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_static.c"
90 #include "inline_invlist.c"
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)
219 #define RExC_flags (pRExC_state->flags)
220 #define RExC_pm_flags (pRExC_state->pm_flags)
221 #define RExC_precomp (pRExC_state->precomp)
222 #define RExC_rx_sv (pRExC_state->rx_sv)
223 #define RExC_rx (pRExC_state->rx)
224 #define RExC_rxi (pRExC_state->rxi)
225 #define RExC_start (pRExC_state->start)
226 #define RExC_end (pRExC_state->end)
227 #define RExC_parse (pRExC_state->parse)
228 #define RExC_whilem_seen (pRExC_state->whilem_seen)
229 #ifdef RE_TRACK_PATTERN_OFFSETS
230 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
233 #define RExC_emit (pRExC_state->emit)
234 #define RExC_emit_dummy (pRExC_state->emit_dummy)
235 #define RExC_emit_start (pRExC_state->emit_start)
236 #define RExC_emit_bound (pRExC_state->emit_bound)
237 #define RExC_sawback (pRExC_state->sawback)
238 #define RExC_seen (pRExC_state->seen)
239 #define RExC_size (pRExC_state->size)
240 #define RExC_maxlen (pRExC_state->maxlen)
241 #define RExC_npar (pRExC_state->npar)
242 #define RExC_nestroot (pRExC_state->nestroot)
243 #define RExC_extralen (pRExC_state->extralen)
244 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
245 #define RExC_utf8 (pRExC_state->utf8)
246 #define RExC_uni_semantics (pRExC_state->uni_semantics)
247 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
248 #define RExC_open_parens (pRExC_state->open_parens)
249 #define RExC_close_parens (pRExC_state->close_parens)
250 #define RExC_opend (pRExC_state->opend)
251 #define RExC_paren_names (pRExC_state->paren_names)
252 #define RExC_recurse (pRExC_state->recurse)
253 #define RExC_recurse_count (pRExC_state->recurse_count)
254 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
255 #define RExC_study_chunk_recursed_bytes \
256 (pRExC_state->study_chunk_recursed_bytes)
257 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
258 #define RExC_contains_locale (pRExC_state->contains_locale)
259 #define RExC_contains_i (pRExC_state->contains_i)
260 #define RExC_override_recoding (pRExC_state->override_recoding)
262 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
264 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
265 #define RExC_frame_head (pRExC_state->frame_head)
266 #define RExC_frame_last (pRExC_state->frame_last)
267 #define RExC_frame_count (pRExC_state->frame_count)
268 #define RExC_strict (pRExC_state->strict)
270 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
271 * a flag to disable back-off on the fixed/floating substrings - if it's
272 * a high complexity pattern we assume the benefit of avoiding a full match
273 * is worth the cost of checking for the substrings even if they rarely help.
275 #define RExC_naughty (pRExC_state->naughty)
276 #define TOO_NAUGHTY (10)
277 #define MARK_NAUGHTY(add) \
278 if (RExC_naughty < TOO_NAUGHTY) \
279 RExC_naughty += (add)
280 #define MARK_NAUGHTY_EXP(exp, add) \
281 if (RExC_naughty < TOO_NAUGHTY) \
282 RExC_naughty += RExC_naughty / (exp) + (add)
284 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
285 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
286 ((*s) == '{' && regcurly(s)))
289 * Flags to be passed up and down.
291 #define WORST 0 /* Worst case. */
292 #define HASWIDTH 0x01 /* Known to match non-null strings. */
294 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
295 * character. (There needs to be a case: in the switch statement in regexec.c
296 * for any node marked SIMPLE.) Note that this is not the same thing as
299 #define SPSTART 0x04 /* Starts with * or + */
300 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
301 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
302 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
304 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
306 /* whether trie related optimizations are enabled */
307 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
308 #define TRIE_STUDY_OPT
309 #define FULL_TRIE_STUDY
315 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
316 #define PBITVAL(paren) (1 << ((paren) & 7))
317 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
318 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
319 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
321 #define REQUIRE_UTF8 STMT_START { \
323 *flagp = RESTART_UTF8; \
328 /* This converts the named class defined in regcomp.h to its equivalent class
329 * number defined in handy.h. */
330 #define namedclass_to_classnum(class) ((int) ((class) / 2))
331 #define classnum_to_namedclass(classnum) ((classnum) * 2)
333 #define _invlist_union_complement_2nd(a, b, output) \
334 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
335 #define _invlist_intersection_complement_2nd(a, b, output) \
336 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
338 /* About scan_data_t.
340 During optimisation we recurse through the regexp program performing
341 various inplace (keyhole style) optimisations. In addition study_chunk
342 and scan_commit populate this data structure with information about
343 what strings MUST appear in the pattern. We look for the longest
344 string that must appear at a fixed location, and we look for the
345 longest string that may appear at a floating location. So for instance
350 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
351 strings (because they follow a .* construct). study_chunk will identify
352 both FOO and BAR as being the longest fixed and floating strings respectively.
354 The strings can be composites, for instance
358 will result in a composite fixed substring 'foo'.
360 For each string some basic information is maintained:
362 - offset or min_offset
363 This is the position the string must appear at, or not before.
364 It also implicitly (when combined with minlenp) tells us how many
365 characters must match before the string we are searching for.
366 Likewise when combined with minlenp and the length of the string it
367 tells us how many characters must appear after the string we have
371 Only used for floating strings. This is the rightmost point that
372 the string can appear at. If set to SSize_t_MAX it indicates that the
373 string can occur infinitely far to the right.
376 A pointer to the minimum number of characters of the pattern that the
377 string was found inside. This is important as in the case of positive
378 lookahead or positive lookbehind we can have multiple patterns
383 The minimum length of the pattern overall is 3, the minimum length
384 of the lookahead part is 3, but the minimum length of the part that
385 will actually match is 1. So 'FOO's minimum length is 3, but the
386 minimum length for the F is 1. This is important as the minimum length
387 is used to determine offsets in front of and behind the string being
388 looked for. Since strings can be composites this is the length of the
389 pattern at the time it was committed with a scan_commit. Note that
390 the length is calculated by study_chunk, so that the minimum lengths
391 are not known until the full pattern has been compiled, thus the
392 pointer to the value.
396 In the case of lookbehind the string being searched for can be
397 offset past the start point of the final matching string.
398 If this value was just blithely removed from the min_offset it would
399 invalidate some of the calculations for how many chars must match
400 before or after (as they are derived from min_offset and minlen and
401 the length of the string being searched for).
402 When the final pattern is compiled and the data is moved from the
403 scan_data_t structure into the regexp structure the information
404 about lookbehind is factored in, with the information that would
405 have been lost precalculated in the end_shift field for the
408 The fields pos_min and pos_delta are used to store the minimum offset
409 and the delta to the maximum offset at the current point in the pattern.
413 typedef struct scan_data_t {
414 /*I32 len_min; unused */
415 /*I32 len_delta; unused */
419 SSize_t last_end; /* min value, <0 unless valid. */
420 SSize_t last_start_min;
421 SSize_t last_start_max;
422 SV **longest; /* Either &l_fixed, or &l_float. */
423 SV *longest_fixed; /* longest fixed string found in pattern */
424 SSize_t offset_fixed; /* offset where it starts */
425 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
426 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
427 SV *longest_float; /* longest floating string found in pattern */
428 SSize_t offset_float_min; /* earliest point in string it can appear */
429 SSize_t offset_float_max; /* latest point in string it can appear */
430 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
431 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
434 SSize_t *last_closep;
435 regnode_ssc *start_class;
439 * Forward declarations for pregcomp()'s friends.
442 static const scan_data_t zero_scan_data =
443 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
445 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
446 #define SF_BEFORE_SEOL 0x0001
447 #define SF_BEFORE_MEOL 0x0002
448 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
449 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
451 #define SF_FIX_SHIFT_EOL (+2)
452 #define SF_FL_SHIFT_EOL (+4)
454 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
455 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
457 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
458 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
459 #define SF_IS_INF 0x0040
460 #define SF_HAS_PAR 0x0080
461 #define SF_IN_PAR 0x0100
462 #define SF_HAS_EVAL 0x0200
463 #define SCF_DO_SUBSTR 0x0400
464 #define SCF_DO_STCLASS_AND 0x0800
465 #define SCF_DO_STCLASS_OR 0x1000
466 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
467 #define SCF_WHILEM_VISITED_POS 0x2000
469 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
470 #define SCF_SEEN_ACCEPT 0x8000
471 #define SCF_TRIE_DOING_RESTUDY 0x10000
472 #define SCF_IN_DEFINE 0x20000
477 #define UTF cBOOL(RExC_utf8)
479 /* The enums for all these are ordered so things work out correctly */
480 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
481 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
482 == REGEX_DEPENDS_CHARSET)
483 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
484 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
485 >= REGEX_UNICODE_CHARSET)
486 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
487 == REGEX_ASCII_RESTRICTED_CHARSET)
488 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
489 >= REGEX_ASCII_RESTRICTED_CHARSET)
490 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
491 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
493 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
495 /* For programs that want to be strictly Unicode compatible by dying if any
496 * attempt is made to match a non-Unicode code point against a Unicode
498 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
500 #define OOB_NAMEDCLASS -1
502 /* There is no code point that is out-of-bounds, so this is problematic. But
503 * its only current use is to initialize a variable that is always set before
505 #define OOB_UNICODE 0xDEADBEEF
507 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
508 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
511 /* length of regex to show in messages that don't mark a position within */
512 #define RegexLengthToShowInErrorMessages 127
515 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
516 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
517 * op/pragma/warn/regcomp.
519 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
520 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
522 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
523 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
525 #define REPORT_LOCATION_ARGS(offset) \
526 UTF8fARG(UTF, offset, RExC_precomp), \
527 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
529 /* Used to point after bad bytes for an error message, but avoid skipping
530 * past a nul byte. */
531 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
534 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
535 * arg. Show regex, up to a maximum length. If it's too long, chop and add
538 #define _FAIL(code) STMT_START { \
539 const char *ellipses = ""; \
540 IV len = RExC_end - RExC_precomp; \
543 SAVEFREESV(RExC_rx_sv); \
544 if (len > RegexLengthToShowInErrorMessages) { \
545 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
546 len = RegexLengthToShowInErrorMessages - 10; \
552 #define FAIL(msg) _FAIL( \
553 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
554 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
556 #define FAIL2(msg,arg) _FAIL( \
557 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
558 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
561 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
563 #define Simple_vFAIL(m) STMT_START { \
565 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
566 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
567 m, REPORT_LOCATION_ARGS(offset)); \
571 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
573 #define vFAIL(m) STMT_START { \
575 SAVEFREESV(RExC_rx_sv); \
580 * Like Simple_vFAIL(), but accepts two arguments.
582 #define Simple_vFAIL2(m,a1) STMT_START { \
583 const IV offset = RExC_parse - RExC_precomp; \
584 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
585 REPORT_LOCATION_ARGS(offset)); \
589 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
591 #define vFAIL2(m,a1) STMT_START { \
593 SAVEFREESV(RExC_rx_sv); \
594 Simple_vFAIL2(m, a1); \
599 * Like Simple_vFAIL(), but accepts three arguments.
601 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
602 const IV offset = RExC_parse - RExC_precomp; \
603 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
604 REPORT_LOCATION_ARGS(offset)); \
608 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
610 #define vFAIL3(m,a1,a2) STMT_START { \
612 SAVEFREESV(RExC_rx_sv); \
613 Simple_vFAIL3(m, a1, a2); \
617 * Like Simple_vFAIL(), but accepts four arguments.
619 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
620 const IV offset = RExC_parse - RExC_precomp; \
621 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
622 REPORT_LOCATION_ARGS(offset)); \
625 #define vFAIL4(m,a1,a2,a3) STMT_START { \
627 SAVEFREESV(RExC_rx_sv); \
628 Simple_vFAIL4(m, a1, a2, a3); \
631 /* A specialized version of vFAIL2 that works with UTF8f */
632 #define vFAIL2utf8f(m, a1) STMT_START { \
633 const IV offset = RExC_parse - RExC_precomp; \
635 SAVEFREESV(RExC_rx_sv); \
636 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
637 REPORT_LOCATION_ARGS(offset)); \
640 /* These have asserts in them because of [perl #122671] Many warnings in
641 * regcomp.c can occur twice. If they get output in pass1 and later in that
642 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
643 * would get output again. So they should be output in pass2, and these
644 * asserts make sure new warnings follow that paradigm. */
646 /* m is not necessarily a "literal string", in this macro */
647 #define reg_warn_non_literal_string(loc, m) STMT_START { \
648 const IV offset = loc - RExC_precomp; \
649 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
650 m, REPORT_LOCATION_ARGS(offset)); \
653 #define ckWARNreg(loc,m) STMT_START { \
654 const IV offset = loc - RExC_precomp; \
655 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
656 REPORT_LOCATION_ARGS(offset)); \
659 #define vWARN(loc, m) STMT_START { \
660 const IV offset = loc - RExC_precomp; \
661 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
662 REPORT_LOCATION_ARGS(offset)); \
665 #define vWARN_dep(loc, m) STMT_START { \
666 const IV offset = loc - RExC_precomp; \
667 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
668 REPORT_LOCATION_ARGS(offset)); \
671 #define ckWARNdep(loc,m) STMT_START { \
672 const IV offset = loc - RExC_precomp; \
673 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
675 REPORT_LOCATION_ARGS(offset)); \
678 #define ckWARNregdep(loc,m) STMT_START { \
679 const IV offset = loc - RExC_precomp; \
680 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
682 REPORT_LOCATION_ARGS(offset)); \
685 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
686 const IV offset = loc - RExC_precomp; \
687 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
689 a1, REPORT_LOCATION_ARGS(offset)); \
692 #define ckWARN2reg(loc, m, a1) STMT_START { \
693 const IV offset = loc - RExC_precomp; \
694 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
695 a1, REPORT_LOCATION_ARGS(offset)); \
698 #define vWARN3(loc, m, a1, a2) STMT_START { \
699 const IV offset = loc - RExC_precomp; \
700 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
701 a1, a2, REPORT_LOCATION_ARGS(offset)); \
704 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
705 const IV offset = loc - RExC_precomp; \
706 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
707 a1, a2, REPORT_LOCATION_ARGS(offset)); \
710 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
711 const IV offset = loc - RExC_precomp; \
712 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
713 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
716 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
717 const IV offset = loc - RExC_precomp; \
718 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
719 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
722 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
723 const IV offset = loc - RExC_precomp; \
724 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
725 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
728 /* Macros for recording node offsets. 20001227 mjd@plover.com
729 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
730 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
731 * Element 0 holds the number n.
732 * Position is 1 indexed.
734 #ifndef RE_TRACK_PATTERN_OFFSETS
735 #define Set_Node_Offset_To_R(node,byte)
736 #define Set_Node_Offset(node,byte)
737 #define Set_Cur_Node_Offset
738 #define Set_Node_Length_To_R(node,len)
739 #define Set_Node_Length(node,len)
740 #define Set_Node_Cur_Length(node,start)
741 #define Node_Offset(n)
742 #define Node_Length(n)
743 #define Set_Node_Offset_Length(node,offset,len)
744 #define ProgLen(ri) ri->u.proglen
745 #define SetProgLen(ri,x) ri->u.proglen = x
747 #define ProgLen(ri) ri->u.offsets[0]
748 #define SetProgLen(ri,x) ri->u.offsets[0] = x
749 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
751 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
752 __LINE__, (int)(node), (int)(byte))); \
754 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
757 RExC_offsets[2*(node)-1] = (byte); \
762 #define Set_Node_Offset(node,byte) \
763 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
764 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
766 #define Set_Node_Length_To_R(node,len) STMT_START { \
768 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
769 __LINE__, (int)(node), (int)(len))); \
771 Perl_croak(aTHX_ "value of node is %d in Length macro", \
774 RExC_offsets[2*(node)] = (len); \
779 #define Set_Node_Length(node,len) \
780 Set_Node_Length_To_R((node)-RExC_emit_start, len)
781 #define Set_Node_Cur_Length(node, start) \
782 Set_Node_Length(node, RExC_parse - start)
784 /* Get offsets and lengths */
785 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
786 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
788 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
789 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
790 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
794 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
795 #define EXPERIMENTAL_INPLACESCAN
796 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
798 #define DEBUG_RExC_seen() \
799 DEBUG_OPTIMISE_MORE_r({ \
800 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
802 if (RExC_seen & REG_ZERO_LEN_SEEN) \
803 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
805 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
806 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
808 if (RExC_seen & REG_GPOS_SEEN) \
809 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
811 if (RExC_seen & REG_CANY_SEEN) \
812 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
814 if (RExC_seen & REG_RECURSE_SEEN) \
815 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
817 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
818 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
820 if (RExC_seen & REG_VERBARG_SEEN) \
821 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
823 if (RExC_seen & REG_CUTGROUP_SEEN) \
824 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
826 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
827 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
829 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
830 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
832 if (RExC_seen & REG_GOSTART_SEEN) \
833 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
835 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
836 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
838 PerlIO_printf(Perl_debug_log,"\n"); \
841 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
842 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
844 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
846 PerlIO_printf(Perl_debug_log, "%s", open_str); \
847 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
848 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
849 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
850 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
851 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
852 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
853 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
854 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
855 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
856 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
857 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
858 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
859 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
860 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
861 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
862 PerlIO_printf(Perl_debug_log, "%s", close_str); \
866 #define DEBUG_STUDYDATA(str,data,depth) \
867 DEBUG_OPTIMISE_MORE_r(if(data){ \
868 PerlIO_printf(Perl_debug_log, \
869 "%*s" str "Pos:%"IVdf"/%"IVdf \
871 (int)(depth)*2, "", \
872 (IV)((data)->pos_min), \
873 (IV)((data)->pos_delta), \
874 (UV)((data)->flags) \
876 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
877 PerlIO_printf(Perl_debug_log, \
878 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
879 (IV)((data)->whilem_c), \
880 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
881 is_inf ? "INF " : "" \
883 if ((data)->last_found) \
884 PerlIO_printf(Perl_debug_log, \
885 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
886 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
887 SvPVX_const((data)->last_found), \
888 (IV)((data)->last_end), \
889 (IV)((data)->last_start_min), \
890 (IV)((data)->last_start_max), \
891 ((data)->longest && \
892 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
893 SvPVX_const((data)->longest_fixed), \
894 (IV)((data)->offset_fixed), \
895 ((data)->longest && \
896 (data)->longest==&((data)->longest_float)) ? "*" : "", \
897 SvPVX_const((data)->longest_float), \
898 (IV)((data)->offset_float_min), \
899 (IV)((data)->offset_float_max) \
901 PerlIO_printf(Perl_debug_log,"\n"); \
904 /* is c a control character for which we have a mnemonic? */
905 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
908 S_cntrl_to_mnemonic(const U8 c)
910 /* Returns the mnemonic string that represents character 'c', if one
911 * exists; NULL otherwise. The only ones that exist for the purposes of
912 * this routine are a few control characters */
915 case '\a': return "\\a";
916 case '\b': return "\\b";
917 case ESC_NATIVE: return "\\e";
918 case '\f': return "\\f";
919 case '\n': return "\\n";
920 case '\r': return "\\r";
921 case '\t': return "\\t";
927 /* Mark that we cannot extend a found fixed substring at this point.
928 Update the longest found anchored substring and the longest found
929 floating substrings if needed. */
932 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
933 SSize_t *minlenp, int is_inf)
935 const STRLEN l = CHR_SVLEN(data->last_found);
936 const STRLEN old_l = CHR_SVLEN(*data->longest);
937 GET_RE_DEBUG_FLAGS_DECL;
939 PERL_ARGS_ASSERT_SCAN_COMMIT;
941 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
942 SvSetMagicSV(*data->longest, data->last_found);
943 if (*data->longest == data->longest_fixed) {
944 data->offset_fixed = l ? data->last_start_min : data->pos_min;
945 if (data->flags & SF_BEFORE_EOL)
947 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
949 data->flags &= ~SF_FIX_BEFORE_EOL;
950 data->minlen_fixed=minlenp;
951 data->lookbehind_fixed=0;
953 else { /* *data->longest == data->longest_float */
954 data->offset_float_min = l ? data->last_start_min : data->pos_min;
955 data->offset_float_max = (l
956 ? data->last_start_max
957 : (data->pos_delta > SSize_t_MAX - data->pos_min
959 : data->pos_min + data->pos_delta));
961 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
962 data->offset_float_max = SSize_t_MAX;
963 if (data->flags & SF_BEFORE_EOL)
965 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
967 data->flags &= ~SF_FL_BEFORE_EOL;
968 data->minlen_float=minlenp;
969 data->lookbehind_float=0;
972 SvCUR_set(data->last_found, 0);
974 SV * const sv = data->last_found;
975 if (SvUTF8(sv) && SvMAGICAL(sv)) {
976 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
982 data->flags &= ~SF_BEFORE_EOL;
983 DEBUG_STUDYDATA("commit: ",data,0);
986 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
987 * list that describes which code points it matches */
990 S_ssc_anything(pTHX_ regnode_ssc *ssc)
992 /* Set the SSC 'ssc' to match an empty string or any code point */
994 PERL_ARGS_ASSERT_SSC_ANYTHING;
996 assert(is_ANYOF_SYNTHETIC(ssc));
998 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
999 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1000 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1004 S_ssc_is_anything(const regnode_ssc *ssc)
1006 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1007 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1008 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1009 * in any way, so there's no point in using it */
1014 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1016 assert(is_ANYOF_SYNTHETIC(ssc));
1018 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1022 /* See if the list consists solely of the range 0 - Infinity */
1023 invlist_iterinit(ssc->invlist);
1024 ret = invlist_iternext(ssc->invlist, &start, &end)
1028 invlist_iterfinish(ssc->invlist);
1034 /* If e.g., both \w and \W are set, matches everything */
1035 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1037 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1038 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1048 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1050 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1051 * string, any code point, or any posix class under locale */
1053 PERL_ARGS_ASSERT_SSC_INIT;
1055 Zero(ssc, 1, regnode_ssc);
1056 set_ANYOF_SYNTHETIC(ssc);
1057 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1060 /* If any portion of the regex is to operate under locale rules that aren't
1061 * fully known at compile time, initialization includes it. The reason
1062 * this isn't done for all regexes is that the optimizer was written under
1063 * the assumption that locale was all-or-nothing. Given the complexity and
1064 * lack of documentation in the optimizer, and that there are inadequate
1065 * test cases for locale, many parts of it may not work properly, it is
1066 * safest to avoid locale unless necessary. */
1067 if (RExC_contains_locale) {
1068 ANYOF_POSIXL_SETALL(ssc);
1071 ANYOF_POSIXL_ZERO(ssc);
1076 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1077 const regnode_ssc *ssc)
1079 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1080 * to the list of code points matched, and locale posix classes; hence does
1081 * not check its flags) */
1086 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1088 assert(is_ANYOF_SYNTHETIC(ssc));
1090 invlist_iterinit(ssc->invlist);
1091 ret = invlist_iternext(ssc->invlist, &start, &end)
1095 invlist_iterfinish(ssc->invlist);
1101 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1109 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1110 const regnode_charclass* const node)
1112 /* Returns a mortal inversion list defining which code points are matched
1113 * by 'node', which is of type ANYOF. Handles complementing the result if
1114 * appropriate. If some code points aren't knowable at this time, the
1115 * returned list must, and will, contain every code point that is a
1118 SV* invlist = sv_2mortal(_new_invlist(0));
1119 SV* only_utf8_locale_invlist = NULL;
1121 const U32 n = ARG(node);
1122 bool new_node_has_latin1 = FALSE;
1124 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1126 /* Look at the data structure created by S_set_ANYOF_arg() */
1127 if (n != ANYOF_ONLY_HAS_BITMAP) {
1128 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1129 AV * const av = MUTABLE_AV(SvRV(rv));
1130 SV **const ary = AvARRAY(av);
1131 assert(RExC_rxi->data->what[n] == 's');
1133 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1134 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1136 else if (ary[0] && ary[0] != &PL_sv_undef) {
1138 /* Here, no compile-time swash, and there are things that won't be
1139 * known until runtime -- we have to assume it could be anything */
1140 return _add_range_to_invlist(invlist, 0, UV_MAX);
1142 else if (ary[3] && ary[3] != &PL_sv_undef) {
1144 /* Here no compile-time swash, and no run-time only data. Use the
1145 * node's inversion list */
1146 invlist = sv_2mortal(invlist_clone(ary[3]));
1149 /* Get the code points valid only under UTF-8 locales */
1150 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1151 && ary[2] && ary[2] != &PL_sv_undef)
1153 only_utf8_locale_invlist = ary[2];
1157 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1158 * code points, and an inversion list for the others, but if there are code
1159 * points that should match only conditionally on the target string being
1160 * UTF-8, those are placed in the inversion list, and not the bitmap.
1161 * Since there are circumstances under which they could match, they are
1162 * included in the SSC. But if the ANYOF node is to be inverted, we have
1163 * to exclude them here, so that when we invert below, the end result
1164 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1165 * have to do this here before we add the unconditionally matched code
1167 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1168 _invlist_intersection_complement_2nd(invlist,
1173 /* Add in the points from the bit map */
1174 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1175 if (ANYOF_BITMAP_TEST(node, i)) {
1176 invlist = add_cp_to_invlist(invlist, i);
1177 new_node_has_latin1 = TRUE;
1181 /* If this can match all upper Latin1 code points, have to add them
1183 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
1184 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1187 /* Similarly for these */
1188 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1189 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1192 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1193 _invlist_invert(invlist);
1195 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1197 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1198 * locale. We can skip this if there are no 0-255 at all. */
1199 _invlist_union(invlist, PL_Latin1, &invlist);
1202 /* Similarly add the UTF-8 locale possible matches. These have to be
1203 * deferred until after the non-UTF-8 locale ones are taken care of just
1204 * above, or it leads to wrong results under ANYOF_INVERT */
1205 if (only_utf8_locale_invlist) {
1206 _invlist_union_maybe_complement_2nd(invlist,
1207 only_utf8_locale_invlist,
1208 ANYOF_FLAGS(node) & ANYOF_INVERT,
1215 /* These two functions currently do the exact same thing */
1216 #define ssc_init_zero ssc_init
1218 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1219 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1221 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1222 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1223 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1226 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1227 const regnode_charclass *and_with)
1229 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1230 * another SSC or a regular ANYOF class. Can create false positives. */
1235 PERL_ARGS_ASSERT_SSC_AND;
1237 assert(is_ANYOF_SYNTHETIC(ssc));
1239 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1240 * the code point inversion list and just the relevant flags */
1241 if (is_ANYOF_SYNTHETIC(and_with)) {
1242 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1243 anded_flags = ANYOF_FLAGS(and_with);
1245 /* XXX This is a kludge around what appears to be deficiencies in the
1246 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1247 * there are paths through the optimizer where it doesn't get weeded
1248 * out when it should. And if we don't make some extra provision for
1249 * it like the code just below, it doesn't get added when it should.
1250 * This solution is to add it only when AND'ing, which is here, and
1251 * only when what is being AND'ed is the pristine, original node
1252 * matching anything. Thus it is like adding it to ssc_anything() but
1253 * only when the result is to be AND'ed. Probably the same solution
1254 * could be adopted for the same problem we have with /l matching,
1255 * which is solved differently in S_ssc_init(), and that would lead to
1256 * fewer false positives than that solution has. But if this solution
1257 * creates bugs, the consequences are only that a warning isn't raised
1258 * that should be; while the consequences for having /l bugs is
1259 * incorrect matches */
1260 if (ssc_is_anything((regnode_ssc *)and_with)) {
1261 anded_flags |= ANYOF_WARN_SUPER;
1265 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1266 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1269 ANYOF_FLAGS(ssc) &= anded_flags;
1271 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1272 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1273 * 'and_with' may be inverted. When not inverted, we have the situation of
1275 * (C1 | P1) & (C2 | P2)
1276 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1277 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1278 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1279 * <= ((C1 & C2) | P1 | P2)
1280 * Alternatively, the last few steps could be:
1281 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1282 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1283 * <= (C1 | C2 | (P1 & P2))
1284 * We favor the second approach if either P1 or P2 is non-empty. This is
1285 * because these components are a barrier to doing optimizations, as what
1286 * they match cannot be known until the moment of matching as they are
1287 * dependent on the current locale, 'AND"ing them likely will reduce or
1289 * But we can do better if we know that C1,P1 are in their initial state (a
1290 * frequent occurrence), each matching everything:
1291 * (<everything>) & (C2 | P2) = C2 | P2
1292 * Similarly, if C2,P2 are in their initial state (again a frequent
1293 * occurrence), the result is a no-op
1294 * (C1 | P1) & (<everything>) = C1 | P1
1297 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1298 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1299 * <= (C1 & ~C2) | (P1 & ~P2)
1302 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1303 && ! is_ANYOF_SYNTHETIC(and_with))
1307 ssc_intersection(ssc,
1309 FALSE /* Has already been inverted */
1312 /* If either P1 or P2 is empty, the intersection will be also; can skip
1314 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1315 ANYOF_POSIXL_ZERO(ssc);
1317 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1319 /* Note that the Posix class component P from 'and_with' actually
1321 * P = Pa | Pb | ... | Pn
1322 * where each component is one posix class, such as in [\w\s].
1324 * ~P = ~(Pa | Pb | ... | Pn)
1325 * = ~Pa & ~Pb & ... & ~Pn
1326 * <= ~Pa | ~Pb | ... | ~Pn
1327 * The last is something we can easily calculate, but unfortunately
1328 * is likely to have many false positives. We could do better
1329 * in some (but certainly not all) instances if two classes in
1330 * P have known relationships. For example
1331 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1333 * :lower: & :print: = :lower:
1334 * And similarly for classes that must be disjoint. For example,
1335 * since \s and \w can have no elements in common based on rules in
1336 * the POSIX standard,
1337 * \w & ^\S = nothing
1338 * Unfortunately, some vendor locales do not meet the Posix
1339 * standard, in particular almost everything by Microsoft.
1340 * The loop below just changes e.g., \w into \W and vice versa */
1342 regnode_charclass_posixl temp;
1343 int add = 1; /* To calculate the index of the complement */
1345 ANYOF_POSIXL_ZERO(&temp);
1346 for (i = 0; i < ANYOF_MAX; i++) {
1348 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1349 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1351 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1352 ANYOF_POSIXL_SET(&temp, i + add);
1354 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1356 ANYOF_POSIXL_AND(&temp, ssc);
1358 } /* else ssc already has no posixes */
1359 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1360 in its initial state */
1361 else if (! is_ANYOF_SYNTHETIC(and_with)
1362 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1364 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1365 * copy it over 'ssc' */
1366 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1367 if (is_ANYOF_SYNTHETIC(and_with)) {
1368 StructCopy(and_with, ssc, regnode_ssc);
1371 ssc->invlist = anded_cp_list;
1372 ANYOF_POSIXL_ZERO(ssc);
1373 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1374 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1378 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1379 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1381 /* One or the other of P1, P2 is non-empty. */
1382 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1383 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1385 ssc_union(ssc, anded_cp_list, FALSE);
1387 else { /* P1 = P2 = empty */
1388 ssc_intersection(ssc, anded_cp_list, FALSE);
1394 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1395 const regnode_charclass *or_with)
1397 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1398 * another SSC or a regular ANYOF class. Can create false positives if
1399 * 'or_with' is to be inverted. */
1404 PERL_ARGS_ASSERT_SSC_OR;
1406 assert(is_ANYOF_SYNTHETIC(ssc));
1408 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1409 * the code point inversion list and just the relevant flags */
1410 if (is_ANYOF_SYNTHETIC(or_with)) {
1411 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1412 ored_flags = ANYOF_FLAGS(or_with);
1415 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1416 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1419 ANYOF_FLAGS(ssc) |= ored_flags;
1421 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1422 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1423 * 'or_with' may be inverted. When not inverted, we have the simple
1424 * situation of computing:
1425 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1426 * If P1|P2 yields a situation with both a class and its complement are
1427 * set, like having both \w and \W, this matches all code points, and we
1428 * can delete these from the P component of the ssc going forward. XXX We
1429 * might be able to delete all the P components, but I (khw) am not certain
1430 * about this, and it is better to be safe.
1433 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1434 * <= (C1 | P1) | ~C2
1435 * <= (C1 | ~C2) | P1
1436 * (which results in actually simpler code than the non-inverted case)
1439 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1440 && ! is_ANYOF_SYNTHETIC(or_with))
1442 /* We ignore P2, leaving P1 going forward */
1443 } /* else Not inverted */
1444 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1445 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1446 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1448 for (i = 0; i < ANYOF_MAX; i += 2) {
1449 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1451 ssc_match_all_cp(ssc);
1452 ANYOF_POSIXL_CLEAR(ssc, i);
1453 ANYOF_POSIXL_CLEAR(ssc, i+1);
1461 FALSE /* Already has been inverted */
1465 PERL_STATIC_INLINE void
1466 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1468 PERL_ARGS_ASSERT_SSC_UNION;
1470 assert(is_ANYOF_SYNTHETIC(ssc));
1472 _invlist_union_maybe_complement_2nd(ssc->invlist,
1478 PERL_STATIC_INLINE void
1479 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1481 const bool invert2nd)
1483 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1485 assert(is_ANYOF_SYNTHETIC(ssc));
1487 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1493 PERL_STATIC_INLINE void
1494 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1496 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1498 assert(is_ANYOF_SYNTHETIC(ssc));
1500 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1503 PERL_STATIC_INLINE void
1504 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1506 /* AND just the single code point 'cp' into the SSC 'ssc' */
1508 SV* cp_list = _new_invlist(2);
1510 PERL_ARGS_ASSERT_SSC_CP_AND;
1512 assert(is_ANYOF_SYNTHETIC(ssc));
1514 cp_list = add_cp_to_invlist(cp_list, cp);
1515 ssc_intersection(ssc, cp_list,
1516 FALSE /* Not inverted */
1518 SvREFCNT_dec_NN(cp_list);
1521 PERL_STATIC_INLINE void
1522 S_ssc_clear_locale(regnode_ssc *ssc)
1524 /* Set the SSC 'ssc' to not match any locale things */
1525 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1527 assert(is_ANYOF_SYNTHETIC(ssc));
1529 ANYOF_POSIXL_ZERO(ssc);
1530 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1533 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1536 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1538 /* The synthetic start class is used to hopefully quickly winnow down
1539 * places where a pattern could start a match in the target string. If it
1540 * doesn't really narrow things down that much, there isn't much point to
1541 * having the overhead of using it. This function uses some very crude
1542 * heuristics to decide if to use the ssc or not.
1544 * It returns TRUE if 'ssc' rules out more than half what it considers to
1545 * be the "likely" possible matches, but of course it doesn't know what the
1546 * actual things being matched are going to be; these are only guesses
1548 * For /l matches, it assumes that the only likely matches are going to be
1549 * in the 0-255 range, uniformly distributed, so half of that is 127
1550 * For /a and /d matches, it assumes that the likely matches will be just
1551 * the ASCII range, so half of that is 63
1552 * For /u and there isn't anything matching above the Latin1 range, it
1553 * assumes that that is the only range likely to be matched, and uses
1554 * half that as the cut-off: 127. If anything matches above Latin1,
1555 * it assumes that all of Unicode could match (uniformly), except for
1556 * non-Unicode code points and things in the General Category "Other"
1557 * (unassigned, private use, surrogates, controls and formats). This
1558 * is a much large number. */
1560 const U32 max_match = (LOC)
1564 : (invlist_highest(ssc->invlist) < 256)
1566 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1567 U32 count = 0; /* Running total of number of code points matched by
1569 UV start, end; /* Start and end points of current range in inversion
1572 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1574 invlist_iterinit(ssc->invlist);
1575 while (invlist_iternext(ssc->invlist, &start, &end)) {
1577 /* /u is the only thing that we expect to match above 255; so if not /u
1578 * and even if there are matches above 255, ignore them. This catches
1579 * things like \d under /d which does match the digits above 255, but
1580 * since the pattern is /d, it is not likely to be expecting them */
1581 if (! UNI_SEMANTICS) {
1585 end = MIN(end, 255);
1587 count += end - start + 1;
1588 if (count > max_match) {
1589 invlist_iterfinish(ssc->invlist);
1599 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1601 /* The inversion list in the SSC is marked mortal; now we need a more
1602 * permanent copy, which is stored the same way that is done in a regular
1603 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1606 SV* invlist = invlist_clone(ssc->invlist);
1608 PERL_ARGS_ASSERT_SSC_FINALIZE;
1610 assert(is_ANYOF_SYNTHETIC(ssc));
1612 /* The code in this file assumes that all but these flags aren't relevant
1613 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1614 * by the time we reach here */
1615 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1617 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1619 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1620 NULL, NULL, NULL, FALSE);
1622 /* Make sure is clone-safe */
1623 ssc->invlist = NULL;
1625 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1626 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1629 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1632 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1633 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1634 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1635 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1636 ? (TRIE_LIST_CUR( idx ) - 1) \
1642 dump_trie(trie,widecharmap,revcharmap)
1643 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1644 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1646 These routines dump out a trie in a somewhat readable format.
1647 The _interim_ variants are used for debugging the interim
1648 tables that are used to generate the final compressed
1649 representation which is what dump_trie expects.
1651 Part of the reason for their existence is to provide a form
1652 of documentation as to how the different representations function.
1657 Dumps the final compressed table form of the trie to Perl_debug_log.
1658 Used for debugging make_trie().
1662 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1663 AV *revcharmap, U32 depth)
1666 SV *sv=sv_newmortal();
1667 int colwidth= widecharmap ? 6 : 4;
1669 GET_RE_DEBUG_FLAGS_DECL;
1671 PERL_ARGS_ASSERT_DUMP_TRIE;
1673 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1674 (int)depth * 2 + 2,"",
1675 "Match","Base","Ofs" );
1677 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1678 SV ** const tmp = av_fetch( revcharmap, state, 0);
1680 PerlIO_printf( Perl_debug_log, "%*s",
1682 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1683 PL_colors[0], PL_colors[1],
1684 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1685 PERL_PV_ESCAPE_FIRSTCHAR
1690 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1691 (int)depth * 2 + 2,"");
1693 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1694 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1695 PerlIO_printf( Perl_debug_log, "\n");
1697 for( state = 1 ; state < trie->statecount ; state++ ) {
1698 const U32 base = trie->states[ state ].trans.base;
1700 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1701 (int)depth * 2 + 2,"", (UV)state);
1703 if ( trie->states[ state ].wordnum ) {
1704 PerlIO_printf( Perl_debug_log, " W%4X",
1705 trie->states[ state ].wordnum );
1707 PerlIO_printf( Perl_debug_log, "%6s", "" );
1710 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1715 while( ( base + ofs < trie->uniquecharcount ) ||
1716 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1717 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1721 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1723 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1724 if ( ( base + ofs >= trie->uniquecharcount )
1725 && ( base + ofs - trie->uniquecharcount
1727 && trie->trans[ base + ofs
1728 - trie->uniquecharcount ].check == state )
1730 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1732 (UV)trie->trans[ base + ofs
1733 - trie->uniquecharcount ].next );
1735 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1739 PerlIO_printf( Perl_debug_log, "]");
1742 PerlIO_printf( Perl_debug_log, "\n" );
1744 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1746 for (word=1; word <= trie->wordcount; word++) {
1747 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1748 (int)word, (int)(trie->wordinfo[word].prev),
1749 (int)(trie->wordinfo[word].len));
1751 PerlIO_printf(Perl_debug_log, "\n" );
1754 Dumps a fully constructed but uncompressed trie in list form.
1755 List tries normally only are used for construction when the number of
1756 possible chars (trie->uniquecharcount) is very high.
1757 Used for debugging make_trie().
1760 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1761 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1765 SV *sv=sv_newmortal();
1766 int colwidth= widecharmap ? 6 : 4;
1767 GET_RE_DEBUG_FLAGS_DECL;
1769 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1771 /* print out the table precompression. */
1772 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1773 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1774 "------:-----+-----------------\n" );
1776 for( state=1 ; state < next_alloc ; state ++ ) {
1779 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1780 (int)depth * 2 + 2,"", (UV)state );
1781 if ( ! trie->states[ state ].wordnum ) {
1782 PerlIO_printf( Perl_debug_log, "%5s| ","");
1784 PerlIO_printf( Perl_debug_log, "W%4x| ",
1785 trie->states[ state ].wordnum
1788 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1789 SV ** const tmp = av_fetch( revcharmap,
1790 TRIE_LIST_ITEM(state,charid).forid, 0);
1792 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1794 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1796 PL_colors[0], PL_colors[1],
1797 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1798 | PERL_PV_ESCAPE_FIRSTCHAR
1800 TRIE_LIST_ITEM(state,charid).forid,
1801 (UV)TRIE_LIST_ITEM(state,charid).newstate
1804 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1805 (int)((depth * 2) + 14), "");
1808 PerlIO_printf( Perl_debug_log, "\n");
1813 Dumps a fully constructed but uncompressed trie in table form.
1814 This is the normal DFA style state transition table, with a few
1815 twists to facilitate compression later.
1816 Used for debugging make_trie().
1819 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1820 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1825 SV *sv=sv_newmortal();
1826 int colwidth= widecharmap ? 6 : 4;
1827 GET_RE_DEBUG_FLAGS_DECL;
1829 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1832 print out the table precompression so that we can do a visual check
1833 that they are identical.
1836 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1838 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1839 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1841 PerlIO_printf( Perl_debug_log, "%*s",
1843 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1844 PL_colors[0], PL_colors[1],
1845 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1846 PERL_PV_ESCAPE_FIRSTCHAR
1852 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1854 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1855 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1858 PerlIO_printf( Perl_debug_log, "\n" );
1860 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1862 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1863 (int)depth * 2 + 2,"",
1864 (UV)TRIE_NODENUM( state ) );
1866 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1867 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1869 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1871 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1873 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1874 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1875 (UV)trie->trans[ state ].check );
1877 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1878 (UV)trie->trans[ state ].check,
1879 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1887 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1888 startbranch: the first branch in the whole branch sequence
1889 first : start branch of sequence of branch-exact nodes.
1890 May be the same as startbranch
1891 last : Thing following the last branch.
1892 May be the same as tail.
1893 tail : item following the branch sequence
1894 count : words in the sequence
1895 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1896 depth : indent depth
1898 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1900 A trie is an N'ary tree where the branches are determined by digital
1901 decomposition of the key. IE, at the root node you look up the 1st character and
1902 follow that branch repeat until you find the end of the branches. Nodes can be
1903 marked as "accepting" meaning they represent a complete word. Eg:
1907 would convert into the following structure. Numbers represent states, letters
1908 following numbers represent valid transitions on the letter from that state, if
1909 the number is in square brackets it represents an accepting state, otherwise it
1910 will be in parenthesis.
1912 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1916 (1) +-i->(6)-+-s->[7]
1918 +-s->(3)-+-h->(4)-+-e->[5]
1920 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1922 This shows that when matching against the string 'hers' we will begin at state 1
1923 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1924 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1925 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1926 single traverse. We store a mapping from accepting to state to which word was
1927 matched, and then when we have multiple possibilities we try to complete the
1928 rest of the regex in the order in which they occurred in the alternation.
1930 The only prior NFA like behaviour that would be changed by the TRIE support is
1931 the silent ignoring of duplicate alternations which are of the form:
1933 / (DUPE|DUPE) X? (?{ ... }) Y /x
1935 Thus EVAL blocks following a trie may be called a different number of times with
1936 and without the optimisation. With the optimisations dupes will be silently
1937 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1938 the following demonstrates:
1940 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1942 which prints out 'word' three times, but
1944 'words'=~/(word|word|word)(?{ print $1 })S/
1946 which doesnt print it out at all. This is due to other optimisations kicking in.
1948 Example of what happens on a structural level:
1950 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1952 1: CURLYM[1] {1,32767}(18)
1963 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1964 and should turn into:
1966 1: CURLYM[1] {1,32767}(18)
1968 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1976 Cases where tail != last would be like /(?foo|bar)baz/:
1986 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1987 and would end up looking like:
1990 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1997 d = uvchr_to_utf8_flags(d, uv, 0);
1999 is the recommended Unicode-aware way of saying
2004 #define TRIE_STORE_REVCHAR(val) \
2007 SV *zlopp = newSV(7); /* XXX: optimize me */ \
2008 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2009 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2010 SvCUR_set(zlopp, kapow - flrbbbbb); \
2013 av_push(revcharmap, zlopp); \
2015 char ooooff = (char)val; \
2016 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2020 /* This gets the next character from the input, folding it if not already
2022 #define TRIE_READ_CHAR STMT_START { \
2025 /* if it is UTF then it is either already folded, or does not need \
2027 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2029 else if (folder == PL_fold_latin1) { \
2030 /* This folder implies Unicode rules, which in the range expressible \
2031 * by not UTF is the lower case, with the two exceptions, one of \
2032 * which should have been taken care of before calling this */ \
2033 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2034 uvc = toLOWER_L1(*uc); \
2035 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2038 /* raw data, will be folded later if needed */ \
2046 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2047 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2048 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2049 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2051 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2052 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2053 TRIE_LIST_CUR( state )++; \
2056 #define TRIE_LIST_NEW(state) STMT_START { \
2057 Newxz( trie->states[ state ].trans.list, \
2058 4, reg_trie_trans_le ); \
2059 TRIE_LIST_CUR( state ) = 1; \
2060 TRIE_LIST_LEN( state ) = 4; \
2063 #define TRIE_HANDLE_WORD(state) STMT_START { \
2064 U16 dupe= trie->states[ state ].wordnum; \
2065 regnode * const noper_next = regnext( noper ); \
2068 /* store the word for dumping */ \
2070 if (OP(noper) != NOTHING) \
2071 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2073 tmp = newSVpvn_utf8( "", 0, UTF ); \
2074 av_push( trie_words, tmp ); \
2078 trie->wordinfo[curword].prev = 0; \
2079 trie->wordinfo[curword].len = wordlen; \
2080 trie->wordinfo[curword].accept = state; \
2082 if ( noper_next < tail ) { \
2084 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2086 trie->jump[curword] = (U16)(noper_next - convert); \
2088 jumper = noper_next; \
2090 nextbranch= regnext(cur); \
2094 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2095 /* chain, so that when the bits of chain are later */\
2096 /* linked together, the dups appear in the chain */\
2097 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2098 trie->wordinfo[dupe].prev = curword; \
2100 /* we haven't inserted this word yet. */ \
2101 trie->states[ state ].wordnum = curword; \
2106 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2107 ( ( base + charid >= ucharcount \
2108 && base + charid < ubound \
2109 && state == trie->trans[ base - ucharcount + charid ].check \
2110 && trie->trans[ base - ucharcount + charid ].next ) \
2111 ? trie->trans[ base - ucharcount + charid ].next \
2112 : ( state==1 ? special : 0 ) \
2116 #define MADE_JUMP_TRIE 2
2117 #define MADE_EXACT_TRIE 4
2120 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2121 regnode *first, regnode *last, regnode *tail,
2122 U32 word_count, U32 flags, U32 depth)
2124 /* first pass, loop through and scan words */
2125 reg_trie_data *trie;
2126 HV *widecharmap = NULL;
2127 AV *revcharmap = newAV();
2133 regnode *jumper = NULL;
2134 regnode *nextbranch = NULL;
2135 regnode *convert = NULL;
2136 U32 *prev_states; /* temp array mapping each state to previous one */
2137 /* we just use folder as a flag in utf8 */
2138 const U8 * folder = NULL;
2141 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2142 AV *trie_words = NULL;
2143 /* along with revcharmap, this only used during construction but both are
2144 * useful during debugging so we store them in the struct when debugging.
2147 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2148 STRLEN trie_charcount=0;
2150 SV *re_trie_maxbuff;
2151 GET_RE_DEBUG_FLAGS_DECL;
2153 PERL_ARGS_ASSERT_MAKE_TRIE;
2155 PERL_UNUSED_ARG(depth);
2159 case EXACT: case EXACTL: break;
2163 case EXACTFLU8: folder = PL_fold_latin1; break;
2164 case EXACTF: folder = PL_fold; break;
2165 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2168 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2170 trie->startstate = 1;
2171 trie->wordcount = word_count;
2172 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2173 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2174 if (flags == EXACT || flags == EXACTL)
2175 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2176 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2177 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2180 trie_words = newAV();
2183 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2184 assert(re_trie_maxbuff);
2185 if (!SvIOK(re_trie_maxbuff)) {
2186 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2188 DEBUG_TRIE_COMPILE_r({
2189 PerlIO_printf( Perl_debug_log,
2190 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2191 (int)depth * 2 + 2, "",
2192 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2193 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2196 /* Find the node we are going to overwrite */
2197 if ( first == startbranch && OP( last ) != BRANCH ) {
2198 /* whole branch chain */
2201 /* branch sub-chain */
2202 convert = NEXTOPER( first );
2205 /* -- First loop and Setup --
2207 We first traverse the branches and scan each word to determine if it
2208 contains widechars, and how many unique chars there are, this is
2209 important as we have to build a table with at least as many columns as we
2212 We use an array of integers to represent the character codes 0..255
2213 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2214 the native representation of the character value as the key and IV's for
2217 *TODO* If we keep track of how many times each character is used we can
2218 remap the columns so that the table compression later on is more
2219 efficient in terms of memory by ensuring the most common value is in the
2220 middle and the least common are on the outside. IMO this would be better
2221 than a most to least common mapping as theres a decent chance the most
2222 common letter will share a node with the least common, meaning the node
2223 will not be compressible. With a middle is most common approach the worst
2224 case is when we have the least common nodes twice.
2228 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2229 regnode *noper = NEXTOPER( cur );
2230 const U8 *uc = (U8*)STRING( noper );
2231 const U8 *e = uc + STR_LEN( noper );
2233 U32 wordlen = 0; /* required init */
2234 STRLEN minchars = 0;
2235 STRLEN maxchars = 0;
2236 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2239 if (OP(noper) == NOTHING) {
2240 regnode *noper_next= regnext(noper);
2241 if (noper_next != tail && OP(noper_next) == flags) {
2243 uc= (U8*)STRING(noper);
2244 e= uc + STR_LEN(noper);
2245 trie->minlen= STR_LEN(noper);
2252 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2253 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2254 regardless of encoding */
2255 if (OP( noper ) == EXACTFU_SS) {
2256 /* false positives are ok, so just set this */
2257 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2260 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2262 TRIE_CHARCOUNT(trie)++;
2265 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2266 * is in effect. Under /i, this character can match itself, or
2267 * anything that folds to it. If not under /i, it can match just
2268 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2269 * all fold to k, and all are single characters. But some folds
2270 * expand to more than one character, so for example LATIN SMALL
2271 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2272 * the string beginning at 'uc' is 'ffi', it could be matched by
2273 * three characters, or just by the one ligature character. (It
2274 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2275 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2276 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2277 * match.) The trie needs to know the minimum and maximum number
2278 * of characters that could match so that it can use size alone to
2279 * quickly reject many match attempts. The max is simple: it is
2280 * the number of folded characters in this branch (since a fold is
2281 * never shorter than what folds to it. */
2285 /* And the min is equal to the max if not under /i (indicated by
2286 * 'folder' being NULL), or there are no multi-character folds. If
2287 * there is a multi-character fold, the min is incremented just
2288 * once, for the character that folds to the sequence. Each
2289 * character in the sequence needs to be added to the list below of
2290 * characters in the trie, but we count only the first towards the
2291 * min number of characters needed. This is done through the
2292 * variable 'foldlen', which is returned by the macros that look
2293 * for these sequences as the number of bytes the sequence
2294 * occupies. Each time through the loop, we decrement 'foldlen' by
2295 * how many bytes the current char occupies. Only when it reaches
2296 * 0 do we increment 'minchars' or look for another multi-character
2298 if (folder == NULL) {
2301 else if (foldlen > 0) {
2302 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2307 /* See if *uc is the beginning of a multi-character fold. If
2308 * so, we decrement the length remaining to look at, to account
2309 * for the current character this iteration. (We can use 'uc'
2310 * instead of the fold returned by TRIE_READ_CHAR because for
2311 * non-UTF, the latin1_safe macro is smart enough to account
2312 * for all the unfolded characters, and because for UTF, the
2313 * string will already have been folded earlier in the
2314 * compilation process */
2316 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2317 foldlen -= UTF8SKIP(uc);
2320 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2325 /* The current character (and any potential folds) should be added
2326 * to the possible matching characters for this position in this
2330 U8 folded= folder[ (U8) uvc ];
2331 if ( !trie->charmap[ folded ] ) {
2332 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2333 TRIE_STORE_REVCHAR( folded );
2336 if ( !trie->charmap[ uvc ] ) {
2337 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2338 TRIE_STORE_REVCHAR( uvc );
2341 /* store the codepoint in the bitmap, and its folded
2343 TRIE_BITMAP_SET(trie, uvc);
2345 /* store the folded codepoint */
2346 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2349 /* store first byte of utf8 representation of
2350 variant codepoints */
2351 if (! UVCHR_IS_INVARIANT(uvc)) {
2352 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2355 set_bit = 0; /* We've done our bit :-) */
2359 /* XXX We could come up with the list of code points that fold
2360 * to this using PL_utf8_foldclosures, except not for
2361 * multi-char folds, as there may be multiple combinations
2362 * there that could work, which needs to wait until runtime to
2363 * resolve (The comment about LIGATURE FFI above is such an
2368 widecharmap = newHV();
2370 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2373 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2375 if ( !SvTRUE( *svpp ) ) {
2376 sv_setiv( *svpp, ++trie->uniquecharcount );
2377 TRIE_STORE_REVCHAR(uvc);
2380 } /* end loop through characters in this branch of the trie */
2382 /* We take the min and max for this branch and combine to find the min
2383 * and max for all branches processed so far */
2384 if( cur == first ) {
2385 trie->minlen = minchars;
2386 trie->maxlen = maxchars;
2387 } else if (minchars < trie->minlen) {
2388 trie->minlen = minchars;
2389 } else if (maxchars > trie->maxlen) {
2390 trie->maxlen = maxchars;
2392 } /* end first pass */
2393 DEBUG_TRIE_COMPILE_r(
2394 PerlIO_printf( Perl_debug_log,
2395 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2396 (int)depth * 2 + 2,"",
2397 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2398 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2399 (int)trie->minlen, (int)trie->maxlen )
2403 We now know what we are dealing with in terms of unique chars and
2404 string sizes so we can calculate how much memory a naive
2405 representation using a flat table will take. If it's over a reasonable
2406 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2407 conservative but potentially much slower representation using an array
2410 At the end we convert both representations into the same compressed
2411 form that will be used in regexec.c for matching with. The latter
2412 is a form that cannot be used to construct with but has memory
2413 properties similar to the list form and access properties similar
2414 to the table form making it both suitable for fast searches and
2415 small enough that its feasable to store for the duration of a program.
2417 See the comment in the code where the compressed table is produced
2418 inplace from the flat tabe representation for an explanation of how
2419 the compression works.
2424 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2427 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2428 > SvIV(re_trie_maxbuff) )
2431 Second Pass -- Array Of Lists Representation
2433 Each state will be represented by a list of charid:state records
2434 (reg_trie_trans_le) the first such element holds the CUR and LEN
2435 points of the allocated array. (See defines above).
2437 We build the initial structure using the lists, and then convert
2438 it into the compressed table form which allows faster lookups
2439 (but cant be modified once converted).
2442 STRLEN transcount = 1;
2444 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2445 "%*sCompiling trie using list compiler\n",
2446 (int)depth * 2 + 2, ""));
2448 trie->states = (reg_trie_state *)
2449 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2450 sizeof(reg_trie_state) );
2454 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2456 regnode *noper = NEXTOPER( cur );
2457 U8 *uc = (U8*)STRING( noper );
2458 const U8 *e = uc + STR_LEN( noper );
2459 U32 state = 1; /* required init */
2460 U16 charid = 0; /* sanity init */
2461 U32 wordlen = 0; /* required init */
2463 if (OP(noper) == NOTHING) {
2464 regnode *noper_next= regnext(noper);
2465 if (noper_next != tail && OP(noper_next) == flags) {
2467 uc= (U8*)STRING(noper);
2468 e= uc + STR_LEN(noper);
2472 if (OP(noper) != NOTHING) {
2473 for ( ; uc < e ; uc += len ) {
2478 charid = trie->charmap[ uvc ];
2480 SV** const svpp = hv_fetch( widecharmap,
2487 charid=(U16)SvIV( *svpp );
2490 /* charid is now 0 if we dont know the char read, or
2491 * nonzero if we do */
2498 if ( !trie->states[ state ].trans.list ) {
2499 TRIE_LIST_NEW( state );
2502 check <= TRIE_LIST_USED( state );
2505 if ( TRIE_LIST_ITEM( state, check ).forid
2508 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2513 newstate = next_alloc++;
2514 prev_states[newstate] = state;
2515 TRIE_LIST_PUSH( state, charid, newstate );
2520 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2524 TRIE_HANDLE_WORD(state);
2526 } /* end second pass */
2528 /* next alloc is the NEXT state to be allocated */
2529 trie->statecount = next_alloc;
2530 trie->states = (reg_trie_state *)
2531 PerlMemShared_realloc( trie->states,
2533 * sizeof(reg_trie_state) );
2535 /* and now dump it out before we compress it */
2536 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2537 revcharmap, next_alloc,
2541 trie->trans = (reg_trie_trans *)
2542 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2549 for( state=1 ; state < next_alloc ; state ++ ) {
2553 DEBUG_TRIE_COMPILE_MORE_r(
2554 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2558 if (trie->states[state].trans.list) {
2559 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2563 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2564 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2565 if ( forid < minid ) {
2567 } else if ( forid > maxid ) {
2571 if ( transcount < tp + maxid - minid + 1) {
2573 trie->trans = (reg_trie_trans *)
2574 PerlMemShared_realloc( trie->trans,
2576 * sizeof(reg_trie_trans) );
2577 Zero( trie->trans + (transcount / 2),
2581 base = trie->uniquecharcount + tp - minid;
2582 if ( maxid == minid ) {
2584 for ( ; zp < tp ; zp++ ) {
2585 if ( ! trie->trans[ zp ].next ) {
2586 base = trie->uniquecharcount + zp - minid;
2587 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2589 trie->trans[ zp ].check = state;
2595 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2597 trie->trans[ tp ].check = state;
2602 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2603 const U32 tid = base
2604 - trie->uniquecharcount
2605 + TRIE_LIST_ITEM( state, idx ).forid;
2606 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2608 trie->trans[ tid ].check = state;
2610 tp += ( maxid - minid + 1 );
2612 Safefree(trie->states[ state ].trans.list);
2615 DEBUG_TRIE_COMPILE_MORE_r(
2616 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2619 trie->states[ state ].trans.base=base;
2621 trie->lasttrans = tp + 1;
2625 Second Pass -- Flat Table Representation.
2627 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2628 each. We know that we will need Charcount+1 trans at most to store
2629 the data (one row per char at worst case) So we preallocate both
2630 structures assuming worst case.
2632 We then construct the trie using only the .next slots of the entry
2635 We use the .check field of the first entry of the node temporarily
2636 to make compression both faster and easier by keeping track of how
2637 many non zero fields are in the node.
2639 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2642 There are two terms at use here: state as a TRIE_NODEIDX() which is
2643 a number representing the first entry of the node, and state as a
2644 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2645 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2646 if there are 2 entrys per node. eg:
2654 The table is internally in the right hand, idx form. However as we
2655 also have to deal with the states array which is indexed by nodenum
2656 we have to use TRIE_NODENUM() to convert.
2659 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2660 "%*sCompiling trie using table compiler\n",
2661 (int)depth * 2 + 2, ""));
2663 trie->trans = (reg_trie_trans *)
2664 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2665 * trie->uniquecharcount + 1,
2666 sizeof(reg_trie_trans) );
2667 trie->states = (reg_trie_state *)
2668 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2669 sizeof(reg_trie_state) );
2670 next_alloc = trie->uniquecharcount + 1;
2673 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2675 regnode *noper = NEXTOPER( cur );
2676 const U8 *uc = (U8*)STRING( noper );
2677 const U8 *e = uc + STR_LEN( noper );
2679 U32 state = 1; /* required init */
2681 U16 charid = 0; /* sanity init */
2682 U32 accept_state = 0; /* sanity init */
2684 U32 wordlen = 0; /* required init */
2686 if (OP(noper) == NOTHING) {
2687 regnode *noper_next= regnext(noper);
2688 if (noper_next != tail && OP(noper_next) == flags) {
2690 uc= (U8*)STRING(noper);
2691 e= uc + STR_LEN(noper);
2695 if ( OP(noper) != NOTHING ) {
2696 for ( ; uc < e ; uc += len ) {
2701 charid = trie->charmap[ uvc ];
2703 SV* const * const svpp = hv_fetch( widecharmap,
2707 charid = svpp ? (U16)SvIV(*svpp) : 0;
2711 if ( !trie->trans[ state + charid ].next ) {
2712 trie->trans[ state + charid ].next = next_alloc;
2713 trie->trans[ state ].check++;
2714 prev_states[TRIE_NODENUM(next_alloc)]
2715 = TRIE_NODENUM(state);
2716 next_alloc += trie->uniquecharcount;
2718 state = trie->trans[ state + charid ].next;
2720 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2722 /* charid is now 0 if we dont know the char read, or
2723 * nonzero if we do */
2726 accept_state = TRIE_NODENUM( state );
2727 TRIE_HANDLE_WORD(accept_state);
2729 } /* end second pass */
2731 /* and now dump it out before we compress it */
2732 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2734 next_alloc, depth+1));
2738 * Inplace compress the table.*
2740 For sparse data sets the table constructed by the trie algorithm will
2741 be mostly 0/FAIL transitions or to put it another way mostly empty.
2742 (Note that leaf nodes will not contain any transitions.)
2744 This algorithm compresses the tables by eliminating most such
2745 transitions, at the cost of a modest bit of extra work during lookup:
2747 - Each states[] entry contains a .base field which indicates the
2748 index in the state[] array wheres its transition data is stored.
2750 - If .base is 0 there are no valid transitions from that node.
2752 - If .base is nonzero then charid is added to it to find an entry in
2755 -If trans[states[state].base+charid].check!=state then the
2756 transition is taken to be a 0/Fail transition. Thus if there are fail
2757 transitions at the front of the node then the .base offset will point
2758 somewhere inside the previous nodes data (or maybe even into a node
2759 even earlier), but the .check field determines if the transition is
2763 The following process inplace converts the table to the compressed
2764 table: We first do not compress the root node 1,and mark all its
2765 .check pointers as 1 and set its .base pointer as 1 as well. This
2766 allows us to do a DFA construction from the compressed table later,
2767 and ensures that any .base pointers we calculate later are greater
2770 - We set 'pos' to indicate the first entry of the second node.
2772 - We then iterate over the columns of the node, finding the first and
2773 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2774 and set the .check pointers accordingly, and advance pos
2775 appropriately and repreat for the next node. Note that when we copy
2776 the next pointers we have to convert them from the original
2777 NODEIDX form to NODENUM form as the former is not valid post
2780 - If a node has no transitions used we mark its base as 0 and do not
2781 advance the pos pointer.
2783 - If a node only has one transition we use a second pointer into the
2784 structure to fill in allocated fail transitions from other states.
2785 This pointer is independent of the main pointer and scans forward
2786 looking for null transitions that are allocated to a state. When it
2787 finds one it writes the single transition into the "hole". If the
2788 pointer doesnt find one the single transition is appended as normal.
2790 - Once compressed we can Renew/realloc the structures to release the
2793 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2794 specifically Fig 3.47 and the associated pseudocode.
2798 const U32 laststate = TRIE_NODENUM( next_alloc );
2801 trie->statecount = laststate;
2803 for ( state = 1 ; state < laststate ; state++ ) {
2805 const U32 stateidx = TRIE_NODEIDX( state );
2806 const U32 o_used = trie->trans[ stateidx ].check;
2807 U32 used = trie->trans[ stateidx ].check;
2808 trie->trans[ stateidx ].check = 0;
2811 used && charid < trie->uniquecharcount;
2814 if ( flag || trie->trans[ stateidx + charid ].next ) {
2815 if ( trie->trans[ stateidx + charid ].next ) {
2817 for ( ; zp < pos ; zp++ ) {
2818 if ( ! trie->trans[ zp ].next ) {
2822 trie->states[ state ].trans.base
2824 + trie->uniquecharcount
2826 trie->trans[ zp ].next
2827 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2829 trie->trans[ zp ].check = state;
2830 if ( ++zp > pos ) pos = zp;
2837 trie->states[ state ].trans.base
2838 = pos + trie->uniquecharcount - charid ;
2840 trie->trans[ pos ].next
2841 = SAFE_TRIE_NODENUM(
2842 trie->trans[ stateidx + charid ].next );
2843 trie->trans[ pos ].check = state;
2848 trie->lasttrans = pos + 1;
2849 trie->states = (reg_trie_state *)
2850 PerlMemShared_realloc( trie->states, laststate
2851 * sizeof(reg_trie_state) );
2852 DEBUG_TRIE_COMPILE_MORE_r(
2853 PerlIO_printf( Perl_debug_log,
2854 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2855 (int)depth * 2 + 2,"",
2856 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2860 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2863 } /* end table compress */
2865 DEBUG_TRIE_COMPILE_MORE_r(
2866 PerlIO_printf(Perl_debug_log,
2867 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2868 (int)depth * 2 + 2, "",
2869 (UV)trie->statecount,
2870 (UV)trie->lasttrans)
2872 /* resize the trans array to remove unused space */
2873 trie->trans = (reg_trie_trans *)
2874 PerlMemShared_realloc( trie->trans, trie->lasttrans
2875 * sizeof(reg_trie_trans) );
2877 { /* Modify the program and insert the new TRIE node */
2878 U8 nodetype =(U8)(flags & 0xFF);
2882 regnode *optimize = NULL;
2883 #ifdef RE_TRACK_PATTERN_OFFSETS
2886 U32 mjd_nodelen = 0;
2887 #endif /* RE_TRACK_PATTERN_OFFSETS */
2888 #endif /* DEBUGGING */
2890 This means we convert either the first branch or the first Exact,
2891 depending on whether the thing following (in 'last') is a branch
2892 or not and whther first is the startbranch (ie is it a sub part of
2893 the alternation or is it the whole thing.)
2894 Assuming its a sub part we convert the EXACT otherwise we convert
2895 the whole branch sequence, including the first.
2897 /* Find the node we are going to overwrite */
2898 if ( first != startbranch || OP( last ) == BRANCH ) {
2899 /* branch sub-chain */
2900 NEXT_OFF( first ) = (U16)(last - first);
2901 #ifdef RE_TRACK_PATTERN_OFFSETS
2903 mjd_offset= Node_Offset((convert));
2904 mjd_nodelen= Node_Length((convert));
2907 /* whole branch chain */
2909 #ifdef RE_TRACK_PATTERN_OFFSETS
2912 const regnode *nop = NEXTOPER( convert );
2913 mjd_offset= Node_Offset((nop));
2914 mjd_nodelen= Node_Length((nop));
2918 PerlIO_printf(Perl_debug_log,
2919 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2920 (int)depth * 2 + 2, "",
2921 (UV)mjd_offset, (UV)mjd_nodelen)
2924 /* But first we check to see if there is a common prefix we can
2925 split out as an EXACT and put in front of the TRIE node. */
2926 trie->startstate= 1;
2927 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2929 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2933 const U32 base = trie->states[ state ].trans.base;
2935 if ( trie->states[state].wordnum )
2938 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2939 if ( ( base + ofs >= trie->uniquecharcount ) &&
2940 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2941 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2943 if ( ++count > 1 ) {
2944 SV **tmp = av_fetch( revcharmap, ofs, 0);
2945 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2946 if ( state == 1 ) break;
2948 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2950 PerlIO_printf(Perl_debug_log,
2951 "%*sNew Start State=%"UVuf" Class: [",
2952 (int)depth * 2 + 2, "",
2955 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2956 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2958 TRIE_BITMAP_SET(trie,*ch);
2960 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2962 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2966 TRIE_BITMAP_SET(trie,*ch);
2968 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2969 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2975 SV **tmp = av_fetch( revcharmap, idx, 0);
2977 char *ch = SvPV( *tmp, len );
2979 SV *sv=sv_newmortal();
2980 PerlIO_printf( Perl_debug_log,
2981 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2982 (int)depth * 2 + 2, "",
2984 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2985 PL_colors[0], PL_colors[1],
2986 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2987 PERL_PV_ESCAPE_FIRSTCHAR
2992 OP( convert ) = nodetype;
2993 str=STRING(convert);
2996 STR_LEN(convert) += len;
3002 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3007 trie->prefixlen = (state-1);
3009 regnode *n = convert+NODE_SZ_STR(convert);
3010 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3011 trie->startstate = state;
3012 trie->minlen -= (state - 1);
3013 trie->maxlen -= (state - 1);
3015 /* At least the UNICOS C compiler choked on this
3016 * being argument to DEBUG_r(), so let's just have
3019 #ifdef PERL_EXT_RE_BUILD
3025 regnode *fix = convert;
3026 U32 word = trie->wordcount;
3028 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3029 while( ++fix < n ) {
3030 Set_Node_Offset_Length(fix, 0, 0);
3033 SV ** const tmp = av_fetch( trie_words, word, 0 );
3035 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3036 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3038 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3046 NEXT_OFF(convert) = (U16)(tail - convert);
3047 DEBUG_r(optimize= n);
3053 if ( trie->maxlen ) {
3054 NEXT_OFF( convert ) = (U16)(tail - convert);
3055 ARG_SET( convert, data_slot );
3056 /* Store the offset to the first unabsorbed branch in
3057 jump[0], which is otherwise unused by the jump logic.
3058 We use this when dumping a trie and during optimisation. */
3060 trie->jump[0] = (U16)(nextbranch - convert);
3062 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3063 * and there is a bitmap
3064 * and the first "jump target" node we found leaves enough room
3065 * then convert the TRIE node into a TRIEC node, with the bitmap
3066 * embedded inline in the opcode - this is hypothetically faster.
3068 if ( !trie->states[trie->startstate].wordnum
3070 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3072 OP( convert ) = TRIEC;
3073 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3074 PerlMemShared_free(trie->bitmap);
3077 OP( convert ) = TRIE;
3079 /* store the type in the flags */
3080 convert->flags = nodetype;
3084 + regarglen[ OP( convert ) ];
3086 /* XXX We really should free up the resource in trie now,
3087 as we won't use them - (which resources?) dmq */
3089 /* needed for dumping*/
3090 DEBUG_r(if (optimize) {
3091 regnode *opt = convert;
3093 while ( ++opt < optimize) {
3094 Set_Node_Offset_Length(opt,0,0);
3097 Try to clean up some of the debris left after the
3100 while( optimize < jumper ) {
3101 mjd_nodelen += Node_Length((optimize));
3102 OP( optimize ) = OPTIMIZED;
3103 Set_Node_Offset_Length(optimize,0,0);
3106 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3108 } /* end node insert */
3110 /* Finish populating the prev field of the wordinfo array. Walk back
3111 * from each accept state until we find another accept state, and if
3112 * so, point the first word's .prev field at the second word. If the
3113 * second already has a .prev field set, stop now. This will be the
3114 * case either if we've already processed that word's accept state,
3115 * or that state had multiple words, and the overspill words were
3116 * already linked up earlier.
3123 for (word=1; word <= trie->wordcount; word++) {
3125 if (trie->wordinfo[word].prev)
3127 state = trie->wordinfo[word].accept;
3129 state = prev_states[state];
3132 prev = trie->states[state].wordnum;
3136 trie->wordinfo[word].prev = prev;
3138 Safefree(prev_states);
3142 /* and now dump out the compressed format */
3143 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3145 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3147 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3148 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3150 SvREFCNT_dec_NN(revcharmap);
3154 : trie->startstate>1
3160 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3162 /* The Trie is constructed and compressed now so we can build a fail array if
3165 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3167 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3171 We find the fail state for each state in the trie, this state is the longest
3172 proper suffix of the current state's 'word' that is also a proper prefix of
3173 another word in our trie. State 1 represents the word '' and is thus the
3174 default fail state. This allows the DFA not to have to restart after its
3175 tried and failed a word at a given point, it simply continues as though it
3176 had been matching the other word in the first place.
3178 'abcdgu'=~/abcdefg|cdgu/
3179 When we get to 'd' we are still matching the first word, we would encounter
3180 'g' which would fail, which would bring us to the state representing 'd' in
3181 the second word where we would try 'g' and succeed, proceeding to match
3184 /* add a fail transition */
3185 const U32 trie_offset = ARG(source);
3186 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3188 const U32 ucharcount = trie->uniquecharcount;
3189 const U32 numstates = trie->statecount;
3190 const U32 ubound = trie->lasttrans + ucharcount;
3194 U32 base = trie->states[ 1 ].trans.base;
3197 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3199 GET_RE_DEBUG_FLAGS_DECL;
3201 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3202 PERL_UNUSED_CONTEXT;
3204 PERL_UNUSED_ARG(depth);
3207 if ( OP(source) == TRIE ) {
3208 struct regnode_1 *op = (struct regnode_1 *)
3209 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3210 StructCopy(source,op,struct regnode_1);
3211 stclass = (regnode *)op;
3213 struct regnode_charclass *op = (struct regnode_charclass *)
3214 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3215 StructCopy(source,op,struct regnode_charclass);
3216 stclass = (regnode *)op;
3218 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3220 ARG_SET( stclass, data_slot );
3221 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3222 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3223 aho->trie=trie_offset;
3224 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3225 Copy( trie->states, aho->states, numstates, reg_trie_state );
3226 Newxz( q, numstates, U32);
3227 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3230 /* initialize fail[0..1] to be 1 so that we always have
3231 a valid final fail state */
3232 fail[ 0 ] = fail[ 1 ] = 1;
3234 for ( charid = 0; charid < ucharcount ; charid++ ) {
3235 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3237 q[ q_write ] = newstate;
3238 /* set to point at the root */
3239 fail[ q[ q_write++ ] ]=1;
3242 while ( q_read < q_write) {
3243 const U32 cur = q[ q_read++ % numstates ];
3244 base = trie->states[ cur ].trans.base;
3246 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3247 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3249 U32 fail_state = cur;
3252 fail_state = fail[ fail_state ];
3253 fail_base = aho->states[ fail_state ].trans.base;
3254 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3256 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3257 fail[ ch_state ] = fail_state;
3258 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3260 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3262 q[ q_write++ % numstates] = ch_state;
3266 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3267 when we fail in state 1, this allows us to use the
3268 charclass scan to find a valid start char. This is based on the principle
3269 that theres a good chance the string being searched contains lots of stuff
3270 that cant be a start char.
3272 fail[ 0 ] = fail[ 1 ] = 0;
3273 DEBUG_TRIE_COMPILE_r({
3274 PerlIO_printf(Perl_debug_log,
3275 "%*sStclass Failtable (%"UVuf" states): 0",
3276 (int)(depth * 2), "", (UV)numstates
3278 for( q_read=1; q_read<numstates; q_read++ ) {
3279 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3281 PerlIO_printf(Perl_debug_log, "\n");
3284 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3289 #define DEBUG_PEEP(str,scan,depth) \
3290 DEBUG_OPTIMISE_r({if (scan){ \
3291 regnode *Next = regnext(scan); \
3292 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3293 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3294 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3295 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3296 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3297 PerlIO_printf(Perl_debug_log, "\n"); \
3300 /* The below joins as many adjacent EXACTish nodes as possible into a single
3301 * one. The regop may be changed if the node(s) contain certain sequences that
3302 * require special handling. The joining is only done if:
3303 * 1) there is room in the current conglomerated node to entirely contain the
3305 * 2) they are the exact same node type
3307 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3308 * these get optimized out
3310 * If a node is to match under /i (folded), the number of characters it matches
3311 * can be different than its character length if it contains a multi-character
3312 * fold. *min_subtract is set to the total delta number of characters of the
3315 * And *unfolded_multi_char is set to indicate whether or not the node contains
3316 * an unfolded multi-char fold. This happens when whether the fold is valid or
3317 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3318 * SMALL LETTER SHARP S, as only if the target string being matched against
3319 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3320 * folding rules depend on the locale in force at runtime. (Multi-char folds
3321 * whose components are all above the Latin1 range are not run-time locale
3322 * dependent, and have already been folded by the time this function is
3325 * This is as good a place as any to discuss the design of handling these
3326 * multi-character fold sequences. It's been wrong in Perl for a very long
3327 * time. There are three code points in Unicode whose multi-character folds
3328 * were long ago discovered to mess things up. The previous designs for
3329 * dealing with these involved assigning a special node for them. This
3330 * approach doesn't always work, as evidenced by this example:
3331 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3332 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3333 * would match just the \xDF, it won't be able to handle the case where a
3334 * successful match would have to cross the node's boundary. The new approach
3335 * that hopefully generally solves the problem generates an EXACTFU_SS node
3336 * that is "sss" in this case.
3338 * It turns out that there are problems with all multi-character folds, and not
3339 * just these three. Now the code is general, for all such cases. The
3340 * approach taken is:
3341 * 1) This routine examines each EXACTFish node that could contain multi-
3342 * character folded sequences. Since a single character can fold into
3343 * such a sequence, the minimum match length for this node is less than
3344 * the number of characters in the node. This routine returns in
3345 * *min_subtract how many characters to subtract from the the actual
3346 * length of the string to get a real minimum match length; it is 0 if
3347 * there are no multi-char foldeds. This delta is used by the caller to
3348 * adjust the min length of the match, and the delta between min and max,
3349 * so that the optimizer doesn't reject these possibilities based on size
3351 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3352 * is used for an EXACTFU node that contains at least one "ss" sequence in
3353 * it. For non-UTF-8 patterns and strings, this is the only case where
3354 * there is a possible fold length change. That means that a regular
3355 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3356 * with length changes, and so can be processed faster. regexec.c takes
3357 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3358 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3359 * known until runtime). This saves effort in regex matching. However,
3360 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3361 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3362 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3363 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3364 * possibilities for the non-UTF8 patterns are quite simple, except for
3365 * the sharp s. All the ones that don't involve a UTF-8 target string are
3366 * members of a fold-pair, and arrays are set up for all of them so that
3367 * the other member of the pair can be found quickly. Code elsewhere in
3368 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3369 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3370 * described in the next item.
3371 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3372 * validity of the fold won't be known until runtime, and so must remain
3373 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3374 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3375 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3376 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3377 * The reason this is a problem is that the optimizer part of regexec.c
3378 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3379 * that a character in the pattern corresponds to at most a single
3380 * character in the target string. (And I do mean character, and not byte
3381 * here, unlike other parts of the documentation that have never been
3382 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3383 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3384 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3385 * nodes, violate the assumption, and they are the only instances where it
3386 * is violated. I'm reluctant to try to change the assumption, as the
3387 * code involved is impenetrable to me (khw), so instead the code here
3388 * punts. This routine examines EXACTFL nodes, and (when the pattern
3389 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3390 * boolean indicating whether or not the node contains such a fold. When
3391 * it is true, the caller sets a flag that later causes the optimizer in
3392 * this file to not set values for the floating and fixed string lengths,
3393 * and thus avoids the optimizer code in regexec.c that makes the invalid
3394 * assumption. Thus, there is no optimization based on string lengths for
3395 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3396 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3397 * assumption is wrong only in these cases is that all other non-UTF-8
3398 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3399 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3400 * EXACTF nodes because we don't know at compile time if it actually
3401 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3402 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3403 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3404 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3405 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3406 * string would require the pattern to be forced into UTF-8, the overhead
3407 * of which we want to avoid. Similarly the unfolded multi-char folds in
3408 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3411 * Similarly, the code that generates tries doesn't currently handle
3412 * not-already-folded multi-char folds, and it looks like a pain to change
3413 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3414 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3415 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3416 * using /iaa matching will be doing so almost entirely with ASCII
3417 * strings, so this should rarely be encountered in practice */
3419 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3420 if (PL_regkind[OP(scan)] == EXACT) \
3421 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3424 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3425 UV *min_subtract, bool *unfolded_multi_char,
3426 U32 flags,regnode *val, U32 depth)
3428 /* Merge several consecutive EXACTish nodes into one. */
3429 regnode *n = regnext(scan);
3431 regnode *next = scan + NODE_SZ_STR(scan);
3435 regnode *stop = scan;
3436 GET_RE_DEBUG_FLAGS_DECL;
3438 PERL_UNUSED_ARG(depth);
3441 PERL_ARGS_ASSERT_JOIN_EXACT;
3442 #ifndef EXPERIMENTAL_INPLACESCAN
3443 PERL_UNUSED_ARG(flags);
3444 PERL_UNUSED_ARG(val);
3446 DEBUG_PEEP("join",scan,depth);
3448 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3449 * EXACT ones that are mergeable to the current one. */
3451 && (PL_regkind[OP(n)] == NOTHING
3452 || (stringok && OP(n) == OP(scan)))
3454 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3457 if (OP(n) == TAIL || n > next)
3459 if (PL_regkind[OP(n)] == NOTHING) {
3460 DEBUG_PEEP("skip:",n,depth);
3461 NEXT_OFF(scan) += NEXT_OFF(n);
3462 next = n + NODE_STEP_REGNODE;
3469 else if (stringok) {
3470 const unsigned int oldl = STR_LEN(scan);
3471 regnode * const nnext = regnext(n);
3473 /* XXX I (khw) kind of doubt that this works on platforms (should
3474 * Perl ever run on one) where U8_MAX is above 255 because of lots
3475 * of other assumptions */
3476 /* Don't join if the sum can't fit into a single node */
3477 if (oldl + STR_LEN(n) > U8_MAX)
3480 DEBUG_PEEP("merg",n,depth);
3483 NEXT_OFF(scan) += NEXT_OFF(n);
3484 STR_LEN(scan) += STR_LEN(n);
3485 next = n + NODE_SZ_STR(n);
3486 /* Now we can overwrite *n : */
3487 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3495 #ifdef EXPERIMENTAL_INPLACESCAN
3496 if (flags && !NEXT_OFF(n)) {
3497 DEBUG_PEEP("atch", val, depth);
3498 if (reg_off_by_arg[OP(n)]) {
3499 ARG_SET(n, val - n);
3502 NEXT_OFF(n) = val - n;
3510 *unfolded_multi_char = FALSE;
3512 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3513 * can now analyze for sequences of problematic code points. (Prior to
3514 * this final joining, sequences could have been split over boundaries, and
3515 * hence missed). The sequences only happen in folding, hence for any
3516 * non-EXACT EXACTish node */
3517 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3518 U8* s0 = (U8*) STRING(scan);
3520 U8* s_end = s0 + STR_LEN(scan);
3522 int total_count_delta = 0; /* Total delta number of characters that
3523 multi-char folds expand to */
3525 /* One pass is made over the node's string looking for all the
3526 * possibilities. To avoid some tests in the loop, there are two main
3527 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3532 if (OP(scan) == EXACTFL) {
3535 /* An EXACTFL node would already have been changed to another
3536 * node type unless there is at least one character in it that
3537 * is problematic; likely a character whose fold definition
3538 * won't be known until runtime, and so has yet to be folded.
3539 * For all but the UTF-8 locale, folds are 1-1 in length, but
3540 * to handle the UTF-8 case, we need to create a temporary
3541 * folded copy using UTF-8 locale rules in order to analyze it.
3542 * This is because our macros that look to see if a sequence is
3543 * a multi-char fold assume everything is folded (otherwise the
3544 * tests in those macros would be too complicated and slow).
3545 * Note that here, the non-problematic folds will have already
3546 * been done, so we can just copy such characters. We actually
3547 * don't completely fold the EXACTFL string. We skip the
3548 * unfolded multi-char folds, as that would just create work
3549 * below to figure out the size they already are */
3551 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3554 STRLEN s_len = UTF8SKIP(s);
3555 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3556 Copy(s, d, s_len, U8);
3559 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3560 *unfolded_multi_char = TRUE;
3561 Copy(s, d, s_len, U8);
3564 else if (isASCII(*s)) {
3565 *(d++) = toFOLD(*s);
3569 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3575 /* Point the remainder of the routine to look at our temporary
3579 } /* End of creating folded copy of EXACTFL string */
3581 /* Examine the string for a multi-character fold sequence. UTF-8
3582 * patterns have all characters pre-folded by the time this code is
3584 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3585 length sequence we are looking for is 2 */
3587 int count = 0; /* How many characters in a multi-char fold */
3588 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3589 if (! len) { /* Not a multi-char fold: get next char */
3594 /* Nodes with 'ss' require special handling, except for
3595 * EXACTFA-ish for which there is no multi-char fold to this */
3596 if (len == 2 && *s == 's' && *(s+1) == 's'
3597 && OP(scan) != EXACTFA
3598 && OP(scan) != EXACTFA_NO_TRIE)
3601 if (OP(scan) != EXACTFL) {
3602 OP(scan) = EXACTFU_SS;
3606 else { /* Here is a generic multi-char fold. */
3607 U8* multi_end = s + len;
3609 /* Count how many characters are in it. In the case of
3610 * /aa, no folds which contain ASCII code points are
3611 * allowed, so check for those, and skip if found. */
3612 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3613 count = utf8_length(s, multi_end);
3617 while (s < multi_end) {
3620 goto next_iteration;
3630 /* The delta is how long the sequence is minus 1 (1 is how long
3631 * the character that folds to the sequence is) */
3632 total_count_delta += count - 1;
3636 /* We created a temporary folded copy of the string in EXACTFL
3637 * nodes. Therefore we need to be sure it doesn't go below zero,
3638 * as the real string could be shorter */
3639 if (OP(scan) == EXACTFL) {
3640 int total_chars = utf8_length((U8*) STRING(scan),
3641 (U8*) STRING(scan) + STR_LEN(scan));
3642 if (total_count_delta > total_chars) {
3643 total_count_delta = total_chars;
3647 *min_subtract += total_count_delta;
3650 else if (OP(scan) == EXACTFA) {
3652 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3653 * fold to the ASCII range (and there are no existing ones in the
3654 * upper latin1 range). But, as outlined in the comments preceding
3655 * this function, we need to flag any occurrences of the sharp s.
3656 * This character forbids trie formation (because of added
3659 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3660 OP(scan) = EXACTFA_NO_TRIE;
3661 *unfolded_multi_char = TRUE;
3670 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3671 * folds that are all Latin1. As explained in the comments
3672 * preceding this function, we look also for the sharp s in EXACTF
3673 * and EXACTFL nodes; it can be in the final position. Otherwise
3674 * we can stop looking 1 byte earlier because have to find at least
3675 * two characters for a multi-fold */
3676 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3681 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3682 if (! len) { /* Not a multi-char fold. */
3683 if (*s == LATIN_SMALL_LETTER_SHARP_S
3684 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3686 *unfolded_multi_char = TRUE;
3693 && isALPHA_FOLD_EQ(*s, 's')
3694 && isALPHA_FOLD_EQ(*(s+1), 's'))
3697 /* EXACTF nodes need to know that the minimum length
3698 * changed so that a sharp s in the string can match this
3699 * ss in the pattern, but they remain EXACTF nodes, as they
3700 * won't match this unless the target string is is UTF-8,
3701 * which we don't know until runtime. EXACTFL nodes can't
3702 * transform into EXACTFU nodes */
3703 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3704 OP(scan) = EXACTFU_SS;
3708 *min_subtract += len - 1;
3715 /* Allow dumping but overwriting the collection of skipped
3716 * ops and/or strings with fake optimized ops */
3717 n = scan + NODE_SZ_STR(scan);
3725 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3729 /* REx optimizer. Converts nodes into quicker variants "in place".
3730 Finds fixed substrings. */
3732 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3733 to the position after last scanned or to NULL. */
3735 #define INIT_AND_WITHP \
3736 assert(!and_withp); \
3737 Newx(and_withp,1, regnode_ssc); \
3738 SAVEFREEPV(and_withp)
3742 S_unwind_scan_frames(pTHX_ const void *p)
3744 scan_frame *f= (scan_frame *)p;
3746 scan_frame *n= f->next_frame;
3754 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3755 SSize_t *minlenp, SSize_t *deltap,
3760 regnode_ssc *and_withp,
3761 U32 flags, U32 depth)
3762 /* scanp: Start here (read-write). */
3763 /* deltap: Write maxlen-minlen here. */
3764 /* last: Stop before this one. */
3765 /* data: string data about the pattern */
3766 /* stopparen: treat close N as END */
3767 /* recursed: which subroutines have we recursed into */
3768 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3770 /* There must be at least this number of characters to match */
3773 regnode *scan = *scanp, *next;
3775 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3776 int is_inf_internal = 0; /* The studied chunk is infinite */
3777 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3778 scan_data_t data_fake;
3779 SV *re_trie_maxbuff = NULL;
3780 regnode *first_non_open = scan;
3781 SSize_t stopmin = SSize_t_MAX;
3782 scan_frame *frame = NULL;
3783 GET_RE_DEBUG_FLAGS_DECL;
3785 PERL_ARGS_ASSERT_STUDY_CHUNK;
3789 while (first_non_open && OP(first_non_open) == OPEN)
3790 first_non_open=regnext(first_non_open);
3796 RExC_study_chunk_recursed_count++;
3798 DEBUG_OPTIMISE_MORE_r(
3800 PerlIO_printf(Perl_debug_log,
3801 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3802 (int)(depth*2), "", (long)stopparen,
3803 (unsigned long)RExC_study_chunk_recursed_count,
3804 (unsigned long)depth, (unsigned long)recursed_depth,
3807 if (recursed_depth) {
3810 for ( j = 0 ; j < recursed_depth ; j++ ) {
3811 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3813 PAREN_TEST(RExC_study_chunk_recursed +
3814 ( j * RExC_study_chunk_recursed_bytes), i )
3817 !PAREN_TEST(RExC_study_chunk_recursed +
3818 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3821 PerlIO_printf(Perl_debug_log," %d",(int)i);
3825 if ( j + 1 < recursed_depth ) {
3826 PerlIO_printf(Perl_debug_log, ",");
3830 PerlIO_printf(Perl_debug_log,"\n");
3833 while ( scan && OP(scan) != END && scan < last ){
3834 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3835 node length to get a real minimum (because
3836 the folded version may be shorter) */
3837 bool unfolded_multi_char = FALSE;
3838 /* Peephole optimizer: */
3839 DEBUG_STUDYDATA("Peep:", data, depth);
3840 DEBUG_PEEP("Peep", scan, depth);
3843 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3844 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3845 * by a different invocation of reg() -- Yves
3847 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3849 /* Follow the next-chain of the current node and optimize
3850 away all the NOTHINGs from it. */
3851 if (OP(scan) != CURLYX) {
3852 const int max = (reg_off_by_arg[OP(scan)]
3854 /* I32 may be smaller than U16 on CRAYs! */
3855 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3856 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3860 /* Skip NOTHING and LONGJMP. */
3861 while ((n = regnext(n))
3862 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3863 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3864 && off + noff < max)
3866 if (reg_off_by_arg[OP(scan)])
3869 NEXT_OFF(scan) = off;
3872 /* The principal pseudo-switch. Cannot be a switch, since we
3873 look into several different things. */
3874 if ( OP(scan) == DEFINEP ) {
3876 SSize_t deltanext = 0;
3877 SSize_t fake_last_close = 0;
3878 I32 f = SCF_IN_DEFINE;
3880 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3881 scan = regnext(scan);
3882 assert( OP(scan) == IFTHEN );
3883 DEBUG_PEEP("expect IFTHEN", scan, depth);
3885 data_fake.last_closep= &fake_last_close;
3887 next = regnext(scan);
3888 scan = NEXTOPER(NEXTOPER(scan));
3889 DEBUG_PEEP("scan", scan, depth);
3890 DEBUG_PEEP("next", next, depth);
3892 /* we suppose the run is continuous, last=next...
3893 * NOTE we dont use the return here! */
3894 (void)study_chunk(pRExC_state, &scan, &minlen,
3895 &deltanext, next, &data_fake, stopparen,
3896 recursed_depth, NULL, f, depth+1);
3901 OP(scan) == BRANCH ||
3902 OP(scan) == BRANCHJ ||
3905 next = regnext(scan);
3908 /* The op(next)==code check below is to see if we
3909 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3910 * IFTHEN is special as it might not appear in pairs.
3911 * Not sure whether BRANCH-BRANCHJ is possible, regardless
3912 * we dont handle it cleanly. */
3913 if (OP(next) == code || code == IFTHEN) {
3914 /* NOTE - There is similar code to this block below for
3915 * handling TRIE nodes on a re-study. If you change stuff here
3916 * check there too. */
3917 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3919 regnode * const startbranch=scan;
3921 if (flags & SCF_DO_SUBSTR) {
3922 /* Cannot merge strings after this. */
3923 scan_commit(pRExC_state, data, minlenp, is_inf);
3926 if (flags & SCF_DO_STCLASS)
3927 ssc_init_zero(pRExC_state, &accum);
3929 while (OP(scan) == code) {
3930 SSize_t deltanext, minnext, fake;
3932 regnode_ssc this_class;
3934 DEBUG_PEEP("Branch", scan, depth);
3937 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3939 data_fake.whilem_c = data->whilem_c;
3940 data_fake.last_closep = data->last_closep;
3943 data_fake.last_closep = &fake;
3945 data_fake.pos_delta = delta;
3946 next = regnext(scan);
3948 scan = NEXTOPER(scan); /* everything */
3949 if (code != BRANCH) /* everything but BRANCH */
3950 scan = NEXTOPER(scan);
3952 if (flags & SCF_DO_STCLASS) {
3953 ssc_init(pRExC_state, &this_class);
3954 data_fake.start_class = &this_class;
3955 f = SCF_DO_STCLASS_AND;
3957 if (flags & SCF_WHILEM_VISITED_POS)
3958 f |= SCF_WHILEM_VISITED_POS;
3960 /* we suppose the run is continuous, last=next...*/
3961 minnext = study_chunk(pRExC_state, &scan, minlenp,
3962 &deltanext, next, &data_fake, stopparen,
3963 recursed_depth, NULL, f,depth+1);
3967 if (deltanext == SSize_t_MAX) {
3968 is_inf = is_inf_internal = 1;
3970 } else if (max1 < minnext + deltanext)
3971 max1 = minnext + deltanext;
3973 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3975 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3976 if ( stopmin > minnext)
3977 stopmin = min + min1;
3978 flags &= ~SCF_DO_SUBSTR;
3980 data->flags |= SCF_SEEN_ACCEPT;
3983 if (data_fake.flags & SF_HAS_EVAL)
3984 data->flags |= SF_HAS_EVAL;
3985 data->whilem_c = data_fake.whilem_c;
3987 if (flags & SCF_DO_STCLASS)
3988 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3990 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3992 if (flags & SCF_DO_SUBSTR) {
3993 data->pos_min += min1;
3994 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3995 data->pos_delta = SSize_t_MAX;
3997 data->pos_delta += max1 - min1;
3998 if (max1 != min1 || is_inf)
3999 data->longest = &(data->longest_float);
4002 if (delta == SSize_t_MAX
4003 || SSize_t_MAX - delta - (max1 - min1) < 0)
4004 delta = SSize_t_MAX;
4006 delta += max1 - min1;
4007 if (flags & SCF_DO_STCLASS_OR) {
4008 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4010 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4011 flags &= ~SCF_DO_STCLASS;
4014 else if (flags & SCF_DO_STCLASS_AND) {
4016 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4017 flags &= ~SCF_DO_STCLASS;
4020 /* Switch to OR mode: cache the old value of
4021 * data->start_class */
4023 StructCopy(data->start_class, and_withp, regnode_ssc);
4024 flags &= ~SCF_DO_STCLASS_AND;
4025 StructCopy(&accum, data->start_class, regnode_ssc);
4026 flags |= SCF_DO_STCLASS_OR;
4030 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4031 OP( startbranch ) == BRANCH )
4035 Assuming this was/is a branch we are dealing with: 'scan'
4036 now points at the item that follows the branch sequence,
4037 whatever it is. We now start at the beginning of the
4038 sequence and look for subsequences of
4044 which would be constructed from a pattern like
4047 If we can find such a subsequence we need to turn the first
4048 element into a trie and then add the subsequent branch exact
4049 strings to the trie.
4053 1. patterns where the whole set of branches can be
4056 2. patterns where only a subset can be converted.
4058 In case 1 we can replace the whole set with a single regop
4059 for the trie. In case 2 we need to keep the start and end
4062 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4063 becomes BRANCH TRIE; BRANCH X;
4065 There is an additional case, that being where there is a
4066 common prefix, which gets split out into an EXACT like node
4067 preceding the TRIE node.
4069 If x(1..n)==tail then we can do a simple trie, if not we make
4070 a "jump" trie, such that when we match the appropriate word
4071 we "jump" to the appropriate tail node. Essentially we turn
4072 a nested if into a case structure of sorts.
4077 if (!re_trie_maxbuff) {
4078 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4079 if (!SvIOK(re_trie_maxbuff))
4080 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4082 if ( SvIV(re_trie_maxbuff)>=0 ) {
4084 regnode *first = (regnode *)NULL;
4085 regnode *last = (regnode *)NULL;
4086 regnode *tail = scan;
4090 /* var tail is used because there may be a TAIL
4091 regop in the way. Ie, the exacts will point to the
4092 thing following the TAIL, but the last branch will
4093 point at the TAIL. So we advance tail. If we
4094 have nested (?:) we may have to move through several
4098 while ( OP( tail ) == TAIL ) {
4099 /* this is the TAIL generated by (?:) */
4100 tail = regnext( tail );
4104 DEBUG_TRIE_COMPILE_r({
4105 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4106 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
4107 (int)depth * 2 + 2, "",
4108 "Looking for TRIE'able sequences. Tail node is: ",
4109 SvPV_nolen_const( RExC_mysv )
4115 Step through the branches
4116 cur represents each branch,
4117 noper is the first thing to be matched as part
4119 noper_next is the regnext() of that node.
4121 We normally handle a case like this
4122 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4123 support building with NOJUMPTRIE, which restricts
4124 the trie logic to structures like /FOO|BAR/.
4126 If noper is a trieable nodetype then the branch is
4127 a possible optimization target. If we are building
4128 under NOJUMPTRIE then we require that noper_next is
4129 the same as scan (our current position in the regex
4132 Once we have two or more consecutive such branches
4133 we can create a trie of the EXACT's contents and
4134 stitch it in place into the program.
4136 If the sequence represents all of the branches in
4137 the alternation we replace the entire thing with a
4140 Otherwise when it is a subsequence we need to
4141 stitch it in place and replace only the relevant
4142 branches. This means the first branch has to remain
4143 as it is used by the alternation logic, and its
4144 next pointer, and needs to be repointed at the item
4145 on the branch chain following the last branch we
4146 have optimized away.
4148 This could be either a BRANCH, in which case the
4149 subsequence is internal, or it could be the item
4150 following the branch sequence in which case the
4151 subsequence is at the end (which does not
4152 necessarily mean the first node is the start of the
4155 TRIE_TYPE(X) is a define which maps the optype to a
4159 ----------------+-----------
4163 EXACTFU_SS | EXACTFU
4166 EXACTFLU8 | EXACTFLU8
4170 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4172 : ( EXACT == (X) ) \
4174 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4176 : ( EXACTFA == (X) ) \
4178 : ( EXACTL == (X) ) \
4180 : ( EXACTFLU8 == (X) ) \
4184 /* dont use tail as the end marker for this traverse */
4185 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4186 regnode * const noper = NEXTOPER( cur );
4187 U8 noper_type = OP( noper );
4188 U8 noper_trietype = TRIE_TYPE( noper_type );
4189 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4190 regnode * const noper_next = regnext( noper );
4191 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
4192 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
4195 DEBUG_TRIE_COMPILE_r({
4196 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4197 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
4198 (int)depth * 2 + 2,"", SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4200 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4201 PerlIO_printf( Perl_debug_log, " -> %s",
4202 SvPV_nolen_const(RExC_mysv));
4205 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4206 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
4207 SvPV_nolen_const(RExC_mysv));
4209 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
4210 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4211 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4215 /* Is noper a trieable nodetype that can be merged
4216 * with the current trie (if there is one)? */
4220 ( noper_trietype == NOTHING)
4221 || ( trietype == NOTHING )
4222 || ( trietype == noper_trietype )
4225 && noper_next == tail
4229 /* Handle mergable triable node Either we are
4230 * the first node in a new trieable sequence,
4231 * in which case we do some bookkeeping,
4232 * otherwise we update the end pointer. */
4235 if ( noper_trietype == NOTHING ) {
4236 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4237 regnode * const noper_next = regnext( noper );
4238 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4239 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4242 if ( noper_next_trietype ) {
4243 trietype = noper_next_trietype;
4244 } else if (noper_next_type) {
4245 /* a NOTHING regop is 1 regop wide.
4246 * We need at least two for a trie
4247 * so we can't merge this in */
4251 trietype = noper_trietype;
4254 if ( trietype == NOTHING )
4255 trietype = noper_trietype;
4260 } /* end handle mergable triable node */
4262 /* handle unmergable node -
4263 * noper may either be a triable node which can
4264 * not be tried together with the current trie,
4265 * or a non triable node */
4267 /* If last is set and trietype is not
4268 * NOTHING then we have found at least two
4269 * triable branch sequences in a row of a
4270 * similar trietype so we can turn them
4271 * into a trie. If/when we allow NOTHING to
4272 * start a trie sequence this condition
4273 * will be required, and it isn't expensive
4274 * so we leave it in for now. */
4275 if ( trietype && trietype != NOTHING )
4276 make_trie( pRExC_state,
4277 startbranch, first, cur, tail,
4278 count, trietype, depth+1 );
4279 last = NULL; /* note: we clear/update
4280 first, trietype etc below,
4281 so we dont do it here */
4285 && noper_next == tail
4288 /* noper is triable, so we can start a new
4292 trietype = noper_trietype;
4294 /* if we already saw a first but the
4295 * current node is not triable then we have
4296 * to reset the first information. */
4301 } /* end handle unmergable node */
4302 } /* loop over branches */
4303 DEBUG_TRIE_COMPILE_r({
4304 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4305 PerlIO_printf( Perl_debug_log,
4306 "%*s- %s (%d) <SCAN FINISHED>\n",
4308 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4311 if ( last && trietype ) {
4312 if ( trietype != NOTHING ) {
4313 /* the last branch of the sequence was part of
4314 * a trie, so we have to construct it here
4315 * outside of the loop */
4316 made= make_trie( pRExC_state, startbranch,
4317 first, scan, tail, count,
4318 trietype, depth+1 );
4319 #ifdef TRIE_STUDY_OPT
4320 if ( ((made == MADE_EXACT_TRIE &&
4321 startbranch == first)
4322 || ( first_non_open == first )) &&
4324 flags |= SCF_TRIE_RESTUDY;
4325 if ( startbranch == first
4328 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4333 /* at this point we know whatever we have is a
4334 * NOTHING sequence/branch AND if 'startbranch'
4335 * is 'first' then we can turn the whole thing
4338 if ( startbranch == first ) {
4340 /* the entire thing is a NOTHING sequence,
4341 * something like this: (?:|) So we can
4342 * turn it into a plain NOTHING op. */
4343 DEBUG_TRIE_COMPILE_r({
4344 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4345 PerlIO_printf( Perl_debug_log,
4346 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4347 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4350 OP(startbranch)= NOTHING;
4351 NEXT_OFF(startbranch)= tail - startbranch;
4352 for ( opt= startbranch + 1; opt < tail ; opt++ )
4356 } /* end if ( last) */
4357 } /* TRIE_MAXBUF is non zero */
4362 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4363 scan = NEXTOPER(NEXTOPER(scan));
4364 } else /* single branch is optimized. */
4365 scan = NEXTOPER(scan);
4367 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4369 regnode *start = NULL;
4370 regnode *end = NULL;
4371 U32 my_recursed_depth= recursed_depth;
4374 if (OP(scan) != SUSPEND) { /* GOSUB/GOSTART */
4375 /* Do setup, note this code has side effects beyond
4376 * the rest of this block. Specifically setting
4377 * RExC_recurse[] must happen at least once during
4379 if (OP(scan) == GOSUB) {
4381 RExC_recurse[ARG2L(scan)] = scan;
4382 start = RExC_open_parens[paren-1];
4383 end = RExC_close_parens[paren-1];
4385 start = RExC_rxi->program + 1;
4388 /* NOTE we MUST always execute the above code, even
4389 * if we do nothing with a GOSUB/GOSTART */
4391 ( flags & SCF_IN_DEFINE )
4394 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4396 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4399 /* no need to do anything here if we are in a define. */
4400 /* or we are after some kind of infinite construct
4401 * so we can skip recursing into this item.
4402 * Since it is infinite we will not change the maxlen
4403 * or delta, and if we miss something that might raise
4404 * the minlen it will merely pessimise a little.
4406 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4407 * might result in a minlen of 1 and not of 4,
4408 * but this doesn't make us mismatch, just try a bit
4409 * harder than we should.
4411 scan= regnext(scan);
4418 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4420 /* it is quite possible that there are more efficient ways
4421 * to do this. We maintain a bitmap per level of recursion
4422 * of which patterns we have entered so we can detect if a
4423 * pattern creates a possible infinite loop. When we
4424 * recurse down a level we copy the previous levels bitmap
4425 * down. When we are at recursion level 0 we zero the top
4426 * level bitmap. It would be nice to implement a different
4427 * more efficient way of doing this. In particular the top
4428 * level bitmap may be unnecessary.
4430 if (!recursed_depth) {
4431 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4433 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4434 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4435 RExC_study_chunk_recursed_bytes, U8);
4437 /* we havent recursed into this paren yet, so recurse into it */
4438 DEBUG_STUDYDATA("set:", data,depth);
4439 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4440 my_recursed_depth= recursed_depth + 1;
4442 DEBUG_STUDYDATA("inf:", data,depth);
4443 /* some form of infinite recursion, assume infinite length
4445 if (flags & SCF_DO_SUBSTR) {
4446 scan_commit(pRExC_state, data, minlenp, is_inf);
4447 data->longest = &(data->longest_float);
4449 is_inf = is_inf_internal = 1;
4450 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4451 ssc_anything(data->start_class);
4452 flags &= ~SCF_DO_STCLASS;
4454 start= NULL; /* reset start so we dont recurse later on. */
4459 end = regnext(scan);
4462 scan_frame *newframe;
4464 if (!RExC_frame_last) {
4465 Newxz(newframe, 1, scan_frame);
4466 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4467 RExC_frame_head= newframe;
4469 } else if (!RExC_frame_last->next_frame) {
4470 Newxz(newframe,1,scan_frame);
4471 RExC_frame_last->next_frame= newframe;
4472 newframe->prev_frame= RExC_frame_last;
4475 newframe= RExC_frame_last->next_frame;
4477 RExC_frame_last= newframe;
4479 newframe->next_regnode = regnext(scan);
4480 newframe->last_regnode = last;
4481 newframe->stopparen = stopparen;
4482 newframe->prev_recursed_depth = recursed_depth;
4483 newframe->this_prev_frame= frame;
4485 DEBUG_STUDYDATA("frame-new:",data,depth);
4486 DEBUG_PEEP("fnew", scan, depth);
4493 recursed_depth= my_recursed_depth;
4498 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4499 SSize_t l = STR_LEN(scan);
4502 const U8 * const s = (U8*)STRING(scan);
4503 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4504 l = utf8_length(s, s + l);
4506 uc = *((U8*)STRING(scan));
4509 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4510 /* The code below prefers earlier match for fixed
4511 offset, later match for variable offset. */
4512 if (data->last_end == -1) { /* Update the start info. */
4513 data->last_start_min = data->pos_min;
4514 data->last_start_max = is_inf
4515 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4517 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4519 SvUTF8_on(data->last_found);
4521 SV * const sv = data->last_found;
4522 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4523 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4524 if (mg && mg->mg_len >= 0)
4525 mg->mg_len += utf8_length((U8*)STRING(scan),
4526 (U8*)STRING(scan)+STR_LEN(scan));
4528 data->last_end = data->pos_min + l;
4529 data->pos_min += l; /* As in the first entry. */
4530 data->flags &= ~SF_BEFORE_EOL;
4533 /* ANDing the code point leaves at most it, and not in locale, and
4534 * can't match null string */
4535 if (flags & SCF_DO_STCLASS_AND) {
4536 ssc_cp_and(data->start_class, uc);
4537 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4538 ssc_clear_locale(data->start_class);
4540 else if (flags & SCF_DO_STCLASS_OR) {
4541 ssc_add_cp(data->start_class, uc);
4542 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4544 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4545 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4547 flags &= ~SCF_DO_STCLASS;
4549 else if (PL_regkind[OP(scan)] == EXACT) {
4550 /* But OP != EXACT!, so is EXACTFish */
4551 SSize_t l = STR_LEN(scan);
4552 const U8 * s = (U8*)STRING(scan);
4554 /* Search for fixed substrings supports EXACT only. */
4555 if (flags & SCF_DO_SUBSTR) {
4557 scan_commit(pRExC_state, data, minlenp, is_inf);
4560 l = utf8_length(s, s + l);
4562 if (unfolded_multi_char) {
4563 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4565 min += l - min_subtract;
4567 delta += min_subtract;
4568 if (flags & SCF_DO_SUBSTR) {
4569 data->pos_min += l - min_subtract;
4570 if (data->pos_min < 0) {
4573 data->pos_delta += min_subtract;
4575 data->longest = &(data->longest_float);
4579 if (flags & SCF_DO_STCLASS) {
4580 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4582 assert(EXACTF_invlist);
4583 if (flags & SCF_DO_STCLASS_AND) {
4584 if (OP(scan) != EXACTFL)
4585 ssc_clear_locale(data->start_class);
4586 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4587 ANYOF_POSIXL_ZERO(data->start_class);
4588 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4590 else { /* SCF_DO_STCLASS_OR */
4591 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4592 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4594 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4595 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4597 flags &= ~SCF_DO_STCLASS;
4598 SvREFCNT_dec(EXACTF_invlist);
4601 else if (REGNODE_VARIES(OP(scan))) {
4602 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4603 I32 fl = 0, f = flags;
4604 regnode * const oscan = scan;
4605 regnode_ssc this_class;
4606 regnode_ssc *oclass = NULL;
4607 I32 next_is_eval = 0;
4609 switch (PL_regkind[OP(scan)]) {
4610 case WHILEM: /* End of (?:...)* . */
4611 scan = NEXTOPER(scan);
4614 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4615 next = NEXTOPER(scan);
4616 if (OP(next) == EXACT
4617 || OP(next) == EXACTL
4618 || (flags & SCF_DO_STCLASS))
4621 maxcount = REG_INFTY;
4622 next = regnext(scan);
4623 scan = NEXTOPER(scan);
4627 if (flags & SCF_DO_SUBSTR)
4632 if (flags & SCF_DO_STCLASS) {
4634 maxcount = REG_INFTY;
4635 next = regnext(scan);
4636 scan = NEXTOPER(scan);
4639 if (flags & SCF_DO_SUBSTR) {
4640 scan_commit(pRExC_state, data, minlenp, is_inf);
4641 /* Cannot extend fixed substrings */
4642 data->longest = &(data->longest_float);
4644 is_inf = is_inf_internal = 1;
4645 scan = regnext(scan);
4646 goto optimize_curly_tail;
4648 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4649 && (scan->flags == stopparen))
4654 mincount = ARG1(scan);
4655 maxcount = ARG2(scan);
4657 next = regnext(scan);
4658 if (OP(scan) == CURLYX) {
4659 I32 lp = (data ? *(data->last_closep) : 0);
4660 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4662 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4663 next_is_eval = (OP(scan) == EVAL);
4665 if (flags & SCF_DO_SUBSTR) {
4667 scan_commit(pRExC_state, data, minlenp, is_inf);
4668 /* Cannot extend fixed substrings */
4669 pos_before = data->pos_min;
4673 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4675 data->flags |= SF_IS_INF;
4677 if (flags & SCF_DO_STCLASS) {
4678 ssc_init(pRExC_state, &this_class);
4679 oclass = data->start_class;
4680 data->start_class = &this_class;
4681 f |= SCF_DO_STCLASS_AND;
4682 f &= ~SCF_DO_STCLASS_OR;
4684 /* Exclude from super-linear cache processing any {n,m}
4685 regops for which the combination of input pos and regex
4686 pos is not enough information to determine if a match
4689 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4690 regex pos at the \s*, the prospects for a match depend not
4691 only on the input position but also on how many (bar\s*)
4692 repeats into the {4,8} we are. */
4693 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4694 f &= ~SCF_WHILEM_VISITED_POS;
4696 /* This will finish on WHILEM, setting scan, or on NULL: */
4697 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4698 last, data, stopparen, recursed_depth, NULL,
4700 ? (f & ~SCF_DO_SUBSTR)
4704 if (flags & SCF_DO_STCLASS)
4705 data->start_class = oclass;
4706 if (mincount == 0 || minnext == 0) {
4707 if (flags & SCF_DO_STCLASS_OR) {
4708 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4710 else if (flags & SCF_DO_STCLASS_AND) {
4711 /* Switch to OR mode: cache the old value of
4712 * data->start_class */
4714 StructCopy(data->start_class, and_withp, regnode_ssc);
4715 flags &= ~SCF_DO_STCLASS_AND;
4716 StructCopy(&this_class, data->start_class, regnode_ssc);
4717 flags |= SCF_DO_STCLASS_OR;
4718 ANYOF_FLAGS(data->start_class)
4719 |= SSC_MATCHES_EMPTY_STRING;
4721 } else { /* Non-zero len */
4722 if (flags & SCF_DO_STCLASS_OR) {
4723 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4724 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4726 else if (flags & SCF_DO_STCLASS_AND)
4727 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4728 flags &= ~SCF_DO_STCLASS;
4730 if (!scan) /* It was not CURLYX, but CURLY. */
4732 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4733 /* ? quantifier ok, except for (?{ ... }) */
4734 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4735 && (minnext == 0) && (deltanext == 0)
4736 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4737 && maxcount <= REG_INFTY/3) /* Complement check for big
4740 /* Fatal warnings may leak the regexp without this: */
4741 SAVEFREESV(RExC_rx_sv);
4742 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
4743 "Quantifier unexpected on zero-length expression "
4744 "in regex m/%"UTF8f"/",
4745 UTF8fARG(UTF, RExC_end - RExC_precomp,
4747 (void)ReREFCNT_inc(RExC_rx_sv);
4750 min += minnext * mincount;
4751 is_inf_internal |= deltanext == SSize_t_MAX
4752 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4753 is_inf |= is_inf_internal;
4755 delta = SSize_t_MAX;
4757 delta += (minnext + deltanext) * maxcount
4758 - minnext * mincount;
4760 /* Try powerful optimization CURLYX => CURLYN. */
4761 if ( OP(oscan) == CURLYX && data
4762 && data->flags & SF_IN_PAR
4763 && !(data->flags & SF_HAS_EVAL)
4764 && !deltanext && minnext == 1 ) {
4765 /* Try to optimize to CURLYN. */
4766 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4767 regnode * const nxt1 = nxt;
4774 if (!REGNODE_SIMPLE(OP(nxt))
4775 && !(PL_regkind[OP(nxt)] == EXACT
4776 && STR_LEN(nxt) == 1))
4782 if (OP(nxt) != CLOSE)
4784 if (RExC_open_parens) {
4785 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4786 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4788 /* Now we know that nxt2 is the only contents: */
4789 oscan->flags = (U8)ARG(nxt);
4791 OP(nxt1) = NOTHING; /* was OPEN. */
4794 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4795 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4796 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4797 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4798 OP(nxt + 1) = OPTIMIZED; /* was count. */
4799 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4804 /* Try optimization CURLYX => CURLYM. */
4805 if ( OP(oscan) == CURLYX && data
4806 && !(data->flags & SF_HAS_PAR)
4807 && !(data->flags & SF_HAS_EVAL)
4808 && !deltanext /* atom is fixed width */
4809 && minnext != 0 /* CURLYM can't handle zero width */
4811 /* Nor characters whose fold at run-time may be
4812 * multi-character */
4813 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4815 /* XXXX How to optimize if data == 0? */
4816 /* Optimize to a simpler form. */
4817 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4821 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4822 && (OP(nxt2) != WHILEM))
4824 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4825 /* Need to optimize away parenths. */
4826 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4827 /* Set the parenth number. */
4828 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4830 oscan->flags = (U8)ARG(nxt);
4831 if (RExC_open_parens) {
4832 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4833 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4835 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4836 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4839 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4840 OP(nxt + 1) = OPTIMIZED; /* was count. */
4841 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4842 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4845 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4846 regnode *nnxt = regnext(nxt1);
4848 if (reg_off_by_arg[OP(nxt1)])
4849 ARG_SET(nxt1, nxt2 - nxt1);
4850 else if (nxt2 - nxt1 < U16_MAX)
4851 NEXT_OFF(nxt1) = nxt2 - nxt1;
4853 OP(nxt) = NOTHING; /* Cannot beautify */
4858 /* Optimize again: */
4859 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4860 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4865 else if ((OP(oscan) == CURLYX)
4866 && (flags & SCF_WHILEM_VISITED_POS)
4867 /* See the comment on a similar expression above.
4868 However, this time it's not a subexpression
4869 we care about, but the expression itself. */
4870 && (maxcount == REG_INFTY)
4871 && data && ++data->whilem_c < 16) {
4872 /* This stays as CURLYX, we can put the count/of pair. */
4873 /* Find WHILEM (as in regexec.c) */
4874 regnode *nxt = oscan + NEXT_OFF(oscan);
4876 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4878 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4879 | (RExC_whilem_seen << 4)); /* On WHILEM */
4881 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4883 if (flags & SCF_DO_SUBSTR) {
4884 SV *last_str = NULL;
4885 STRLEN last_chrs = 0;
4886 int counted = mincount != 0;
4888 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4890 SSize_t b = pos_before >= data->last_start_min
4891 ? pos_before : data->last_start_min;
4893 const char * const s = SvPV_const(data->last_found, l);
4894 SSize_t old = b - data->last_start_min;
4897 old = utf8_hop((U8*)s, old) - (U8*)s;
4899 /* Get the added string: */
4900 last_str = newSVpvn_utf8(s + old, l, UTF);
4901 last_chrs = UTF ? utf8_length((U8*)(s + old),
4902 (U8*)(s + old + l)) : l;
4903 if (deltanext == 0 && pos_before == b) {
4904 /* What was added is a constant string */
4907 SvGROW(last_str, (mincount * l) + 1);
4908 repeatcpy(SvPVX(last_str) + l,
4909 SvPVX_const(last_str), l,
4911 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4912 /* Add additional parts. */
4913 SvCUR_set(data->last_found,
4914 SvCUR(data->last_found) - l);
4915 sv_catsv(data->last_found, last_str);
4917 SV * sv = data->last_found;
4919 SvUTF8(sv) && SvMAGICAL(sv) ?
4920 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4921 if (mg && mg->mg_len >= 0)
4922 mg->mg_len += last_chrs * (mincount-1);
4924 last_chrs *= mincount;
4925 data->last_end += l * (mincount - 1);
4928 /* start offset must point into the last copy */
4929 data->last_start_min += minnext * (mincount - 1);
4930 data->last_start_max =
4933 : data->last_start_max +
4934 (maxcount - 1) * (minnext + data->pos_delta);
4937 /* It is counted once already... */
4938 data->pos_min += minnext * (mincount - counted);
4940 PerlIO_printf(Perl_debug_log, "counted=%"UVuf" deltanext=%"UVuf
4941 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
4942 " maxcount=%"UVuf" mincount=%"UVuf"\n",
4943 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4945 if (deltanext != SSize_t_MAX)
4946 PerlIO_printf(Perl_debug_log, "LHS=%"UVuf" RHS=%"UVuf"\n",
4947 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4948 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4950 if (deltanext == SSize_t_MAX
4951 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4952 data->pos_delta = SSize_t_MAX;
4954 data->pos_delta += - counted * deltanext +
4955 (minnext + deltanext) * maxcount - minnext * mincount;
4956 if (mincount != maxcount) {
4957 /* Cannot extend fixed substrings found inside
4959 scan_commit(pRExC_state, data, minlenp, is_inf);
4960 if (mincount && last_str) {
4961 SV * const sv = data->last_found;
4962 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4963 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4967 sv_setsv(sv, last_str);
4968 data->last_end = data->pos_min;
4969 data->last_start_min = data->pos_min - last_chrs;
4970 data->last_start_max = is_inf
4972 : data->pos_min + data->pos_delta - last_chrs;
4974 data->longest = &(data->longest_float);
4976 SvREFCNT_dec(last_str);
4978 if (data && (fl & SF_HAS_EVAL))
4979 data->flags |= SF_HAS_EVAL;
4980 optimize_curly_tail:
4981 if (OP(oscan) != CURLYX) {
4982 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4984 NEXT_OFF(oscan) += NEXT_OFF(next);
4990 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4995 if (flags & SCF_DO_SUBSTR) {
4996 /* Cannot expect anything... */
4997 scan_commit(pRExC_state, data, minlenp, is_inf);
4998 data->longest = &(data->longest_float);
5000 is_inf = is_inf_internal = 1;
5001 if (flags & SCF_DO_STCLASS_OR) {
5002 if (OP(scan) == CLUMP) {
5003 /* Actually is any start char, but very few code points
5004 * aren't start characters */
5005 ssc_match_all_cp(data->start_class);
5008 ssc_anything(data->start_class);
5011 flags &= ~SCF_DO_STCLASS;
5015 else if (OP(scan) == LNBREAK) {
5016 if (flags & SCF_DO_STCLASS) {
5017 if (flags & SCF_DO_STCLASS_AND) {
5018 ssc_intersection(data->start_class,
5019 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5020 ssc_clear_locale(data->start_class);
5021 ANYOF_FLAGS(data->start_class)
5022 &= ~SSC_MATCHES_EMPTY_STRING;
5024 else if (flags & SCF_DO_STCLASS_OR) {
5025 ssc_union(data->start_class,
5026 PL_XPosix_ptrs[_CC_VERTSPACE],
5028 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5030 /* See commit msg for
5031 * 749e076fceedeb708a624933726e7989f2302f6a */
5032 ANYOF_FLAGS(data->start_class)
5033 &= ~SSC_MATCHES_EMPTY_STRING;
5035 flags &= ~SCF_DO_STCLASS;
5038 if (delta != SSize_t_MAX)
5039 delta++; /* Because of the 2 char string cr-lf */
5040 if (flags & SCF_DO_SUBSTR) {
5041 /* Cannot expect anything... */
5042 scan_commit(pRExC_state, data, minlenp, is_inf);
5044 data->pos_delta += 1;
5045 data->longest = &(data->longest_float);
5048 else if (REGNODE_SIMPLE(OP(scan))) {
5050 if (flags & SCF_DO_SUBSTR) {
5051 scan_commit(pRExC_state, data, minlenp, is_inf);
5055 if (flags & SCF_DO_STCLASS) {
5057 SV* my_invlist = NULL;
5060 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5061 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5063 /* Some of the logic below assumes that switching
5064 locale on will only add false positives. */
5069 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5074 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5075 ssc_match_all_cp(data->start_class);
5080 SV* REG_ANY_invlist = _new_invlist(2);
5081 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5083 if (flags & SCF_DO_STCLASS_OR) {
5084 ssc_union(data->start_class,
5086 TRUE /* TRUE => invert, hence all but \n
5090 else if (flags & SCF_DO_STCLASS_AND) {
5091 ssc_intersection(data->start_class,
5093 TRUE /* TRUE => invert */
5095 ssc_clear_locale(data->start_class);
5097 SvREFCNT_dec_NN(REG_ANY_invlist);
5103 if (flags & SCF_DO_STCLASS_AND)
5104 ssc_and(pRExC_state, data->start_class,
5105 (regnode_charclass *) scan);
5107 ssc_or(pRExC_state, data->start_class,
5108 (regnode_charclass *) scan);
5116 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5117 if (flags & SCF_DO_STCLASS_AND) {
5118 bool was_there = cBOOL(
5119 ANYOF_POSIXL_TEST(data->start_class,
5121 ANYOF_POSIXL_ZERO(data->start_class);
5122 if (was_there) { /* Do an AND */
5123 ANYOF_POSIXL_SET(data->start_class, namedclass);
5125 /* No individual code points can now match */
5126 data->start_class->invlist
5127 = sv_2mortal(_new_invlist(0));
5130 int complement = namedclass + ((invert) ? -1 : 1);
5132 assert(flags & SCF_DO_STCLASS_OR);
5134 /* If the complement of this class was already there,
5135 * the result is that they match all code points,
5136 * (\d + \D == everything). Remove the classes from
5137 * future consideration. Locale is not relevant in
5139 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5140 ssc_match_all_cp(data->start_class);
5141 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5142 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5144 else { /* The usual case; just add this class to the
5146 ANYOF_POSIXL_SET(data->start_class, namedclass);
5151 case NPOSIXA: /* For these, we always know the exact set of
5156 if (FLAGS(scan) == _CC_ASCII) {
5157 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5160 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5161 PL_XPosix_ptrs[_CC_ASCII],
5172 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5174 /* NPOSIXD matches all upper Latin1 code points unless the
5175 * target string being matched is UTF-8, which is
5176 * unknowable until match time. Since we are going to
5177 * invert, we want to get rid of all of them so that the
5178 * inversion will match all */
5179 if (OP(scan) == NPOSIXD) {
5180 _invlist_subtract(my_invlist, PL_UpperLatin1,
5186 if (flags & SCF_DO_STCLASS_AND) {
5187 ssc_intersection(data->start_class, my_invlist, invert);
5188 ssc_clear_locale(data->start_class);
5191 assert(flags & SCF_DO_STCLASS_OR);
5192 ssc_union(data->start_class, my_invlist, invert);
5194 SvREFCNT_dec(my_invlist);
5196 if (flags & SCF_DO_STCLASS_OR)
5197 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5198 flags &= ~SCF_DO_STCLASS;
5201 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5202 data->flags |= (OP(scan) == MEOL
5205 scan_commit(pRExC_state, data, minlenp, is_inf);
5208 else if ( PL_regkind[OP(scan)] == BRANCHJ
5209 /* Lookbehind, or need to calculate parens/evals/stclass: */
5210 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5211 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5213 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5214 || OP(scan) == UNLESSM )
5216 /* Negative Lookahead/lookbehind
5217 In this case we can't do fixed string optimisation.
5220 SSize_t deltanext, minnext, fake = 0;
5225 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5227 data_fake.whilem_c = data->whilem_c;
5228 data_fake.last_closep = data->last_closep;
5231 data_fake.last_closep = &fake;
5232 data_fake.pos_delta = delta;
5233 if ( flags & SCF_DO_STCLASS && !scan->flags
5234 && OP(scan) == IFMATCH ) { /* Lookahead */
5235 ssc_init(pRExC_state, &intrnl);
5236 data_fake.start_class = &intrnl;
5237 f |= SCF_DO_STCLASS_AND;
5239 if (flags & SCF_WHILEM_VISITED_POS)
5240 f |= SCF_WHILEM_VISITED_POS;
5241 next = regnext(scan);
5242 nscan = NEXTOPER(NEXTOPER(scan));
5243 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5244 last, &data_fake, stopparen,
5245 recursed_depth, NULL, f, depth+1);
5248 FAIL("Variable length lookbehind not implemented");
5250 else if (minnext > (I32)U8_MAX) {
5251 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5254 scan->flags = (U8)minnext;
5257 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5259 if (data_fake.flags & SF_HAS_EVAL)
5260 data->flags |= SF_HAS_EVAL;
5261 data->whilem_c = data_fake.whilem_c;
5263 if (f & SCF_DO_STCLASS_AND) {
5264 if (flags & SCF_DO_STCLASS_OR) {
5265 /* OR before, AND after: ideally we would recurse with
5266 * data_fake to get the AND applied by study of the
5267 * remainder of the pattern, and then derecurse;
5268 * *** HACK *** for now just treat as "no information".
5269 * See [perl #56690].
5271 ssc_init(pRExC_state, data->start_class);
5273 /* AND before and after: combine and continue. These
5274 * assertions are zero-length, so can match an EMPTY
5276 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5277 ANYOF_FLAGS(data->start_class)
5278 |= SSC_MATCHES_EMPTY_STRING;
5282 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5284 /* Positive Lookahead/lookbehind
5285 In this case we can do fixed string optimisation,
5286 but we must be careful about it. Note in the case of
5287 lookbehind the positions will be offset by the minimum
5288 length of the pattern, something we won't know about
5289 until after the recurse.
5291 SSize_t deltanext, fake = 0;
5295 /* We use SAVEFREEPV so that when the full compile
5296 is finished perl will clean up the allocated
5297 minlens when it's all done. This way we don't
5298 have to worry about freeing them when we know
5299 they wont be used, which would be a pain.
5302 Newx( minnextp, 1, SSize_t );
5303 SAVEFREEPV(minnextp);
5306 StructCopy(data, &data_fake, scan_data_t);
5307 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5310 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5311 data_fake.last_found=newSVsv(data->last_found);
5315 data_fake.last_closep = &fake;
5316 data_fake.flags = 0;
5317 data_fake.pos_delta = delta;
5319 data_fake.flags |= SF_IS_INF;
5320 if ( flags & SCF_DO_STCLASS && !scan->flags
5321 && OP(scan) == IFMATCH ) { /* Lookahead */
5322 ssc_init(pRExC_state, &intrnl);
5323 data_fake.start_class = &intrnl;
5324 f |= SCF_DO_STCLASS_AND;
5326 if (flags & SCF_WHILEM_VISITED_POS)
5327 f |= SCF_WHILEM_VISITED_POS;
5328 next = regnext(scan);
5329 nscan = NEXTOPER(NEXTOPER(scan));
5331 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5332 &deltanext, last, &data_fake,
5333 stopparen, recursed_depth, NULL,
5337 FAIL("Variable length lookbehind not implemented");
5339 else if (*minnextp > (I32)U8_MAX) {
5340 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5343 scan->flags = (U8)*minnextp;
5348 if (f & SCF_DO_STCLASS_AND) {
5349 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5350 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5353 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5355 if (data_fake.flags & SF_HAS_EVAL)
5356 data->flags |= SF_HAS_EVAL;
5357 data->whilem_c = data_fake.whilem_c;
5358 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5359 if (RExC_rx->minlen<*minnextp)
5360 RExC_rx->minlen=*minnextp;
5361 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5362 SvREFCNT_dec_NN(data_fake.last_found);
5364 if ( data_fake.minlen_fixed != minlenp )
5366 data->offset_fixed= data_fake.offset_fixed;
5367 data->minlen_fixed= data_fake.minlen_fixed;
5368 data->lookbehind_fixed+= scan->flags;
5370 if ( data_fake.minlen_float != minlenp )
5372 data->minlen_float= data_fake.minlen_float;
5373 data->offset_float_min=data_fake.offset_float_min;
5374 data->offset_float_max=data_fake.offset_float_max;
5375 data->lookbehind_float+= scan->flags;
5382 else if (OP(scan) == OPEN) {
5383 if (stopparen != (I32)ARG(scan))
5386 else if (OP(scan) == CLOSE) {
5387 if (stopparen == (I32)ARG(scan)) {
5390 if ((I32)ARG(scan) == is_par) {
5391 next = regnext(scan);
5393 if ( next && (OP(next) != WHILEM) && next < last)
5394 is_par = 0; /* Disable optimization */
5397 *(data->last_closep) = ARG(scan);
5399 else if (OP(scan) == EVAL) {
5401 data->flags |= SF_HAS_EVAL;
5403 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5404 if (flags & SCF_DO_SUBSTR) {
5405 scan_commit(pRExC_state, data, minlenp, is_inf);
5406 flags &= ~SCF_DO_SUBSTR;
5408 if (data && OP(scan)==ACCEPT) {
5409 data->flags |= SCF_SEEN_ACCEPT;
5414 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5416 if (flags & SCF_DO_SUBSTR) {
5417 scan_commit(pRExC_state, data, minlenp, is_inf);
5418 data->longest = &(data->longest_float);
5420 is_inf = is_inf_internal = 1;
5421 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5422 ssc_anything(data->start_class);
5423 flags &= ~SCF_DO_STCLASS;
5425 else if (OP(scan) == GPOS) {
5426 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5427 !(delta || is_inf || (data && data->pos_delta)))
5429 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5430 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5431 if (RExC_rx->gofs < (STRLEN)min)
5432 RExC_rx->gofs = min;
5434 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5438 #ifdef TRIE_STUDY_OPT
5439 #ifdef FULL_TRIE_STUDY
5440 else if (PL_regkind[OP(scan)] == TRIE) {
5441 /* NOTE - There is similar code to this block above for handling
5442 BRANCH nodes on the initial study. If you change stuff here
5444 regnode *trie_node= scan;
5445 regnode *tail= regnext(scan);
5446 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5447 SSize_t max1 = 0, min1 = SSize_t_MAX;
5450 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5451 /* Cannot merge strings after this. */
5452 scan_commit(pRExC_state, data, minlenp, is_inf);
5454 if (flags & SCF_DO_STCLASS)
5455 ssc_init_zero(pRExC_state, &accum);
5461 const regnode *nextbranch= NULL;
5464 for ( word=1 ; word <= trie->wordcount ; word++)
5466 SSize_t deltanext=0, minnext=0, f = 0, fake;
5467 regnode_ssc this_class;
5469 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5471 data_fake.whilem_c = data->whilem_c;
5472 data_fake.last_closep = data->last_closep;
5475 data_fake.last_closep = &fake;
5476 data_fake.pos_delta = delta;
5477 if (flags & SCF_DO_STCLASS) {
5478 ssc_init(pRExC_state, &this_class);
5479 data_fake.start_class = &this_class;
5480 f = SCF_DO_STCLASS_AND;
5482 if (flags & SCF_WHILEM_VISITED_POS)
5483 f |= SCF_WHILEM_VISITED_POS;
5485 if (trie->jump[word]) {
5487 nextbranch = trie_node + trie->jump[0];
5488 scan= trie_node + trie->jump[word];
5489 /* We go from the jump point to the branch that follows
5490 it. Note this means we need the vestigal unused
5491 branches even though they arent otherwise used. */
5492 minnext = study_chunk(pRExC_state, &scan, minlenp,
5493 &deltanext, (regnode *)nextbranch, &data_fake,
5494 stopparen, recursed_depth, NULL, f,depth+1);
5496 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5497 nextbranch= regnext((regnode*)nextbranch);
5499 if (min1 > (SSize_t)(minnext + trie->minlen))
5500 min1 = minnext + trie->minlen;
5501 if (deltanext == SSize_t_MAX) {
5502 is_inf = is_inf_internal = 1;
5504 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5505 max1 = minnext + deltanext + trie->maxlen;
5507 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5509 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5510 if ( stopmin > min + min1)
5511 stopmin = min + min1;
5512 flags &= ~SCF_DO_SUBSTR;
5514 data->flags |= SCF_SEEN_ACCEPT;
5517 if (data_fake.flags & SF_HAS_EVAL)
5518 data->flags |= SF_HAS_EVAL;
5519 data->whilem_c = data_fake.whilem_c;
5521 if (flags & SCF_DO_STCLASS)
5522 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5525 if (flags & SCF_DO_SUBSTR) {
5526 data->pos_min += min1;
5527 data->pos_delta += max1 - min1;
5528 if (max1 != min1 || is_inf)
5529 data->longest = &(data->longest_float);
5532 if (delta != SSize_t_MAX)
5533 delta += max1 - min1;
5534 if (flags & SCF_DO_STCLASS_OR) {
5535 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5537 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5538 flags &= ~SCF_DO_STCLASS;
5541 else if (flags & SCF_DO_STCLASS_AND) {
5543 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5544 flags &= ~SCF_DO_STCLASS;
5547 /* Switch to OR mode: cache the old value of
5548 * data->start_class */
5550 StructCopy(data->start_class, and_withp, regnode_ssc);
5551 flags &= ~SCF_DO_STCLASS_AND;
5552 StructCopy(&accum, data->start_class, regnode_ssc);
5553 flags |= SCF_DO_STCLASS_OR;
5560 else if (PL_regkind[OP(scan)] == TRIE) {
5561 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5564 min += trie->minlen;
5565 delta += (trie->maxlen - trie->minlen);
5566 flags &= ~SCF_DO_STCLASS; /* xxx */
5567 if (flags & SCF_DO_SUBSTR) {
5568 /* Cannot expect anything... */
5569 scan_commit(pRExC_state, data, minlenp, is_inf);
5570 data->pos_min += trie->minlen;
5571 data->pos_delta += (trie->maxlen - trie->minlen);
5572 if (trie->maxlen != trie->minlen)
5573 data->longest = &(data->longest_float);
5575 if (trie->jump) /* no more substrings -- for now /grr*/
5576 flags &= ~SCF_DO_SUBSTR;
5578 #endif /* old or new */
5579 #endif /* TRIE_STUDY_OPT */
5581 /* Else: zero-length, ignore. */
5582 scan = regnext(scan);
5584 /* If we are exiting a recursion we can unset its recursed bit
5585 * and allow ourselves to enter it again - no danger of an
5586 * infinite loop there.
5587 if (stopparen > -1 && recursed) {
5588 DEBUG_STUDYDATA("unset:", data,depth);
5589 PAREN_UNSET( recursed, stopparen);
5595 DEBUG_STUDYDATA("frame-end:",data,depth);
5596 DEBUG_PEEP("fend", scan, depth);
5598 /* restore previous context */
5599 last = frame->last_regnode;
5600 scan = frame->next_regnode;
5601 stopparen = frame->stopparen;
5602 recursed_depth = frame->prev_recursed_depth;
5604 RExC_frame_last = frame->prev_frame;
5605 frame = frame->this_prev_frame;
5606 goto fake_study_recurse;
5611 DEBUG_STUDYDATA("pre-fin:",data,depth);
5614 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5616 if (flags & SCF_DO_SUBSTR && is_inf)
5617 data->pos_delta = SSize_t_MAX - data->pos_min;
5618 if (is_par > (I32)U8_MAX)
5620 if (is_par && pars==1 && data) {
5621 data->flags |= SF_IN_PAR;
5622 data->flags &= ~SF_HAS_PAR;
5624 else if (pars && data) {
5625 data->flags |= SF_HAS_PAR;
5626 data->flags &= ~SF_IN_PAR;
5628 if (flags & SCF_DO_STCLASS_OR)
5629 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5630 if (flags & SCF_TRIE_RESTUDY)
5631 data->flags |= SCF_TRIE_RESTUDY;
5633 DEBUG_STUDYDATA("post-fin:",data,depth);
5636 SSize_t final_minlen= min < stopmin ? min : stopmin;
5638 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5639 if (final_minlen > SSize_t_MAX - delta)
5640 RExC_maxlen = SSize_t_MAX;
5641 else if (RExC_maxlen < final_minlen + delta)
5642 RExC_maxlen = final_minlen + delta;
5644 return final_minlen;
5646 NOT_REACHED; /* NOTREACHED */
5650 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5652 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5654 PERL_ARGS_ASSERT_ADD_DATA;
5656 Renewc(RExC_rxi->data,
5657 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5658 char, struct reg_data);
5660 Renew(RExC_rxi->data->what, count + n, U8);
5662 Newx(RExC_rxi->data->what, n, U8);
5663 RExC_rxi->data->count = count + n;
5664 Copy(s, RExC_rxi->data->what + count, n, U8);
5668 /*XXX: todo make this not included in a non debugging perl, but appears to be
5669 * used anyway there, in 'use re' */
5670 #ifndef PERL_IN_XSUB_RE
5672 Perl_reginitcolors(pTHX)
5674 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5676 char *t = savepv(s);
5680 t = strchr(t, '\t');
5686 PL_colors[i] = t = (char *)"";
5691 PL_colors[i++] = (char *)"";
5698 #ifdef TRIE_STUDY_OPT
5699 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5702 (data.flags & SCF_TRIE_RESTUDY) \
5710 #define CHECK_RESTUDY_GOTO_butfirst
5714 * pregcomp - compile a regular expression into internal code
5716 * Decides which engine's compiler to call based on the hint currently in
5720 #ifndef PERL_IN_XSUB_RE
5722 /* return the currently in-scope regex engine (or the default if none) */
5724 regexp_engine const *
5725 Perl_current_re_engine(pTHX)
5727 if (IN_PERL_COMPILETIME) {
5728 HV * const table = GvHV(PL_hintgv);
5731 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5732 return &PL_core_reg_engine;
5733 ptr = hv_fetchs(table, "regcomp", FALSE);
5734 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5735 return &PL_core_reg_engine;
5736 return INT2PTR(regexp_engine*,SvIV(*ptr));
5740 if (!PL_curcop->cop_hints_hash)
5741 return &PL_core_reg_engine;
5742 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5743 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5744 return &PL_core_reg_engine;
5745 return INT2PTR(regexp_engine*,SvIV(ptr));
5751 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5753 regexp_engine const *eng = current_re_engine();
5754 GET_RE_DEBUG_FLAGS_DECL;
5756 PERL_ARGS_ASSERT_PREGCOMP;
5758 /* Dispatch a request to compile a regexp to correct regexp engine. */
5760 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5763 return CALLREGCOMP_ENG(eng, pattern, flags);
5767 /* public(ish) entry point for the perl core's own regex compiling code.
5768 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5769 * pattern rather than a list of OPs, and uses the internal engine rather
5770 * than the current one */
5773 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5775 SV *pat = pattern; /* defeat constness! */
5776 PERL_ARGS_ASSERT_RE_COMPILE;
5777 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5778 #ifdef PERL_IN_XSUB_RE
5781 &PL_core_reg_engine,
5783 NULL, NULL, rx_flags, 0);
5787 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5788 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5789 * point to the realloced string and length.
5791 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5795 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5796 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5798 U8 *const src = (U8*)*pat_p;
5803 GET_RE_DEBUG_FLAGS_DECL;
5805 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5806 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5808 Newx(dst, *plen_p * 2 + 1, U8);
5811 while (s < *plen_p) {
5812 append_utf8_from_native_byte(src[s], &d);
5813 if (n < num_code_blocks) {
5814 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5815 pRExC_state->code_blocks[n].start = d - dst - 1;
5816 assert(*(d - 1) == '(');
5819 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5820 pRExC_state->code_blocks[n].end = d - dst - 1;
5821 assert(*(d - 1) == ')');
5830 *pat_p = (char*) dst;
5832 RExC_orig_utf8 = RExC_utf8 = 1;
5837 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5838 * while recording any code block indices, and handling overloading,
5839 * nested qr// objects etc. If pat is null, it will allocate a new
5840 * string, or just return the first arg, if there's only one.
5842 * Returns the malloced/updated pat.
5843 * patternp and pat_count is the array of SVs to be concatted;
5844 * oplist is the optional list of ops that generated the SVs;
5845 * recompile_p is a pointer to a boolean that will be set if
5846 * the regex will need to be recompiled.
5847 * delim, if non-null is an SV that will be inserted between each element
5851 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5852 SV *pat, SV ** const patternp, int pat_count,
5853 OP *oplist, bool *recompile_p, SV *delim)
5857 bool use_delim = FALSE;
5858 bool alloced = FALSE;
5860 /* if we know we have at least two args, create an empty string,
5861 * then concatenate args to that. For no args, return an empty string */
5862 if (!pat && pat_count != 1) {
5868 for (svp = patternp; svp < patternp + pat_count; svp++) {
5871 STRLEN orig_patlen = 0;
5873 SV *msv = use_delim ? delim : *svp;
5874 if (!msv) msv = &PL_sv_undef;
5876 /* if we've got a delimiter, we go round the loop twice for each
5877 * svp slot (except the last), using the delimiter the second
5886 if (SvTYPE(msv) == SVt_PVAV) {
5887 /* we've encountered an interpolated array within
5888 * the pattern, e.g. /...@a..../. Expand the list of elements,
5889 * then recursively append elements.
5890 * The code in this block is based on S_pushav() */
5892 AV *const av = (AV*)msv;
5893 const SSize_t maxarg = AvFILL(av) + 1;
5897 assert(oplist->op_type == OP_PADAV
5898 || oplist->op_type == OP_RV2AV);
5899 oplist = OpSIBLING(oplist);
5902 if (SvRMAGICAL(av)) {
5905 Newx(array, maxarg, SV*);
5907 for (i=0; i < maxarg; i++) {
5908 SV ** const svp = av_fetch(av, i, FALSE);
5909 array[i] = svp ? *svp : &PL_sv_undef;
5913 array = AvARRAY(av);
5915 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5916 array, maxarg, NULL, recompile_p,
5918 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5924 /* we make the assumption here that each op in the list of
5925 * op_siblings maps to one SV pushed onto the stack,
5926 * except for code blocks, with have both an OP_NULL and
5928 * This allows us to match up the list of SVs against the
5929 * list of OPs to find the next code block.
5931 * Note that PUSHMARK PADSV PADSV ..
5933 * PADRANGE PADSV PADSV ..
5934 * so the alignment still works. */
5937 if (oplist->op_type == OP_NULL
5938 && (oplist->op_flags & OPf_SPECIAL))
5940 assert(n < pRExC_state->num_code_blocks);
5941 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5942 pRExC_state->code_blocks[n].block = oplist;
5943 pRExC_state->code_blocks[n].src_regex = NULL;
5946 oplist = OpSIBLING(oplist); /* skip CONST */
5949 oplist = OpSIBLING(oplist);;
5952 /* apply magic and QR overloading to arg */
5955 if (SvROK(msv) && SvAMAGIC(msv)) {
5956 SV *sv = AMG_CALLunary(msv, regexp_amg);
5960 if (SvTYPE(sv) != SVt_REGEXP)
5961 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5966 /* try concatenation overload ... */
5967 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5968 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5971 /* overloading involved: all bets are off over literal
5972 * code. Pretend we haven't seen it */
5973 pRExC_state->num_code_blocks -= n;
5977 /* ... or failing that, try "" overload */
5978 while (SvAMAGIC(msv)
5979 && (sv = AMG_CALLunary(msv, string_amg))
5983 && SvRV(msv) == SvRV(sv))
5988 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5992 /* this is a partially unrolled
5993 * sv_catsv_nomg(pat, msv);
5994 * that allows us to adjust code block indices if
5997 char *dst = SvPV_force_nomg(pat, dlen);
5999 if (SvUTF8(msv) && !SvUTF8(pat)) {
6000 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6001 sv_setpvn(pat, dst, dlen);
6004 sv_catsv_nomg(pat, msv);
6011 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6014 /* extract any code blocks within any embedded qr//'s */
6015 if (rx && SvTYPE(rx) == SVt_REGEXP
6016 && RX_ENGINE((REGEXP*)rx)->op_comp)
6019 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6020 if (ri->num_code_blocks) {
6022 /* the presence of an embedded qr// with code means
6023 * we should always recompile: the text of the
6024 * qr// may not have changed, but it may be a
6025 * different closure than last time */
6027 Renew(pRExC_state->code_blocks,
6028 pRExC_state->num_code_blocks + ri->num_code_blocks,
6029 struct reg_code_block);
6030 pRExC_state->num_code_blocks += ri->num_code_blocks;
6032 for (i=0; i < ri->num_code_blocks; i++) {
6033 struct reg_code_block *src, *dst;
6034 STRLEN offset = orig_patlen
6035 + ReANY((REGEXP *)rx)->pre_prefix;
6036 assert(n < pRExC_state->num_code_blocks);
6037 src = &ri->code_blocks[i];
6038 dst = &pRExC_state->code_blocks[n];
6039 dst->start = src->start + offset;
6040 dst->end = src->end + offset;
6041 dst->block = src->block;
6042 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6051 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6060 /* see if there are any run-time code blocks in the pattern.
6061 * False positives are allowed */
6064 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6065 char *pat, STRLEN plen)
6070 PERL_UNUSED_CONTEXT;
6072 for (s = 0; s < plen; s++) {
6073 if (n < pRExC_state->num_code_blocks
6074 && s == pRExC_state->code_blocks[n].start)
6076 s = pRExC_state->code_blocks[n].end;
6080 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6082 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6084 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6091 /* Handle run-time code blocks. We will already have compiled any direct
6092 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6093 * copy of it, but with any literal code blocks blanked out and
6094 * appropriate chars escaped; then feed it into
6096 * eval "qr'modified_pattern'"
6100 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6104 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6106 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6107 * and merge them with any code blocks of the original regexp.
6109 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6110 * instead, just save the qr and return FALSE; this tells our caller that
6111 * the original pattern needs upgrading to utf8.
6115 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6116 char *pat, STRLEN plen)
6120 GET_RE_DEBUG_FLAGS_DECL;
6122 if (pRExC_state->runtime_code_qr) {
6123 /* this is the second time we've been called; this should
6124 * only happen if the main pattern got upgraded to utf8
6125 * during compilation; re-use the qr we compiled first time
6126 * round (which should be utf8 too)
6128 qr = pRExC_state->runtime_code_qr;
6129 pRExC_state->runtime_code_qr = NULL;
6130 assert(RExC_utf8 && SvUTF8(qr));
6136 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6140 /* determine how many extra chars we need for ' and \ escaping */
6141 for (s = 0; s < plen; s++) {
6142 if (pat[s] == '\'' || pat[s] == '\\')
6146 Newx(newpat, newlen, char);
6148 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6150 for (s = 0; s < plen; s++) {
6151 if (n < pRExC_state->num_code_blocks
6152 && s == pRExC_state->code_blocks[n].start)
6154 /* blank out literal code block */
6155 assert(pat[s] == '(');
6156 while (s <= pRExC_state->code_blocks[n].end) {
6164 if (pat[s] == '\'' || pat[s] == '\\')
6169 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6173 PerlIO_printf(Perl_debug_log,
6174 "%sre-parsing pattern for runtime code:%s %s\n",
6175 PL_colors[4],PL_colors[5],newpat);
6178 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6183 PUSHSTACKi(PERLSI_REQUIRE);
6184 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6185 * parsing qr''; normally only q'' does this. It also alters
6187 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6188 SvREFCNT_dec_NN(sv);
6193 SV * const errsv = ERRSV;
6194 if (SvTRUE_NN(errsv))
6196 Safefree(pRExC_state->code_blocks);
6197 /* use croak_sv ? */
6198 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6201 assert(SvROK(qr_ref));
6203 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6204 /* the leaving below frees the tmp qr_ref.
6205 * Give qr a life of its own */
6213 if (!RExC_utf8 && SvUTF8(qr)) {
6214 /* first time through; the pattern got upgraded; save the
6215 * qr for the next time through */
6216 assert(!pRExC_state->runtime_code_qr);
6217 pRExC_state->runtime_code_qr = qr;
6222 /* extract any code blocks within the returned qr// */
6225 /* merge the main (r1) and run-time (r2) code blocks into one */
6227 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6228 struct reg_code_block *new_block, *dst;
6229 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6232 if (!r2->num_code_blocks) /* we guessed wrong */
6234 SvREFCNT_dec_NN(qr);
6239 r1->num_code_blocks + r2->num_code_blocks,
6240 struct reg_code_block);
6243 while ( i1 < r1->num_code_blocks
6244 || i2 < r2->num_code_blocks)
6246 struct reg_code_block *src;
6249 if (i1 == r1->num_code_blocks) {
6250 src = &r2->code_blocks[i2++];
6253 else if (i2 == r2->num_code_blocks)
6254 src = &r1->code_blocks[i1++];
6255 else if ( r1->code_blocks[i1].start
6256 < r2->code_blocks[i2].start)
6258 src = &r1->code_blocks[i1++];
6259 assert(src->end < r2->code_blocks[i2].start);
6262 assert( r1->code_blocks[i1].start
6263 > r2->code_blocks[i2].start);
6264 src = &r2->code_blocks[i2++];
6266 assert(src->end < r1->code_blocks[i1].start);
6269 assert(pat[src->start] == '(');
6270 assert(pat[src->end] == ')');
6271 dst->start = src->start;
6272 dst->end = src->end;
6273 dst->block = src->block;
6274 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6278 r1->num_code_blocks += r2->num_code_blocks;
6279 Safefree(r1->code_blocks);
6280 r1->code_blocks = new_block;
6283 SvREFCNT_dec_NN(qr);
6289 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6290 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6291 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6292 STRLEN longest_length, bool eol, bool meol)
6294 /* This is the common code for setting up the floating and fixed length
6295 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6296 * as to whether succeeded or not */
6301 if (! (longest_length
6302 || (eol /* Can't have SEOL and MULTI */
6303 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6305 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6306 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6311 /* copy the information about the longest from the reg_scan_data
6312 over to the program. */
6313 if (SvUTF8(sv_longest)) {
6314 *rx_utf8 = sv_longest;
6317 *rx_substr = sv_longest;
6320 /* end_shift is how many chars that must be matched that
6321 follow this item. We calculate it ahead of time as once the
6322 lookbehind offset is added in we lose the ability to correctly
6324 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6325 *rx_end_shift = ml - offset
6326 - longest_length + (SvTAIL(sv_longest) != 0)
6329 t = (eol/* Can't have SEOL and MULTI */
6330 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6331 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6337 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6338 * regular expression into internal code.
6339 * The pattern may be passed either as:
6340 * a list of SVs (patternp plus pat_count)
6341 * a list of OPs (expr)
6342 * If both are passed, the SV list is used, but the OP list indicates
6343 * which SVs are actually pre-compiled code blocks
6345 * The SVs in the list have magic and qr overloading applied to them (and
6346 * the list may be modified in-place with replacement SVs in the latter
6349 * If the pattern hasn't changed from old_re, then old_re will be
6352 * eng is the current engine. If that engine has an op_comp method, then
6353 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6354 * do the initial concatenation of arguments and pass on to the external
6357 * If is_bare_re is not null, set it to a boolean indicating whether the
6358 * arg list reduced (after overloading) to a single bare regex which has
6359 * been returned (i.e. /$qr/).
6361 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6363 * pm_flags contains the PMf_* flags, typically based on those from the
6364 * pm_flags field of the related PMOP. Currently we're only interested in
6365 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6367 * We can't allocate space until we know how big the compiled form will be,
6368 * but we can't compile it (and thus know how big it is) until we've got a
6369 * place to put the code. So we cheat: we compile it twice, once with code
6370 * generation turned off and size counting turned on, and once "for real".
6371 * This also means that we don't allocate space until we are sure that the
6372 * thing really will compile successfully, and we never have to move the
6373 * code and thus invalidate pointers into it. (Note that it has to be in
6374 * one piece because free() must be able to free it all.) [NB: not true in perl]
6376 * Beware that the optimization-preparation code in here knows about some
6377 * of the structure of the compiled regexp. [I'll say.]
6381 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6382 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6383 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6387 regexp_internal *ri;
6395 SV *code_blocksv = NULL;
6396 SV** new_patternp = patternp;
6398 /* these are all flags - maybe they should be turned
6399 * into a single int with different bit masks */
6400 I32 sawlookahead = 0;
6405 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6407 bool runtime_code = 0;
6409 RExC_state_t RExC_state;
6410 RExC_state_t * const pRExC_state = &RExC_state;
6411 #ifdef TRIE_STUDY_OPT
6413 RExC_state_t copyRExC_state;
6415 GET_RE_DEBUG_FLAGS_DECL;
6417 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6419 DEBUG_r(if (!PL_colorset) reginitcolors());
6421 /* Initialize these here instead of as-needed, as is quick and avoids
6422 * having to test them each time otherwise */
6423 if (! PL_AboveLatin1) {
6424 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6425 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6426 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6427 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6428 PL_HasMultiCharFold =
6429 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6431 /* This is calculated here, because the Perl program that generates the
6432 * static global ones doesn't currently have access to
6433 * NUM_ANYOF_CODE_POINTS */
6434 PL_InBitmap = _new_invlist(2);
6435 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6436 NUM_ANYOF_CODE_POINTS - 1);
6439 pRExC_state->code_blocks = NULL;
6440 pRExC_state->num_code_blocks = 0;
6443 *is_bare_re = FALSE;
6445 if (expr && (expr->op_type == OP_LIST ||
6446 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6447 /* allocate code_blocks if needed */
6451 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6452 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6453 ncode++; /* count of DO blocks */
6455 pRExC_state->num_code_blocks = ncode;
6456 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6461 /* compile-time pattern with just OP_CONSTs and DO blocks */
6466 /* find how many CONSTs there are */
6469 if (expr->op_type == OP_CONST)
6472 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6473 if (o->op_type == OP_CONST)
6477 /* fake up an SV array */
6479 assert(!new_patternp);
6480 Newx(new_patternp, n, SV*);
6481 SAVEFREEPV(new_patternp);
6485 if (expr->op_type == OP_CONST)
6486 new_patternp[n] = cSVOPx_sv(expr);
6488 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6489 if (o->op_type == OP_CONST)
6490 new_patternp[n++] = cSVOPo_sv;
6495 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6496 "Assembling pattern from %d elements%s\n", pat_count,
6497 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6499 /* set expr to the first arg op */
6501 if (pRExC_state->num_code_blocks
6502 && expr->op_type != OP_CONST)
6504 expr = cLISTOPx(expr)->op_first;
6505 assert( expr->op_type == OP_PUSHMARK
6506 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6507 || expr->op_type == OP_PADRANGE);
6508 expr = OpSIBLING(expr);
6511 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6512 expr, &recompile, NULL);
6514 /* handle bare (possibly after overloading) regex: foo =~ $re */
6519 if (SvTYPE(re) == SVt_REGEXP) {
6523 Safefree(pRExC_state->code_blocks);
6524 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6525 "Precompiled pattern%s\n",
6526 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6532 exp = SvPV_nomg(pat, plen);
6534 if (!eng->op_comp) {
6535 if ((SvUTF8(pat) && IN_BYTES)
6536 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6538 /* make a temporary copy; either to convert to bytes,
6539 * or to avoid repeating get-magic / overloaded stringify */
6540 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6541 (IN_BYTES ? 0 : SvUTF8(pat)));
6543 Safefree(pRExC_state->code_blocks);
6544 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6547 /* ignore the utf8ness if the pattern is 0 length */
6548 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6549 RExC_uni_semantics = 0;
6550 RExC_contains_locale = 0;
6551 RExC_contains_i = 0;
6552 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6553 pRExC_state->runtime_code_qr = NULL;
6554 RExC_frame_head= NULL;
6555 RExC_frame_last= NULL;
6556 RExC_frame_count= 0;
6559 RExC_mysv1= sv_newmortal();
6560 RExC_mysv2= sv_newmortal();
6563 SV *dsv= sv_newmortal();
6564 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6565 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6566 PL_colors[4],PL_colors[5],s);
6570 /* we jump here if we upgrade the pattern to utf8 and have to
6573 if ((pm_flags & PMf_USE_RE_EVAL)
6574 /* this second condition covers the non-regex literal case,
6575 * i.e. $foo =~ '(?{})'. */
6576 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6578 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6580 /* return old regex if pattern hasn't changed */
6581 /* XXX: note in the below we have to check the flags as well as the
6584 * Things get a touch tricky as we have to compare the utf8 flag
6585 * independently from the compile flags. */
6589 && !!RX_UTF8(old_re) == !!RExC_utf8
6590 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6591 && RX_PRECOMP(old_re)
6592 && RX_PRELEN(old_re) == plen
6593 && memEQ(RX_PRECOMP(old_re), exp, plen)
6594 && !runtime_code /* with runtime code, always recompile */ )
6596 Safefree(pRExC_state->code_blocks);
6600 rx_flags = orig_rx_flags;
6602 if (rx_flags & PMf_FOLD) {
6603 RExC_contains_i = 1;
6605 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6607 /* Set to use unicode semantics if the pattern is in utf8 and has the
6608 * 'depends' charset specified, as it means unicode when utf8 */
6609 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6613 RExC_flags = rx_flags;
6614 RExC_pm_flags = pm_flags;
6617 if (TAINTING_get && TAINT_get)
6618 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6620 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6621 /* whoops, we have a non-utf8 pattern, whilst run-time code
6622 * got compiled as utf8. Try again with a utf8 pattern */
6623 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6624 pRExC_state->num_code_blocks);
6625 goto redo_first_pass;
6628 assert(!pRExC_state->runtime_code_qr);
6634 RExC_in_lookbehind = 0;
6635 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6637 RExC_override_recoding = 0;
6639 RExC_recode_x_to_native = 0;
6641 RExC_in_multi_char_class = 0;
6643 /* First pass: determine size, legality. */
6646 RExC_end = exp + plen;
6651 RExC_emit = (regnode *) &RExC_emit_dummy;
6652 RExC_whilem_seen = 0;
6653 RExC_open_parens = NULL;
6654 RExC_close_parens = NULL;
6656 RExC_paren_names = NULL;
6658 RExC_paren_name_list = NULL;
6660 RExC_recurse = NULL;
6661 RExC_study_chunk_recursed = NULL;
6662 RExC_study_chunk_recursed_bytes= 0;
6663 RExC_recurse_count = 0;
6664 pRExC_state->code_index = 0;
6667 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6669 RExC_lastparse=NULL;
6671 /* reg may croak on us, not giving us a chance to free
6672 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6673 need it to survive as long as the regexp (qr/(?{})/).
6674 We must check that code_blocksv is not already set, because we may
6675 have jumped back to restart the sizing pass. */
6676 if (pRExC_state->code_blocks && !code_blocksv) {
6677 code_blocksv = newSV_type(SVt_PV);
6678 SAVEFREESV(code_blocksv);
6679 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6680 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6682 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6683 /* It's possible to write a regexp in ascii that represents Unicode
6684 codepoints outside of the byte range, such as via \x{100}. If we
6685 detect such a sequence we have to convert the entire pattern to utf8
6686 and then recompile, as our sizing calculation will have been based
6687 on 1 byte == 1 character, but we will need to use utf8 to encode
6688 at least some part of the pattern, and therefore must convert the whole
6691 if (flags & RESTART_UTF8) {
6692 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6693 pRExC_state->num_code_blocks);
6694 goto redo_first_pass;
6696 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6699 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6702 PerlIO_printf(Perl_debug_log,
6703 "Required size %"IVdf" nodes\n"
6704 "Starting second pass (creation)\n",
6707 RExC_lastparse=NULL;
6710 /* The first pass could have found things that force Unicode semantics */
6711 if ((RExC_utf8 || RExC_uni_semantics)
6712 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6714 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6717 /* Small enough for pointer-storage convention?
6718 If extralen==0, this means that we will not need long jumps. */
6719 if (RExC_size >= 0x10000L && RExC_extralen)
6720 RExC_size += RExC_extralen;
6723 if (RExC_whilem_seen > 15)
6724 RExC_whilem_seen = 15;
6726 /* Allocate space and zero-initialize. Note, the two step process
6727 of zeroing when in debug mode, thus anything assigned has to
6728 happen after that */
6729 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6731 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6732 char, regexp_internal);
6733 if ( r == NULL || ri == NULL )
6734 FAIL("Regexp out of space");
6736 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6737 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6740 /* bulk initialize base fields with 0. */
6741 Zero(ri, sizeof(regexp_internal), char);
6744 /* non-zero initialization begins here */
6747 r->extflags = rx_flags;
6748 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6750 if (pm_flags & PMf_IS_QR) {
6751 ri->code_blocks = pRExC_state->code_blocks;
6752 ri->num_code_blocks = pRExC_state->num_code_blocks;
6757 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6758 if (pRExC_state->code_blocks[n].src_regex)
6759 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6760 SAVEFREEPV(pRExC_state->code_blocks);
6764 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6765 bool has_charset = (get_regex_charset(r->extflags)
6766 != REGEX_DEPENDS_CHARSET);
6768 /* The caret is output if there are any defaults: if not all the STD
6769 * flags are set, or if no character set specifier is needed */
6771 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6773 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6774 == REG_RUN_ON_COMMENT_SEEN);
6775 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6776 >> RXf_PMf_STD_PMMOD_SHIFT);
6777 const char *fptr = STD_PAT_MODS; /*"msixn"*/
6779 /* Allocate for the worst case, which is all the std flags are turned
6780 * on. If more precision is desired, we could do a population count of
6781 * the flags set. This could be done with a small lookup table, or by
6782 * shifting, masking and adding, or even, when available, assembly
6783 * language for a machine-language population count.
6784 * We never output a minus, as all those are defaults, so are
6785 * covered by the caret */
6786 const STRLEN wraplen = plen + has_p + has_runon
6787 + has_default /* If needs a caret */
6789 /* If needs a character set specifier */
6790 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6791 + (sizeof(STD_PAT_MODS) - 1)
6792 + (sizeof("(?:)") - 1);
6794 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6795 r->xpv_len_u.xpvlenu_pv = p;
6797 SvFLAGS(rx) |= SVf_UTF8;
6800 /* If a default, cover it using the caret */
6802 *p++= DEFAULT_PAT_MOD;
6806 const char* const name = get_regex_charset_name(r->extflags, &len);
6807 Copy(name, p, len, char);
6811 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6814 while((ch = *fptr++)) {
6822 Copy(RExC_precomp, p, plen, char);
6823 assert ((RX_WRAPPED(rx) - p) < 16);
6824 r->pre_prefix = p - RX_WRAPPED(rx);
6830 SvCUR_set(rx, p - RX_WRAPPED(rx));
6834 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6836 /* setup various meta data about recursion, this all requires
6837 * RExC_npar to be correctly set, and a bit later on we clear it */
6838 if (RExC_seen & REG_RECURSE_SEEN) {
6839 Newxz(RExC_open_parens, RExC_npar,regnode *);
6840 SAVEFREEPV(RExC_open_parens);
6841 Newxz(RExC_close_parens,RExC_npar,regnode *);
6842 SAVEFREEPV(RExC_close_parens);
6844 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6845 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6846 * So its 1 if there are no parens. */
6847 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6848 ((RExC_npar & 0x07) != 0);
6849 Newx(RExC_study_chunk_recursed,
6850 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6851 SAVEFREEPV(RExC_study_chunk_recursed);
6854 /* Useful during FAIL. */
6855 #ifdef RE_TRACK_PATTERN_OFFSETS
6856 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6857 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6858 "%s %"UVuf" bytes for offset annotations.\n",
6859 ri->u.offsets ? "Got" : "Couldn't get",
6860 (UV)((2*RExC_size+1) * sizeof(U32))));
6862 SetProgLen(ri,RExC_size);
6867 /* Second pass: emit code. */
6868 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6869 RExC_pm_flags = pm_flags;
6871 RExC_end = exp + plen;
6874 RExC_emit_start = ri->program;
6875 RExC_emit = ri->program;
6876 RExC_emit_bound = ri->program + RExC_size + 1;
6877 pRExC_state->code_index = 0;
6879 *((char*) RExC_emit++) = (char) REG_MAGIC;
6880 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6882 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6884 /* XXXX To minimize changes to RE engine we always allocate
6885 3-units-long substrs field. */
6886 Newx(r->substrs, 1, struct reg_substr_data);
6887 if (RExC_recurse_count) {
6888 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6889 SAVEFREEPV(RExC_recurse);
6893 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6895 RExC_study_chunk_recursed_count= 0;
6897 Zero(r->substrs, 1, struct reg_substr_data);
6898 if (RExC_study_chunk_recursed) {
6899 Zero(RExC_study_chunk_recursed,
6900 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6904 #ifdef TRIE_STUDY_OPT
6906 StructCopy(&zero_scan_data, &data, scan_data_t);
6907 copyRExC_state = RExC_state;
6910 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6912 RExC_state = copyRExC_state;
6913 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6914 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6916 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6917 StructCopy(&zero_scan_data, &data, scan_data_t);
6920 StructCopy(&zero_scan_data, &data, scan_data_t);
6923 /* Dig out information for optimizations. */
6924 r->extflags = RExC_flags; /* was pm_op */
6925 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6928 SvUTF8_on(rx); /* Unicode in it? */
6929 ri->regstclass = NULL;
6930 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
6931 r->intflags |= PREGf_NAUGHTY;
6932 scan = ri->program + 1; /* First BRANCH. */
6934 /* testing for BRANCH here tells us whether there is "must appear"
6935 data in the pattern. If there is then we can use it for optimisations */
6936 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6939 STRLEN longest_float_length, longest_fixed_length;
6940 regnode_ssc ch_class; /* pointed to by data */
6942 SSize_t last_close = 0; /* pointed to by data */
6943 regnode *first= scan;
6944 regnode *first_next= regnext(first);
6946 * Skip introductions and multiplicators >= 1
6947 * so that we can extract the 'meat' of the pattern that must
6948 * match in the large if() sequence following.
6949 * NOTE that EXACT is NOT covered here, as it is normally
6950 * picked up by the optimiser separately.
6952 * This is unfortunate as the optimiser isnt handling lookahead
6953 * properly currently.
6956 while ((OP(first) == OPEN && (sawopen = 1)) ||
6957 /* An OR of *one* alternative - should not happen now. */
6958 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6959 /* for now we can't handle lookbehind IFMATCH*/
6960 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6961 (OP(first) == PLUS) ||
6962 (OP(first) == MINMOD) ||
6963 /* An {n,m} with n>0 */
6964 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6965 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6968 * the only op that could be a regnode is PLUS, all the rest
6969 * will be regnode_1 or regnode_2.
6971 * (yves doesn't think this is true)
6973 if (OP(first) == PLUS)
6976 if (OP(first) == MINMOD)
6978 first += regarglen[OP(first)];
6980 first = NEXTOPER(first);
6981 first_next= regnext(first);
6984 /* Starting-point info. */
6986 DEBUG_PEEP("first:",first,0);
6987 /* Ignore EXACT as we deal with it later. */
6988 if (PL_regkind[OP(first)] == EXACT) {
6989 if (OP(first) == EXACT || OP(first) == EXACTL)
6990 NOOP; /* Empty, get anchored substr later. */
6992 ri->regstclass = first;
6995 else if (PL_regkind[OP(first)] == TRIE &&
6996 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6998 /* this can happen only on restudy */
6999 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7002 else if (REGNODE_SIMPLE(OP(first)))
7003 ri->regstclass = first;
7004 else if (PL_regkind[OP(first)] == BOUND ||
7005 PL_regkind[OP(first)] == NBOUND)
7006 ri->regstclass = first;
7007 else if (PL_regkind[OP(first)] == BOL) {
7008 r->intflags |= (OP(first) == MBOL
7011 first = NEXTOPER(first);
7014 else if (OP(first) == GPOS) {
7015 r->intflags |= PREGf_ANCH_GPOS;
7016 first = NEXTOPER(first);
7019 else if ((!sawopen || !RExC_sawback) &&
7021 (OP(first) == STAR &&
7022 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7023 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7025 /* turn .* into ^.* with an implied $*=1 */
7027 (OP(NEXTOPER(first)) == REG_ANY)
7030 r->intflags |= (type | PREGf_IMPLICIT);
7031 first = NEXTOPER(first);
7034 if (sawplus && !sawminmod && !sawlookahead
7035 && (!sawopen || !RExC_sawback)
7036 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7037 /* x+ must match at the 1st pos of run of x's */
7038 r->intflags |= PREGf_SKIP;
7040 /* Scan is after the zeroth branch, first is atomic matcher. */
7041 #ifdef TRIE_STUDY_OPT
7044 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7045 (IV)(first - scan + 1))
7049 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7050 (IV)(first - scan + 1))
7056 * If there's something expensive in the r.e., find the
7057 * longest literal string that must appear and make it the
7058 * regmust. Resolve ties in favor of later strings, since
7059 * the regstart check works with the beginning of the r.e.
7060 * and avoiding duplication strengthens checking. Not a
7061 * strong reason, but sufficient in the absence of others.
7062 * [Now we resolve ties in favor of the earlier string if
7063 * it happens that c_offset_min has been invalidated, since the
7064 * earlier string may buy us something the later one won't.]
7067 data.longest_fixed = newSVpvs("");
7068 data.longest_float = newSVpvs("");
7069 data.last_found = newSVpvs("");
7070 data.longest = &(data.longest_fixed);
7071 ENTER_with_name("study_chunk");
7072 SAVEFREESV(data.longest_fixed);
7073 SAVEFREESV(data.longest_float);
7074 SAVEFREESV(data.last_found);
7076 if (!ri->regstclass) {
7077 ssc_init(pRExC_state, &ch_class);
7078 data.start_class = &ch_class;
7079 stclass_flag = SCF_DO_STCLASS_AND;
7080 } else /* XXXX Check for BOUND? */
7082 data.last_closep = &last_close;
7085 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7086 scan + RExC_size, /* Up to end */
7088 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7089 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7093 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7096 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7097 && data.last_start_min == 0 && data.last_end > 0
7098 && !RExC_seen_zerolen
7099 && !(RExC_seen & REG_VERBARG_SEEN)
7100 && !(RExC_seen & REG_GPOS_SEEN)
7102 r->extflags |= RXf_CHECK_ALL;
7104 scan_commit(pRExC_state, &data,&minlen,0);
7106 longest_float_length = CHR_SVLEN(data.longest_float);
7108 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7109 && data.offset_fixed == data.offset_float_min
7110 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7111 && S_setup_longest (aTHX_ pRExC_state,
7115 &(r->float_end_shift),
7116 data.lookbehind_float,
7117 data.offset_float_min,
7119 longest_float_length,
7120 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7121 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7123 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7124 r->float_max_offset = data.offset_float_max;
7125 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7126 r->float_max_offset -= data.lookbehind_float;
7127 SvREFCNT_inc_simple_void_NN(data.longest_float);
7130 r->float_substr = r->float_utf8 = NULL;
7131 longest_float_length = 0;
7134 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7136 if (S_setup_longest (aTHX_ pRExC_state,
7138 &(r->anchored_utf8),
7139 &(r->anchored_substr),
7140 &(r->anchored_end_shift),
7141 data.lookbehind_fixed,
7144 longest_fixed_length,
7145 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7146 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7148 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7149 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7152 r->anchored_substr = r->anchored_utf8 = NULL;
7153 longest_fixed_length = 0;
7155 LEAVE_with_name("study_chunk");
7158 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7159 ri->regstclass = NULL;
7161 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7163 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7164 && is_ssc_worth_it(pRExC_state, data.start_class))
7166 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7168 ssc_finalize(pRExC_state, data.start_class);
7170 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7171 StructCopy(data.start_class,
7172 (regnode_ssc*)RExC_rxi->data->data[n],
7174 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7175 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7176 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7177 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7178 PerlIO_printf(Perl_debug_log,
7179 "synthetic stclass \"%s\".\n",
7180 SvPVX_const(sv));});
7181 data.start_class = NULL;
7184 /* A temporary algorithm prefers floated substr to fixed one to dig
7186 if (longest_fixed_length > longest_float_length) {
7187 r->substrs->check_ix = 0;
7188 r->check_end_shift = r->anchored_end_shift;
7189 r->check_substr = r->anchored_substr;
7190 r->check_utf8 = r->anchored_utf8;
7191 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7192 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7193 r->intflags |= PREGf_NOSCAN;
7196 r->substrs->check_ix = 1;
7197 r->check_end_shift = r->float_end_shift;
7198 r->check_substr = r->float_substr;
7199 r->check_utf8 = r->float_utf8;
7200 r->check_offset_min = r->float_min_offset;
7201 r->check_offset_max = r->float_max_offset;
7203 if ((r->check_substr || r->check_utf8) ) {
7204 r->extflags |= RXf_USE_INTUIT;
7205 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7206 r->extflags |= RXf_INTUIT_TAIL;
7208 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7210 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7211 if ( (STRLEN)minlen < longest_float_length )
7212 minlen= longest_float_length;
7213 if ( (STRLEN)minlen < longest_fixed_length )
7214 minlen= longest_fixed_length;
7218 /* Several toplevels. Best we can is to set minlen. */
7220 regnode_ssc ch_class;
7221 SSize_t last_close = 0;
7223 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7225 scan = ri->program + 1;
7226 ssc_init(pRExC_state, &ch_class);
7227 data.start_class = &ch_class;
7228 data.last_closep = &last_close;
7231 minlen = study_chunk(pRExC_state,
7232 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7233 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7234 ? SCF_TRIE_DOING_RESTUDY
7238 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7240 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7241 = r->float_substr = r->float_utf8 = NULL;
7243 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7244 && is_ssc_worth_it(pRExC_state, data.start_class))
7246 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7248 ssc_finalize(pRExC_state, data.start_class);
7250 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7251 StructCopy(data.start_class,
7252 (regnode_ssc*)RExC_rxi->data->data[n],
7254 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7255 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7256 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7257 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7258 PerlIO_printf(Perl_debug_log,
7259 "synthetic stclass \"%s\".\n",
7260 SvPVX_const(sv));});
7261 data.start_class = NULL;
7265 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7266 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7267 r->maxlen = REG_INFTY;
7270 r->maxlen = RExC_maxlen;
7273 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7274 the "real" pattern. */
7276 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%"IVdf"\n",
7277 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7279 r->minlenret = minlen;
7280 if (r->minlen < minlen)
7283 if (RExC_seen & REG_GPOS_SEEN)
7284 r->intflags |= PREGf_GPOS_SEEN;
7285 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7286 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7288 if (pRExC_state->num_code_blocks)
7289 r->extflags |= RXf_EVAL_SEEN;
7290 if (RExC_seen & REG_CANY_SEEN)
7291 r->intflags |= PREGf_CANY_SEEN;
7292 if (RExC_seen & REG_VERBARG_SEEN)
7294 r->intflags |= PREGf_VERBARG_SEEN;
7295 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7297 if (RExC_seen & REG_CUTGROUP_SEEN)
7298 r->intflags |= PREGf_CUTGROUP_SEEN;
7299 if (pm_flags & PMf_USE_RE_EVAL)
7300 r->intflags |= PREGf_USE_RE_EVAL;
7301 if (RExC_paren_names)
7302 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7304 RXp_PAREN_NAMES(r) = NULL;
7306 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7307 * so it can be used in pp.c */
7308 if (r->intflags & PREGf_ANCH)
7309 r->extflags |= RXf_IS_ANCHORED;
7313 /* this is used to identify "special" patterns that might result
7314 * in Perl NOT calling the regex engine and instead doing the match "itself",
7315 * particularly special cases in split//. By having the regex compiler
7316 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7317 * we avoid weird issues with equivalent patterns resulting in different behavior,
7318 * AND we allow non Perl engines to get the same optimizations by the setting the
7319 * flags appropriately - Yves */
7320 regnode *first = ri->program + 1;
7322 regnode *next = NEXTOPER(first);
7325 if (PL_regkind[fop] == NOTHING && nop == END)
7326 r->extflags |= RXf_NULL;
7327 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7328 /* when fop is SBOL first->flags will be true only when it was
7329 * produced by parsing /\A/, and not when parsing /^/. This is
7330 * very important for the split code as there we want to
7331 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7332 * See rt #122761 for more details. -- Yves */
7333 r->extflags |= RXf_START_ONLY;
7334 else if (fop == PLUS
7335 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7336 && OP(regnext(first)) == END)
7337 r->extflags |= RXf_WHITE;
7338 else if ( r->extflags & RXf_SPLIT
7339 && (fop == EXACT || fop == EXACTL)
7340 && STR_LEN(first) == 1
7341 && *(STRING(first)) == ' '
7342 && OP(regnext(first)) == END )
7343 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7347 if (RExC_contains_locale) {
7348 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7352 if (RExC_paren_names) {
7353 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7354 ri->data->data[ri->name_list_idx]
7355 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7358 ri->name_list_idx = 0;
7360 if (RExC_recurse_count) {
7361 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7362 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7363 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7366 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7367 /* assume we don't need to swap parens around before we match */
7369 PerlIO_printf(Perl_debug_log,"study_chunk_recursed_count: %lu\n",
7370 (unsigned long)RExC_study_chunk_recursed_count);
7374 PerlIO_printf(Perl_debug_log,"Final program:\n");
7377 #ifdef RE_TRACK_PATTERN_OFFSETS
7378 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7379 const STRLEN len = ri->u.offsets[0];
7381 GET_RE_DEBUG_FLAGS_DECL;
7382 PerlIO_printf(Perl_debug_log,
7383 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7384 for (i = 1; i <= len; i++) {
7385 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7386 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7387 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7389 PerlIO_printf(Perl_debug_log, "\n");
7394 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7395 * by setting the regexp SV to readonly-only instead. If the
7396 * pattern's been recompiled, the USEDness should remain. */
7397 if (old_re && SvREADONLY(old_re))
7405 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7408 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7410 PERL_UNUSED_ARG(value);
7412 if (flags & RXapif_FETCH) {
7413 return reg_named_buff_fetch(rx, key, flags);
7414 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7415 Perl_croak_no_modify();
7417 } else if (flags & RXapif_EXISTS) {
7418 return reg_named_buff_exists(rx, key, flags)
7421 } else if (flags & RXapif_REGNAMES) {
7422 return reg_named_buff_all(rx, flags);
7423 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7424 return reg_named_buff_scalar(rx, flags);
7426 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7432 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7435 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7436 PERL_UNUSED_ARG(lastkey);
7438 if (flags & RXapif_FIRSTKEY)
7439 return reg_named_buff_firstkey(rx, flags);
7440 else if (flags & RXapif_NEXTKEY)
7441 return reg_named_buff_nextkey(rx, flags);
7443 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7450 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7453 AV *retarray = NULL;
7455 struct regexp *const rx = ReANY(r);
7457 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7459 if (flags & RXapif_ALL)
7462 if (rx && RXp_PAREN_NAMES(rx)) {
7463 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7466 SV* sv_dat=HeVAL(he_str);
7467 I32 *nums=(I32*)SvPVX(sv_dat);
7468 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7469 if ((I32)(rx->nparens) >= nums[i]
7470 && rx->offs[nums[i]].start != -1
7471 && rx->offs[nums[i]].end != -1)
7474 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7479 ret = newSVsv(&PL_sv_undef);
7482 av_push(retarray, ret);
7485 return newRV_noinc(MUTABLE_SV(retarray));
7492 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7495 struct regexp *const rx = ReANY(r);
7497 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7499 if (rx && RXp_PAREN_NAMES(rx)) {
7500 if (flags & RXapif_ALL) {
7501 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7503 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7505 SvREFCNT_dec_NN(sv);
7517 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7519 struct regexp *const rx = ReANY(r);
7521 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7523 if ( rx && RXp_PAREN_NAMES(rx) ) {
7524 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7526 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7533 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7535 struct regexp *const rx = ReANY(r);
7536 GET_RE_DEBUG_FLAGS_DECL;
7538 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7540 if (rx && RXp_PAREN_NAMES(rx)) {
7541 HV *hv = RXp_PAREN_NAMES(rx);
7543 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7546 SV* sv_dat = HeVAL(temphe);
7547 I32 *nums = (I32*)SvPVX(sv_dat);
7548 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7549 if ((I32)(rx->lastparen) >= nums[i] &&
7550 rx->offs[nums[i]].start != -1 &&
7551 rx->offs[nums[i]].end != -1)
7557 if (parno || flags & RXapif_ALL) {
7558 return newSVhek(HeKEY_hek(temphe));
7566 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7571 struct regexp *const rx = ReANY(r);
7573 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7575 if (rx && RXp_PAREN_NAMES(rx)) {
7576 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7577 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7578 } else if (flags & RXapif_ONE) {
7579 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7580 av = MUTABLE_AV(SvRV(ret));
7581 length = av_tindex(av);
7582 SvREFCNT_dec_NN(ret);
7583 return newSViv(length + 1);
7585 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7590 return &PL_sv_undef;
7594 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7596 struct regexp *const rx = ReANY(r);
7599 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7601 if (rx && RXp_PAREN_NAMES(rx)) {
7602 HV *hv= RXp_PAREN_NAMES(rx);
7604 (void)hv_iterinit(hv);
7605 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7608 SV* sv_dat = HeVAL(temphe);
7609 I32 *nums = (I32*)SvPVX(sv_dat);
7610 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7611 if ((I32)(rx->lastparen) >= nums[i] &&
7612 rx->offs[nums[i]].start != -1 &&
7613 rx->offs[nums[i]].end != -1)
7619 if (parno || flags & RXapif_ALL) {
7620 av_push(av, newSVhek(HeKEY_hek(temphe)));
7625 return newRV_noinc(MUTABLE_SV(av));
7629 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7632 struct regexp *const rx = ReANY(r);
7638 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7640 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7641 || n == RX_BUFF_IDX_CARET_FULLMATCH
7642 || n == RX_BUFF_IDX_CARET_POSTMATCH
7645 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7647 /* on something like
7650 * the KEEPCOPY is set on the PMOP rather than the regex */
7651 if (PL_curpm && r == PM_GETRE(PL_curpm))
7652 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7661 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7662 /* no need to distinguish between them any more */
7663 n = RX_BUFF_IDX_FULLMATCH;
7665 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7666 && rx->offs[0].start != -1)
7668 /* $`, ${^PREMATCH} */
7669 i = rx->offs[0].start;
7673 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7674 && rx->offs[0].end != -1)
7676 /* $', ${^POSTMATCH} */
7677 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7678 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7681 if ( 0 <= n && n <= (I32)rx->nparens &&
7682 (s1 = rx->offs[n].start) != -1 &&
7683 (t1 = rx->offs[n].end) != -1)
7685 /* $&, ${^MATCH}, $1 ... */
7687 s = rx->subbeg + s1 - rx->suboffset;
7692 assert(s >= rx->subbeg);
7693 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7695 #ifdef NO_TAINT_SUPPORT
7696 sv_setpvn(sv, s, i);
7698 const int oldtainted = TAINT_get;
7700 sv_setpvn(sv, s, i);
7701 TAINT_set(oldtainted);
7703 if ( (rx->intflags & PREGf_CANY_SEEN)
7704 ? (RXp_MATCH_UTF8(rx)
7705 && (!i || is_utf8_string((U8*)s, i)))
7706 : (RXp_MATCH_UTF8(rx)) )
7713 if (RXp_MATCH_TAINTED(rx)) {
7714 if (SvTYPE(sv) >= SVt_PVMG) {
7715 MAGIC* const mg = SvMAGIC(sv);
7718 SvMAGIC_set(sv, mg->mg_moremagic);
7720 if ((mgt = SvMAGIC(sv))) {
7721 mg->mg_moremagic = mgt;
7722 SvMAGIC_set(sv, mg);
7733 sv_setsv(sv,&PL_sv_undef);
7739 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7740 SV const * const value)
7742 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7744 PERL_UNUSED_ARG(rx);
7745 PERL_UNUSED_ARG(paren);
7746 PERL_UNUSED_ARG(value);
7749 Perl_croak_no_modify();
7753 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7756 struct regexp *const rx = ReANY(r);
7760 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7762 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7763 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7764 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7767 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7769 /* on something like
7772 * the KEEPCOPY is set on the PMOP rather than the regex */
7773 if (PL_curpm && r == PM_GETRE(PL_curpm))
7774 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7780 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7782 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7783 case RX_BUFF_IDX_PREMATCH: /* $` */
7784 if (rx->offs[0].start != -1) {
7785 i = rx->offs[0].start;
7794 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7795 case RX_BUFF_IDX_POSTMATCH: /* $' */
7796 if (rx->offs[0].end != -1) {
7797 i = rx->sublen - rx->offs[0].end;
7799 s1 = rx->offs[0].end;
7806 default: /* $& / ${^MATCH}, $1, $2, ... */
7807 if (paren <= (I32)rx->nparens &&
7808 (s1 = rx->offs[paren].start) != -1 &&
7809 (t1 = rx->offs[paren].end) != -1)
7815 if (ckWARN(WARN_UNINITIALIZED))
7816 report_uninit((const SV *)sv);
7821 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7822 const char * const s = rx->subbeg - rx->suboffset + s1;
7827 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7834 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7836 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7837 PERL_UNUSED_ARG(rx);
7841 return newSVpvs("Regexp");
7844 /* Scans the name of a named buffer from the pattern.
7845 * If flags is REG_RSN_RETURN_NULL returns null.
7846 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7847 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7848 * to the parsed name as looked up in the RExC_paren_names hash.
7849 * If there is an error throws a vFAIL().. type exception.
7852 #define REG_RSN_RETURN_NULL 0
7853 #define REG_RSN_RETURN_NAME 1
7854 #define REG_RSN_RETURN_DATA 2
7857 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7859 char *name_start = RExC_parse;
7861 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7863 assert (RExC_parse <= RExC_end);
7864 if (RExC_parse == RExC_end) NOOP;
7865 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7866 /* skip IDFIRST by using do...while */
7869 RExC_parse += UTF8SKIP(RExC_parse);
7870 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7874 } while (isWORDCHAR(*RExC_parse));
7876 RExC_parse++; /* so the <- from the vFAIL is after the offending
7878 vFAIL("Group name must start with a non-digit word character");
7882 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7883 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7884 if ( flags == REG_RSN_RETURN_NAME)
7886 else if (flags==REG_RSN_RETURN_DATA) {
7889 if ( ! sv_name ) /* should not happen*/
7890 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7891 if (RExC_paren_names)
7892 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7894 sv_dat = HeVAL(he_str);
7896 vFAIL("Reference to nonexistent named group");
7900 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7901 (unsigned long) flags);
7903 NOT_REACHED; /* NOTREACHED */
7908 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7910 if (RExC_lastparse!=RExC_parse) { \
7911 PerlIO_printf(Perl_debug_log, "%s", \
7912 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
7913 RExC_end - RExC_parse, 16, \
7915 PERL_PV_ESCAPE_UNI_DETECT | \
7916 PERL_PV_PRETTY_ELLIPSES | \
7917 PERL_PV_PRETTY_LTGT | \
7918 PERL_PV_ESCAPE_RE | \
7919 PERL_PV_PRETTY_EXACTSIZE \
7923 PerlIO_printf(Perl_debug_log,"%16s",""); \
7926 num = RExC_size + 1; \
7928 num=REG_NODE_NUM(RExC_emit); \
7929 if (RExC_lastnum!=num) \
7930 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7932 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7933 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7934 (int)((depth*2)), "", \
7938 RExC_lastparse=RExC_parse; \
7943 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7944 DEBUG_PARSE_MSG((funcname)); \
7945 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7947 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7948 DEBUG_PARSE_MSG((funcname)); \
7949 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7952 /* This section of code defines the inversion list object and its methods. The
7953 * interfaces are highly subject to change, so as much as possible is static to
7954 * this file. An inversion list is here implemented as a malloc'd C UV array
7955 * as an SVt_INVLIST scalar.
7957 * An inversion list for Unicode is an array of code points, sorted by ordinal
7958 * number. The zeroth element is the first code point in the list. The 1th
7959 * element is the first element beyond that not in the list. In other words,
7960 * the first range is
7961 * invlist[0]..(invlist[1]-1)
7962 * The other ranges follow. Thus every element whose index is divisible by two
7963 * marks the beginning of a range that is in the list, and every element not
7964 * divisible by two marks the beginning of a range not in the list. A single
7965 * element inversion list that contains the single code point N generally
7966 * consists of two elements
7969 * (The exception is when N is the highest representable value on the
7970 * machine, in which case the list containing just it would be a single
7971 * element, itself. By extension, if the last range in the list extends to
7972 * infinity, then the first element of that range will be in the inversion list
7973 * at a position that is divisible by two, and is the final element in the
7975 * Taking the complement (inverting) an inversion list is quite simple, if the
7976 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7977 * This implementation reserves an element at the beginning of each inversion
7978 * list to always contain 0; there is an additional flag in the header which
7979 * indicates if the list begins at the 0, or is offset to begin at the next
7982 * More about inversion lists can be found in "Unicode Demystified"
7983 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7984 * More will be coming when functionality is added later.
7986 * The inversion list data structure is currently implemented as an SV pointing
7987 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7988 * array of UV whose memory management is automatically handled by the existing
7989 * facilities for SV's.
7991 * Some of the methods should always be private to the implementation, and some
7992 * should eventually be made public */
7994 /* The header definitions are in F<inline_invlist.c> */
7996 PERL_STATIC_INLINE UV*
7997 S__invlist_array_init(SV* const invlist, const bool will_have_0)
7999 /* Returns a pointer to the first element in the inversion list's array.
8000 * This is called upon initialization of an inversion list. Where the
8001 * array begins depends on whether the list has the code point U+0000 in it
8002 * or not. The other parameter tells it whether the code that follows this
8003 * call is about to put a 0 in the inversion list or not. The first
8004 * element is either the element reserved for 0, if TRUE, or the element
8005 * after it, if FALSE */
8007 bool* offset = get_invlist_offset_addr(invlist);
8008 UV* zero_addr = (UV *) SvPVX(invlist);
8010 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8013 assert(! _invlist_len(invlist));
8017 /* 1^1 = 0; 1^0 = 1 */
8018 *offset = 1 ^ will_have_0;
8019 return zero_addr + *offset;
8022 PERL_STATIC_INLINE void
8023 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8025 /* Sets the current number of elements stored in the inversion list.
8026 * Updates SvCUR correspondingly */
8027 PERL_UNUSED_CONTEXT;
8028 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8030 assert(SvTYPE(invlist) == SVt_INVLIST);
8035 : TO_INTERNAL_SIZE(len + offset));
8036 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8039 #ifndef PERL_IN_XSUB_RE
8041 PERL_STATIC_INLINE IV*
8042 S_get_invlist_previous_index_addr(SV* invlist)
8044 /* Return the address of the IV that is reserved to hold the cached index
8046 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8048 assert(SvTYPE(invlist) == SVt_INVLIST);
8050 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8053 PERL_STATIC_INLINE IV
8054 S_invlist_previous_index(SV* const invlist)
8056 /* Returns cached index of previous search */
8058 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8060 return *get_invlist_previous_index_addr(invlist);
8063 PERL_STATIC_INLINE void
8064 S_invlist_set_previous_index(SV* const invlist, const IV index)
8066 /* Caches <index> for later retrieval */
8068 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8070 assert(index == 0 || index < (int) _invlist_len(invlist));
8072 *get_invlist_previous_index_addr(invlist) = index;
8075 PERL_STATIC_INLINE void
8076 S_invlist_trim(SV* const invlist)
8078 PERL_ARGS_ASSERT_INVLIST_TRIM;
8080 assert(SvTYPE(invlist) == SVt_INVLIST);
8082 /* Change the length of the inversion list to how many entries it currently
8084 SvPV_shrink_to_cur((SV *) invlist);
8087 PERL_STATIC_INLINE bool
8088 S_invlist_is_iterating(SV* const invlist)
8090 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8092 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8095 #endif /* ifndef PERL_IN_XSUB_RE */
8097 PERL_STATIC_INLINE UV
8098 S_invlist_max(SV* const invlist)
8100 /* Returns the maximum number of elements storable in the inversion list's
8101 * array, without having to realloc() */
8103 PERL_ARGS_ASSERT_INVLIST_MAX;
8105 assert(SvTYPE(invlist) == SVt_INVLIST);
8107 /* Assumes worst case, in which the 0 element is not counted in the
8108 * inversion list, so subtracts 1 for that */
8109 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8110 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8111 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8114 #ifndef PERL_IN_XSUB_RE
8116 Perl__new_invlist(pTHX_ IV initial_size)
8119 /* Return a pointer to a newly constructed inversion list, with enough
8120 * space to store 'initial_size' elements. If that number is negative, a
8121 * system default is used instead */
8125 if (initial_size < 0) {
8129 /* Allocate the initial space */
8130 new_list = newSV_type(SVt_INVLIST);
8132 /* First 1 is in case the zero element isn't in the list; second 1 is for
8134 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8135 invlist_set_len(new_list, 0, 0);
8137 /* Force iterinit() to be used to get iteration to work */
8138 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8140 *get_invlist_previous_index_addr(new_list) = 0;
8146 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8148 /* Return a pointer to a newly constructed inversion list, initialized to
8149 * point to <list>, which has to be in the exact correct inversion list
8150 * form, including internal fields. Thus this is a dangerous routine that
8151 * should not be used in the wrong hands. The passed in 'list' contains
8152 * several header fields at the beginning that are not part of the
8153 * inversion list body proper */
8155 const STRLEN length = (STRLEN) list[0];
8156 const UV version_id = list[1];
8157 const bool offset = cBOOL(list[2]);
8158 #define HEADER_LENGTH 3
8159 /* If any of the above changes in any way, you must change HEADER_LENGTH
8160 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8161 * perl -E 'say int(rand 2**31-1)'
8163 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8164 data structure type, so that one being
8165 passed in can be validated to be an
8166 inversion list of the correct vintage.
8169 SV* invlist = newSV_type(SVt_INVLIST);
8171 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8173 if (version_id != INVLIST_VERSION_ID) {
8174 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8177 /* The generated array passed in includes header elements that aren't part
8178 * of the list proper, so start it just after them */
8179 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8181 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8182 shouldn't touch it */
8184 *(get_invlist_offset_addr(invlist)) = offset;
8186 /* The 'length' passed to us is the physical number of elements in the
8187 * inversion list. But if there is an offset the logical number is one
8189 invlist_set_len(invlist, length - offset, offset);
8191 invlist_set_previous_index(invlist, 0);
8193 /* Initialize the iteration pointer. */
8194 invlist_iterfinish(invlist);
8196 SvREADONLY_on(invlist);
8200 #endif /* ifndef PERL_IN_XSUB_RE */
8203 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8205 /* Grow the maximum size of an inversion list */
8207 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8209 assert(SvTYPE(invlist) == SVt_INVLIST);
8211 /* Add one to account for the zero element at the beginning which may not
8212 * be counted by the calling parameters */
8213 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8217 S__append_range_to_invlist(pTHX_ SV* const invlist,
8218 const UV start, const UV end)
8220 /* Subject to change or removal. Append the range from 'start' to 'end' at
8221 * the end of the inversion list. The range must be above any existing
8225 UV max = invlist_max(invlist);
8226 UV len = _invlist_len(invlist);
8229 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8231 if (len == 0) { /* Empty lists must be initialized */
8232 offset = start != 0;
8233 array = _invlist_array_init(invlist, ! offset);
8236 /* Here, the existing list is non-empty. The current max entry in the
8237 * list is generally the first value not in the set, except when the
8238 * set extends to the end of permissible values, in which case it is
8239 * the first entry in that final set, and so this call is an attempt to
8240 * append out-of-order */
8242 UV final_element = len - 1;
8243 array = invlist_array(invlist);
8244 if (array[final_element] > start
8245 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8247 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",
8248 array[final_element], start,
8249 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8252 /* Here, it is a legal append. If the new range begins with the first
8253 * value not in the set, it is extending the set, so the new first
8254 * value not in the set is one greater than the newly extended range.
8256 offset = *get_invlist_offset_addr(invlist);
8257 if (array[final_element] == start) {
8258 if (end != UV_MAX) {
8259 array[final_element] = end + 1;
8262 /* But if the end is the maximum representable on the machine,
8263 * just let the range that this would extend to have no end */
8264 invlist_set_len(invlist, len - 1, offset);
8270 /* Here the new range doesn't extend any existing set. Add it */
8272 len += 2; /* Includes an element each for the start and end of range */
8274 /* If wll overflow the existing space, extend, which may cause the array to
8277 invlist_extend(invlist, len);
8279 /* Have to set len here to avoid assert failure in invlist_array() */
8280 invlist_set_len(invlist, len, offset);
8282 array = invlist_array(invlist);
8285 invlist_set_len(invlist, len, offset);
8288 /* The next item on the list starts the range, the one after that is
8289 * one past the new range. */
8290 array[len - 2] = start;
8291 if (end != UV_MAX) {
8292 array[len - 1] = end + 1;
8295 /* But if the end is the maximum representable on the machine, just let
8296 * the range have no end */
8297 invlist_set_len(invlist, len - 1, offset);
8301 #ifndef PERL_IN_XSUB_RE
8304 Perl__invlist_search(SV* const invlist, const UV cp)
8306 /* Searches the inversion list for the entry that contains the input code
8307 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8308 * return value is the index into the list's array of the range that
8313 IV high = _invlist_len(invlist);
8314 const IV highest_element = high - 1;
8317 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8319 /* If list is empty, return failure. */
8324 /* (We can't get the array unless we know the list is non-empty) */
8325 array = invlist_array(invlist);
8327 mid = invlist_previous_index(invlist);
8328 assert(mid >=0 && mid <= highest_element);
8330 /* <mid> contains the cache of the result of the previous call to this
8331 * function (0 the first time). See if this call is for the same result,
8332 * or if it is for mid-1. This is under the theory that calls to this
8333 * function will often be for related code points that are near each other.
8334 * And benchmarks show that caching gives better results. We also test
8335 * here if the code point is within the bounds of the list. These tests
8336 * replace others that would have had to be made anyway to make sure that
8337 * the array bounds were not exceeded, and these give us extra information
8338 * at the same time */
8339 if (cp >= array[mid]) {
8340 if (cp >= array[highest_element]) {
8341 return highest_element;
8344 /* Here, array[mid] <= cp < array[highest_element]. This means that
8345 * the final element is not the answer, so can exclude it; it also
8346 * means that <mid> is not the final element, so can refer to 'mid + 1'
8348 if (cp < array[mid + 1]) {
8354 else { /* cp < aray[mid] */
8355 if (cp < array[0]) { /* Fail if outside the array */
8359 if (cp >= array[mid - 1]) {
8364 /* Binary search. What we are looking for is <i> such that
8365 * array[i] <= cp < array[i+1]
8366 * The loop below converges on the i+1. Note that there may not be an
8367 * (i+1)th element in the array, and things work nonetheless */
8368 while (low < high) {
8369 mid = (low + high) / 2;
8370 assert(mid <= highest_element);
8371 if (array[mid] <= cp) { /* cp >= array[mid] */
8374 /* We could do this extra test to exit the loop early.
8375 if (cp < array[low]) {
8380 else { /* cp < array[mid] */
8387 invlist_set_previous_index(invlist, high);
8392 Perl__invlist_populate_swatch(SV* const invlist,
8393 const UV start, const UV end, U8* swatch)
8395 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8396 * but is used when the swash has an inversion list. This makes this much
8397 * faster, as it uses a binary search instead of a linear one. This is
8398 * intimately tied to that function, and perhaps should be in utf8.c,
8399 * except it is intimately tied to inversion lists as well. It assumes
8400 * that <swatch> is all 0's on input */
8403 const IV len = _invlist_len(invlist);
8407 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8409 if (len == 0) { /* Empty inversion list */
8413 array = invlist_array(invlist);
8415 /* Find which element it is */
8416 i = _invlist_search(invlist, start);
8418 /* We populate from <start> to <end> */
8419 while (current < end) {
8422 /* The inversion list gives the results for every possible code point
8423 * after the first one in the list. Only those ranges whose index is
8424 * even are ones that the inversion list matches. For the odd ones,
8425 * and if the initial code point is not in the list, we have to skip
8426 * forward to the next element */
8427 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8429 if (i >= len) { /* Finished if beyond the end of the array */
8433 if (current >= end) { /* Finished if beyond the end of what we
8435 if (LIKELY(end < UV_MAX)) {
8439 /* We get here when the upper bound is the maximum
8440 * representable on the machine, and we are looking for just
8441 * that code point. Have to special case it */
8443 goto join_end_of_list;
8446 assert(current >= start);
8448 /* The current range ends one below the next one, except don't go past
8451 upper = (i < len && array[i] < end) ? array[i] : end;
8453 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8454 * for each code point in it */
8455 for (; current < upper; current++) {
8456 const STRLEN offset = (STRLEN)(current - start);
8457 swatch[offset >> 3] |= 1 << (offset & 7);
8462 /* Quit if at the end of the list */
8465 /* But first, have to deal with the highest possible code point on
8466 * the platform. The previous code assumes that <end> is one
8467 * beyond where we want to populate, but that is impossible at the
8468 * platform's infinity, so have to handle it specially */
8469 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8471 const STRLEN offset = (STRLEN)(end - start);
8472 swatch[offset >> 3] |= 1 << (offset & 7);
8477 /* Advance to the next range, which will be for code points not in the
8486 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8487 const bool complement_b, SV** output)
8489 /* Take the union of two inversion lists and point <output> to it. *output
8490 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8491 * the reference count to that list will be decremented if not already a
8492 * temporary (mortal); otherwise *output will be made correspondingly
8493 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8494 * second list is returned. If <complement_b> is TRUE, the union is taken
8495 * of the complement (inversion) of <b> instead of b itself.
8497 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8498 * Richard Gillam, published by Addison-Wesley, and explained at some
8499 * length there. The preface says to incorporate its examples into your
8500 * code at your own risk.
8502 * The algorithm is like a merge sort.
8504 * XXX A potential performance improvement is to keep track as we go along
8505 * if only one of the inputs contributes to the result, meaning the other
8506 * is a subset of that one. In that case, we can skip the final copy and
8507 * return the larger of the input lists, but then outside code might need
8508 * to keep track of whether to free the input list or not */
8510 const UV* array_a; /* a's array */
8512 UV len_a; /* length of a's array */
8515 SV* u; /* the resulting union */
8519 UV i_a = 0; /* current index into a's array */
8523 /* running count, as explained in the algorithm source book; items are
8524 * stopped accumulating and are output when the count changes to/from 0.
8525 * The count is incremented when we start a range that's in the set, and
8526 * decremented when we start a range that's not in the set. So its range
8527 * is 0 to 2. Only when the count is zero is something not in the set.
8531 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8534 /* If either one is empty, the union is the other one */
8535 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8536 bool make_temp = FALSE; /* Should we mortalize the result? */
8540 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8546 *output = invlist_clone(b);
8548 _invlist_invert(*output);
8550 } /* else *output already = b; */
8553 sv_2mortal(*output);
8557 else if ((len_b = _invlist_len(b)) == 0) {
8558 bool make_temp = FALSE;
8560 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8565 /* The complement of an empty list is a list that has everything in it,
8566 * so the union with <a> includes everything too */
8569 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8573 *output = _new_invlist(1);
8574 _append_range_to_invlist(*output, 0, UV_MAX);
8576 else if (*output != a) {
8577 *output = invlist_clone(a);
8579 /* else *output already = a; */
8582 sv_2mortal(*output);
8587 /* Here both lists exist and are non-empty */
8588 array_a = invlist_array(a);
8589 array_b = invlist_array(b);
8591 /* If are to take the union of 'a' with the complement of b, set it
8592 * up so are looking at b's complement. */
8595 /* To complement, we invert: if the first element is 0, remove it. To
8596 * do this, we just pretend the array starts one later */
8597 if (array_b[0] == 0) {
8603 /* But if the first element is not zero, we pretend the list starts
8604 * at the 0 that is always stored immediately before the array. */
8610 /* Size the union for the worst case: that the sets are completely
8612 u = _new_invlist(len_a + len_b);
8614 /* Will contain U+0000 if either component does */
8615 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8616 || (len_b > 0 && array_b[0] == 0));
8618 /* Go through each list item by item, stopping when exhausted one of
8620 while (i_a < len_a && i_b < len_b) {
8621 UV cp; /* The element to potentially add to the union's array */
8622 bool cp_in_set; /* is it in the the input list's set or not */
8624 /* We need to take one or the other of the two inputs for the union.
8625 * Since we are merging two sorted lists, we take the smaller of the
8626 * next items. In case of a tie, we take the one that is in its set
8627 * first. If we took one not in the set first, it would decrement the
8628 * count, possibly to 0 which would cause it to be output as ending the
8629 * range, and the next time through we would take the same number, and
8630 * output it again as beginning the next range. By doing it the
8631 * opposite way, there is no possibility that the count will be
8632 * momentarily decremented to 0, and thus the two adjoining ranges will
8633 * be seamlessly merged. (In a tie and both are in the set or both not
8634 * in the set, it doesn't matter which we take first.) */
8635 if (array_a[i_a] < array_b[i_b]
8636 || (array_a[i_a] == array_b[i_b]
8637 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8639 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8643 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8644 cp = array_b[i_b++];
8647 /* Here, have chosen which of the two inputs to look at. Only output
8648 * if the running count changes to/from 0, which marks the
8649 * beginning/end of a range in that's in the set */
8652 array_u[i_u++] = cp;
8659 array_u[i_u++] = cp;
8664 /* Here, we are finished going through at least one of the lists, which
8665 * means there is something remaining in at most one. We check if the list
8666 * that hasn't been exhausted is positioned such that we are in the middle
8667 * of a range in its set or not. (i_a and i_b point to the element beyond
8668 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8669 * is potentially more to output.
8670 * There are four cases:
8671 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8672 * in the union is entirely from the non-exhausted set.
8673 * 2) Both were in their sets, count is 2. Nothing further should
8674 * be output, as everything that remains will be in the exhausted
8675 * list's set, hence in the union; decrementing to 1 but not 0 insures
8677 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8678 * Nothing further should be output because the union includes
8679 * everything from the exhausted set. Not decrementing ensures that.
8680 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8681 * decrementing to 0 insures that we look at the remainder of the
8682 * non-exhausted set */
8683 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8684 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8689 /* The final length is what we've output so far, plus what else is about to
8690 * be output. (If 'count' is non-zero, then the input list we exhausted
8691 * has everything remaining up to the machine's limit in its set, and hence
8692 * in the union, so there will be no further output. */
8695 /* At most one of the subexpressions will be non-zero */
8696 len_u += (len_a - i_a) + (len_b - i_b);
8699 /* Set result to final length, which can change the pointer to array_u, so
8701 if (len_u != _invlist_len(u)) {
8702 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8704 array_u = invlist_array(u);
8707 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8708 * the other) ended with everything above it not in its set. That means
8709 * that the remaining part of the union is precisely the same as the
8710 * non-exhausted list, so can just copy it unchanged. (If both list were
8711 * exhausted at the same time, then the operations below will be both 0.)
8714 IV copy_count; /* At most one will have a non-zero copy count */
8715 if ((copy_count = len_a - i_a) > 0) {
8716 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8718 else if ((copy_count = len_b - i_b) > 0) {
8719 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8723 /* We may be removing a reference to one of the inputs. If so, the output
8724 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8725 * count decremented) */
8726 if (a == *output || b == *output) {
8727 assert(! invlist_is_iterating(*output));
8728 if ((SvTEMP(*output))) {
8732 SvREFCNT_dec_NN(*output);
8742 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8743 const bool complement_b, SV** i)
8745 /* Take the intersection of two inversion lists and point <i> to it. *i
8746 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8747 * the reference count to that list will be decremented if not already a
8748 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8749 * The first list, <a>, may be NULL, in which case an empty list is
8750 * returned. If <complement_b> is TRUE, the result will be the
8751 * intersection of <a> and the complement (or inversion) of <b> instead of
8754 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8755 * Richard Gillam, published by Addison-Wesley, and explained at some
8756 * length there. The preface says to incorporate its examples into your
8757 * code at your own risk. In fact, it had bugs
8759 * The algorithm is like a merge sort, and is essentially the same as the
8763 const UV* array_a; /* a's array */
8765 UV len_a; /* length of a's array */
8768 SV* r; /* the resulting intersection */
8772 UV i_a = 0; /* current index into a's array */
8776 /* running count, as explained in the algorithm source book; items are
8777 * stopped accumulating and are output when the count changes to/from 2.
8778 * The count is incremented when we start a range that's in the set, and
8779 * decremented when we start a range that's not in the set. So its range
8780 * is 0 to 2. Only when the count is 2 is something in the intersection.
8784 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8787 /* Special case if either one is empty */
8788 len_a = (a == NULL) ? 0 : _invlist_len(a);
8789 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8790 bool make_temp = FALSE;
8792 if (len_a != 0 && complement_b) {
8794 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8795 * be empty. Here, also we are using 'b's complement, which hence
8796 * must be every possible code point. Thus the intersection is
8800 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8805 *i = invlist_clone(a);
8807 /* else *i is already 'a' */
8815 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8816 * intersection must be empty */
8818 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8823 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8827 *i = _new_invlist(0);
8835 /* Here both lists exist and are non-empty */
8836 array_a = invlist_array(a);
8837 array_b = invlist_array(b);
8839 /* If are to take the intersection of 'a' with the complement of b, set it
8840 * up so are looking at b's complement. */
8843 /* To complement, we invert: if the first element is 0, remove it. To
8844 * do this, we just pretend the array starts one later */
8845 if (array_b[0] == 0) {
8851 /* But if the first element is not zero, we pretend the list starts
8852 * at the 0 that is always stored immediately before the array. */
8858 /* Size the intersection for the worst case: that the intersection ends up
8859 * fragmenting everything to be completely disjoint */
8860 r= _new_invlist(len_a + len_b);
8862 /* Will contain U+0000 iff both components do */
8863 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8864 && len_b > 0 && array_b[0] == 0);
8866 /* Go through each list item by item, stopping when exhausted one of
8868 while (i_a < len_a && i_b < len_b) {
8869 UV cp; /* The element to potentially add to the intersection's
8871 bool cp_in_set; /* Is it in the input list's set or not */
8873 /* We need to take one or the other of the two inputs for the
8874 * intersection. Since we are merging two sorted lists, we take the
8875 * smaller of the next items. In case of a tie, we take the one that
8876 * is not in its set first (a difference from the union algorithm). If
8877 * we took one in the set first, it would increment the count, possibly
8878 * to 2 which would cause it to be output as starting a range in the
8879 * intersection, and the next time through we would take that same
8880 * number, and output it again as ending the set. By doing it the
8881 * opposite of this, there is no possibility that the count will be
8882 * momentarily incremented to 2. (In a tie and both are in the set or
8883 * both not in the set, it doesn't matter which we take first.) */
8884 if (array_a[i_a] < array_b[i_b]
8885 || (array_a[i_a] == array_b[i_b]
8886 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8888 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8892 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8896 /* Here, have chosen which of the two inputs to look at. Only output
8897 * if the running count changes to/from 2, which marks the
8898 * beginning/end of a range that's in the intersection */
8902 array_r[i_r++] = cp;
8907 array_r[i_r++] = cp;
8913 /* Here, we are finished going through at least one of the lists, which
8914 * means there is something remaining in at most one. We check if the list
8915 * that has been exhausted is positioned such that we are in the middle
8916 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8917 * the ones we care about.) There are four cases:
8918 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8919 * nothing left in the intersection.
8920 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8921 * above 2. What should be output is exactly that which is in the
8922 * non-exhausted set, as everything it has is also in the intersection
8923 * set, and everything it doesn't have can't be in the intersection
8924 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8925 * gets incremented to 2. Like the previous case, the intersection is
8926 * everything that remains in the non-exhausted set.
8927 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8928 * remains 1. And the intersection has nothing more. */
8929 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8930 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8935 /* The final length is what we've output so far plus what else is in the
8936 * intersection. At most one of the subexpressions below will be non-zero
8940 len_r += (len_a - i_a) + (len_b - i_b);
8943 /* Set result to final length, which can change the pointer to array_r, so
8945 if (len_r != _invlist_len(r)) {
8946 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8948 array_r = invlist_array(r);
8951 /* Finish outputting any remaining */
8952 if (count >= 2) { /* At most one will have a non-zero copy count */
8954 if ((copy_count = len_a - i_a) > 0) {
8955 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8957 else if ((copy_count = len_b - i_b) > 0) {
8958 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8962 /* We may be removing a reference to one of the inputs. If so, the output
8963 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8964 * count decremented) */
8965 if (a == *i || b == *i) {
8966 assert(! invlist_is_iterating(*i));
8971 SvREFCNT_dec_NN(*i);
8981 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8983 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8984 * set. A pointer to the inversion list is returned. This may actually be
8985 * a new list, in which case the passed in one has been destroyed. The
8986 * passed-in inversion list can be NULL, in which case a new one is created
8987 * with just the one range in it */
8992 if (invlist == NULL) {
8993 invlist = _new_invlist(2);
8997 len = _invlist_len(invlist);
9000 /* If comes after the final entry actually in the list, can just append it
9003 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
9004 && start >= invlist_array(invlist)[len - 1]))
9006 _append_range_to_invlist(invlist, start, end);
9010 /* Here, can't just append things, create and return a new inversion list
9011 * which is the union of this range and the existing inversion list */
9012 range_invlist = _new_invlist(2);
9013 _append_range_to_invlist(range_invlist, start, end);
9015 _invlist_union(invlist, range_invlist, &invlist);
9017 /* The temporary can be freed */
9018 SvREFCNT_dec_NN(range_invlist);
9024 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9025 UV** other_elements_ptr)
9027 /* Create and return an inversion list whose contents are to be populated
9028 * by the caller. The caller gives the number of elements (in 'size') and
9029 * the very first element ('element0'). This function will set
9030 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9033 * Obviously there is some trust involved that the caller will properly
9034 * fill in the other elements of the array.
9036 * (The first element needs to be passed in, as the underlying code does
9037 * things differently depending on whether it is zero or non-zero) */
9039 SV* invlist = _new_invlist(size);
9042 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9044 _append_range_to_invlist(invlist, element0, element0);
9045 offset = *get_invlist_offset_addr(invlist);
9047 invlist_set_len(invlist, size, offset);
9048 *other_elements_ptr = invlist_array(invlist) + 1;
9054 PERL_STATIC_INLINE SV*
9055 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9056 return _add_range_to_invlist(invlist, cp, cp);
9059 #ifndef PERL_IN_XSUB_RE
9061 Perl__invlist_invert(pTHX_ SV* const invlist)
9063 /* Complement the input inversion list. This adds a 0 if the list didn't
9064 * have a zero; removes it otherwise. As described above, the data
9065 * structure is set up so that this is very efficient */
9067 PERL_ARGS_ASSERT__INVLIST_INVERT;
9069 assert(! invlist_is_iterating(invlist));
9071 /* The inverse of matching nothing is matching everything */
9072 if (_invlist_len(invlist) == 0) {
9073 _append_range_to_invlist(invlist, 0, UV_MAX);
9077 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9082 PERL_STATIC_INLINE SV*
9083 S_invlist_clone(pTHX_ SV* const invlist)
9086 /* Return a new inversion list that is a copy of the input one, which is
9087 * unchanged. The new list will not be mortal even if the old one was. */
9089 /* Need to allocate extra space to accommodate Perl's addition of a
9090 * trailing NUL to SvPV's, since it thinks they are always strings */
9091 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9092 STRLEN physical_length = SvCUR(invlist);
9093 bool offset = *(get_invlist_offset_addr(invlist));
9095 PERL_ARGS_ASSERT_INVLIST_CLONE;
9097 *(get_invlist_offset_addr(new_invlist)) = offset;
9098 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9099 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9104 PERL_STATIC_INLINE STRLEN*
9105 S_get_invlist_iter_addr(SV* invlist)
9107 /* Return the address of the UV that contains the current iteration
9110 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9112 assert(SvTYPE(invlist) == SVt_INVLIST);
9114 return &(((XINVLIST*) SvANY(invlist))->iterator);
9117 PERL_STATIC_INLINE void
9118 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9120 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9122 *get_invlist_iter_addr(invlist) = 0;
9125 PERL_STATIC_INLINE void
9126 S_invlist_iterfinish(SV* invlist)
9128 /* Terminate iterator for invlist. This is to catch development errors.
9129 * Any iteration that is interrupted before completed should call this
9130 * function. Functions that add code points anywhere else but to the end
9131 * of an inversion list assert that they are not in the middle of an
9132 * iteration. If they were, the addition would make the iteration
9133 * problematical: if the iteration hadn't reached the place where things
9134 * were being added, it would be ok */
9136 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9138 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9142 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9144 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9145 * This call sets in <*start> and <*end>, the next range in <invlist>.
9146 * Returns <TRUE> if successful and the next call will return the next
9147 * range; <FALSE> if was already at the end of the list. If the latter,
9148 * <*start> and <*end> are unchanged, and the next call to this function
9149 * will start over at the beginning of the list */
9151 STRLEN* pos = get_invlist_iter_addr(invlist);
9152 UV len = _invlist_len(invlist);
9155 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9158 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9162 array = invlist_array(invlist);
9164 *start = array[(*pos)++];
9170 *end = array[(*pos)++] - 1;
9176 PERL_STATIC_INLINE UV
9177 S_invlist_highest(SV* const invlist)
9179 /* Returns the highest code point that matches an inversion list. This API
9180 * has an ambiguity, as it returns 0 under either the highest is actually
9181 * 0, or if the list is empty. If this distinction matters to you, check
9182 * for emptiness before calling this function */
9184 UV len = _invlist_len(invlist);
9187 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9193 array = invlist_array(invlist);
9195 /* The last element in the array in the inversion list always starts a
9196 * range that goes to infinity. That range may be for code points that are
9197 * matched in the inversion list, or it may be for ones that aren't
9198 * matched. In the latter case, the highest code point in the set is one
9199 * less than the beginning of this range; otherwise it is the final element
9200 * of this range: infinity */
9201 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9203 : array[len - 1] - 1;
9206 #ifndef PERL_IN_XSUB_RE
9208 Perl__invlist_contents(pTHX_ SV* const invlist)
9210 /* Get the contents of an inversion list into a string SV so that they can
9211 * be printed out. It uses the format traditionally done for debug tracing
9215 SV* output = newSVpvs("\n");
9217 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9219 assert(! invlist_is_iterating(invlist));
9221 invlist_iterinit(invlist);
9222 while (invlist_iternext(invlist, &start, &end)) {
9223 if (end == UV_MAX) {
9224 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9226 else if (end != start) {
9227 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9231 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9239 #ifndef PERL_IN_XSUB_RE
9241 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9242 const char * const indent, SV* const invlist)
9244 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9245 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9246 * the string 'indent'. The output looks like this:
9247 [0] 0x000A .. 0x000D
9249 [4] 0x2028 .. 0x2029
9250 [6] 0x3104 .. INFINITY
9251 * This means that the first range of code points matched by the list are
9252 * 0xA through 0xD; the second range contains only the single code point
9253 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9254 * are used to define each range (except if the final range extends to
9255 * infinity, only a single element is needed). The array index of the
9256 * first element for the corresponding range is given in brackets. */
9261 PERL_ARGS_ASSERT__INVLIST_DUMP;
9263 if (invlist_is_iterating(invlist)) {
9264 Perl_dump_indent(aTHX_ level, file,
9265 "%sCan't dump inversion list because is in middle of iterating\n",
9270 invlist_iterinit(invlist);
9271 while (invlist_iternext(invlist, &start, &end)) {
9272 if (end == UV_MAX) {
9273 Perl_dump_indent(aTHX_ level, file,
9274 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9275 indent, (UV)count, start);
9277 else if (end != start) {
9278 Perl_dump_indent(aTHX_ level, file,
9279 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9280 indent, (UV)count, start, end);
9283 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9284 indent, (UV)count, start);
9291 Perl__load_PL_utf8_foldclosures (pTHX)
9293 assert(! PL_utf8_foldclosures);
9295 /* If the folds haven't been read in, call a fold function
9297 if (! PL_utf8_tofold) {
9298 U8 dummy[UTF8_MAXBYTES_CASE+1];
9300 /* This string is just a short named one above \xff */
9301 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
9302 assert(PL_utf8_tofold); /* Verify that worked */
9304 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
9308 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9310 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9312 /* Return a boolean as to if the two passed in inversion lists are
9313 * identical. The final argument, if TRUE, says to take the complement of
9314 * the second inversion list before doing the comparison */
9316 const UV* array_a = invlist_array(a);
9317 const UV* array_b = invlist_array(b);
9318 UV len_a = _invlist_len(a);
9319 UV len_b = _invlist_len(b);
9321 UV i = 0; /* current index into the arrays */
9322 bool retval = TRUE; /* Assume are identical until proven otherwise */
9324 PERL_ARGS_ASSERT__INVLISTEQ;
9326 /* If are to compare 'a' with the complement of b, set it
9327 * up so are looking at b's complement. */
9330 /* The complement of nothing is everything, so <a> would have to have
9331 * just one element, starting at zero (ending at infinity) */
9333 return (len_a == 1 && array_a[0] == 0);
9335 else if (array_b[0] == 0) {
9337 /* Otherwise, to complement, we invert. Here, the first element is
9338 * 0, just remove it. To do this, we just pretend the array starts
9346 /* But if the first element is not zero, we pretend the list starts
9347 * at the 0 that is always stored immediately before the array. */
9353 /* Make sure that the lengths are the same, as well as the final element
9354 * before looping through the remainder. (Thus we test the length, final,
9355 * and first elements right off the bat) */
9356 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9359 else for (i = 0; i < len_a - 1; i++) {
9360 if (array_a[i] != array_b[i]) {
9371 * As best we can, determine the characters that can match the start of
9372 * the given EXACTF-ish node.
9374 * Returns the invlist as a new SV*; it is the caller's responsibility to
9375 * call SvREFCNT_dec() when done with it.
9378 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
9380 const U8 * s = (U8*)STRING(node);
9381 SSize_t bytelen = STR_LEN(node);
9383 /* Start out big enough for 2 separate code points */
9384 SV* invlist = _new_invlist(4);
9386 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
9391 /* We punt and assume can match anything if the node begins
9392 * with a multi-character fold. Things are complicated. For
9393 * example, /ffi/i could match any of:
9394 * "\N{LATIN SMALL LIGATURE FFI}"
9395 * "\N{LATIN SMALL LIGATURE FF}I"
9396 * "F\N{LATIN SMALL LIGATURE FI}"
9397 * plus several other things; and making sure we have all the
9398 * possibilities is hard. */
9399 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
9400 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9403 /* Any Latin1 range character can potentially match any
9404 * other depending on the locale */
9405 if (OP(node) == EXACTFL) {
9406 _invlist_union(invlist, PL_Latin1, &invlist);
9409 /* But otherwise, it matches at least itself. We can
9410 * quickly tell if it has a distinct fold, and if so,
9411 * it matches that as well */
9412 invlist = add_cp_to_invlist(invlist, uc);
9413 if (IS_IN_SOME_FOLD_L1(uc))
9414 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
9417 /* Some characters match above-Latin1 ones under /i. This
9418 * is true of EXACTFL ones when the locale is UTF-8 */
9419 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
9420 && (! isASCII(uc) || (OP(node) != EXACTFA
9421 && OP(node) != EXACTFA_NO_TRIE)))
9423 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
9427 else { /* Pattern is UTF-8 */
9428 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
9429 STRLEN foldlen = UTF8SKIP(s);
9430 const U8* e = s + bytelen;
9433 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
9435 /* The only code points that aren't folded in a UTF EXACTFish
9436 * node are are the problematic ones in EXACTFL nodes */
9437 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
9438 /* We need to check for the possibility that this EXACTFL
9439 * node begins with a multi-char fold. Therefore we fold
9440 * the first few characters of it so that we can make that
9445 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
9447 *(d++) = (U8) toFOLD(*s);
9452 to_utf8_fold(s, d, &len);
9458 /* And set up so the code below that looks in this folded
9459 * buffer instead of the node's string */
9461 foldlen = UTF8SKIP(folded);
9465 /* When we reach here 's' points to the fold of the first
9466 * character(s) of the node; and 'e' points to far enough along
9467 * the folded string to be just past any possible multi-char
9468 * fold. 'foldlen' is the length in bytes of the first
9471 * Unlike the non-UTF-8 case, the macro for determining if a
9472 * string is a multi-char fold requires all the characters to
9473 * already be folded. This is because of all the complications
9474 * if not. Note that they are folded anyway, except in EXACTFL
9475 * nodes. Like the non-UTF case above, we punt if the node
9476 * begins with a multi-char fold */
9478 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
9479 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9481 else { /* Single char fold */
9483 /* It matches all the things that fold to it, which are
9484 * found in PL_utf8_foldclosures (including itself) */
9485 invlist = add_cp_to_invlist(invlist, uc);
9486 if (! PL_utf8_foldclosures)
9487 _load_PL_utf8_foldclosures();
9488 if ((listp = hv_fetch(PL_utf8_foldclosures,
9489 (char *) s, foldlen, FALSE)))
9491 AV* list = (AV*) *listp;
9493 for (k = 0; k <= av_tindex(list); k++) {
9494 SV** c_p = av_fetch(list, k, FALSE);
9500 /* /aa doesn't allow folds between ASCII and non- */
9501 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
9502 && isASCII(c) != isASCII(uc))
9507 invlist = add_cp_to_invlist(invlist, c);
9516 #undef HEADER_LENGTH
9517 #undef TO_INTERNAL_SIZE
9518 #undef FROM_INTERNAL_SIZE
9519 #undef INVLIST_VERSION_ID
9521 /* End of inversion list object */
9524 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9526 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9527 * constructs, and updates RExC_flags with them. On input, RExC_parse
9528 * should point to the first flag; it is updated on output to point to the
9529 * final ')' or ':'. There needs to be at least one flag, or this will
9532 /* for (?g), (?gc), and (?o) warnings; warning
9533 about (?c) will warn about (?g) -- japhy */
9535 #define WASTED_O 0x01
9536 #define WASTED_G 0x02
9537 #define WASTED_C 0x04
9538 #define WASTED_GC (WASTED_G|WASTED_C)
9539 I32 wastedflags = 0x00;
9540 U32 posflags = 0, negflags = 0;
9541 U32 *flagsp = &posflags;
9542 char has_charset_modifier = '\0';
9544 bool has_use_defaults = FALSE;
9545 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9546 int x_mod_count = 0;
9548 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9550 /* '^' as an initial flag sets certain defaults */
9551 if (UCHARAT(RExC_parse) == '^') {
9553 has_use_defaults = TRUE;
9554 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9555 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9556 ? REGEX_UNICODE_CHARSET
9557 : REGEX_DEPENDS_CHARSET);
9560 cs = get_regex_charset(RExC_flags);
9561 if (cs == REGEX_DEPENDS_CHARSET
9562 && (RExC_utf8 || RExC_uni_semantics))
9564 cs = REGEX_UNICODE_CHARSET;
9567 while (*RExC_parse) {
9568 /* && strchr("iogcmsx", *RExC_parse) */
9569 /* (?g), (?gc) and (?o) are useless here
9570 and must be globally applied -- japhy */
9571 switch (*RExC_parse) {
9573 /* Code for the imsxn flags */
9574 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
9576 case LOCALE_PAT_MOD:
9577 if (has_charset_modifier) {
9578 goto excess_modifier;
9580 else if (flagsp == &negflags) {
9583 cs = REGEX_LOCALE_CHARSET;
9584 has_charset_modifier = LOCALE_PAT_MOD;
9586 case UNICODE_PAT_MOD:
9587 if (has_charset_modifier) {
9588 goto excess_modifier;
9590 else if (flagsp == &negflags) {
9593 cs = REGEX_UNICODE_CHARSET;
9594 has_charset_modifier = UNICODE_PAT_MOD;
9596 case ASCII_RESTRICT_PAT_MOD:
9597 if (flagsp == &negflags) {
9600 if (has_charset_modifier) {
9601 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9602 goto excess_modifier;
9604 /* Doubled modifier implies more restricted */
9605 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9608 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9610 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9612 case DEPENDS_PAT_MOD:
9613 if (has_use_defaults) {
9614 goto fail_modifiers;
9616 else if (flagsp == &negflags) {
9619 else if (has_charset_modifier) {
9620 goto excess_modifier;
9623 /* The dual charset means unicode semantics if the
9624 * pattern (or target, not known until runtime) are
9625 * utf8, or something in the pattern indicates unicode
9627 cs = (RExC_utf8 || RExC_uni_semantics)
9628 ? REGEX_UNICODE_CHARSET
9629 : REGEX_DEPENDS_CHARSET;
9630 has_charset_modifier = DEPENDS_PAT_MOD;
9634 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9635 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9637 else if (has_charset_modifier == *(RExC_parse - 1)) {
9638 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9642 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9644 NOT_REACHED; /*NOTREACHED*/
9647 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9649 NOT_REACHED; /*NOTREACHED*/
9650 case ONCE_PAT_MOD: /* 'o' */
9651 case GLOBAL_PAT_MOD: /* 'g' */
9652 if (PASS2 && ckWARN(WARN_REGEXP)) {
9653 const I32 wflagbit = *RExC_parse == 'o'
9656 if (! (wastedflags & wflagbit) ) {
9657 wastedflags |= wflagbit;
9658 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9661 "Useless (%s%c) - %suse /%c modifier",
9662 flagsp == &negflags ? "?-" : "?",
9664 flagsp == &negflags ? "don't " : "",
9671 case CONTINUE_PAT_MOD: /* 'c' */
9672 if (PASS2 && ckWARN(WARN_REGEXP)) {
9673 if (! (wastedflags & WASTED_C) ) {
9674 wastedflags |= WASTED_GC;
9675 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9678 "Useless (%sc) - %suse /gc modifier",
9679 flagsp == &negflags ? "?-" : "?",
9680 flagsp == &negflags ? "don't " : ""
9685 case KEEPCOPY_PAT_MOD: /* 'p' */
9686 if (flagsp == &negflags) {
9688 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9690 *flagsp |= RXf_PMf_KEEPCOPY;
9694 /* A flag is a default iff it is following a minus, so
9695 * if there is a minus, it means will be trying to
9696 * re-specify a default which is an error */
9697 if (has_use_defaults || flagsp == &negflags) {
9698 goto fail_modifiers;
9701 wastedflags = 0; /* reset so (?g-c) warns twice */
9705 RExC_flags |= posflags;
9706 RExC_flags &= ~negflags;
9707 set_regex_charset(&RExC_flags, cs);
9708 if (RExC_flags & RXf_PMf_FOLD) {
9709 RExC_contains_i = 1;
9712 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9718 RExC_parse += SKIP_IF_CHAR(RExC_parse);
9719 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9720 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9721 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9722 NOT_REACHED; /*NOTREACHED*/
9729 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9734 - reg - regular expression, i.e. main body or parenthesized thing
9736 * Caller must absorb opening parenthesis.
9738 * Combining parenthesis handling with the base level of regular expression
9739 * is a trifle forced, but the need to tie the tails of the branches to what
9740 * follows makes it hard to avoid.
9742 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9744 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9746 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9749 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9750 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9751 needs to be restarted.
9752 Otherwise would only return NULL if regbranch() returns NULL, which
9755 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9756 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9757 * 2 is like 1, but indicates that nextchar() has been called to advance
9758 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9759 * this flag alerts us to the need to check for that */
9761 regnode *ret; /* Will be the head of the group. */
9764 regnode *ender = NULL;
9767 U32 oregflags = RExC_flags;
9768 bool have_branch = 0;
9770 I32 freeze_paren = 0;
9771 I32 after_freeze = 0;
9772 I32 num; /* numeric backreferences */
9774 char * parse_start = RExC_parse; /* MJD */
9775 char * const oregcomp_parse = RExC_parse;
9777 GET_RE_DEBUG_FLAGS_DECL;
9779 PERL_ARGS_ASSERT_REG;
9780 DEBUG_PARSE("reg ");
9782 *flagp = 0; /* Tentatively. */
9785 /* Make an OPEN node, if parenthesized. */
9788 /* Under /x, space and comments can be gobbled up between the '(' and
9789 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9790 * intervening space, as the sequence is a token, and a token should be
9792 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9794 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9795 char *start_verb = RExC_parse;
9796 STRLEN verb_len = 0;
9797 char *start_arg = NULL;
9798 unsigned char op = 0;
9800 int internal_argval = 0; /* internal_argval is only useful if
9803 if (has_intervening_patws) {
9805 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
9807 while ( *RExC_parse && *RExC_parse != ')' ) {
9808 if ( *RExC_parse == ':' ) {
9809 start_arg = RExC_parse + 1;
9815 verb_len = RExC_parse - start_verb;
9818 while ( *RExC_parse && *RExC_parse != ')' )
9820 if ( *RExC_parse != ')' )
9821 vFAIL("Unterminated verb pattern argument");
9822 if ( RExC_parse == start_arg )
9825 if ( *RExC_parse != ')' )
9826 vFAIL("Unterminated verb pattern");
9829 switch ( *start_verb ) {
9830 case 'A': /* (*ACCEPT) */
9831 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9833 internal_argval = RExC_nestroot;
9836 case 'C': /* (*COMMIT) */
9837 if ( memEQs(start_verb,verb_len,"COMMIT") )
9840 case 'F': /* (*FAIL) */
9841 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9846 case ':': /* (*:NAME) */
9847 case 'M': /* (*MARK:NAME) */
9848 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9853 case 'P': /* (*PRUNE) */
9854 if ( memEQs(start_verb,verb_len,"PRUNE") )
9857 case 'S': /* (*SKIP) */
9858 if ( memEQs(start_verb,verb_len,"SKIP") )
9861 case 'T': /* (*THEN) */
9862 /* [19:06] <TimToady> :: is then */
9863 if ( memEQs(start_verb,verb_len,"THEN") ) {
9865 RExC_seen |= REG_CUTGROUP_SEEN;
9870 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9872 "Unknown verb pattern '%"UTF8f"'",
9873 UTF8fARG(UTF, verb_len, start_verb));
9876 if ( start_arg && internal_argval ) {
9877 vFAIL3("Verb pattern '%.*s' may not have an argument",
9878 verb_len, start_verb);
9879 } else if ( argok < 0 && !start_arg ) {
9880 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9881 verb_len, start_verb);
9883 ret = reganode(pRExC_state, op, internal_argval);
9884 if ( ! internal_argval && ! SIZE_ONLY ) {
9886 SV *sv = newSVpvn( start_arg,
9887 RExC_parse - start_arg);
9888 ARG(ret) = add_data( pRExC_state,
9890 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9897 if (!internal_argval)
9898 RExC_seen |= REG_VERBARG_SEEN;
9899 } else if ( start_arg ) {
9900 vFAIL3("Verb pattern '%.*s' may not have an argument",
9901 verb_len, start_verb);
9903 ret = reg_node(pRExC_state, op);
9905 nextchar(pRExC_state);
9908 else if (*RExC_parse == '?') { /* (?...) */
9909 bool is_logical = 0;
9910 const char * const seqstart = RExC_parse;
9911 const char * endptr;
9912 if (has_intervening_patws) {
9914 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
9918 paren = *RExC_parse++;
9919 ret = NULL; /* For look-ahead/behind. */
9922 case 'P': /* (?P...) variants for those used to PCRE/Python */
9923 paren = *RExC_parse++;
9924 if ( paren == '<') /* (?P<...>) named capture */
9926 else if (paren == '>') { /* (?P>name) named recursion */
9927 goto named_recursion;
9929 else if (paren == '=') { /* (?P=...) named backref */
9930 /* this pretty much dupes the code for \k<NAME> in
9931 * regatom(), if you change this make sure you change that
9933 char* name_start = RExC_parse;
9935 SV *sv_dat = reg_scan_name(pRExC_state,
9936 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9937 if (RExC_parse == name_start || *RExC_parse != ')')
9938 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9939 vFAIL2("Sequence %.3s... not terminated",parse_start);
9942 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9943 RExC_rxi->data->data[num]=(void*)sv_dat;
9944 SvREFCNT_inc_simple_void(sv_dat);
9947 ret = reganode(pRExC_state,
9950 : (ASCII_FOLD_RESTRICTED)
9952 : (AT_LEAST_UNI_SEMANTICS)
9960 Set_Node_Offset(ret, parse_start+1);
9961 Set_Node_Cur_Length(ret, parse_start);
9963 nextchar(pRExC_state);
9967 RExC_parse += SKIP_IF_CHAR(RExC_parse);
9968 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9969 vFAIL3("Sequence (%.*s...) not recognized",
9970 RExC_parse-seqstart, seqstart);
9971 NOT_REACHED; /*NOTREACHED*/
9972 case '<': /* (?<...) */
9973 if (*RExC_parse == '!')
9975 else if (*RExC_parse != '=')
9981 case '\'': /* (?'...') */
9982 name_start= RExC_parse;
9983 svname = reg_scan_name(pRExC_state,
9984 SIZE_ONLY /* reverse test from the others */
9985 ? REG_RSN_RETURN_NAME
9986 : REG_RSN_RETURN_NULL);
9987 if (RExC_parse == name_start || *RExC_parse != paren)
9988 vFAIL2("Sequence (?%c... not terminated",
9989 paren=='>' ? '<' : paren);
9993 if (!svname) /* shouldn't happen */
9995 "panic: reg_scan_name returned NULL");
9996 if (!RExC_paren_names) {
9997 RExC_paren_names= newHV();
9998 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10000 RExC_paren_name_list= newAV();
10001 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10004 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10006 sv_dat = HeVAL(he_str);
10008 /* croak baby croak */
10010 "panic: paren_name hash element allocation failed");
10011 } else if ( SvPOK(sv_dat) ) {
10012 /* (?|...) can mean we have dupes so scan to check
10013 its already been stored. Maybe a flag indicating
10014 we are inside such a construct would be useful,
10015 but the arrays are likely to be quite small, so
10016 for now we punt -- dmq */
10017 IV count = SvIV(sv_dat);
10018 I32 *pv = (I32*)SvPVX(sv_dat);
10020 for ( i = 0 ; i < count ; i++ ) {
10021 if ( pv[i] == RExC_npar ) {
10027 pv = (I32*)SvGROW(sv_dat,
10028 SvCUR(sv_dat) + sizeof(I32)+1);
10029 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10030 pv[count] = RExC_npar;
10031 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10034 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10035 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10038 SvIV_set(sv_dat, 1);
10041 /* Yes this does cause a memory leak in debugging Perls
10043 if (!av_store(RExC_paren_name_list,
10044 RExC_npar, SvREFCNT_inc(svname)))
10045 SvREFCNT_dec_NN(svname);
10048 /*sv_dump(sv_dat);*/
10050 nextchar(pRExC_state);
10052 goto capturing_parens;
10054 RExC_seen |= REG_LOOKBEHIND_SEEN;
10055 RExC_in_lookbehind++;
10058 case '=': /* (?=...) */
10059 RExC_seen_zerolen++;
10061 case '!': /* (?!...) */
10062 RExC_seen_zerolen++;
10063 /* check if we're really just a "FAIL" assertion */
10065 nextchar(pRExC_state);
10066 if (*RExC_parse == ')') {
10067 ret=reg_node(pRExC_state, OPFAIL);
10068 nextchar(pRExC_state);
10072 case '|': /* (?|...) */
10073 /* branch reset, behave like a (?:...) except that
10074 buffers in alternations share the same numbers */
10076 after_freeze = freeze_paren = RExC_npar;
10078 case ':': /* (?:...) */
10079 case '>': /* (?>...) */
10081 case '$': /* (?$...) */
10082 case '@': /* (?@...) */
10083 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10085 case '0' : /* (?0) */
10086 case 'R' : /* (?R) */
10087 if (*RExC_parse != ')')
10088 FAIL("Sequence (?R) not terminated");
10089 ret = reg_node(pRExC_state, GOSTART);
10090 RExC_seen |= REG_GOSTART_SEEN;
10091 *flagp |= POSTPONED;
10092 nextchar(pRExC_state);
10095 /* named and numeric backreferences */
10096 case '&': /* (?&NAME) */
10097 parse_start = RExC_parse - 1;
10100 SV *sv_dat = reg_scan_name(pRExC_state,
10101 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10102 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10104 if (RExC_parse == RExC_end || *RExC_parse != ')')
10105 vFAIL("Sequence (?&... not terminated");
10106 goto gen_recurse_regop;
10109 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10111 vFAIL("Illegal pattern");
10113 goto parse_recursion;
10115 case '-': /* (?-1) */
10116 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10117 RExC_parse--; /* rewind to let it be handled later */
10121 case '1': case '2': case '3': case '4': /* (?1) */
10122 case '5': case '6': case '7': case '8': case '9':
10126 bool is_neg = FALSE;
10128 parse_start = RExC_parse - 1; /* MJD */
10129 if (*RExC_parse == '-') {
10133 if (grok_atoUV(RExC_parse, &unum, &endptr)
10137 RExC_parse = (char*)endptr;
10141 /* Some limit for num? */
10145 if (*RExC_parse!=')')
10146 vFAIL("Expecting close bracket");
10149 if ( paren == '-' ) {
10151 Diagram of capture buffer numbering.
10152 Top line is the normal capture buffer numbers
10153 Bottom line is the negative indexing as from
10157 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10161 num = RExC_npar + num;
10164 vFAIL("Reference to nonexistent group");
10166 } else if ( paren == '+' ) {
10167 num = RExC_npar + num - 1;
10170 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10172 if (num > (I32)RExC_rx->nparens) {
10174 vFAIL("Reference to nonexistent group");
10176 RExC_recurse_count++;
10177 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10178 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
10179 22, "| |", (int)(depth * 2 + 1), "",
10180 (UV)ARG(ret), (IV)ARG2L(ret)));
10182 RExC_seen |= REG_RECURSE_SEEN;
10183 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10184 Set_Node_Offset(ret, parse_start); /* MJD */
10186 *flagp |= POSTPONED;
10187 nextchar(pRExC_state);
10192 case '?': /* (??...) */
10194 if (*RExC_parse != '{') {
10195 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10196 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10198 "Sequence (%"UTF8f"...) not recognized",
10199 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10200 NOT_REACHED; /*NOTREACHED*/
10202 *flagp |= POSTPONED;
10203 paren = *RExC_parse++;
10205 case '{': /* (?{...}) */
10208 struct reg_code_block *cb;
10210 RExC_seen_zerolen++;
10212 if ( !pRExC_state->num_code_blocks
10213 || pRExC_state->code_index >= pRExC_state->num_code_blocks
10214 || pRExC_state->code_blocks[pRExC_state->code_index].start
10215 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
10218 if (RExC_pm_flags & PMf_USE_RE_EVAL)
10219 FAIL("panic: Sequence (?{...}): no code block found\n");
10220 FAIL("Eval-group not allowed at runtime, use re 'eval'");
10222 /* this is a pre-compiled code block (?{...}) */
10223 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
10224 RExC_parse = RExC_start + cb->end;
10227 if (cb->src_regex) {
10228 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
10229 RExC_rxi->data->data[n] =
10230 (void*)SvREFCNT_inc((SV*)cb->src_regex);
10231 RExC_rxi->data->data[n+1] = (void*)o;
10234 n = add_data(pRExC_state,
10235 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
10236 RExC_rxi->data->data[n] = (void*)o;
10239 pRExC_state->code_index++;
10240 nextchar(pRExC_state);
10244 ret = reg_node(pRExC_state, LOGICAL);
10246 eval = reg2Lanode(pRExC_state, EVAL,
10249 /* for later propagation into (??{})
10251 RExC_flags & RXf_PMf_COMPILETIME
10256 REGTAIL(pRExC_state, ret, eval);
10257 /* deal with the length of this later - MJD */
10260 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
10261 Set_Node_Length(ret, RExC_parse - parse_start + 1);
10262 Set_Node_Offset(ret, parse_start);
10265 case '(': /* (?(?{...})...) and (?(?=...)...) */
10268 const int DEFINE_len = sizeof("DEFINE") - 1;
10269 if (RExC_parse[0] == '?') { /* (?(?...)) */
10270 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
10271 || RExC_parse[1] == '<'
10272 || RExC_parse[1] == '{') { /* Lookahead or eval. */
10276 ret = reg_node(pRExC_state, LOGICAL);
10280 tail = reg(pRExC_state, 1, &flag, depth+1);
10281 if (flag & RESTART_UTF8) {
10282 *flagp = RESTART_UTF8;
10285 REGTAIL(pRExC_state, ret, tail);
10288 /* Fall through to ‘Unknown switch condition’ at the
10289 end of the if/else chain. */
10291 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
10292 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
10294 char ch = RExC_parse[0] == '<' ? '>' : '\'';
10295 char *name_start= RExC_parse++;
10297 SV *sv_dat=reg_scan_name(pRExC_state,
10298 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10299 if (RExC_parse == name_start || *RExC_parse != ch)
10300 vFAIL2("Sequence (?(%c... not terminated",
10301 (ch == '>' ? '<' : ch));
10304 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10305 RExC_rxi->data->data[num]=(void*)sv_dat;
10306 SvREFCNT_inc_simple_void(sv_dat);
10308 ret = reganode(pRExC_state,NGROUPP,num);
10309 goto insert_if_check_paren;
10311 else if (RExC_end - RExC_parse >= DEFINE_len
10312 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
10314 ret = reganode(pRExC_state,DEFINEP,0);
10315 RExC_parse += DEFINE_len;
10317 goto insert_if_check_paren;
10319 else if (RExC_parse[0] == 'R') {
10322 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10324 if (grok_atoUV(RExC_parse, &uv, &endptr)
10328 RExC_parse = (char*)endptr;
10330 /* else "Switch condition not recognized" below */
10331 } else if (RExC_parse[0] == '&') {
10334 sv_dat = reg_scan_name(pRExC_state,
10336 ? REG_RSN_RETURN_NULL
10337 : REG_RSN_RETURN_DATA);
10338 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10340 ret = reganode(pRExC_state,INSUBP,parno);
10341 goto insert_if_check_paren;
10343 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10348 if (grok_atoUV(RExC_parse, &uv, &endptr)
10352 RExC_parse = (char*)endptr;
10354 /* XXX else what? */
10355 ret = reganode(pRExC_state, GROUPP, parno);
10357 insert_if_check_paren:
10358 if (*(tmp = nextchar(pRExC_state)) != ')') {
10359 /* nextchar also skips comments, so undo its work
10360 * and skip over the the next character.
10363 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10364 vFAIL("Switch condition not recognized");
10367 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
10368 br = regbranch(pRExC_state, &flags, 1,depth+1);
10370 if (flags & RESTART_UTF8) {
10371 *flagp = RESTART_UTF8;
10374 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10377 REGTAIL(pRExC_state, br, reganode(pRExC_state,
10379 c = *nextchar(pRExC_state);
10380 if (flags&HASWIDTH)
10381 *flagp |= HASWIDTH;
10384 vFAIL("(?(DEFINE)....) does not allow branches");
10386 /* Fake one for optimizer. */
10387 lastbr = reganode(pRExC_state, IFTHEN, 0);
10389 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
10390 if (flags & RESTART_UTF8) {
10391 *flagp = RESTART_UTF8;
10394 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10397 REGTAIL(pRExC_state, ret, lastbr);
10398 if (flags&HASWIDTH)
10399 *flagp |= HASWIDTH;
10400 c = *nextchar(pRExC_state);
10405 if (RExC_parse>RExC_end)
10406 vFAIL("Switch (?(condition)... not terminated");
10408 vFAIL("Switch (?(condition)... contains too many branches");
10410 ender = reg_node(pRExC_state, TAIL);
10411 REGTAIL(pRExC_state, br, ender);
10413 REGTAIL(pRExC_state, lastbr, ender);
10414 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10417 REGTAIL(pRExC_state, ret, ender);
10418 RExC_size++; /* XXX WHY do we need this?!!
10419 For large programs it seems to be required
10420 but I can't figure out why. -- dmq*/
10423 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10424 vFAIL("Unknown switch condition (?(...))");
10426 case '[': /* (?[ ... ]) */
10427 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10430 RExC_parse--; /* for vFAIL to print correctly */
10431 vFAIL("Sequence (? incomplete");
10433 default: /* e.g., (?i) */
10436 parse_lparen_question_flags(pRExC_state);
10437 if (UCHARAT(RExC_parse) != ':') {
10439 nextchar(pRExC_state);
10444 nextchar(pRExC_state);
10449 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
10454 ret = reganode(pRExC_state, OPEN, parno);
10456 if (!RExC_nestroot)
10457 RExC_nestroot = parno;
10458 if (RExC_seen & REG_RECURSE_SEEN
10459 && !RExC_open_parens[parno-1])
10461 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10462 "%*s%*s Setting open paren #%"IVdf" to %d\n",
10463 22, "| |", (int)(depth * 2 + 1), "",
10464 (IV)parno, REG_NODE_NUM(ret)));
10465 RExC_open_parens[parno-1]= ret;
10468 Set_Node_Length(ret, 1); /* MJD */
10469 Set_Node_Offset(ret, RExC_parse); /* MJD */
10472 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
10481 /* Pick up the branches, linking them together. */
10482 parse_start = RExC_parse; /* MJD */
10483 br = regbranch(pRExC_state, &flags, 1,depth+1);
10485 /* branch_len = (paren != 0); */
10488 if (flags & RESTART_UTF8) {
10489 *flagp = RESTART_UTF8;
10492 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10494 if (*RExC_parse == '|') {
10495 if (!SIZE_ONLY && RExC_extralen) {
10496 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10499 reginsert(pRExC_state, BRANCH, br, depth+1);
10500 Set_Node_Length(br, paren != 0);
10501 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10505 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10507 else if (paren == ':') {
10508 *flagp |= flags&SIMPLE;
10510 if (is_open) { /* Starts with OPEN. */
10511 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10513 else if (paren != '?') /* Not Conditional */
10515 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10517 while (*RExC_parse == '|') {
10518 if (!SIZE_ONLY && RExC_extralen) {
10519 ender = reganode(pRExC_state, LONGJMP,0);
10521 /* Append to the previous. */
10522 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10525 RExC_extralen += 2; /* Account for LONGJMP. */
10526 nextchar(pRExC_state);
10527 if (freeze_paren) {
10528 if (RExC_npar > after_freeze)
10529 after_freeze = RExC_npar;
10530 RExC_npar = freeze_paren;
10532 br = regbranch(pRExC_state, &flags, 0, depth+1);
10535 if (flags & RESTART_UTF8) {
10536 *flagp = RESTART_UTF8;
10539 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10541 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10543 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10546 if (have_branch || paren != ':') {
10547 /* Make a closing node, and hook it on the end. */
10550 ender = reg_node(pRExC_state, TAIL);
10553 ender = reganode(pRExC_state, CLOSE, parno);
10554 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10555 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10556 "%*s%*s Setting close paren #%"IVdf" to %d\n",
10557 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
10558 RExC_close_parens[parno-1]= ender;
10559 if (RExC_nestroot == parno)
10562 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10563 Set_Node_Length(ender,1); /* MJD */
10569 *flagp &= ~HASWIDTH;
10572 ender = reg_node(pRExC_state, SUCCEED);
10575 ender = reg_node(pRExC_state, END);
10577 assert(!RExC_opend); /* there can only be one! */
10578 RExC_opend = ender;
10582 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10583 DEBUG_PARSE_MSG("lsbr");
10584 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
10585 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10586 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10587 SvPV_nolen_const(RExC_mysv1),
10588 (IV)REG_NODE_NUM(lastbr),
10589 SvPV_nolen_const(RExC_mysv2),
10590 (IV)REG_NODE_NUM(ender),
10591 (IV)(ender - lastbr)
10594 REGTAIL(pRExC_state, lastbr, ender);
10596 if (have_branch && !SIZE_ONLY) {
10597 char is_nothing= 1;
10599 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10601 /* Hook the tails of the branches to the closing node. */
10602 for (br = ret; br; br = regnext(br)) {
10603 const U8 op = PL_regkind[OP(br)];
10604 if (op == BRANCH) {
10605 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10606 if ( OP(NEXTOPER(br)) != NOTHING
10607 || regnext(NEXTOPER(br)) != ender)
10610 else if (op == BRANCHJ) {
10611 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10612 /* for now we always disable this optimisation * /
10613 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10614 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10620 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10621 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10622 DEBUG_PARSE_MSG("NADA");
10623 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
10624 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10625 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10626 SvPV_nolen_const(RExC_mysv1),
10627 (IV)REG_NODE_NUM(ret),
10628 SvPV_nolen_const(RExC_mysv2),
10629 (IV)REG_NODE_NUM(ender),
10634 if (OP(ender) == TAIL) {
10639 for ( opt= br + 1; opt < ender ; opt++ )
10640 OP(opt)= OPTIMIZED;
10641 NEXT_OFF(br)= ender - br;
10649 static const char parens[] = "=!<,>";
10651 if (paren && (p = strchr(parens, paren))) {
10652 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10653 int flag = (p - parens) > 1;
10656 node = SUSPEND, flag = 0;
10657 reginsert(pRExC_state, node,ret, depth+1);
10658 Set_Node_Cur_Length(ret, parse_start);
10659 Set_Node_Offset(ret, parse_start + 1);
10661 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10665 /* Check for proper termination. */
10667 /* restore original flags, but keep (?p) */
10668 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10669 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10670 RExC_parse = oregcomp_parse;
10671 vFAIL("Unmatched (");
10674 else if (!paren && RExC_parse < RExC_end) {
10675 if (*RExC_parse == ')') {
10677 vFAIL("Unmatched )");
10680 FAIL("Junk on end of regexp"); /* "Can't happen". */
10681 NOT_REACHED; /* NOTREACHED */
10684 if (RExC_in_lookbehind) {
10685 RExC_in_lookbehind--;
10687 if (after_freeze > RExC_npar)
10688 RExC_npar = after_freeze;
10693 - regbranch - one alternative of an | operator
10695 * Implements the concatenation operator.
10697 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10701 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10704 regnode *chain = NULL;
10706 I32 flags = 0, c = 0;
10707 GET_RE_DEBUG_FLAGS_DECL;
10709 PERL_ARGS_ASSERT_REGBRANCH;
10711 DEBUG_PARSE("brnc");
10716 if (!SIZE_ONLY && RExC_extralen)
10717 ret = reganode(pRExC_state, BRANCHJ,0);
10719 ret = reg_node(pRExC_state, BRANCH);
10720 Set_Node_Length(ret, 1);
10724 if (!first && SIZE_ONLY)
10725 RExC_extralen += 1; /* BRANCHJ */
10727 *flagp = WORST; /* Tentatively. */
10730 nextchar(pRExC_state);
10731 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10732 flags &= ~TRYAGAIN;
10733 latest = regpiece(pRExC_state, &flags,depth+1);
10734 if (latest == NULL) {
10735 if (flags & TRYAGAIN)
10737 if (flags & RESTART_UTF8) {
10738 *flagp = RESTART_UTF8;
10741 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10743 else if (ret == NULL)
10745 *flagp |= flags&(HASWIDTH|POSTPONED);
10746 if (chain == NULL) /* First piece. */
10747 *flagp |= flags&SPSTART;
10749 /* FIXME adding one for every branch after the first is probably
10750 * excessive now we have TRIE support. (hv) */
10752 REGTAIL(pRExC_state, chain, latest);
10757 if (chain == NULL) { /* Loop ran zero times. */
10758 chain = reg_node(pRExC_state, NOTHING);
10763 *flagp |= flags&SIMPLE;
10770 - regpiece - something followed by possible [*+?]
10772 * Note that the branching code sequences used for ? and the general cases
10773 * of * and + are somewhat optimized: they use the same NOTHING node as
10774 * both the endmarker for their branch list and the body of the last branch.
10775 * It might seem that this node could be dispensed with entirely, but the
10776 * endmarker role is not redundant.
10778 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10780 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10784 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10790 const char * const origparse = RExC_parse;
10792 I32 max = REG_INFTY;
10793 #ifdef RE_TRACK_PATTERN_OFFSETS
10796 const char *maxpos = NULL;
10799 /* Save the original in case we change the emitted regop to a FAIL. */
10800 regnode * const orig_emit = RExC_emit;
10802 GET_RE_DEBUG_FLAGS_DECL;
10804 PERL_ARGS_ASSERT_REGPIECE;
10806 DEBUG_PARSE("piec");
10808 ret = regatom(pRExC_state, &flags,depth+1);
10810 if (flags & (TRYAGAIN|RESTART_UTF8))
10811 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10813 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10819 if (op == '{' && regcurly(RExC_parse)) {
10821 #ifdef RE_TRACK_PATTERN_OFFSETS
10822 parse_start = RExC_parse; /* MJD */
10824 next = RExC_parse + 1;
10825 while (isDIGIT(*next) || *next == ',') {
10826 if (*next == ',') {
10834 if (*next == '}') { /* got one */
10835 const char* endptr;
10839 if (isDIGIT(*RExC_parse)) {
10840 if (!grok_atoUV(RExC_parse, &uv, &endptr))
10841 vFAIL("Invalid quantifier in {,}");
10842 if (uv >= REG_INFTY)
10843 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10848 if (*maxpos == ',')
10851 maxpos = RExC_parse;
10852 if (isDIGIT(*maxpos)) {
10853 if (!grok_atoUV(maxpos, &uv, &endptr))
10854 vFAIL("Invalid quantifier in {,}");
10855 if (uv >= REG_INFTY)
10856 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10859 max = REG_INFTY; /* meaning "infinity" */
10862 nextchar(pRExC_state);
10863 if (max < min) { /* If can't match, warn and optimize to fail
10867 /* We can't back off the size because we have to reserve
10868 * enough space for all the things we are about to throw
10869 * away, but we can shrink it by the ammount we are about
10870 * to re-use here */
10871 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10874 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10875 RExC_emit = orig_emit;
10877 ret = reg_node(pRExC_state, OPFAIL);
10880 else if (min == max
10881 && RExC_parse < RExC_end
10882 && (*RExC_parse == '?' || *RExC_parse == '+'))
10885 ckWARN2reg(RExC_parse + 1,
10886 "Useless use of greediness modifier '%c'",
10889 /* Absorb the modifier, so later code doesn't see nor use
10891 nextchar(pRExC_state);
10895 if ((flags&SIMPLE)) {
10896 MARK_NAUGHTY_EXP(2, 2);
10897 reginsert(pRExC_state, CURLY, ret, depth+1);
10898 Set_Node_Offset(ret, parse_start+1); /* MJD */
10899 Set_Node_Cur_Length(ret, parse_start);
10902 regnode * const w = reg_node(pRExC_state, WHILEM);
10905 REGTAIL(pRExC_state, ret, w);
10906 if (!SIZE_ONLY && RExC_extralen) {
10907 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10908 reginsert(pRExC_state, NOTHING,ret, depth+1);
10909 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10911 reginsert(pRExC_state, CURLYX,ret, depth+1);
10913 Set_Node_Offset(ret, parse_start+1);
10914 Set_Node_Length(ret,
10915 op == '{' ? (RExC_parse - parse_start) : 1);
10917 if (!SIZE_ONLY && RExC_extralen)
10918 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10919 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10921 RExC_whilem_seen++, RExC_extralen += 3;
10922 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
10929 *flagp |= HASWIDTH;
10931 ARG1_SET(ret, (U16)min);
10932 ARG2_SET(ret, (U16)max);
10934 if (max == REG_INFTY)
10935 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10941 if (!ISMULT1(op)) {
10946 #if 0 /* Now runtime fix should be reliable. */
10948 /* if this is reinstated, don't forget to put this back into perldiag:
10950 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10952 (F) The part of the regexp subject to either the * or + quantifier
10953 could match an empty string. The {#} shows in the regular
10954 expression about where the problem was discovered.
10958 if (!(flags&HASWIDTH) && op != '?')
10959 vFAIL("Regexp *+ operand could be empty");
10962 #ifdef RE_TRACK_PATTERN_OFFSETS
10963 parse_start = RExC_parse;
10965 nextchar(pRExC_state);
10967 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10969 if (op == '*' && (flags&SIMPLE)) {
10970 reginsert(pRExC_state, STAR, ret, depth+1);
10973 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10975 else if (op == '*') {
10979 else if (op == '+' && (flags&SIMPLE)) {
10980 reginsert(pRExC_state, PLUS, ret, depth+1);
10983 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10985 else if (op == '+') {
10989 else if (op == '?') {
10994 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10995 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10996 ckWARN2reg(RExC_parse,
10997 "%"UTF8f" matches null string many times",
10998 UTF8fARG(UTF, (RExC_parse >= origparse
10999 ? RExC_parse - origparse
11002 (void)ReREFCNT_inc(RExC_rx_sv);
11005 if (RExC_parse < RExC_end && *RExC_parse == '?') {
11006 nextchar(pRExC_state);
11007 reginsert(pRExC_state, MINMOD, ret, depth+1);
11008 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11011 if (RExC_parse < RExC_end && *RExC_parse == '+') {
11013 nextchar(pRExC_state);
11014 ender = reg_node(pRExC_state, SUCCEED);
11015 REGTAIL(pRExC_state, ret, ender);
11016 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11018 ender = reg_node(pRExC_state, TAIL);
11019 REGTAIL(pRExC_state, ret, ender);
11022 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
11024 vFAIL("Nested quantifiers");
11031 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11039 /* This routine teases apart the various meanings of \N and returns
11040 * accordingly. The input parameters constrain which meaning(s) is/are valid
11041 * in the current context.
11043 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11045 * If <code_point_p> is not NULL, the context is expecting the result to be a
11046 * single code point. If this \N instance turns out to a single code point,
11047 * the function returns TRUE and sets *code_point_p to that code point.
11049 * If <node_p> is not NULL, the context is expecting the result to be one of
11050 * the things representable by a regnode. If this \N instance turns out to be
11051 * one such, the function generates the regnode, returns TRUE and sets *node_p
11052 * to point to that regnode.
11054 * If this instance of \N isn't legal in any context, this function will
11055 * generate a fatal error and not return.
11057 * On input, RExC_parse should point to the first char following the \N at the
11058 * time of the call. On successful return, RExC_parse will have been updated
11059 * to point to just after the sequence identified by this routine. Also
11060 * *flagp has been updated as needed.
11062 * When there is some problem with the current context and this \N instance,
11063 * the function returns FALSE, without advancing RExC_parse, nor setting
11064 * *node_p, nor *code_point_p, nor *flagp.
11066 * If <cp_count> is not NULL, the caller wants to know the length (in code
11067 * points) that this \N sequence matches. This is set even if the function
11068 * returns FALSE, as detailed below.
11070 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11072 * Probably the most common case is for the \N to specify a single code point.
11073 * *cp_count will be set to 1, and *code_point_p will be set to that code
11076 * Another possibility is for the input to be an empty \N{}, which for
11077 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11078 * will be set to a generated NOTHING node.
11080 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11081 * set to 0. *node_p will be set to a generated REG_ANY node.
11083 * The fourth possibility is that \N resolves to a sequence of more than one
11084 * code points. *cp_count will be set to the number of code points in the
11085 * sequence. *node_p * will be set to a generated node returned by this
11086 * function calling S_reg().
11088 * The final possibility, which happens only when the fourth one would
11089 * otherwise be in effect, is that one of those code points requires the
11090 * pattern to be recompiled as UTF-8. The function returns FALSE, and sets
11091 * the RESTART_UTF8 flag in *flagp. When this happens, the caller needs to
11092 * desist from continuing parsing, and return this information to its caller.
11093 * This is not set for when there is only one code point, as this can be
11094 * called as part of an ANYOF node, and they can store above-Latin1 code
11095 * points without the pattern having to be in UTF-8.
11097 * For non-single-quoted regexes, the tokenizer has resolved character and
11098 * sequence names inside \N{...} into their Unicode values, normalizing the
11099 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11100 * hex-represented code points in the sequence. This is done there because
11101 * the names can vary based on what charnames pragma is in scope at the time,
11102 * so we need a way to take a snapshot of what they resolve to at the time of
11103 * the original parse. [perl #56444].
11105 * That parsing is skipped for single-quoted regexes, so we may here get
11106 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11107 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11108 * is legal and handled here. The code point is Unicode, and has to be
11109 * translated into the native character set for non-ASCII platforms.
11110 * the tokenizer passes the \N sequence through unchanged; this code will not
11111 * attempt to determine this nor expand those, instead raising a syntax error.
11114 char * endbrace; /* points to '}' following the name */
11115 char *endchar; /* Points to '.' or '}' ending cur char in the input
11117 char* p; /* Temporary */
11119 GET_RE_DEBUG_FLAGS_DECL;
11121 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11123 GET_RE_DEBUG_FLAGS;
11125 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11126 assert(! (node_p && cp_count)); /* At most 1 should be set */
11128 if (cp_count) { /* Initialize return for the most common case */
11132 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11133 * modifier. The other meanings do not, so use a temporary until we find
11134 * out which we are being called with */
11135 p = (RExC_flags & RXf_PMf_EXTENDED)
11136 ? regpatws(pRExC_state, RExC_parse,
11137 TRUE) /* means recognize comments */
11140 /* Disambiguate between \N meaning a named character versus \N meaning
11141 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11142 * quantifier, or there is no a '{' at all */
11143 if (*p != '{' || regcurly(p)) {
11152 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
11154 nextchar(pRExC_state);
11155 *node_p = reg_node(pRExC_state, REG_ANY);
11156 *flagp |= HASWIDTH|SIMPLE;
11158 Set_Node_Length(*node_p, 1); /* MJD */
11162 /* Here, we have decided it should be a named character or sequence */
11164 /* The test above made sure that the next real character is a '{', but
11165 * under the /x modifier, it could be separated by space (or a comment and
11166 * \n) and this is not allowed (for consistency with \x{...} and the
11167 * tokenizer handling of \N{NAME}). */
11168 if (*RExC_parse != '{') {
11169 vFAIL("Missing braces on \\N{}");
11172 RExC_parse++; /* Skip past the '{' */
11174 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
11175 || ! (endbrace == RExC_parse /* nothing between the {} */
11176 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11177 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
11180 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
11181 vFAIL("\\N{NAME} must be resolved by the lexer");
11184 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
11186 if (endbrace == RExC_parse) { /* empty: \N{} */
11190 nextchar(pRExC_state);
11195 *node_p = reg_node(pRExC_state,NOTHING);
11199 RExC_parse += 2; /* Skip past the 'U+' */
11201 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11203 /* Code points are separated by dots. If none, there is only one code
11204 * point, and is terminated by the brace */
11206 if (endchar >= endbrace) {
11207 STRLEN length_of_hex;
11208 I32 grok_hex_flags;
11210 /* Here, exactly one code point. If that isn't what is wanted, fail */
11211 if (! code_point_p) {
11216 /* Convert code point from hex */
11217 length_of_hex = (STRLEN)(endchar - RExC_parse);
11218 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
11219 | PERL_SCAN_DISALLOW_PREFIX
11221 /* No errors in the first pass (See [perl
11222 * #122671].) We let the code below find the
11223 * errors when there are multiple chars. */
11225 ? PERL_SCAN_SILENT_ILLDIGIT
11228 /* This routine is the one place where both single- and double-quotish
11229 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
11230 * must be converted to native. */
11231 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
11236 /* The tokenizer should have guaranteed validity, but it's possible to
11237 * bypass it by using single quoting, so check. Don't do the check
11238 * here when there are multiple chars; we do it below anyway. */
11239 if (length_of_hex == 0
11240 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
11242 RExC_parse += length_of_hex; /* Includes all the valid */
11243 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
11244 ? UTF8SKIP(RExC_parse)
11246 /* Guard against malformed utf8 */
11247 if (RExC_parse >= endchar) {
11248 RExC_parse = endchar;
11250 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11253 RExC_parse = endbrace + 1;
11256 else { /* Is a multiple character sequence */
11257 SV * substitute_parse;
11259 char *orig_end = RExC_end;
11262 /* Count the code points, if desired, in the sequence */
11265 while (RExC_parse < endbrace) {
11266 /* Point to the beginning of the next character in the sequence. */
11267 RExC_parse = endchar + 1;
11268 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11273 /* Fail if caller doesn't want to handle a multi-code-point sequence.
11274 * But don't backup up the pointer if the caller want to know how many
11275 * code points there are (they can then handle things) */
11283 /* What is done here is to convert this to a sub-pattern of the form
11284 * \x{char1}\x{char2}... and then call reg recursively to parse it
11285 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
11286 * while not having to worry about special handling that some code
11287 * points may have. */
11289 substitute_parse = newSVpvs("?:");
11291 while (RExC_parse < endbrace) {
11293 /* Convert to notation the rest of the code understands */
11294 sv_catpv(substitute_parse, "\\x{");
11295 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
11296 sv_catpv(substitute_parse, "}");
11298 /* Point to the beginning of the next character in the sequence. */
11299 RExC_parse = endchar + 1;
11300 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11303 sv_catpv(substitute_parse, ")");
11305 RExC_parse = SvPV(substitute_parse, len);
11307 /* Don't allow empty number */
11308 if (len < (STRLEN) 8) {
11309 RExC_parse = endbrace;
11310 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11312 RExC_end = RExC_parse + len;
11314 /* The values are Unicode, and therefore not subject to recoding, but
11315 * have to be converted to native on a non-Unicode (meaning non-ASCII)
11317 RExC_override_recoding = 1;
11319 RExC_recode_x_to_native = 1;
11323 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
11324 if (flags & RESTART_UTF8) {
11325 *flagp = RESTART_UTF8;
11328 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
11331 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11334 /* Restore the saved values */
11335 RExC_parse = endbrace;
11336 RExC_end = orig_end;
11337 RExC_override_recoding = 0;
11339 RExC_recode_x_to_native = 0;
11342 SvREFCNT_dec_NN(substitute_parse);
11343 nextchar(pRExC_state);
11353 * It returns the code point in utf8 for the value in *encp.
11354 * value: a code value in the source encoding
11355 * encp: a pointer to an Encode object
11357 * If the result from Encode is not a single character,
11358 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
11361 S_reg_recode(pTHX_ const char value, SV **encp)
11364 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
11365 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
11366 const STRLEN newlen = SvCUR(sv);
11367 UV uv = UNICODE_REPLACEMENT;
11369 PERL_ARGS_ASSERT_REG_RECODE;
11373 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
11376 if (!newlen || numlen != newlen) {
11377 uv = UNICODE_REPLACEMENT;
11383 PERL_STATIC_INLINE U8
11384 S_compute_EXACTish(RExC_state_t *pRExC_state)
11388 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
11396 op = get_regex_charset(RExC_flags);
11397 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
11398 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
11399 been, so there is no hole */
11402 return op + EXACTF;
11405 PERL_STATIC_INLINE void
11406 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
11407 regnode *node, I32* flagp, STRLEN len, UV code_point,
11410 /* This knows the details about sizing an EXACTish node, setting flags for
11411 * it (by setting <*flagp>, and potentially populating it with a single
11414 * If <len> (the length in bytes) is non-zero, this function assumes that
11415 * the node has already been populated, and just does the sizing. In this
11416 * case <code_point> should be the final code point that has already been
11417 * placed into the node. This value will be ignored except that under some
11418 * circumstances <*flagp> is set based on it.
11420 * If <len> is zero, the function assumes that the node is to contain only
11421 * the single character given by <code_point> and calculates what <len>
11422 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
11423 * additionally will populate the node's STRING with <code_point> or its
11426 * In both cases <*flagp> is appropriately set
11428 * It knows that under FOLD, the Latin Sharp S and UTF characters above
11429 * 255, must be folded (the former only when the rules indicate it can
11432 * When it does the populating, it looks at the flag 'downgradable'. If
11433 * true with a node that folds, it checks if the single code point
11434 * participates in a fold, and if not downgrades the node to an EXACT.
11435 * This helps the optimizer */
11437 bool len_passed_in = cBOOL(len != 0);
11438 U8 character[UTF8_MAXBYTES_CASE+1];
11440 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
11442 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
11443 * sizing difference, and is extra work that is thrown away */
11444 if (downgradable && ! PASS2) {
11445 downgradable = FALSE;
11448 if (! len_passed_in) {
11450 if (UVCHR_IS_INVARIANT(code_point)) {
11451 if (LOC || ! FOLD) { /* /l defers folding until runtime */
11452 *character = (U8) code_point;
11454 else { /* Here is /i and not /l. (toFOLD() is defined on just
11455 ASCII, which isn't the same thing as INVARIANT on
11456 EBCDIC, but it works there, as the extra invariants
11457 fold to themselves) */
11458 *character = toFOLD((U8) code_point);
11460 /* We can downgrade to an EXACT node if this character
11461 * isn't a folding one. Note that this assumes that
11462 * nothing above Latin1 folds to some other invariant than
11463 * one of these alphabetics; otherwise we would also have
11465 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11466 * || ASCII_FOLD_RESTRICTED))
11468 if (downgradable && PL_fold[code_point] == code_point) {
11474 else if (FOLD && (! LOC
11475 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
11476 { /* Folding, and ok to do so now */
11477 UV folded = _to_uni_fold_flags(
11481 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11482 ? FOLD_FLAGS_NOMIX_ASCII
11485 && folded == code_point /* This quickly rules out many
11486 cases, avoiding the
11487 _invlist_contains_cp() overhead
11489 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11496 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11498 /* Not folding this cp, and can output it directly */
11499 *character = UTF8_TWO_BYTE_HI(code_point);
11500 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11504 uvchr_to_utf8( character, code_point);
11505 len = UTF8SKIP(character);
11507 } /* Else pattern isn't UTF8. */
11509 *character = (U8) code_point;
11511 } /* Else is folded non-UTF8 */
11512 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11514 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11515 * comments at join_exact()); */
11516 *character = (U8) code_point;
11519 /* Can turn into an EXACT node if we know the fold at compile time,
11520 * and it folds to itself and doesn't particpate in other folds */
11523 && PL_fold_latin1[code_point] == code_point
11524 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11525 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11529 } /* else is Sharp s. May need to fold it */
11530 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11532 *(character + 1) = 's';
11536 *character = LATIN_SMALL_LETTER_SHARP_S;
11542 RExC_size += STR_SZ(len);
11545 RExC_emit += STR_SZ(len);
11546 STR_LEN(node) = len;
11547 if (! len_passed_in) {
11548 Copy((char *) character, STRING(node), len, char);
11552 *flagp |= HASWIDTH;
11554 /* A single character node is SIMPLE, except for the special-cased SHARP S
11556 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11557 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11558 || ! FOLD || ! DEPENDS_SEMANTICS))
11563 /* The OP may not be well defined in PASS1 */
11564 if (PASS2 && OP(node) == EXACTFL) {
11565 RExC_contains_locale = 1;
11570 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
11571 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11574 S_backref_value(char *p)
11576 const char* endptr;
11578 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
11585 - regatom - the lowest level
11587 Try to identify anything special at the start of the pattern. If there
11588 is, then handle it as required. This may involve generating a single regop,
11589 such as for an assertion; or it may involve recursing, such as to
11590 handle a () structure.
11592 If the string doesn't start with something special then we gobble up
11593 as much literal text as we can.
11595 Once we have been able to handle whatever type of thing started the
11596 sequence, we return.
11598 Note: we have to be careful with escapes, as they can be both literal
11599 and special, and in the case of \10 and friends, context determines which.
11601 A summary of the code structure is:
11603 switch (first_byte) {
11604 cases for each special:
11605 handle this special;
11608 switch (2nd byte) {
11609 cases for each unambiguous special:
11610 handle this special;
11612 cases for each ambigous special/literal:
11614 if (special) handle here
11616 default: // unambiguously literal:
11619 default: // is a literal char
11622 create EXACTish node for literal;
11623 while (more input and node isn't full) {
11624 switch (input_byte) {
11625 cases for each special;
11626 make sure parse pointer is set so that the next call to
11627 regatom will see this special first
11628 goto loopdone; // EXACTish node terminated by prev. char
11630 append char to EXACTISH node;
11632 get next input byte;
11636 return the generated node;
11638 Specifically there are two separate switches for handling
11639 escape sequences, with the one for handling literal escapes requiring
11640 a dummy entry for all of the special escapes that are actually handled
11643 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11645 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11647 Otherwise does not return NULL.
11651 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11653 regnode *ret = NULL;
11655 char *parse_start = RExC_parse;
11660 GET_RE_DEBUG_FLAGS_DECL;
11662 *flagp = WORST; /* Tentatively. */
11664 DEBUG_PARSE("atom");
11666 PERL_ARGS_ASSERT_REGATOM;
11669 switch ((U8)*RExC_parse) {
11671 RExC_seen_zerolen++;
11672 nextchar(pRExC_state);
11673 if (RExC_flags & RXf_PMf_MULTILINE)
11674 ret = reg_node(pRExC_state, MBOL);
11676 ret = reg_node(pRExC_state, SBOL);
11677 Set_Node_Length(ret, 1); /* MJD */
11680 nextchar(pRExC_state);
11682 RExC_seen_zerolen++;
11683 if (RExC_flags & RXf_PMf_MULTILINE)
11684 ret = reg_node(pRExC_state, MEOL);
11686 ret = reg_node(pRExC_state, SEOL);
11687 Set_Node_Length(ret, 1); /* MJD */
11690 nextchar(pRExC_state);
11691 if (RExC_flags & RXf_PMf_SINGLELINE)
11692 ret = reg_node(pRExC_state, SANY);
11694 ret = reg_node(pRExC_state, REG_ANY);
11695 *flagp |= HASWIDTH|SIMPLE;
11697 Set_Node_Length(ret, 1); /* MJD */
11701 char * const oregcomp_parse = ++RExC_parse;
11702 ret = regclass(pRExC_state, flagp,depth+1,
11703 FALSE, /* means parse the whole char class */
11704 TRUE, /* allow multi-char folds */
11705 FALSE, /* don't silence non-portable warnings. */
11706 (bool) RExC_strict,
11708 if (*RExC_parse != ']') {
11709 RExC_parse = oregcomp_parse;
11710 vFAIL("Unmatched [");
11713 if (*flagp & RESTART_UTF8)
11715 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11718 nextchar(pRExC_state);
11719 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11723 nextchar(pRExC_state);
11724 ret = reg(pRExC_state, 2, &flags,depth+1);
11726 if (flags & TRYAGAIN) {
11727 if (RExC_parse == RExC_end) {
11728 /* Make parent create an empty node if needed. */
11729 *flagp |= TRYAGAIN;
11734 if (flags & RESTART_UTF8) {
11735 *flagp = RESTART_UTF8;
11738 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11741 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11745 if (flags & TRYAGAIN) {
11746 *flagp |= TRYAGAIN;
11749 vFAIL("Internal urp");
11750 /* Supposed to be caught earlier. */
11756 vFAIL("Quantifier follows nothing");
11761 This switch handles escape sequences that resolve to some kind
11762 of special regop and not to literal text. Escape sequnces that
11763 resolve to literal text are handled below in the switch marked
11766 Every entry in this switch *must* have a corresponding entry
11767 in the literal escape switch. However, the opposite is not
11768 required, as the default for this switch is to jump to the
11769 literal text handling code.
11771 switch ((U8)*++RExC_parse) {
11772 /* Special Escapes */
11774 RExC_seen_zerolen++;
11775 ret = reg_node(pRExC_state, SBOL);
11776 /* SBOL is shared with /^/ so we set the flags so we can tell
11777 * /\A/ from /^/ in split. We check ret because first pass we
11778 * have no regop struct to set the flags on. */
11782 goto finish_meta_pat;
11784 ret = reg_node(pRExC_state, GPOS);
11785 RExC_seen |= REG_GPOS_SEEN;
11787 goto finish_meta_pat;
11789 RExC_seen_zerolen++;
11790 ret = reg_node(pRExC_state, KEEPS);
11792 /* XXX:dmq : disabling in-place substitution seems to
11793 * be necessary here to avoid cases of memory corruption, as
11794 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11796 RExC_seen |= REG_LOOKBEHIND_SEEN;
11797 goto finish_meta_pat;
11799 ret = reg_node(pRExC_state, SEOL);
11801 RExC_seen_zerolen++; /* Do not optimize RE away */
11802 goto finish_meta_pat;
11804 ret = reg_node(pRExC_state, EOS);
11806 RExC_seen_zerolen++; /* Do not optimize RE away */
11807 goto finish_meta_pat;
11809 ret = reg_node(pRExC_state, CANY);
11810 RExC_seen |= REG_CANY_SEEN;
11811 *flagp |= HASWIDTH|SIMPLE;
11813 ckWARNdep(RExC_parse+1, "\\C is deprecated");
11815 goto finish_meta_pat;
11817 ret = reg_node(pRExC_state, CLUMP);
11818 *flagp |= HASWIDTH;
11819 goto finish_meta_pat;
11825 arg = ANYOF_WORDCHAR;
11833 regex_charset charset = get_regex_charset(RExC_flags);
11835 RExC_seen_zerolen++;
11836 RExC_seen |= REG_LOOKBEHIND_SEEN;
11837 op = BOUND + charset;
11839 if (op == BOUNDL) {
11840 RExC_contains_locale = 1;
11843 ret = reg_node(pRExC_state, op);
11845 if (*(RExC_parse + 1) != '{') {
11846 FLAGS(ret) = TRADITIONAL_BOUND;
11847 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
11853 char name = *RExC_parse;
11856 endbrace = strchr(RExC_parse, '}');
11859 vFAIL2("Missing right brace on \\%c{}", name);
11861 /* XXX Need to decide whether to take spaces or not. Should be
11862 * consistent with \p{}, but that currently is SPACE, which
11863 * means vertical too, which seems wrong
11864 * while (isBLANK(*RExC_parse)) {
11867 if (endbrace == RExC_parse) {
11868 RExC_parse++; /* After the '}' */
11869 vFAIL2("Empty \\%c{}", name);
11871 length = endbrace - RExC_parse;
11872 /*while (isBLANK(*(RExC_parse + length - 1))) {
11875 switch (*RExC_parse) {
11878 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
11880 goto bad_bound_type;
11882 FLAGS(ret) = GCB_BOUND;
11885 if (length != 2 || *(RExC_parse + 1) != 'b') {
11886 goto bad_bound_type;
11888 FLAGS(ret) = SB_BOUND;
11891 if (length != 2 || *(RExC_parse + 1) != 'b') {
11892 goto bad_bound_type;
11894 FLAGS(ret) = WB_BOUND;
11898 RExC_parse = endbrace;
11900 "'%"UTF8f"' is an unknown bound type",
11901 UTF8fARG(UTF, length, endbrace - length));
11902 NOT_REACHED; /*NOTREACHED*/
11904 RExC_parse = endbrace;
11905 RExC_uni_semantics = 1;
11907 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
11911 /* Don't have to worry about UTF-8, in this message because
11912 * to get here the contents of the \b must be ASCII */
11913 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
11914 "Using /u for '%.*s' instead of /%s",
11916 endbrace - length + 1,
11917 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
11918 ? ASCII_RESTRICT_PAT_MODS
11919 : ASCII_MORE_RESTRICT_PAT_MODS);
11923 if (PASS2 && invert) {
11924 OP(ret) += NBOUND - BOUND;
11926 goto finish_meta_pat;
11934 if (! DEPENDS_SEMANTICS) {
11938 /* \d doesn't have any matches in the upper Latin1 range, hence /d
11939 * is equivalent to /u. Changing to /u saves some branches at
11942 goto join_posix_op_known;
11945 ret = reg_node(pRExC_state, LNBREAK);
11946 *flagp |= HASWIDTH|SIMPLE;
11947 goto finish_meta_pat;
11955 goto join_posix_op_known;
11961 arg = ANYOF_VERTWS;
11963 goto join_posix_op_known;
11973 op = POSIXD + get_regex_charset(RExC_flags);
11974 if (op > POSIXA) { /* /aa is same as /a */
11977 else if (op == POSIXL) {
11978 RExC_contains_locale = 1;
11981 join_posix_op_known:
11984 op += NPOSIXD - POSIXD;
11987 ret = reg_node(pRExC_state, op);
11989 FLAGS(ret) = namedclass_to_classnum(arg);
11992 *flagp |= HASWIDTH|SIMPLE;
11996 nextchar(pRExC_state);
11997 Set_Node_Length(ret, 2); /* MJD */
12003 char* parse_start = RExC_parse - 2;
12008 ret = regclass(pRExC_state, flagp,depth+1,
12009 TRUE, /* means just parse this element */
12010 FALSE, /* don't allow multi-char folds */
12011 FALSE, /* don't silence non-portable warnings.
12012 It would be a bug if these returned
12014 (bool) RExC_strict,
12016 /* regclass() can only return RESTART_UTF8 if multi-char folds
12019 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12024 Set_Node_Offset(ret, parse_start + 2);
12025 Set_Node_Cur_Length(ret, parse_start);
12026 nextchar(pRExC_state);
12030 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12031 * \N{...} evaluates to a sequence of more than one code points).
12032 * The function call below returns a regnode, which is our result.
12033 * The parameters cause it to fail if the \N{} evaluates to a
12034 * single code point; we handle those like any other literal. The
12035 * reason that the multicharacter case is handled here and not as
12036 * part of the EXACtish code is because of quantifiers. In
12037 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12038 * this way makes that Just Happen. dmq.
12039 * join_exact() will join this up with adjacent EXACTish nodes
12040 * later on, if appropriate. */
12042 if (grok_bslash_N(pRExC_state,
12043 &ret, /* Want a regnode returned */
12044 NULL, /* Fail if evaluates to a single code
12046 NULL, /* Don't need a count of how many code
12054 if (*flagp & RESTART_UTF8)
12059 case 'k': /* Handle \k<NAME> and \k'NAME' */
12062 char ch= RExC_parse[1];
12063 if (ch != '<' && ch != '\'' && ch != '{') {
12065 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12066 vFAIL2("Sequence %.2s... not terminated",parse_start);
12068 /* this pretty much dupes the code for (?P=...) in reg(), if
12069 you change this make sure you change that */
12070 char* name_start = (RExC_parse += 2);
12072 SV *sv_dat = reg_scan_name(pRExC_state,
12073 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
12074 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
12075 if (RExC_parse == name_start || *RExC_parse != ch)
12076 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12077 vFAIL2("Sequence %.3s... not terminated",parse_start);
12080 num = add_data( pRExC_state, STR_WITH_LEN("S"));
12081 RExC_rxi->data->data[num]=(void*)sv_dat;
12082 SvREFCNT_inc_simple_void(sv_dat);
12086 ret = reganode(pRExC_state,
12089 : (ASCII_FOLD_RESTRICTED)
12091 : (AT_LEAST_UNI_SEMANTICS)
12097 *flagp |= HASWIDTH;
12099 /* override incorrect value set in reganode MJD */
12100 Set_Node_Offset(ret, parse_start+1);
12101 Set_Node_Cur_Length(ret, parse_start);
12102 nextchar(pRExC_state);
12108 case '1': case '2': case '3': case '4':
12109 case '5': case '6': case '7': case '8': case '9':
12114 if (*RExC_parse == 'g') {
12118 if (*RExC_parse == '{') {
12122 if (*RExC_parse == '-') {
12126 if (hasbrace && !isDIGIT(*RExC_parse)) {
12127 if (isrel) RExC_parse--;
12129 goto parse_named_seq;
12132 num = S_backref_value(RExC_parse);
12134 vFAIL("Reference to invalid group 0");
12135 else if (num == I32_MAX) {
12136 if (isDIGIT(*RExC_parse))
12137 vFAIL("Reference to nonexistent group");
12139 vFAIL("Unterminated \\g... pattern");
12143 num = RExC_npar - num;
12145 vFAIL("Reference to nonexistent or unclosed group");
12149 num = S_backref_value(RExC_parse);
12150 /* bare \NNN might be backref or octal - if it is larger
12151 * than or equal RExC_npar then it is assumed to be an
12152 * octal escape. Note RExC_npar is +1 from the actual
12153 * number of parens. */
12154 /* Note we do NOT check if num == I32_MAX here, as that is
12155 * handled by the RExC_npar check */
12158 /* any numeric escape < 10 is always a backref */
12160 /* any numeric escape < RExC_npar is a backref */
12161 && num >= RExC_npar
12162 /* cannot be an octal escape if it starts with 8 */
12163 && *RExC_parse != '8'
12164 /* cannot be an octal escape it it starts with 9 */
12165 && *RExC_parse != '9'
12168 /* Probably not a backref, instead likely to be an
12169 * octal character escape, e.g. \35 or \777.
12170 * The above logic should make it obvious why using
12171 * octal escapes in patterns is problematic. - Yves */
12176 /* At this point RExC_parse points at a numeric escape like
12177 * \12 or \88 or something similar, which we should NOT treat
12178 * as an octal escape. It may or may not be a valid backref
12179 * escape. For instance \88888888 is unlikely to be a valid
12182 #ifdef RE_TRACK_PATTERN_OFFSETS
12183 char * const parse_start = RExC_parse - 1; /* MJD */
12185 while (isDIGIT(*RExC_parse))
12188 if (*RExC_parse != '}')
12189 vFAIL("Unterminated \\g{...} pattern");
12193 if (num > (I32)RExC_rx->nparens)
12194 vFAIL("Reference to nonexistent group");
12197 ret = reganode(pRExC_state,
12200 : (ASCII_FOLD_RESTRICTED)
12202 : (AT_LEAST_UNI_SEMANTICS)
12208 *flagp |= HASWIDTH;
12210 /* override incorrect value set in reganode MJD */
12211 Set_Node_Offset(ret, parse_start+1);
12212 Set_Node_Cur_Length(ret, parse_start);
12214 nextchar(pRExC_state);
12219 if (RExC_parse >= RExC_end)
12220 FAIL("Trailing \\");
12223 /* Do not generate "unrecognized" warnings here, we fall
12224 back into the quick-grab loop below */
12231 if (RExC_flags & RXf_PMf_EXTENDED) {
12232 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
12233 if (RExC_parse < RExC_end)
12240 parse_start = RExC_parse - 1;
12249 #define MAX_NODE_STRING_SIZE 127
12250 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
12252 U8 upper_parse = MAX_NODE_STRING_SIZE;
12253 U8 node_type = compute_EXACTish(pRExC_state);
12254 bool next_is_quantifier;
12255 char * oldp = NULL;
12257 /* We can convert EXACTF nodes to EXACTFU if they contain only
12258 * characters that match identically regardless of the target
12259 * string's UTF8ness. The reason to do this is that EXACTF is not
12260 * trie-able, EXACTFU is.
12262 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
12263 * contain only above-Latin1 characters (hence must be in UTF8),
12264 * which don't participate in folds with Latin1-range characters,
12265 * as the latter's folds aren't known until runtime. (We don't
12266 * need to figure this out until pass 2) */
12267 bool maybe_exactfu = PASS2
12268 && (node_type == EXACTF || node_type == EXACTFL);
12270 /* If a folding node contains only code points that don't
12271 * participate in folds, it can be changed into an EXACT node,
12272 * which allows the optimizer more things to look for */
12275 ret = reg_node(pRExC_state, node_type);
12277 /* In pass1, folded, we use a temporary buffer instead of the
12278 * actual node, as the node doesn't exist yet */
12279 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
12285 /* We do the EXACTFish to EXACT node only if folding. (And we
12286 * don't need to figure this out until pass 2) */
12287 maybe_exact = FOLD && PASS2;
12289 /* XXX The node can hold up to 255 bytes, yet this only goes to
12290 * 127. I (khw) do not know why. Keeping it somewhat less than
12291 * 255 allows us to not have to worry about overflow due to
12292 * converting to utf8 and fold expansion, but that value is
12293 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
12294 * split up by this limit into a single one using the real max of
12295 * 255. Even at 127, this breaks under rare circumstances. If
12296 * folding, we do not want to split a node at a character that is a
12297 * non-final in a multi-char fold, as an input string could just
12298 * happen to want to match across the node boundary. The join
12299 * would solve that problem if the join actually happens. But a
12300 * series of more than two nodes in a row each of 127 would cause
12301 * the first join to succeed to get to 254, but then there wouldn't
12302 * be room for the next one, which could at be one of those split
12303 * multi-char folds. I don't know of any fool-proof solution. One
12304 * could back off to end with only a code point that isn't such a
12305 * non-final, but it is possible for there not to be any in the
12307 for (p = RExC_parse - 1;
12308 len < upper_parse && p < RExC_end;
12313 if (RExC_flags & RXf_PMf_EXTENDED)
12314 p = regpatws(pRExC_state, p,
12315 TRUE); /* means recognize comments */
12326 /* Literal Escapes Switch
12328 This switch is meant to handle escape sequences that
12329 resolve to a literal character.
12331 Every escape sequence that represents something
12332 else, like an assertion or a char class, is handled
12333 in the switch marked 'Special Escapes' above in this
12334 routine, but also has an entry here as anything that
12335 isn't explicitly mentioned here will be treated as
12336 an unescaped equivalent literal.
12339 switch ((U8)*++p) {
12340 /* These are all the special escapes. */
12341 case 'A': /* Start assertion */
12342 case 'b': case 'B': /* Word-boundary assertion*/
12343 case 'C': /* Single char !DANGEROUS! */
12344 case 'd': case 'D': /* digit class */
12345 case 'g': case 'G': /* generic-backref, pos assertion */
12346 case 'h': case 'H': /* HORIZWS */
12347 case 'k': case 'K': /* named backref, keep marker */
12348 case 'p': case 'P': /* Unicode property */
12349 case 'R': /* LNBREAK */
12350 case 's': case 'S': /* space class */
12351 case 'v': case 'V': /* VERTWS */
12352 case 'w': case 'W': /* word class */
12353 case 'X': /* eXtended Unicode "combining
12354 character sequence" */
12355 case 'z': case 'Z': /* End of line/string assertion */
12359 /* Anything after here is an escape that resolves to a
12360 literal. (Except digits, which may or may not)
12366 case 'N': /* Handle a single-code point named character. */
12367 RExC_parse = p + 1;
12368 if (! grok_bslash_N(pRExC_state,
12369 NULL, /* Fail if evaluates to
12370 anything other than a
12371 single code point */
12372 &ender, /* The returned single code
12374 NULL, /* Don't need a count of
12375 how many code points */
12379 if (*flagp & RESTART_UTF8)
12380 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12382 /* Here, it wasn't a single code point. Go close
12383 * up this EXACTish node. The switch() prior to
12384 * this switch handles the other cases */
12385 RExC_parse = p = oldp;
12389 if (ender > 0xff) {
12406 ender = ESC_NATIVE;
12416 const char* error_msg;
12418 bool valid = grok_bslash_o(&p,
12421 PASS2, /* out warnings */
12422 (bool) RExC_strict,
12423 TRUE, /* Output warnings
12428 RExC_parse = p; /* going to die anyway; point
12429 to exact spot of failure */
12433 if (IN_ENCODING && ender < 0x100) {
12434 goto recode_encoding;
12436 if (ender > 0xff) {
12443 UV result = UV_MAX; /* initialize to erroneous
12445 const char* error_msg;
12447 bool valid = grok_bslash_x(&p,
12450 PASS2, /* out warnings */
12451 (bool) RExC_strict,
12452 TRUE, /* Silence warnings
12457 RExC_parse = p; /* going to die anyway; point
12458 to exact spot of failure */
12463 if (ender < 0x100) {
12465 if (RExC_recode_x_to_native) {
12466 ender = LATIN1_TO_NATIVE(ender);
12471 goto recode_encoding;
12481 ender = grok_bslash_c(*p++, PASS2);
12483 case '8': case '9': /* must be a backreference */
12485 /* we have an escape like \8 which cannot be an octal escape
12486 * so we exit the loop, and let the outer loop handle this
12487 * escape which may or may not be a legitimate backref. */
12489 case '1': case '2': case '3':case '4':
12490 case '5': case '6': case '7':
12491 /* When we parse backslash escapes there is ambiguity
12492 * between backreferences and octal escapes. Any escape
12493 * from \1 - \9 is a backreference, any multi-digit
12494 * escape which does not start with 0 and which when
12495 * evaluated as decimal could refer to an already
12496 * parsed capture buffer is a back reference. Anything
12499 * Note this implies that \118 could be interpreted as
12500 * 118 OR as "\11" . "8" depending on whether there
12501 * were 118 capture buffers defined already in the
12504 /* NOTE, RExC_npar is 1 more than the actual number of
12505 * parens we have seen so far, hence the < RExC_npar below. */
12507 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
12508 { /* Not to be treated as an octal constant, go
12516 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12518 ender = grok_oct(p, &numlen, &flags, NULL);
12519 if (ender > 0xff) {
12523 if (PASS2 /* like \08, \178 */
12526 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
12528 reg_warn_non_literal_string(
12530 form_short_octal_warning(p, numlen));
12533 if (IN_ENCODING && ender < 0x100)
12534 goto recode_encoding;
12537 if (! RExC_override_recoding) {
12538 SV* enc = _get_encoding();
12539 ender = reg_recode((const char)(U8)ender, &enc);
12541 ckWARNreg(p, "Invalid escape in the specified encoding");
12547 FAIL("Trailing \\");
12550 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
12551 /* Include any { following the alpha to emphasize
12552 * that it could be part of an escape at some point
12554 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
12555 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
12557 goto normal_default;
12558 } /* End of switch on '\' */
12561 /* Currently we don't warn when the lbrace is at the start
12562 * of a construct. This catches it in the middle of a
12563 * literal string, or when its the first thing after
12564 * something like "\b" */
12566 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
12568 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
12571 default: /* A literal character */
12573 if (UTF8_IS_START(*p) && UTF) {
12575 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
12576 &numlen, UTF8_ALLOW_DEFAULT);
12582 } /* End of switch on the literal */
12584 /* Here, have looked at the literal character and <ender>
12585 * contains its ordinal, <p> points to the character after it
12588 if ( RExC_flags & RXf_PMf_EXTENDED)
12589 p = regpatws(pRExC_state, p,
12590 TRUE); /* means recognize comments */
12592 /* If the next thing is a quantifier, it applies to this
12593 * character only, which means that this character has to be in
12594 * its own node and can't just be appended to the string in an
12595 * existing node, so if there are already other characters in
12596 * the node, close the node with just them, and set up to do
12597 * this character again next time through, when it will be the
12598 * only thing in its new node */
12599 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
12605 if (! FOLD) { /* The simple case, just append the literal */
12607 /* In the sizing pass, we need only the size of the
12608 * character we are appending, hence we can delay getting
12609 * its representation until PASS2. */
12612 const STRLEN unilen = UNISKIP(ender);
12615 /* We have to subtract 1 just below (and again in
12616 * the corresponding PASS2 code) because the loop
12617 * increments <len> each time, as all but this path
12618 * (and one other) through it add a single byte to
12619 * the EXACTish node. But these paths would change
12620 * len to be the correct final value, so cancel out
12621 * the increment that follows */
12627 } else { /* PASS2 */
12630 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
12631 len += (char *) new_s - s - 1;
12632 s = (char *) new_s;
12635 *(s++) = (char) ender;
12639 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
12641 /* Here are folding under /l, and the code point is
12642 * problematic. First, we know we can't simplify things */
12643 maybe_exact = FALSE;
12644 maybe_exactfu = FALSE;
12646 /* A problematic code point in this context means that its
12647 * fold isn't known until runtime, so we can't fold it now.
12648 * (The non-problematic code points are the above-Latin1
12649 * ones that fold to also all above-Latin1. Their folds
12650 * don't vary no matter what the locale is.) But here we
12651 * have characters whose fold depends on the locale.
12652 * Unlike the non-folding case above, we have to keep track
12653 * of these in the sizing pass, so that we can make sure we
12654 * don't split too-long nodes in the middle of a potential
12655 * multi-char fold. And unlike the regular fold case
12656 * handled in the else clauses below, we don't actually
12657 * fold and don't have special cases to consider. What we
12658 * do for both passes is the PASS2 code for non-folding */
12659 goto not_fold_common;
12661 else /* A regular FOLD code point */
12663 /* See comments for join_exact() as to why we fold this
12664 * non-UTF at compile time */
12665 || (node_type == EXACTFU
12666 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12668 /* Here, are folding and are not UTF-8 encoded; therefore
12669 * the character must be in the range 0-255, and is not /l
12670 * (Not /l because we already handled these under /l in
12671 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
12672 if (IS_IN_SOME_FOLD_L1(ender)) {
12673 maybe_exact = FALSE;
12675 /* See if the character's fold differs between /d and
12676 * /u. This includes the multi-char fold SHARP S to
12679 && (PL_fold[ender] != PL_fold_latin1[ender]
12680 || ender == LATIN_SMALL_LETTER_SHARP_S
12682 && isALPHA_FOLD_EQ(ender, 's')
12683 && isALPHA_FOLD_EQ(*(s-1), 's'))))
12685 maybe_exactfu = FALSE;
12689 /* Even when folding, we store just the input character, as
12690 * we have an array that finds its fold quickly */
12691 *(s++) = (char) ender;
12693 else { /* FOLD and UTF */
12694 /* Unlike the non-fold case, we do actually have to
12695 * calculate the results here in pass 1. This is for two
12696 * reasons, the folded length may be longer than the
12697 * unfolded, and we have to calculate how many EXACTish
12698 * nodes it will take; and we may run out of room in a node
12699 * in the middle of a potential multi-char fold, and have
12700 * to back off accordingly. */
12703 if (isASCII_uni(ender)) {
12704 folded = toFOLD(ender);
12705 *(s)++ = (U8) folded;
12710 folded = _to_uni_fold_flags(
12714 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12715 ? FOLD_FLAGS_NOMIX_ASCII
12719 /* The loop increments <len> each time, as all but this
12720 * path (and one other) through it add a single byte to
12721 * the EXACTish node. But this one has changed len to
12722 * be the correct final value, so subtract one to
12723 * cancel out the increment that follows */
12724 len += foldlen - 1;
12726 /* If this node only contains non-folding code points so
12727 * far, see if this new one is also non-folding */
12729 if (folded != ender) {
12730 maybe_exact = FALSE;
12733 /* Here the fold is the original; we have to check
12734 * further to see if anything folds to it */
12735 if (_invlist_contains_cp(PL_utf8_foldable,
12738 maybe_exact = FALSE;
12745 if (next_is_quantifier) {
12747 /* Here, the next input is a quantifier, and to get here,
12748 * the current character is the only one in the node.
12749 * Also, here <len> doesn't include the final byte for this
12755 } /* End of loop through literal characters */
12757 /* Here we have either exhausted the input or ran out of room in
12758 * the node. (If we encountered a character that can't be in the
12759 * node, transfer is made directly to <loopdone>, and so we
12760 * wouldn't have fallen off the end of the loop.) In the latter
12761 * case, we artificially have to split the node into two, because
12762 * we just don't have enough space to hold everything. This
12763 * creates a problem if the final character participates in a
12764 * multi-character fold in the non-final position, as a match that
12765 * should have occurred won't, due to the way nodes are matched,
12766 * and our artificial boundary. So back off until we find a non-
12767 * problematic character -- one that isn't at the beginning or
12768 * middle of such a fold. (Either it doesn't participate in any
12769 * folds, or appears only in the final position of all the folds it
12770 * does participate in.) A better solution with far fewer false
12771 * positives, and that would fill the nodes more completely, would
12772 * be to actually have available all the multi-character folds to
12773 * test against, and to back-off only far enough to be sure that
12774 * this node isn't ending with a partial one. <upper_parse> is set
12775 * further below (if we need to reparse the node) to include just
12776 * up through that final non-problematic character that this code
12777 * identifies, so when it is set to less than the full node, we can
12778 * skip the rest of this */
12779 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12781 const STRLEN full_len = len;
12783 assert(len >= MAX_NODE_STRING_SIZE);
12785 /* Here, <s> points to the final byte of the final character.
12786 * Look backwards through the string until find a non-
12787 * problematic character */
12791 /* This has no multi-char folds to non-UTF characters */
12792 if (ASCII_FOLD_RESTRICTED) {
12796 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12800 if (! PL_NonL1NonFinalFold) {
12801 PL_NonL1NonFinalFold = _new_invlist_C_array(
12802 NonL1_Perl_Non_Final_Folds_invlist);
12805 /* Point to the first byte of the final character */
12806 s = (char *) utf8_hop((U8 *) s, -1);
12808 while (s >= s0) { /* Search backwards until find
12809 non-problematic char */
12810 if (UTF8_IS_INVARIANT(*s)) {
12812 /* There are no ascii characters that participate
12813 * in multi-char folds under /aa. In EBCDIC, the
12814 * non-ascii invariants are all control characters,
12815 * so don't ever participate in any folds. */
12816 if (ASCII_FOLD_RESTRICTED
12817 || ! IS_NON_FINAL_FOLD(*s))
12822 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12823 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12829 else if (! _invlist_contains_cp(
12830 PL_NonL1NonFinalFold,
12831 valid_utf8_to_uvchr((U8 *) s, NULL)))
12836 /* Here, the current character is problematic in that
12837 * it does occur in the non-final position of some
12838 * fold, so try the character before it, but have to
12839 * special case the very first byte in the string, so
12840 * we don't read outside the string */
12841 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12842 } /* End of loop backwards through the string */
12844 /* If there were only problematic characters in the string,
12845 * <s> will point to before s0, in which case the length
12846 * should be 0, otherwise include the length of the
12847 * non-problematic character just found */
12848 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12851 /* Here, have found the final character, if any, that is
12852 * non-problematic as far as ending the node without splitting
12853 * it across a potential multi-char fold. <len> contains the
12854 * number of bytes in the node up-to and including that
12855 * character, or is 0 if there is no such character, meaning
12856 * the whole node contains only problematic characters. In
12857 * this case, give up and just take the node as-is. We can't
12862 /* If the node ends in an 's' we make sure it stays EXACTF,
12863 * as if it turns into an EXACTFU, it could later get
12864 * joined with another 's' that would then wrongly match
12866 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
12868 maybe_exactfu = FALSE;
12872 /* Here, the node does contain some characters that aren't
12873 * problematic. If one such is the final character in the
12874 * node, we are done */
12875 if (len == full_len) {
12878 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12880 /* If the final character is problematic, but the
12881 * penultimate is not, back-off that last character to
12882 * later start a new node with it */
12887 /* Here, the final non-problematic character is earlier
12888 * in the input than the penultimate character. What we do
12889 * is reparse from the beginning, going up only as far as
12890 * this final ok one, thus guaranteeing that the node ends
12891 * in an acceptable character. The reason we reparse is
12892 * that we know how far in the character is, but we don't
12893 * know how to correlate its position with the input parse.
12894 * An alternate implementation would be to build that
12895 * correlation as we go along during the original parse,
12896 * but that would entail extra work for every node, whereas
12897 * this code gets executed only when the string is too
12898 * large for the node, and the final two characters are
12899 * problematic, an infrequent occurrence. Yet another
12900 * possible strategy would be to save the tail of the
12901 * string, and the next time regatom is called, initialize
12902 * with that. The problem with this is that unless you
12903 * back off one more character, you won't be guaranteed
12904 * regatom will get called again, unless regbranch,
12905 * regpiece ... are also changed. If you do back off that
12906 * extra character, so that there is input guaranteed to
12907 * force calling regatom, you can't handle the case where
12908 * just the first character in the node is acceptable. I
12909 * (khw) decided to try this method which doesn't have that
12910 * pitfall; if performance issues are found, we can do a
12911 * combination of the current approach plus that one */
12917 } /* End of verifying node ends with an appropriate char */
12919 loopdone: /* Jumped to when encounters something that shouldn't be
12922 /* I (khw) don't know if you can get here with zero length, but the
12923 * old code handled this situation by creating a zero-length EXACT
12924 * node. Might as well be NOTHING instead */
12930 /* If 'maybe_exact' is still set here, means there are no
12931 * code points in the node that participate in folds;
12932 * similarly for 'maybe_exactfu' and code points that match
12933 * differently depending on UTF8ness of the target string
12934 * (for /u), or depending on locale for /l */
12940 else if (maybe_exactfu) {
12946 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12947 FALSE /* Don't look to see if could
12948 be turned into an EXACT
12949 node, as we have already
12954 RExC_parse = p - 1;
12955 Set_Node_Cur_Length(ret, parse_start);
12956 nextchar(pRExC_state);
12958 /* len is STRLEN which is unsigned, need to copy to signed */
12961 vFAIL("Internal disaster");
12964 } /* End of label 'defchar:' */
12966 } /* End of giant switch on input character */
12972 S_regpatws(RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12974 /* Returns the next non-pattern-white space, non-comment character (the
12975 * latter only if 'recognize_comment is true) in the string p, which is
12976 * ended by RExC_end. See also reg_skipcomment */
12977 const char *e = RExC_end;
12979 PERL_ARGS_ASSERT_REGPATWS;
12983 if ((len = is_PATWS_safe(p, e, UTF))) {
12986 else if (recognize_comment && *p == '#') {
12987 p = reg_skipcomment(pRExC_state, p);
12996 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12998 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12999 * sets up the bitmap and any flags, removing those code points from the
13000 * inversion list, setting it to NULL should it become completely empty */
13002 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13003 assert(PL_regkind[OP(node)] == ANYOF);
13005 ANYOF_BITMAP_ZERO(node);
13006 if (*invlist_ptr) {
13008 /* This gets set if we actually need to modify things */
13009 bool change_invlist = FALSE;
13013 /* Start looking through *invlist_ptr */
13014 invlist_iterinit(*invlist_ptr);
13015 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13019 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13020 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13022 else if (end >= NUM_ANYOF_CODE_POINTS) {
13023 ANYOF_FLAGS(node) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
13026 /* Quit if are above what we should change */
13027 if (start >= NUM_ANYOF_CODE_POINTS) {
13031 change_invlist = TRUE;
13033 /* Set all the bits in the range, up to the max that we are doing */
13034 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13036 : NUM_ANYOF_CODE_POINTS - 1;
13037 for (i = start; i <= (int) high; i++) {
13038 if (! ANYOF_BITMAP_TEST(node, i)) {
13039 ANYOF_BITMAP_SET(node, i);
13043 invlist_iterfinish(*invlist_ptr);
13045 /* Done with loop; remove any code points that are in the bitmap from
13046 * *invlist_ptr; similarly for code points above the bitmap if we have
13047 * a flag to match all of them anyways */
13048 if (change_invlist) {
13049 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13051 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13052 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13055 /* If have completely emptied it, remove it completely */
13056 if (_invlist_len(*invlist_ptr) == 0) {
13057 SvREFCNT_dec_NN(*invlist_ptr);
13058 *invlist_ptr = NULL;
13063 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13064 Character classes ([:foo:]) can also be negated ([:^foo:]).
13065 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13066 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13067 but trigger failures because they are currently unimplemented. */
13069 #define POSIXCC_DONE(c) ((c) == ':')
13070 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13071 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13073 PERL_STATIC_INLINE I32
13074 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
13076 I32 namedclass = OOB_NAMEDCLASS;
13078 PERL_ARGS_ASSERT_REGPPOSIXCC;
13080 if (value == '[' && RExC_parse + 1 < RExC_end &&
13081 /* I smell either [: or [= or [. -- POSIX has been here, right? */
13082 POSIXCC(UCHARAT(RExC_parse)))
13084 const char c = UCHARAT(RExC_parse);
13085 char* const s = RExC_parse++;
13087 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
13089 if (RExC_parse == RExC_end) {
13092 /* Try to give a better location for the error (than the end of
13093 * the string) by looking for the matching ']' */
13095 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
13098 vFAIL2("Unmatched '%c' in POSIX class", c);
13100 /* Grandfather lone [:, [=, [. */
13104 const char* const t = RExC_parse++; /* skip over the c */
13107 if (UCHARAT(RExC_parse) == ']') {
13108 const char *posixcc = s + 1;
13109 RExC_parse++; /* skip over the ending ] */
13112 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
13113 const I32 skip = t - posixcc;
13115 /* Initially switch on the length of the name. */
13118 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
13119 this is the Perl \w
13121 namedclass = ANYOF_WORDCHAR;
13124 /* Names all of length 5. */
13125 /* alnum alpha ascii blank cntrl digit graph lower
13126 print punct space upper */
13127 /* Offset 4 gives the best switch position. */
13128 switch (posixcc[4]) {
13130 if (memEQ(posixcc, "alph", 4)) /* alpha */
13131 namedclass = ANYOF_ALPHA;
13134 if (memEQ(posixcc, "spac", 4)) /* space */
13135 namedclass = ANYOF_SPACE;
13138 if (memEQ(posixcc, "grap", 4)) /* graph */
13139 namedclass = ANYOF_GRAPH;
13142 if (memEQ(posixcc, "asci", 4)) /* ascii */
13143 namedclass = ANYOF_ASCII;
13146 if (memEQ(posixcc, "blan", 4)) /* blank */
13147 namedclass = ANYOF_BLANK;
13150 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
13151 namedclass = ANYOF_CNTRL;
13154 if (memEQ(posixcc, "alnu", 4)) /* alnum */
13155 namedclass = ANYOF_ALPHANUMERIC;
13158 if (memEQ(posixcc, "lowe", 4)) /* lower */
13159 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
13160 else if (memEQ(posixcc, "uppe", 4)) /* upper */
13161 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
13164 if (memEQ(posixcc, "digi", 4)) /* digit */
13165 namedclass = ANYOF_DIGIT;
13166 else if (memEQ(posixcc, "prin", 4)) /* print */
13167 namedclass = ANYOF_PRINT;
13168 else if (memEQ(posixcc, "punc", 4)) /* punct */
13169 namedclass = ANYOF_PUNCT;
13174 if (memEQ(posixcc, "xdigit", 6))
13175 namedclass = ANYOF_XDIGIT;
13179 if (namedclass == OOB_NAMEDCLASS)
13181 "POSIX class [:%"UTF8f":] unknown",
13182 UTF8fARG(UTF, t - s - 1, s + 1));
13184 /* The #defines are structured so each complement is +1 to
13185 * the normal one */
13189 assert (posixcc[skip] == ':');
13190 assert (posixcc[skip+1] == ']');
13191 } else if (!SIZE_ONLY) {
13192 /* [[=foo=]] and [[.foo.]] are still future. */
13194 /* adjust RExC_parse so the warning shows after
13195 the class closes */
13196 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
13198 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
13201 /* Maternal grandfather:
13202 * "[:" ending in ":" but not in ":]" */
13204 vFAIL("Unmatched '[' in POSIX class");
13207 /* Grandfather lone [:, [=, [. */
13217 S_could_it_be_a_POSIX_class(RExC_state_t *pRExC_state)
13219 /* This applies some heuristics at the current parse position (which should
13220 * be at a '[') to see if what follows might be intended to be a [:posix:]
13221 * class. It returns true if it really is a posix class, of course, but it
13222 * also can return true if it thinks that what was intended was a posix
13223 * class that didn't quite make it.
13225 * It will return true for
13227 * [:alphanumerics] (as long as the ] isn't followed immediately by a
13228 * ')' indicating the end of the (?[
13229 * [:any garbage including %^&$ punctuation:]
13231 * This is designed to be called only from S_handle_regex_sets; it could be
13232 * easily adapted to be called from the spot at the beginning of regclass()
13233 * that checks to see in a normal bracketed class if the surrounding []
13234 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
13235 * change long-standing behavior, so I (khw) didn't do that */
13236 char* p = RExC_parse + 1;
13237 char first_char = *p;
13239 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
13241 assert(*(p - 1) == '[');
13243 if (! POSIXCC(first_char)) {
13248 while (p < RExC_end && isWORDCHAR(*p)) p++;
13250 if (p >= RExC_end) {
13254 if (p - RExC_parse > 2 /* Got at least 1 word character */
13255 && (*p == first_char
13256 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
13261 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
13264 && p - RExC_parse > 2 /* [:] evaluates to colon;
13265 [::] is a bad posix class. */
13266 && first_char == *(p - 1));
13270 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
13271 I32 *flagp, U32 depth,
13272 char * const oregcomp_parse)
13274 /* Handle the (?[...]) construct to do set operations */
13277 UV start, end; /* End points of code point ranges */
13279 char *save_end, *save_parse;
13284 const bool save_fold = FOLD;
13286 GET_RE_DEBUG_FLAGS_DECL;
13288 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
13291 vFAIL("(?[...]) not valid in locale");
13293 RExC_uni_semantics = 1;
13295 /* This will return only an ANYOF regnode, or (unlikely) something smaller
13296 * (such as EXACT). Thus we can skip most everything if just sizing. We
13297 * call regclass to handle '[]' so as to not have to reinvent its parsing
13298 * rules here (throwing away the size it computes each time). And, we exit
13299 * upon an unescaped ']' that isn't one ending a regclass. To do both
13300 * these things, we need to realize that something preceded by a backslash
13301 * is escaped, so we have to keep track of backslashes */
13303 Perl_ck_warner_d(aTHX_
13304 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
13305 "The regex_sets feature is experimental" REPORT_LOCATION,
13306 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
13308 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
13309 RExC_precomp + (RExC_parse - RExC_precomp)));
13312 UV depth = 0; /* how many nested (?[...]) constructs */
13314 while (RExC_parse < RExC_end) {
13315 SV* current = NULL;
13316 RExC_parse = regpatws(pRExC_state, RExC_parse,
13317 TRUE); /* means recognize comments */
13318 switch (*RExC_parse) {
13320 if (RExC_parse[1] == '[') depth++, RExC_parse++;
13325 /* Skip the next byte (which could cause us to end up in
13326 * the middle of a UTF-8 character, but since none of those
13327 * are confusable with anything we currently handle in this
13328 * switch (invariants all), it's safe. We'll just hit the
13329 * default: case next time and keep on incrementing until
13330 * we find one of the invariants we do handle. */
13335 /* If this looks like it is a [:posix:] class, leave the
13336 * parse pointer at the '[' to fool regclass() into
13337 * thinking it is part of a '[[:posix:]]'. That function
13338 * will use strict checking to force a syntax error if it
13339 * doesn't work out to a legitimate class */
13340 bool is_posix_class
13341 = could_it_be_a_POSIX_class(pRExC_state);
13342 if (! is_posix_class) {
13346 /* regclass() can only return RESTART_UTF8 if multi-char
13347 folds are allowed. */
13348 if (!regclass(pRExC_state, flagp,depth+1,
13349 is_posix_class, /* parse the whole char
13350 class only if not a
13352 FALSE, /* don't allow multi-char folds */
13353 TRUE, /* silence non-portable warnings. */
13357 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13360 /* function call leaves parse pointing to the ']', except
13361 * if we faked it */
13362 if (is_posix_class) {
13366 SvREFCNT_dec(current); /* In case it returned something */
13371 if (depth--) break;
13373 if (RExC_parse < RExC_end
13374 && *RExC_parse == ')')
13376 node = reganode(pRExC_state, ANYOF, 0);
13377 RExC_size += ANYOF_SKIP;
13378 nextchar(pRExC_state);
13379 Set_Node_Length(node,
13380 RExC_parse - oregcomp_parse + 1); /* MJD */
13389 FAIL("Syntax error in (?[...])");
13392 /* Pass 2 only after this. Everything in this construct is a
13393 * metacharacter. Operands begin with either a '\' (for an escape
13394 * sequence), or a '[' for a bracketed character class. Any other
13395 * character should be an operator, or parenthesis for grouping. Both
13396 * types of operands are handled by calling regclass() to parse them. It
13397 * is called with a parameter to indicate to return the computed inversion
13398 * list. The parsing here is implemented via a stack. Each entry on the
13399 * stack is a single character representing one of the operators, or the
13400 * '('; or else a pointer to an operand inversion list. */
13402 #define IS_OPERAND(a) (! SvIOK(a))
13404 /* The stack starts empty. It is a syntax error if the first thing parsed
13405 * is a binary operator; everything else is pushed on the stack. When an
13406 * operand is parsed, the top of the stack is examined. If it is a binary
13407 * operator, the item before it should be an operand, and both are replaced
13408 * by the result of doing that operation on the new operand and the one on
13409 * the stack. Thus a sequence of binary operands is reduced to a single
13410 * one before the next one is parsed.
13412 * A unary operator may immediately follow a binary in the input, for
13415 * When an operand is parsed and the top of the stack is a unary operator,
13416 * the operation is performed, and then the stack is rechecked to see if
13417 * this new operand is part of a binary operation; if so, it is handled as
13420 * A '(' is simply pushed on the stack; it is valid only if the stack is
13421 * empty, or the top element of the stack is an operator or another '('
13422 * (for which the parenthesized expression will become an operand). By the
13423 * time the corresponding ')' is parsed everything in between should have
13424 * been parsed and evaluated to a single operand (or else is a syntax
13425 * error), and is handled as a regular operand */
13427 sv_2mortal((SV *)(stack = newAV()));
13429 while (RExC_parse < RExC_end) {
13430 I32 top_index = av_tindex(stack);
13432 SV* current = NULL;
13434 /* Skip white space */
13435 RExC_parse = regpatws(pRExC_state, RExC_parse,
13436 TRUE /* means recognize comments */ );
13437 if (RExC_parse >= RExC_end) {
13438 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
13440 if ((curchar = UCHARAT(RExC_parse)) == ']') {
13447 if (av_tindex(stack) >= 0 /* This makes sure that we can
13448 safely subtract 1 from
13449 RExC_parse in the next clause.
13450 If we have something on the
13451 stack, we have parsed something
13453 && UCHARAT(RExC_parse - 1) == '('
13454 && RExC_parse < RExC_end)
13456 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
13457 * This happens when we have some thing like
13459 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
13461 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
13463 * Here we would be handling the interpolated
13464 * '$thai_or_lao'. We handle this by a recursive call to
13465 * ourselves which returns the inversion list the
13466 * interpolated expression evaluates to. We use the flags
13467 * from the interpolated pattern. */
13468 U32 save_flags = RExC_flags;
13469 const char * const save_parse = ++RExC_parse;
13471 parse_lparen_question_flags(pRExC_state);
13473 if (RExC_parse == save_parse /* Makes sure there was at
13474 least one flag (or this
13475 embedding wasn't compiled)
13477 || RExC_parse >= RExC_end - 4
13478 || UCHARAT(RExC_parse) != ':'
13479 || UCHARAT(++RExC_parse) != '('
13480 || UCHARAT(++RExC_parse) != '?'
13481 || UCHARAT(++RExC_parse) != '[')
13484 /* In combination with the above, this moves the
13485 * pointer to the point just after the first erroneous
13486 * character (or if there are no flags, to where they
13487 * should have been) */
13488 if (RExC_parse >= RExC_end - 4) {
13489 RExC_parse = RExC_end;
13491 else if (RExC_parse != save_parse) {
13492 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13494 vFAIL("Expecting '(?flags:(?[...'");
13497 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
13498 depth+1, oregcomp_parse);
13500 /* Here, 'current' contains the embedded expression's
13501 * inversion list, and RExC_parse points to the trailing
13502 * ']'; the next character should be the ')' which will be
13503 * paired with the '(' that has been put on the stack, so
13504 * the whole embedded expression reduces to '(operand)' */
13507 RExC_flags = save_flags;
13508 goto handle_operand;
13513 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13514 vFAIL("Unexpected character");
13517 /* regclass() can only return RESTART_UTF8 if multi-char
13518 folds are allowed. */
13519 if (!regclass(pRExC_state, flagp,depth+1,
13520 TRUE, /* means parse just the next thing */
13521 FALSE, /* don't allow multi-char folds */
13522 FALSE, /* don't silence non-portable warnings. */
13526 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13528 /* regclass() will return with parsing just the \ sequence,
13529 * leaving the parse pointer at the next thing to parse */
13531 goto handle_operand;
13533 case '[': /* Is a bracketed character class */
13535 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
13537 if (! is_posix_class) {
13541 /* regclass() can only return RESTART_UTF8 if multi-char
13542 folds are allowed. */
13543 if(!regclass(pRExC_state, flagp,depth+1,
13544 is_posix_class, /* parse the whole char class
13545 only if not a posix class */
13546 FALSE, /* don't allow multi-char folds */
13547 FALSE, /* don't silence non-portable warnings. */
13551 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
13553 /* function call leaves parse pointing to the ']', except if we
13555 if (is_posix_class) {
13559 goto handle_operand;
13568 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
13569 || ! IS_OPERAND(*top_ptr))
13572 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
13574 av_push(stack, newSVuv(curchar));
13578 av_push(stack, newSVuv(curchar));
13582 if (top_index >= 0) {
13583 top_ptr = av_fetch(stack, top_index, FALSE);
13585 if (IS_OPERAND(*top_ptr)) {
13587 vFAIL("Unexpected '(' with no preceding operator");
13590 av_push(stack, newSVuv(curchar));
13597 || ! (current = av_pop(stack))
13598 || ! IS_OPERAND(current)
13599 || ! (lparen = av_pop(stack))
13600 || IS_OPERAND(lparen)
13601 || SvUV(lparen) != '(')
13603 SvREFCNT_dec(current);
13605 vFAIL("Unexpected ')'");
13608 SvREFCNT_dec_NN(lparen);
13615 /* Here, we have an operand to process, in 'current' */
13617 if (top_index < 0) { /* Just push if stack is empty */
13618 av_push(stack, current);
13621 SV* top = av_pop(stack);
13623 char current_operator;
13625 if (IS_OPERAND(top)) {
13626 SvREFCNT_dec_NN(top);
13627 SvREFCNT_dec_NN(current);
13628 vFAIL("Operand with no preceding operator");
13630 current_operator = (char) SvUV(top);
13631 switch (current_operator) {
13632 case '(': /* Push the '(' back on followed by the new
13634 av_push(stack, top);
13635 av_push(stack, current);
13636 SvREFCNT_inc(top); /* Counters the '_dec' done
13637 just after the 'break', so
13638 it doesn't get wrongly freed
13643 _invlist_invert(current);
13645 /* Unlike binary operators, the top of the stack,
13646 * now that this unary one has been popped off, may
13647 * legally be an operator, and we now have operand
13650 SvREFCNT_dec_NN(top);
13651 goto handle_operand;
13654 prev = av_pop(stack);
13655 _invlist_intersection(prev,
13658 av_push(stack, current);
13663 prev = av_pop(stack);
13664 _invlist_union(prev, current, ¤t);
13665 av_push(stack, current);
13669 prev = av_pop(stack);;
13670 _invlist_subtract(prev, current, ¤t);
13671 av_push(stack, current);
13674 case '^': /* The union minus the intersection */
13680 prev = av_pop(stack);
13681 _invlist_union(prev, current, &u);
13682 _invlist_intersection(prev, current, &i);
13683 /* _invlist_subtract will overwrite current
13684 without freeing what it already contains */
13686 _invlist_subtract(u, i, ¤t);
13687 av_push(stack, current);
13688 SvREFCNT_dec_NN(i);
13689 SvREFCNT_dec_NN(u);
13690 SvREFCNT_dec_NN(element);
13695 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13697 SvREFCNT_dec_NN(top);
13698 SvREFCNT_dec(prev);
13702 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13705 if (av_tindex(stack) < 0 /* Was empty */
13706 || ((final = av_pop(stack)) == NULL)
13707 || ! IS_OPERAND(final)
13708 || av_tindex(stack) >= 0) /* More left on stack */
13710 vFAIL("Incomplete expression within '(?[ ])'");
13713 /* Here, 'final' is the resultant inversion list from evaluating the
13714 * expression. Return it if so requested */
13715 if (return_invlist) {
13716 *return_invlist = final;
13720 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13721 * expecting a string of ranges and individual code points */
13722 invlist_iterinit(final);
13723 result_string = newSVpvs("");
13724 while (invlist_iternext(final, &start, &end)) {
13725 if (start == end) {
13726 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13729 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13734 save_parse = RExC_parse;
13735 RExC_parse = SvPV(result_string, len);
13736 save_end = RExC_end;
13737 RExC_end = RExC_parse + len;
13739 /* We turn off folding around the call, as the class we have constructed
13740 * already has all folding taken into consideration, and we don't want
13741 * regclass() to add to that */
13742 RExC_flags &= ~RXf_PMf_FOLD;
13743 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13745 node = regclass(pRExC_state, flagp,depth+1,
13746 FALSE, /* means parse the whole char class */
13747 FALSE, /* don't allow multi-char folds */
13748 TRUE, /* silence non-portable warnings. The above may very
13749 well have generated non-portable code points, but
13750 they're valid on this machine */
13751 FALSE, /* similarly, no need for strict */
13755 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13758 RExC_flags |= RXf_PMf_FOLD;
13760 RExC_parse = save_parse + 1;
13761 RExC_end = save_end;
13762 SvREFCNT_dec_NN(final);
13763 SvREFCNT_dec_NN(result_string);
13765 nextchar(pRExC_state);
13766 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13772 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
13774 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
13775 * innocent-looking character class, like /[ks]/i won't have to go out to
13776 * disk to find the possible matches.
13778 * This should be called only for a Latin1-range code points, cp, which is
13779 * known to be involved in a simple fold with other code points above
13780 * Latin1. It would give false results if /aa has been specified.
13781 * Multi-char folds are outside the scope of this, and must be handled
13784 * XXX It would be better to generate these via regen, in case a new
13785 * version of the Unicode standard adds new mappings, though that is not
13786 * really likely, and may be caught by the default: case of the switch
13789 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
13791 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
13797 add_cp_to_invlist(*invlist, KELVIN_SIGN);
13801 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
13804 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
13805 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
13807 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13808 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13809 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
13811 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13812 *invlist = add_cp_to_invlist(*invlist,
13813 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13815 case LATIN_SMALL_LETTER_SHARP_S:
13816 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
13819 /* Use deprecated warning to increase the chances of this being
13822 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
13829 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
13831 /* This adds the string scalar <multi_string> to the array
13832 * <multi_char_matches>. <multi_string> is known to have exactly
13833 * <cp_count> code points in it. This is used when constructing a
13834 * bracketed character class and we find something that needs to match more
13835 * than a single character.
13837 * <multi_char_matches> is actually an array of arrays. Each top-level
13838 * element is an array that contains all the strings known so far that are
13839 * the same length. And that length (in number of code points) is the same
13840 * as the index of the top-level array. Hence, the [2] element is an
13841 * array, each element thereof is a string containing TWO code points;
13842 * while element [3] is for strings of THREE characters, and so on. Since
13843 * this is for multi-char strings there can never be a [0] nor [1] element.
13845 * When we rewrite the character class below, we will do so such that the
13846 * longest strings are written first, so that it prefers the longest
13847 * matching strings first. This is done even if it turns out that any
13848 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
13849 * Christiansen has agreed that this is ok. This makes the test for the
13850 * ligature 'ffi' come before the test for 'ff', for example */
13853 AV** this_array_ptr;
13855 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
13857 if (! multi_char_matches) {
13858 multi_char_matches = newAV();
13861 if (av_exists(multi_char_matches, cp_count)) {
13862 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
13863 this_array = *this_array_ptr;
13866 this_array = newAV();
13867 av_store(multi_char_matches, cp_count,
13870 av_push(this_array, multi_string);
13872 return multi_char_matches;
13875 /* The names of properties whose definitions are not known at compile time are
13876 * stored in this SV, after a constant heading. So if the length has been
13877 * changed since initialization, then there is a run-time definition. */
13878 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13879 (SvCUR(listsv) != initial_listsv_len)
13882 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13883 const bool stop_at_1, /* Just parse the next thing, don't
13884 look for a full character class */
13885 bool allow_multi_folds,
13886 const bool silence_non_portable, /* Don't output warnings
13890 SV** ret_invlist /* Return an inversion list, not a node */
13893 /* parse a bracketed class specification. Most of these will produce an
13894 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13895 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13896 * under /i with multi-character folds: it will be rewritten following the
13897 * paradigm of this example, where the <multi-fold>s are characters which
13898 * fold to multiple character sequences:
13899 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13900 * gets effectively rewritten as:
13901 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13902 * reg() gets called (recursively) on the rewritten version, and this
13903 * function will return what it constructs. (Actually the <multi-fold>s
13904 * aren't physically removed from the [abcdefghi], it's just that they are
13905 * ignored in the recursion by means of a flag:
13906 * <RExC_in_multi_char_class>.)
13908 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
13909 * characters, with the corresponding bit set if that character is in the
13910 * list. For characters above this, a range list or swash is used. There
13911 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
13912 * determinable at compile time
13914 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13915 * to be restarted. This can only happen if ret_invlist is non-NULL.
13918 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13920 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13923 IV namedclass = OOB_NAMEDCLASS;
13924 char *rangebegin = NULL;
13925 bool need_class = 0;
13927 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13928 than just initialized. */
13929 SV* properties = NULL; /* Code points that match \p{} \P{} */
13930 SV* posixes = NULL; /* Code points that match classes like [:word:],
13931 extended beyond the Latin1 range. These have to
13932 be kept separate from other code points for much
13933 of this function because their handling is
13934 different under /i, and for most classes under
13936 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13937 separate for a while from the non-complemented
13938 versions because of complications with /d
13940 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
13941 treated more simply than the general case,
13942 leading to less compilation and execution
13944 UV element_count = 0; /* Number of distinct elements in the class.
13945 Optimizations may be possible if this is tiny */
13946 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13947 character; used under /i */
13949 char * stop_ptr = RExC_end; /* where to stop parsing */
13950 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13953 /* Unicode properties are stored in a swash; this holds the current one
13954 * being parsed. If this swash is the only above-latin1 component of the
13955 * character class, an optimization is to pass it directly on to the
13956 * execution engine. Otherwise, it is set to NULL to indicate that there
13957 * are other things in the class that have to be dealt with at execution
13959 SV* swash = NULL; /* Code points that match \p{} \P{} */
13961 /* Set if a component of this character class is user-defined; just passed
13962 * on to the engine */
13963 bool has_user_defined_property = FALSE;
13965 /* inversion list of code points this node matches only when the target
13966 * string is in UTF-8. (Because is under /d) */
13967 SV* depends_list = NULL;
13969 /* Inversion list of code points this node matches regardless of things
13970 * like locale, folding, utf8ness of the target string */
13971 SV* cp_list = NULL;
13973 /* Like cp_list, but code points on this list need to be checked for things
13974 * that fold to/from them under /i */
13975 SV* cp_foldable_list = NULL;
13977 /* Like cp_list, but code points on this list are valid only when the
13978 * runtime locale is UTF-8 */
13979 SV* only_utf8_locale_list = NULL;
13981 /* In a range, if one of the endpoints is non-character-set portable,
13982 * meaning that it hard-codes a code point that may mean a different
13983 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
13984 * mnemonic '\t' which each mean the same character no matter which
13985 * character set the platform is on. */
13986 unsigned int non_portable_endpoint = 0;
13988 /* Is the range unicode? which means on a platform that isn't 1-1 native
13989 * to Unicode (i.e. non-ASCII), each code point in it should be considered
13990 * to be a Unicode value. */
13991 bool unicode_range = FALSE;
13992 bool invert = FALSE; /* Is this class to be complemented */
13994 bool warn_super = ALWAYS_WARN_SUPER;
13996 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13997 case we need to change the emitted regop to an EXACT. */
13998 const char * orig_parse = RExC_parse;
13999 const SSize_t orig_size = RExC_size;
14000 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
14001 GET_RE_DEBUG_FLAGS_DECL;
14003 PERL_ARGS_ASSERT_REGCLASS;
14005 PERL_UNUSED_ARG(depth);
14008 DEBUG_PARSE("clas");
14010 /* Assume we are going to generate an ANYOF node. */
14011 ret = reganode(pRExC_state,
14018 RExC_size += ANYOF_SKIP;
14019 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
14022 ANYOF_FLAGS(ret) = 0;
14024 RExC_emit += ANYOF_SKIP;
14025 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
14026 initial_listsv_len = SvCUR(listsv);
14027 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
14031 RExC_parse = regpatws(pRExC_state, RExC_parse,
14032 FALSE /* means don't recognize comments */ );
14035 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
14038 allow_multi_folds = FALSE;
14041 RExC_parse = regpatws(pRExC_state, RExC_parse,
14042 FALSE /* means don't recognize comments */ );
14046 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
14047 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
14048 const char *s = RExC_parse;
14049 const char c = *s++;
14054 while (isWORDCHAR(*s))
14056 if (*s && c == *s && s[1] == ']') {
14057 SAVEFREESV(RExC_rx_sv);
14059 "POSIX syntax [%c %c] belongs inside character classes",
14061 (void)ReREFCNT_inc(RExC_rx_sv);
14065 /* If the caller wants us to just parse a single element, accomplish this
14066 * by faking the loop ending condition */
14067 if (stop_at_1 && RExC_end > RExC_parse) {
14068 stop_ptr = RExC_parse + 1;
14071 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
14072 if (UCHARAT(RExC_parse) == ']')
14073 goto charclassloop;
14076 if (RExC_parse >= stop_ptr) {
14081 RExC_parse = regpatws(pRExC_state, RExC_parse,
14082 FALSE /* means don't recognize comments */ );
14085 if (UCHARAT(RExC_parse) == ']') {
14091 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
14092 save_value = value;
14093 save_prevvalue = prevvalue;
14096 rangebegin = RExC_parse;
14098 non_portable_endpoint = 0;
14101 value = utf8n_to_uvchr((U8*)RExC_parse,
14102 RExC_end - RExC_parse,
14103 &numlen, UTF8_ALLOW_DEFAULT);
14104 RExC_parse += numlen;
14107 value = UCHARAT(RExC_parse++);
14110 && RExC_parse < RExC_end
14111 && POSIXCC(UCHARAT(RExC_parse)))
14113 namedclass = regpposixcc(pRExC_state, value, strict);
14115 else if (value == '\\') {
14116 /* Is a backslash; get the code point of the char after it */
14117 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
14118 value = utf8n_to_uvchr((U8*)RExC_parse,
14119 RExC_end - RExC_parse,
14120 &numlen, UTF8_ALLOW_DEFAULT);
14121 RExC_parse += numlen;
14124 value = UCHARAT(RExC_parse++);
14126 /* Some compilers cannot handle switching on 64-bit integer
14127 * values, therefore value cannot be an UV. Yes, this will
14128 * be a problem later if we want switch on Unicode.
14129 * A similar issue a little bit later when switching on
14130 * namedclass. --jhi */
14132 /* If the \ is escaping white space when white space is being
14133 * skipped, it means that that white space is wanted literally, and
14134 * is already in 'value'. Otherwise, need to translate the escape
14135 * into what it signifies. */
14136 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
14138 case 'w': namedclass = ANYOF_WORDCHAR; break;
14139 case 'W': namedclass = ANYOF_NWORDCHAR; break;
14140 case 's': namedclass = ANYOF_SPACE; break;
14141 case 'S': namedclass = ANYOF_NSPACE; break;
14142 case 'd': namedclass = ANYOF_DIGIT; break;
14143 case 'D': namedclass = ANYOF_NDIGIT; break;
14144 case 'v': namedclass = ANYOF_VERTWS; break;
14145 case 'V': namedclass = ANYOF_NVERTWS; break;
14146 case 'h': namedclass = ANYOF_HORIZWS; break;
14147 case 'H': namedclass = ANYOF_NHORIZWS; break;
14148 case 'N': /* Handle \N{NAME} in class */
14150 const char * const backslash_N_beg = RExC_parse - 2;
14153 if (! grok_bslash_N(pRExC_state,
14154 NULL, /* No regnode */
14155 &value, /* Yes single value */
14156 &cp_count, /* Multiple code pt count */
14161 if (*flagp & RESTART_UTF8)
14162 FAIL("panic: grok_bslash_N set RESTART_UTF8");
14164 if (cp_count < 0) {
14165 vFAIL("\\N in a character class must be a named character: \\N{...}");
14167 else if (cp_count == 0) {
14169 RExC_parse++; /* Position after the "}" */
14170 vFAIL("Zero length \\N{}");
14173 ckWARNreg(RExC_parse,
14174 "Ignoring zero length \\N{} in character class");
14177 else { /* cp_count > 1 */
14178 if (! RExC_in_multi_char_class) {
14179 if (invert || range || *RExC_parse == '-') {
14182 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
14185 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
14187 break; /* <value> contains the first code
14188 point. Drop out of the switch to
14192 SV * multi_char_N = newSVpvn(backslash_N_beg,
14193 RExC_parse - backslash_N_beg);
14195 = add_multi_match(multi_char_matches,
14200 } /* End of cp_count != 1 */
14202 /* This element should not be processed further in this
14205 value = save_value;
14206 prevvalue = save_prevvalue;
14207 continue; /* Back to top of loop to get next char */
14210 /* Here, is a single code point, and <value> contains it */
14211 unicode_range = TRUE; /* \N{} are Unicode */
14219 /* We will handle any undefined properties ourselves */
14220 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
14221 /* And we actually would prefer to get
14222 * the straight inversion list of the
14223 * swash, since we will be accessing it
14224 * anyway, to save a little time */
14225 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
14227 if (RExC_parse >= RExC_end)
14228 vFAIL2("Empty \\%c{}", (U8)value);
14229 if (*RExC_parse == '{') {
14230 const U8 c = (U8)value;
14231 e = strchr(RExC_parse++, '}');
14233 vFAIL2("Missing right brace on \\%c{}", c);
14234 while (isSPACE(*RExC_parse))
14236 if (e == RExC_parse)
14237 vFAIL2("Empty \\%c{}", c);
14238 n = e - RExC_parse;
14239 while (isSPACE(*(RExC_parse + n - 1)))
14250 if (UCHARAT(RExC_parse) == '^') {
14253 /* toggle. (The rhs xor gets the single bit that
14254 * differs between P and p; the other xor inverts just
14256 value ^= 'P' ^ 'p';
14258 while (isSPACE(*RExC_parse)) {
14263 /* Try to get the definition of the property into
14264 * <invlist>. If /i is in effect, the effective property
14265 * will have its name be <__NAME_i>. The design is
14266 * discussed in commit
14267 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
14268 name = savepv(Perl_form(aTHX_
14270 (FOLD) ? "__" : "",
14276 /* Look up the property name, and get its swash and
14277 * inversion list, if the property is found */
14279 SvREFCNT_dec_NN(swash);
14281 swash = _core_swash_init("utf8", name, &PL_sv_undef,
14284 NULL, /* No inversion list */
14287 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
14288 HV* curpkg = (IN_PERL_COMPILETIME)
14290 : CopSTASH(PL_curcop);
14292 SvREFCNT_dec_NN(swash);
14296 /* Here didn't find it. It could be a user-defined
14297 * property that will be available at run-time. If we
14298 * accept only compile-time properties, is an error;
14299 * otherwise add it to the list for run-time look up */
14301 RExC_parse = e + 1;
14303 "Property '%"UTF8f"' is unknown",
14304 UTF8fARG(UTF, n, name));
14307 /* If the property name doesn't already have a package
14308 * name, add the current one to it so that it can be
14309 * referred to outside it. [perl #121777] */
14310 if (curpkg && ! instr(name, "::")) {
14311 char* pkgname = HvNAME(curpkg);
14312 if (strNE(pkgname, "main")) {
14313 char* full_name = Perl_form(aTHX_
14317 n = strlen(full_name);
14319 name = savepvn(full_name, n);
14322 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
14323 (value == 'p' ? '+' : '!'),
14324 UTF8fARG(UTF, n, name));
14325 has_user_defined_property = TRUE;
14327 /* We don't know yet, so have to assume that the
14328 * property could match something in the Latin1 range,
14329 * hence something that isn't utf8. Note that this
14330 * would cause things in <depends_list> to match
14331 * inappropriately, except that any \p{}, including
14332 * this one forces Unicode semantics, which means there
14333 * is no <depends_list> */
14335 |= ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES;
14339 /* Here, did get the swash and its inversion list. If
14340 * the swash is from a user-defined property, then this
14341 * whole character class should be regarded as such */
14342 if (swash_init_flags
14343 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
14345 has_user_defined_property = TRUE;
14348 /* We warn on matching an above-Unicode code point
14349 * if the match would return true, except don't
14350 * warn for \p{All}, which has exactly one element
14352 (_invlist_contains_cp(invlist, 0x110000)
14353 && (! (_invlist_len(invlist) == 1
14354 && *invlist_array(invlist) == 0)))
14360 /* Invert if asking for the complement */
14361 if (value == 'P') {
14362 _invlist_union_complement_2nd(properties,
14366 /* The swash can't be used as-is, because we've
14367 * inverted things; delay removing it to here after
14368 * have copied its invlist above */
14369 SvREFCNT_dec_NN(swash);
14373 _invlist_union(properties, invlist, &properties);
14378 RExC_parse = e + 1;
14379 namedclass = ANYOF_UNIPROP; /* no official name, but it's
14382 /* \p means they want Unicode semantics */
14383 RExC_uni_semantics = 1;
14386 case 'n': value = '\n'; break;
14387 case 'r': value = '\r'; break;
14388 case 't': value = '\t'; break;
14389 case 'f': value = '\f'; break;
14390 case 'b': value = '\b'; break;
14391 case 'e': value = ESC_NATIVE; break;
14392 case 'a': value = '\a'; break;
14394 RExC_parse--; /* function expects to be pointed at the 'o' */
14396 const char* error_msg;
14397 bool valid = grok_bslash_o(&RExC_parse,
14400 PASS2, /* warnings only in
14403 silence_non_portable,
14409 non_portable_endpoint++;
14410 if (IN_ENCODING && value < 0x100) {
14411 goto recode_encoding;
14415 RExC_parse--; /* function expects to be pointed at the 'x' */
14417 const char* error_msg;
14418 bool valid = grok_bslash_x(&RExC_parse,
14421 PASS2, /* Output warnings */
14423 silence_non_portable,
14429 non_portable_endpoint++;
14430 if (IN_ENCODING && value < 0x100)
14431 goto recode_encoding;
14434 value = grok_bslash_c(*RExC_parse++, PASS2);
14435 non_portable_endpoint++;
14437 case '0': case '1': case '2': case '3': case '4':
14438 case '5': case '6': case '7':
14440 /* Take 1-3 octal digits */
14441 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14442 numlen = (strict) ? 4 : 3;
14443 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
14444 RExC_parse += numlen;
14447 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
14448 vFAIL("Need exactly 3 octal digits");
14450 else if (! SIZE_ONLY /* like \08, \178 */
14452 && RExC_parse < RExC_end
14453 && isDIGIT(*RExC_parse)
14454 && ckWARN(WARN_REGEXP))
14456 SAVEFREESV(RExC_rx_sv);
14457 reg_warn_non_literal_string(
14459 form_short_octal_warning(RExC_parse, numlen));
14460 (void)ReREFCNT_inc(RExC_rx_sv);
14463 non_portable_endpoint++;
14464 if (IN_ENCODING && value < 0x100)
14465 goto recode_encoding;
14469 if (! RExC_override_recoding) {
14470 SV* enc = _get_encoding();
14471 value = reg_recode((const char)(U8)value, &enc);
14474 vFAIL("Invalid escape in the specified encoding");
14477 ckWARNreg(RExC_parse,
14478 "Invalid escape in the specified encoding");
14484 /* Allow \_ to not give an error */
14485 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
14487 vFAIL2("Unrecognized escape \\%c in character class",
14491 SAVEFREESV(RExC_rx_sv);
14492 ckWARN2reg(RExC_parse,
14493 "Unrecognized escape \\%c in character class passed through",
14495 (void)ReREFCNT_inc(RExC_rx_sv);
14499 } /* End of switch on char following backslash */
14500 } /* end of handling backslash escape sequences */
14502 /* Here, we have the current token in 'value' */
14504 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
14507 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
14508 * literal, as is the character that began the false range, i.e.
14509 * the 'a' in the examples */
14512 const int w = (RExC_parse >= rangebegin)
14513 ? RExC_parse - rangebegin
14517 "False [] range \"%"UTF8f"\"",
14518 UTF8fARG(UTF, w, rangebegin));
14521 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
14522 ckWARN2reg(RExC_parse,
14523 "False [] range \"%"UTF8f"\"",
14524 UTF8fARG(UTF, w, rangebegin));
14525 (void)ReREFCNT_inc(RExC_rx_sv);
14526 cp_list = add_cp_to_invlist(cp_list, '-');
14527 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
14532 range = 0; /* this was not a true range */
14533 element_count += 2; /* So counts for three values */
14536 classnum = namedclass_to_classnum(namedclass);
14538 if (LOC && namedclass < ANYOF_POSIXL_MAX
14539 #ifndef HAS_ISASCII
14540 && classnum != _CC_ASCII
14543 /* What the Posix classes (like \w, [:space:]) match in locale
14544 * isn't knowable under locale until actual match time. Room
14545 * must be reserved (one time per outer bracketed class) to
14546 * store such classes. The space will contain a bit for each
14547 * named class that is to be matched against. This isn't
14548 * needed for \p{} and pseudo-classes, as they are not affected
14549 * by locale, and hence are dealt with separately */
14550 if (! need_class) {
14553 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14556 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14558 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
14559 ANYOF_POSIXL_ZERO(ret);
14562 /* Coverity thinks it is possible for this to be negative; both
14563 * jhi and khw think it's not, but be safer */
14564 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14565 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
14567 /* See if it already matches the complement of this POSIX
14569 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14570 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
14574 posixl_matches_all = TRUE;
14575 break; /* No need to continue. Since it matches both
14576 e.g., \w and \W, it matches everything, and the
14577 bracketed class can be optimized into qr/./s */
14580 /* Add this class to those that should be checked at runtime */
14581 ANYOF_POSIXL_SET(ret, namedclass);
14583 /* The above-Latin1 characters are not subject to locale rules.
14584 * Just add them, in the second pass, to the
14585 * unconditionally-matched list */
14587 SV* scratch_list = NULL;
14589 /* Get the list of the above-Latin1 code points this
14591 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
14592 PL_XPosix_ptrs[classnum],
14594 /* Odd numbers are complements, like
14595 * NDIGIT, NASCII, ... */
14596 namedclass % 2 != 0,
14598 /* Checking if 'cp_list' is NULL first saves an extra
14599 * clone. Its reference count will be decremented at the
14600 * next union, etc, or if this is the only instance, at the
14601 * end of the routine */
14603 cp_list = scratch_list;
14606 _invlist_union(cp_list, scratch_list, &cp_list);
14607 SvREFCNT_dec_NN(scratch_list);
14609 continue; /* Go get next character */
14612 else if (! SIZE_ONLY) {
14614 /* Here, not in pass1 (in that pass we skip calculating the
14615 * contents of this class), and is /l, or is a POSIX class for
14616 * which /l doesn't matter (or is a Unicode property, which is
14617 * skipped here). */
14618 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
14619 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
14621 /* Here, should be \h, \H, \v, or \V. None of /d, /i
14622 * nor /l make a difference in what these match,
14623 * therefore we just add what they match to cp_list. */
14624 if (classnum != _CC_VERTSPACE) {
14625 assert( namedclass == ANYOF_HORIZWS
14626 || namedclass == ANYOF_NHORIZWS);
14628 /* It turns out that \h is just a synonym for
14630 classnum = _CC_BLANK;
14633 _invlist_union_maybe_complement_2nd(
14635 PL_XPosix_ptrs[classnum],
14636 namedclass % 2 != 0, /* Complement if odd
14637 (NHORIZWS, NVERTWS)
14642 else if (UNI_SEMANTICS
14643 || classnum == _CC_ASCII
14644 || (DEPENDS_SEMANTICS && (classnum == _CC_DIGIT
14645 || classnum == _CC_XDIGIT)))
14647 /* We usually have to worry about /d and /a affecting what
14648 * POSIX classes match, with special code needed for /d
14649 * because we won't know until runtime what all matches.
14650 * But there is no extra work needed under /u, and
14651 * [:ascii:] is unaffected by /a and /d; and :digit: and
14652 * :xdigit: don't have runtime differences under /d. So we
14653 * can special case these, and avoid some extra work below,
14654 * and at runtime. */
14655 _invlist_union_maybe_complement_2nd(
14657 PL_XPosix_ptrs[classnum],
14658 namedclass % 2 != 0,
14661 else { /* Garden variety class. If is NUPPER, NALPHA, ...
14662 complement and use nposixes */
14663 SV** posixes_ptr = namedclass % 2 == 0
14666 _invlist_union_maybe_complement_2nd(
14668 PL_XPosix_ptrs[classnum],
14669 namedclass % 2 != 0,
14673 } /* end of namedclass \blah */
14676 RExC_parse = regpatws(pRExC_state, RExC_parse,
14677 FALSE /* means don't recognize comments */ );
14680 /* If 'range' is set, 'value' is the ending of a range--check its
14681 * validity. (If value isn't a single code point in the case of a
14682 * range, we should have figured that out above in the code that
14683 * catches false ranges). Later, we will handle each individual code
14684 * point in the range. If 'range' isn't set, this could be the
14685 * beginning of a range, so check for that by looking ahead to see if
14686 * the next real character to be processed is the range indicator--the
14691 /* For unicode ranges, we have to test that the Unicode as opposed
14692 * to the native values are not decreasing. (Above 255, there is
14693 * no difference between native and Unicode) */
14694 if (unicode_range && prevvalue < 255 && value < 255) {
14695 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
14696 goto backwards_range;
14701 if (prevvalue > value) /* b-a */ {
14706 w = RExC_parse - rangebegin;
14708 "Invalid [] range \"%"UTF8f"\"",
14709 UTF8fARG(UTF, w, rangebegin));
14710 NOT_REACHED; /* NOTREACHED */
14714 prevvalue = value; /* save the beginning of the potential range */
14715 if (! stop_at_1 /* Can't be a range if parsing just one thing */
14716 && *RExC_parse == '-')
14718 char* next_char_ptr = RExC_parse + 1;
14719 if (skip_white) { /* Get the next real char after the '-' */
14720 next_char_ptr = regpatws(pRExC_state,
14722 FALSE); /* means don't recognize
14726 /* If the '-' is at the end of the class (just before the ']',
14727 * it is a literal minus; otherwise it is a range */
14728 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
14729 RExC_parse = next_char_ptr;
14731 /* a bad range like \w-, [:word:]- ? */
14732 if (namedclass > OOB_NAMEDCLASS) {
14733 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
14734 const int w = RExC_parse >= rangebegin
14735 ? RExC_parse - rangebegin
14738 vFAIL4("False [] range \"%*.*s\"",
14743 "False [] range \"%*.*s\"",
14748 cp_list = add_cp_to_invlist(cp_list, '-');
14752 range = 1; /* yeah, it's a range! */
14753 continue; /* but do it the next time */
14758 if (namedclass > OOB_NAMEDCLASS) {
14762 /* Here, we have a single value this time through the loop, and
14763 * <prevvalue> is the beginning of the range, if any; or <value> if
14766 /* non-Latin1 code point implies unicode semantics. Must be set in
14767 * pass1 so is there for the whole of pass 2 */
14769 RExC_uni_semantics = 1;
14772 /* Ready to process either the single value, or the completed range.
14773 * For single-valued non-inverted ranges, we consider the possibility
14774 * of multi-char folds. (We made a conscious decision to not do this
14775 * for the other cases because it can often lead to non-intuitive
14776 * results. For example, you have the peculiar case that:
14777 * "s s" =~ /^[^\xDF]+$/i => Y
14778 * "ss" =~ /^[^\xDF]+$/i => N
14780 * See [perl #89750] */
14781 if (FOLD && allow_multi_folds && value == prevvalue) {
14782 if (value == LATIN_SMALL_LETTER_SHARP_S
14783 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
14786 /* Here <value> is indeed a multi-char fold. Get what it is */
14788 U8 foldbuf[UTF8_MAXBYTES_CASE];
14791 UV folded = _to_uni_fold_flags(
14795 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
14796 ? FOLD_FLAGS_NOMIX_ASCII
14800 /* Here, <folded> should be the first character of the
14801 * multi-char fold of <value>, with <foldbuf> containing the
14802 * whole thing. But, if this fold is not allowed (because of
14803 * the flags), <fold> will be the same as <value>, and should
14804 * be processed like any other character, so skip the special
14806 if (folded != value) {
14808 /* Skip if we are recursed, currently parsing the class
14809 * again. Otherwise add this character to the list of
14810 * multi-char folds. */
14811 if (! RExC_in_multi_char_class) {
14812 STRLEN cp_count = utf8_length(foldbuf,
14813 foldbuf + foldlen);
14814 SV* multi_fold = sv_2mortal(newSVpvs(""));
14816 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
14819 = add_multi_match(multi_char_matches,
14825 /* This element should not be processed further in this
14828 value = save_value;
14829 prevvalue = save_prevvalue;
14835 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
14838 /* If the range starts above 255, everything is portable and
14839 * likely to be so for any forseeable character set, so don't
14841 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
14842 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
14844 else if (prevvalue != value) {
14846 /* Under strict, ranges that stop and/or end in an ASCII
14847 * printable should have each end point be a portable value
14848 * for it (preferably like 'A', but we don't warn if it is
14849 * a (portable) Unicode name or code point), and the range
14850 * must be be all digits or all letters of the same case.
14851 * Otherwise, the range is non-portable and unclear as to
14852 * what it contains */
14853 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
14854 && (non_portable_endpoint
14855 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
14856 || (isLOWER_A(prevvalue) && isLOWER_A(value))
14857 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
14859 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
14861 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
14863 /* But the nature of Unicode and languages mean we
14864 * can't do the same checks for above-ASCII ranges,
14865 * except in the case of digit ones. These should
14866 * contain only digits from the same group of 10. The
14867 * ASCII case is handled just above. 0x660 is the
14868 * first digit character beyond ASCII. Hence here, the
14869 * range could be a range of digits. Find out. */
14870 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
14872 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
14875 /* If the range start and final points are in the same
14876 * inversion list element, it means that either both
14877 * are not digits, or both are digits in a consecutive
14878 * sequence of digits. (So far, Unicode has kept all
14879 * such sequences as distinct groups of 10, but assert
14880 * to make sure). If the end points are not in the
14881 * same element, neither should be a digit. */
14882 if (index_start == index_final) {
14883 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
14884 || invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
14885 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
14888 else if ((index_start >= 0
14889 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
14890 || (index_final >= 0
14891 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
14893 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
14898 if ((! range || prevvalue == value) && non_portable_endpoint) {
14899 if (isPRINT_A(value)) {
14902 if (isBACKSLASHED_PUNCT(value)) {
14903 literal[d++] = '\\';
14905 literal[d++] = (char) value;
14906 literal[d++] = '\0';
14909 "\"%.*s\" is more clearly written simply as \"%s\"",
14910 (int) (RExC_parse - rangebegin),
14915 else if isMNEMONIC_CNTRL(value) {
14917 "\"%.*s\" is more clearly written simply as \"%s\"",
14918 (int) (RExC_parse - rangebegin),
14920 cntrl_to_mnemonic((char) value)
14926 /* Deal with this element of the class */
14930 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14933 /* On non-ASCII platforms, for ranges that span all of 0..255, and
14934 * ones that don't require special handling, we can just add the
14935 * range like we do for ASCII platforms */
14936 if ((UNLIKELY(prevvalue == 0) && value >= 255)
14937 || ! (prevvalue < 256
14939 || (! non_portable_endpoint
14940 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
14941 || (isUPPER_A(prevvalue)
14942 && isUPPER_A(value)))))))
14944 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14948 /* Here, requires special handling. This can be because it is
14949 * a range whose code points are considered to be Unicode, and
14950 * so must be individually translated into native, or because
14951 * its a subrange of 'A-Z' or 'a-z' which each aren't
14952 * contiguous in EBCDIC, but we have defined them to include
14953 * only the "expected" upper or lower case ASCII alphabetics.
14954 * Subranges above 255 are the same in native and Unicode, so
14955 * can be added as a range */
14956 U8 start = NATIVE_TO_LATIN1(prevvalue);
14958 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
14959 for (j = start; j <= end; j++) {
14960 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
14963 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14970 range = 0; /* this range (if it was one) is done now */
14971 } /* End of loop through all the text within the brackets */
14973 /* If anything in the class expands to more than one character, we have to
14974 * deal with them by building up a substitute parse string, and recursively
14975 * calling reg() on it, instead of proceeding */
14976 if (multi_char_matches) {
14977 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14980 char *save_end = RExC_end;
14981 char *save_parse = RExC_parse;
14982 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14987 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14988 because too confusing */
14990 sv_catpv(substitute_parse, "(?:");
14994 /* Look at the longest folds first */
14995 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14997 if (av_exists(multi_char_matches, cp_count)) {
14998 AV** this_array_ptr;
15001 this_array_ptr = (AV**) av_fetch(multi_char_matches,
15003 while ((this_sequence = av_pop(*this_array_ptr)) !=
15006 if (! first_time) {
15007 sv_catpv(substitute_parse, "|");
15009 first_time = FALSE;
15011 sv_catpv(substitute_parse, SvPVX(this_sequence));
15016 /* If the character class contains anything else besides these
15017 * multi-character folds, have to include it in recursive parsing */
15018 if (element_count) {
15019 sv_catpv(substitute_parse, "|[");
15020 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
15021 sv_catpv(substitute_parse, "]");
15024 sv_catpv(substitute_parse, ")");
15027 /* This is a way to get the parse to skip forward a whole named
15028 * sequence instead of matching the 2nd character when it fails the
15030 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
15034 RExC_parse = SvPV(substitute_parse, len);
15035 RExC_end = RExC_parse + len;
15036 RExC_in_multi_char_class = 1;
15037 RExC_override_recoding = 1;
15038 RExC_emit = (regnode *)orig_emit;
15040 ret = reg(pRExC_state, 1, ®_flags, depth+1);
15042 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
15044 RExC_parse = save_parse;
15045 RExC_end = save_end;
15046 RExC_in_multi_char_class = 0;
15047 RExC_override_recoding = 0;
15048 SvREFCNT_dec_NN(multi_char_matches);
15052 /* Here, we've gone through the entire class and dealt with multi-char
15053 * folds. We are now in a position that we can do some checks to see if we
15054 * can optimize this ANYOF node into a simpler one, even in Pass 1.
15055 * Currently we only do two checks:
15056 * 1) is in the unlikely event that the user has specified both, eg. \w and
15057 * \W under /l, then the class matches everything. (This optimization
15058 * is done only to make the optimizer code run later work.)
15059 * 2) if the character class contains only a single element (including a
15060 * single range), we see if there is an equivalent node for it.
15061 * Other checks are possible */
15062 if (! ret_invlist /* Can't optimize if returning the constructed
15064 && (UNLIKELY(posixl_matches_all) || element_count == 1))
15069 if (UNLIKELY(posixl_matches_all)) {
15072 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
15073 \w or [:digit:] or \p{foo}
15076 /* All named classes are mapped into POSIXish nodes, with its FLAG
15077 * argument giving which class it is */
15078 switch ((I32)namedclass) {
15079 case ANYOF_UNIPROP:
15082 /* These don't depend on the charset modifiers. They always
15083 * match under /u rules */
15084 case ANYOF_NHORIZWS:
15085 case ANYOF_HORIZWS:
15086 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
15089 case ANYOF_NVERTWS:
15094 /* The actual POSIXish node for all the rest depends on the
15095 * charset modifier. The ones in the first set depend only on
15096 * ASCII or, if available on this platform, also locale */
15100 op = (LOC) ? POSIXL : POSIXA;
15106 /* The following don't have any matches in the upper Latin1
15107 * range, hence /d is equivalent to /u for them. Making it /u
15108 * saves some branches at runtime */
15112 case ANYOF_NXDIGIT:
15113 if (! DEPENDS_SEMANTICS) {
15114 goto treat_as_default;
15120 /* The following change to CASED under /i */
15126 namedclass = ANYOF_CASED + (namedclass % 2);
15130 /* The rest have more possibilities depending on the charset.
15131 * We take advantage of the enum ordering of the charset
15132 * modifiers to get the exact node type, */
15135 op = POSIXD + get_regex_charset(RExC_flags);
15136 if (op > POSIXA) { /* /aa is same as /a */
15141 /* The odd numbered ones are the complements of the
15142 * next-lower even number one */
15143 if (namedclass % 2 == 1) {
15147 arg = namedclass_to_classnum(namedclass);
15151 else if (value == prevvalue) {
15153 /* Here, the class consists of just a single code point */
15156 if (! LOC && value == '\n') {
15157 op = REG_ANY; /* Optimize [^\n] */
15158 *flagp |= HASWIDTH|SIMPLE;
15162 else if (value < 256 || UTF) {
15164 /* Optimize a single value into an EXACTish node, but not if it
15165 * would require converting the pattern to UTF-8. */
15166 op = compute_EXACTish(pRExC_state);
15168 } /* Otherwise is a range */
15169 else if (! LOC) { /* locale could vary these */
15170 if (prevvalue == '0') {
15171 if (value == '9') {
15176 else if (AT_LEAST_ASCII_RESTRICTED || ! FOLD) {
15177 /* We can optimize A-Z or a-z, but not if they could match
15178 * something like the KELVIN SIGN under /i (/a means they
15180 if (prevvalue == 'A') {
15183 && ! non_portable_endpoint
15186 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
15190 else if (prevvalue == 'a') {
15193 && ! non_portable_endpoint
15196 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
15203 /* Here, we have changed <op> away from its initial value iff we found
15204 * an optimization */
15207 /* Throw away this ANYOF regnode, and emit the calculated one,
15208 * which should correspond to the beginning, not current, state of
15210 const char * cur_parse = RExC_parse;
15211 RExC_parse = (char *)orig_parse;
15215 /* To get locale nodes to not use the full ANYOF size would
15216 * require moving the code above that writes the portions
15217 * of it that aren't in other nodes to after this point.
15218 * e.g. ANYOF_POSIXL_SET */
15219 RExC_size = orig_size;
15223 RExC_emit = (regnode *)orig_emit;
15224 if (PL_regkind[op] == POSIXD) {
15225 if (op == POSIXL) {
15226 RExC_contains_locale = 1;
15229 op += NPOSIXD - POSIXD;
15234 ret = reg_node(pRExC_state, op);
15236 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
15240 *flagp |= HASWIDTH|SIMPLE;
15242 else if (PL_regkind[op] == EXACT) {
15243 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15244 TRUE /* downgradable to EXACT */
15248 RExC_parse = (char *) cur_parse;
15250 SvREFCNT_dec(posixes);
15251 SvREFCNT_dec(nposixes);
15252 SvREFCNT_dec(simple_posixes);
15253 SvREFCNT_dec(cp_list);
15254 SvREFCNT_dec(cp_foldable_list);
15261 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
15263 /* If folding, we calculate all characters that could fold to or from the
15264 * ones already on the list */
15265 if (cp_foldable_list) {
15267 UV start, end; /* End points of code point ranges */
15269 SV* fold_intersection = NULL;
15272 /* Our calculated list will be for Unicode rules. For locale
15273 * matching, we have to keep a separate list that is consulted at
15274 * runtime only when the locale indicates Unicode rules. For
15275 * non-locale, we just use to the general list */
15277 use_list = &only_utf8_locale_list;
15280 use_list = &cp_list;
15283 /* Only the characters in this class that participate in folds need
15284 * be checked. Get the intersection of this class and all the
15285 * possible characters that are foldable. This can quickly narrow
15286 * down a large class */
15287 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
15288 &fold_intersection);
15290 /* The folds for all the Latin1 characters are hard-coded into this
15291 * program, but we have to go out to disk to get the others. */
15292 if (invlist_highest(cp_foldable_list) >= 256) {
15294 /* This is a hash that for a particular fold gives all
15295 * characters that are involved in it */
15296 if (! PL_utf8_foldclosures) {
15297 _load_PL_utf8_foldclosures();
15301 /* Now look at the foldable characters in this class individually */
15302 invlist_iterinit(fold_intersection);
15303 while (invlist_iternext(fold_intersection, &start, &end)) {
15306 /* Look at every character in the range */
15307 for (j = start; j <= end; j++) {
15308 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
15314 if (IS_IN_SOME_FOLD_L1(j)) {
15316 /* ASCII is always matched; non-ASCII is matched
15317 * only under Unicode rules (which could happen
15318 * under /l if the locale is a UTF-8 one */
15319 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
15320 *use_list = add_cp_to_invlist(*use_list,
15321 PL_fold_latin1[j]);
15325 add_cp_to_invlist(depends_list,
15326 PL_fold_latin1[j]);
15330 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
15331 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
15333 add_above_Latin1_folds(pRExC_state,
15340 /* Here is an above Latin1 character. We don't have the
15341 * rules hard-coded for it. First, get its fold. This is
15342 * the simple fold, as the multi-character folds have been
15343 * handled earlier and separated out */
15344 _to_uni_fold_flags(j, foldbuf, &foldlen,
15345 (ASCII_FOLD_RESTRICTED)
15346 ? FOLD_FLAGS_NOMIX_ASCII
15349 /* Single character fold of above Latin1. Add everything in
15350 * its fold closure to the list that this node should match.
15351 * The fold closures data structure is a hash with the keys
15352 * being the UTF-8 of every character that is folded to, like
15353 * 'k', and the values each an array of all code points that
15354 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
15355 * Multi-character folds are not included */
15356 if ((listp = hv_fetch(PL_utf8_foldclosures,
15357 (char *) foldbuf, foldlen, FALSE)))
15359 AV* list = (AV*) *listp;
15361 for (k = 0; k <= av_tindex(list); k++) {
15362 SV** c_p = av_fetch(list, k, FALSE);
15368 /* /aa doesn't allow folds between ASCII and non- */
15369 if ((ASCII_FOLD_RESTRICTED
15370 && (isASCII(c) != isASCII(j))))
15375 /* Folds under /l which cross the 255/256 boundary
15376 * are added to a separate list. (These are valid
15377 * only when the locale is UTF-8.) */
15378 if (c < 256 && LOC) {
15379 *use_list = add_cp_to_invlist(*use_list, c);
15383 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
15385 cp_list = add_cp_to_invlist(cp_list, c);
15388 /* Similarly folds involving non-ascii Latin1
15389 * characters under /d are added to their list */
15390 depends_list = add_cp_to_invlist(depends_list,
15397 SvREFCNT_dec_NN(fold_intersection);
15400 /* Now that we have finished adding all the folds, there is no reason
15401 * to keep the foldable list separate */
15402 _invlist_union(cp_list, cp_foldable_list, &cp_list);
15403 SvREFCNT_dec_NN(cp_foldable_list);
15406 /* And combine the result (if any) with any inversion list from posix
15407 * classes. The lists are kept separate up to now because we don't want to
15408 * fold the classes (folding of those is automatically handled by the swash
15409 * fetching code) */
15410 if (simple_posixes) {
15411 _invlist_union(cp_list, simple_posixes, &cp_list);
15412 SvREFCNT_dec_NN(simple_posixes);
15414 if (posixes || nposixes) {
15415 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
15416 /* Under /a and /aa, nothing above ASCII matches these */
15417 _invlist_intersection(posixes,
15418 PL_XPosix_ptrs[_CC_ASCII],
15422 if (DEPENDS_SEMANTICS) {
15423 /* Under /d, everything in the upper half of the Latin1 range
15424 * matches these complements */
15425 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII;
15427 else if (AT_LEAST_ASCII_RESTRICTED) {
15428 /* Under /a and /aa, everything above ASCII matches these
15430 _invlist_union_complement_2nd(nposixes,
15431 PL_XPosix_ptrs[_CC_ASCII],
15435 _invlist_union(posixes, nposixes, &posixes);
15436 SvREFCNT_dec_NN(nposixes);
15439 posixes = nposixes;
15442 if (! DEPENDS_SEMANTICS) {
15444 _invlist_union(cp_list, posixes, &cp_list);
15445 SvREFCNT_dec_NN(posixes);
15452 /* Under /d, we put into a separate list the Latin1 things that
15453 * match only when the target string is utf8 */
15454 SV* nonascii_but_latin1_properties = NULL;
15455 _invlist_intersection(posixes, PL_UpperLatin1,
15456 &nonascii_but_latin1_properties);
15457 _invlist_subtract(posixes, nonascii_but_latin1_properties,
15460 _invlist_union(cp_list, posixes, &cp_list);
15461 SvREFCNT_dec_NN(posixes);
15467 if (depends_list) {
15468 _invlist_union(depends_list, nonascii_but_latin1_properties,
15470 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
15473 depends_list = nonascii_but_latin1_properties;
15478 /* And combine the result (if any) with any inversion list from properties.
15479 * The lists are kept separate up to now so that we can distinguish the two
15480 * in regards to matching above-Unicode. A run-time warning is generated
15481 * if a Unicode property is matched against a non-Unicode code point. But,
15482 * we allow user-defined properties to match anything, without any warning,
15483 * and we also suppress the warning if there is a portion of the character
15484 * class that isn't a Unicode property, and which matches above Unicode, \W
15485 * or [\x{110000}] for example.
15486 * (Note that in this case, unlike the Posix one above, there is no
15487 * <depends_list>, because having a Unicode property forces Unicode
15492 /* If it matters to the final outcome, see if a non-property
15493 * component of the class matches above Unicode. If so, the
15494 * warning gets suppressed. This is true even if just a single
15495 * such code point is specified, as though not strictly correct if
15496 * another such code point is matched against, the fact that they
15497 * are using above-Unicode code points indicates they should know
15498 * the issues involved */
15500 warn_super = ! (invert
15501 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
15504 _invlist_union(properties, cp_list, &cp_list);
15505 SvREFCNT_dec_NN(properties);
15508 cp_list = properties;
15512 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
15516 /* Here, we have calculated what code points should be in the character
15519 * Now we can see about various optimizations. Fold calculation (which we
15520 * did above) needs to take place before inversion. Otherwise /[^k]/i
15521 * would invert to include K, which under /i would match k, which it
15522 * shouldn't. Therefore we can't invert folded locale now, as it won't be
15523 * folded until runtime */
15525 /* If we didn't do folding, it's because some information isn't available
15526 * until runtime; set the run-time fold flag for these. (We don't have to
15527 * worry about properties folding, as that is taken care of by the swash
15528 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
15529 * locales, or the class matches at least one 0-255 range code point */
15531 if (only_utf8_locale_list) {
15532 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15534 else if (cp_list) { /* Look to see if there a 0-255 code point is in
15537 invlist_iterinit(cp_list);
15538 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
15539 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15541 invlist_iterfinish(cp_list);
15545 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
15546 * at compile time. Besides not inverting folded locale now, we can't
15547 * invert if there are things such as \w, which aren't known until runtime
15551 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15553 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15555 _invlist_invert(cp_list);
15557 /* Any swash can't be used as-is, because we've inverted things */
15559 SvREFCNT_dec_NN(swash);
15563 /* Clear the invert flag since have just done it here */
15570 *ret_invlist = cp_list;
15571 SvREFCNT_dec(swash);
15573 /* Discard the generated node */
15575 RExC_size = orig_size;
15578 RExC_emit = orig_emit;
15583 /* Some character classes are equivalent to other nodes. Such nodes take
15584 * up less room and generally fewer operations to execute than ANYOF nodes.
15585 * Above, we checked for and optimized into some such equivalents for
15586 * certain common classes that are easy to test. Getting to this point in
15587 * the code means that the class didn't get optimized there. Since this
15588 * code is only executed in Pass 2, it is too late to save space--it has
15589 * been allocated in Pass 1, and currently isn't given back. But turning
15590 * things into an EXACTish node can allow the optimizer to join it to any
15591 * adjacent such nodes. And if the class is equivalent to things like /./,
15592 * expensive run-time swashes can be avoided. Now that we have more
15593 * complete information, we can find things necessarily missed by the
15594 * earlier code. I (khw) am not sure how much to look for here. It would
15595 * be easy, but perhaps too slow, to check any candidates against all the
15596 * node types they could possibly match using _invlistEQ(). */
15601 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15602 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15604 /* We don't optimize if we are supposed to make sure all non-Unicode
15605 * code points raise a warning, as only ANYOF nodes have this check.
15607 && ! ((ANYOF_FLAGS(ret) & ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
15610 U8 op = END; /* The optimzation node-type */
15611 const char * cur_parse= RExC_parse;
15613 invlist_iterinit(cp_list);
15614 if (! invlist_iternext(cp_list, &start, &end)) {
15616 /* Here, the list is empty. This happens, for example, when a
15617 * Unicode property is the only thing in the character class, and
15618 * it doesn't match anything. (perluniprops.pod notes such
15621 *flagp |= HASWIDTH|SIMPLE;
15623 else if (start == end) { /* The range is a single code point */
15624 if (! invlist_iternext(cp_list, &start, &end)
15626 /* Don't do this optimization if it would require changing
15627 * the pattern to UTF-8 */
15628 && (start < 256 || UTF))
15630 /* Here, the list contains a single code point. Can optimize
15631 * into an EXACTish node */
15642 /* A locale node under folding with one code point can be
15643 * an EXACTFL, as its fold won't be calculated until
15649 /* Here, we are generally folding, but there is only one
15650 * code point to match. If we have to, we use an EXACT
15651 * node, but it would be better for joining with adjacent
15652 * nodes in the optimization pass if we used the same
15653 * EXACTFish node that any such are likely to be. We can
15654 * do this iff the code point doesn't participate in any
15655 * folds. For example, an EXACTF of a colon is the same as
15656 * an EXACT one, since nothing folds to or from a colon. */
15658 if (IS_IN_SOME_FOLD_L1(value)) {
15663 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
15668 /* If we haven't found the node type, above, it means we
15669 * can use the prevailing one */
15671 op = compute_EXACTish(pRExC_state);
15676 else if (start == 0) {
15677 if (end == UV_MAX) {
15679 *flagp |= HASWIDTH|SIMPLE;
15682 else if (end == '\n' - 1
15683 && invlist_iternext(cp_list, &start, &end)
15684 && start == '\n' + 1 && end == UV_MAX)
15687 *flagp |= HASWIDTH|SIMPLE;
15691 invlist_iterfinish(cp_list);
15694 RExC_parse = (char *)orig_parse;
15695 RExC_emit = (regnode *)orig_emit;
15697 ret = reg_node(pRExC_state, op);
15699 RExC_parse = (char *)cur_parse;
15701 if (PL_regkind[op] == EXACT) {
15702 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15703 TRUE /* downgradable to EXACT */
15707 SvREFCNT_dec_NN(cp_list);
15712 /* Here, <cp_list> contains all the code points we can determine at
15713 * compile time that match under all conditions. Go through it, and
15714 * for things that belong in the bitmap, put them there, and delete from
15715 * <cp_list>. While we are at it, see if everything above 255 is in the
15716 * list, and if so, set a flag to speed up execution */
15718 populate_ANYOF_from_invlist(ret, &cp_list);
15721 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
15724 /* Here, the bitmap has been populated with all the Latin1 code points that
15725 * always match. Can now add to the overall list those that match only
15726 * when the target string is UTF-8 (<depends_list>). */
15727 if (depends_list) {
15729 _invlist_union(cp_list, depends_list, &cp_list);
15730 SvREFCNT_dec_NN(depends_list);
15733 cp_list = depends_list;
15735 ANYOF_FLAGS(ret) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
15738 /* If there is a swash and more than one element, we can't use the swash in
15739 * the optimization below. */
15740 if (swash && element_count > 1) {
15741 SvREFCNT_dec_NN(swash);
15745 /* Note that the optimization of using 'swash' if it is the only thing in
15746 * the class doesn't have us change swash at all, so it can include things
15747 * that are also in the bitmap; otherwise we have purposely deleted that
15748 * duplicate information */
15749 set_ANYOF_arg(pRExC_state, ret, cp_list,
15750 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15752 only_utf8_locale_list,
15753 swash, has_user_defined_property);
15755 *flagp |= HASWIDTH|SIMPLE;
15757 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
15758 RExC_contains_locale = 1;
15764 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15767 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
15768 regnode* const node,
15770 SV* const runtime_defns,
15771 SV* const only_utf8_locale_list,
15773 const bool has_user_defined_property)
15775 /* Sets the arg field of an ANYOF-type node 'node', using information about
15776 * the node passed-in. If there is nothing outside the node's bitmap, the
15777 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
15778 * the count returned by add_data(), having allocated and stored an array,
15779 * av, that that count references, as follows:
15780 * av[0] stores the character class description in its textual form.
15781 * This is used later (regexec.c:Perl_regclass_swash()) to
15782 * initialize the appropriate swash, and is also useful for dumping
15783 * the regnode. This is set to &PL_sv_undef if the textual
15784 * description is not needed at run-time (as happens if the other
15785 * elements completely define the class)
15786 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
15787 * computed from av[0]. But if no further computation need be done,
15788 * the swash is stored here now (and av[0] is &PL_sv_undef).
15789 * av[2] stores the inversion list of code points that match only if the
15790 * current locale is UTF-8
15791 * av[3] stores the cp_list inversion list for use in addition or instead
15792 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
15793 * (Otherwise everything needed is already in av[0] and av[1])
15794 * av[4] is set if any component of the class is from a user-defined
15795 * property; used only if av[3] exists */
15799 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
15801 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
15802 assert(! (ANYOF_FLAGS(node)
15803 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15804 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES)));
15805 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
15808 AV * const av = newAV();
15811 assert(ANYOF_FLAGS(node)
15812 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15813 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
15815 av_store(av, 0, (runtime_defns)
15816 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
15819 av_store(av, 1, swash);
15820 SvREFCNT_dec_NN(cp_list);
15823 av_store(av, 1, &PL_sv_undef);
15825 av_store(av, 3, cp_list);
15826 av_store(av, 4, newSVuv(has_user_defined_property));
15830 if (only_utf8_locale_list) {
15831 av_store(av, 2, only_utf8_locale_list);
15834 av_store(av, 2, &PL_sv_undef);
15837 rv = newRV_noinc(MUTABLE_SV(av));
15838 n = add_data(pRExC_state, STR_WITH_LEN("s"));
15839 RExC_rxi->data->data[n] = (void*)rv;
15844 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
15846 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
15847 const regnode* node,
15850 SV** only_utf8_locale_ptr,
15854 /* For internal core use only.
15855 * Returns the swash for the input 'node' in the regex 'prog'.
15856 * If <doinit> is 'true', will attempt to create the swash if not already
15858 * If <listsvp> is non-null, will return the printable contents of the
15859 * swash. This can be used to get debugging information even before the
15860 * swash exists, by calling this function with 'doinit' set to false, in
15861 * which case the components that will be used to eventually create the
15862 * swash are returned (in a printable form).
15863 * If <exclude_list> is not NULL, it is an inversion list of things to
15864 * exclude from what's returned in <listsvp>.
15865 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
15866 * that, in spite of this function's name, the swash it returns may include
15867 * the bitmap data as well */
15870 SV *si = NULL; /* Input swash initialization string */
15871 SV* invlist = NULL;
15873 RXi_GET_DECL(prog,progi);
15874 const struct reg_data * const data = prog ? progi->data : NULL;
15876 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
15878 assert(ANYOF_FLAGS(node)
15879 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
15880 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
15882 if (data && data->count) {
15883 const U32 n = ARG(node);
15885 if (data->what[n] == 's') {
15886 SV * const rv = MUTABLE_SV(data->data[n]);
15887 AV * const av = MUTABLE_AV(SvRV(rv));
15888 SV **const ary = AvARRAY(av);
15889 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
15891 si = *ary; /* ary[0] = the string to initialize the swash with */
15893 /* Elements 3 and 4 are either both present or both absent. [3] is
15894 * any inversion list generated at compile time; [4] indicates if
15895 * that inversion list has any user-defined properties in it. */
15896 if (av_tindex(av) >= 2) {
15897 if (only_utf8_locale_ptr
15899 && ary[2] != &PL_sv_undef)
15901 *only_utf8_locale_ptr = ary[2];
15904 assert(only_utf8_locale_ptr);
15905 *only_utf8_locale_ptr = NULL;
15908 if (av_tindex(av) >= 3) {
15910 if (SvUV(ary[4])) {
15911 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
15919 /* Element [1] is reserved for the set-up swash. If already there,
15920 * return it; if not, create it and store it there */
15921 if (ary[1] && SvROK(ary[1])) {
15924 else if (doinit && ((si && si != &PL_sv_undef)
15925 || (invlist && invlist != &PL_sv_undef))) {
15927 sw = _core_swash_init("utf8", /* the utf8 package */
15931 0, /* not from tr/// */
15933 &swash_init_flags);
15934 (void)av_store(av, 1, sw);
15939 /* If requested, return a printable version of what this swash matches */
15941 SV* matches_string = newSVpvs("");
15943 /* The swash should be used, if possible, to get the data, as it
15944 * contains the resolved data. But this function can be called at
15945 * compile-time, before everything gets resolved, in which case we
15946 * return the currently best available information, which is the string
15947 * that will eventually be used to do that resolving, 'si' */
15948 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
15949 && (si && si != &PL_sv_undef))
15951 sv_catsv(matches_string, si);
15954 /* Add the inversion list to whatever we have. This may have come from
15955 * the swash, or from an input parameter */
15957 if (exclude_list) {
15958 SV* clone = invlist_clone(invlist);
15959 _invlist_subtract(clone, exclude_list, &clone);
15960 sv_catsv(matches_string, _invlist_contents(clone));
15961 SvREFCNT_dec_NN(clone);
15964 sv_catsv(matches_string, _invlist_contents(invlist));
15967 *listsvp = matches_string;
15972 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
15974 /* reg_skipcomment()
15976 Absorbs an /x style # comment from the input stream,
15977 returning a pointer to the first character beyond the comment, or if the
15978 comment terminates the pattern without anything following it, this returns
15979 one past the final character of the pattern (in other words, RExC_end) and
15980 sets the REG_RUN_ON_COMMENT_SEEN flag.
15982 Note it's the callers responsibility to ensure that we are
15983 actually in /x mode
15987 PERL_STATIC_INLINE char*
15988 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
15990 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
15994 while (p < RExC_end) {
15995 if (*(++p) == '\n') {
16000 /* we ran off the end of the pattern without ending the comment, so we have
16001 * to add an \n when wrapping */
16002 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
16008 Advances the parse position, and optionally absorbs
16009 "whitespace" from the inputstream.
16011 Without /x "whitespace" means (?#...) style comments only,
16012 with /x this means (?#...) and # comments and whitespace proper.
16014 Returns the RExC_parse point from BEFORE the scan occurs.
16016 This is the /x friendly way of saying RExC_parse++.
16020 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
16022 char* const retval = RExC_parse++;
16024 PERL_ARGS_ASSERT_NEXTCHAR;
16027 if (RExC_end - RExC_parse >= 3
16028 && *RExC_parse == '('
16029 && RExC_parse[1] == '?'
16030 && RExC_parse[2] == '#')
16032 while (*RExC_parse != ')') {
16033 if (RExC_parse == RExC_end)
16034 FAIL("Sequence (?#... not terminated");
16040 if (RExC_flags & RXf_PMf_EXTENDED) {
16041 char * p = regpatws(pRExC_state, RExC_parse,
16042 TRUE); /* means recognize comments */
16043 if (p != RExC_parse) {
16053 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
16055 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
16056 * space. In pass1, it aligns and increments RExC_size; in pass2,
16059 regnode * const ret = RExC_emit;
16060 GET_RE_DEBUG_FLAGS_DECL;
16062 PERL_ARGS_ASSERT_REGNODE_GUTS;
16064 assert(extra_size >= regarglen[op]);
16067 SIZE_ALIGN(RExC_size);
16068 RExC_size += 1 + extra_size;
16071 if (RExC_emit >= RExC_emit_bound)
16072 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
16073 op, (void*)RExC_emit, (void*)RExC_emit_bound);
16075 NODE_ALIGN_FILL(ret);
16076 #ifndef RE_TRACK_PATTERN_OFFSETS
16077 PERL_UNUSED_ARG(name);
16079 if (RExC_offsets) { /* MJD */
16081 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
16084 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
16085 ? "Overwriting end of array!\n" : "OK",
16086 (UV)(RExC_emit - RExC_emit_start),
16087 (UV)(RExC_parse - RExC_start),
16088 (UV)RExC_offsets[0]));
16089 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
16096 - reg_node - emit a node
16098 STATIC regnode * /* Location. */
16099 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
16101 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
16103 PERL_ARGS_ASSERT_REG_NODE;
16105 assert(regarglen[op] == 0);
16108 regnode *ptr = ret;
16109 FILL_ADVANCE_NODE(ptr, op);
16116 - reganode - emit a node with an argument
16118 STATIC regnode * /* Location. */
16119 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
16121 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
16123 PERL_ARGS_ASSERT_REGANODE;
16125 assert(regarglen[op] == 1);
16128 regnode *ptr = ret;
16129 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
16136 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
16138 /* emit a node with U32 and I32 arguments */
16140 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
16142 PERL_ARGS_ASSERT_REG2LANODE;
16144 assert(regarglen[op] == 2);
16147 regnode *ptr = ret;
16148 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
16155 - reginsert - insert an operator in front of already-emitted operand
16157 * Means relocating the operand.
16160 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
16165 const int offset = regarglen[(U8)op];
16166 const int size = NODE_STEP_REGNODE + offset;
16167 GET_RE_DEBUG_FLAGS_DECL;
16169 PERL_ARGS_ASSERT_REGINSERT;
16170 PERL_UNUSED_CONTEXT;
16171 PERL_UNUSED_ARG(depth);
16172 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
16173 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
16182 if (RExC_open_parens) {
16184 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
16185 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
16186 if ( RExC_open_parens[paren] >= opnd ) {
16187 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
16188 RExC_open_parens[paren] += size;
16190 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
16192 if ( RExC_close_parens[paren] >= opnd ) {
16193 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
16194 RExC_close_parens[paren] += size;
16196 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
16201 while (src > opnd) {
16202 StructCopy(--src, --dst, regnode);
16203 #ifdef RE_TRACK_PATTERN_OFFSETS
16204 if (RExC_offsets) { /* MJD 20010112 */
16206 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
16210 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
16211 ? "Overwriting end of array!\n" : "OK",
16212 (UV)(src - RExC_emit_start),
16213 (UV)(dst - RExC_emit_start),
16214 (UV)RExC_offsets[0]));
16215 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
16216 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
16222 place = opnd; /* Op node, where operand used to be. */
16223 #ifdef RE_TRACK_PATTERN_OFFSETS
16224 if (RExC_offsets) { /* MJD */
16226 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
16230 (UV)(place - RExC_emit_start) > RExC_offsets[0]
16231 ? "Overwriting end of array!\n" : "OK",
16232 (UV)(place - RExC_emit_start),
16233 (UV)(RExC_parse - RExC_start),
16234 (UV)RExC_offsets[0]));
16235 Set_Node_Offset(place, RExC_parse);
16236 Set_Node_Length(place, 1);
16239 src = NEXTOPER(place);
16240 FILL_ADVANCE_NODE(place, op);
16241 Zero(src, offset, regnode);
16245 - regtail - set the next-pointer at the end of a node chain of p to val.
16246 - SEE ALSO: regtail_study
16248 /* TODO: All three parms should be const */
16250 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16251 const regnode *val,U32 depth)
16254 GET_RE_DEBUG_FLAGS_DECL;
16256 PERL_ARGS_ASSERT_REGTAIL;
16258 PERL_UNUSED_ARG(depth);
16264 /* Find last node. */
16267 regnode * const temp = regnext(scan);
16269 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
16270 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16271 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
16272 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
16273 (temp == NULL ? "->" : ""),
16274 (temp == NULL ? PL_reg_name[OP(val)] : "")
16282 if (reg_off_by_arg[OP(scan)]) {
16283 ARG_SET(scan, val - scan);
16286 NEXT_OFF(scan) = val - scan;
16292 - regtail_study - set the next-pointer at the end of a node chain of p to val.
16293 - Look for optimizable sequences at the same time.
16294 - currently only looks for EXACT chains.
16296 This is experimental code. The idea is to use this routine to perform
16297 in place optimizations on branches and groups as they are constructed,
16298 with the long term intention of removing optimization from study_chunk so
16299 that it is purely analytical.
16301 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
16302 to control which is which.
16305 /* TODO: All four parms should be const */
16308 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16309 const regnode *val,U32 depth)
16313 #ifdef EXPERIMENTAL_INPLACESCAN
16316 GET_RE_DEBUG_FLAGS_DECL;
16318 PERL_ARGS_ASSERT_REGTAIL_STUDY;
16324 /* Find last node. */
16328 regnode * const temp = regnext(scan);
16329 #ifdef EXPERIMENTAL_INPLACESCAN
16330 if (PL_regkind[OP(scan)] == EXACT) {
16331 bool unfolded_multi_char; /* Unexamined in this routine */
16332 if (join_exact(pRExC_state, scan, &min,
16333 &unfolded_multi_char, 1, val, depth+1))
16338 switch (OP(scan)) {
16342 case EXACTFA_NO_TRIE:
16348 if( exact == PSEUDO )
16350 else if ( exact != OP(scan) )
16359 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
16360 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16361 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
16362 SvPV_nolen_const(RExC_mysv),
16363 REG_NODE_NUM(scan),
16364 PL_reg_name[exact]);
16371 DEBUG_PARSE_MSG("");
16372 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
16373 PerlIO_printf(Perl_debug_log,
16374 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
16375 SvPV_nolen_const(RExC_mysv),
16376 (IV)REG_NODE_NUM(val),
16380 if (reg_off_by_arg[OP(scan)]) {
16381 ARG_SET(scan, val - scan);
16384 NEXT_OFF(scan) = val - scan;
16392 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
16397 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
16402 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16404 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
16405 if (flags & (1<<bit)) {
16406 if (!set++ && lead)
16407 PerlIO_printf(Perl_debug_log, "%s",lead);
16408 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
16413 PerlIO_printf(Perl_debug_log, "\n");
16415 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16420 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
16426 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16428 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
16429 if (flags & (1<<bit)) {
16430 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
16433 if (!set++ && lead)
16434 PerlIO_printf(Perl_debug_log, "%s",lead);
16435 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
16438 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
16439 if (!set++ && lead) {
16440 PerlIO_printf(Perl_debug_log, "%s",lead);
16443 case REGEX_UNICODE_CHARSET:
16444 PerlIO_printf(Perl_debug_log, "UNICODE");
16446 case REGEX_LOCALE_CHARSET:
16447 PerlIO_printf(Perl_debug_log, "LOCALE");
16449 case REGEX_ASCII_RESTRICTED_CHARSET:
16450 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
16452 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
16453 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
16456 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
16462 PerlIO_printf(Perl_debug_log, "\n");
16464 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16470 Perl_regdump(pTHX_ const regexp *r)
16473 SV * const sv = sv_newmortal();
16474 SV *dsv= sv_newmortal();
16475 RXi_GET_DECL(r,ri);
16476 GET_RE_DEBUG_FLAGS_DECL;
16478 PERL_ARGS_ASSERT_REGDUMP;
16480 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
16482 /* Header fields of interest. */
16483 if (r->anchored_substr) {
16484 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
16485 RE_SV_DUMPLEN(r->anchored_substr), 30);
16486 PerlIO_printf(Perl_debug_log,
16487 "anchored %s%s at %"IVdf" ",
16488 s, RE_SV_TAIL(r->anchored_substr),
16489 (IV)r->anchored_offset);
16490 } else if (r->anchored_utf8) {
16491 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
16492 RE_SV_DUMPLEN(r->anchored_utf8), 30);
16493 PerlIO_printf(Perl_debug_log,
16494 "anchored utf8 %s%s at %"IVdf" ",
16495 s, RE_SV_TAIL(r->anchored_utf8),
16496 (IV)r->anchored_offset);
16498 if (r->float_substr) {
16499 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
16500 RE_SV_DUMPLEN(r->float_substr), 30);
16501 PerlIO_printf(Perl_debug_log,
16502 "floating %s%s at %"IVdf"..%"UVuf" ",
16503 s, RE_SV_TAIL(r->float_substr),
16504 (IV)r->float_min_offset, (UV)r->float_max_offset);
16505 } else if (r->float_utf8) {
16506 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
16507 RE_SV_DUMPLEN(r->float_utf8), 30);
16508 PerlIO_printf(Perl_debug_log,
16509 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
16510 s, RE_SV_TAIL(r->float_utf8),
16511 (IV)r->float_min_offset, (UV)r->float_max_offset);
16513 if (r->check_substr || r->check_utf8)
16514 PerlIO_printf(Perl_debug_log,
16516 (r->check_substr == r->float_substr
16517 && r->check_utf8 == r->float_utf8
16518 ? "(checking floating" : "(checking anchored"));
16519 if (r->intflags & PREGf_NOSCAN)
16520 PerlIO_printf(Perl_debug_log, " noscan");
16521 if (r->extflags & RXf_CHECK_ALL)
16522 PerlIO_printf(Perl_debug_log, " isall");
16523 if (r->check_substr || r->check_utf8)
16524 PerlIO_printf(Perl_debug_log, ") ");
16526 if (ri->regstclass) {
16527 regprop(r, sv, ri->regstclass, NULL, NULL);
16528 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
16530 if (r->intflags & PREGf_ANCH) {
16531 PerlIO_printf(Perl_debug_log, "anchored");
16532 if (r->intflags & PREGf_ANCH_MBOL)
16533 PerlIO_printf(Perl_debug_log, "(MBOL)");
16534 if (r->intflags & PREGf_ANCH_SBOL)
16535 PerlIO_printf(Perl_debug_log, "(SBOL)");
16536 if (r->intflags & PREGf_ANCH_GPOS)
16537 PerlIO_printf(Perl_debug_log, "(GPOS)");
16538 PerlIO_putc(Perl_debug_log, ' ');
16540 if (r->intflags & PREGf_GPOS_SEEN)
16541 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
16542 if (r->intflags & PREGf_SKIP)
16543 PerlIO_printf(Perl_debug_log, "plus ");
16544 if (r->intflags & PREGf_IMPLICIT)
16545 PerlIO_printf(Perl_debug_log, "implicit ");
16546 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
16547 if (r->extflags & RXf_EVAL_SEEN)
16548 PerlIO_printf(Perl_debug_log, "with eval ");
16549 PerlIO_printf(Perl_debug_log, "\n");
16551 regdump_extflags("r->extflags: ",r->extflags);
16552 regdump_intflags("r->intflags: ",r->intflags);
16555 PERL_ARGS_ASSERT_REGDUMP;
16556 PERL_UNUSED_CONTEXT;
16557 PERL_UNUSED_ARG(r);
16558 #endif /* DEBUGGING */
16562 - regprop - printable representation of opcode, with run time support
16566 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
16571 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
16572 static const char * const anyofs[] = {
16573 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
16574 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
16575 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
16576 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
16577 || _CC_CNTRL != 13 || _CC_ASCII != 14 || _CC_VERTSPACE != 15
16578 #error Need to adjust order of anyofs[]
16613 RXi_GET_DECL(prog,progi);
16614 GET_RE_DEBUG_FLAGS_DECL;
16616 PERL_ARGS_ASSERT_REGPROP;
16618 sv_setpvn(sv, "", 0);
16620 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
16621 /* It would be nice to FAIL() here, but this may be called from
16622 regexec.c, and it would be hard to supply pRExC_state. */
16623 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16624 (int)OP(o), (int)REGNODE_MAX);
16625 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
16627 k = PL_regkind[OP(o)];
16630 sv_catpvs(sv, " ");
16631 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
16632 * is a crude hack but it may be the best for now since
16633 * we have no flag "this EXACTish node was UTF-8"
16635 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
16636 PERL_PV_ESCAPE_UNI_DETECT |
16637 PERL_PV_ESCAPE_NONASCII |
16638 PERL_PV_PRETTY_ELLIPSES |
16639 PERL_PV_PRETTY_LTGT |
16640 PERL_PV_PRETTY_NOCLEAR
16642 } else if (k == TRIE) {
16643 /* print the details of the trie in dumpuntil instead, as
16644 * progi->data isn't available here */
16645 const char op = OP(o);
16646 const U32 n = ARG(o);
16647 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
16648 (reg_ac_data *)progi->data->data[n] :
16650 const reg_trie_data * const trie
16651 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
16653 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
16654 DEBUG_TRIE_COMPILE_r(
16655 Perl_sv_catpvf(aTHX_ sv,
16656 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
16657 (UV)trie->startstate,
16658 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
16659 (UV)trie->wordcount,
16662 (UV)TRIE_CHARCOUNT(trie),
16663 (UV)trie->uniquecharcount
16666 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
16667 sv_catpvs(sv, "[");
16668 (void) put_charclass_bitmap_innards(sv,
16669 (IS_ANYOF_TRIE(op))
16671 : TRIE_BITMAP(trie),
16673 sv_catpvs(sv, "]");
16676 } else if (k == CURLY) {
16677 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
16678 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
16679 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
16681 else if (k == WHILEM && o->flags) /* Ordinal/of */
16682 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
16683 else if (k == REF || k == OPEN || k == CLOSE
16684 || k == GROUPP || OP(o)==ACCEPT)
16686 AV *name_list= NULL;
16687 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
16688 if ( RXp_PAREN_NAMES(prog) ) {
16689 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16690 } else if ( pRExC_state ) {
16691 name_list= RExC_paren_name_list;
16694 if ( k != REF || (OP(o) < NREF)) {
16695 SV **name= av_fetch(name_list, ARG(o), 0 );
16697 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16700 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
16701 I32 *nums=(I32*)SvPVX(sv_dat);
16702 SV **name= av_fetch(name_list, nums[0], 0 );
16705 for ( n=0; n<SvIVX(sv_dat); n++ ) {
16706 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
16707 (n ? "," : ""), (IV)nums[n]);
16709 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16713 if ( k == REF && reginfo) {
16714 U32 n = ARG(o); /* which paren pair */
16715 I32 ln = prog->offs[n].start;
16716 if (prog->lastparen < n || ln == -1)
16717 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
16718 else if (ln == prog->offs[n].end)
16719 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
16721 const char *s = reginfo->strbeg + ln;
16722 Perl_sv_catpvf(aTHX_ sv, ": ");
16723 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
16724 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
16727 } else if (k == GOSUB) {
16728 AV *name_list= NULL;
16729 if ( RXp_PAREN_NAMES(prog) ) {
16730 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16731 } else if ( pRExC_state ) {
16732 name_list= RExC_paren_name_list;
16735 /* Paren and offset */
16736 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
16738 SV **name= av_fetch(name_list, ARG(o), 0 );
16740 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16743 else if (k == VERB) {
16745 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
16746 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
16747 } else if (k == LOGICAL)
16748 /* 2: embedded, otherwise 1 */
16749 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
16750 else if (k == ANYOF) {
16751 const U8 flags = ANYOF_FLAGS(o);
16753 SV* bitmap_invlist; /* Will hold what the bit map contains */
16756 if (OP(o) == ANYOFL)
16757 sv_catpvs(sv, "{loc}");
16758 if (flags & ANYOF_LOC_FOLD)
16759 sv_catpvs(sv, "{i}");
16760 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
16761 if (flags & ANYOF_INVERT)
16762 sv_catpvs(sv, "^");
16764 /* output what the standard cp 0-NUM_ANYOF_CODE_POINTS-1 bitmap matches
16766 do_sep = put_charclass_bitmap_innards(sv, ANYOF_BITMAP(o),
16769 /* output any special charclass tests (used entirely under use
16771 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
16773 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
16774 if (ANYOF_POSIXL_TEST(o,i)) {
16775 sv_catpv(sv, anyofs[i]);
16781 if ((flags & (ANYOF_MATCHES_ALL_ABOVE_BITMAP
16782 |ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16783 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES
16787 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
16788 if (flags & ANYOF_INVERT)
16789 /*make sure the invert info is in each */
16790 sv_catpvs(sv, "^");
16793 if (flags & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
16794 sv_catpvs(sv, "{non-utf8-latin1-all}");
16797 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP)
16798 sv_catpvs(sv, "{above_bitmap_all}");
16800 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
16801 SV *lv; /* Set if there is something outside the bit map. */
16802 bool byte_output = FALSE; /* If something has been output */
16803 SV *only_utf8_locale;
16805 /* Get the stuff that wasn't in the bitmap. 'bitmap_invlist'
16806 * is used to guarantee that nothing in the bitmap gets
16808 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
16809 &lv, &only_utf8_locale,
16811 if (lv && lv != &PL_sv_undef) {
16812 char *s = savesvpv(lv);
16813 char * const origs = s;
16815 while (*s && *s != '\n')
16819 const char * const t = ++s;
16821 if (flags & ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES) {
16822 sv_catpvs(sv, "{outside bitmap}");
16825 sv_catpvs(sv, "{utf8}");
16829 sv_catpvs(sv, " ");
16835 /* Truncate very long output */
16836 if (s - origs > 256) {
16837 Perl_sv_catpvf(aTHX_ sv,
16839 (int) (s - origs - 1),
16845 else if (*s == '\t') {
16859 SvREFCNT_dec_NN(lv);
16862 if ((flags & ANYOF_LOC_FOLD)
16863 && only_utf8_locale
16864 && only_utf8_locale != &PL_sv_undef)
16867 int max_entries = 256;
16869 sv_catpvs(sv, "{utf8 locale}");
16870 invlist_iterinit(only_utf8_locale);
16871 while (invlist_iternext(only_utf8_locale,
16873 put_range(sv, start, end, FALSE);
16875 if (max_entries < 0) {
16876 sv_catpvs(sv, "...");
16880 invlist_iterfinish(only_utf8_locale);
16884 SvREFCNT_dec(bitmap_invlist);
16887 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
16889 else if (k == POSIXD || k == NPOSIXD) {
16890 U8 index = FLAGS(o) * 2;
16891 if (index < C_ARRAY_LENGTH(anyofs)) {
16892 if (*anyofs[index] != '[') {
16895 sv_catpv(sv, anyofs[index]);
16896 if (*anyofs[index] != '[') {
16901 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
16904 else if (k == BOUND || k == NBOUND) {
16905 /* Must be synced with order of 'bound_type' in regcomp.h */
16906 const char * const bounds[] = {
16907 "", /* Traditional */
16912 sv_catpv(sv, bounds[FLAGS(o)]);
16914 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
16915 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
16916 else if (OP(o) == SBOL)
16917 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
16919 PERL_UNUSED_CONTEXT;
16920 PERL_UNUSED_ARG(sv);
16921 PERL_UNUSED_ARG(o);
16922 PERL_UNUSED_ARG(prog);
16923 PERL_UNUSED_ARG(reginfo);
16924 PERL_UNUSED_ARG(pRExC_state);
16925 #endif /* DEBUGGING */
16931 Perl_re_intuit_string(pTHX_ REGEXP * const r)
16932 { /* Assume that RE_INTUIT is set */
16933 struct regexp *const prog = ReANY(r);
16934 GET_RE_DEBUG_FLAGS_DECL;
16936 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
16937 PERL_UNUSED_CONTEXT;
16941 const char * const s = SvPV_nolen_const(RX_UTF8(r)
16942 ? prog->check_utf8 : prog->check_substr);
16944 if (!PL_colorset) reginitcolors();
16945 PerlIO_printf(Perl_debug_log,
16946 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
16948 RX_UTF8(r) ? "utf8 " : "",
16949 PL_colors[5],PL_colors[0],
16952 (strlen(s) > 60 ? "..." : ""));
16955 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
16956 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
16962 handles refcounting and freeing the perl core regexp structure. When
16963 it is necessary to actually free the structure the first thing it
16964 does is call the 'free' method of the regexp_engine associated to
16965 the regexp, allowing the handling of the void *pprivate; member
16966 first. (This routine is not overridable by extensions, which is why
16967 the extensions free is called first.)
16969 See regdupe and regdupe_internal if you change anything here.
16971 #ifndef PERL_IN_XSUB_RE
16973 Perl_pregfree(pTHX_ REGEXP *r)
16979 Perl_pregfree2(pTHX_ REGEXP *rx)
16981 struct regexp *const r = ReANY(rx);
16982 GET_RE_DEBUG_FLAGS_DECL;
16984 PERL_ARGS_ASSERT_PREGFREE2;
16986 if (r->mother_re) {
16987 ReREFCNT_dec(r->mother_re);
16989 CALLREGFREE_PVT(rx); /* free the private data */
16990 SvREFCNT_dec(RXp_PAREN_NAMES(r));
16991 Safefree(r->xpv_len_u.xpvlenu_pv);
16994 SvREFCNT_dec(r->anchored_substr);
16995 SvREFCNT_dec(r->anchored_utf8);
16996 SvREFCNT_dec(r->float_substr);
16997 SvREFCNT_dec(r->float_utf8);
16998 Safefree(r->substrs);
17000 RX_MATCH_COPY_FREE(rx);
17001 #ifdef PERL_ANY_COW
17002 SvREFCNT_dec(r->saved_copy);
17005 SvREFCNT_dec(r->qr_anoncv);
17006 rx->sv_u.svu_rx = 0;
17011 This is a hacky workaround to the structural issue of match results
17012 being stored in the regexp structure which is in turn stored in
17013 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
17014 could be PL_curpm in multiple contexts, and could require multiple
17015 result sets being associated with the pattern simultaneously, such
17016 as when doing a recursive match with (??{$qr})
17018 The solution is to make a lightweight copy of the regexp structure
17019 when a qr// is returned from the code executed by (??{$qr}) this
17020 lightweight copy doesn't actually own any of its data except for
17021 the starp/end and the actual regexp structure itself.
17027 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
17029 struct regexp *ret;
17030 struct regexp *const r = ReANY(rx);
17031 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
17033 PERL_ARGS_ASSERT_REG_TEMP_COPY;
17036 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
17038 SvOK_off((SV *)ret_x);
17040 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
17041 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
17042 made both spots point to the same regexp body.) */
17043 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
17044 assert(!SvPVX(ret_x));
17045 ret_x->sv_u.svu_rx = temp->sv_any;
17046 temp->sv_any = NULL;
17047 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
17048 SvREFCNT_dec_NN(temp);
17049 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
17050 ing below will not set it. */
17051 SvCUR_set(ret_x, SvCUR(rx));
17054 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
17055 sv_force_normal(sv) is called. */
17057 ret = ReANY(ret_x);
17059 SvFLAGS(ret_x) |= SvUTF8(rx);
17060 /* We share the same string buffer as the original regexp, on which we
17061 hold a reference count, incremented when mother_re is set below.
17062 The string pointer is copied here, being part of the regexp struct.
17064 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
17065 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
17067 const I32 npar = r->nparens+1;
17068 Newx(ret->offs, npar, regexp_paren_pair);
17069 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17072 Newx(ret->substrs, 1, struct reg_substr_data);
17073 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17075 SvREFCNT_inc_void(ret->anchored_substr);
17076 SvREFCNT_inc_void(ret->anchored_utf8);
17077 SvREFCNT_inc_void(ret->float_substr);
17078 SvREFCNT_inc_void(ret->float_utf8);
17080 /* check_substr and check_utf8, if non-NULL, point to either their
17081 anchored or float namesakes, and don't hold a second reference. */
17083 RX_MATCH_COPIED_off(ret_x);
17084 #ifdef PERL_ANY_COW
17085 ret->saved_copy = NULL;
17087 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
17088 SvREFCNT_inc_void(ret->qr_anoncv);
17094 /* regfree_internal()
17096 Free the private data in a regexp. This is overloadable by
17097 extensions. Perl takes care of the regexp structure in pregfree(),
17098 this covers the *pprivate pointer which technically perl doesn't
17099 know about, however of course we have to handle the
17100 regexp_internal structure when no extension is in use.
17102 Note this is called before freeing anything in the regexp
17107 Perl_regfree_internal(pTHX_ REGEXP * const rx)
17109 struct regexp *const r = ReANY(rx);
17110 RXi_GET_DECL(r,ri);
17111 GET_RE_DEBUG_FLAGS_DECL;
17113 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
17119 SV *dsv= sv_newmortal();
17120 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
17121 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
17122 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
17123 PL_colors[4],PL_colors[5],s);
17126 #ifdef RE_TRACK_PATTERN_OFFSETS
17128 Safefree(ri->u.offsets); /* 20010421 MJD */
17130 if (ri->code_blocks) {
17132 for (n = 0; n < ri->num_code_blocks; n++)
17133 SvREFCNT_dec(ri->code_blocks[n].src_regex);
17134 Safefree(ri->code_blocks);
17138 int n = ri->data->count;
17141 /* If you add a ->what type here, update the comment in regcomp.h */
17142 switch (ri->data->what[n]) {
17148 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
17151 Safefree(ri->data->data[n]);
17157 { /* Aho Corasick add-on structure for a trie node.
17158 Used in stclass optimization only */
17160 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
17161 #ifdef USE_ITHREADS
17165 refcount = --aho->refcount;
17168 PerlMemShared_free(aho->states);
17169 PerlMemShared_free(aho->fail);
17170 /* do this last!!!! */
17171 PerlMemShared_free(ri->data->data[n]);
17172 /* we should only ever get called once, so
17173 * assert as much, and also guard the free
17174 * which /might/ happen twice. At the least
17175 * it will make code anlyzers happy and it
17176 * doesn't cost much. - Yves */
17177 assert(ri->regstclass);
17178 if (ri->regstclass) {
17179 PerlMemShared_free(ri->regstclass);
17180 ri->regstclass = 0;
17187 /* trie structure. */
17189 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
17190 #ifdef USE_ITHREADS
17194 refcount = --trie->refcount;
17197 PerlMemShared_free(trie->charmap);
17198 PerlMemShared_free(trie->states);
17199 PerlMemShared_free(trie->trans);
17201 PerlMemShared_free(trie->bitmap);
17203 PerlMemShared_free(trie->jump);
17204 PerlMemShared_free(trie->wordinfo);
17205 /* do this last!!!! */
17206 PerlMemShared_free(ri->data->data[n]);
17211 Perl_croak(aTHX_ "panic: regfree data code '%c'",
17212 ri->data->what[n]);
17215 Safefree(ri->data->what);
17216 Safefree(ri->data);
17222 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
17223 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
17224 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
17227 re_dup - duplicate a regexp.
17229 This routine is expected to clone a given regexp structure. It is only
17230 compiled under USE_ITHREADS.
17232 After all of the core data stored in struct regexp is duplicated
17233 the regexp_engine.dupe method is used to copy any private data
17234 stored in the *pprivate pointer. This allows extensions to handle
17235 any duplication it needs to do.
17237 See pregfree() and regfree_internal() if you change anything here.
17239 #if defined(USE_ITHREADS)
17240 #ifndef PERL_IN_XSUB_RE
17242 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
17246 const struct regexp *r = ReANY(sstr);
17247 struct regexp *ret = ReANY(dstr);
17249 PERL_ARGS_ASSERT_RE_DUP_GUTS;
17251 npar = r->nparens+1;
17252 Newx(ret->offs, npar, regexp_paren_pair);
17253 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17255 if (ret->substrs) {
17256 /* Do it this way to avoid reading from *r after the StructCopy().
17257 That way, if any of the sv_dup_inc()s dislodge *r from the L1
17258 cache, it doesn't matter. */
17259 const bool anchored = r->check_substr
17260 ? r->check_substr == r->anchored_substr
17261 : r->check_utf8 == r->anchored_utf8;
17262 Newx(ret->substrs, 1, struct reg_substr_data);
17263 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17265 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
17266 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
17267 ret->float_substr = sv_dup_inc(ret->float_substr, param);
17268 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
17270 /* check_substr and check_utf8, if non-NULL, point to either their
17271 anchored or float namesakes, and don't hold a second reference. */
17273 if (ret->check_substr) {
17275 assert(r->check_utf8 == r->anchored_utf8);
17276 ret->check_substr = ret->anchored_substr;
17277 ret->check_utf8 = ret->anchored_utf8;
17279 assert(r->check_substr == r->float_substr);
17280 assert(r->check_utf8 == r->float_utf8);
17281 ret->check_substr = ret->float_substr;
17282 ret->check_utf8 = ret->float_utf8;
17284 } else if (ret->check_utf8) {
17286 ret->check_utf8 = ret->anchored_utf8;
17288 ret->check_utf8 = ret->float_utf8;
17293 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
17294 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
17297 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
17299 if (RX_MATCH_COPIED(dstr))
17300 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
17302 ret->subbeg = NULL;
17303 #ifdef PERL_ANY_COW
17304 ret->saved_copy = NULL;
17307 /* Whether mother_re be set or no, we need to copy the string. We
17308 cannot refrain from copying it when the storage points directly to
17309 our mother regexp, because that's
17310 1: a buffer in a different thread
17311 2: something we no longer hold a reference on
17312 so we need to copy it locally. */
17313 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
17314 ret->mother_re = NULL;
17316 #endif /* PERL_IN_XSUB_RE */
17321 This is the internal complement to regdupe() which is used to copy
17322 the structure pointed to by the *pprivate pointer in the regexp.
17323 This is the core version of the extension overridable cloning hook.
17324 The regexp structure being duplicated will be copied by perl prior
17325 to this and will be provided as the regexp *r argument, however
17326 with the /old/ structures pprivate pointer value. Thus this routine
17327 may override any copying normally done by perl.
17329 It returns a pointer to the new regexp_internal structure.
17333 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
17336 struct regexp *const r = ReANY(rx);
17337 regexp_internal *reti;
17339 RXi_GET_DECL(r,ri);
17341 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
17345 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
17346 char, regexp_internal);
17347 Copy(ri->program, reti->program, len+1, regnode);
17349 reti->num_code_blocks = ri->num_code_blocks;
17350 if (ri->code_blocks) {
17352 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
17353 struct reg_code_block);
17354 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
17355 struct reg_code_block);
17356 for (n = 0; n < ri->num_code_blocks; n++)
17357 reti->code_blocks[n].src_regex = (REGEXP*)
17358 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
17361 reti->code_blocks = NULL;
17363 reti->regstclass = NULL;
17366 struct reg_data *d;
17367 const int count = ri->data->count;
17370 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
17371 char, struct reg_data);
17372 Newx(d->what, count, U8);
17375 for (i = 0; i < count; i++) {
17376 d->what[i] = ri->data->what[i];
17377 switch (d->what[i]) {
17378 /* see also regcomp.h and regfree_internal() */
17379 case 'a': /* actually an AV, but the dup function is identical. */
17383 case 'u': /* actually an HV, but the dup function is identical. */
17384 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
17387 /* This is cheating. */
17388 Newx(d->data[i], 1, regnode_ssc);
17389 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
17390 reti->regstclass = (regnode*)d->data[i];
17393 /* Trie stclasses are readonly and can thus be shared
17394 * without duplication. We free the stclass in pregfree
17395 * when the corresponding reg_ac_data struct is freed.
17397 reti->regstclass= ri->regstclass;
17401 ((reg_trie_data*)ri->data->data[i])->refcount++;
17406 d->data[i] = ri->data->data[i];
17409 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
17410 ri->data->what[i]);
17419 reti->name_list_idx = ri->name_list_idx;
17421 #ifdef RE_TRACK_PATTERN_OFFSETS
17422 if (ri->u.offsets) {
17423 Newx(reti->u.offsets, 2*len+1, U32);
17424 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
17427 SetProgLen(reti,len);
17430 return (void*)reti;
17433 #endif /* USE_ITHREADS */
17435 #ifndef PERL_IN_XSUB_RE
17438 - regnext - dig the "next" pointer out of a node
17441 Perl_regnext(pTHX_ regnode *p)
17448 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
17449 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
17450 (int)OP(p), (int)REGNODE_MAX);
17453 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
17462 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
17465 STRLEN l1 = strlen(pat1);
17466 STRLEN l2 = strlen(pat2);
17469 const char *message;
17471 PERL_ARGS_ASSERT_RE_CROAK2;
17477 Copy(pat1, buf, l1 , char);
17478 Copy(pat2, buf + l1, l2 , char);
17479 buf[l1 + l2] = '\n';
17480 buf[l1 + l2 + 1] = '\0';
17481 va_start(args, pat2);
17482 msv = vmess(buf, &args);
17484 message = SvPV_const(msv,l1);
17487 Copy(message, buf, l1 , char);
17488 /* l1-1 to avoid \n */
17489 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
17495 S_put_code_point(pTHX_ SV *sv, UV c)
17497 PERL_ARGS_ASSERT_PUT_CODE_POINT;
17500 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
17502 else if (isPRINT(c)) {
17503 const char string = (char) c;
17504 if (isBACKSLASHED_PUNCT(c))
17505 sv_catpvs(sv, "\\");
17506 sv_catpvn(sv, &string, 1);
17509 const char * const mnemonic = cntrl_to_mnemonic((char) c);
17511 Perl_sv_catpvf(aTHX_ sv, "%s", mnemonic);
17514 Perl_sv_catpvf(aTHX_ sv, "\\x{%02X}", (U8) c);
17519 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
17522 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
17524 /* Appends to 'sv' a displayable version of the range of code points from
17525 * 'start' to 'end'. It assumes that only ASCII printables are displayable
17526 * as-is (though some of these will be escaped by put_code_point()). */
17528 const unsigned int min_range_count = 3;
17530 assert(start <= end);
17532 PERL_ARGS_ASSERT_PUT_RANGE;
17534 while (start <= end) {
17536 const char * format;
17538 if (end - start < min_range_count) {
17540 /* Individual chars in short ranges */
17541 for (; start <= end; start++) {
17542 put_code_point(sv, start);
17547 /* If permitted by the input options, and there is a possibility that
17548 * this range contains a printable literal, look to see if there is
17550 if (allow_literals && start <= MAX_PRINT_A) {
17552 /* If the range begin isn't an ASCII printable, effectively split
17553 * the range into two parts:
17554 * 1) the portion before the first such printable,
17556 * and output them separately. */
17557 if (! isPRINT_A(start)) {
17558 UV temp_end = start + 1;
17560 /* There is no point looking beyond the final possible
17561 * printable, in MAX_PRINT_A */
17562 UV max = MIN(end, MAX_PRINT_A);
17564 while (temp_end <= max && ! isPRINT_A(temp_end)) {
17568 /* Here, temp_end points to one beyond the first printable if
17569 * found, or to one beyond 'max' if not. If none found, make
17570 * sure that we use the entire range */
17571 if (temp_end > MAX_PRINT_A) {
17572 temp_end = end + 1;
17575 /* Output the first part of the split range, the part that
17576 * doesn't have printables, with no looking for literals
17577 * (otherwise we would infinitely recurse) */
17578 put_range(sv, start, temp_end - 1, FALSE);
17580 /* The 2nd part of the range (if any) starts here. */
17583 /* We continue instead of dropping down because even if the 2nd
17584 * part is non-empty, it could be so short that we want to
17585 * output it specially, as tested for at the top of this loop.
17590 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
17591 * output a sub-range of just the digits or letters, then process
17592 * the remaining portion as usual. */
17593 if (isALPHANUMERIC_A(start)) {
17594 UV mask = (isDIGIT_A(start))
17599 UV temp_end = start + 1;
17601 /* Find the end of the sub-range that includes just the
17602 * characters in the same class as the first character in it */
17603 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
17608 /* For short ranges, don't duplicate the code above to output
17609 * them; just call recursively */
17610 if (temp_end - start < min_range_count) {
17611 put_range(sv, start, temp_end, FALSE);
17613 else { /* Output as a range */
17614 put_code_point(sv, start);
17615 sv_catpvs(sv, "-");
17616 put_code_point(sv, temp_end);
17618 start = temp_end + 1;
17622 /* We output any other printables as individual characters */
17623 if (isPUNCT_A(start) || isSPACE_A(start)) {
17624 while (start <= end && (isPUNCT_A(start)
17625 || isSPACE_A(start)))
17627 put_code_point(sv, start);
17632 } /* End of looking for literals */
17634 /* Here is not to output as a literal. Some control characters have
17635 * mnemonic names. Split off any of those at the beginning and end of
17636 * the range to print mnemonically. It isn't possible for many of
17637 * these to be in a row, so this won't overwhelm with output */
17638 while (isMNEMONIC_CNTRL(start) && start <= end) {
17639 put_code_point(sv, start);
17642 if (start < end && isMNEMONIC_CNTRL(end)) {
17644 /* Here, the final character in the range has a mnemonic name.
17645 * Work backwards from the end to find the final non-mnemonic */
17646 UV temp_end = end - 1;
17647 while (isMNEMONIC_CNTRL(temp_end)) {
17651 /* And separately output the range that doesn't have mnemonics */
17652 put_range(sv, start, temp_end, FALSE);
17654 /* Then output the mnemonic trailing controls */
17655 start = temp_end + 1;
17656 while (start <= end) {
17657 put_code_point(sv, start);
17663 /* As a final resort, output the range or subrange as hex. */
17665 this_end = (end < NUM_ANYOF_CODE_POINTS)
17667 : NUM_ANYOF_CODE_POINTS - 1;
17668 format = (this_end < 256)
17669 ? "\\x{%02"UVXf"}-\\x{%02"UVXf"}"
17670 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
17671 GCC_DIAG_IGNORE(-Wformat-nonliteral);
17672 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
17679 S_put_charclass_bitmap_innards(pTHX_ SV *sv, char *bitmap, SV** bitmap_invlist)
17681 /* Appends to 'sv' a displayable version of the innards of the bracketed
17682 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
17683 * output anything, and bitmap_invlist, if not NULL, will point to an
17684 * inversion list of what is in the bit map */
17688 unsigned int punct_count = 0;
17689 SV* invlist = NULL;
17690 SV** invlist_ptr; /* Temporary, in case bitmap_invlist is NULL */
17691 bool allow_literals = TRUE;
17693 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
17695 invlist_ptr = (bitmap_invlist) ? bitmap_invlist : &invlist;
17697 /* Worst case is exactly every-other code point is in the list */
17698 *invlist_ptr = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
17700 /* Convert the bit map to an inversion list, keeping track of how many
17701 * ASCII puncts are set, including an extra amount for the backslashed
17703 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
17704 if (BITMAP_TEST(bitmap, i)) {
17705 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, i);
17706 if (isPUNCT_A(i)) {
17708 if isBACKSLASHED_PUNCT(i) {
17715 /* Nothing to output */
17716 if (_invlist_len(*invlist_ptr) == 0) {
17717 SvREFCNT_dec(invlist);
17721 /* Generally, it is more readable if printable characters are output as
17722 * literals, but if a range (nearly) spans all of them, it's best to output
17723 * it as a single range. This code will use a single range if all but 2
17724 * printables are in it */
17725 invlist_iterinit(*invlist_ptr);
17726 while (invlist_iternext(*invlist_ptr, &start, &end)) {
17728 /* If range starts beyond final printable, it doesn't have any in it */
17729 if (start > MAX_PRINT_A) {
17733 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
17734 * all but two, the range must start and end no later than 2 from
17736 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
17737 if (end > MAX_PRINT_A) {
17743 if (end - start >= MAX_PRINT_A - ' ' - 2) {
17744 allow_literals = FALSE;
17749 invlist_iterfinish(*invlist_ptr);
17751 /* The legibility of the output depends mostly on how many punctuation
17752 * characters are output. There are 32 possible ASCII ones, and some have
17753 * an additional backslash, bringing it to currently 36, so if any more
17754 * than 18 are to be output, we can instead output it as its complement,
17755 * yielding fewer puncts, and making it more legible. But give some weight
17756 * to the fact that outputting it as a complement is less legible than a
17757 * straight output, so don't complement unless we are somewhat over the 18
17759 if (allow_literals && punct_count > 22) {
17760 sv_catpvs(sv, "^");
17762 /* Add everything remaining to the list, so when we invert it just
17763 * below, it will be excluded */
17764 _invlist_union_complement_2nd(*invlist_ptr, PL_InBitmap, invlist_ptr);
17765 _invlist_invert(*invlist_ptr);
17768 /* Here we have figured things out. Output each range */
17769 invlist_iterinit(*invlist_ptr);
17770 while (invlist_iternext(*invlist_ptr, &start, &end)) {
17771 if (start >= NUM_ANYOF_CODE_POINTS) {
17774 put_range(sv, start, end, allow_literals);
17776 invlist_iterfinish(*invlist_ptr);
17781 #define CLEAR_OPTSTART \
17782 if (optstart) STMT_START { \
17783 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
17784 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
17788 #define DUMPUNTIL(b,e) \
17790 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
17792 STATIC const regnode *
17793 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
17794 const regnode *last, const regnode *plast,
17795 SV* sv, I32 indent, U32 depth)
17797 U8 op = PSEUDO; /* Arbitrary non-END op. */
17798 const regnode *next;
17799 const regnode *optstart= NULL;
17801 RXi_GET_DECL(r,ri);
17802 GET_RE_DEBUG_FLAGS_DECL;
17804 PERL_ARGS_ASSERT_DUMPUNTIL;
17806 #ifdef DEBUG_DUMPUNTIL
17807 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
17808 last ? last-start : 0,plast ? plast-start : 0);
17811 if (plast && plast < last)
17814 while (PL_regkind[op] != END && (!last || node < last)) {
17816 /* While that wasn't END last time... */
17819 if (op == CLOSE || op == WHILEM)
17821 next = regnext((regnode *)node);
17824 if (OP(node) == OPTIMIZED) {
17825 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
17832 regprop(r, sv, node, NULL, NULL);
17833 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
17834 (int)(2*indent + 1), "", SvPVX_const(sv));
17836 if (OP(node) != OPTIMIZED) {
17837 if (next == NULL) /* Next ptr. */
17838 PerlIO_printf(Perl_debug_log, " (0)");
17839 else if (PL_regkind[(U8)op] == BRANCH
17840 && PL_regkind[OP(next)] != BRANCH )
17841 PerlIO_printf(Perl_debug_log, " (FAIL)");
17843 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
17844 (void)PerlIO_putc(Perl_debug_log, '\n');
17848 if (PL_regkind[(U8)op] == BRANCHJ) {
17851 const regnode *nnode = (OP(next) == LONGJMP
17852 ? regnext((regnode *)next)
17854 if (last && nnode > last)
17856 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
17859 else if (PL_regkind[(U8)op] == BRANCH) {
17861 DUMPUNTIL(NEXTOPER(node), next);
17863 else if ( PL_regkind[(U8)op] == TRIE ) {
17864 const regnode *this_trie = node;
17865 const char op = OP(node);
17866 const U32 n = ARG(node);
17867 const reg_ac_data * const ac = op>=AHOCORASICK ?
17868 (reg_ac_data *)ri->data->data[n] :
17870 const reg_trie_data * const trie =
17871 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
17873 AV *const trie_words
17874 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
17876 const regnode *nextbranch= NULL;
17879 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
17880 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
17882 PerlIO_printf(Perl_debug_log, "%*s%s ",
17883 (int)(2*(indent+3)), "",
17885 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
17886 SvCUR(*elem_ptr), 60,
17887 PL_colors[0], PL_colors[1],
17889 ? PERL_PV_ESCAPE_UNI
17891 | PERL_PV_PRETTY_ELLIPSES
17892 | PERL_PV_PRETTY_LTGT
17897 U16 dist= trie->jump[word_idx+1];
17898 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
17899 (UV)((dist ? this_trie + dist : next) - start));
17902 nextbranch= this_trie + trie->jump[0];
17903 DUMPUNTIL(this_trie + dist, nextbranch);
17905 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
17906 nextbranch= regnext((regnode *)nextbranch);
17908 PerlIO_printf(Perl_debug_log, "\n");
17911 if (last && next > last)
17916 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
17917 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
17918 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
17920 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
17922 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
17924 else if ( op == PLUS || op == STAR) {
17925 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
17927 else if (PL_regkind[(U8)op] == ANYOF) {
17928 /* arglen 1 + class block */
17929 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
17930 ? ANYOF_POSIXL_SKIP
17932 node = NEXTOPER(node);
17934 else if (PL_regkind[(U8)op] == EXACT) {
17935 /* Literal string, where present. */
17936 node += NODE_SZ_STR(node) - 1;
17937 node = NEXTOPER(node);
17940 node = NEXTOPER(node);
17941 node += regarglen[(U8)op];
17943 if (op == CURLYX || op == OPEN)
17947 #ifdef DEBUG_DUMPUNTIL
17948 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
17953 #endif /* DEBUGGING */
17957 * c-indentation-style: bsd
17958 * c-basic-offset: 4
17959 * indent-tabs-mode: nil
17962 * ex: set ts=8 sts=4 sw=4 et: