') and
9093 * the very first element ('element0'). This function will set
9094 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9097 * Obviously there is some trust involved that the caller will properly
9098 * fill in the other elements of the array.
9100 * (The first element needs to be passed in, as the underlying code does
9101 * things differently depending on whether it is zero or non-zero) */
9103 SV* invlist = _new_invlist(size);
9106 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9108 _append_range_to_invlist(invlist, element0, element0);
9109 offset = *get_invlist_offset_addr(invlist);
9111 invlist_set_len(invlist, size, offset);
9112 *other_elements_ptr = invlist_array(invlist) + 1;
9118 PERL_STATIC_INLINE SV*
9119 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9120 return _add_range_to_invlist(invlist, cp, cp);
9123 #ifndef PERL_IN_XSUB_RE
9125 Perl__invlist_invert(pTHX_ SV* const invlist)
9127 /* Complement the input inversion list. This adds a 0 if the list didn't
9128 * have a zero; removes it otherwise. As described above, the data
9129 * structure is set up so that this is very efficient */
9131 PERL_ARGS_ASSERT__INVLIST_INVERT;
9133 assert(! invlist_is_iterating(invlist));
9135 /* The inverse of matching nothing is matching everything */
9136 if (_invlist_len(invlist) == 0) {
9137 _append_range_to_invlist(invlist, 0, UV_MAX);
9141 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9146 PERL_STATIC_INLINE SV*
9147 S_invlist_clone(pTHX_ SV* const invlist)
9150 /* Return a new inversion list that is a copy of the input one, which is
9151 * unchanged. The new list will not be mortal even if the old one was. */
9153 /* Need to allocate extra space to accommodate Perl's addition of a
9154 * trailing NUL to SvPV's, since it thinks they are always strings */
9155 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9156 STRLEN physical_length = SvCUR(invlist);
9157 bool offset = *(get_invlist_offset_addr(invlist));
9159 PERL_ARGS_ASSERT_INVLIST_CLONE;
9161 *(get_invlist_offset_addr(new_invlist)) = offset;
9162 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9163 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9168 PERL_STATIC_INLINE STRLEN*
9169 S_get_invlist_iter_addr(SV* invlist)
9171 /* Return the address of the UV that contains the current iteration
9174 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9176 assert(SvTYPE(invlist) == SVt_INVLIST);
9178 return &(((XINVLIST*) SvANY(invlist))->iterator);
9181 PERL_STATIC_INLINE void
9182 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9184 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9186 *get_invlist_iter_addr(invlist) = 0;
9189 PERL_STATIC_INLINE void
9190 S_invlist_iterfinish(SV* invlist)
9192 /* Terminate iterator for invlist. This is to catch development errors.
9193 * Any iteration that is interrupted before completed should call this
9194 * function. Functions that add code points anywhere else but to the end
9195 * of an inversion list assert that they are not in the middle of an
9196 * iteration. If they were, the addition would make the iteration
9197 * problematical: if the iteration hadn't reached the place where things
9198 * were being added, it would be ok */
9200 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9202 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9206 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9208 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9209 * This call sets in <*start> and <*end>, the next range in <invlist>.
9210 * Returns <TRUE> if successful and the next call will return the next
9211 * range; <FALSE> if was already at the end of the list. If the latter,
9212 * <*start> and <*end> are unchanged, and the next call to this function
9213 * will start over at the beginning of the list */
9215 STRLEN* pos = get_invlist_iter_addr(invlist);
9216 UV len = _invlist_len(invlist);
9219 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9222 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9226 array = invlist_array(invlist);
9228 *start = array[(*pos)++];
9234 *end = array[(*pos)++] - 1;
9240 PERL_STATIC_INLINE UV
9241 S_invlist_highest(SV* const invlist)
9243 /* Returns the highest code point that matches an inversion list. This API
9244 * has an ambiguity, as it returns 0 under either the highest is actually
9245 * 0, or if the list is empty. If this distinction matters to you, check
9246 * for emptiness before calling this function */
9248 UV len = _invlist_len(invlist);
9251 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9257 array = invlist_array(invlist);
9259 /* The last element in the array in the inversion list always starts a
9260 * range that goes to infinity. That range may be for code points that are
9261 * matched in the inversion list, or it may be for ones that aren't
9262 * matched. In the latter case, the highest code point in the set is one
9263 * less than the beginning of this range; otherwise it is the final element
9264 * of this range: infinity */
9265 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9267 : array[len - 1] - 1;
9270 #ifndef PERL_IN_XSUB_RE
9272 Perl__invlist_contents(pTHX_ SV* const invlist)
9274 /* Get the contents of an inversion list into a string SV so that they can
9275 * be printed out. It uses the format traditionally done for debug tracing
9279 SV* output = newSVpvs("\n");
9281 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9283 assert(! invlist_is_iterating(invlist));
9285 invlist_iterinit(invlist);
9286 while (invlist_iternext(invlist, &start, &end)) {
9287 if (end == UV_MAX) {
9288 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9290 else if (end != start) {
9291 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9295 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9303 #ifndef PERL_IN_XSUB_RE
9305 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9306 const char * const indent, SV* const invlist)
9308 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9309 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9310 * the string 'indent'. The output looks like this:
9311 [0] 0x000A .. 0x000D
9313 [4] 0x2028 .. 0x2029
9314 [6] 0x3104 .. INFINITY
9315 * This means that the first range of code points matched by the list are
9316 * 0xA through 0xD; the second range contains only the single code point
9317 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9318 * are used to define each range (except if the final range extends to
9319 * infinity, only a single element is needed). The array index of the
9320 * first element for the corresponding range is given in brackets. */
9325 PERL_ARGS_ASSERT__INVLIST_DUMP;
9327 if (invlist_is_iterating(invlist)) {
9328 Perl_dump_indent(aTHX_ level, file,
9329 "%sCan't dump inversion list because is in middle of iterating\n",
9334 invlist_iterinit(invlist);
9335 while (invlist_iternext(invlist, &start, &end)) {
9336 if (end == UV_MAX) {
9337 Perl_dump_indent(aTHX_ level, file,
9338 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9339 indent, (UV)count, start);
9341 else if (end != start) {
9342 Perl_dump_indent(aTHX_ level, file,
9343 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9344 indent, (UV)count, start, end);
9347 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9348 indent, (UV)count, start);
9355 Perl__load_PL_utf8_foldclosures (pTHX)
9357 assert(! PL_utf8_foldclosures);
9359 /* If the folds haven't been read in, call a fold function
9361 if (! PL_utf8_tofold) {
9362 U8 dummy[UTF8_MAXBYTES_CASE+1];
9364 /* This string is just a short named one above \xff */
9365 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
9366 assert(PL_utf8_tofold); /* Verify that worked */
9368 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
9372 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9374 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9376 /* Return a boolean as to if the two passed in inversion lists are
9377 * identical. The final argument, if TRUE, says to take the complement of
9378 * the second inversion list before doing the comparison */
9380 const UV* array_a = invlist_array(a);
9381 const UV* array_b = invlist_array(b);
9382 UV len_a = _invlist_len(a);
9383 UV len_b = _invlist_len(b);
9385 UV i = 0; /* current index into the arrays */
9386 bool retval = TRUE; /* Assume are identical until proven otherwise */
9388 PERL_ARGS_ASSERT__INVLISTEQ;
9390 /* If are to compare 'a' with the complement of b, set it
9391 * up so are looking at b's complement. */
9394 /* The complement of nothing is everything, so <a> would have to have
9395 * just one element, starting at zero (ending at infinity) */
9397 return (len_a == 1 && array_a[0] == 0);
9399 else if (array_b[0] == 0) {
9401 /* Otherwise, to complement, we invert. Here, the first element is
9402 * 0, just remove it. To do this, we just pretend the array starts
9410 /* But if the first element is not zero, we pretend the list starts
9411 * at the 0 that is always stored immediately before the array. */
9417 /* Make sure that the lengths are the same, as well as the final element
9418 * before looping through the remainder. (Thus we test the length, final,
9419 * and first elements right off the bat) */
9420 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9423 else for (i = 0; i < len_a - 1; i++) {
9424 if (array_a[i] != array_b[i]) {
9435 * As best we can, determine the characters that can match the start of
9436 * the given EXACTF-ish node.
9438 * Returns the invlist as a new SV*; it is the caller's responsibility to
9439 * call SvREFCNT_dec() when done with it.
9442 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
9444 const U8 * s = (U8*)STRING(node);
9445 SSize_t bytelen = STR_LEN(node);
9447 /* Start out big enough for 2 separate code points */
9448 SV* invlist = _new_invlist(4);
9450 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
9455 /* We punt and assume can match anything if the node begins
9456 * with a multi-character fold. Things are complicated. For
9457 * example, /ffi/i could match any of:
9458 * "\N{LATIN SMALL LIGATURE FFI}"
9459 * "\N{LATIN SMALL LIGATURE FF}I"
9460 * "F\N{LATIN SMALL LIGATURE FI}"
9461 * plus several other things; and making sure we have all the
9462 * possibilities is hard. */
9463 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
9464 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9467 /* Any Latin1 range character can potentially match any
9468 * other depending on the locale */
9469 if (OP(node) == EXACTFL) {
9470 _invlist_union(invlist, PL_Latin1, &invlist);
9473 /* But otherwise, it matches at least itself. We can
9474 * quickly tell if it has a distinct fold, and if so,
9475 * it matches that as well */
9476 invlist = add_cp_to_invlist(invlist, uc);
9477 if (IS_IN_SOME_FOLD_L1(uc))
9478 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
9481 /* Some characters match above-Latin1 ones under /i. This
9482 * is true of EXACTFL ones when the locale is UTF-8 */
9483 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
9484 && (! isASCII(uc) || (OP(node) != EXACTFA
9485 && OP(node) != EXACTFA_NO_TRIE)))
9487 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
9491 else { /* Pattern is UTF-8 */
9492 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
9493 STRLEN foldlen = UTF8SKIP(s);
9494 const U8* e = s + bytelen;
9497 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
9499 /* The only code points that aren't folded in a UTF EXACTFish
9500 * node are are the problematic ones in EXACTFL nodes */
9501 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
9502 /* We need to check for the possibility that this EXACTFL
9503 * node begins with a multi-char fold. Therefore we fold
9504 * the first few characters of it so that we can make that
9509 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
9511 *(d++) = (U8) toFOLD(*s);
9516 to_utf8_fold(s, d, &len);
9522 /* And set up so the code below that looks in this folded
9523 * buffer instead of the node's string */
9525 foldlen = UTF8SKIP(folded);
9529 /* When we reach here 's' points to the fold of the first
9530 * character(s) of the node; and 'e' points to far enough along
9531 * the folded string to be just past any possible multi-char
9532 * fold. 'foldlen' is the length in bytes of the first
9535 * Unlike the non-UTF-8 case, the macro for determining if a
9536 * string is a multi-char fold requires all the characters to
9537 * already be folded. This is because of all the complications
9538 * if not. Note that they are folded anyway, except in EXACTFL
9539 * nodes. Like the non-UTF case above, we punt if the node
9540 * begins with a multi-char fold */
9542 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
9543 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9545 else { /* Single char fold */
9547 /* It matches all the things that fold to it, which are
9548 * found in PL_utf8_foldclosures (including itself) */
9549 invlist = add_cp_to_invlist(invlist, uc);
9550 if (! PL_utf8_foldclosures)
9551 _load_PL_utf8_foldclosures();
9552 if ((listp = hv_fetch(PL_utf8_foldclosures,
9553 (char *) s, foldlen, FALSE)))
9555 AV* list = (AV*) *listp;
9557 for (k = 0; k <= av_tindex(list); k++) {
9558 SV** c_p = av_fetch(list, k, FALSE);
9564 /* /aa doesn't allow folds between ASCII and non- */
9565 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
9566 && isASCII(c) != isASCII(uc))
9571 invlist = add_cp_to_invlist(invlist, c);
9580 #undef HEADER_LENGTH
9581 #undef TO_INTERNAL_SIZE
9582 #undef FROM_INTERNAL_SIZE
9583 #undef INVLIST_VERSION_ID
9585 /* End of inversion list object */
9588 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9590 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9591 * constructs, and updates RExC_flags with them. On input, RExC_parse
9592 * should point to the first flag; it is updated on output to point to the
9593 * final ')' or ':'. There needs to be at least one flag, or this will
9596 /* for (?g), (?gc), and (?o) warnings; warning
9597 about (?c) will warn about (?g) -- japhy */
9599 #define WASTED_O 0x01
9600 #define WASTED_G 0x02
9601 #define WASTED_C 0x04
9602 #define WASTED_GC (WASTED_G|WASTED_C)
9603 I32 wastedflags = 0x00;
9604 U32 posflags = 0, negflags = 0;
9605 U32 *flagsp = &posflags;
9606 char has_charset_modifier = '\0';
9608 bool has_use_defaults = FALSE;
9609 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9610 int x_mod_count = 0;
9612 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9614 /* '^' as an initial flag sets certain defaults */
9615 if (UCHARAT(RExC_parse) == '^') {
9617 has_use_defaults = TRUE;
9618 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9619 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9620 ? REGEX_UNICODE_CHARSET
9621 : REGEX_DEPENDS_CHARSET);
9624 cs = get_regex_charset(RExC_flags);
9625 if (cs == REGEX_DEPENDS_CHARSET
9626 && (RExC_utf8 || RExC_uni_semantics))
9628 cs = REGEX_UNICODE_CHARSET;
9631 while (*RExC_parse) {
9632 /* && strchr("iogcmsx", *RExC_parse) */
9633 /* (?g), (?gc) and (?o) are useless here
9634 and must be globally applied -- japhy */
9635 switch (*RExC_parse) {
9637 /* Code for the imsxn flags */
9638 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
9640 case LOCALE_PAT_MOD:
9641 if (has_charset_modifier) {
9642 goto excess_modifier;
9644 else if (flagsp == &negflags) {
9647 cs = REGEX_LOCALE_CHARSET;
9648 has_charset_modifier = LOCALE_PAT_MOD;
9650 case UNICODE_PAT_MOD:
9651 if (has_charset_modifier) {
9652 goto excess_modifier;
9654 else if (flagsp == &negflags) {
9657 cs = REGEX_UNICODE_CHARSET;
9658 has_charset_modifier = UNICODE_PAT_MOD;
9660 case ASCII_RESTRICT_PAT_MOD:
9661 if (flagsp == &negflags) {
9664 if (has_charset_modifier) {
9665 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9666 goto excess_modifier;
9668 /* Doubled modifier implies more restricted */
9669 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9672 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9674 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9676 case DEPENDS_PAT_MOD:
9677 if (has_use_defaults) {
9678 goto fail_modifiers;
9680 else if (flagsp == &negflags) {
9683 else if (has_charset_modifier) {
9684 goto excess_modifier;
9687 /* The dual charset means unicode semantics if the
9688 * pattern (or target, not known until runtime) are
9689 * utf8, or something in the pattern indicates unicode
9691 cs = (RExC_utf8 || RExC_uni_semantics)
9692 ? REGEX_UNICODE_CHARSET
9693 : REGEX_DEPENDS_CHARSET;
9694 has_charset_modifier = DEPENDS_PAT_MOD;
9698 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9699 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9701 else if (has_charset_modifier == *(RExC_parse - 1)) {
9702 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9706 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9708 NOT_REACHED; /*NOTREACHED*/
9711 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9713 NOT_REACHED; /*NOTREACHED*/
9714 case ONCE_PAT_MOD: /* 'o' */
9715 case GLOBAL_PAT_MOD: /* 'g' */
9716 if (PASS2 && ckWARN(WARN_REGEXP)) {
9717 const I32 wflagbit = *RExC_parse == 'o'
9720 if (! (wastedflags & wflagbit) ) {
9721 wastedflags |= wflagbit;
9722 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9725 "Useless (%s%c) - %suse /%c modifier",
9726 flagsp == &negflags ? "?-" : "?",
9728 flagsp == &negflags ? "don't " : "",
9735 case CONTINUE_PAT_MOD: /* 'c' */
9736 if (PASS2 && ckWARN(WARN_REGEXP)) {
9737 if (! (wastedflags & WASTED_C) ) {
9738 wastedflags |= WASTED_GC;
9739 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9742 "Useless (%sc) - %suse /gc modifier",
9743 flagsp == &negflags ? "?-" : "?",
9744 flagsp == &negflags ? "don't " : ""
9749 case KEEPCOPY_PAT_MOD: /* 'p' */
9750 if (flagsp == &negflags) {
9752 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9754 *flagsp |= RXf_PMf_KEEPCOPY;
9758 /* A flag is a default iff it is following a minus, so
9759 * if there is a minus, it means will be trying to
9760 * re-specify a default which is an error */
9761 if (has_use_defaults || flagsp == &negflags) {
9762 goto fail_modifiers;
9765 wastedflags = 0; /* reset so (?g-c) warns twice */
9769 RExC_flags |= posflags;
9770 RExC_flags &= ~negflags;
9771 set_regex_charset(&RExC_flags, cs);
9772 if (RExC_flags & RXf_PMf_FOLD) {
9773 RExC_contains_i = 1;
9776 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9782 RExC_parse += SKIP_IF_CHAR(RExC_parse);
9783 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9784 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9785 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9786 NOT_REACHED; /*NOTREACHED*/
9793 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9798 - reg - regular expression, i.e. main body or parenthesized thing
9800 * Caller must absorb opening parenthesis.
9802 * Combining parenthesis handling with the base level of regular expression
9803 * is a trifle forced, but the need to tie the tails of the branches to what
9804 * follows makes it hard to avoid.
9806 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9808 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9810 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9813 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9814 flags. Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan
9815 needs to be restarted, or'd with NEED_UTF8 if the pattern needs to be
9816 upgraded to UTF-8. Otherwise would only return NULL if regbranch() returns
9817 NULL, which cannot happen. */
9819 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9820 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9821 * 2 is like 1, but indicates that nextchar() has been called to advance
9822 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9823 * this flag alerts us to the need to check for that */
9825 regnode *ret; /* Will be the head of the group. */
9828 regnode *ender = NULL;
9831 U32 oregflags = RExC_flags;
9832 bool have_branch = 0;
9834 I32 freeze_paren = 0;
9835 I32 after_freeze = 0;
9836 I32 num; /* numeric backreferences */
9838 char * parse_start = RExC_parse; /* MJD */
9839 char * const oregcomp_parse = RExC_parse;
9841 GET_RE_DEBUG_FLAGS_DECL;
9843 PERL_ARGS_ASSERT_REG;
9844 DEBUG_PARSE("reg ");
9846 *flagp = 0; /* Tentatively. */
9849 /* Make an OPEN node, if parenthesized. */
9852 /* Under /x, space and comments can be gobbled up between the '(' and
9853 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9854 * intervening space, as the sequence is a token, and a token should be
9856 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9858 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9859 char *start_verb = RExC_parse;
9860 STRLEN verb_len = 0;
9861 char *start_arg = NULL;
9862 unsigned char op = 0;
9863 int arg_required = 0;
9864 int internal_argval = -1; /* if >-1 we are not allowed an argument*/
9866 if (has_intervening_patws) {
9868 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
9870 while ( *RExC_parse && *RExC_parse != ')' ) {
9871 if ( *RExC_parse == ':' ) {
9872 start_arg = RExC_parse + 1;
9878 verb_len = RExC_parse - start_verb;
9881 while ( *RExC_parse && *RExC_parse != ')' )
9883 if ( *RExC_parse != ')' )
9884 vFAIL("Unterminated verb pattern argument");
9885 if ( RExC_parse == start_arg )
9888 if ( *RExC_parse != ')' )
9889 vFAIL("Unterminated verb pattern");
9892 switch ( *start_verb ) {
9893 case 'A': /* (*ACCEPT) */
9894 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9896 internal_argval = RExC_nestroot;
9899 case 'C': /* (*COMMIT) */
9900 if ( memEQs(start_verb,verb_len,"COMMIT") )
9903 case 'F': /* (*FAIL) */
9904 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9908 case ':': /* (*:NAME) */
9909 case 'M': /* (*MARK:NAME) */
9910 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9915 case 'P': /* (*PRUNE) */
9916 if ( memEQs(start_verb,verb_len,"PRUNE") )
9919 case 'S': /* (*SKIP) */
9920 if ( memEQs(start_verb,verb_len,"SKIP") )
9923 case 'T': /* (*THEN) */
9924 /* [19:06] <TimToady> :: is then */
9925 if ( memEQs(start_verb,verb_len,"THEN") ) {
9927 RExC_seen |= REG_CUTGROUP_SEEN;
9932 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9934 "Unknown verb pattern '%"UTF8f"'",
9935 UTF8fARG(UTF, verb_len, start_verb));
9937 if ( arg_required && !start_arg ) {
9938 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9939 verb_len, start_verb);
9941 if (internal_argval == -1) {
9942 ret = reganode(pRExC_state, op, 0);
9944 ret = reg2Lanode(pRExC_state, op, 0, internal_argval);
9946 RExC_seen |= REG_VERBARG_SEEN;
9947 if ( ! SIZE_ONLY ) {
9949 SV *sv = newSVpvn( start_arg,
9950 RExC_parse - start_arg);
9951 ARG(ret) = add_data( pRExC_state,
9953 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9958 if ( internal_argval != -1 )
9959 ARG2L_SET(ret, internal_argval);
9961 nextchar(pRExC_state);
9964 else if (*RExC_parse == '?') { /* (?...) */
9965 bool is_logical = 0;
9966 const char * const seqstart = RExC_parse;
9967 const char * endptr;
9968 if (has_intervening_patws) {
9970 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
9974 paren = *RExC_parse++;
9975 ret = NULL; /* For look-ahead/behind. */
9978 case 'P': /* (?P...) variants for those used to PCRE/Python */
9979 paren = *RExC_parse++;
9980 if ( paren == '<') /* (?P<...>) named capture */
9982 else if (paren == '>') { /* (?P>name) named recursion */
9983 goto named_recursion;
9985 else if (paren == '=') { /* (?P=...) named backref */
9986 /* this pretty much dupes the code for \k<NAME> in
9987 * regatom(), if you change this make sure you change that
9989 char* name_start = RExC_parse;
9991 SV *sv_dat = reg_scan_name(pRExC_state,
9992 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9993 if (RExC_parse == name_start || *RExC_parse != ')')
9994 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9995 vFAIL2("Sequence %.3s... not terminated",parse_start);
9998 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9999 RExC_rxi->data->data[num]=(void*)sv_dat;
10000 SvREFCNT_inc_simple_void(sv_dat);
10003 ret = reganode(pRExC_state,
10006 : (ASCII_FOLD_RESTRICTED)
10008 : (AT_LEAST_UNI_SEMANTICS)
10014 *flagp |= HASWIDTH;
10016 Set_Node_Offset(ret, parse_start+1);
10017 Set_Node_Cur_Length(ret, parse_start);
10019 nextchar(pRExC_state);
10023 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10024 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10025 vFAIL3("Sequence (%.*s...) not recognized",
10026 RExC_parse-seqstart, seqstart);
10027 NOT_REACHED; /*NOTREACHED*/
10028 case '<': /* (?<...) */
10029 if (*RExC_parse == '!')
10031 else if (*RExC_parse != '=')
10037 case '\'': /* (?'...') */
10038 name_start= RExC_parse;
10039 svname = reg_scan_name(pRExC_state,
10040 SIZE_ONLY /* reverse test from the others */
10041 ? REG_RSN_RETURN_NAME
10042 : REG_RSN_RETURN_NULL);
10043 if (RExC_parse == name_start || *RExC_parse != paren)
10044 vFAIL2("Sequence (?%c... not terminated",
10045 paren=='>' ? '<' : paren);
10049 if (!svname) /* shouldn't happen */
10051 "panic: reg_scan_name returned NULL");
10052 if (!RExC_paren_names) {
10053 RExC_paren_names= newHV();
10054 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10056 RExC_paren_name_list= newAV();
10057 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10060 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10062 sv_dat = HeVAL(he_str);
10064 /* croak baby croak */
10066 "panic: paren_name hash element allocation failed");
10067 } else if ( SvPOK(sv_dat) ) {
10068 /* (?|...) can mean we have dupes so scan to check
10069 its already been stored. Maybe a flag indicating
10070 we are inside such a construct would be useful,
10071 but the arrays are likely to be quite small, so
10072 for now we punt -- dmq */
10073 IV count = SvIV(sv_dat);
10074 I32 *pv = (I32*)SvPVX(sv_dat);
10076 for ( i = 0 ; i < count ; i++ ) {
10077 if ( pv[i] == RExC_npar ) {
10083 pv = (I32*)SvGROW(sv_dat,
10084 SvCUR(sv_dat) + sizeof(I32)+1);
10085 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10086 pv[count] = RExC_npar;
10087 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10090 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10091 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10094 SvIV_set(sv_dat, 1);
10097 /* Yes this does cause a memory leak in debugging Perls
10099 if (!av_store(RExC_paren_name_list,
10100 RExC_npar, SvREFCNT_inc(svname)))
10101 SvREFCNT_dec_NN(svname);
10104 /*sv_dump(sv_dat);*/
10106 nextchar(pRExC_state);
10108 goto capturing_parens;
10110 RExC_seen |= REG_LOOKBEHIND_SEEN;
10111 RExC_in_lookbehind++;
10114 case '=': /* (?=...) */
10115 RExC_seen_zerolen++;
10117 case '!': /* (?!...) */
10118 RExC_seen_zerolen++;
10119 /* check if we're really just a "FAIL" assertion */
10120 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10121 FALSE /* Don't force to /x */ );
10122 if (*RExC_parse == ')') {
10123 ret=reganode(pRExC_state, OPFAIL, 0);
10124 nextchar(pRExC_state);
10128 case '|': /* (?|...) */
10129 /* branch reset, behave like a (?:...) except that
10130 buffers in alternations share the same numbers */
10132 after_freeze = freeze_paren = RExC_npar;
10134 case ':': /* (?:...) */
10135 case '>': /* (?>...) */
10137 case '$': /* (?$...) */
10138 case '@': /* (?@...) */
10139 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10141 case '0' : /* (?0) */
10142 case 'R' : /* (?R) */
10143 if (*RExC_parse != ')')
10144 FAIL("Sequence (?R) not terminated");
10145 ret = reg_node(pRExC_state, GOSTART);
10146 RExC_seen |= REG_GOSTART_SEEN;
10147 *flagp |= POSTPONED;
10148 nextchar(pRExC_state);
10151 /* named and numeric backreferences */
10152 case '&': /* (?&NAME) */
10153 parse_start = RExC_parse - 1;
10156 SV *sv_dat = reg_scan_name(pRExC_state,
10157 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10158 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10160 if (RExC_parse == RExC_end || *RExC_parse != ')')
10161 vFAIL("Sequence (?&... not terminated");
10162 goto gen_recurse_regop;
10165 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10167 vFAIL("Illegal pattern");
10169 goto parse_recursion;
10171 case '-': /* (?-1) */
10172 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10173 RExC_parse--; /* rewind to let it be handled later */
10177 case '1': case '2': case '3': case '4': /* (?1) */
10178 case '5': case '6': case '7': case '8': case '9':
10182 bool is_neg = FALSE;
10184 parse_start = RExC_parse - 1; /* MJD */
10185 if (*RExC_parse == '-') {
10189 if (grok_atoUV(RExC_parse, &unum, &endptr)
10193 RExC_parse = (char*)endptr;
10197 /* Some limit for num? */
10201 if (*RExC_parse!=')')
10202 vFAIL("Expecting close bracket");
10205 if ( paren == '-' ) {
10207 Diagram of capture buffer numbering.
10208 Top line is the normal capture buffer numbers
10209 Bottom line is the negative indexing as from
10213 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10217 num = RExC_npar + num;
10220 vFAIL("Reference to nonexistent group");
10222 } else if ( paren == '+' ) {
10223 num = RExC_npar + num - 1;
10226 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10228 if (num > (I32)RExC_rx->nparens) {
10230 vFAIL("Reference to nonexistent group");
10232 RExC_recurse_count++;
10233 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10234 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
10235 22, "| |", (int)(depth * 2 + 1), "",
10236 (UV)ARG(ret), (IV)ARG2L(ret)));
10238 RExC_seen |= REG_RECURSE_SEEN;
10239 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10240 Set_Node_Offset(ret, parse_start); /* MJD */
10242 *flagp |= POSTPONED;
10243 nextchar(pRExC_state);
10248 case '?': /* (??...) */
10250 if (*RExC_parse != '{') {
10251 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10252 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10254 "Sequence (%"UTF8f"...) not recognized",
10255 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10256 NOT_REACHED; /*NOTREACHED*/
10258 *flagp |= POSTPONED;
10259 paren = *RExC_parse++;
10261 case '{': /* (?{...}) */
10264 struct reg_code_block *cb;
10266 RExC_seen_zerolen++;
10268 if ( !pRExC_state->num_code_blocks
10269 || pRExC_state->code_index >= pRExC_state->num_code_blocks
10270 || pRExC_state->code_blocks[pRExC_state->code_index].start
10271 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
10274 if (RExC_pm_flags & PMf_USE_RE_EVAL)
10275 FAIL("panic: Sequence (?{...}): no code block found\n");
10276 FAIL("Eval-group not allowed at runtime, use re 'eval'");
10278 /* this is a pre-compiled code block (?{...}) */
10279 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
10280 RExC_parse = RExC_start + cb->end;
10283 if (cb->src_regex) {
10284 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
10285 RExC_rxi->data->data[n] =
10286 (void*)SvREFCNT_inc((SV*)cb->src_regex);
10287 RExC_rxi->data->data[n+1] = (void*)o;
10290 n = add_data(pRExC_state,
10291 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
10292 RExC_rxi->data->data[n] = (void*)o;
10295 pRExC_state->code_index++;
10296 nextchar(pRExC_state);
10300 ret = reg_node(pRExC_state, LOGICAL);
10302 eval = reg2Lanode(pRExC_state, EVAL,
10305 /* for later propagation into (??{})
10307 RExC_flags & RXf_PMf_COMPILETIME
10312 REGTAIL(pRExC_state, ret, eval);
10313 /* deal with the length of this later - MJD */
10316 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
10317 Set_Node_Length(ret, RExC_parse - parse_start + 1);
10318 Set_Node_Offset(ret, parse_start);
10321 case '(': /* (?(?{...})...) and (?(?=...)...) */
10324 const int DEFINE_len = sizeof("DEFINE") - 1;
10325 if (RExC_parse[0] == '?') { /* (?(?...)) */
10327 RExC_parse[1] == '=' ||
10328 RExC_parse[1] == '!' ||
10329 RExC_parse[1] == '<' ||
10330 RExC_parse[1] == '{'
10331 ) { /* Lookahead or eval. */
10335 ret = reg_node(pRExC_state, LOGICAL);
10339 tail = reg(pRExC_state, 1, &flag, depth+1);
10340 if (flag & (RESTART_PASS1|NEED_UTF8)) {
10341 *flagp = flag & (RESTART_PASS1|NEED_UTF8);
10344 REGTAIL(pRExC_state, ret, tail);
10347 /* Fall through to ‘Unknown switch condition’ at the
10348 end of the if/else chain. */
10350 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
10351 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
10353 char ch = RExC_parse[0] == '<' ? '>' : '\'';
10354 char *name_start= RExC_parse++;
10356 SV *sv_dat=reg_scan_name(pRExC_state,
10357 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10358 if (RExC_parse == name_start || *RExC_parse != ch)
10359 vFAIL2("Sequence (?(%c... not terminated",
10360 (ch == '>' ? '<' : ch));
10363 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10364 RExC_rxi->data->data[num]=(void*)sv_dat;
10365 SvREFCNT_inc_simple_void(sv_dat);
10367 ret = reganode(pRExC_state,NGROUPP,num);
10368 goto insert_if_check_paren;
10370 else if (RExC_end - RExC_parse >= DEFINE_len
10371 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
10373 ret = reganode(pRExC_state,DEFINEP,0);
10374 RExC_parse += DEFINE_len;
10376 goto insert_if_check_paren;
10378 else if (RExC_parse[0] == 'R') {
10381 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10383 if (grok_atoUV(RExC_parse, &uv, &endptr)
10387 RExC_parse = (char*)endptr;
10389 /* else "Switch condition not recognized" below */
10390 } else if (RExC_parse[0] == '&') {
10393 sv_dat = reg_scan_name(pRExC_state,
10395 ? REG_RSN_RETURN_NULL
10396 : REG_RSN_RETURN_DATA);
10397 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10399 ret = reganode(pRExC_state,INSUBP,parno);
10400 goto insert_if_check_paren;
10402 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10406 if (grok_atoUV(RExC_parse, &uv, &endptr)
10410 RExC_parse = (char*)endptr;
10413 vFAIL("panic: grok_atoUV returned FALSE");
10415 ret = reganode(pRExC_state, GROUPP, parno);
10417 insert_if_check_paren:
10418 if (UCHARAT(RExC_parse) != ')') {
10419 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10420 vFAIL("Switch condition not recognized");
10422 nextchar(pRExC_state);
10424 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
10425 br = regbranch(pRExC_state, &flags, 1,depth+1);
10427 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10428 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10431 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10434 REGTAIL(pRExC_state, br, reganode(pRExC_state,
10436 c = UCHARAT(RExC_parse);
10437 nextchar(pRExC_state);
10438 if (flags&HASWIDTH)
10439 *flagp |= HASWIDTH;
10442 vFAIL("(?(DEFINE)....) does not allow branches");
10444 /* Fake one for optimizer. */
10445 lastbr = reganode(pRExC_state, IFTHEN, 0);
10447 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
10448 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10449 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10452 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10455 REGTAIL(pRExC_state, ret, lastbr);
10456 if (flags&HASWIDTH)
10457 *flagp |= HASWIDTH;
10458 c = UCHARAT(RExC_parse);
10459 nextchar(pRExC_state);
10464 if (RExC_parse>RExC_end)
10465 vFAIL("Switch (?(condition)... not terminated");
10467 vFAIL("Switch (?(condition)... contains too many branches");
10469 ender = reg_node(pRExC_state, TAIL);
10470 REGTAIL(pRExC_state, br, ender);
10472 REGTAIL(pRExC_state, lastbr, ender);
10473 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10476 REGTAIL(pRExC_state, ret, ender);
10477 RExC_size++; /* XXX WHY do we need this?!!
10478 For large programs it seems to be required
10479 but I can't figure out why. -- dmq*/
10482 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10483 vFAIL("Unknown switch condition (?(...))");
10485 case '[': /* (?[ ... ]) */
10486 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10489 RExC_parse--; /* for vFAIL to print correctly */
10490 vFAIL("Sequence (? incomplete");
10492 default: /* e.g., (?i) */
10495 parse_lparen_question_flags(pRExC_state);
10496 if (UCHARAT(RExC_parse) != ':') {
10498 nextchar(pRExC_state);
10503 nextchar(pRExC_state);
10508 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
10513 ret = reganode(pRExC_state, OPEN, parno);
10515 if (!RExC_nestroot)
10516 RExC_nestroot = parno;
10517 if (RExC_seen & REG_RECURSE_SEEN
10518 && !RExC_open_parens[parno-1])
10520 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10521 "%*s%*s Setting open paren #%"IVdf" to %d\n",
10522 22, "| |", (int)(depth * 2 + 1), "",
10523 (IV)parno, REG_NODE_NUM(ret)));
10524 RExC_open_parens[parno-1]= ret;
10527 Set_Node_Length(ret, 1); /* MJD */
10528 Set_Node_Offset(ret, RExC_parse); /* MJD */
10531 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
10540 /* Pick up the branches, linking them together. */
10541 parse_start = RExC_parse; /* MJD */
10542 br = regbranch(pRExC_state, &flags, 1,depth+1);
10544 /* branch_len = (paren != 0); */
10547 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10548 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10551 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10553 if (*RExC_parse == '|') {
10554 if (!SIZE_ONLY && RExC_extralen) {
10555 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10558 reginsert(pRExC_state, BRANCH, br, depth+1);
10559 Set_Node_Length(br, paren != 0);
10560 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10564 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10566 else if (paren == ':') {
10567 *flagp |= flags&SIMPLE;
10569 if (is_open) { /* Starts with OPEN. */
10570 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10572 else if (paren != '?') /* Not Conditional */
10574 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10576 while (*RExC_parse == '|') {
10577 if (!SIZE_ONLY && RExC_extralen) {
10578 ender = reganode(pRExC_state, LONGJMP,0);
10580 /* Append to the previous. */
10581 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10584 RExC_extralen += 2; /* Account for LONGJMP. */
10585 nextchar(pRExC_state);
10586 if (freeze_paren) {
10587 if (RExC_npar > after_freeze)
10588 after_freeze = RExC_npar;
10589 RExC_npar = freeze_paren;
10591 br = regbranch(pRExC_state, &flags, 0, depth+1);
10594 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10595 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10598 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10600 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10602 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10605 if (have_branch || paren != ':') {
10606 /* Make a closing node, and hook it on the end. */
10609 ender = reg_node(pRExC_state, TAIL);
10612 ender = reganode(pRExC_state, CLOSE, parno);
10613 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10614 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10615 "%*s%*s Setting close paren #%"IVdf" to %d\n",
10616 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
10617 RExC_close_parens[parno-1]= ender;
10618 if (RExC_nestroot == parno)
10621 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10622 Set_Node_Length(ender,1); /* MJD */
10628 *flagp &= ~HASWIDTH;
10631 ender = reg_node(pRExC_state, SUCCEED);
10634 ender = reg_node(pRExC_state, END);
10636 assert(!RExC_opend); /* there can only be one! */
10637 RExC_opend = ender;
10641 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10642 DEBUG_PARSE_MSG("lsbr");
10643 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
10644 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10645 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10646 SvPV_nolen_const(RExC_mysv1),
10647 (IV)REG_NODE_NUM(lastbr),
10648 SvPV_nolen_const(RExC_mysv2),
10649 (IV)REG_NODE_NUM(ender),
10650 (IV)(ender - lastbr)
10653 REGTAIL(pRExC_state, lastbr, ender);
10655 if (have_branch && !SIZE_ONLY) {
10656 char is_nothing= 1;
10658 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10660 /* Hook the tails of the branches to the closing node. */
10661 for (br = ret; br; br = regnext(br)) {
10662 const U8 op = PL_regkind[OP(br)];
10663 if (op == BRANCH) {
10664 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10665 if ( OP(NEXTOPER(br)) != NOTHING
10666 || regnext(NEXTOPER(br)) != ender)
10669 else if (op == BRANCHJ) {
10670 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10671 /* for now we always disable this optimisation * /
10672 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10673 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10679 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10680 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10681 DEBUG_PARSE_MSG("NADA");
10682 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
10683 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10684 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10685 SvPV_nolen_const(RExC_mysv1),
10686 (IV)REG_NODE_NUM(ret),
10687 SvPV_nolen_const(RExC_mysv2),
10688 (IV)REG_NODE_NUM(ender),
10693 if (OP(ender) == TAIL) {
10698 for ( opt= br + 1; opt < ender ; opt++ )
10699 OP(opt)= OPTIMIZED;
10700 NEXT_OFF(br)= ender - br;
10708 static const char parens[] = "=!<,>";
10710 if (paren && (p = strchr(parens, paren))) {
10711 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10712 int flag = (p - parens) > 1;
10715 node = SUSPEND, flag = 0;
10716 reginsert(pRExC_state, node,ret, depth+1);
10717 Set_Node_Cur_Length(ret, parse_start);
10718 Set_Node_Offset(ret, parse_start + 1);
10720 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10724 /* Check for proper termination. */
10726 /* restore original flags, but keep (?p) and, if we've changed from /d
10727 * rules to /u, keep the /u */
10728 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10729 if (DEPENDS_SEMANTICS && RExC_uni_semantics) {
10730 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
10732 if (RExC_parse >= RExC_end || UCHARAT(RExC_parse) != ')') {
10733 RExC_parse = oregcomp_parse;
10734 vFAIL("Unmatched (");
10736 nextchar(pRExC_state);
10738 else if (!paren && RExC_parse < RExC_end) {
10739 if (*RExC_parse == ')') {
10741 vFAIL("Unmatched )");
10744 FAIL("Junk on end of regexp"); /* "Can't happen". */
10745 NOT_REACHED; /* NOTREACHED */
10748 if (RExC_in_lookbehind) {
10749 RExC_in_lookbehind--;
10751 if (after_freeze > RExC_npar)
10752 RExC_npar = after_freeze;
10757 - regbranch - one alternative of an | operator
10759 * Implements the concatenation operator.
10761 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
10762 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
10765 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10768 regnode *chain = NULL;
10770 I32 flags = 0, c = 0;
10771 GET_RE_DEBUG_FLAGS_DECL;
10773 PERL_ARGS_ASSERT_REGBRANCH;
10775 DEBUG_PARSE("brnc");
10780 if (!SIZE_ONLY && RExC_extralen)
10781 ret = reganode(pRExC_state, BRANCHJ,0);
10783 ret = reg_node(pRExC_state, BRANCH);
10784 Set_Node_Length(ret, 1);
10788 if (!first && SIZE_ONLY)
10789 RExC_extralen += 1; /* BRANCHJ */
10791 *flagp = WORST; /* Tentatively. */
10793 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
10794 FALSE /* Don't force to /x */ );
10795 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10796 flags &= ~TRYAGAIN;
10797 latest = regpiece(pRExC_state, &flags,depth+1);
10798 if (latest == NULL) {
10799 if (flags & TRYAGAIN)
10801 if (flags & (RESTART_PASS1|NEED_UTF8)) {
10802 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
10805 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10807 else if (ret == NULL)
10809 *flagp |= flags&(HASWIDTH|POSTPONED);
10810 if (chain == NULL) /* First piece. */
10811 *flagp |= flags&SPSTART;
10813 /* FIXME adding one for every branch after the first is probably
10814 * excessive now we have TRIE support. (hv) */
10816 REGTAIL(pRExC_state, chain, latest);
10821 if (chain == NULL) { /* Loop ran zero times. */
10822 chain = reg_node(pRExC_state, NOTHING);
10827 *flagp |= flags&SIMPLE;
10834 - regpiece - something followed by possible [*+?]
10836 * Note that the branching code sequences used for ? and the general cases
10837 * of * and + are somewhat optimized: they use the same NOTHING node as
10838 * both the endmarker for their branch list and the body of the last branch.
10839 * It might seem that this node could be dispensed with entirely, but the
10840 * endmarker role is not redundant.
10842 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10844 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
10845 * restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
10848 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10854 const char * const origparse = RExC_parse;
10856 I32 max = REG_INFTY;
10857 #ifdef RE_TRACK_PATTERN_OFFSETS
10860 const char *maxpos = NULL;
10863 /* Save the original in case we change the emitted regop to a FAIL. */
10864 regnode * const orig_emit = RExC_emit;
10866 GET_RE_DEBUG_FLAGS_DECL;
10868 PERL_ARGS_ASSERT_REGPIECE;
10870 DEBUG_PARSE("piec");
10872 ret = regatom(pRExC_state, &flags,depth+1);
10874 if (flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8))
10875 *flagp |= flags & (TRYAGAIN|RESTART_PASS1|NEED_UTF8);
10877 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10883 if (op == '{' && regcurly(RExC_parse)) {
10885 #ifdef RE_TRACK_PATTERN_OFFSETS
10886 parse_start = RExC_parse; /* MJD */
10888 next = RExC_parse + 1;
10889 while (isDIGIT(*next) || *next == ',') {
10890 if (*next == ',') {
10898 if (*next == '}') { /* got one */
10899 const char* endptr;
10903 if (isDIGIT(*RExC_parse)) {
10904 if (!grok_atoUV(RExC_parse, &uv, &endptr))
10905 vFAIL("Invalid quantifier in {,}");
10906 if (uv >= REG_INFTY)
10907 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10912 if (*maxpos == ',')
10915 maxpos = RExC_parse;
10916 if (isDIGIT(*maxpos)) {
10917 if (!grok_atoUV(maxpos, &uv, &endptr))
10918 vFAIL("Invalid quantifier in {,}");
10919 if (uv >= REG_INFTY)
10920 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10923 max = REG_INFTY; /* meaning "infinity" */
10926 nextchar(pRExC_state);
10927 if (max < min) { /* If can't match, warn and optimize to fail
10931 /* We can't back off the size because we have to reserve
10932 * enough space for all the things we are about to throw
10933 * away, but we can shrink it by the ammount we are about
10934 * to re-use here */
10935 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10938 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10939 RExC_emit = orig_emit;
10941 ret = reganode(pRExC_state, OPFAIL, 0);
10944 else if (min == max && RExC_parse < RExC_end && *RExC_parse == '?')
10947 ckWARN2reg(RExC_parse + 1,
10948 "Useless use of greediness modifier '%c'",
10951 /* Absorb the modifier, so later code doesn't see nor use it */
10952 nextchar(pRExC_state);
10956 if ((flags&SIMPLE)) {
10957 if (min == 0 && max == REG_INFTY) {
10958 reginsert(pRExC_state, STAR, ret, depth+1);
10961 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10964 if (min == 1 && max == REG_INFTY) {
10965 reginsert(pRExC_state, PLUS, ret, depth+1);
10968 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10971 MARK_NAUGHTY_EXP(2, 2);
10972 reginsert(pRExC_state, CURLY, ret, depth+1);
10973 Set_Node_Offset(ret, parse_start+1); /* MJD */
10974 Set_Node_Cur_Length(ret, parse_start);
10977 regnode * const w = reg_node(pRExC_state, WHILEM);
10980 REGTAIL(pRExC_state, ret, w);
10981 if (!SIZE_ONLY && RExC_extralen) {
10982 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10983 reginsert(pRExC_state, NOTHING,ret, depth+1);
10984 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10986 reginsert(pRExC_state, CURLYX,ret, depth+1);
10988 Set_Node_Offset(ret, parse_start+1);
10989 Set_Node_Length(ret,
10990 op == '{' ? (RExC_parse - parse_start) : 1);
10992 if (!SIZE_ONLY && RExC_extralen)
10993 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10994 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10996 RExC_whilem_seen++, RExC_extralen += 3;
10997 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
11004 *flagp |= HASWIDTH;
11006 ARG1_SET(ret, (U16)min);
11007 ARG2_SET(ret, (U16)max);
11009 if (max == REG_INFTY)
11010 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
11016 if (!ISMULT1(op)) {
11021 #if 0 /* Now runtime fix should be reliable. */
11023 /* if this is reinstated, don't forget to put this back into perldiag:
11025 =item Regexp *+ operand could be empty at {#} in regex m/%s/
11027 (F) The part of the regexp subject to either the * or + quantifier
11028 could match an empty string. The {#} shows in the regular
11029 expression about where the problem was discovered.
11033 if (!(flags&HASWIDTH) && op != '?')
11034 vFAIL("Regexp *+ operand could be empty");
11037 #ifdef RE_TRACK_PATTERN_OFFSETS
11038 parse_start = RExC_parse;
11040 nextchar(pRExC_state);
11042 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
11048 else if (op == '+') {
11052 else if (op == '?') {
11057 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
11058 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
11059 ckWARN2reg(RExC_parse,
11060 "%"UTF8f" matches null string many times",
11061 UTF8fARG(UTF, (RExC_parse >= origparse
11062 ? RExC_parse - origparse
11065 (void)ReREFCNT_inc(RExC_rx_sv);
11068 if (RExC_parse < RExC_end && *RExC_parse == '?') {
11069 nextchar(pRExC_state);
11070 reginsert(pRExC_state, MINMOD, ret, depth+1);
11071 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11074 if (RExC_parse < RExC_end && *RExC_parse == '+') {
11076 nextchar(pRExC_state);
11077 ender = reg_node(pRExC_state, SUCCEED);
11078 REGTAIL(pRExC_state, ret, ender);
11079 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11081 ender = reg_node(pRExC_state, TAIL);
11082 REGTAIL(pRExC_state, ret, ender);
11085 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
11087 vFAIL("Nested quantifiers");
11094 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11102 /* This routine teases apart the various meanings of \N and returns
11103 * accordingly. The input parameters constrain which meaning(s) is/are valid
11104 * in the current context.
11106 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11108 * If <code_point_p> is not NULL, the context is expecting the result to be a
11109 * single code point. If this \N instance turns out to a single code point,
11110 * the function returns TRUE and sets *code_point_p to that code point.
11112 * If <node_p> is not NULL, the context is expecting the result to be one of
11113 * the things representable by a regnode. If this \N instance turns out to be
11114 * one such, the function generates the regnode, returns TRUE and sets *node_p
11115 * to point to that regnode.
11117 * If this instance of \N isn't legal in any context, this function will
11118 * generate a fatal error and not return.
11120 * On input, RExC_parse should point to the first char following the \N at the
11121 * time of the call. On successful return, RExC_parse will have been updated
11122 * to point to just after the sequence identified by this routine. Also
11123 * *flagp has been updated as needed.
11125 * When there is some problem with the current context and this \N instance,
11126 * the function returns FALSE, without advancing RExC_parse, nor setting
11127 * *node_p, nor *code_point_p, nor *flagp.
11129 * If <cp_count> is not NULL, the caller wants to know the length (in code
11130 * points) that this \N sequence matches. This is set even if the function
11131 * returns FALSE, as detailed below.
11133 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11135 * Probably the most common case is for the \N to specify a single code point.
11136 * *cp_count will be set to 1, and *code_point_p will be set to that code
11139 * Another possibility is for the input to be an empty \N{}, which for
11140 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11141 * will be set to a generated NOTHING node.
11143 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11144 * set to 0. *node_p will be set to a generated REG_ANY node.
11146 * The fourth possibility is that \N resolves to a sequence of more than one
11147 * code points. *cp_count will be set to the number of code points in the
11148 * sequence. *node_p * will be set to a generated node returned by this
11149 * function calling S_reg().
11151 * The final possibility is that it is premature to be calling this function;
11152 * that pass1 needs to be restarted. This can happen when this changes from
11153 * /d to /u rules, or when the pattern needs to be upgraded to UTF-8. The
11154 * latter occurs only when the fourth possibility would otherwise be in
11155 * effect, and is because one of those code points requires the pattern to be
11156 * recompiled as UTF-8. The function returns FALSE, and sets the
11157 * RESTART_PASS1 and NEED_UTF8 flags in *flagp, as appropriate. When this
11158 * happens, the caller needs to desist from continuing parsing, and return
11159 * this information to its caller. This is not set for when there is only one
11160 * code point, as this can be called as part of an ANYOF node, and they can
11161 * store above-Latin1 code points without the pattern having to be in UTF-8.
11163 * For non-single-quoted regexes, the tokenizer has resolved character and
11164 * sequence names inside \N{...} into their Unicode values, normalizing the
11165 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11166 * hex-represented code points in the sequence. This is done there because
11167 * the names can vary based on what charnames pragma is in scope at the time,
11168 * so we need a way to take a snapshot of what they resolve to at the time of
11169 * the original parse. [perl #56444].
11171 * That parsing is skipped for single-quoted regexes, so we may here get
11172 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11173 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11174 * is legal and handled here. The code point is Unicode, and has to be
11175 * translated into the native character set for non-ASCII platforms.
11178 char * endbrace; /* points to '}' following the name */
11179 char *endchar; /* Points to '.' or '}' ending cur char in the input
11181 char* p; /* Temporary */
11183 GET_RE_DEBUG_FLAGS_DECL;
11185 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11187 GET_RE_DEBUG_FLAGS;
11189 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11190 assert(! (node_p && cp_count)); /* At most 1 should be set */
11192 if (cp_count) { /* Initialize return for the most common case */
11196 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11197 * modifier. The other meanings do not, so use a temporary until we find
11198 * out which we are being called with */
11199 p = (RExC_flags & RXf_PMf_EXTENDED)
11200 ? regpatws(pRExC_state, RExC_parse,
11201 TRUE) /* means recognize comments */
11204 /* Disambiguate between \N meaning a named character versus \N meaning
11205 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11206 * quantifier, or there is no '{' at all */
11207 if (*p != '{' || regcurly(p)) {
11216 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
11217 FALSE /* Don't force to /x */ );
11218 *node_p = reg_node(pRExC_state, REG_ANY);
11219 *flagp |= HASWIDTH|SIMPLE;
11221 Set_Node_Length(*node_p, 1); /* MJD */
11225 /* Here, we have decided it should be a named character or sequence */
11227 /* The test above made sure that the next real character is a '{', but
11228 * under the /x modifier, it could be separated by space (or a comment and
11229 * \n) and this is not allowed (for consistency with \x{...} and the
11230 * tokenizer handling of \N{NAME}). */
11231 if (*RExC_parse != '{') {
11232 vFAIL("Missing braces on \\N{}");
11235 RExC_parse++; /* Skip past the '{' */
11237 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
11238 || ! (endbrace == RExC_parse /* nothing between the {} */
11239 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11240 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
11243 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
11244 vFAIL("\\N{NAME} must be resolved by the lexer");
11247 REQUIRE_UNI_RULES(flagp, FALSE); /* Unicode named chars imply Unicode
11250 if (endbrace == RExC_parse) { /* empty: \N{} */
11254 nextchar(pRExC_state);
11259 *node_p = reg_node(pRExC_state,NOTHING);
11263 RExC_parse += 2; /* Skip past the 'U+' */
11265 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11267 /* Code points are separated by dots. If none, there is only one code
11268 * point, and is terminated by the brace */
11270 if (endchar >= endbrace) {
11271 STRLEN length_of_hex;
11272 I32 grok_hex_flags;
11274 /* Here, exactly one code point. If that isn't what is wanted, fail */
11275 if (! code_point_p) {
11280 /* Convert code point from hex */
11281 length_of_hex = (STRLEN)(endchar - RExC_parse);
11282 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
11283 | PERL_SCAN_DISALLOW_PREFIX
11285 /* No errors in the first pass (See [perl
11286 * #122671].) We let the code below find the
11287 * errors when there are multiple chars. */
11289 ? PERL_SCAN_SILENT_ILLDIGIT
11292 /* This routine is the one place where both single- and double-quotish
11293 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
11294 * must be converted to native. */
11295 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
11300 /* The tokenizer should have guaranteed validity, but it's possible to
11301 * bypass it by using single quoting, so check. Don't do the check
11302 * here when there are multiple chars; we do it below anyway. */
11303 if (length_of_hex == 0
11304 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
11306 RExC_parse += length_of_hex; /* Includes all the valid */
11307 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
11308 ? UTF8SKIP(RExC_parse)
11310 /* Guard against malformed utf8 */
11311 if (RExC_parse >= endchar) {
11312 RExC_parse = endchar;
11314 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11317 RExC_parse = endbrace + 1;
11320 else { /* Is a multiple character sequence */
11321 SV * substitute_parse;
11323 char *orig_end = RExC_end;
11326 /* Count the code points, if desired, in the sequence */
11329 while (RExC_parse < endbrace) {
11330 /* Point to the beginning of the next character in the sequence. */
11331 RExC_parse = endchar + 1;
11332 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11337 /* Fail if caller doesn't want to handle a multi-code-point sequence.
11338 * But don't backup up the pointer if the caller want to know how many
11339 * code points there are (they can then handle things) */
11347 /* What is done here is to convert this to a sub-pattern of the form
11348 * \x{char1}\x{char2}... and then call reg recursively to parse it
11349 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
11350 * while not having to worry about special handling that some code
11351 * points may have. */
11353 substitute_parse = newSVpvs("?:");
11355 while (RExC_parse < endbrace) {
11357 /* Convert to notation the rest of the code understands */
11358 sv_catpv(substitute_parse, "\\x{");
11359 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
11360 sv_catpv(substitute_parse, "}");
11362 /* Point to the beginning of the next character in the sequence. */
11363 RExC_parse = endchar + 1;
11364 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11367 sv_catpv(substitute_parse, ")");
11369 RExC_parse = SvPV(substitute_parse, len);
11371 /* Don't allow empty number */
11372 if (len < (STRLEN) 8) {
11373 RExC_parse = endbrace;
11374 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11376 RExC_end = RExC_parse + len;
11378 /* The values are Unicode, and therefore not subject to recoding, but
11379 * have to be converted to native on a non-Unicode (meaning non-ASCII)
11381 RExC_override_recoding = 1;
11383 RExC_recode_x_to_native = 1;
11387 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
11388 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11389 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11392 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
11395 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11398 /* Restore the saved values */
11399 RExC_parse = endbrace;
11400 RExC_end = orig_end;
11401 RExC_override_recoding = 0;
11403 RExC_recode_x_to_native = 0;
11406 SvREFCNT_dec_NN(substitute_parse);
11407 nextchar(pRExC_state);
11417 * It returns the code point in utf8 for the value in *encp.
11418 * value: a code value in the source encoding
11419 * encp: a pointer to an Encode object
11421 * If the result from Encode is not a single character,
11422 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
11425 S_reg_recode(pTHX_ const char value, SV **encp)
11428 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
11429 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
11430 const STRLEN newlen = SvCUR(sv);
11431 UV uv = UNICODE_REPLACEMENT;
11433 PERL_ARGS_ASSERT_REG_RECODE;
11437 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
11440 if (!newlen || numlen != newlen) {
11441 uv = UNICODE_REPLACEMENT;
11447 PERL_STATIC_INLINE U8
11448 S_compute_EXACTish(RExC_state_t *pRExC_state)
11452 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
11460 op = get_regex_charset(RExC_flags);
11461 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
11462 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
11463 been, so there is no hole */
11466 return op + EXACTF;
11469 PERL_STATIC_INLINE void
11470 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
11471 regnode *node, I32* flagp, STRLEN len, UV code_point,
11474 /* This knows the details about sizing an EXACTish node, setting flags for
11475 * it (by setting <*flagp>, and potentially populating it with a single
11478 * If <len> (the length in bytes) is non-zero, this function assumes that
11479 * the node has already been populated, and just does the sizing. In this
11480 * case <code_point> should be the final code point that has already been
11481 * placed into the node. This value will be ignored except that under some
11482 * circumstances <*flagp> is set based on it.
11484 * If <len> is zero, the function assumes that the node is to contain only
11485 * the single character given by <code_point> and calculates what <len>
11486 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
11487 * additionally will populate the node's STRING with <code_point> or its
11490 * In both cases <*flagp> is appropriately set
11492 * It knows that under FOLD, the Latin Sharp S and UTF characters above
11493 * 255, must be folded (the former only when the rules indicate it can
11496 * When it does the populating, it looks at the flag 'downgradable'. If
11497 * true with a node that folds, it checks if the single code point
11498 * participates in a fold, and if not downgrades the node to an EXACT.
11499 * This helps the optimizer */
11501 bool len_passed_in = cBOOL(len != 0);
11502 U8 character[UTF8_MAXBYTES_CASE+1];
11504 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
11506 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
11507 * sizing difference, and is extra work that is thrown away */
11508 if (downgradable && ! PASS2) {
11509 downgradable = FALSE;
11512 if (! len_passed_in) {
11514 if (UVCHR_IS_INVARIANT(code_point)) {
11515 if (LOC || ! FOLD) { /* /l defers folding until runtime */
11516 *character = (U8) code_point;
11518 else { /* Here is /i and not /l. (toFOLD() is defined on just
11519 ASCII, which isn't the same thing as INVARIANT on
11520 EBCDIC, but it works there, as the extra invariants
11521 fold to themselves) */
11522 *character = toFOLD((U8) code_point);
11524 /* We can downgrade to an EXACT node if this character
11525 * isn't a folding one. Note that this assumes that
11526 * nothing above Latin1 folds to some other invariant than
11527 * one of these alphabetics; otherwise we would also have
11529 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11530 * || ASCII_FOLD_RESTRICTED))
11532 if (downgradable && PL_fold[code_point] == code_point) {
11538 else if (FOLD && (! LOC
11539 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
11540 { /* Folding, and ok to do so now */
11541 UV folded = _to_uni_fold_flags(
11545 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11546 ? FOLD_FLAGS_NOMIX_ASCII
11549 && folded == code_point /* This quickly rules out many
11550 cases, avoiding the
11551 _invlist_contains_cp() overhead
11553 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11560 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11562 /* Not folding this cp, and can output it directly */
11563 *character = UTF8_TWO_BYTE_HI(code_point);
11564 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11568 uvchr_to_utf8( character, code_point);
11569 len = UTF8SKIP(character);
11571 } /* Else pattern isn't UTF8. */
11573 *character = (U8) code_point;
11575 } /* Else is folded non-UTF8 */
11576 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
11577 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
11578 || UNICODE_DOT_DOT_VERSION > 0)
11579 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11583 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11584 * comments at join_exact()); */
11585 *character = (U8) code_point;
11588 /* Can turn into an EXACT node if we know the fold at compile time,
11589 * and it folds to itself and doesn't particpate in other folds */
11592 && PL_fold_latin1[code_point] == code_point
11593 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11594 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11598 } /* else is Sharp s. May need to fold it */
11599 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11601 *(character + 1) = 's';
11605 *character = LATIN_SMALL_LETTER_SHARP_S;
11611 RExC_size += STR_SZ(len);
11614 RExC_emit += STR_SZ(len);
11615 STR_LEN(node) = len;
11616 if (! len_passed_in) {
11617 Copy((char *) character, STRING(node), len, char);
11621 *flagp |= HASWIDTH;
11623 /* A single character node is SIMPLE, except for the special-cased SHARP S
11625 if ((len == 1 || (UTF && len == UVCHR_SKIP(code_point)))
11626 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
11627 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
11628 || UNICODE_DOT_DOT_VERSION > 0)
11629 && ( code_point != LATIN_SMALL_LETTER_SHARP_S
11630 || ! FOLD || ! DEPENDS_SEMANTICS)
11636 /* The OP may not be well defined in PASS1 */
11637 if (PASS2 && OP(node) == EXACTFL) {
11638 RExC_contains_locale = 1;
11643 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
11644 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11647 S_backref_value(char *p)
11649 const char* endptr;
11651 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
11658 - regatom - the lowest level
11660 Try to identify anything special at the start of the pattern. If there
11661 is, then handle it as required. This may involve generating a single regop,
11662 such as for an assertion; or it may involve recursing, such as to
11663 handle a () structure.
11665 If the string doesn't start with something special then we gobble up
11666 as much literal text as we can.
11668 Once we have been able to handle whatever type of thing started the
11669 sequence, we return.
11671 Note: we have to be careful with escapes, as they can be both literal
11672 and special, and in the case of \10 and friends, context determines which.
11674 A summary of the code structure is:
11676 switch (first_byte) {
11677 cases for each special:
11678 handle this special;
11681 switch (2nd byte) {
11682 cases for each unambiguous special:
11683 handle this special;
11685 cases for each ambigous special/literal:
11687 if (special) handle here
11689 default: // unambiguously literal:
11692 default: // is a literal char
11695 create EXACTish node for literal;
11696 while (more input and node isn't full) {
11697 switch (input_byte) {
11698 cases for each special;
11699 make sure parse pointer is set so that the next call to
11700 regatom will see this special first
11701 goto loopdone; // EXACTish node terminated by prev. char
11703 append char to EXACTISH node;
11705 get next input byte;
11709 return the generated node;
11711 Specifically there are two separate switches for handling
11712 escape sequences, with the one for handling literal escapes requiring
11713 a dummy entry for all of the special escapes that are actually handled
11716 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11718 Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs to be
11719 restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded to UTF-8
11720 Otherwise does not return NULL.
11724 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11726 regnode *ret = NULL;
11728 char *parse_start = RExC_parse;
11733 GET_RE_DEBUG_FLAGS_DECL;
11735 *flagp = WORST; /* Tentatively. */
11737 DEBUG_PARSE("atom");
11739 PERL_ARGS_ASSERT_REGATOM;
11742 switch ((U8)*RExC_parse) {
11744 RExC_seen_zerolen++;
11745 nextchar(pRExC_state);
11746 if (RExC_flags & RXf_PMf_MULTILINE)
11747 ret = reg_node(pRExC_state, MBOL);
11749 ret = reg_node(pRExC_state, SBOL);
11750 Set_Node_Length(ret, 1); /* MJD */
11753 nextchar(pRExC_state);
11755 RExC_seen_zerolen++;
11756 if (RExC_flags & RXf_PMf_MULTILINE)
11757 ret = reg_node(pRExC_state, MEOL);
11759 ret = reg_node(pRExC_state, SEOL);
11760 Set_Node_Length(ret, 1); /* MJD */
11763 nextchar(pRExC_state);
11764 if (RExC_flags & RXf_PMf_SINGLELINE)
11765 ret = reg_node(pRExC_state, SANY);
11767 ret = reg_node(pRExC_state, REG_ANY);
11768 *flagp |= HASWIDTH|SIMPLE;
11770 Set_Node_Length(ret, 1); /* MJD */
11774 char * const oregcomp_parse = ++RExC_parse;
11775 ret = regclass(pRExC_state, flagp,depth+1,
11776 FALSE, /* means parse the whole char class */
11777 TRUE, /* allow multi-char folds */
11778 FALSE, /* don't silence non-portable warnings. */
11779 (bool) RExC_strict,
11780 TRUE, /* Allow an optimized regnode result */
11783 if (*flagp & (RESTART_PASS1|NEED_UTF8))
11785 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11788 if (*RExC_parse != ']') {
11789 RExC_parse = oregcomp_parse;
11790 vFAIL("Unmatched [");
11792 nextchar(pRExC_state);
11793 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11797 nextchar(pRExC_state);
11798 ret = reg(pRExC_state, 2, &flags,depth+1);
11800 if (flags & TRYAGAIN) {
11801 if (RExC_parse == RExC_end) {
11802 /* Make parent create an empty node if needed. */
11803 *flagp |= TRYAGAIN;
11808 if (flags & (RESTART_PASS1|NEED_UTF8)) {
11809 *flagp = flags & (RESTART_PASS1|NEED_UTF8);
11812 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11815 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11819 if (flags & TRYAGAIN) {
11820 *flagp |= TRYAGAIN;
11823 vFAIL("Internal urp");
11824 /* Supposed to be caught earlier. */
11830 vFAIL("Quantifier follows nothing");
11835 This switch handles escape sequences that resolve to some kind
11836 of special regop and not to literal text. Escape sequnces that
11837 resolve to literal text are handled below in the switch marked
11840 Every entry in this switch *must* have a corresponding entry
11841 in the literal escape switch. However, the opposite is not
11842 required, as the default for this switch is to jump to the
11843 literal text handling code.
11845 switch ((U8)*++RExC_parse) {
11846 /* Special Escapes */
11848 RExC_seen_zerolen++;
11849 ret = reg_node(pRExC_state, SBOL);
11850 /* SBOL is shared with /^/ so we set the flags so we can tell
11851 * /\A/ from /^/ in split. We check ret because first pass we
11852 * have no regop struct to set the flags on. */
11856 goto finish_meta_pat;
11858 ret = reg_node(pRExC_state, GPOS);
11859 RExC_seen |= REG_GPOS_SEEN;
11861 goto finish_meta_pat;
11863 RExC_seen_zerolen++;
11864 ret = reg_node(pRExC_state, KEEPS);
11866 /* XXX:dmq : disabling in-place substitution seems to
11867 * be necessary here to avoid cases of memory corruption, as
11868 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11870 RExC_seen |= REG_LOOKBEHIND_SEEN;
11871 goto finish_meta_pat;
11873 ret = reg_node(pRExC_state, SEOL);
11875 RExC_seen_zerolen++; /* Do not optimize RE away */
11876 goto finish_meta_pat;
11878 ret = reg_node(pRExC_state, EOS);
11880 RExC_seen_zerolen++; /* Do not optimize RE away */
11881 goto finish_meta_pat;
11883 vFAIL("\\C no longer supported");
11885 ret = reg_node(pRExC_state, CLUMP);
11886 *flagp |= HASWIDTH;
11887 goto finish_meta_pat;
11893 arg = ANYOF_WORDCHAR;
11901 regex_charset charset = get_regex_charset(RExC_flags);
11903 RExC_seen_zerolen++;
11904 RExC_seen |= REG_LOOKBEHIND_SEEN;
11905 op = BOUND + charset;
11907 if (op == BOUNDL) {
11908 RExC_contains_locale = 1;
11911 ret = reg_node(pRExC_state, op);
11913 if (*(RExC_parse + 1) != '{') {
11914 FLAGS(ret) = TRADITIONAL_BOUND;
11915 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
11921 char name = *RExC_parse;
11924 endbrace = strchr(RExC_parse, '}');
11927 vFAIL2("Missing right brace on \\%c{}", name);
11929 /* XXX Need to decide whether to take spaces or not. Should be
11930 * consistent with \p{}, but that currently is SPACE, which
11931 * means vertical too, which seems wrong
11932 * while (isBLANK(*RExC_parse)) {
11935 if (endbrace == RExC_parse) {
11936 RExC_parse++; /* After the '}' */
11937 vFAIL2("Empty \\%c{}", name);
11939 length = endbrace - RExC_parse;
11940 /*while (isBLANK(*(RExC_parse + length - 1))) {
11943 switch (*RExC_parse) {
11946 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
11948 goto bad_bound_type;
11950 FLAGS(ret) = GCB_BOUND;
11953 if (length != 2 || *(RExC_parse + 1) != 'b') {
11954 goto bad_bound_type;
11956 FLAGS(ret) = SB_BOUND;
11959 if (length != 2 || *(RExC_parse + 1) != 'b') {
11960 goto bad_bound_type;
11962 FLAGS(ret) = WB_BOUND;
11966 RExC_parse = endbrace;
11968 "'%"UTF8f"' is an unknown bound type",
11969 UTF8fARG(UTF, length, endbrace - length));
11970 NOT_REACHED; /*NOTREACHED*/
11972 RExC_parse = endbrace;
11973 REQUIRE_UNI_RULES(flagp, NULL);
11975 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
11979 /* Don't have to worry about UTF-8, in this message because
11980 * to get here the contents of the \b must be ASCII */
11981 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
11982 "Using /u for '%.*s' instead of /%s",
11984 endbrace - length + 1,
11985 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
11986 ? ASCII_RESTRICT_PAT_MODS
11987 : ASCII_MORE_RESTRICT_PAT_MODS);
11991 if (PASS2 && invert) {
11992 OP(ret) += NBOUND - BOUND;
11994 goto finish_meta_pat;
12002 if (! DEPENDS_SEMANTICS) {
12006 /* \d doesn't have any matches in the upper Latin1 range, hence /d
12007 * is equivalent to /u. Changing to /u saves some branches at
12010 goto join_posix_op_known;
12013 ret = reg_node(pRExC_state, LNBREAK);
12014 *flagp |= HASWIDTH|SIMPLE;
12015 goto finish_meta_pat;
12023 goto join_posix_op_known;
12029 arg = ANYOF_VERTWS;
12031 goto join_posix_op_known;
12041 op = POSIXD + get_regex_charset(RExC_flags);
12042 if (op > POSIXA) { /* /aa is same as /a */
12045 else if (op == POSIXL) {
12046 RExC_contains_locale = 1;
12049 join_posix_op_known:
12052 op += NPOSIXD - POSIXD;
12055 ret = reg_node(pRExC_state, op);
12057 FLAGS(ret) = namedclass_to_classnum(arg);
12060 *flagp |= HASWIDTH|SIMPLE;
12064 nextchar(pRExC_state);
12065 Set_Node_Length(ret, 2); /* MJD */
12071 char* parse_start = RExC_parse - 2;
12076 ret = regclass(pRExC_state, flagp,depth+1,
12077 TRUE, /* means just parse this element */
12078 FALSE, /* don't allow multi-char folds */
12079 FALSE, /* don't silence non-portable warnings.
12080 It would be a bug if these returned
12082 (bool) RExC_strict,
12083 TRUE, /* Allow an optimized regnode result */
12085 if (*flagp & RESTART_PASS1)
12087 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
12088 * multi-char folds are allowed. */
12090 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12095 Set_Node_Offset(ret, parse_start + 2);
12096 Set_Node_Cur_Length(ret, parse_start);
12097 nextchar(pRExC_state);
12101 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12102 * \N{...} evaluates to a sequence of more than one code points).
12103 * The function call below returns a regnode, which is our result.
12104 * The parameters cause it to fail if the \N{} evaluates to a
12105 * single code point; we handle those like any other literal. The
12106 * reason that the multicharacter case is handled here and not as
12107 * part of the EXACtish code is because of quantifiers. In
12108 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12109 * this way makes that Just Happen. dmq.
12110 * join_exact() will join this up with adjacent EXACTish nodes
12111 * later on, if appropriate. */
12113 if (grok_bslash_N(pRExC_state,
12114 &ret, /* Want a regnode returned */
12115 NULL, /* Fail if evaluates to a single code
12117 NULL, /* Don't need a count of how many code
12125 if (*flagp & RESTART_PASS1)
12129 /* Here, evaluates to a single code point. Go get that */
12132 case 'k': /* Handle \k<NAME> and \k'NAME' */
12135 char ch= RExC_parse[1];
12136 if (ch != '<' && ch != '\'' && ch != '{') {
12138 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12139 vFAIL2("Sequence %.2s... not terminated",parse_start);
12141 /* this pretty much dupes the code for (?P=...) in reg(), if
12142 you change this make sure you change that */
12143 char* name_start = (RExC_parse += 2);
12145 SV *sv_dat = reg_scan_name(pRExC_state,
12146 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
12147 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
12148 if (RExC_parse == name_start || *RExC_parse != ch)
12149 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12150 vFAIL2("Sequence %.3s... not terminated",parse_start);
12153 num = add_data( pRExC_state, STR_WITH_LEN("S"));
12154 RExC_rxi->data->data[num]=(void*)sv_dat;
12155 SvREFCNT_inc_simple_void(sv_dat);
12159 ret = reganode(pRExC_state,
12162 : (ASCII_FOLD_RESTRICTED)
12164 : (AT_LEAST_UNI_SEMANTICS)
12170 *flagp |= HASWIDTH;
12172 /* override incorrect value set in reganode MJD */
12173 Set_Node_Offset(ret, parse_start+1);
12174 Set_Node_Cur_Length(ret, parse_start);
12175 nextchar(pRExC_state);
12181 case '1': case '2': case '3': case '4':
12182 case '5': case '6': case '7': case '8': case '9':
12187 if (*RExC_parse == 'g') {
12191 if (*RExC_parse == '{') {
12195 if (*RExC_parse == '-') {
12199 if (hasbrace && !isDIGIT(*RExC_parse)) {
12200 if (isrel) RExC_parse--;
12202 goto parse_named_seq;
12205 num = S_backref_value(RExC_parse);
12207 vFAIL("Reference to invalid group 0");
12208 else if (num == I32_MAX) {
12209 if (isDIGIT(*RExC_parse))
12210 vFAIL("Reference to nonexistent group");
12212 vFAIL("Unterminated \\g... pattern");
12216 num = RExC_npar - num;
12218 vFAIL("Reference to nonexistent or unclosed group");
12222 num = S_backref_value(RExC_parse);
12223 /* bare \NNN might be backref or octal - if it is larger
12224 * than or equal RExC_npar then it is assumed to be an
12225 * octal escape. Note RExC_npar is +1 from the actual
12226 * number of parens. */
12227 /* Note we do NOT check if num == I32_MAX here, as that is
12228 * handled by the RExC_npar check */
12231 /* any numeric escape < 10 is always a backref */
12233 /* any numeric escape < RExC_npar is a backref */
12234 && num >= RExC_npar
12235 /* cannot be an octal escape if it starts with 8 */
12236 && *RExC_parse != '8'
12237 /* cannot be an octal escape it it starts with 9 */
12238 && *RExC_parse != '9'
12241 /* Probably not a backref, instead likely to be an
12242 * octal character escape, e.g. \35 or \777.
12243 * The above logic should make it obvious why using
12244 * octal escapes in patterns is problematic. - Yves */
12249 /* At this point RExC_parse points at a numeric escape like
12250 * \12 or \88 or something similar, which we should NOT treat
12251 * as an octal escape. It may or may not be a valid backref
12252 * escape. For instance \88888888 is unlikely to be a valid
12255 #ifdef RE_TRACK_PATTERN_OFFSETS
12256 char * const parse_start = RExC_parse - 1; /* MJD */
12258 while (isDIGIT(*RExC_parse))
12261 if (*RExC_parse != '}')
12262 vFAIL("Unterminated \\g{...} pattern");
12266 if (num > (I32)RExC_rx->nparens)
12267 vFAIL("Reference to nonexistent group");
12270 ret = reganode(pRExC_state,
12273 : (ASCII_FOLD_RESTRICTED)
12275 : (AT_LEAST_UNI_SEMANTICS)
12281 *flagp |= HASWIDTH;
12283 /* override incorrect value set in reganode MJD */
12284 Set_Node_Offset(ret, parse_start+1);
12285 Set_Node_Cur_Length(ret, parse_start);
12286 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
12287 FALSE /* Don't force to /x */ );
12292 if (RExC_parse >= RExC_end)
12293 FAIL("Trailing \\");
12296 /* Do not generate "unrecognized" warnings here, we fall
12297 back into the quick-grab loop below */
12300 } /* end of switch on a \foo sequence */
12304 if (RExC_flags & RXf_PMf_EXTENDED) {
12305 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
12306 if (RExC_parse < RExC_end)
12313 parse_start = RExC_parse - 1;
12319 /* Here, we have determined that the next thing is probably a
12320 * literal character. (It still may be an escape sequence that
12321 * evaluates to a single character) */
12327 #define MAX_NODE_STRING_SIZE 127
12328 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
12330 U8 upper_parse = MAX_NODE_STRING_SIZE;
12331 U8 node_type = compute_EXACTish(pRExC_state);
12332 bool next_is_quantifier;
12333 char * oldp = NULL;
12335 /* We can convert EXACTF nodes to EXACTFU if they contain only
12336 * characters that match identically regardless of the target
12337 * string's UTF8ness. The reason to do this is that EXACTF is not
12338 * trie-able, EXACTFU is.
12340 * Similarly, we can convert EXACTFL nodes to EXACTFLU8 if they
12341 * contain only above-Latin1 characters (hence must be in UTF8),
12342 * which don't participate in folds with Latin1-range characters,
12343 * as the latter's folds aren't known until runtime. (We don't
12344 * need to figure this out until pass 2) */
12345 bool maybe_exactfu = PASS2
12346 && (node_type == EXACTF || node_type == EXACTFL);
12348 /* If a folding node contains only code points that don't
12349 * participate in folds, it can be changed into an EXACT node,
12350 * which allows the optimizer more things to look for */
12353 ret = reg_node(pRExC_state, node_type);
12355 /* In pass1, folded, we use a temporary buffer instead of the
12356 * actual node, as the node doesn't exist yet */
12357 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
12363 /* We look for the EXACTFish to EXACT node optimizaton only if
12364 * folding. (And we don't need to figure this out until pass 2) */
12365 maybe_exact = FOLD && PASS2;
12367 /* XXX The node can hold up to 255 bytes, yet this only goes to
12368 * 127. I (khw) do not know why. Keeping it somewhat less than
12369 * 255 allows us to not have to worry about overflow due to
12370 * converting to utf8 and fold expansion, but that value is
12371 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
12372 * split up by this limit into a single one using the real max of
12373 * 255. Even at 127, this breaks under rare circumstances. If
12374 * folding, we do not want to split a node at a character that is a
12375 * non-final in a multi-char fold, as an input string could just
12376 * happen to want to match across the node boundary. The join
12377 * would solve that problem if the join actually happens. But a
12378 * series of more than two nodes in a row each of 127 would cause
12379 * the first join to succeed to get to 254, but then there wouldn't
12380 * be room for the next one, which could at be one of those split
12381 * multi-char folds. I don't know of any fool-proof solution. One
12382 * could back off to end with only a code point that isn't such a
12383 * non-final, but it is possible for there not to be any in the
12385 for (p = RExC_parse - 1;
12386 len < upper_parse && p < RExC_end;
12391 if (RExC_flags & RXf_PMf_EXTENDED)
12392 p = regpatws(pRExC_state, p,
12393 TRUE); /* means recognize comments */
12404 /* Literal Escapes Switch
12406 This switch is meant to handle escape sequences that
12407 resolve to a literal character.
12409 Every escape sequence that represents something
12410 else, like an assertion or a char class, is handled
12411 in the switch marked 'Special Escapes' above in this
12412 routine, but also has an entry here as anything that
12413 isn't explicitly mentioned here will be treated as
12414 an unescaped equivalent literal.
12417 switch ((U8)*++p) {
12418 /* These are all the special escapes. */
12419 case 'A': /* Start assertion */
12420 case 'b': case 'B': /* Word-boundary assertion*/
12421 case 'C': /* Single char !DANGEROUS! */
12422 case 'd': case 'D': /* digit class */
12423 case 'g': case 'G': /* generic-backref, pos assertion */
12424 case 'h': case 'H': /* HORIZWS */
12425 case 'k': case 'K': /* named backref, keep marker */
12426 case 'p': case 'P': /* Unicode property */
12427 case 'R': /* LNBREAK */
12428 case 's': case 'S': /* space class */
12429 case 'v': case 'V': /* VERTWS */
12430 case 'w': case 'W': /* word class */
12431 case 'X': /* eXtended Unicode "combining
12432 character sequence" */
12433 case 'z': case 'Z': /* End of line/string assertion */
12437 /* Anything after here is an escape that resolves to a
12438 literal. (Except digits, which may or may not)
12444 case 'N': /* Handle a single-code point named character. */
12445 RExC_parse = p + 1;
12446 if (! grok_bslash_N(pRExC_state,
12447 NULL, /* Fail if evaluates to
12448 anything other than a
12449 single code point */
12450 &ender, /* The returned single code
12452 NULL, /* Don't need a count of
12453 how many code points */
12457 if (*flagp & NEED_UTF8)
12458 FAIL("panic: grok_bslash_N set NEED_UTF8");
12459 if (*flagp & RESTART_PASS1)
12462 /* Here, it wasn't a single code point. Go close
12463 * up this EXACTish node. The switch() prior to
12464 * this switch handles the other cases */
12465 RExC_parse = p = oldp;
12469 if (ender > 0xff) {
12470 REQUIRE_UTF8(flagp);
12486 ender = ESC_NATIVE;
12496 const char* error_msg;
12498 bool valid = grok_bslash_o(&p,
12501 PASS2, /* out warnings */
12502 (bool) RExC_strict,
12503 TRUE, /* Output warnings
12508 RExC_parse = p; /* going to die anyway; point
12509 to exact spot of failure */
12513 if (IN_ENCODING && ender < 0x100) {
12514 goto recode_encoding;
12516 if (ender > 0xff) {
12517 REQUIRE_UTF8(flagp);
12523 UV result = UV_MAX; /* initialize to erroneous
12525 const char* error_msg;
12527 bool valid = grok_bslash_x(&p,
12530 PASS2, /* out warnings */
12531 (bool) RExC_strict,
12532 TRUE, /* Silence warnings
12537 RExC_parse = p; /* going to die anyway; point
12538 to exact spot of failure */
12543 if (ender < 0x100) {
12545 if (RExC_recode_x_to_native) {
12546 ender = LATIN1_TO_NATIVE(ender);
12551 goto recode_encoding;
12555 REQUIRE_UTF8(flagp);
12561 ender = grok_bslash_c(*p++, PASS2);
12563 case '8': case '9': /* must be a backreference */
12565 /* we have an escape like \8 which cannot be an octal escape
12566 * so we exit the loop, and let the outer loop handle this
12567 * escape which may or may not be a legitimate backref. */
12569 case '1': case '2': case '3':case '4':
12570 case '5': case '6': case '7':
12571 /* When we parse backslash escapes there is ambiguity
12572 * between backreferences and octal escapes. Any escape
12573 * from \1 - \9 is a backreference, any multi-digit
12574 * escape which does not start with 0 and which when
12575 * evaluated as decimal could refer to an already
12576 * parsed capture buffer is a back reference. Anything
12579 * Note this implies that \118 could be interpreted as
12580 * 118 OR as "\11" . "8" depending on whether there
12581 * were 118 capture buffers defined already in the
12584 /* NOTE, RExC_npar is 1 more than the actual number of
12585 * parens we have seen so far, hence the < RExC_npar below. */
12587 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
12588 { /* Not to be treated as an octal constant, go
12596 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12598 ender = grok_oct(p, &numlen, &flags, NULL);
12599 if (ender > 0xff) {
12600 REQUIRE_UTF8(flagp);
12603 if (PASS2 /* like \08, \178 */
12606 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
12608 reg_warn_non_literal_string(
12610 form_short_octal_warning(p, numlen));
12613 if (IN_ENCODING && ender < 0x100)
12614 goto recode_encoding;
12617 if (! RExC_override_recoding) {
12618 SV* enc = _get_encoding();
12619 ender = reg_recode((const char)(U8)ender, &enc);
12621 ckWARNreg(p, "Invalid escape in the specified encoding");
12622 REQUIRE_UTF8(flagp);
12627 FAIL("Trailing \\");
12630 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
12631 /* Include any left brace following the alpha to emphasize
12632 * that it could be part of an escape at some point
12634 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
12635 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
12637 goto normal_default;
12638 } /* End of switch on '\' */
12641 /* Currently we don't warn when the lbrace is at the start
12642 * of a construct. This catches it in the middle of a
12643 * literal string, or when it's the first thing after
12644 * something like "\b" */
12646 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
12648 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
12651 default: /* A literal character */
12653 if (! UTF8_IS_INVARIANT(*p) && UTF) {
12655 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
12656 &numlen, UTF8_ALLOW_DEFAULT);
12662 } /* End of switch on the literal */
12664 /* Here, have looked at the literal character and <ender>
12665 * contains its ordinal, <p> points to the character after it.
12666 * We need to check if the next non-ignored thing is a
12667 * quantifier. Move <p> to after anything that should be
12668 * ignored, which, as a side effect, positions <p> for the next
12669 * loop iteration */
12670 if ( RExC_flags & RXf_PMf_EXTENDED)
12671 p = regpatws(pRExC_state, p,
12672 TRUE); /* means recognize comments */
12674 /* If the next thing is a quantifier, it applies to this
12675 * character only, which means that this character has to be in
12676 * its own node and can't just be appended to the string in an
12677 * existing node, so if there are already other characters in
12678 * the node, close the node with just them, and set up to do
12679 * this character again next time through, when it will be the
12680 * only thing in its new node */
12681 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
12687 /* Ready to add 'ender' to the node */
12689 if (! FOLD) { /* The simple case, just append the literal */
12691 /* In the sizing pass, we need only the size of the
12692 * character we are appending, hence we can delay getting
12693 * its representation until PASS2. */
12696 const STRLEN unilen = UVCHR_SKIP(ender);
12699 /* We have to subtract 1 just below (and again in
12700 * the corresponding PASS2 code) because the loop
12701 * increments <len> each time, as all but this path
12702 * (and one other) through it add a single byte to
12703 * the EXACTish node. But these paths would change
12704 * len to be the correct final value, so cancel out
12705 * the increment that follows */
12711 } else { /* PASS2 */
12714 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
12715 len += (char *) new_s - s - 1;
12716 s = (char *) new_s;
12719 *(s++) = (char) ender;
12723 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
12725 /* Here are folding under /l, and the code point is
12726 * problematic. First, we know we can't simplify things */
12727 maybe_exact = FALSE;
12728 maybe_exactfu = FALSE;
12730 /* A problematic code point in this context means that its
12731 * fold isn't known until runtime, so we can't fold it now.
12732 * (The non-problematic code points are the above-Latin1
12733 * ones that fold to also all above-Latin1. Their folds
12734 * don't vary no matter what the locale is.) But here we
12735 * have characters whose fold depends on the locale.
12736 * Unlike the non-folding case above, we have to keep track
12737 * of these in the sizing pass, so that we can make sure we
12738 * don't split too-long nodes in the middle of a potential
12739 * multi-char fold. And unlike the regular fold case
12740 * handled in the else clauses below, we don't actually
12741 * fold and don't have special cases to consider. What we
12742 * do for both passes is the PASS2 code for non-folding */
12743 goto not_fold_common;
12745 else /* A regular FOLD code point */
12747 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12748 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12749 || UNICODE_DOT_DOT_VERSION > 0)
12750 /* See comments for join_exact() as to why we fold this
12751 * non-UTF at compile time */
12752 || (node_type == EXACTFU
12753 && ender == LATIN_SMALL_LETTER_SHARP_S)
12756 /* Here, are folding and are not UTF-8 encoded; therefore
12757 * the character must be in the range 0-255, and is not /l
12758 * (Not /l because we already handled these under /l in
12759 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
12760 if (IS_IN_SOME_FOLD_L1(ender)) {
12761 maybe_exact = FALSE;
12763 /* See if the character's fold differs between /d and
12764 * /u. This includes the multi-char fold SHARP S to
12766 if (UNLIKELY(ender == LATIN_SMALL_LETTER_SHARP_S)) {
12767 RExC_seen_unfolded_sharp_s = 1;
12768 maybe_exactfu = FALSE;
12770 else if (maybe_exactfu
12771 && (PL_fold[ender] != PL_fold_latin1[ender]
12772 #if UNICODE_MAJOR_VERSION > 3 /* no multifolds in early Unicode */ \
12773 || (UNICODE_MAJOR_VERSION == 3 && ( UNICODE_DOT_VERSION > 0) \
12774 || UNICODE_DOT_DOT_VERSION > 0)
12776 && isALPHA_FOLD_EQ(ender, 's')
12777 && isALPHA_FOLD_EQ(*(s-1), 's'))
12780 maybe_exactfu = FALSE;
12784 /* Even when folding, we store just the input character, as
12785 * we have an array that finds its fold quickly */
12786 *(s++) = (char) ender;
12788 else { /* FOLD, and UTF (or sharp s) */
12789 /* Unlike the non-fold case, we do actually have to
12790 * calculate the results here in pass 1. This is for two
12791 * reasons, the folded length may be longer than the
12792 * unfolded, and we have to calculate how many EXACTish
12793 * nodes it will take; and we may run out of room in a node
12794 * in the middle of a potential multi-char fold, and have
12795 * to back off accordingly. */
12798 if (isASCII_uni(ender)) {
12799 folded = toFOLD(ender);
12800 *(s)++ = (U8) folded;
12805 folded = _to_uni_fold_flags(
12809 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12810 ? FOLD_FLAGS_NOMIX_ASCII
12814 /* The loop increments <len> each time, as all but this
12815 * path (and one other) through it add a single byte to
12816 * the EXACTish node. But this one has changed len to
12817 * be the correct final value, so subtract one to
12818 * cancel out the increment that follows */
12819 len += foldlen - 1;
12821 /* If this node only contains non-folding code points so
12822 * far, see if this new one is also non-folding */
12824 if (folded != ender) {
12825 maybe_exact = FALSE;
12828 /* Here the fold is the original; we have to check
12829 * further to see if anything folds to it */
12830 if (_invlist_contains_cp(PL_utf8_foldable,
12833 maybe_exact = FALSE;
12840 if (next_is_quantifier) {
12842 /* Here, the next input is a quantifier, and to get here,
12843 * the current character is the only one in the node.
12844 * Also, here <len> doesn't include the final byte for this
12850 } /* End of loop through literal characters */
12852 /* Here we have either exhausted the input or ran out of room in
12853 * the node. (If we encountered a character that can't be in the
12854 * node, transfer is made directly to <loopdone>, and so we
12855 * wouldn't have fallen off the end of the loop.) In the latter
12856 * case, we artificially have to split the node into two, because
12857 * we just don't have enough space to hold everything. This
12858 * creates a problem if the final character participates in a
12859 * multi-character fold in the non-final position, as a match that
12860 * should have occurred won't, due to the way nodes are matched,
12861 * and our artificial boundary. So back off until we find a non-
12862 * problematic character -- one that isn't at the beginning or
12863 * middle of such a fold. (Either it doesn't participate in any
12864 * folds, or appears only in the final position of all the folds it
12865 * does participate in.) A better solution with far fewer false
12866 * positives, and that would fill the nodes more completely, would
12867 * be to actually have available all the multi-character folds to
12868 * test against, and to back-off only far enough to be sure that
12869 * this node isn't ending with a partial one. <upper_parse> is set
12870 * further below (if we need to reparse the node) to include just
12871 * up through that final non-problematic character that this code
12872 * identifies, so when it is set to less than the full node, we can
12873 * skip the rest of this */
12874 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12876 const STRLEN full_len = len;
12878 assert(len >= MAX_NODE_STRING_SIZE);
12880 /* Here, <s> points to the final byte of the final character.
12881 * Look backwards through the string until find a non-
12882 * problematic character */
12886 /* This has no multi-char folds to non-UTF characters */
12887 if (ASCII_FOLD_RESTRICTED) {
12891 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12895 if (! PL_NonL1NonFinalFold) {
12896 PL_NonL1NonFinalFold = _new_invlist_C_array(
12897 NonL1_Perl_Non_Final_Folds_invlist);
12900 /* Point to the first byte of the final character */
12901 s = (char *) utf8_hop((U8 *) s, -1);
12903 while (s >= s0) { /* Search backwards until find
12904 non-problematic char */
12905 if (UTF8_IS_INVARIANT(*s)) {
12907 /* There are no ascii characters that participate
12908 * in multi-char folds under /aa. In EBCDIC, the
12909 * non-ascii invariants are all control characters,
12910 * so don't ever participate in any folds. */
12911 if (ASCII_FOLD_RESTRICTED
12912 || ! IS_NON_FINAL_FOLD(*s))
12917 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12918 if (! IS_NON_FINAL_FOLD(EIGHT_BIT_UTF8_TO_NATIVE(
12924 else if (! _invlist_contains_cp(
12925 PL_NonL1NonFinalFold,
12926 valid_utf8_to_uvchr((U8 *) s, NULL)))
12931 /* Here, the current character is problematic in that
12932 * it does occur in the non-final position of some
12933 * fold, so try the character before it, but have to
12934 * special case the very first byte in the string, so
12935 * we don't read outside the string */
12936 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12937 } /* End of loop backwards through the string */
12939 /* If there were only problematic characters in the string,
12940 * <s> will point to before s0, in which case the length
12941 * should be 0, otherwise include the length of the
12942 * non-problematic character just found */
12943 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12946 /* Here, have found the final character, if any, that is
12947 * non-problematic as far as ending the node without splitting
12948 * it across a potential multi-char fold. <len> contains the
12949 * number of bytes in the node up-to and including that
12950 * character, or is 0 if there is no such character, meaning
12951 * the whole node contains only problematic characters. In
12952 * this case, give up and just take the node as-is. We can't
12957 /* If the node ends in an 's' we make sure it stays EXACTF,
12958 * as if it turns into an EXACTFU, it could later get
12959 * joined with another 's' that would then wrongly match
12961 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
12963 maybe_exactfu = FALSE;
12967 /* Here, the node does contain some characters that aren't
12968 * problematic. If one such is the final character in the
12969 * node, we are done */
12970 if (len == full_len) {
12973 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12975 /* If the final character is problematic, but the
12976 * penultimate is not, back-off that last character to
12977 * later start a new node with it */
12982 /* Here, the final non-problematic character is earlier
12983 * in the input than the penultimate character. What we do
12984 * is reparse from the beginning, going up only as far as
12985 * this final ok one, thus guaranteeing that the node ends
12986 * in an acceptable character. The reason we reparse is
12987 * that we know how far in the character is, but we don't
12988 * know how to correlate its position with the input parse.
12989 * An alternate implementation would be to build that
12990 * correlation as we go along during the original parse,
12991 * but that would entail extra work for every node, whereas
12992 * this code gets executed only when the string is too
12993 * large for the node, and the final two characters are
12994 * problematic, an infrequent occurrence. Yet another
12995 * possible strategy would be to save the tail of the
12996 * string, and the next time regatom is called, initialize
12997 * with that. The problem with this is that unless you
12998 * back off one more character, you won't be guaranteed
12999 * regatom will get called again, unless regbranch,
13000 * regpiece ... are also changed. If you do back off that
13001 * extra character, so that there is input guaranteed to
13002 * force calling regatom, you can't handle the case where
13003 * just the first character in the node is acceptable. I
13004 * (khw) decided to try this method which doesn't have that
13005 * pitfall; if performance issues are found, we can do a
13006 * combination of the current approach plus that one */
13012 } /* End of verifying node ends with an appropriate char */
13014 loopdone: /* Jumped to when encounters something that shouldn't be
13017 /* I (khw) don't know if you can get here with zero length, but the
13018 * old code handled this situation by creating a zero-length EXACT
13019 * node. Might as well be NOTHING instead */
13025 /* If 'maybe_exact' is still set here, means there are no
13026 * code points in the node that participate in folds;
13027 * similarly for 'maybe_exactfu' and code points that match
13028 * differently depending on UTF8ness of the target string
13029 * (for /u), or depending on locale for /l */
13035 else if (maybe_exactfu) {
13041 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
13042 FALSE /* Don't look to see if could
13043 be turned into an EXACT
13044 node, as we have already
13049 RExC_parse = p - 1;
13050 Set_Node_Cur_Length(ret, parse_start);
13052 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
13053 FALSE /* Don't force to /x */ );
13055 /* len is STRLEN which is unsigned, need to copy to signed */
13058 vFAIL("Internal disaster");
13061 } /* End of label 'defchar:' */
13063 } /* End of giant switch on input character */
13069 S_regpatws(RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
13071 /* Returns the next non-pattern-white space, non-comment character (the
13072 * latter only if 'recognize_comment is true) in the string p, which is
13073 * ended by RExC_end. See also reg_skipcomment */
13074 const char *e = RExC_end;
13076 PERL_ARGS_ASSERT_REGPATWS;
13080 if ((len = is_PATWS_safe(p, e, UTF))) {
13083 else if (recognize_comment && *p == '#') {
13084 p = reg_skipcomment(pRExC_state, p);
13093 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13095 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13096 * sets up the bitmap and any flags, removing those code points from the
13097 * inversion list, setting it to NULL should it become completely empty */
13099 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13100 assert(PL_regkind[OP(node)] == ANYOF);
13102 ANYOF_BITMAP_ZERO(node);
13103 if (*invlist_ptr) {
13105 /* This gets set if we actually need to modify things */
13106 bool change_invlist = FALSE;
13110 /* Start looking through *invlist_ptr */
13111 invlist_iterinit(*invlist_ptr);
13112 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13116 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13117 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13119 else if (end >= NUM_ANYOF_CODE_POINTS) {
13120 ANYOF_FLAGS(node) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
13123 /* Quit if are above what we should change */
13124 if (start >= NUM_ANYOF_CODE_POINTS) {
13128 change_invlist = TRUE;
13130 /* Set all the bits in the range, up to the max that we are doing */
13131 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13133 : NUM_ANYOF_CODE_POINTS - 1;
13134 for (i = start; i <= (int) high; i++) {
13135 if (! ANYOF_BITMAP_TEST(node, i)) {
13136 ANYOF_BITMAP_SET(node, i);
13140 invlist_iterfinish(*invlist_ptr);
13142 /* Done with loop; remove any code points that are in the bitmap from
13143 * *invlist_ptr; similarly for code points above the bitmap if we have
13144 * a flag to match all of them anyways */
13145 if (change_invlist) {
13146 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13148 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13149 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13152 /* If have completely emptied it, remove it completely */
13153 if (_invlist_len(*invlist_ptr) == 0) {
13154 SvREFCNT_dec_NN(*invlist_ptr);
13155 *invlist_ptr = NULL;
13160 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13161 Character classes ([:foo:]) can also be negated ([:^foo:]).
13162 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13163 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13164 but trigger failures because they are currently unimplemented. */
13166 #define POSIXCC_DONE(c) ((c) == ':')
13167 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13168 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13170 PERL_STATIC_INLINE I32
13171 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
13173 I32 namedclass = OOB_NAMEDCLASS;
13175 PERL_ARGS_ASSERT_REGPPOSIXCC;
13177 if (value == '[' && RExC_parse + 1 < RExC_end &&
13178 /* I smell either [: or [= or [. -- POSIX has been here, right? */
13179 POSIXCC(UCHARAT(RExC_parse)))
13181 const char c = UCHARAT(RExC_parse);
13182 char* const s = RExC_parse++;
13184 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
13186 if (RExC_parse == RExC_end) {
13189 /* Try to give a better location for the error (than the end of
13190 * the string) by looking for the matching ']' */
13192 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
13195 vFAIL2("Unmatched '%c' in POSIX class", c);
13197 /* Grandfather lone [:, [=, [. */
13201 const char* const t = RExC_parse++; /* skip over the c */
13204 if (UCHARAT(RExC_parse) == ']') {
13205 const char *posixcc = s + 1;
13206 RExC_parse++; /* skip over the ending ] */
13209 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
13210 const I32 skip = t - posixcc;
13212 /* Initially switch on the length of the name. */
13215 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
13216 this is the Perl \w
13218 namedclass = ANYOF_WORDCHAR;
13221 /* Names all of length 5. */
13222 /* alnum alpha ascii blank cntrl digit graph lower
13223 print punct space upper */
13224 /* Offset 4 gives the best switch position. */
13225 switch (posixcc[4]) {
13227 if (memEQ(posixcc, "alph", 4)) /* alpha */
13228 namedclass = ANYOF_ALPHA;
13231 if (memEQ(posixcc, "spac", 4)) /* space */
13232 namedclass = ANYOF_SPACE;
13235 if (memEQ(posixcc, "grap", 4)) /* graph */
13236 namedclass = ANYOF_GRAPH;
13239 if (memEQ(posixcc, "asci", 4)) /* ascii */
13240 namedclass = ANYOF_ASCII;
13243 if (memEQ(posixcc, "blan", 4)) /* blank */
13244 namedclass = ANYOF_BLANK;
13247 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
13248 namedclass = ANYOF_CNTRL;
13251 if (memEQ(posixcc, "alnu", 4)) /* alnum */
13252 namedclass = ANYOF_ALPHANUMERIC;
13255 if (memEQ(posixcc, "lowe", 4)) /* lower */
13256 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
13257 else if (memEQ(posixcc, "uppe", 4)) /* upper */
13258 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
13261 if (memEQ(posixcc, "digi", 4)) /* digit */
13262 namedclass = ANYOF_DIGIT;
13263 else if (memEQ(posixcc, "prin", 4)) /* print */
13264 namedclass = ANYOF_PRINT;
13265 else if (memEQ(posixcc, "punc", 4)) /* punct */
13266 namedclass = ANYOF_PUNCT;
13271 if (memEQ(posixcc, "xdigit", 6))
13272 namedclass = ANYOF_XDIGIT;
13276 if (namedclass == OOB_NAMEDCLASS)
13278 "POSIX class [:%"UTF8f":] unknown",
13279 UTF8fARG(UTF, t - s - 1, s + 1));
13281 /* The #defines are structured so each complement is +1 to
13282 * the normal one */
13286 assert (posixcc[skip] == ':');
13287 assert (posixcc[skip+1] == ']');
13288 } else if (!SIZE_ONLY) {
13289 /* [[=foo=]] and [[.foo.]] are still future. */
13291 /* adjust RExC_parse so the warning shows after
13292 the class closes */
13293 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
13295 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
13298 /* Maternal grandfather:
13299 * "[:" ending in ":" but not in ":]" */
13301 vFAIL("Unmatched '[' in POSIX class");
13304 /* Grandfather lone [:, [=, [. */
13314 S_could_it_be_a_POSIX_class(RExC_state_t *pRExC_state)
13316 /* This applies some heuristics at the current parse position (which should
13317 * be at a '[') to see if what follows might be intended to be a [:posix:]
13318 * class. It returns true if it really is a posix class, of course, but it
13319 * also can return true if it thinks that what was intended was a posix
13320 * class that didn't quite make it.
13322 * It will return true for
13324 * [:alphanumerics] (as long as the ] isn't followed immediately by a
13325 * ')' indicating the end of the (?[
13326 * [:any garbage including %^&$ punctuation:]
13328 * This is designed to be called only from S_handle_regex_sets; it could be
13329 * easily adapted to be called from the spot at the beginning of regclass()
13330 * that checks to see in a normal bracketed class if the surrounding []
13331 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
13332 * change long-standing behavior, so I (khw) didn't do that */
13333 char* p = RExC_parse + 1;
13334 char first_char = *p;
13336 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
13338 assert(*(p - 1) == '[');
13340 if (! POSIXCC(first_char)) {
13345 while (p < RExC_end && isWORDCHAR(*p)) p++;
13347 if (p >= RExC_end) {
13351 if (p - RExC_parse > 2 /* Got at least 1 word character */
13352 && (*p == first_char
13353 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
13358 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
13361 && p - RExC_parse > 2 /* [:] evaluates to colon;
13362 [::] is a bad posix class. */
13363 && first_char == *(p - 1));
13366 STATIC unsigned int
13367 S_regex_set_precedence(const U8 my_operator) {
13369 /* Returns the precedence in the (?[...]) construct of the input operator,
13370 * specified by its character representation. The precedence follows
13371 * general Perl rules, but it extends this so that ')' and ']' have (low)
13372 * precedence even though they aren't really operators */
13374 switch (my_operator) {
13390 NOT_REACHED; /* NOTREACHED */
13391 return 0; /* Silence compiler warning */
13395 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
13396 I32 *flagp, U32 depth,
13397 char * const oregcomp_parse)
13399 /* Handle the (?[...]) construct to do set operations */
13401 U8 curchar; /* Current character being parsed */
13402 UV start, end; /* End points of code point ranges */
13403 SV* final = NULL; /* The end result inversion list */
13404 SV* result_string; /* 'final' stringified */
13405 AV* stack; /* stack of operators and operands not yet
13407 AV* fence_stack = NULL; /* A stack containing the positions in
13408 'stack' of where the undealt-with left
13409 parens would be if they were actually
13411 IV fence = 0; /* Position of where most recent undealt-
13412 with left paren in stack is; -1 if none.
13414 STRLEN len; /* Temporary */
13415 regnode* node; /* Temporary, and final regnode returned by
13417 const bool save_fold = FOLD; /* Temporary */
13418 char *save_end, *save_parse; /* Temporaries */
13419 const bool in_locale = LOC; /* we turn off /l during processing */
13421 GET_RE_DEBUG_FLAGS_DECL;
13423 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
13426 set_regex_charset(&RExC_flags, REGEX_UNICODE_CHARSET);
13429 REQUIRE_UNI_RULES(flagp, NULL); /* The use of this operator implies /u.
13430 This is required so that the compile
13431 time values are valid in all runtime
13434 /* This will return only an ANYOF regnode, or (unlikely) something smaller
13435 * (such as EXACT). Thus we can skip most everything if just sizing. We
13436 * call regclass to handle '[]' so as to not have to reinvent its parsing
13437 * rules here (throwing away the size it computes each time). And, we exit
13438 * upon an unescaped ']' that isn't one ending a regclass. To do both
13439 * these things, we need to realize that something preceded by a backslash
13440 * is escaped, so we have to keep track of backslashes */
13442 UV depth = 0; /* how many nested (?[...]) constructs */
13444 while (RExC_parse < RExC_end) {
13445 SV* current = NULL;
13446 RExC_parse = regpatws(pRExC_state, RExC_parse,
13447 TRUE); /* means recognize comments */
13448 switch (*RExC_parse) {
13450 if (RExC_parse[1] == '[') depth++, RExC_parse++;
13455 /* Skip the next byte (which could cause us to end up in
13456 * the middle of a UTF-8 character, but since none of those
13457 * are confusable with anything we currently handle in this
13458 * switch (invariants all), it's safe. We'll just hit the
13459 * default: case next time and keep on incrementing until
13460 * we find one of the invariants we do handle. */
13465 /* If this looks like it is a [:posix:] class, leave the
13466 * parse pointer at the '[' to fool regclass() into
13467 * thinking it is part of a '[[:posix:]]'. That function
13468 * will use strict checking to force a syntax error if it
13469 * doesn't work out to a legitimate class */
13470 bool is_posix_class
13471 = could_it_be_a_POSIX_class(pRExC_state);
13472 if (! is_posix_class) {
13476 /* regclass() can only return RESTART_PASS1 and NEED_UTF8
13477 * if multi-char folds are allowed. */
13478 if (!regclass(pRExC_state, flagp,depth+1,
13479 is_posix_class, /* parse the whole char
13480 class only if not a
13482 FALSE, /* don't allow multi-char folds */
13483 TRUE, /* silence non-portable warnings. */
13485 FALSE, /* Require return to be an ANYOF */
13488 FAIL2("panic: regclass returned NULL to handle_sets, "
13489 "flags=%#"UVxf"", (UV) *flagp);
13491 /* function call leaves parse pointing to the ']', except
13492 * if we faked it */
13493 if (is_posix_class) {
13497 SvREFCNT_dec(current); /* In case it returned something */
13502 if (depth--) break;
13504 if (RExC_parse < RExC_end
13505 && *RExC_parse == ')')
13507 node = reganode(pRExC_state, ANYOF, 0);
13508 RExC_size += ANYOF_SKIP;
13509 nextchar(pRExC_state);
13510 Set_Node_Length(node,
13511 RExC_parse - oregcomp_parse + 1); /* MJD */
13513 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
13521 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
13525 FAIL("Syntax error in (?[...])");
13528 /* Pass 2 only after this. */
13529 Perl_ck_warner_d(aTHX_
13530 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
13531 "The regex_sets feature is experimental" REPORT_LOCATION,
13532 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
13534 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
13535 RExC_precomp + (RExC_parse - RExC_precomp)));
13537 /* Everything in this construct is a metacharacter. Operands begin with
13538 * either a '\' (for an escape sequence), or a '[' for a bracketed
13539 * character class. Any other character should be an operator, or
13540 * parenthesis for grouping. Both types of operands are handled by calling
13541 * regclass() to parse them. It is called with a parameter to indicate to
13542 * return the computed inversion list. The parsing here is implemented via
13543 * a stack. Each entry on the stack is a single character representing one
13544 * of the operators; or else a pointer to an operand inversion list. */
13546 #define IS_OPERATOR(a) SvIOK(a)
13547 #define IS_OPERAND(a) (! IS_OPERATOR(a))
13549 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
13550 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
13551 * with pronouncing it called it Reverse Polish instead, but now that YOU
13552 * know how to pronounce it you can use the correct term, thus giving due
13553 * credit to the person who invented it, and impressing your geek friends.
13554 * Wikipedia says that the pronounciation of "Ł" has been changing so that
13555 * it is now more like an English initial W (as in wonk) than an L.)
13557 * This means that, for example, 'a | b & c' is stored on the stack as
13565 * where the numbers in brackets give the stack [array] element number.
13566 * In this implementation, parentheses are not stored on the stack.
13567 * Instead a '(' creates a "fence" so that the part of the stack below the
13568 * fence is invisible except to the corresponding ')' (this allows us to
13569 * replace testing for parens, by using instead subtraction of the fence
13570 * position). As new operands are processed they are pushed onto the stack
13571 * (except as noted in the next paragraph). New operators of higher
13572 * precedence than the current final one are inserted on the stack before
13573 * the lhs operand (so that when the rhs is pushed next, everything will be
13574 * in the correct positions shown above. When an operator of equal or
13575 * lower precedence is encountered in parsing, all the stacked operations
13576 * of equal or higher precedence are evaluated, leaving the result as the
13577 * top entry on the stack. This makes higher precedence operations
13578 * evaluate before lower precedence ones, and causes operations of equal
13579 * precedence to left associate.
13581 * The only unary operator '!' is immediately pushed onto the stack when
13582 * encountered. When an operand is encountered, if the top of the stack is
13583 * a '!", the complement is immediately performed, and the '!' popped. The
13584 * resulting value is treated as a new operand, and the logic in the
13585 * previous paragraph is executed. Thus in the expression
13587 * the stack looks like
13593 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
13600 * A ')' is treated as an operator with lower precedence than all the
13601 * aforementioned ones, which causes all operations on the stack above the
13602 * corresponding '(' to be evaluated down to a single resultant operand.
13603 * Then the fence for the '(' is removed, and the operand goes through the
13604 * algorithm above, without the fence.
13606 * A separate stack is kept of the fence positions, so that the position of
13607 * the latest so-far unbalanced '(' is at the top of it.
13609 * The ']' ending the construct is treated as the lowest operator of all,
13610 * so that everything gets evaluated down to a single operand, which is the
13613 sv_2mortal((SV *)(stack = newAV()));
13614 sv_2mortal((SV *)(fence_stack = newAV()));
13616 while (RExC_parse < RExC_end) {
13617 I32 top_index; /* Index of top-most element in 'stack' */
13618 SV** top_ptr; /* Pointer to top 'stack' element */
13619 SV* current = NULL; /* To contain the current inversion list
13621 SV* only_to_avoid_leaks;
13623 /* Skip white space */
13624 RExC_parse = regpatws(pRExC_state, RExC_parse,
13625 TRUE /* means recognize comments */ );
13626 if (RExC_parse >= RExC_end) {
13627 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
13630 curchar = UCHARAT(RExC_parse);
13634 top_index = av_tindex(stack);
13637 SV** stacked_ptr; /* Ptr to something already on 'stack' */
13638 char stacked_operator; /* The topmost operator on the 'stack'. */
13639 SV* lhs; /* Operand to the left of the operator */
13640 SV* rhs; /* Operand to the right of the operator */
13641 SV* fence_ptr; /* Pointer to top element of the fence
13646 if (RExC_parse < RExC_end && (UCHARAT(RExC_parse + 1) == '?'))
13648 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
13649 * This happens when we have some thing like
13651 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
13653 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
13655 * Here we would be handling the interpolated
13656 * '$thai_or_lao'. We handle this by a recursive call to
13657 * ourselves which returns the inversion list the
13658 * interpolated expression evaluates to. We use the flags
13659 * from the interpolated pattern. */
13660 U32 save_flags = RExC_flags;
13661 const char * save_parse;
13663 RExC_parse += 2; /* Skip past the '(?' */
13664 save_parse = RExC_parse;
13666 /* Parse any flags for the '(?' */
13667 parse_lparen_question_flags(pRExC_state);
13669 if (RExC_parse == save_parse /* Makes sure there was at
13670 least one flag (or else
13671 this embedding wasn't
13673 || RExC_parse >= RExC_end - 4
13674 || UCHARAT(RExC_parse) != ':'
13675 || UCHARAT(++RExC_parse) != '('
13676 || UCHARAT(++RExC_parse) != '?'
13677 || UCHARAT(++RExC_parse) != '[')
13680 /* In combination with the above, this moves the
13681 * pointer to the point just after the first erroneous
13682 * character (or if there are no flags, to where they
13683 * should have been) */
13684 if (RExC_parse >= RExC_end - 4) {
13685 RExC_parse = RExC_end;
13687 else if (RExC_parse != save_parse) {
13688 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13690 vFAIL("Expecting '(?flags:(?[...'");
13693 /* Recurse, with the meat of the embedded expression */
13695 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
13696 depth+1, oregcomp_parse);
13698 /* Here, 'current' contains the embedded expression's
13699 * inversion list, and RExC_parse points to the trailing
13700 * ']'; the next character should be the ')' */
13702 assert(RExC_parse < RExC_end && UCHARAT(RExC_parse) == ')');
13704 /* Then the ')' matching the original '(' handled by this
13705 * case: statement */
13707 assert(RExC_parse < RExC_end && UCHARAT(RExC_parse) == ')');
13710 RExC_flags = save_flags;
13711 goto handle_operand;
13714 /* A regular '('. Look behind for illegal syntax */
13715 if (top_index - fence >= 0) {
13716 /* If the top entry on the stack is an operator, it had
13717 * better be a '!', otherwise the entry below the top
13718 * operand should be an operator */
13719 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
13720 || (IS_OPERATOR(*top_ptr) && SvUV(*top_ptr) != '!')
13721 || ( IS_OPERAND(*top_ptr)
13722 && ( top_index - fence < 1
13723 || ! (stacked_ptr = av_fetch(stack,
13726 || ! IS_OPERATOR(*stacked_ptr))))
13729 vFAIL("Unexpected '(' with no preceding operator");
13733 /* Stack the position of this undealt-with left paren */
13734 fence = top_index + 1;
13735 av_push(fence_stack, newSViv(fence));
13739 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
13740 * multi-char folds are allowed. */
13741 if (!regclass(pRExC_state, flagp,depth+1,
13742 TRUE, /* means parse just the next thing */
13743 FALSE, /* don't allow multi-char folds */
13744 FALSE, /* don't silence non-portable warnings. */
13746 FALSE, /* Require return to be an ANYOF */
13749 FAIL2("panic: regclass returned NULL to handle_sets, "
13750 "flags=%#"UVxf"", (UV) *flagp);
13753 /* regclass() will return with parsing just the \ sequence,
13754 * leaving the parse pointer at the next thing to parse */
13756 goto handle_operand;
13758 case '[': /* Is a bracketed character class */
13760 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
13762 if (! is_posix_class) {
13766 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if
13767 * multi-char folds are allowed. */
13768 if(!regclass(pRExC_state, flagp,depth+1,
13769 is_posix_class, /* parse the whole char class
13770 only if not a posix class */
13771 FALSE, /* don't allow multi-char folds */
13772 FALSE, /* don't silence non-portable warnings. */
13774 FALSE, /* Require return to be an ANYOF */
13778 FAIL2("panic: regclass returned NULL to handle_sets, "
13779 "flags=%#"UVxf"", (UV) *flagp);
13782 /* function call leaves parse pointing to the ']', except if we
13784 if (is_posix_class) {
13788 goto handle_operand;
13792 if (top_index >= 1) {
13793 goto join_operators;
13796 /* Only a single operand on the stack: are done */
13800 if (av_tindex(fence_stack) < 0) {
13802 vFAIL("Unexpected ')'");
13805 /* If at least two thing on the stack, treat this as an
13807 if (top_index - fence >= 1) {
13808 goto join_operators;
13811 /* Here only a single thing on the fenced stack, and there is a
13812 * fence. Get rid of it */
13813 fence_ptr = av_pop(fence_stack);
13815 fence = SvIV(fence_ptr) - 1;
13816 SvREFCNT_dec_NN(fence_ptr);
13823 /* Having gotten rid of the fence, we pop the operand at the
13824 * stack top and process it as a newly encountered operand */
13825 current = av_pop(stack);
13826 assert(IS_OPERAND(current));
13827 goto handle_operand;
13835 /* These binary operators should have a left operand already
13837 if ( top_index - fence < 0
13838 || top_index - fence == 1
13839 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
13840 || ! IS_OPERAND(*top_ptr))
13842 goto unexpected_binary;
13845 /* If only the one operand is on the part of the stack visible
13846 * to us, we just place this operator in the proper position */
13847 if (top_index - fence < 2) {
13849 /* Place the operator before the operand */
13851 SV* lhs = av_pop(stack);
13852 av_push(stack, newSVuv(curchar));
13853 av_push(stack, lhs);
13857 /* But if there is something else on the stack, we need to
13858 * process it before this new operator if and only if the
13859 * stacked operation has equal or higher precedence than the
13864 /* The operator on the stack is supposed to be below both its
13866 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
13867 || IS_OPERAND(*stacked_ptr))
13869 /* But if not, it's legal and indicates we are completely
13870 * done if and only if we're currently processing a ']',
13871 * which should be the final thing in the expression */
13872 if (curchar == ']') {
13878 vFAIL2("Unexpected binary operator '%c' with no "
13879 "preceding operand", curchar);
13881 stacked_operator = (char) SvUV(*stacked_ptr);
13883 if (regex_set_precedence(curchar)
13884 > regex_set_precedence(stacked_operator))
13886 /* Here, the new operator has higher precedence than the
13887 * stacked one. This means we need to add the new one to
13888 * the stack to await its rhs operand (and maybe more
13889 * stuff). We put it before the lhs operand, leaving
13890 * untouched the stacked operator and everything below it
13892 lhs = av_pop(stack);
13893 assert(IS_OPERAND(lhs));
13895 av_push(stack, newSVuv(curchar));
13896 av_push(stack, lhs);
13900 /* Here, the new operator has equal or lower precedence than
13901 * what's already there. This means the operation already
13902 * there should be performed now, before the new one. */
13903 rhs = av_pop(stack);
13904 lhs = av_pop(stack);
13906 assert(IS_OPERAND(rhs));
13907 assert(IS_OPERAND(lhs));
13909 switch (stacked_operator) {
13911 _invlist_intersection(lhs, rhs, &rhs);
13916 _invlist_union(lhs, rhs, &rhs);
13920 _invlist_subtract(lhs, rhs, &rhs);
13923 case '^': /* The union minus the intersection */
13929 _invlist_union(lhs, rhs, &u);
13930 _invlist_intersection(lhs, rhs, &i);
13931 /* _invlist_subtract will overwrite rhs
13932 without freeing what it already contains */
13934 _invlist_subtract(u, i, &rhs);
13935 SvREFCNT_dec_NN(i);
13936 SvREFCNT_dec_NN(u);
13937 SvREFCNT_dec_NN(element);
13943 /* Here, the higher precedence operation has been done, and the
13944 * result is in 'rhs'. We overwrite the stacked operator with
13945 * the result. Then we redo this code to either push the new
13946 * operator onto the stack or perform any higher precedence
13947 * stacked operation */
13948 only_to_avoid_leaks = av_pop(stack);
13949 SvREFCNT_dec(only_to_avoid_leaks);
13950 av_push(stack, rhs);
13953 case '!': /* Highest priority, right associative, so just push
13955 av_push(stack, newSVuv(curchar));
13959 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13960 vFAIL("Unexpected character");
13964 /* Here 'current' is the operand. If something is already on the
13965 * stack, we have to check if it is a !. */
13966 top_index = av_tindex(stack); /* Code above may have altered the
13967 * stack in the time since we
13968 * earlier set 'top_index'. */
13969 if (top_index - fence >= 0) {
13970 /* If the top entry on the stack is an operator, it had better
13971 * be a '!', otherwise the entry below the top operand should
13972 * be an operator */
13973 top_ptr = av_fetch(stack, top_index, FALSE);
13975 if (IS_OPERATOR(*top_ptr)) {
13977 /* The only permissible operator at the top of the stack is
13978 * '!', which is applied immediately to this operand. */
13979 curchar = (char) SvUV(*top_ptr);
13980 if (curchar != '!') {
13981 SvREFCNT_dec(current);
13982 vFAIL2("Unexpected binary operator '%c' with no "
13983 "preceding operand", curchar);
13986 _invlist_invert(current);
13988 only_to_avoid_leaks = av_pop(stack);
13989 SvREFCNT_dec(only_to_avoid_leaks);
13990 top_index = av_tindex(stack);
13992 /* And we redo with the inverted operand. This allows
13993 * handling multiple ! in a row */
13994 goto handle_operand;
13996 /* Single operand is ok only for the non-binary ')'
13998 else if ((top_index - fence == 0 && curchar != ')')
13999 || (top_index - fence > 0
14000 && (! (stacked_ptr = av_fetch(stack,
14003 || IS_OPERAND(*stacked_ptr))))
14005 SvREFCNT_dec(current);
14006 vFAIL("Operand with no preceding operator");
14010 /* Here there was nothing on the stack or the top element was
14011 * another operand. Just add this new one */
14012 av_push(stack, current);
14014 } /* End of switch on next parse token */
14016 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
14017 } /* End of loop parsing through the construct */
14020 if (av_tindex(fence_stack) >= 0) {
14021 vFAIL("Unmatched (");
14024 if (av_tindex(stack) < 0 /* Was empty */
14025 || ((final = av_pop(stack)) == NULL)
14026 || ! IS_OPERAND(final)
14027 || SvTYPE(final) != SVt_INVLIST
14028 || av_tindex(stack) >= 0) /* More left on stack */
14030 SvREFCNT_dec(final);
14031 vFAIL("Incomplete expression within '(?[ ])'");
14034 /* Here, 'final' is the resultant inversion list from evaluating the
14035 * expression. Return it if so requested */
14036 if (return_invlist) {
14037 *return_invlist = final;
14041 /* Otherwise generate a resultant node, based on 'final'. regclass() is
14042 * expecting a string of ranges and individual code points */
14043 invlist_iterinit(final);
14044 result_string = newSVpvs("");
14045 while (invlist_iternext(final, &start, &end)) {
14046 if (start == end) {
14047 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
14050 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
14055 /* About to generate an ANYOF (or similar) node from the inversion list we
14056 * have calculated */
14057 save_parse = RExC_parse;
14058 RExC_parse = SvPV(result_string, len);
14059 save_end = RExC_end;
14060 RExC_end = RExC_parse + len;
14062 /* We turn off folding around the call, as the class we have constructed
14063 * already has all folding taken into consideration, and we don't want
14064 * regclass() to add to that */
14065 RExC_flags &= ~RXf_PMf_FOLD;
14066 /* regclass() can only return RESTART_PASS1 and NEED_UTF8 if multi-char
14067 * folds are allowed. */
14068 node = regclass(pRExC_state, flagp,depth+1,
14069 FALSE, /* means parse the whole char class */
14070 FALSE, /* don't allow multi-char folds */
14071 TRUE, /* silence non-portable warnings. The above may very
14072 well have generated non-portable code points, but
14073 they're valid on this machine */
14074 FALSE, /* similarly, no need for strict */
14075 FALSE, /* Require return to be an ANYOF */
14079 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
14082 /* Fix up the node type if we are in locale. (We have pretended we are
14083 * under /u for the purposes of regclass(), as this construct will only
14084 * work under UTF-8 locales. But now we change the opcode to be ANYOFL (so
14085 * as to cause any warnings about bad locales to be output in regexec.c),
14086 * and add the flag that indicates to check if not in a UTF-8 locale. The
14087 * reason we above forbid optimization into something other than an ANYOF
14088 * node is simply to minimize the number of code changes in regexec.c.
14089 * Otherwise we would have to create new EXACTish node types and deal with
14090 * them. This decision could be revisited should this construct become
14093 * (One might think we could look at the resulting ANYOF node and suppress
14094 * the flag if everything is above 255, as those would be UTF-8 only,
14095 * but this isn't true, as the components that led to that result could
14096 * have been locale-affected, and just happen to cancel each other out
14097 * under UTF-8 locales.) */
14099 set_regex_charset(&RExC_flags, REGEX_LOCALE_CHARSET);
14101 assert(OP(node) == ANYOF);
14104 ANYOF_FLAGS(node) |= ANYOF_LOC_REQ_UTF8;
14108 RExC_flags |= RXf_PMf_FOLD;
14111 RExC_parse = save_parse + 1;
14112 RExC_end = save_end;
14113 SvREFCNT_dec_NN(final);
14114 SvREFCNT_dec_NN(result_string);
14116 nextchar(pRExC_state);
14117 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
14124 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
14126 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
14127 * innocent-looking character class, like /[ks]/i won't have to go out to
14128 * disk to find the possible matches.
14130 * This should be called only for a Latin1-range code points, cp, which is
14131 * known to be involved in a simple fold with other code points above
14132 * Latin1. It would give false results if /aa has been specified.
14133 * Multi-char folds are outside the scope of this, and must be handled
14136 * XXX It would be better to generate these via regen, in case a new
14137 * version of the Unicode standard adds new mappings, though that is not
14138 * really likely, and may be caught by the default: case of the switch
14141 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
14143 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
14149 add_cp_to_invlist(*invlist, KELVIN_SIGN);
14153 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
14156 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
14157 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
14159 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14160 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14161 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
14163 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14164 *invlist = add_cp_to_invlist(*invlist,
14165 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14168 #ifdef LATIN_CAPITAL_LETTER_SHARP_S /* not defined in early Unicode releases */
14170 case LATIN_SMALL_LETTER_SHARP_S:
14171 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
14176 #if UNICODE_MAJOR_VERSION < 3 \
14177 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0)
14179 /* In 3.0 and earlier, U+0130 folded simply to 'i'; and in 3.0.1 so did
14184 add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_I_WITH_DOT_ABOVE);
14185 # if UNICODE_DOT_DOT_VERSION == 1
14186 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_DOTLESS_I);
14192 /* Use deprecated warning to increase the chances of this being
14195 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
14202 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
14204 /* This adds the string scalar <multi_string> to the array
14205 * <multi_char_matches>. <multi_string> is known to have exactly
14206 * <cp_count> code points in it. This is used when constructing a
14207 * bracketed character class and we find something that needs to match more
14208 * than a single character.
14210 * <multi_char_matches> is actually an array of arrays. Each top-level
14211 * element is an array that contains all the strings known so far that are
14212 * the same length. And that length (in number of code points) is the same
14213 * as the index of the top-level array. Hence, the [2] element is an
14214 * array, each element thereof is a string containing TWO code points;
14215 * while element [3] is for strings of THREE characters, and so on. Since
14216 * this is for multi-char strings there can never be a [0] nor [1] element.
14218 * When we rewrite the character class below, we will do so such that the
14219 * longest strings are written first, so that it prefers the longest
14220 * matching strings first. This is done even if it turns out that any
14221 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
14222 * Christiansen has agreed that this is ok. This makes the test for the
14223 * ligature 'ffi' come before the test for 'ff', for example */
14226 AV** this_array_ptr;
14228 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
14230 if (! multi_char_matches) {
14231 multi_char_matches = newAV();
14234 if (av_exists(multi_char_matches, cp_count)) {
14235 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
14236 this_array = *this_array_ptr;
14239 this_array = newAV();
14240 av_store(multi_char_matches, cp_count,
14243 av_push(this_array, multi_string);
14245 return multi_char_matches;
14248 /* The names of properties whose definitions are not known at compile time are
14249 * stored in this SV, after a constant heading. So if the length has been
14250 * changed since initialization, then there is a run-time definition. */
14251 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
14252 (SvCUR(listsv) != initial_listsv_len)
14255 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
14256 const bool stop_at_1, /* Just parse the next thing, don't
14257 look for a full character class */
14258 bool allow_multi_folds,
14259 const bool silence_non_portable, /* Don't output warnings
14263 bool optimizable, /* ? Allow a non-ANYOF return
14265 SV** ret_invlist /* Return an inversion list, not a node */
14268 /* parse a bracketed class specification. Most of these will produce an
14269 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
14270 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
14271 * under /i with multi-character folds: it will be rewritten following the
14272 * paradigm of this example, where the <multi-fold>s are characters which
14273 * fold to multiple character sequences:
14274 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
14275 * gets effectively rewritten as:
14276 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
14277 * reg() gets called (recursively) on the rewritten version, and this
14278 * function will return what it constructs. (Actually the <multi-fold>s
14279 * aren't physically removed from the [abcdefghi], it's just that they are
14280 * ignored in the recursion by means of a flag:
14281 * <RExC_in_multi_char_class>.)
14283 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
14284 * characters, with the corresponding bit set if that character is in the
14285 * list. For characters above this, a range list or swash is used. There
14286 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
14287 * determinable at compile time
14289 * Returns NULL, setting *flagp to RESTART_PASS1 if the sizing scan needs
14290 * to be restarted, or'd with NEED_UTF8 if the pattern needs to be upgraded
14291 * to UTF-8. This can only happen if ret_invlist is non-NULL.
14294 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
14296 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
14299 IV namedclass = OOB_NAMEDCLASS;
14300 char *rangebegin = NULL;
14301 bool need_class = 0;
14303 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
14304 than just initialized. */
14305 SV* properties = NULL; /* Code points that match \p{} \P{} */
14306 SV* posixes = NULL; /* Code points that match classes like [:word:],
14307 extended beyond the Latin1 range. These have to
14308 be kept separate from other code points for much
14309 of this function because their handling is
14310 different under /i, and for most classes under
14312 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
14313 separate for a while from the non-complemented
14314 versions because of complications with /d
14316 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
14317 treated more simply than the general case,
14318 leading to less compilation and execution
14320 UV element_count = 0; /* Number of distinct elements in the class.
14321 Optimizations may be possible if this is tiny */
14322 AV * multi_char_matches = NULL; /* Code points that fold to more than one
14323 character; used under /i */
14325 char * stop_ptr = RExC_end; /* where to stop parsing */
14326 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
14329 /* Unicode properties are stored in a swash; this holds the current one
14330 * being parsed. If this swash is the only above-latin1 component of the
14331 * character class, an optimization is to pass it directly on to the
14332 * execution engine. Otherwise, it is set to NULL to indicate that there
14333 * are other things in the class that have to be dealt with at execution
14335 SV* swash = NULL; /* Code points that match \p{} \P{} */
14337 /* Set if a component of this character class is user-defined; just passed
14338 * on to the engine */
14339 bool has_user_defined_property = FALSE;
14341 /* inversion list of code points this node matches only when the target
14342 * string is in UTF-8. (Because is under /d) */
14343 SV* depends_list = NULL;
14345 /* Inversion list of code points this node matches regardless of things
14346 * like locale, folding, utf8ness of the target string */
14347 SV* cp_list = NULL;
14349 /* Like cp_list, but code points on this list need to be checked for things
14350 * that fold to/from them under /i */
14351 SV* cp_foldable_list = NULL;
14353 /* Like cp_list, but code points on this list are valid only when the
14354 * runtime locale is UTF-8 */
14355 SV* only_utf8_locale_list = NULL;
14357 /* In a range, if one of the endpoints is non-character-set portable,
14358 * meaning that it hard-codes a code point that may mean a different
14359 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
14360 * mnemonic '\t' which each mean the same character no matter which
14361 * character set the platform is on. */
14362 unsigned int non_portable_endpoint = 0;
14364 /* Is the range unicode? which means on a platform that isn't 1-1 native
14365 * to Unicode (i.e. non-ASCII), each code point in it should be considered
14366 * to be a Unicode value. */
14367 bool unicode_range = FALSE;
14368 bool invert = FALSE; /* Is this class to be complemented */
14370 bool warn_super = ALWAYS_WARN_SUPER;
14372 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
14373 case we need to change the emitted regop to an EXACT. */
14374 const char * orig_parse = RExC_parse;
14375 const SSize_t orig_size = RExC_size;
14376 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
14377 GET_RE_DEBUG_FLAGS_DECL;
14379 PERL_ARGS_ASSERT_REGCLASS;
14381 PERL_UNUSED_ARG(depth);
14384 DEBUG_PARSE("clas");
14386 #if UNICODE_MAJOR_VERSION < 3 /* no multifolds in early Unicode */ \
14387 || (UNICODE_MAJOR_VERSION == 3 && UNICODE_DOT_VERSION == 0 \
14388 && UNICODE_DOT_DOT_VERSION == 0)
14389 allow_multi_folds = FALSE;
14392 /* Assume we are going to generate an ANYOF node. */
14393 ret = reganode(pRExC_state,
14396 : (DEPENDS_SEMANTICS)
14402 RExC_size += ANYOF_SKIP;
14403 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
14406 ANYOF_FLAGS(ret) = 0;
14408 RExC_emit += ANYOF_SKIP;
14409 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
14410 initial_listsv_len = SvCUR(listsv);
14411 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
14415 RExC_parse = regpatws(pRExC_state, RExC_parse,
14416 FALSE /* means don't recognize comments */ );
14419 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
14422 allow_multi_folds = FALSE;
14425 RExC_parse = regpatws(pRExC_state, RExC_parse,
14426 FALSE /* means don't recognize comments */ );
14430 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
14431 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
14432 const char *s = RExC_parse;
14433 const char c = *s++;
14438 while (isWORDCHAR(*s))
14440 if (*s && c == *s && s[1] == ']') {
14441 SAVEFREESV(RExC_rx_sv);
14443 "POSIX syntax [%c %c] belongs inside character classes",
14445 (void)ReREFCNT_inc(RExC_rx_sv);
14449 /* If the caller wants us to just parse a single element, accomplish this
14450 * by faking the loop ending condition */
14451 if (stop_at_1 && RExC_end > RExC_parse) {
14452 stop_ptr = RExC_parse + 1;
14455 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
14456 if (UCHARAT(RExC_parse) == ']')
14457 goto charclassloop;
14460 if (RExC_parse >= stop_ptr) {
14465 RExC_parse = regpatws(pRExC_state, RExC_parse,
14466 FALSE /* means don't recognize comments */ );
14469 if (UCHARAT(RExC_parse) == ']') {
14475 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
14476 save_value = value;
14477 save_prevvalue = prevvalue;
14480 rangebegin = RExC_parse;
14482 non_portable_endpoint = 0;
14485 value = utf8n_to_uvchr((U8*)RExC_parse,
14486 RExC_end - RExC_parse,
14487 &numlen, UTF8_ALLOW_DEFAULT);
14488 RExC_parse += numlen;
14491 value = UCHARAT(RExC_parse++);
14494 && RExC_parse < RExC_end
14495 && POSIXCC(UCHARAT(RExC_parse)))
14497 namedclass = regpposixcc(pRExC_state, value, strict);
14499 else if (value == '\\') {
14500 /* Is a backslash; get the code point of the char after it */
14501 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
14502 value = utf8n_to_uvchr((U8*)RExC_parse,
14503 RExC_end - RExC_parse,
14504 &numlen, UTF8_ALLOW_DEFAULT);
14505 RExC_parse += numlen;
14508 value = UCHARAT(RExC_parse++);
14510 /* Some compilers cannot handle switching on 64-bit integer
14511 * values, therefore value cannot be an UV. Yes, this will
14512 * be a problem later if we want switch on Unicode.
14513 * A similar issue a little bit later when switching on
14514 * namedclass. --jhi */
14516 /* If the \ is escaping white space when white space is being
14517 * skipped, it means that that white space is wanted literally, and
14518 * is already in 'value'. Otherwise, need to translate the escape
14519 * into what it signifies. */
14520 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
14522 case 'w': namedclass = ANYOF_WORDCHAR; break;
14523 case 'W': namedclass = ANYOF_NWORDCHAR; break;
14524 case 's': namedclass = ANYOF_SPACE; break;
14525 case 'S': namedclass = ANYOF_NSPACE; break;
14526 case 'd': namedclass = ANYOF_DIGIT; break;
14527 case 'D': namedclass = ANYOF_NDIGIT; break;
14528 case 'v': namedclass = ANYOF_VERTWS; break;
14529 case 'V': namedclass = ANYOF_NVERTWS; break;
14530 case 'h': namedclass = ANYOF_HORIZWS; break;
14531 case 'H': namedclass = ANYOF_NHORIZWS; break;
14532 case 'N': /* Handle \N{NAME} in class */
14534 const char * const backslash_N_beg = RExC_parse - 2;
14537 if (! grok_bslash_N(pRExC_state,
14538 NULL, /* No regnode */
14539 &value, /* Yes single value */
14540 &cp_count, /* Multiple code pt count */
14545 if (*flagp & NEED_UTF8)
14546 FAIL("panic: grok_bslash_N set NEED_UTF8");
14547 if (*flagp & RESTART_PASS1)
14550 if (cp_count < 0) {
14551 vFAIL("\\N in a character class must be a named character: \\N{...}");
14553 else if (cp_count == 0) {
14555 RExC_parse++; /* Position after the "}" */
14556 vFAIL("Zero length \\N{}");
14559 ckWARNreg(RExC_parse,
14560 "Ignoring zero length \\N{} in character class");
14563 else { /* cp_count > 1 */
14564 if (! RExC_in_multi_char_class) {
14565 if (invert || range || *RExC_parse == '-') {
14568 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
14571 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
14573 break; /* <value> contains the first code
14574 point. Drop out of the switch to
14578 SV * multi_char_N = newSVpvn(backslash_N_beg,
14579 RExC_parse - backslash_N_beg);
14581 = add_multi_match(multi_char_matches,
14586 } /* End of cp_count != 1 */
14588 /* This element should not be processed further in this
14591 value = save_value;
14592 prevvalue = save_prevvalue;
14593 continue; /* Back to top of loop to get next char */
14596 /* Here, is a single code point, and <value> contains it */
14597 unicode_range = TRUE; /* \N{} are Unicode */
14605 /* We will handle any undefined properties ourselves */
14606 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
14607 /* And we actually would prefer to get
14608 * the straight inversion list of the
14609 * swash, since we will be accessing it
14610 * anyway, to save a little time */
14611 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
14613 if (RExC_parse >= RExC_end)
14614 vFAIL2("Empty \\%c{}", (U8)value);
14615 if (*RExC_parse == '{') {
14616 const U8 c = (U8)value;
14617 e = strchr(RExC_parse++, '}');
14619 vFAIL2("Missing right brace on \\%c{}", c);
14620 while (isSPACE(*RExC_parse))
14622 if (e == RExC_parse)
14623 vFAIL2("Empty \\%c{}", c);
14624 n = e - RExC_parse;
14625 while (isSPACE(*(RExC_parse + n - 1)))
14636 if (UCHARAT(RExC_parse) == '^') {
14639 /* toggle. (The rhs xor gets the single bit that
14640 * differs between P and p; the other xor inverts just
14642 value ^= 'P' ^ 'p';
14644 while (isSPACE(*RExC_parse)) {
14649 /* Try to get the definition of the property into
14650 * <invlist>. If /i is in effect, the effective property
14651 * will have its name be <__NAME_i>. The design is
14652 * discussed in commit
14653 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
14654 name = savepv(Perl_form(aTHX_
14656 (FOLD) ? "__" : "",
14662 /* Look up the property name, and get its swash and
14663 * inversion list, if the property is found */
14665 SvREFCNT_dec_NN(swash);
14667 swash = _core_swash_init("utf8", name, &PL_sv_undef,
14670 NULL, /* No inversion list */
14673 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
14674 HV* curpkg = (IN_PERL_COMPILETIME)
14676 : CopSTASH(PL_curcop);
14678 SvREFCNT_dec_NN(swash);
14682 /* Here didn't find it. It could be a user-defined
14683 * property that will be available at run-time. If we
14684 * accept only compile-time properties, is an error;
14685 * otherwise add it to the list for run-time look up */
14687 RExC_parse = e + 1;
14689 "Property '%"UTF8f"' is unknown",
14690 UTF8fARG(UTF, n, name));
14693 /* If the property name doesn't already have a package
14694 * name, add the current one to it so that it can be
14695 * referred to outside it. [perl #121777] */
14696 if (curpkg && ! instr(name, "::")) {
14697 char* pkgname = HvNAME(curpkg);
14698 if (strNE(pkgname, "main")) {
14699 char* full_name = Perl_form(aTHX_
14703 n = strlen(full_name);
14705 name = savepvn(full_name, n);
14708 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
14709 (value == 'p' ? '+' : '!'),
14710 UTF8fARG(UTF, n, name));
14711 has_user_defined_property = TRUE;
14712 optimizable = FALSE; /* Will have to leave this an
14715 /* We don't know yet, so have to assume that the
14716 * property could match something in the Latin1 range,
14717 * hence something that isn't utf8. Note that this
14718 * would cause things in <depends_list> to match
14719 * inappropriately, except that any \p{}, including
14720 * this one forces Unicode semantics, which means there
14721 * is no <depends_list> */
14723 |= ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES;
14727 /* Here, did get the swash and its inversion list. If
14728 * the swash is from a user-defined property, then this
14729 * whole character class should be regarded as such */
14730 if (swash_init_flags
14731 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
14733 has_user_defined_property = TRUE;
14736 /* We warn on matching an above-Unicode code point
14737 * if the match would return true, except don't
14738 * warn for \p{All}, which has exactly one element
14740 (_invlist_contains_cp(invlist, 0x110000)
14741 && (! (_invlist_len(invlist) == 1
14742 && *invlist_array(invlist) == 0)))
14748 /* Invert if asking for the complement */
14749 if (value == 'P') {
14750 _invlist_union_complement_2nd(properties,
14754 /* The swash can't be used as-is, because we've
14755 * inverted things; delay removing it to here after
14756 * have copied its invlist above */
14757 SvREFCNT_dec_NN(swash);
14761 _invlist_union(properties, invlist, &properties);
14766 RExC_parse = e + 1;
14767 namedclass = ANYOF_UNIPROP; /* no official name, but it's
14770 /* \p means they want Unicode semantics */
14771 REQUIRE_UNI_RULES(flagp, NULL);
14774 case 'n': value = '\n'; break;
14775 case 'r': value = '\r'; break;
14776 case 't': value = '\t'; break;
14777 case 'f': value = '\f'; break;
14778 case 'b': value = '\b'; break;
14779 case 'e': value = ESC_NATIVE; break;
14780 case 'a': value = '\a'; break;
14782 RExC_parse--; /* function expects to be pointed at the 'o' */
14784 const char* error_msg;
14785 bool valid = grok_bslash_o(&RExC_parse,
14788 PASS2, /* warnings only in
14791 silence_non_portable,
14797 non_portable_endpoint++;
14798 if (IN_ENCODING && value < 0x100) {
14799 goto recode_encoding;
14803 RExC_parse--; /* function expects to be pointed at the 'x' */
14805 const char* error_msg;
14806 bool valid = grok_bslash_x(&RExC_parse,
14809 PASS2, /* Output warnings */
14811 silence_non_portable,
14817 non_portable_endpoint++;
14818 if (IN_ENCODING && value < 0x100)
14819 goto recode_encoding;
14822 value = grok_bslash_c(*RExC_parse++, PASS2);
14823 non_portable_endpoint++;
14825 case '0': case '1': case '2': case '3': case '4':
14826 case '5': case '6': case '7':
14828 /* Take 1-3 octal digits */
14829 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14830 numlen = (strict) ? 4 : 3;
14831 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
14832 RExC_parse += numlen;
14835 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
14836 vFAIL("Need exactly 3 octal digits");
14838 else if (! SIZE_ONLY /* like \08, \178 */
14840 && RExC_parse < RExC_end
14841 && isDIGIT(*RExC_parse)
14842 && ckWARN(WARN_REGEXP))
14844 SAVEFREESV(RExC_rx_sv);
14845 reg_warn_non_literal_string(
14847 form_short_octal_warning(RExC_parse, numlen));
14848 (void)ReREFCNT_inc(RExC_rx_sv);
14851 non_portable_endpoint++;
14852 if (IN_ENCODING && value < 0x100)
14853 goto recode_encoding;
14857 if (! RExC_override_recoding) {
14858 SV* enc = _get_encoding();
14859 value = reg_recode((const char)(U8)value, &enc);
14862 vFAIL("Invalid escape in the specified encoding");
14865 ckWARNreg(RExC_parse,
14866 "Invalid escape in the specified encoding");
14872 /* Allow \_ to not give an error */
14873 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
14875 vFAIL2("Unrecognized escape \\%c in character class",
14879 SAVEFREESV(RExC_rx_sv);
14880 ckWARN2reg(RExC_parse,
14881 "Unrecognized escape \\%c in character class passed through",
14883 (void)ReREFCNT_inc(RExC_rx_sv);
14887 } /* End of switch on char following backslash */
14888 } /* end of handling backslash escape sequences */
14890 /* Here, we have the current token in 'value' */
14892 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
14895 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
14896 * literal, as is the character that began the false range, i.e.
14897 * the 'a' in the examples */
14900 const int w = (RExC_parse >= rangebegin)
14901 ? RExC_parse - rangebegin
14905 "False [] range \"%"UTF8f"\"",
14906 UTF8fARG(UTF, w, rangebegin));
14909 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
14910 ckWARN2reg(RExC_parse,
14911 "False [] range \"%"UTF8f"\"",
14912 UTF8fARG(UTF, w, rangebegin));
14913 (void)ReREFCNT_inc(RExC_rx_sv);
14914 cp_list = add_cp_to_invlist(cp_list, '-');
14915 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
14920 range = 0; /* this was not a true range */
14921 element_count += 2; /* So counts for three values */
14924 classnum = namedclass_to_classnum(namedclass);
14926 if (LOC && namedclass < ANYOF_POSIXL_MAX
14927 #ifndef HAS_ISASCII
14928 && classnum != _CC_ASCII
14931 /* What the Posix classes (like \w, [:space:]) match in locale
14932 * isn't knowable under locale until actual match time. Room
14933 * must be reserved (one time per outer bracketed class) to
14934 * store such classes. The space will contain a bit for each
14935 * named class that is to be matched against. This isn't
14936 * needed for \p{} and pseudo-classes, as they are not affected
14937 * by locale, and hence are dealt with separately */
14938 if (! need_class) {
14941 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14944 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14946 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
14947 ANYOF_POSIXL_ZERO(ret);
14949 /* We can't change this into some other type of node
14950 * (unless this is the only element, in which case there
14951 * are nodes that mean exactly this) as has runtime
14953 optimizable = FALSE;
14956 /* Coverity thinks it is possible for this to be negative; both
14957 * jhi and khw think it's not, but be safer */
14958 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14959 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
14961 /* See if it already matches the complement of this POSIX
14963 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14964 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
14968 posixl_matches_all = TRUE;
14969 break; /* No need to continue. Since it matches both
14970 e.g., \w and \W, it matches everything, and the
14971 bracketed class can be optimized into qr/./s */
14974 /* Add this class to those that should be checked at runtime */
14975 ANYOF_POSIXL_SET(ret, namedclass);
14977 /* The above-Latin1 characters are not subject to locale rules.
14978 * Just add them, in the second pass, to the
14979 * unconditionally-matched list */
14981 SV* scratch_list = NULL;
14983 /* Get the list of the above-Latin1 code points this
14985 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
14986 PL_XPosix_ptrs[classnum],
14988 /* Odd numbers are complements, like
14989 * NDIGIT, NASCII, ... */
14990 namedclass % 2 != 0,
14992 /* Checking if 'cp_list' is NULL first saves an extra
14993 * clone. Its reference count will be decremented at the
14994 * next union, etc, or if this is the only instance, at the
14995 * end of the routine */
14997 cp_list = scratch_list;
15000 _invlist_union(cp_list, scratch_list, &cp_list);
15001 SvREFCNT_dec_NN(scratch_list);
15003 continue; /* Go get next character */
15006 else if (! SIZE_ONLY) {
15008 /* Here, not in pass1 (in that pass we skip calculating the
15009 * contents of this class), and is /l, or is a POSIX class for
15010 * which /l doesn't matter (or is a Unicode property, which is
15011 * skipped here). */
15012 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
15013 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
15015 /* Here, should be \h, \H, \v, or \V. None of /d, /i
15016 * nor /l make a difference in what these match,
15017 * therefore we just add what they match to cp_list. */
15018 if (classnum != _CC_VERTSPACE) {
15019 assert( namedclass == ANYOF_HORIZWS
15020 || namedclass == ANYOF_NHORIZWS);
15022 /* It turns out that \h is just a synonym for
15024 classnum = _CC_BLANK;
15027 _invlist_union_maybe_complement_2nd(
15029 PL_XPosix_ptrs[classnum],
15030 namedclass % 2 != 0, /* Complement if odd
15031 (NHORIZWS, NVERTWS)
15036 else if (UNI_SEMANTICS
15037 || classnum == _CC_ASCII
15038 || (DEPENDS_SEMANTICS && (classnum == _CC_DIGIT
15039 || classnum == _CC_XDIGIT)))
15041 /* We usually have to worry about /d and /a affecting what
15042 * POSIX classes match, with special code needed for /d
15043 * because we won't know until runtime what all matches.
15044 * But there is no extra work needed under /u, and
15045 * [:ascii:] is unaffected by /a and /d; and :digit: and
15046 * :xdigit: don't have runtime differences under /d. So we
15047 * can special case these, and avoid some extra work below,
15048 * and at runtime. */
15049 _invlist_union_maybe_complement_2nd(
15051 PL_XPosix_ptrs[classnum],
15052 namedclass % 2 != 0,
15055 else { /* Garden variety class. If is NUPPER, NALPHA, ...
15056 complement and use nposixes */
15057 SV** posixes_ptr = namedclass % 2 == 0
15060 _invlist_union_maybe_complement_2nd(
15062 PL_XPosix_ptrs[classnum],
15063 namedclass % 2 != 0,
15067 } /* end of namedclass \blah */
15070 RExC_parse = regpatws(pRExC_state, RExC_parse,
15071 FALSE /* means don't recognize comments */ );
15074 /* If 'range' is set, 'value' is the ending of a range--check its
15075 * validity. (If value isn't a single code point in the case of a
15076 * range, we should have figured that out above in the code that
15077 * catches false ranges). Later, we will handle each individual code
15078 * point in the range. If 'range' isn't set, this could be the
15079 * beginning of a range, so check for that by looking ahead to see if
15080 * the next real character to be processed is the range indicator--the
15085 /* For unicode ranges, we have to test that the Unicode as opposed
15086 * to the native values are not decreasing. (Above 255, there is
15087 * no difference between native and Unicode) */
15088 if (unicode_range && prevvalue < 255 && value < 255) {
15089 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
15090 goto backwards_range;
15095 if (prevvalue > value) /* b-a */ {
15100 w = RExC_parse - rangebegin;
15102 "Invalid [] range \"%"UTF8f"\"",
15103 UTF8fARG(UTF, w, rangebegin));
15104 NOT_REACHED; /* NOTREACHED */
15108 prevvalue = value; /* save the beginning of the potential range */
15109 if (! stop_at_1 /* Can't be a range if parsing just one thing */
15110 && *RExC_parse == '-')
15112 char* next_char_ptr = RExC_parse + 1;
15113 if (skip_white) { /* Get the next real char after the '-' */
15114 next_char_ptr = regpatws(pRExC_state,
15116 FALSE); /* means don't recognize
15120 /* If the '-' is at the end of the class (just before the ']',
15121 * it is a literal minus; otherwise it is a range */
15122 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
15123 RExC_parse = next_char_ptr;
15125 /* a bad range like \w-, [:word:]- ? */
15126 if (namedclass > OOB_NAMEDCLASS) {
15127 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
15128 const int w = RExC_parse >= rangebegin
15129 ? RExC_parse - rangebegin
15132 vFAIL4("False [] range \"%*.*s\"",
15137 "False [] range \"%*.*s\"",
15142 cp_list = add_cp_to_invlist(cp_list, '-');
15146 range = 1; /* yeah, it's a range! */
15147 continue; /* but do it the next time */
15152 if (namedclass > OOB_NAMEDCLASS) {
15156 /* Here, we have a single value this time through the loop, and
15157 * <prevvalue> is the beginning of the range, if any; or <value> if
15160 /* non-Latin1 code point implies unicode semantics. Must be set in
15161 * pass1 so is there for the whole of pass 2 */
15163 REQUIRE_UNI_RULES(flagp, NULL);
15166 /* Ready to process either the single value, or the completed range.
15167 * For single-valued non-inverted ranges, we consider the possibility
15168 * of multi-char folds. (We made a conscious decision to not do this
15169 * for the other cases because it can often lead to non-intuitive
15170 * results. For example, you have the peculiar case that:
15171 * "s s" =~ /^[^\xDF]+$/i => Y
15172 * "ss" =~ /^[^\xDF]+$/i => N
15174 * See [perl #89750] */
15175 if (FOLD && allow_multi_folds && value == prevvalue) {
15176 if (value == LATIN_SMALL_LETTER_SHARP_S
15177 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
15180 /* Here <value> is indeed a multi-char fold. Get what it is */
15182 U8 foldbuf[UTF8_MAXBYTES_CASE];
15185 UV folded = _to_uni_fold_flags(
15189 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
15190 ? FOLD_FLAGS_NOMIX_ASCII
15194 /* Here, <folded> should be the first character of the
15195 * multi-char fold of <value>, with <foldbuf> containing the
15196 * whole thing. But, if this fold is not allowed (because of
15197 * the flags), <fold> will be the same as <value>, and should
15198 * be processed like any other character, so skip the special
15200 if (folded != value) {
15202 /* Skip if we are recursed, currently parsing the class
15203 * again. Otherwise add this character to the list of
15204 * multi-char folds. */
15205 if (! RExC_in_multi_char_class) {
15206 STRLEN cp_count = utf8_length(foldbuf,
15207 foldbuf + foldlen);
15208 SV* multi_fold = sv_2mortal(newSVpvs(""));
15210 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
15213 = add_multi_match(multi_char_matches,
15219 /* This element should not be processed further in this
15222 value = save_value;
15223 prevvalue = save_prevvalue;
15229 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
15232 /* If the range starts above 255, everything is portable and
15233 * likely to be so for any forseeable character set, so don't
15235 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
15236 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
15238 else if (prevvalue != value) {
15240 /* Under strict, ranges that stop and/or end in an ASCII
15241 * printable should have each end point be a portable value
15242 * for it (preferably like 'A', but we don't warn if it is
15243 * a (portable) Unicode name or code point), and the range
15244 * must be be all digits or all letters of the same case.
15245 * Otherwise, the range is non-portable and unclear as to
15246 * what it contains */
15247 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
15248 && (non_portable_endpoint
15249 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
15250 || (isLOWER_A(prevvalue) && isLOWER_A(value))
15251 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
15253 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
15255 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
15257 /* But the nature of Unicode and languages mean we
15258 * can't do the same checks for above-ASCII ranges,
15259 * except in the case of digit ones. These should
15260 * contain only digits from the same group of 10. The
15261 * ASCII case is handled just above. 0x660 is the
15262 * first digit character beyond ASCII. Hence here, the
15263 * range could be a range of digits. Find out. */
15264 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
15266 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
15269 /* If the range start and final points are in the same
15270 * inversion list element, it means that either both
15271 * are not digits, or both are digits in a consecutive
15272 * sequence of digits. (So far, Unicode has kept all
15273 * such sequences as distinct groups of 10, but assert
15274 * to make sure). If the end points are not in the
15275 * same element, neither should be a digit. */
15276 if (index_start == index_final) {
15277 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
15278 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
15279 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15281 /* But actually Unicode did have one group of 11
15282 * 'digits' in 5.2, so in case we are operating
15283 * on that version, let that pass */
15284 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
15285 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15287 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15291 else if ((index_start >= 0
15292 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
15293 || (index_final >= 0
15294 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
15296 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
15301 if ((! range || prevvalue == value) && non_portable_endpoint) {
15302 if (isPRINT_A(value)) {
15305 if (isBACKSLASHED_PUNCT(value)) {
15306 literal[d++] = '\\';
15308 literal[d++] = (char) value;
15309 literal[d++] = '\0';
15312 "\"%.*s\" is more clearly written simply as \"%s\"",
15313 (int) (RExC_parse - rangebegin),
15318 else if isMNEMONIC_CNTRL(value) {
15320 "\"%.*s\" is more clearly written simply as \"%s\"",
15321 (int) (RExC_parse - rangebegin),
15323 cntrl_to_mnemonic((char) value)
15329 /* Deal with this element of the class */
15333 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15336 /* On non-ASCII platforms, for ranges that span all of 0..255, and
15337 * ones that don't require special handling, we can just add the
15338 * range like we do for ASCII platforms */
15339 if ((UNLIKELY(prevvalue == 0) && value >= 255)
15340 || ! (prevvalue < 256
15342 || (! non_portable_endpoint
15343 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
15344 || (isUPPER_A(prevvalue)
15345 && isUPPER_A(value)))))))
15347 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15351 /* Here, requires special handling. This can be because it is
15352 * a range whose code points are considered to be Unicode, and
15353 * so must be individually translated into native, or because
15354 * its a subrange of 'A-Z' or 'a-z' which each aren't
15355 * contiguous in EBCDIC, but we have defined them to include
15356 * only the "expected" upper or lower case ASCII alphabetics.
15357 * Subranges above 255 are the same in native and Unicode, so
15358 * can be added as a range */
15359 U8 start = NATIVE_TO_LATIN1(prevvalue);
15361 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
15362 for (j = start; j <= end; j++) {
15363 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
15366 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15373 range = 0; /* this range (if it was one) is done now */
15374 } /* End of loop through all the text within the brackets */
15376 /* If anything in the class expands to more than one character, we have to
15377 * deal with them by building up a substitute parse string, and recursively
15378 * calling reg() on it, instead of proceeding */
15379 if (multi_char_matches) {
15380 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
15383 char *save_end = RExC_end;
15384 char *save_parse = RExC_parse;
15385 bool first_time = TRUE; /* First multi-char occurrence doesn't get
15390 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
15391 because too confusing */
15393 sv_catpv(substitute_parse, "(?:");
15397 /* Look at the longest folds first */
15398 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
15400 if (av_exists(multi_char_matches, cp_count)) {
15401 AV** this_array_ptr;
15404 this_array_ptr = (AV**) av_fetch(multi_char_matches,
15406 while ((this_sequence = av_pop(*this_array_ptr)) !=
15409 if (! first_time) {
15410 sv_catpv(substitute_parse, "|");
15412 first_time = FALSE;
15414 sv_catpv(substitute_parse, SvPVX(this_sequence));
15419 /* If the character class contains anything else besides these
15420 * multi-character folds, have to include it in recursive parsing */
15421 if (element_count) {
15422 sv_catpv(substitute_parse, "|[");
15423 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
15424 sv_catpv(substitute_parse, "]");
15427 sv_catpv(substitute_parse, ")");
15430 /* This is a way to get the parse to skip forward a whole named
15431 * sequence instead of matching the 2nd character when it fails the
15433 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
15437 RExC_parse = SvPV(substitute_parse, len);
15438 RExC_end = RExC_parse + len;
15439 RExC_in_multi_char_class = 1;
15440 RExC_override_recoding = 1;
15441 RExC_emit = (regnode *)orig_emit;
15443 ret = reg(pRExC_state, 1, ®_flags, depth+1);
15445 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_PASS1|NEED_UTF8);
15447 RExC_parse = save_parse;
15448 RExC_end = save_end;
15449 RExC_in_multi_char_class = 0;
15450 RExC_override_recoding = 0;
15451 SvREFCNT_dec_NN(multi_char_matches);
15455 /* Here, we've gone through the entire class and dealt with multi-char
15456 * folds. We are now in a position that we can do some checks to see if we
15457 * can optimize this ANYOF node into a simpler one, even in Pass 1.
15458 * Currently we only do two checks:
15459 * 1) is in the unlikely event that the user has specified both, eg. \w and
15460 * \W under /l, then the class matches everything. (This optimization
15461 * is done only to make the optimizer code run later work.)
15462 * 2) if the character class contains only a single element (including a
15463 * single range), we see if there is an equivalent node for it.
15464 * Other checks are possible */
15466 && ! ret_invlist /* Can't optimize if returning the constructed
15468 && (UNLIKELY(posixl_matches_all) || element_count == 1))
15473 if (UNLIKELY(posixl_matches_all)) {
15476 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
15477 \w or [:digit:] or \p{foo}
15480 /* All named classes are mapped into POSIXish nodes, with its FLAG
15481 * argument giving which class it is */
15482 switch ((I32)namedclass) {
15483 case ANYOF_UNIPROP:
15486 /* These don't depend on the charset modifiers. They always
15487 * match under /u rules */
15488 case ANYOF_NHORIZWS:
15489 case ANYOF_HORIZWS:
15490 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
15493 case ANYOF_NVERTWS:
15498 /* The actual POSIXish node for all the rest depends on the
15499 * charset modifier. The ones in the first set depend only on
15500 * ASCII or, if available on this platform, also locale */
15504 op = (LOC) ? POSIXL : POSIXA;
15510 /* The following don't have any matches in the upper Latin1
15511 * range, hence /d is equivalent to /u for them. Making it /u
15512 * saves some branches at runtime */
15516 case ANYOF_NXDIGIT:
15517 if (! DEPENDS_SEMANTICS) {
15518 goto treat_as_default;
15524 /* The following change to CASED under /i */
15530 namedclass = ANYOF_CASED + (namedclass % 2);
15534 /* The rest have more possibilities depending on the charset.
15535 * We take advantage of the enum ordering of the charset
15536 * modifiers to get the exact node type, */
15539 op = POSIXD + get_regex_charset(RExC_flags);
15540 if (op > POSIXA) { /* /aa is same as /a */
15545 /* The odd numbered ones are the complements of the
15546 * next-lower even number one */
15547 if (namedclass % 2 == 1) {
15551 arg = namedclass_to_classnum(namedclass);
15555 else if (value == prevvalue) {
15557 /* Here, the class consists of just a single code point */
15560 if (! LOC && value == '\n') {
15561 op = REG_ANY; /* Optimize [^\n] */
15562 *flagp |= HASWIDTH|SIMPLE;
15566 else if (value < 256 || UTF) {
15568 /* Optimize a single value into an EXACTish node, but not if it
15569 * would require converting the pattern to UTF-8. */
15570 op = compute_EXACTish(pRExC_state);
15572 } /* Otherwise is a range */
15573 else if (! LOC) { /* locale could vary these */
15574 if (prevvalue == '0') {
15575 if (value == '9') {
15580 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
15581 /* We can optimize A-Z or a-z, but not if they could match
15582 * something like the KELVIN SIGN under /i. */
15583 if (prevvalue == 'A') {
15586 && ! non_portable_endpoint
15589 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
15593 else if (prevvalue == 'a') {
15596 && ! non_portable_endpoint
15599 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
15606 /* Here, we have changed <op> away from its initial value iff we found
15607 * an optimization */
15610 /* Throw away this ANYOF regnode, and emit the calculated one,
15611 * which should correspond to the beginning, not current, state of
15613 const char * cur_parse = RExC_parse;
15614 RExC_parse = (char *)orig_parse;
15618 /* To get locale nodes to not use the full ANYOF size would
15619 * require moving the code above that writes the portions
15620 * of it that aren't in other nodes to after this point.
15621 * e.g. ANYOF_POSIXL_SET */
15622 RExC_size = orig_size;
15626 RExC_emit = (regnode *)orig_emit;
15627 if (PL_regkind[op] == POSIXD) {
15628 if (op == POSIXL) {
15629 RExC_contains_locale = 1;
15632 op += NPOSIXD - POSIXD;
15637 ret = reg_node(pRExC_state, op);
15639 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
15643 *flagp |= HASWIDTH|SIMPLE;
15645 else if (PL_regkind[op] == EXACT) {
15646 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15647 TRUE /* downgradable to EXACT */
15651 RExC_parse = (char *) cur_parse;
15653 SvREFCNT_dec(posixes);
15654 SvREFCNT_dec(nposixes);
15655 SvREFCNT_dec(simple_posixes);
15656 SvREFCNT_dec(cp_list);
15657 SvREFCNT_dec(cp_foldable_list);
15664 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
15666 /* If folding, we calculate all characters that could fold to or from the
15667 * ones already on the list */
15668 if (cp_foldable_list) {
15670 UV start, end; /* End points of code point ranges */
15672 SV* fold_intersection = NULL;
15675 /* Our calculated list will be for Unicode rules. For locale
15676 * matching, we have to keep a separate list that is consulted at
15677 * runtime only when the locale indicates Unicode rules. For
15678 * non-locale, we just use the general list */
15680 use_list = &only_utf8_locale_list;
15683 use_list = &cp_list;
15686 /* Only the characters in this class that participate in folds need
15687 * be checked. Get the intersection of this class and all the
15688 * possible characters that are foldable. This can quickly narrow
15689 * down a large class */
15690 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
15691 &fold_intersection);
15693 /* The folds for all the Latin1 characters are hard-coded into this
15694 * program, but we have to go out to disk to get the others. */
15695 if (invlist_highest(cp_foldable_list) >= 256) {
15697 /* This is a hash that for a particular fold gives all
15698 * characters that are involved in it */
15699 if (! PL_utf8_foldclosures) {
15700 _load_PL_utf8_foldclosures();
15704 /* Now look at the foldable characters in this class individually */
15705 invlist_iterinit(fold_intersection);
15706 while (invlist_iternext(fold_intersection, &start, &end)) {
15709 /* Look at every character in the range */
15710 for (j = start; j <= end; j++) {
15711 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
15717 if (IS_IN_SOME_FOLD_L1(j)) {
15719 /* ASCII is always matched; non-ASCII is matched
15720 * only under Unicode rules (which could happen
15721 * under /l if the locale is a UTF-8 one */
15722 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
15723 *use_list = add_cp_to_invlist(*use_list,
15724 PL_fold_latin1[j]);
15728 add_cp_to_invlist(depends_list,
15729 PL_fold_latin1[j]);
15733 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
15734 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
15736 add_above_Latin1_folds(pRExC_state,
15743 /* Here is an above Latin1 character. We don't have the
15744 * rules hard-coded for it. First, get its fold. This is
15745 * the simple fold, as the multi-character folds have been
15746 * handled earlier and separated out */
15747 _to_uni_fold_flags(j, foldbuf, &foldlen,
15748 (ASCII_FOLD_RESTRICTED)
15749 ? FOLD_FLAGS_NOMIX_ASCII
15752 /* Single character fold of above Latin1. Add everything in
15753 * its fold closure to the list that this node should match.
15754 * The fold closures data structure is a hash with the keys
15755 * being the UTF-8 of every character that is folded to, like
15756 * 'k', and the values each an array of all code points that
15757 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
15758 * Multi-character folds are not included */
15759 if ((listp = hv_fetch(PL_utf8_foldclosures,
15760 (char *) foldbuf, foldlen, FALSE)))
15762 AV* list = (AV*) *listp;
15764 for (k = 0; k <= av_tindex(list); k++) {
15765 SV** c_p = av_fetch(list, k, FALSE);
15771 /* /aa doesn't allow folds between ASCII and non- */
15772 if ((ASCII_FOLD_RESTRICTED
15773 && (isASCII(c) != isASCII(j))))
15778 /* Folds under /l which cross the 255/256 boundary
15779 * are added to a separate list. (These are valid
15780 * only when the locale is UTF-8.) */
15781 if (c < 256 && LOC) {
15782 *use_list = add_cp_to_invlist(*use_list, c);
15786 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
15788 cp_list = add_cp_to_invlist(cp_list, c);
15791 /* Similarly folds involving non-ascii Latin1
15792 * characters under /d are added to their list */
15793 depends_list = add_cp_to_invlist(depends_list,
15800 SvREFCNT_dec_NN(fold_intersection);
15803 /* Now that we have finished adding all the folds, there is no reason
15804 * to keep the foldable list separate */
15805 _invlist_union(cp_list, cp_foldable_list, &cp_list);
15806 SvREFCNT_dec_NN(cp_foldable_list);
15809 /* And combine the result (if any) with any inversion list from posix
15810 * classes. The lists are kept separate up to now because we don't want to
15811 * fold the classes (folding of those is automatically handled by the swash
15812 * fetching code) */
15813 if (simple_posixes) {
15814 _invlist_union(cp_list, simple_posixes, &cp_list);
15815 SvREFCNT_dec_NN(simple_posixes);
15817 if (posixes || nposixes) {
15818 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
15819 /* Under /a and /aa, nothing above ASCII matches these */
15820 _invlist_intersection(posixes,
15821 PL_XPosix_ptrs[_CC_ASCII],
15825 if (DEPENDS_SEMANTICS) {
15826 /* Under /d, everything in the upper half of the Latin1 range
15827 * matches these complements */
15828 ANYOF_FLAGS(ret) |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
15830 else if (AT_LEAST_ASCII_RESTRICTED) {
15831 /* Under /a and /aa, everything above ASCII matches these
15833 _invlist_union_complement_2nd(nposixes,
15834 PL_XPosix_ptrs[_CC_ASCII],
15838 _invlist_union(posixes, nposixes, &posixes);
15839 SvREFCNT_dec_NN(nposixes);
15842 posixes = nposixes;
15845 if (! DEPENDS_SEMANTICS) {
15847 _invlist_union(cp_list, posixes, &cp_list);
15848 SvREFCNT_dec_NN(posixes);
15855 /* Under /d, we put into a separate list the Latin1 things that
15856 * match only when the target string is utf8 */
15857 SV* nonascii_but_latin1_properties = NULL;
15858 _invlist_intersection(posixes, PL_UpperLatin1,
15859 &nonascii_but_latin1_properties);
15860 _invlist_subtract(posixes, nonascii_but_latin1_properties,
15863 _invlist_union(cp_list, posixes, &cp_list);
15864 SvREFCNT_dec_NN(posixes);
15870 if (depends_list) {
15871 _invlist_union(depends_list, nonascii_but_latin1_properties,
15873 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
15876 depends_list = nonascii_but_latin1_properties;
15881 /* And combine the result (if any) with any inversion list from properties.
15882 * The lists are kept separate up to now so that we can distinguish the two
15883 * in regards to matching above-Unicode. A run-time warning is generated
15884 * if a Unicode property is matched against a non-Unicode code point. But,
15885 * we allow user-defined properties to match anything, without any warning,
15886 * and we also suppress the warning if there is a portion of the character
15887 * class that isn't a Unicode property, and which matches above Unicode, \W
15888 * or [\x{110000}] for example.
15889 * (Note that in this case, unlike the Posix one above, there is no
15890 * <depends_list>, because having a Unicode property forces Unicode
15895 /* If it matters to the final outcome, see if a non-property
15896 * component of the class matches above Unicode. If so, the
15897 * warning gets suppressed. This is true even if just a single
15898 * such code point is specified, as though not strictly correct if
15899 * another such code point is matched against, the fact that they
15900 * are using above-Unicode code points indicates they should know
15901 * the issues involved */
15903 warn_super = ! (invert
15904 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
15907 _invlist_union(properties, cp_list, &cp_list);
15908 SvREFCNT_dec_NN(properties);
15911 cp_list = properties;
15916 |= ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER;
15918 /* Because an ANYOF node is the only one that warns, this node
15919 * can't be optimized into something else */
15920 optimizable = FALSE;
15924 /* Here, we have calculated what code points should be in the character
15927 * Now we can see about various optimizations. Fold calculation (which we
15928 * did above) needs to take place before inversion. Otherwise /[^k]/i
15929 * would invert to include K, which under /i would match k, which it
15930 * shouldn't. Therefore we can't invert folded locale now, as it won't be
15931 * folded until runtime */
15933 /* If we didn't do folding, it's because some information isn't available
15934 * until runtime; set the run-time fold flag for these. (We don't have to
15935 * worry about properties folding, as that is taken care of by the swash
15936 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
15937 * locales, or the class matches at least one 0-255 range code point */
15939 if (only_utf8_locale_list) {
15940 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15942 else if (cp_list) { /* Look to see if a 0-255 code point is in list */
15944 invlist_iterinit(cp_list);
15945 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
15946 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15948 invlist_iterfinish(cp_list);
15952 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
15953 * at compile time. Besides not inverting folded locale now, we can't
15954 * invert if there are things such as \w, which aren't known until runtime
15958 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15960 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15962 _invlist_invert(cp_list);
15964 /* Any swash can't be used as-is, because we've inverted things */
15966 SvREFCNT_dec_NN(swash);
15970 /* Clear the invert flag since have just done it here */
15977 *ret_invlist = cp_list;
15978 SvREFCNT_dec(swash);
15980 /* Discard the generated node */
15982 RExC_size = orig_size;
15985 RExC_emit = orig_emit;
15990 /* Some character classes are equivalent to other nodes. Such nodes take
15991 * up less room and generally fewer operations to execute than ANYOF nodes.
15992 * Above, we checked for and optimized into some such equivalents for
15993 * certain common classes that are easy to test. Getting to this point in
15994 * the code means that the class didn't get optimized there. Since this
15995 * code is only executed in Pass 2, it is too late to save space--it has
15996 * been allocated in Pass 1, and currently isn't given back. But turning
15997 * things into an EXACTish node can allow the optimizer to join it to any
15998 * adjacent such nodes. And if the class is equivalent to things like /./,
15999 * expensive run-time swashes can be avoided. Now that we have more
16000 * complete information, we can find things necessarily missed by the
16001 * earlier code. I (khw) did some benchmarks and found essentially no
16002 * speed difference between using a POSIXA node versus an ANYOF node, so
16003 * there is no reason to optimize, for example [A-Za-z0-9_] into
16004 * [[:word:]]/a (although if we did it in the sizing pass it would save
16005 * space). _invlistEQ() could be used if one ever wanted to do something
16006 * like this at this point in the code */
16008 if (optimizable && cp_list && ! invert && ! depends_list) {
16010 U8 op = END; /* The optimzation node-type */
16011 const char * cur_parse= RExC_parse;
16013 invlist_iterinit(cp_list);
16014 if (! invlist_iternext(cp_list, &start, &end)) {
16016 /* Here, the list is empty. This happens, for example, when a
16017 * Unicode property that doesn't match anything is the only element
16018 * in the character class (perluniprops.pod notes such properties).
16021 *flagp |= HASWIDTH|SIMPLE;
16023 else if (start == end) { /* The range is a single code point */
16024 if (! invlist_iternext(cp_list, &start, &end)
16026 /* Don't do this optimization if it would require changing
16027 * the pattern to UTF-8 */
16028 && (start < 256 || UTF))
16030 /* Here, the list contains a single code point. Can optimize
16031 * into an EXACTish node */
16042 /* A locale node under folding with one code point can be
16043 * an EXACTFL, as its fold won't be calculated until
16049 /* Here, we are generally folding, but there is only one
16050 * code point to match. If we have to, we use an EXACT
16051 * node, but it would be better for joining with adjacent
16052 * nodes in the optimization pass if we used the same
16053 * EXACTFish node that any such are likely to be. We can
16054 * do this iff the code point doesn't participate in any
16055 * folds. For example, an EXACTF of a colon is the same as
16056 * an EXACT one, since nothing folds to or from a colon. */
16058 if (IS_IN_SOME_FOLD_L1(value)) {
16063 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
16068 /* If we haven't found the node type, above, it means we
16069 * can use the prevailing one */
16071 op = compute_EXACTish(pRExC_state);
16075 } /* End of first range contains just a single code point */
16076 else if (start == 0) {
16077 if (end == UV_MAX) {
16079 *flagp |= HASWIDTH|SIMPLE;
16082 else if (end == '\n' - 1
16083 && invlist_iternext(cp_list, &start, &end)
16084 && start == '\n' + 1 && end == UV_MAX)
16087 *flagp |= HASWIDTH|SIMPLE;
16091 invlist_iterfinish(cp_list);
16094 RExC_parse = (char *)orig_parse;
16095 RExC_emit = (regnode *)orig_emit;
16097 if (regarglen[op]) {
16098 ret = reganode(pRExC_state, op, 0);
16100 ret = reg_node(pRExC_state, op);
16103 RExC_parse = (char *)cur_parse;
16105 if (PL_regkind[op] == EXACT) {
16106 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
16107 TRUE /* downgradable to EXACT */
16111 SvREFCNT_dec_NN(cp_list);
16116 /* Here, <cp_list> contains all the code points we can determine at
16117 * compile time that match under all conditions. Go through it, and
16118 * for things that belong in the bitmap, put them there, and delete from
16119 * <cp_list>. While we are at it, see if everything above 255 is in the
16120 * list, and if so, set a flag to speed up execution */
16122 populate_ANYOF_from_invlist(ret, &cp_list);
16125 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
16128 /* Here, the bitmap has been populated with all the Latin1 code points that
16129 * always match. Can now add to the overall list those that match only
16130 * when the target string is UTF-8 (<depends_list>). */
16131 if (depends_list) {
16133 _invlist_union(cp_list, depends_list, &cp_list);
16134 SvREFCNT_dec_NN(depends_list);
16137 cp_list = depends_list;
16139 ANYOF_FLAGS(ret) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
16142 /* If there is a swash and more than one element, we can't use the swash in
16143 * the optimization below. */
16144 if (swash && element_count > 1) {
16145 SvREFCNT_dec_NN(swash);
16149 /* Note that the optimization of using 'swash' if it is the only thing in
16150 * the class doesn't have us change swash at all, so it can include things
16151 * that are also in the bitmap; otherwise we have purposely deleted that
16152 * duplicate information */
16153 set_ANYOF_arg(pRExC_state, ret, cp_list,
16154 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
16156 only_utf8_locale_list,
16157 swash, has_user_defined_property);
16159 *flagp |= HASWIDTH|SIMPLE;
16161 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
16162 RExC_contains_locale = 1;
16168 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
16171 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
16172 regnode* const node,
16174 SV* const runtime_defns,
16175 SV* const only_utf8_locale_list,
16177 const bool has_user_defined_property)
16179 /* Sets the arg field of an ANYOF-type node 'node', using information about
16180 * the node passed-in. If there is nothing outside the node's bitmap, the
16181 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
16182 * the count returned by add_data(), having allocated and stored an array,
16183 * av, that that count references, as follows:
16184 * av[0] stores the character class description in its textual form.
16185 * This is used later (regexec.c:Perl_regclass_swash()) to
16186 * initialize the appropriate swash, and is also useful for dumping
16187 * the regnode. This is set to &PL_sv_undef if the textual
16188 * description is not needed at run-time (as happens if the other
16189 * elements completely define the class)
16190 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
16191 * computed from av[0]. But if no further computation need be done,
16192 * the swash is stored here now (and av[0] is &PL_sv_undef).
16193 * av[2] stores the inversion list of code points that match only if the
16194 * current locale is UTF-8
16195 * av[3] stores the cp_list inversion list for use in addition or instead
16196 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
16197 * (Otherwise everything needed is already in av[0] and av[1])
16198 * av[4] is set if any component of the class is from a user-defined
16199 * property; used only if av[3] exists */
16203 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
16205 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
16206 assert(! (ANYOF_FLAGS(node)
16207 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16208 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES)));
16209 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
16212 AV * const av = newAV();
16215 assert(ANYOF_FLAGS(node)
16216 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16217 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
16219 av_store(av, 0, (runtime_defns)
16220 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
16223 av_store(av, 1, swash);
16224 SvREFCNT_dec_NN(cp_list);
16227 av_store(av, 1, &PL_sv_undef);
16229 av_store(av, 3, cp_list);
16230 av_store(av, 4, newSVuv(has_user_defined_property));
16234 if (only_utf8_locale_list) {
16235 av_store(av, 2, only_utf8_locale_list);
16238 av_store(av, 2, &PL_sv_undef);
16241 rv = newRV_noinc(MUTABLE_SV(av));
16242 n = add_data(pRExC_state, STR_WITH_LEN("s"));
16243 RExC_rxi->data->data[n] = (void*)rv;
16248 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
16250 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
16251 const regnode* node,
16254 SV** only_utf8_locale_ptr,
16258 /* For internal core use only.
16259 * Returns the swash for the input 'node' in the regex 'prog'.
16260 * If <doinit> is 'true', will attempt to create the swash if not already
16262 * If <listsvp> is non-null, will return the printable contents of the
16263 * swash. This can be used to get debugging information even before the
16264 * swash exists, by calling this function with 'doinit' set to false, in
16265 * which case the components that will be used to eventually create the
16266 * swash are returned (in a printable form).
16267 * If <exclude_list> is not NULL, it is an inversion list of things to
16268 * exclude from what's returned in <listsvp>.
16269 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
16270 * that, in spite of this function's name, the swash it returns may include
16271 * the bitmap data as well */
16274 SV *si = NULL; /* Input swash initialization string */
16275 SV* invlist = NULL;
16277 RXi_GET_DECL(prog,progi);
16278 const struct reg_data * const data = prog ? progi->data : NULL;
16280 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
16282 assert(ANYOF_FLAGS(node)
16283 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16284 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
16286 if (data && data->count) {
16287 const U32 n = ARG(node);
16289 if (data->what[n] == 's') {
16290 SV * const rv = MUTABLE_SV(data->data[n]);
16291 AV * const av = MUTABLE_AV(SvRV(rv));
16292 SV **const ary = AvARRAY(av);
16293 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
16295 si = *ary; /* ary[0] = the string to initialize the swash with */
16297 /* Elements 3 and 4 are either both present or both absent. [3] is
16298 * any inversion list generated at compile time; [4] indicates if
16299 * that inversion list has any user-defined properties in it. */
16300 if (av_tindex(av) >= 2) {
16301 if (only_utf8_locale_ptr
16303 && ary[2] != &PL_sv_undef)
16305 *only_utf8_locale_ptr = ary[2];
16308 assert(only_utf8_locale_ptr);
16309 *only_utf8_locale_ptr = NULL;
16312 if (av_tindex(av) >= 3) {
16314 if (SvUV(ary[4])) {
16315 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
16323 /* Element [1] is reserved for the set-up swash. If already there,
16324 * return it; if not, create it and store it there */
16325 if (ary[1] && SvROK(ary[1])) {
16328 else if (doinit && ((si && si != &PL_sv_undef)
16329 || (invlist && invlist != &PL_sv_undef))) {
16331 sw = _core_swash_init("utf8", /* the utf8 package */
16335 0, /* not from tr/// */
16337 &swash_init_flags);
16338 (void)av_store(av, 1, sw);
16343 /* If requested, return a printable version of what this swash matches */
16345 SV* matches_string = newSVpvs("");
16347 /* The swash should be used, if possible, to get the data, as it
16348 * contains the resolved data. But this function can be called at
16349 * compile-time, before everything gets resolved, in which case we
16350 * return the currently best available information, which is the string
16351 * that will eventually be used to do that resolving, 'si' */
16352 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
16353 && (si && si != &PL_sv_undef))
16355 sv_catsv(matches_string, si);
16358 /* Add the inversion list to whatever we have. This may have come from
16359 * the swash, or from an input parameter */
16361 if (exclude_list) {
16362 SV* clone = invlist_clone(invlist);
16363 _invlist_subtract(clone, exclude_list, &clone);
16364 sv_catsv(matches_string, _invlist_contents(clone));
16365 SvREFCNT_dec_NN(clone);
16368 sv_catsv(matches_string, _invlist_contents(invlist));
16371 *listsvp = matches_string;
16376 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
16378 /* reg_skipcomment()
16380 Absorbs an /x style # comment from the input stream,
16381 returning a pointer to the first character beyond the comment, or if the
16382 comment terminates the pattern without anything following it, this returns
16383 one past the final character of the pattern (in other words, RExC_end) and
16384 sets the REG_RUN_ON_COMMENT_SEEN flag.
16386 Note it's the callers responsibility to ensure that we are
16387 actually in /x mode
16391 PERL_STATIC_INLINE char*
16392 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
16394 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
16398 while (p < RExC_end) {
16399 if (*(++p) == '\n') {
16404 /* we ran off the end of the pattern without ending the comment, so we have
16405 * to add an \n when wrapping */
16406 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
16411 S_skip_to_be_ignored_text(pTHX_ RExC_state_t *pRExC_state,
16413 const bool force_to_xmod
16416 /* If the text at the current parse position '*p' is a '(?#...)' comment,
16417 * or if we are under /x or 'force_to_xmod' is TRUE, and the text at '*p'
16418 * is /x whitespace, advance '*p' so that on exit it points to the first
16419 * byte past all such white space and comments */
16421 const bool use_xmod = force_to_xmod || (RExC_flags & RXf_PMf_EXTENDED);
16423 PERL_ARGS_ASSERT_SKIP_TO_BE_IGNORED_TEXT;
16425 assert( ! UTF || UTF8_IS_INVARIANT(**p) || UTF8_IS_START(**p));
16428 if (RExC_end - (*p) >= 3
16430 && *(*p + 1) == '?'
16431 && *(*p + 2) == '#')
16433 while (*(*p) != ')') {
16434 if ((*p) == RExC_end)
16435 FAIL("Sequence (?#... not terminated");
16443 char * new_p = regpatws(pRExC_state, *p,
16444 TRUE); /* means recognize comments */
16459 Advances the parse position by one byte, unless that byte is the beginning
16460 of a '(?#...)' style comment, or is /x whitespace and /x is in effect. In
16461 those two cases, the parse position is advanced beyond all such comments and
16464 This is the UTF, (?#...), and /x friendly way of saying RExC_parse++.
16468 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
16470 PERL_ARGS_ASSERT_NEXTCHAR;
16473 || UTF8_IS_INVARIANT(*RExC_parse)
16474 || UTF8_IS_START(*RExC_parse));
16476 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
16478 skip_to_be_ignored_text(pRExC_state, &RExC_parse,
16479 FALSE /* Don't assume /x */ );
16483 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
16485 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
16486 * space. In pass1, it aligns and increments RExC_size; in pass2,
16489 regnode * const ret = RExC_emit;
16490 GET_RE_DEBUG_FLAGS_DECL;
16492 PERL_ARGS_ASSERT_REGNODE_GUTS;
16494 assert(extra_size >= regarglen[op]);
16497 SIZE_ALIGN(RExC_size);
16498 RExC_size += 1 + extra_size;
16501 if (RExC_emit >= RExC_emit_bound)
16502 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
16503 op, (void*)RExC_emit, (void*)RExC_emit_bound);
16505 NODE_ALIGN_FILL(ret);
16506 #ifndef RE_TRACK_PATTERN_OFFSETS
16507 PERL_UNUSED_ARG(name);
16509 if (RExC_offsets) { /* MJD */
16511 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
16514 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
16515 ? "Overwriting end of array!\n" : "OK",
16516 (UV)(RExC_emit - RExC_emit_start),
16517 (UV)(RExC_parse - RExC_start),
16518 (UV)RExC_offsets[0]));
16519 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
16526 - reg_node - emit a node
16528 STATIC regnode * /* Location. */
16529 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
16531 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
16533 PERL_ARGS_ASSERT_REG_NODE;
16535 assert(regarglen[op] == 0);
16538 regnode *ptr = ret;
16539 FILL_ADVANCE_NODE(ptr, op);
16546 - reganode - emit a node with an argument
16548 STATIC regnode * /* Location. */
16549 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
16551 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
16553 PERL_ARGS_ASSERT_REGANODE;
16555 assert(regarglen[op] == 1);
16558 regnode *ptr = ret;
16559 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
16566 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
16568 /* emit a node with U32 and I32 arguments */
16570 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
16572 PERL_ARGS_ASSERT_REG2LANODE;
16574 assert(regarglen[op] == 2);
16577 regnode *ptr = ret;
16578 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
16585 - reginsert - insert an operator in front of already-emitted operand
16587 * Means relocating the operand.
16590 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
16595 const int offset = regarglen[(U8)op];
16596 const int size = NODE_STEP_REGNODE + offset;
16597 GET_RE_DEBUG_FLAGS_DECL;
16599 PERL_ARGS_ASSERT_REGINSERT;
16600 PERL_UNUSED_CONTEXT;
16601 PERL_UNUSED_ARG(depth);
16602 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
16603 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
16612 if (RExC_open_parens) {
16614 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
16615 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
16616 if ( RExC_open_parens[paren] >= opnd ) {
16617 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
16618 RExC_open_parens[paren] += size;
16620 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
16622 if ( RExC_close_parens[paren] >= opnd ) {
16623 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
16624 RExC_close_parens[paren] += size;
16626 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
16631 while (src > opnd) {
16632 StructCopy(--src, --dst, regnode);
16633 #ifdef RE_TRACK_PATTERN_OFFSETS
16634 if (RExC_offsets) { /* MJD 20010112 */
16636 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
16640 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
16641 ? "Overwriting end of array!\n" : "OK",
16642 (UV)(src - RExC_emit_start),
16643 (UV)(dst - RExC_emit_start),
16644 (UV)RExC_offsets[0]));
16645 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
16646 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
16652 place = opnd; /* Op node, where operand used to be. */
16653 #ifdef RE_TRACK_PATTERN_OFFSETS
16654 if (RExC_offsets) { /* MJD */
16656 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
16660 (UV)(place - RExC_emit_start) > RExC_offsets[0]
16661 ? "Overwriting end of array!\n" : "OK",
16662 (UV)(place - RExC_emit_start),
16663 (UV)(RExC_parse - RExC_start),
16664 (UV)RExC_offsets[0]));
16665 Set_Node_Offset(place, RExC_parse);
16666 Set_Node_Length(place, 1);
16669 src = NEXTOPER(place);
16670 FILL_ADVANCE_NODE(place, op);
16671 Zero(src, offset, regnode);
16675 - regtail - set the next-pointer at the end of a node chain of p to val.
16676 - SEE ALSO: regtail_study
16678 /* TODO: All three parms should be const */
16680 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16681 const regnode *val,U32 depth)
16684 GET_RE_DEBUG_FLAGS_DECL;
16686 PERL_ARGS_ASSERT_REGTAIL;
16688 PERL_UNUSED_ARG(depth);
16694 /* Find last node. */
16697 regnode * const temp = regnext(scan);
16699 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
16700 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16701 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
16702 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
16703 (temp == NULL ? "->" : ""),
16704 (temp == NULL ? PL_reg_name[OP(val)] : "")
16712 if (reg_off_by_arg[OP(scan)]) {
16713 ARG_SET(scan, val - scan);
16716 NEXT_OFF(scan) = val - scan;
16722 - regtail_study - set the next-pointer at the end of a node chain of p to val.
16723 - Look for optimizable sequences at the same time.
16724 - currently only looks for EXACT chains.
16726 This is experimental code. The idea is to use this routine to perform
16727 in place optimizations on branches and groups as they are constructed,
16728 with the long term intention of removing optimization from study_chunk so
16729 that it is purely analytical.
16731 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
16732 to control which is which.
16735 /* TODO: All four parms should be const */
16738 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16739 const regnode *val,U32 depth)
16743 #ifdef EXPERIMENTAL_INPLACESCAN
16746 GET_RE_DEBUG_FLAGS_DECL;
16748 PERL_ARGS_ASSERT_REGTAIL_STUDY;
16754 /* Find last node. */
16758 regnode * const temp = regnext(scan);
16759 #ifdef EXPERIMENTAL_INPLACESCAN
16760 if (PL_regkind[OP(scan)] == EXACT) {
16761 bool unfolded_multi_char; /* Unexamined in this routine */
16762 if (join_exact(pRExC_state, scan, &min,
16763 &unfolded_multi_char, 1, val, depth+1))
16768 switch (OP(scan)) {
16772 case EXACTFA_NO_TRIE:
16778 if( exact == PSEUDO )
16780 else if ( exact != OP(scan) )
16789 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
16790 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16791 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
16792 SvPV_nolen_const(RExC_mysv),
16793 REG_NODE_NUM(scan),
16794 PL_reg_name[exact]);
16801 DEBUG_PARSE_MSG("");
16802 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
16803 PerlIO_printf(Perl_debug_log,
16804 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
16805 SvPV_nolen_const(RExC_mysv),
16806 (IV)REG_NODE_NUM(val),
16810 if (reg_off_by_arg[OP(scan)]) {
16811 ARG_SET(scan, val - scan);
16814 NEXT_OFF(scan) = val - scan;
16822 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
16827 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
16832 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16834 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
16835 if (flags & (1<<bit)) {
16836 if (!set++ && lead)
16837 PerlIO_printf(Perl_debug_log, "%s",lead);
16838 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
16843 PerlIO_printf(Perl_debug_log, "\n");
16845 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16850 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
16856 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16858 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
16859 if (flags & (1<<bit)) {
16860 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
16863 if (!set++ && lead)
16864 PerlIO_printf(Perl_debug_log, "%s",lead);
16865 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
16868 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
16869 if (!set++ && lead) {
16870 PerlIO_printf(Perl_debug_log, "%s",lead);
16873 case REGEX_UNICODE_CHARSET:
16874 PerlIO_printf(Perl_debug_log, "UNICODE");
16876 case REGEX_LOCALE_CHARSET:
16877 PerlIO_printf(Perl_debug_log, "LOCALE");
16879 case REGEX_ASCII_RESTRICTED_CHARSET:
16880 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
16882 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
16883 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
16886 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
16892 PerlIO_printf(Perl_debug_log, "\n");
16894 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16900 Perl_regdump(pTHX_ const regexp *r)
16903 SV * const sv = sv_newmortal();
16904 SV *dsv= sv_newmortal();
16905 RXi_GET_DECL(r,ri);
16906 GET_RE_DEBUG_FLAGS_DECL;
16908 PERL_ARGS_ASSERT_REGDUMP;
16910 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
16912 /* Header fields of interest. */
16913 if (r->anchored_substr) {
16914 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
16915 RE_SV_DUMPLEN(r->anchored_substr), 30);
16916 PerlIO_printf(Perl_debug_log,
16917 "anchored %s%s at %"IVdf" ",
16918 s, RE_SV_TAIL(r->anchored_substr),
16919 (IV)r->anchored_offset);
16920 } else if (r->anchored_utf8) {
16921 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
16922 RE_SV_DUMPLEN(r->anchored_utf8), 30);
16923 PerlIO_printf(Perl_debug_log,
16924 "anchored utf8 %s%s at %"IVdf" ",
16925 s, RE_SV_TAIL(r->anchored_utf8),
16926 (IV)r->anchored_offset);
16928 if (r->float_substr) {
16929 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
16930 RE_SV_DUMPLEN(r->float_substr), 30);
16931 PerlIO_printf(Perl_debug_log,
16932 "floating %s%s at %"IVdf"..%"UVuf" ",
16933 s, RE_SV_TAIL(r->float_substr),
16934 (IV)r->float_min_offset, (UV)r->float_max_offset);
16935 } else if (r->float_utf8) {
16936 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
16937 RE_SV_DUMPLEN(r->float_utf8), 30);
16938 PerlIO_printf(Perl_debug_log,
16939 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
16940 s, RE_SV_TAIL(r->float_utf8),
16941 (IV)r->float_min_offset, (UV)r->float_max_offset);
16943 if (r->check_substr || r->check_utf8)
16944 PerlIO_printf(Perl_debug_log,
16946 (r->check_substr == r->float_substr
16947 && r->check_utf8 == r->float_utf8
16948 ? "(checking floating" : "(checking anchored"));
16949 if (r->intflags & PREGf_NOSCAN)
16950 PerlIO_printf(Perl_debug_log, " noscan");
16951 if (r->extflags & RXf_CHECK_ALL)
16952 PerlIO_printf(Perl_debug_log, " isall");
16953 if (r->check_substr || r->check_utf8)
16954 PerlIO_printf(Perl_debug_log, ") ");
16956 if (ri->regstclass) {
16957 regprop(r, sv, ri->regstclass, NULL, NULL);
16958 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
16960 if (r->intflags & PREGf_ANCH) {
16961 PerlIO_printf(Perl_debug_log, "anchored");
16962 if (r->intflags & PREGf_ANCH_MBOL)
16963 PerlIO_printf(Perl_debug_log, "(MBOL)");
16964 if (r->intflags & PREGf_ANCH_SBOL)
16965 PerlIO_printf(Perl_debug_log, "(SBOL)");
16966 if (r->intflags & PREGf_ANCH_GPOS)
16967 PerlIO_printf(Perl_debug_log, "(GPOS)");
16968 (void)PerlIO_putc(Perl_debug_log, ' ');
16970 if (r->intflags & PREGf_GPOS_SEEN)
16971 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
16972 if (r->intflags & PREGf_SKIP)
16973 PerlIO_printf(Perl_debug_log, "plus ");
16974 if (r->intflags & PREGf_IMPLICIT)
16975 PerlIO_printf(Perl_debug_log, "implicit ");
16976 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
16977 if (r->extflags & RXf_EVAL_SEEN)
16978 PerlIO_printf(Perl_debug_log, "with eval ");
16979 PerlIO_printf(Perl_debug_log, "\n");
16981 regdump_extflags("r->extflags: ",r->extflags);
16982 regdump_intflags("r->intflags: ",r->intflags);
16985 PERL_ARGS_ASSERT_REGDUMP;
16986 PERL_UNUSED_CONTEXT;
16987 PERL_UNUSED_ARG(r);
16988 #endif /* DEBUGGING */
16992 - regprop - printable representation of opcode, with run time support
16996 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
17001 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
17002 static const char * const anyofs[] = {
17003 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
17004 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
17005 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
17006 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
17007 || _CC_CNTRL != 13 || _CC_ASCII != 14 || _CC_VERTSPACE != 15
17008 #error Need to adjust order of anyofs[]
17043 RXi_GET_DECL(prog,progi);
17044 GET_RE_DEBUG_FLAGS_DECL;
17046 PERL_ARGS_ASSERT_REGPROP;
17048 sv_setpvn(sv, "", 0);
17050 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
17051 /* It would be nice to FAIL() here, but this may be called from
17052 regexec.c, and it would be hard to supply pRExC_state. */
17053 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
17054 (int)OP(o), (int)REGNODE_MAX);
17055 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
17057 k = PL_regkind[OP(o)];
17060 sv_catpvs(sv, " ");
17061 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
17062 * is a crude hack but it may be the best for now since
17063 * we have no flag "this EXACTish node was UTF-8"
17065 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
17066 PERL_PV_ESCAPE_UNI_DETECT |
17067 PERL_PV_ESCAPE_NONASCII |
17068 PERL_PV_PRETTY_ELLIPSES |
17069 PERL_PV_PRETTY_LTGT |
17070 PERL_PV_PRETTY_NOCLEAR
17072 } else if (k == TRIE) {
17073 /* print the details of the trie in dumpuntil instead, as
17074 * progi->data isn't available here */
17075 const char op = OP(o);
17076 const U32 n = ARG(o);
17077 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
17078 (reg_ac_data *)progi->data->data[n] :
17080 const reg_trie_data * const trie
17081 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
17083 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
17084 DEBUG_TRIE_COMPILE_r(
17085 Perl_sv_catpvf(aTHX_ sv,
17086 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
17087 (UV)trie->startstate,
17088 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
17089 (UV)trie->wordcount,
17092 (UV)TRIE_CHARCOUNT(trie),
17093 (UV)trie->uniquecharcount
17096 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
17097 sv_catpvs(sv, "[");
17098 (void) put_charclass_bitmap_innards(sv,
17099 (IS_ANYOF_TRIE(op))
17101 : TRIE_BITMAP(trie),
17103 sv_catpvs(sv, "]");
17106 } else if (k == CURLY) {
17107 U32 lo = ARG1(o), hi = ARG2(o);
17108 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
17109 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
17110 Perl_sv_catpvf(aTHX_ sv, "{%u,", (unsigned) lo);
17111 if (hi == REG_INFTY)
17112 sv_catpvs(sv, "INFTY");
17114 Perl_sv_catpvf(aTHX_ sv, "%u", (unsigned) hi);
17115 sv_catpvs(sv, "}");
17117 else if (k == WHILEM && o->flags) /* Ordinal/of */
17118 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
17119 else if (k == REF || k == OPEN || k == CLOSE
17120 || k == GROUPP || OP(o)==ACCEPT)
17122 AV *name_list= NULL;
17123 U32 parno= OP(o) == ACCEPT ? (U32)ARG2L(o) : ARG(o);
17124 Perl_sv_catpvf(aTHX_ sv, "%"UVuf, (UV)parno); /* Parenth number */
17125 if ( RXp_PAREN_NAMES(prog) ) {
17126 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
17127 } else if ( pRExC_state ) {
17128 name_list= RExC_paren_name_list;
17131 if ( k != REF || (OP(o) < NREF)) {
17132 SV **name= av_fetch(name_list, parno, 0 );
17134 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
17137 SV *sv_dat= MUTABLE_SV(progi->data->data[ parno ]);
17138 I32 *nums=(I32*)SvPVX(sv_dat);
17139 SV **name= av_fetch(name_list, nums[0], 0 );
17142 for ( n=0; n<SvIVX(sv_dat); n++ ) {
17143 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
17144 (n ? "," : ""), (IV)nums[n]);
17146 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
17150 if ( k == REF && reginfo) {
17151 U32 n = ARG(o); /* which paren pair */
17152 I32 ln = prog->offs[n].start;
17153 if (prog->lastparen < n || ln == -1)
17154 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
17155 else if (ln == prog->offs[n].end)
17156 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
17158 const char *s = reginfo->strbeg + ln;
17159 Perl_sv_catpvf(aTHX_ sv, ": ");
17160 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
17161 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
17164 } else if (k == GOSUB) {
17165 AV *name_list= NULL;
17166 if ( RXp_PAREN_NAMES(prog) ) {
17167 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
17168 } else if ( pRExC_state ) {
17169 name_list= RExC_paren_name_list;
17172 /* Paren and offset */
17173 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
17175 SV **name= av_fetch(name_list, ARG(o), 0 );
17177 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
17180 else if (k == LOGICAL)
17181 /* 2: embedded, otherwise 1 */
17182 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
17183 else if (k == ANYOF) {
17184 const U8 flags = ANYOF_FLAGS(o);
17186 SV* bitmap_invlist; /* Will hold what the bit map contains */
17189 if (OP(o) == ANYOFL) {
17190 if (flags & ANYOF_LOC_REQ_UTF8) {
17191 sv_catpvs(sv, "{utf8-loc}");
17194 sv_catpvs(sv, "{loc}");
17197 if (flags & ANYOF_LOC_FOLD)
17198 sv_catpvs(sv, "{i}");
17199 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
17200 if (flags & ANYOF_INVERT)
17201 sv_catpvs(sv, "^");
17203 /* output what the standard cp 0-NUM_ANYOF_CODE_POINTS-1 bitmap matches
17205 do_sep = put_charclass_bitmap_innards(sv, ANYOF_BITMAP(o),
17208 /* output any special charclass tests (used entirely under use
17210 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
17212 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
17213 if (ANYOF_POSIXL_TEST(o,i)) {
17214 sv_catpv(sv, anyofs[i]);
17220 if ((flags & (ANYOF_MATCHES_ALL_ABOVE_BITMAP
17221 |ANYOF_HAS_UTF8_NONBITMAP_MATCHES
17222 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES
17226 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
17227 if (flags & ANYOF_INVERT)
17228 /*make sure the invert info is in each */
17229 sv_catpvs(sv, "^");
17232 if (OP(o) == ANYOFD
17233 && (flags & ANYOF_SHARED_d_MATCHES_ALL_NON_UTF8_NON_ASCII_non_d_WARN_SUPER))
17235 sv_catpvs(sv, "{non-utf8-latin1-all}");
17238 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP)
17239 sv_catpvs(sv, "{above_bitmap_all}");
17241 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
17242 SV *lv; /* Set if there is something outside the bit map. */
17243 bool byte_output = FALSE; /* If something has been output */
17244 SV *only_utf8_locale;
17246 /* Get the stuff that wasn't in the bitmap. 'bitmap_invlist'
17247 * is used to guarantee that nothing in the bitmap gets
17249 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
17250 &lv, &only_utf8_locale,
17252 if (lv && lv != &PL_sv_undef) {
17253 char *s = savesvpv(lv);
17254 char * const origs = s;
17256 while (*s && *s != '\n')
17260 const char * const t = ++s;
17262 if (flags & ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES) {
17263 sv_catpvs(sv, "{outside bitmap}");
17266 sv_catpvs(sv, "{utf8}");
17270 sv_catpvs(sv, " ");
17276 /* Truncate very long output */
17277 if (s - origs > 256) {
17278 Perl_sv_catpvf(aTHX_ sv,
17280 (int) (s - origs - 1),
17286 else if (*s == '\t') {
17300 SvREFCNT_dec_NN(lv);
17303 if ((flags & ANYOF_LOC_FOLD)
17304 && only_utf8_locale
17305 && only_utf8_locale != &PL_sv_undef)
17308 int max_entries = 256;
17310 sv_catpvs(sv, "{utf8 locale}");
17311 invlist_iterinit(only_utf8_locale);
17312 while (invlist_iternext(only_utf8_locale,
17314 put_range(sv, start, end, FALSE);
17316 if (max_entries < 0) {
17317 sv_catpvs(sv, "...");
17321 invlist_iterfinish(only_utf8_locale);
17325 SvREFCNT_dec(bitmap_invlist);
17328 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
17330 else if (k == POSIXD || k == NPOSIXD) {
17331 U8 index = FLAGS(o) * 2;
17332 if (index < C_ARRAY_LENGTH(anyofs)) {
17333 if (*anyofs[index] != '[') {
17336 sv_catpv(sv, anyofs[index]);
17337 if (*anyofs[index] != '[') {
17342 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
17345 else if (k == BOUND || k == NBOUND) {
17346 /* Must be synced with order of 'bound_type' in regcomp.h */
17347 const char * const bounds[] = {
17348 "", /* Traditional */
17353 sv_catpv(sv, bounds[FLAGS(o)]);
17355 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
17356 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
17357 else if (OP(o) == SBOL)
17358 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
17360 /* add on the verb argument if there is one */
17361 if ( ( k == VERB || OP(o) == ACCEPT || OP(o) == OPFAIL ) && o->flags) {
17362 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
17363 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
17366 PERL_UNUSED_CONTEXT;
17367 PERL_UNUSED_ARG(sv);
17368 PERL_UNUSED_ARG(o);
17369 PERL_UNUSED_ARG(prog);
17370 PERL_UNUSED_ARG(reginfo);
17371 PERL_UNUSED_ARG(pRExC_state);
17372 #endif /* DEBUGGING */
17378 Perl_re_intuit_string(pTHX_ REGEXP * const r)
17379 { /* Assume that RE_INTUIT is set */
17380 struct regexp *const prog = ReANY(r);
17381 GET_RE_DEBUG_FLAGS_DECL;
17383 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
17384 PERL_UNUSED_CONTEXT;
17388 const char * const s = SvPV_nolen_const(RX_UTF8(r)
17389 ? prog->check_utf8 : prog->check_substr);
17391 if (!PL_colorset) reginitcolors();
17392 PerlIO_printf(Perl_debug_log,
17393 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
17395 RX_UTF8(r) ? "utf8 " : "",
17396 PL_colors[5],PL_colors[0],
17399 (strlen(s) > 60 ? "..." : ""));
17402 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
17403 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
17409 handles refcounting and freeing the perl core regexp structure. When
17410 it is necessary to actually free the structure the first thing it
17411 does is call the 'free' method of the regexp_engine associated to
17412 the regexp, allowing the handling of the void *pprivate; member
17413 first. (This routine is not overridable by extensions, which is why
17414 the extensions free is called first.)
17416 See regdupe and regdupe_internal if you change anything here.
17418 #ifndef PERL_IN_XSUB_RE
17420 Perl_pregfree(pTHX_ REGEXP *r)
17426 Perl_pregfree2(pTHX_ REGEXP *rx)
17428 struct regexp *const r = ReANY(rx);
17429 GET_RE_DEBUG_FLAGS_DECL;
17431 PERL_ARGS_ASSERT_PREGFREE2;
17433 if (r->mother_re) {
17434 ReREFCNT_dec(r->mother_re);
17436 CALLREGFREE_PVT(rx); /* free the private data */
17437 SvREFCNT_dec(RXp_PAREN_NAMES(r));
17438 Safefree(r->xpv_len_u.xpvlenu_pv);
17441 SvREFCNT_dec(r->anchored_substr);
17442 SvREFCNT_dec(r->anchored_utf8);
17443 SvREFCNT_dec(r->float_substr);
17444 SvREFCNT_dec(r->float_utf8);
17445 Safefree(r->substrs);
17447 RX_MATCH_COPY_FREE(rx);
17448 #ifdef PERL_ANY_COW
17449 SvREFCNT_dec(r->saved_copy);
17452 SvREFCNT_dec(r->qr_anoncv);
17453 rx->sv_u.svu_rx = 0;
17458 This is a hacky workaround to the structural issue of match results
17459 being stored in the regexp structure which is in turn stored in
17460 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
17461 could be PL_curpm in multiple contexts, and could require multiple
17462 result sets being associated with the pattern simultaneously, such
17463 as when doing a recursive match with (??{$qr})
17465 The solution is to make a lightweight copy of the regexp structure
17466 when a qr// is returned from the code executed by (??{$qr}) this
17467 lightweight copy doesn't actually own any of its data except for
17468 the starp/end and the actual regexp structure itself.
17474 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
17476 struct regexp *ret;
17477 struct regexp *const r = ReANY(rx);
17478 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
17480 PERL_ARGS_ASSERT_REG_TEMP_COPY;
17483 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
17485 SvOK_off((SV *)ret_x);
17487 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
17488 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
17489 made both spots point to the same regexp body.) */
17490 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
17491 assert(!SvPVX(ret_x));
17492 ret_x->sv_u.svu_rx = temp->sv_any;
17493 temp->sv_any = NULL;
17494 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
17495 SvREFCNT_dec_NN(temp);
17496 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
17497 ing below will not set it. */
17498 SvCUR_set(ret_x, SvCUR(rx));
17501 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
17502 sv_force_normal(sv) is called. */
17504 ret = ReANY(ret_x);
17506 SvFLAGS(ret_x) |= SvUTF8(rx);
17507 /* We share the same string buffer as the original regexp, on which we
17508 hold a reference count, incremented when mother_re is set below.
17509 The string pointer is copied here, being part of the regexp struct.
17511 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
17512 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
17514 const I32 npar = r->nparens+1;
17515 Newx(ret->offs, npar, regexp_paren_pair);
17516 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17519 Newx(ret->substrs, 1, struct reg_substr_data);
17520 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17522 SvREFCNT_inc_void(ret->anchored_substr);
17523 SvREFCNT_inc_void(ret->anchored_utf8);
17524 SvREFCNT_inc_void(ret->float_substr);
17525 SvREFCNT_inc_void(ret->float_utf8);
17527 /* check_substr and check_utf8, if non-NULL, point to either their
17528 anchored or float namesakes, and don't hold a second reference. */
17530 RX_MATCH_COPIED_off(ret_x);
17531 #ifdef PERL_ANY_COW
17532 ret->saved_copy = NULL;
17534 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
17535 SvREFCNT_inc_void(ret->qr_anoncv);
17541 /* regfree_internal()
17543 Free the private data in a regexp. This is overloadable by
17544 extensions. Perl takes care of the regexp structure in pregfree(),
17545 this covers the *pprivate pointer which technically perl doesn't
17546 know about, however of course we have to handle the
17547 regexp_internal structure when no extension is in use.
17549 Note this is called before freeing anything in the regexp
17554 Perl_regfree_internal(pTHX_ REGEXP * const rx)
17556 struct regexp *const r = ReANY(rx);
17557 RXi_GET_DECL(r,ri);
17558 GET_RE_DEBUG_FLAGS_DECL;
17560 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
17566 SV *dsv= sv_newmortal();
17567 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
17568 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
17569 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
17570 PL_colors[4],PL_colors[5],s);
17573 #ifdef RE_TRACK_PATTERN_OFFSETS
17575 Safefree(ri->u.offsets); /* 20010421 MJD */
17577 if (ri->code_blocks) {
17579 for (n = 0; n < ri->num_code_blocks; n++)
17580 SvREFCNT_dec(ri->code_blocks[n].src_regex);
17581 Safefree(ri->code_blocks);
17585 int n = ri->data->count;
17588 /* If you add a ->what type here, update the comment in regcomp.h */
17589 switch (ri->data->what[n]) {
17595 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
17598 Safefree(ri->data->data[n]);
17604 { /* Aho Corasick add-on structure for a trie node.
17605 Used in stclass optimization only */
17607 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
17608 #ifdef USE_ITHREADS
17612 refcount = --aho->refcount;
17615 PerlMemShared_free(aho->states);
17616 PerlMemShared_free(aho->fail);
17617 /* do this last!!!! */
17618 PerlMemShared_free(ri->data->data[n]);
17619 /* we should only ever get called once, so
17620 * assert as much, and also guard the free
17621 * which /might/ happen twice. At the least
17622 * it will make code anlyzers happy and it
17623 * doesn't cost much. - Yves */
17624 assert(ri->regstclass);
17625 if (ri->regstclass) {
17626 PerlMemShared_free(ri->regstclass);
17627 ri->regstclass = 0;
17634 /* trie structure. */
17636 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
17637 #ifdef USE_ITHREADS
17641 refcount = --trie->refcount;
17644 PerlMemShared_free(trie->charmap);
17645 PerlMemShared_free(trie->states);
17646 PerlMemShared_free(trie->trans);
17648 PerlMemShared_free(trie->bitmap);
17650 PerlMemShared_free(trie->jump);
17651 PerlMemShared_free(trie->wordinfo);
17652 /* do this last!!!! */
17653 PerlMemShared_free(ri->data->data[n]);
17658 Perl_croak(aTHX_ "panic: regfree data code '%c'",
17659 ri->data->what[n]);
17662 Safefree(ri->data->what);
17663 Safefree(ri->data);
17669 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
17670 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
17671 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
17674 re_dup - duplicate a regexp.
17676 This routine is expected to clone a given regexp structure. It is only
17677 compiled under USE_ITHREADS.
17679 After all of the core data stored in struct regexp is duplicated
17680 the regexp_engine.dupe method is used to copy any private data
17681 stored in the *pprivate pointer. This allows extensions to handle
17682 any duplication it needs to do.
17684 See pregfree() and regfree_internal() if you change anything here.
17686 #if defined(USE_ITHREADS)
17687 #ifndef PERL_IN_XSUB_RE
17689 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
17693 const struct regexp *r = ReANY(sstr);
17694 struct regexp *ret = ReANY(dstr);
17696 PERL_ARGS_ASSERT_RE_DUP_GUTS;
17698 npar = r->nparens+1;
17699 Newx(ret->offs, npar, regexp_paren_pair);
17700 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17702 if (ret->substrs) {
17703 /* Do it this way to avoid reading from *r after the StructCopy().
17704 That way, if any of the sv_dup_inc()s dislodge *r from the L1
17705 cache, it doesn't matter. */
17706 const bool anchored = r->check_substr
17707 ? r->check_substr == r->anchored_substr
17708 : r->check_utf8 == r->anchored_utf8;
17709 Newx(ret->substrs, 1, struct reg_substr_data);
17710 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17712 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
17713 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
17714 ret->float_substr = sv_dup_inc(ret->float_substr, param);
17715 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
17717 /* check_substr and check_utf8, if non-NULL, point to either their
17718 anchored or float namesakes, and don't hold a second reference. */
17720 if (ret->check_substr) {
17722 assert(r->check_utf8 == r->anchored_utf8);
17723 ret->check_substr = ret->anchored_substr;
17724 ret->check_utf8 = ret->anchored_utf8;
17726 assert(r->check_substr == r->float_substr);
17727 assert(r->check_utf8 == r->float_utf8);
17728 ret->check_substr = ret->float_substr;
17729 ret->check_utf8 = ret->float_utf8;
17731 } else if (ret->check_utf8) {
17733 ret->check_utf8 = ret->anchored_utf8;
17735 ret->check_utf8 = ret->float_utf8;
17740 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
17741 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
17744 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
17746 if (RX_MATCH_COPIED(dstr))
17747 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
17749 ret->subbeg = NULL;
17750 #ifdef PERL_ANY_COW
17751 ret->saved_copy = NULL;
17754 /* Whether mother_re be set or no, we need to copy the string. We
17755 cannot refrain from copying it when the storage points directly to
17756 our mother regexp, because that's
17757 1: a buffer in a different thread
17758 2: something we no longer hold a reference on
17759 so we need to copy it locally. */
17760 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
17761 ret->mother_re = NULL;
17763 #endif /* PERL_IN_XSUB_RE */
17768 This is the internal complement to regdupe() which is used to copy
17769 the structure pointed to by the *pprivate pointer in the regexp.
17770 This is the core version of the extension overridable cloning hook.
17771 The regexp structure being duplicated will be copied by perl prior
17772 to this and will be provided as the regexp *r argument, however
17773 with the /old/ structures pprivate pointer value. Thus this routine
17774 may override any copying normally done by perl.
17776 It returns a pointer to the new regexp_internal structure.
17780 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
17783 struct regexp *const r = ReANY(rx);
17784 regexp_internal *reti;
17786 RXi_GET_DECL(r,ri);
17788 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
17792 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
17793 char, regexp_internal);
17794 Copy(ri->program, reti->program, len+1, regnode);
17796 reti->num_code_blocks = ri->num_code_blocks;
17797 if (ri->code_blocks) {
17799 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
17800 struct reg_code_block);
17801 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
17802 struct reg_code_block);
17803 for (n = 0; n < ri->num_code_blocks; n++)
17804 reti->code_blocks[n].src_regex = (REGEXP*)
17805 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
17808 reti->code_blocks = NULL;
17810 reti->regstclass = NULL;
17813 struct reg_data *d;
17814 const int count = ri->data->count;
17817 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
17818 char, struct reg_data);
17819 Newx(d->what, count, U8);
17822 for (i = 0; i < count; i++) {
17823 d->what[i] = ri->data->what[i];
17824 switch (d->what[i]) {
17825 /* see also regcomp.h and regfree_internal() */
17826 case 'a': /* actually an AV, but the dup function is identical. */
17830 case 'u': /* actually an HV, but the dup function is identical. */
17831 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
17834 /* This is cheating. */
17835 Newx(d->data[i], 1, regnode_ssc);
17836 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
17837 reti->regstclass = (regnode*)d->data[i];
17840 /* Trie stclasses are readonly and can thus be shared
17841 * without duplication. We free the stclass in pregfree
17842 * when the corresponding reg_ac_data struct is freed.
17844 reti->regstclass= ri->regstclass;
17848 ((reg_trie_data*)ri->data->data[i])->refcount++;
17853 d->data[i] = ri->data->data[i];
17856 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
17857 ri->data->what[i]);
17866 reti->name_list_idx = ri->name_list_idx;
17868 #ifdef RE_TRACK_PATTERN_OFFSETS
17869 if (ri->u.offsets) {
17870 Newx(reti->u.offsets, 2*len+1, U32);
17871 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
17874 SetProgLen(reti,len);
17877 return (void*)reti;
17880 #endif /* USE_ITHREADS */
17882 #ifndef PERL_IN_XSUB_RE
17885 - regnext - dig the "next" pointer out of a node
17888 Perl_regnext(pTHX_ regnode *p)
17895 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
17896 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
17897 (int)OP(p), (int)REGNODE_MAX);
17900 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
17909 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
17912 STRLEN l1 = strlen(pat1);
17913 STRLEN l2 = strlen(pat2);
17916 const char *message;
17918 PERL_ARGS_ASSERT_RE_CROAK2;
17924 Copy(pat1, buf, l1 , char);
17925 Copy(pat2, buf + l1, l2 , char);
17926 buf[l1 + l2] = '\n';
17927 buf[l1 + l2 + 1] = '\0';
17928 va_start(args, pat2);
17929 msv = vmess(buf, &args);
17931 message = SvPV_const(msv,l1);
17934 Copy(message, buf, l1 , char);
17935 /* l1-1 to avoid \n */
17936 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
17939 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
17941 #ifndef PERL_IN_XSUB_RE
17943 Perl_save_re_context(pTHX)
17948 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
17951 const REGEXP * const rx = PM_GETRE(PL_curpm);
17953 nparens = RX_NPARENS(rx);
17956 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
17957 * that PL_curpm will be null, but that utf8.pm and the modules it
17958 * loads will only use $1..$3.
17959 * The t/porting/re_context.t test file checks this assumption.
17964 for (i = 1; i <= nparens; i++) {
17965 char digits[TYPE_CHARS(long)];
17966 const STRLEN len = my_snprintf(digits, sizeof(digits),
17968 GV *const *const gvp
17969 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
17972 GV * const gv = *gvp;
17973 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
17983 S_put_code_point(pTHX_ SV *sv, UV c)
17985 PERL_ARGS_ASSERT_PUT_CODE_POINT;
17988 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
17990 else if (isPRINT(c)) {
17991 const char string = (char) c;
17992 if (isBACKSLASHED_PUNCT(c))
17993 sv_catpvs(sv, "\\");
17994 sv_catpvn(sv, &string, 1);
17997 const char * const mnemonic = cntrl_to_mnemonic((char) c);
17999 Perl_sv_catpvf(aTHX_ sv, "%s", mnemonic);
18002 Perl_sv_catpvf(aTHX_ sv, "\\x{%02X}", (U8) c);
18007 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
18010 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
18012 /* Appends to 'sv' a displayable version of the range of code points from
18013 * 'start' to 'end'. It assumes that only ASCII printables are displayable
18014 * as-is (though some of these will be escaped by put_code_point()). */
18016 const unsigned int min_range_count = 3;
18018 assert(start <= end);
18020 PERL_ARGS_ASSERT_PUT_RANGE;
18022 while (start <= end) {
18024 const char * format;
18026 if (end - start < min_range_count) {
18028 /* Individual chars in short ranges */
18029 for (; start <= end; start++) {
18030 put_code_point(sv, start);
18035 /* If permitted by the input options, and there is a possibility that
18036 * this range contains a printable literal, look to see if there is
18038 if (allow_literals && start <= MAX_PRINT_A) {
18040 /* If the range begin isn't an ASCII printable, effectively split
18041 * the range into two parts:
18042 * 1) the portion before the first such printable,
18044 * and output them separately. */
18045 if (! isPRINT_A(start)) {
18046 UV temp_end = start + 1;
18048 /* There is no point looking beyond the final possible
18049 * printable, in MAX_PRINT_A */
18050 UV max = MIN(end, MAX_PRINT_A);
18052 while (temp_end <= max && ! isPRINT_A(temp_end)) {
18056 /* Here, temp_end points to one beyond the first printable if
18057 * found, or to one beyond 'max' if not. If none found, make
18058 * sure that we use the entire range */
18059 if (temp_end > MAX_PRINT_A) {
18060 temp_end = end + 1;
18063 /* Output the first part of the split range, the part that
18064 * doesn't have printables, with no looking for literals
18065 * (otherwise we would infinitely recurse) */
18066 put_range(sv, start, temp_end - 1, FALSE);
18068 /* The 2nd part of the range (if any) starts here. */
18071 /* We continue instead of dropping down because even if the 2nd
18072 * part is non-empty, it could be so short that we want to
18073 * output it specially, as tested for at the top of this loop.
18078 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
18079 * output a sub-range of just the digits or letters, then process
18080 * the remaining portion as usual. */
18081 if (isALPHANUMERIC_A(start)) {
18082 UV mask = (isDIGIT_A(start))
18087 UV temp_end = start + 1;
18089 /* Find the end of the sub-range that includes just the
18090 * characters in the same class as the first character in it */
18091 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
18096 /* For short ranges, don't duplicate the code above to output
18097 * them; just call recursively */
18098 if (temp_end - start < min_range_count) {
18099 put_range(sv, start, temp_end, FALSE);
18101 else { /* Output as a range */
18102 put_code_point(sv, start);
18103 sv_catpvs(sv, "-");
18104 put_code_point(sv, temp_end);
18106 start = temp_end + 1;
18110 /* We output any other printables as individual characters */
18111 if (isPUNCT_A(start) || isSPACE_A(start)) {
18112 while (start <= end && (isPUNCT_A(start)
18113 || isSPACE_A(start)))
18115 put_code_point(sv, start);
18120 } /* End of looking for literals */
18122 /* Here is not to output as a literal. Some control characters have
18123 * mnemonic names. Split off any of those at the beginning and end of
18124 * the range to print mnemonically. It isn't possible for many of
18125 * these to be in a row, so this won't overwhelm with output */
18126 while (isMNEMONIC_CNTRL(start) && start <= end) {
18127 put_code_point(sv, start);
18130 if (start < end && isMNEMONIC_CNTRL(end)) {
18132 /* Here, the final character in the range has a mnemonic name.
18133 * Work backwards from the end to find the final non-mnemonic */
18134 UV temp_end = end - 1;
18135 while (isMNEMONIC_CNTRL(temp_end)) {
18139 /* And separately output the range that doesn't have mnemonics */
18140 put_range(sv, start, temp_end, FALSE);
18142 /* Then output the mnemonic trailing controls */
18143 start = temp_end + 1;
18144 while (start <= end) {
18145 put_code_point(sv, start);
18151 /* As a final resort, output the range or subrange as hex. */
18153 this_end = (end < NUM_ANYOF_CODE_POINTS)
18155 : NUM_ANYOF_CODE_POINTS - 1;
18156 #if NUM_ANYOF_CODE_POINTS > 256
18157 format = (this_end < 256)
18158 ? "\\x{%02"UVXf"}-\\x{%02"UVXf"}"
18159 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
18161 format = "\\x{%02"UVXf"}-\\x{%02"UVXf"}";
18163 GCC_DIAG_IGNORE(-Wformat-nonliteral);
18164 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
18171 S_put_charclass_bitmap_innards(pTHX_ SV *sv, char *bitmap, SV** bitmap_invlist)
18173 /* Appends to 'sv' a displayable version of the innards of the bracketed
18174 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
18175 * output anything, and bitmap_invlist, if not NULL, will point to an
18176 * inversion list of what is in the bit map */
18180 unsigned int punct_count = 0;
18181 SV* invlist = NULL;
18182 SV** invlist_ptr; /* Temporary, in case bitmap_invlist is NULL */
18183 bool allow_literals = TRUE;
18185 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
18187 invlist_ptr = (bitmap_invlist) ? bitmap_invlist : &invlist;
18189 /* Worst case is exactly every-other code point is in the list */
18190 *invlist_ptr = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
18192 /* Convert the bit map to an inversion list, keeping track of how many
18193 * ASCII puncts are set, including an extra amount for the backslashed
18195 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
18196 if (BITMAP_TEST(bitmap, i)) {
18197 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, i);
18198 if (isPUNCT_A(i)) {
18200 if isBACKSLASHED_PUNCT(i) {
18207 /* Nothing to output */
18208 if (_invlist_len(*invlist_ptr) == 0) {
18209 SvREFCNT_dec(invlist);
18213 /* Generally, it is more readable if printable characters are output as
18214 * literals, but if a range (nearly) spans all of them, it's best to output
18215 * it as a single range. This code will use a single range if all but 2
18216 * printables are in it */
18217 invlist_iterinit(*invlist_ptr);
18218 while (invlist_iternext(*invlist_ptr, &start, &end)) {
18220 /* If range starts beyond final printable, it doesn't have any in it */
18221 if (start > MAX_PRINT_A) {
18225 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
18226 * all but two, the range must start and end no later than 2 from
18228 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
18229 if (end > MAX_PRINT_A) {
18235 if (end - start >= MAX_PRINT_A - ' ' - 2) {
18236 allow_literals = FALSE;
18241 invlist_iterfinish(*invlist_ptr);
18243 /* The legibility of the output depends mostly on how many punctuation
18244 * characters are output. There are 32 possible ASCII ones, and some have
18245 * an additional backslash, bringing it to currently 36, so if any more
18246 * than 18 are to be output, we can instead output it as its complement,
18247 * yielding fewer puncts, and making it more legible. But give some weight
18248 * to the fact that outputting it as a complement is less legible than a
18249 * straight output, so don't complement unless we are somewhat over the 18
18251 if (allow_literals && punct_count > 22) {
18252 sv_catpvs(sv, "^");
18254 /* Add everything remaining to the list, so when we invert it just
18255 * below, it will be excluded */
18256 _invlist_union_complement_2nd(*invlist_ptr, PL_InBitmap, invlist_ptr);
18257 _invlist_invert(*invlist_ptr);
18260 /* Here we have figured things out. Output each range */
18261 invlist_iterinit(*invlist_ptr);
18262 while (invlist_iternext(*invlist_ptr, &start, &end)) {
18263 if (start >= NUM_ANYOF_CODE_POINTS) {
18266 put_range(sv, start, end, allow_literals);
18268 invlist_iterfinish(*invlist_ptr);
18273 #define CLEAR_OPTSTART \
18274 if (optstart) STMT_START { \
18275 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
18276 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
18280 #define DUMPUNTIL(b,e) \
18282 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
18284 STATIC const regnode *
18285 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
18286 const regnode *last, const regnode *plast,
18287 SV* sv, I32 indent, U32 depth)
18289 U8 op = PSEUDO; /* Arbitrary non-END op. */
18290 const regnode *next;
18291 const regnode *optstart= NULL;
18293 RXi_GET_DECL(r,ri);
18294 GET_RE_DEBUG_FLAGS_DECL;
18296 PERL_ARGS_ASSERT_DUMPUNTIL;
18298 #ifdef DEBUG_DUMPUNTIL
18299 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
18300 last ? last-start : 0,plast ? plast-start : 0);
18303 if (plast && plast < last)
18306 while (PL_regkind[op] != END && (!last || node < last)) {
18308 /* While that wasn't END last time... */
18311 if (op == CLOSE || op == WHILEM)
18313 next = regnext((regnode *)node);
18316 if (OP(node) == OPTIMIZED) {
18317 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
18324 regprop(r, sv, node, NULL, NULL);
18325 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
18326 (int)(2*indent + 1), "", SvPVX_const(sv));
18328 if (OP(node) != OPTIMIZED) {
18329 if (next == NULL) /* Next ptr. */
18330 PerlIO_printf(Perl_debug_log, " (0)");
18331 else if (PL_regkind[(U8)op] == BRANCH
18332 && PL_regkind[OP(next)] != BRANCH )
18333 PerlIO_printf(Perl_debug_log, " (FAIL)");
18335 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
18336 (void)PerlIO_putc(Perl_debug_log, '\n');
18340 if (PL_regkind[(U8)op] == BRANCHJ) {
18343 const regnode *nnode = (OP(next) == LONGJMP
18344 ? regnext((regnode *)next)
18346 if (last && nnode > last)
18348 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
18351 else if (PL_regkind[(U8)op] == BRANCH) {
18353 DUMPUNTIL(NEXTOPER(node), next);
18355 else if ( PL_regkind[(U8)op] == TRIE ) {
18356 const regnode *this_trie = node;
18357 const char op = OP(node);
18358 const U32 n = ARG(node);
18359 const reg_ac_data * const ac = op>=AHOCORASICK ?
18360 (reg_ac_data *)ri->data->data[n] :
18362 const reg_trie_data * const trie =
18363 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
18365 AV *const trie_words
18366 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
18368 const regnode *nextbranch= NULL;
18371 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
18372 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
18374 PerlIO_printf(Perl_debug_log, "%*s%s ",
18375 (int)(2*(indent+3)), "",
18377 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
18378 SvCUR(*elem_ptr), 60,
18379 PL_colors[0], PL_colors[1],
18381 ? PERL_PV_ESCAPE_UNI
18383 | PERL_PV_PRETTY_ELLIPSES
18384 | PERL_PV_PRETTY_LTGT
18389 U16 dist= trie->jump[word_idx+1];
18390 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
18391 (UV)((dist ? this_trie + dist : next) - start));
18394 nextbranch= this_trie + trie->jump[0];
18395 DUMPUNTIL(this_trie + dist, nextbranch);
18397 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
18398 nextbranch= regnext((regnode *)nextbranch);
18400 PerlIO_printf(Perl_debug_log, "\n");
18403 if (last && next > last)
18408 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
18409 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
18410 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
18412 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
18414 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
18416 else if ( op == PLUS || op == STAR) {
18417 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
18419 else if (PL_regkind[(U8)op] == ANYOF) {
18420 /* arglen 1 + class block */
18421 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
18422 ? ANYOF_POSIXL_SKIP
18424 node = NEXTOPER(node);
18426 else if (PL_regkind[(U8)op] == EXACT) {
18427 /* Literal string, where present. */
18428 node += NODE_SZ_STR(node) - 1;
18429 node = NEXTOPER(node);
18432 node = NEXTOPER(node);
18433 node += regarglen[(U8)op];
18435 if (op == CURLYX || op == OPEN)
18439 #ifdef DEBUG_DUMPUNTIL
18440 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
18445 #endif /* DEBUGGING */
18448 * ex: set ts=8 sts=4 sw=4 et: