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 "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
95 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
96 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
97 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
100 #define STATIC static
104 struct RExC_state_t {
105 U32 flags; /* RXf_* are we folding, multilining? */
106 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
107 char *precomp; /* uncompiled string. */
108 REGEXP *rx_sv; /* The SV that is the regexp. */
109 regexp *rx; /* perl core regexp structure */
110 regexp_internal *rxi; /* internal data for regexp object
112 char *start; /* Start of input for compile */
113 char *end; /* End of input for compile */
114 char *parse; /* Input-scan pointer. */
115 SSize_t whilem_seen; /* number of WHILEM in this expr */
116 regnode *emit_start; /* Start of emitted-code area */
117 regnode *emit_bound; /* First regnode outside of the
119 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
120 implies compiling, so don't emit */
121 regnode_ssc emit_dummy; /* placeholder for emit to point to;
122 large enough for the largest
123 non-EXACTish node, so can use it as
125 I32 naughty; /* How bad is this pattern? */
126 I32 sawback; /* Did we see \1, ...? */
128 SSize_t size; /* Code size. */
129 I32 npar; /* Capture buffer count, (OPEN) plus
130 one. ("par" 0 is the whole
132 I32 nestroot; /* root parens we are in - used by
136 regnode **open_parens; /* pointers to open parens */
137 regnode **close_parens; /* pointers to close parens */
138 regnode *opend; /* END node in program */
139 I32 utf8; /* whether the pattern is utf8 or not */
140 I32 orig_utf8; /* whether the pattern was originally in utf8 */
141 /* XXX use this for future optimisation of case
142 * where pattern must be upgraded to utf8. */
143 I32 uni_semantics; /* If a d charset modifier should use unicode
144 rules, even if the pattern is not in
146 HV *paren_names; /* Paren names */
148 regnode **recurse; /* Recurse regops */
149 I32 recurse_count; /* Number of recurse regops */
150 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
152 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
156 I32 override_recoding;
157 I32 in_multi_char_class;
158 struct reg_code_block *code_blocks; /* positions of literal (?{})
160 int num_code_blocks; /* size of code_blocks[] */
161 int code_index; /* next code_blocks[] slot */
162 SSize_t maxlen; /* mininum possible number of chars in string to match */
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
192 #define RExC_emit (pRExC_state->emit)
193 #define RExC_emit_dummy (pRExC_state->emit_dummy)
194 #define RExC_emit_start (pRExC_state->emit_start)
195 #define RExC_emit_bound (pRExC_state->emit_bound)
196 #define RExC_naughty (pRExC_state->naughty)
197 #define RExC_sawback (pRExC_state->sawback)
198 #define RExC_seen (pRExC_state->seen)
199 #define RExC_size (pRExC_state->size)
200 #define RExC_maxlen (pRExC_state->maxlen)
201 #define RExC_npar (pRExC_state->npar)
202 #define RExC_nestroot (pRExC_state->nestroot)
203 #define RExC_extralen (pRExC_state->extralen)
204 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
205 #define RExC_utf8 (pRExC_state->utf8)
206 #define RExC_uni_semantics (pRExC_state->uni_semantics)
207 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
208 #define RExC_open_parens (pRExC_state->open_parens)
209 #define RExC_close_parens (pRExC_state->close_parens)
210 #define RExC_opend (pRExC_state->opend)
211 #define RExC_paren_names (pRExC_state->paren_names)
212 #define RExC_recurse (pRExC_state->recurse)
213 #define RExC_recurse_count (pRExC_state->recurse_count)
214 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
215 #define RExC_study_chunk_recursed_bytes \
216 (pRExC_state->study_chunk_recursed_bytes)
217 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
218 #define RExC_contains_locale (pRExC_state->contains_locale)
219 #define RExC_contains_i (pRExC_state->contains_i)
220 #define RExC_override_recoding (pRExC_state->override_recoding)
221 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
224 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
225 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
226 ((*s) == '{' && regcurly(s, FALSE)))
229 * Flags to be passed up and down.
231 #define WORST 0 /* Worst case. */
232 #define HASWIDTH 0x01 /* Known to match non-null strings. */
234 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
235 * character. (There needs to be a case: in the switch statement in regexec.c
236 * for any node marked SIMPLE.) Note that this is not the same thing as
239 #define SPSTART 0x04 /* Starts with * or + */
240 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
241 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
242 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
244 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
246 /* whether trie related optimizations are enabled */
247 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
248 #define TRIE_STUDY_OPT
249 #define FULL_TRIE_STUDY
255 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
256 #define PBITVAL(paren) (1 << ((paren) & 7))
257 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
258 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
259 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
261 #define REQUIRE_UTF8 STMT_START { \
263 *flagp = RESTART_UTF8; \
268 /* This converts the named class defined in regcomp.h to its equivalent class
269 * number defined in handy.h. */
270 #define namedclass_to_classnum(class) ((int) ((class) / 2))
271 #define classnum_to_namedclass(classnum) ((classnum) * 2)
273 #define _invlist_union_complement_2nd(a, b, output) \
274 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
275 #define _invlist_intersection_complement_2nd(a, b, output) \
276 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
278 /* About scan_data_t.
280 During optimisation we recurse through the regexp program performing
281 various inplace (keyhole style) optimisations. In addition study_chunk
282 and scan_commit populate this data structure with information about
283 what strings MUST appear in the pattern. We look for the longest
284 string that must appear at a fixed location, and we look for the
285 longest string that may appear at a floating location. So for instance
290 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
291 strings (because they follow a .* construct). study_chunk will identify
292 both FOO and BAR as being the longest fixed and floating strings respectively.
294 The strings can be composites, for instance
298 will result in a composite fixed substring 'foo'.
300 For each string some basic information is maintained:
302 - offset or min_offset
303 This is the position the string must appear at, or not before.
304 It also implicitly (when combined with minlenp) tells us how many
305 characters must match before the string we are searching for.
306 Likewise when combined with minlenp and the length of the string it
307 tells us how many characters must appear after the string we have
311 Only used for floating strings. This is the rightmost point that
312 the string can appear at. If set to SSize_t_MAX it indicates that the
313 string can occur infinitely far to the right.
316 A pointer to the minimum number of characters of the pattern that the
317 string was found inside. This is important as in the case of positive
318 lookahead or positive lookbehind we can have multiple patterns
323 The minimum length of the pattern overall is 3, the minimum length
324 of the lookahead part is 3, but the minimum length of the part that
325 will actually match is 1. So 'FOO's minimum length is 3, but the
326 minimum length for the F is 1. This is important as the minimum length
327 is used to determine offsets in front of and behind the string being
328 looked for. Since strings can be composites this is the length of the
329 pattern at the time it was committed with a scan_commit. Note that
330 the length is calculated by study_chunk, so that the minimum lengths
331 are not known until the full pattern has been compiled, thus the
332 pointer to the value.
336 In the case of lookbehind the string being searched for can be
337 offset past the start point of the final matching string.
338 If this value was just blithely removed from the min_offset it would
339 invalidate some of the calculations for how many chars must match
340 before or after (as they are derived from min_offset and minlen and
341 the length of the string being searched for).
342 When the final pattern is compiled and the data is moved from the
343 scan_data_t structure into the regexp structure the information
344 about lookbehind is factored in, with the information that would
345 have been lost precalculated in the end_shift field for the
348 The fields pos_min and pos_delta are used to store the minimum offset
349 and the delta to the maximum offset at the current point in the pattern.
353 typedef struct scan_data_t {
354 /*I32 len_min; unused */
355 /*I32 len_delta; unused */
359 SSize_t last_end; /* min value, <0 unless valid. */
360 SSize_t last_start_min;
361 SSize_t last_start_max;
362 SV **longest; /* Either &l_fixed, or &l_float. */
363 SV *longest_fixed; /* longest fixed string found in pattern */
364 SSize_t offset_fixed; /* offset where it starts */
365 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
366 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
367 SV *longest_float; /* longest floating string found in pattern */
368 SSize_t offset_float_min; /* earliest point in string it can appear */
369 SSize_t offset_float_max; /* latest point in string it can appear */
370 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
371 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
374 SSize_t *last_closep;
375 regnode_ssc *start_class;
378 /* The below is perhaps overboard, but this allows us to save a test at the
379 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
380 * and 'a' differ by a single bit; the same with the upper and lower case of
381 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
382 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
383 * then inverts it to form a mask, with just a single 0, in the bit position
384 * where the upper- and lowercase differ. XXX There are about 40 other
385 * instances in the Perl core where this micro-optimization could be used.
386 * Should decide if maintenance cost is worse, before changing those
388 * Returns a boolean as to whether or not 'v' is either a lowercase or
389 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
390 * compile-time constant, the generated code is better than some optimizing
391 * compilers figure out, amounting to a mask and test. The results are
392 * meaningless if 'c' is not one of [A-Za-z] */
393 #define isARG2_lower_or_UPPER_ARG1(c, v) \
394 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
397 * Forward declarations for pregcomp()'s friends.
400 static const scan_data_t zero_scan_data =
401 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
403 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
404 #define SF_BEFORE_SEOL 0x0001
405 #define SF_BEFORE_MEOL 0x0002
406 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
407 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
409 #define SF_FIX_SHIFT_EOL (+2)
410 #define SF_FL_SHIFT_EOL (+4)
412 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
413 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
415 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
416 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
417 #define SF_IS_INF 0x0040
418 #define SF_HAS_PAR 0x0080
419 #define SF_IN_PAR 0x0100
420 #define SF_HAS_EVAL 0x0200
421 #define SCF_DO_SUBSTR 0x0400
422 #define SCF_DO_STCLASS_AND 0x0800
423 #define SCF_DO_STCLASS_OR 0x1000
424 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
425 #define SCF_WHILEM_VISITED_POS 0x2000
427 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
428 #define SCF_SEEN_ACCEPT 0x8000
429 #define SCF_TRIE_DOING_RESTUDY 0x10000
431 #define UTF cBOOL(RExC_utf8)
433 /* The enums for all these are ordered so things work out correctly */
434 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
435 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
436 == REGEX_DEPENDS_CHARSET)
437 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
438 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
439 >= REGEX_UNICODE_CHARSET)
440 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
441 == REGEX_ASCII_RESTRICTED_CHARSET)
442 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
443 >= REGEX_ASCII_RESTRICTED_CHARSET)
444 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
445 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
447 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
449 /* For programs that want to be strictly Unicode compatible by dying if any
450 * attempt is made to match a non-Unicode code point against a Unicode
452 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
454 #define OOB_NAMEDCLASS -1
456 /* There is no code point that is out-of-bounds, so this is problematic. But
457 * its only current use is to initialize a variable that is always set before
459 #define OOB_UNICODE 0xDEADBEEF
461 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
462 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
465 /* length of regex to show in messages that don't mark a position within */
466 #define RegexLengthToShowInErrorMessages 127
469 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
470 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
471 * op/pragma/warn/regcomp.
473 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
474 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
476 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
477 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
479 #define REPORT_LOCATION_ARGS(offset) \
480 UTF8fARG(UTF, offset, RExC_precomp), \
481 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
484 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
485 * arg. Show regex, up to a maximum length. If it's too long, chop and add
488 #define _FAIL(code) STMT_START { \
489 const char *ellipses = ""; \
490 IV len = RExC_end - RExC_precomp; \
493 SAVEFREESV(RExC_rx_sv); \
494 if (len > RegexLengthToShowInErrorMessages) { \
495 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
496 len = RegexLengthToShowInErrorMessages - 10; \
502 #define FAIL(msg) _FAIL( \
503 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
504 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
506 #define FAIL2(msg,arg) _FAIL( \
507 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
508 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
511 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
513 #define Simple_vFAIL(m) STMT_START { \
514 const IV offset = RExC_parse - RExC_precomp; \
515 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
516 m, REPORT_LOCATION_ARGS(offset)); \
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
522 #define vFAIL(m) STMT_START { \
524 SAVEFREESV(RExC_rx_sv); \
529 * Like Simple_vFAIL(), but accepts two arguments.
531 #define Simple_vFAIL2(m,a1) STMT_START { \
532 const IV offset = RExC_parse - RExC_precomp; \
533 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
534 REPORT_LOCATION_ARGS(offset)); \
538 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
540 #define vFAIL2(m,a1) STMT_START { \
542 SAVEFREESV(RExC_rx_sv); \
543 Simple_vFAIL2(m, a1); \
548 * Like Simple_vFAIL(), but accepts three arguments.
550 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
551 const IV offset = RExC_parse - RExC_precomp; \
552 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
553 REPORT_LOCATION_ARGS(offset)); \
557 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
559 #define vFAIL3(m,a1,a2) STMT_START { \
561 SAVEFREESV(RExC_rx_sv); \
562 Simple_vFAIL3(m, a1, a2); \
566 * Like Simple_vFAIL(), but accepts four arguments.
568 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
569 const IV offset = RExC_parse - RExC_precomp; \
570 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
571 REPORT_LOCATION_ARGS(offset)); \
574 #define vFAIL4(m,a1,a2,a3) STMT_START { \
576 SAVEFREESV(RExC_rx_sv); \
577 Simple_vFAIL4(m, a1, a2, a3); \
580 /* A specialized version of vFAIL2 that works with UTF8f */
581 #define vFAIL2utf8f(m, a1) STMT_START { \
582 const IV offset = RExC_parse - RExC_precomp; \
584 SAVEFREESV(RExC_rx_sv); \
585 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
586 REPORT_LOCATION_ARGS(offset)); \
590 /* m is not necessarily a "literal string", in this macro */
591 #define reg_warn_non_literal_string(loc, m) STMT_START { \
592 const IV offset = loc - RExC_precomp; \
593 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
594 m, REPORT_LOCATION_ARGS(offset)); \
597 #define ckWARNreg(loc,m) STMT_START { \
598 const IV offset = loc - RExC_precomp; \
599 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
600 REPORT_LOCATION_ARGS(offset)); \
603 #define vWARN_dep(loc, m) STMT_START { \
604 const IV offset = loc - RExC_precomp; \
605 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
606 REPORT_LOCATION_ARGS(offset)); \
609 #define ckWARNdep(loc,m) STMT_START { \
610 const IV offset = loc - RExC_precomp; \
611 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
613 REPORT_LOCATION_ARGS(offset)); \
616 #define ckWARNregdep(loc,m) STMT_START { \
617 const IV offset = loc - RExC_precomp; \
618 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
620 REPORT_LOCATION_ARGS(offset)); \
623 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
624 const IV offset = loc - RExC_precomp; \
625 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
627 a1, REPORT_LOCATION_ARGS(offset)); \
630 #define ckWARN2reg(loc, m, a1) STMT_START { \
631 const IV offset = loc - RExC_precomp; \
632 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
633 a1, REPORT_LOCATION_ARGS(offset)); \
636 #define vWARN3(loc, m, a1, a2) STMT_START { \
637 const IV offset = loc - RExC_precomp; \
638 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
639 a1, a2, REPORT_LOCATION_ARGS(offset)); \
642 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
643 const IV offset = loc - RExC_precomp; \
644 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
645 a1, a2, REPORT_LOCATION_ARGS(offset)); \
648 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
649 const IV offset = loc - RExC_precomp; \
650 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
651 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
654 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
655 const IV offset = loc - RExC_precomp; \
656 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
657 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
660 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
661 const IV offset = loc - RExC_precomp; \
662 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
663 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
667 /* Allow for side effects in s */
668 #define REGC(c,s) STMT_START { \
669 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
672 /* Macros for recording node offsets. 20001227 mjd@plover.com
673 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
674 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
675 * Element 0 holds the number n.
676 * Position is 1 indexed.
678 #ifndef RE_TRACK_PATTERN_OFFSETS
679 #define Set_Node_Offset_To_R(node,byte)
680 #define Set_Node_Offset(node,byte)
681 #define Set_Cur_Node_Offset
682 #define Set_Node_Length_To_R(node,len)
683 #define Set_Node_Length(node,len)
684 #define Set_Node_Cur_Length(node,start)
685 #define Node_Offset(n)
686 #define Node_Length(n)
687 #define Set_Node_Offset_Length(node,offset,len)
688 #define ProgLen(ri) ri->u.proglen
689 #define SetProgLen(ri,x) ri->u.proglen = x
691 #define ProgLen(ri) ri->u.offsets[0]
692 #define SetProgLen(ri,x) ri->u.offsets[0] = x
693 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
695 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
696 __LINE__, (int)(node), (int)(byte))); \
698 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
701 RExC_offsets[2*(node)-1] = (byte); \
706 #define Set_Node_Offset(node,byte) \
707 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
708 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
710 #define Set_Node_Length_To_R(node,len) STMT_START { \
712 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
713 __LINE__, (int)(node), (int)(len))); \
715 Perl_croak(aTHX_ "value of node is %d in Length macro", \
718 RExC_offsets[2*(node)] = (len); \
723 #define Set_Node_Length(node,len) \
724 Set_Node_Length_To_R((node)-RExC_emit_start, len)
725 #define Set_Node_Cur_Length(node, start) \
726 Set_Node_Length(node, RExC_parse - start)
728 /* Get offsets and lengths */
729 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
730 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
732 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
733 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
734 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
738 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
739 #define EXPERIMENTAL_INPLACESCAN
740 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
742 #define DEBUG_RExC_seen() \
743 DEBUG_OPTIMISE_MORE_r({ \
744 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
746 if (RExC_seen & REG_ZERO_LEN_SEEN) \
747 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
749 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
750 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
752 if (RExC_seen & REG_GPOS_SEEN) \
753 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
755 if (RExC_seen & REG_CANY_SEEN) \
756 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
758 if (RExC_seen & REG_RECURSE_SEEN) \
759 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
761 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
762 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
764 if (RExC_seen & REG_VERBARG_SEEN) \
765 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
767 if (RExC_seen & REG_CUTGROUP_SEEN) \
768 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
770 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
771 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
773 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
774 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
776 if (RExC_seen & REG_GOSTART_SEEN) \
777 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
779 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
780 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
782 PerlIO_printf(Perl_debug_log,"\n"); \
785 #define DEBUG_STUDYDATA(str,data,depth) \
786 DEBUG_OPTIMISE_MORE_r(if(data){ \
787 PerlIO_printf(Perl_debug_log, \
788 "%*s" str "Pos:%"IVdf"/%"IVdf \
789 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
790 (int)(depth)*2, "", \
791 (IV)((data)->pos_min), \
792 (IV)((data)->pos_delta), \
793 (UV)((data)->flags), \
794 (IV)((data)->whilem_c), \
795 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
796 is_inf ? "INF " : "" \
798 if ((data)->last_found) \
799 PerlIO_printf(Perl_debug_log, \
800 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
801 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
802 SvPVX_const((data)->last_found), \
803 (IV)((data)->last_end), \
804 (IV)((data)->last_start_min), \
805 (IV)((data)->last_start_max), \
806 ((data)->longest && \
807 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
808 SvPVX_const((data)->longest_fixed), \
809 (IV)((data)->offset_fixed), \
810 ((data)->longest && \
811 (data)->longest==&((data)->longest_float)) ? "*" : "", \
812 SvPVX_const((data)->longest_float), \
813 (IV)((data)->offset_float_min), \
814 (IV)((data)->offset_float_max) \
816 PerlIO_printf(Perl_debug_log,"\n"); \
819 /* Mark that we cannot extend a found fixed substring at this point.
820 Update the longest found anchored substring and the longest found
821 floating substrings if needed. */
824 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
825 SSize_t *minlenp, int is_inf)
827 const STRLEN l = CHR_SVLEN(data->last_found);
828 const STRLEN old_l = CHR_SVLEN(*data->longest);
829 GET_RE_DEBUG_FLAGS_DECL;
831 PERL_ARGS_ASSERT_SCAN_COMMIT;
833 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
834 SvSetMagicSV(*data->longest, data->last_found);
835 if (*data->longest == data->longest_fixed) {
836 data->offset_fixed = l ? data->last_start_min : data->pos_min;
837 if (data->flags & SF_BEFORE_EOL)
839 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
841 data->flags &= ~SF_FIX_BEFORE_EOL;
842 data->minlen_fixed=minlenp;
843 data->lookbehind_fixed=0;
845 else { /* *data->longest == data->longest_float */
846 data->offset_float_min = l ? data->last_start_min : data->pos_min;
847 data->offset_float_max = (l
848 ? data->last_start_max
849 : (data->pos_delta == SSize_t_MAX
851 : data->pos_min + data->pos_delta));
853 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
854 data->offset_float_max = SSize_t_MAX;
855 if (data->flags & SF_BEFORE_EOL)
857 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
859 data->flags &= ~SF_FL_BEFORE_EOL;
860 data->minlen_float=minlenp;
861 data->lookbehind_float=0;
864 SvCUR_set(data->last_found, 0);
866 SV * const sv = data->last_found;
867 if (SvUTF8(sv) && SvMAGICAL(sv)) {
868 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
874 data->flags &= ~SF_BEFORE_EOL;
875 DEBUG_STUDYDATA("commit: ",data,0);
878 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
879 * list that describes which code points it matches */
882 S_ssc_anything(pTHX_ regnode_ssc *ssc)
884 /* Set the SSC 'ssc' to match an empty string or any code point */
886 PERL_ARGS_ASSERT_SSC_ANYTHING;
888 assert(is_ANYOF_SYNTHETIC(ssc));
890 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
891 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
892 ANYOF_FLAGS(ssc) |= ANYOF_EMPTY_STRING; /* Plus match empty string */
896 S_ssc_is_anything(pTHX_ const regnode_ssc *ssc)
898 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
899 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
900 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
901 * in any way, so there's no point in using it */
906 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
908 assert(is_ANYOF_SYNTHETIC(ssc));
910 if (! (ANYOF_FLAGS(ssc) & ANYOF_EMPTY_STRING)) {
914 /* See if the list consists solely of the range 0 - Infinity */
915 invlist_iterinit(ssc->invlist);
916 ret = invlist_iternext(ssc->invlist, &start, &end)
920 invlist_iterfinish(ssc->invlist);
926 /* If e.g., both \w and \W are set, matches everything */
927 if (ANYOF_FLAGS(ssc) & ANYOF_POSIXL) {
929 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
930 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
940 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
942 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
943 * string, any code point, or any posix class under locale */
945 PERL_ARGS_ASSERT_SSC_INIT;
947 Zero(ssc, 1, regnode_ssc);
948 set_ANYOF_SYNTHETIC(ssc);
949 ARG_SET(ssc, ANYOF_NONBITMAP_EMPTY);
952 /* If any portion of the regex is to operate under locale rules,
953 * initialization includes it. The reason this isn't done for all regexes
954 * is that the optimizer was written under the assumption that locale was
955 * all-or-nothing. Given the complexity and lack of documentation in the
956 * optimizer, and that there are inadequate test cases for locale, many
957 * parts of it may not work properly, it is safest to avoid locale unless
959 if (RExC_contains_locale) {
960 ANYOF_POSIXL_SETALL(ssc);
961 ANYOF_FLAGS(ssc) |= ANYOF_LOCALE|ANYOF_POSIXL;
964 ANYOF_POSIXL_ZERO(ssc);
969 S_ssc_is_cp_posixl_init(pTHX_ const RExC_state_t *pRExC_state,
970 const regnode_ssc *ssc)
972 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
973 * to the list of code points matched, and locale posix classes; hence does
974 * not check its flags) */
979 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
981 assert(is_ANYOF_SYNTHETIC(ssc));
983 invlist_iterinit(ssc->invlist);
984 ret = invlist_iternext(ssc->invlist, &start, &end)
988 invlist_iterfinish(ssc->invlist);
994 if (RExC_contains_locale
995 && ! ((ANYOF_FLAGS(ssc) & ANYOF_LOCALE)
996 || ! (ANYOF_FLAGS(ssc) & ANYOF_POSIXL)
997 || ! ANYOF_POSIXL_TEST_ALL_SET(ssc)))
1006 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1007 const regnode_charclass_posixl_fold* const node)
1009 /* Returns a mortal inversion list defining which code points are matched
1010 * by 'node', which is of type ANYOF. Handles complementing the result if
1011 * appropriate. If some code points aren't knowable at this time, the
1012 * returned list must, and will, contain every code point that is a
1015 SV* invlist = sv_2mortal(_new_invlist(0));
1017 const U32 n = ARG(node);
1018 bool new_node_has_latin1 = FALSE;
1020 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1022 /* Look at the data structure created by S_set_ANYOF_arg() */
1023 if (n != ANYOF_NONBITMAP_EMPTY) {
1024 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1025 AV * const av = MUTABLE_AV(SvRV(rv));
1026 SV **const ary = AvARRAY(av);
1027 assert(RExC_rxi->data->what[n] == 's');
1029 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1030 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1032 else if (ary[0] && ary[0] != &PL_sv_undef) {
1034 /* Here, no compile-time swash, and there are things that won't be
1035 * known until runtime -- we have to assume it could be anything */
1036 return _add_range_to_invlist(invlist, 0, UV_MAX);
1040 /* Here no compile-time swash, and no run-time only data. Use the
1041 * node's inversion list */
1042 invlist = sv_2mortal(invlist_clone(ary[2]));
1046 /* An ANYOF node contains a bitmap for the first 256 code points, and an
1047 * inversion list for the others, but if there are code points that should
1048 * match only conditionally on the target string being UTF-8, those are
1049 * placed in the inversion list, and not the bitmap. Since there are
1050 * circumstances under which they could match, they are included in the
1051 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1052 * here, so that when we invert below, the end result actually does include
1053 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1054 * before we add the unconditionally matched code points */
1055 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1056 _invlist_intersection_complement_2nd(invlist,
1061 /* Add in the points from the bit map */
1062 for (i = 0; i < 256; i++) {
1063 if (ANYOF_BITMAP_TEST(node, i)) {
1064 invlist = add_cp_to_invlist(invlist, i);
1065 new_node_has_latin1 = TRUE;
1069 /* If this can match all upper Latin1 code points, have to add them
1071 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1072 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1075 /* Similarly for these */
1076 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1077 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1080 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1081 _invlist_invert(invlist);
1083 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1085 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1086 * locale. We can skip this if there are no 0-255 at all. */
1087 _invlist_union(invlist, PL_Latin1, &invlist);
1090 /* Similarly add the UTF-8 locale possible matches */
1091 if (ANYOF_FLAGS(node) & ANYOF_LOC_FOLD && ANYOF_UTF8_LOCALE_INVLIST(node))
1093 _invlist_union_maybe_complement_2nd(invlist,
1094 ANYOF_UTF8_LOCALE_INVLIST(node),
1095 ANYOF_FLAGS(node) & ANYOF_INVERT,
1102 /* These two functions currently do the exact same thing */
1103 #define ssc_init_zero ssc_init
1105 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1106 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1109 S_ssc_flags_and(regnode_ssc *ssc, const U8 and_with)
1111 /* Take the flags 'and_with' and accumulate them anded into the flags for
1112 * the SSC 'ssc'. The non-SSC related flags in 'and_with' are ignored.
1113 * The flags 'and_with' should not come from another SSC (otherwise the
1114 * EMPTY_STRING flag won't work) */
1116 const U8 ssc_only_flags = ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS;
1118 PERL_ARGS_ASSERT_SSC_FLAGS_AND;
1120 /* Use just the SSC-related flags from 'and_with' */
1121 ANYOF_FLAGS(ssc) &= (and_with & ANYOF_COMMON_FLAGS);
1122 ANYOF_FLAGS(ssc) |= ssc_only_flags;
1125 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1126 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1127 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1130 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1131 const regnode_ssc *and_with)
1133 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1134 * another SSC or a regular ANYOF class. Can create false positives. */
1139 PERL_ARGS_ASSERT_SSC_AND;
1141 assert(is_ANYOF_SYNTHETIC(ssc));
1143 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1144 * the code point inversion list and just the relevant flags */
1145 if (is_ANYOF_SYNTHETIC(and_with)) {
1146 anded_cp_list = and_with->invlist;
1147 anded_flags = ANYOF_FLAGS(and_with);
1149 /* XXX This is a kludge around what appears to be deficiencies in the
1150 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1151 * there are paths through the optimizer where it doesn't get weeded
1152 * out when it should. And if we don't make some extra provision for
1153 * it like the code just below, it doesn't get added when it should.
1154 * This solution is to add it only when AND'ing, which is here, and
1155 * only when what is being AND'ed is the pristine, original node
1156 * matching anything. Thus it is like adding it to ssc_anything() but
1157 * only when the result is to be AND'ed. Probably the same solution
1158 * could be adopted for the same problem we have with /l matching,
1159 * which is solved differently in S_ssc_init(), and that would lead to
1160 * fewer false positives than that solution has. But if this solution
1161 * creates bugs, the consequences are only that a warning isn't raised
1162 * that should be; while the consequences for having /l bugs is
1163 * incorrect matches */
1164 if (ssc_is_anything(and_with)) {
1165 anded_flags |= ANYOF_WARN_SUPER;
1169 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state,
1170 (regnode_charclass_posixl_fold*) and_with);
1171 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1174 ANYOF_FLAGS(ssc) &= anded_flags;
1176 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1177 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1178 * 'and_with' may be inverted. When not inverted, we have the situation of
1180 * (C1 | P1) & (C2 | P2)
1181 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1182 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1183 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1184 * <= ((C1 & C2) | P1 | P2)
1185 * Alternatively, the last few steps could be:
1186 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1187 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1188 * <= (C1 | C2 | (P1 & P2))
1189 * We favor the second approach if either P1 or P2 is non-empty. This is
1190 * because these components are a barrier to doing optimizations, as what
1191 * they match cannot be known until the moment of matching as they are
1192 * dependent on the current locale, 'AND"ing them likely will reduce or
1194 * But we can do better if we know that C1,P1 are in their initial state (a
1195 * frequent occurrence), each matching everything:
1196 * (<everything>) & (C2 | P2) = C2 | P2
1197 * Similarly, if C2,P2 are in their initial state (again a frequent
1198 * occurrence), the result is a no-op
1199 * (C1 | P1) & (<everything>) = C1 | P1
1202 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1203 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1204 * <= (C1 & ~C2) | (P1 & ~P2)
1207 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1208 && ! is_ANYOF_SYNTHETIC(and_with))
1212 ssc_intersection(ssc,
1214 FALSE /* Has already been inverted */
1217 /* If either P1 or P2 is empty, the intersection will be also; can skip
1219 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1220 ANYOF_POSIXL_ZERO(ssc);
1222 else if (ANYOF_POSIXL_TEST_ANY_SET(ssc)) {
1224 /* Note that the Posix class component P from 'and_with' actually
1226 * P = Pa | Pb | ... | Pn
1227 * where each component is one posix class, such as in [\w\s].
1229 * ~P = ~(Pa | Pb | ... | Pn)
1230 * = ~Pa & ~Pb & ... & ~Pn
1231 * <= ~Pa | ~Pb | ... | ~Pn
1232 * The last is something we can easily calculate, but unfortunately
1233 * is likely to have many false positives. We could do better
1234 * in some (but certainly not all) instances if two classes in
1235 * P have known relationships. For example
1236 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1238 * :lower: & :print: = :lower:
1239 * And similarly for classes that must be disjoint. For example,
1240 * since \s and \w can have no elements in common based on rules in
1241 * the POSIX standard,
1242 * \w & ^\S = nothing
1243 * Unfortunately, some vendor locales do not meet the Posix
1244 * standard, in particular almost everything by Microsoft.
1245 * The loop below just changes e.g., \w into \W and vice versa */
1247 regnode_charclass_posixl_fold temp;
1248 int add = 1; /* To calculate the index of the complement */
1250 ANYOF_POSIXL_ZERO(&temp);
1251 for (i = 0; i < ANYOF_MAX; i++) {
1253 || ! ANYOF_POSIXL_TEST(and_with, i)
1254 || ! ANYOF_POSIXL_TEST(and_with, i + 1));
1256 if (ANYOF_POSIXL_TEST(and_with, i)) {
1257 ANYOF_POSIXL_SET(&temp, i + add);
1259 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1261 ANYOF_POSIXL_AND(&temp, ssc);
1263 } /* else ssc already has no posixes */
1264 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1265 in its initial state */
1266 else if (! is_ANYOF_SYNTHETIC(and_with)
1267 || ! ssc_is_cp_posixl_init(pRExC_state, and_with))
1269 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1270 * copy it over 'ssc' */
1271 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1272 if (is_ANYOF_SYNTHETIC(and_with)) {
1273 StructCopy(and_with, ssc, regnode_ssc);
1276 ssc->invlist = anded_cp_list;
1277 ANYOF_POSIXL_ZERO(ssc);
1278 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1279 ANYOF_POSIXL_OR(and_with, ssc);
1283 else if ((ANYOF_FLAGS(ssc) & ANYOF_POSIXL)
1284 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1286 /* One or the other of P1, P2 is non-empty. */
1287 ANYOF_POSIXL_AND(and_with, ssc);
1288 ssc_union(ssc, anded_cp_list, FALSE);
1290 else { /* P1 = P2 = empty */
1291 ssc_intersection(ssc, anded_cp_list, FALSE);
1297 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1298 const regnode_ssc *or_with)
1300 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1301 * another SSC or a regular ANYOF class. Can create false positives if
1302 * 'or_with' is to be inverted. */
1307 PERL_ARGS_ASSERT_SSC_OR;
1309 assert(is_ANYOF_SYNTHETIC(ssc));
1311 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1312 * the code point inversion list and just the relevant flags */
1313 if (is_ANYOF_SYNTHETIC(or_with)) {
1314 ored_cp_list = or_with->invlist;
1315 ored_flags = ANYOF_FLAGS(or_with);
1318 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state,
1319 (regnode_charclass_posixl_fold*) or_with);
1320 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1323 ANYOF_FLAGS(ssc) |= ored_flags;
1325 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1326 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1327 * 'or_with' may be inverted. When not inverted, we have the simple
1328 * situation of computing:
1329 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1330 * If P1|P2 yields a situation with both a class and its complement are
1331 * set, like having both \w and \W, this matches all code points, and we
1332 * can delete these from the P component of the ssc going forward. XXX We
1333 * might be able to delete all the P components, but I (khw) am not certain
1334 * about this, and it is better to be safe.
1337 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1338 * <= (C1 | P1) | ~C2
1339 * <= (C1 | ~C2) | P1
1340 * (which results in actually simpler code than the non-inverted case)
1343 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1344 && ! is_ANYOF_SYNTHETIC(or_with))
1346 /* We ignore P2, leaving P1 going forward */
1348 else { /* Not inverted */
1349 ANYOF_POSIXL_OR(or_with, ssc);
1350 if (ANYOF_POSIXL_TEST_ANY_SET(ssc)) {
1352 for (i = 0; i < ANYOF_MAX; i += 2) {
1353 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1355 ssc_match_all_cp(ssc);
1356 ANYOF_POSIXL_CLEAR(ssc, i);
1357 ANYOF_POSIXL_CLEAR(ssc, i+1);
1358 if (! ANYOF_POSIXL_TEST_ANY_SET(ssc)) {
1359 ANYOF_FLAGS(ssc) &= ~ANYOF_POSIXL;
1368 FALSE /* Already has been inverted */
1372 PERL_STATIC_INLINE void
1373 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1375 PERL_ARGS_ASSERT_SSC_UNION;
1377 assert(is_ANYOF_SYNTHETIC(ssc));
1379 _invlist_union_maybe_complement_2nd(ssc->invlist,
1385 PERL_STATIC_INLINE void
1386 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1388 const bool invert2nd)
1390 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1392 assert(is_ANYOF_SYNTHETIC(ssc));
1394 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1400 PERL_STATIC_INLINE void
1401 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1403 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1405 assert(is_ANYOF_SYNTHETIC(ssc));
1407 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1410 PERL_STATIC_INLINE void
1411 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1413 /* AND just the single code point 'cp' into the SSC 'ssc' */
1415 SV* cp_list = _new_invlist(2);
1417 PERL_ARGS_ASSERT_SSC_CP_AND;
1419 assert(is_ANYOF_SYNTHETIC(ssc));
1421 cp_list = add_cp_to_invlist(cp_list, cp);
1422 ssc_intersection(ssc, cp_list,
1423 FALSE /* Not inverted */
1425 SvREFCNT_dec_NN(cp_list);
1428 PERL_STATIC_INLINE void
1429 S_ssc_clear_locale(pTHX_ regnode_ssc *ssc)
1431 /* Set the SSC 'ssc' to not match any locale things */
1433 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1435 assert(is_ANYOF_SYNTHETIC(ssc));
1437 ANYOF_POSIXL_ZERO(ssc);
1438 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1442 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1444 /* The inversion list in the SSC is marked mortal; now we need a more
1445 * permanent copy, which is stored the same way that is done in a regular
1446 * ANYOF node, with the first 256 code points in a bit map */
1448 SV* invlist = invlist_clone(ssc->invlist);
1450 PERL_ARGS_ASSERT_SSC_FINALIZE;
1452 assert(is_ANYOF_SYNTHETIC(ssc));
1454 /* The code in this file assumes that all but these flags aren't relevant
1455 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1456 * time we reach here */
1457 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1459 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1461 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist, NULL, NULL, FALSE);
1463 /* The code points that could match under /li are already incorporated into
1464 * the inversion list and bit map */
1465 ANYOF_FLAGS(ssc) &= ~ANYOF_LOC_FOLD;
1467 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE) || RExC_contains_locale);
1470 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1471 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1472 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1473 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1474 ? (TRIE_LIST_CUR( idx ) - 1) \
1480 dump_trie(trie,widecharmap,revcharmap)
1481 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1482 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1484 These routines dump out a trie in a somewhat readable format.
1485 The _interim_ variants are used for debugging the interim
1486 tables that are used to generate the final compressed
1487 representation which is what dump_trie expects.
1489 Part of the reason for their existence is to provide a form
1490 of documentation as to how the different representations function.
1495 Dumps the final compressed table form of the trie to Perl_debug_log.
1496 Used for debugging make_trie().
1500 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1501 AV *revcharmap, U32 depth)
1504 SV *sv=sv_newmortal();
1505 int colwidth= widecharmap ? 6 : 4;
1507 GET_RE_DEBUG_FLAGS_DECL;
1509 PERL_ARGS_ASSERT_DUMP_TRIE;
1511 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1512 (int)depth * 2 + 2,"",
1513 "Match","Base","Ofs" );
1515 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1516 SV ** const tmp = av_fetch( revcharmap, state, 0);
1518 PerlIO_printf( Perl_debug_log, "%*s",
1520 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1521 PL_colors[0], PL_colors[1],
1522 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1523 PERL_PV_ESCAPE_FIRSTCHAR
1528 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1529 (int)depth * 2 + 2,"");
1531 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1532 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1533 PerlIO_printf( Perl_debug_log, "\n");
1535 for( state = 1 ; state < trie->statecount ; state++ ) {
1536 const U32 base = trie->states[ state ].trans.base;
1538 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1539 (int)depth * 2 + 2,"", (UV)state);
1541 if ( trie->states[ state ].wordnum ) {
1542 PerlIO_printf( Perl_debug_log, " W%4X",
1543 trie->states[ state ].wordnum );
1545 PerlIO_printf( Perl_debug_log, "%6s", "" );
1548 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1553 while( ( base + ofs < trie->uniquecharcount ) ||
1554 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1555 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1559 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1561 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1562 if ( ( base + ofs >= trie->uniquecharcount )
1563 && ( base + ofs - trie->uniquecharcount
1565 && trie->trans[ base + ofs
1566 - trie->uniquecharcount ].check == state )
1568 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1570 (UV)trie->trans[ base + ofs
1571 - trie->uniquecharcount ].next );
1573 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1577 PerlIO_printf( Perl_debug_log, "]");
1580 PerlIO_printf( Perl_debug_log, "\n" );
1582 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1584 for (word=1; word <= trie->wordcount; word++) {
1585 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1586 (int)word, (int)(trie->wordinfo[word].prev),
1587 (int)(trie->wordinfo[word].len));
1589 PerlIO_printf(Perl_debug_log, "\n" );
1592 Dumps a fully constructed but uncompressed trie in list form.
1593 List tries normally only are used for construction when the number of
1594 possible chars (trie->uniquecharcount) is very high.
1595 Used for debugging make_trie().
1598 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1599 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1603 SV *sv=sv_newmortal();
1604 int colwidth= widecharmap ? 6 : 4;
1605 GET_RE_DEBUG_FLAGS_DECL;
1607 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1609 /* print out the table precompression. */
1610 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1611 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1612 "------:-----+-----------------\n" );
1614 for( state=1 ; state < next_alloc ; state ++ ) {
1617 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1618 (int)depth * 2 + 2,"", (UV)state );
1619 if ( ! trie->states[ state ].wordnum ) {
1620 PerlIO_printf( Perl_debug_log, "%5s| ","");
1622 PerlIO_printf( Perl_debug_log, "W%4x| ",
1623 trie->states[ state ].wordnum
1626 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1627 SV ** const tmp = av_fetch( revcharmap,
1628 TRIE_LIST_ITEM(state,charid).forid, 0);
1630 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1632 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1634 PL_colors[0], PL_colors[1],
1635 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1636 | PERL_PV_ESCAPE_FIRSTCHAR
1638 TRIE_LIST_ITEM(state,charid).forid,
1639 (UV)TRIE_LIST_ITEM(state,charid).newstate
1642 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1643 (int)((depth * 2) + 14), "");
1646 PerlIO_printf( Perl_debug_log, "\n");
1651 Dumps a fully constructed but uncompressed trie in table form.
1652 This is the normal DFA style state transition table, with a few
1653 twists to facilitate compression later.
1654 Used for debugging make_trie().
1657 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1658 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1663 SV *sv=sv_newmortal();
1664 int colwidth= widecharmap ? 6 : 4;
1665 GET_RE_DEBUG_FLAGS_DECL;
1667 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1670 print out the table precompression so that we can do a visual check
1671 that they are identical.
1674 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1676 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1677 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1679 PerlIO_printf( Perl_debug_log, "%*s",
1681 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1682 PL_colors[0], PL_colors[1],
1683 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1684 PERL_PV_ESCAPE_FIRSTCHAR
1690 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1692 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1693 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1696 PerlIO_printf( Perl_debug_log, "\n" );
1698 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1700 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1701 (int)depth * 2 + 2,"",
1702 (UV)TRIE_NODENUM( state ) );
1704 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1705 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1707 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1709 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1711 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1712 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1713 (UV)trie->trans[ state ].check );
1715 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1716 (UV)trie->trans[ state ].check,
1717 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1725 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1726 startbranch: the first branch in the whole branch sequence
1727 first : start branch of sequence of branch-exact nodes.
1728 May be the same as startbranch
1729 last : Thing following the last branch.
1730 May be the same as tail.
1731 tail : item following the branch sequence
1732 count : words in the sequence
1733 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1734 depth : indent depth
1736 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1738 A trie is an N'ary tree where the branches are determined by digital
1739 decomposition of the key. IE, at the root node you look up the 1st character and
1740 follow that branch repeat until you find the end of the branches. Nodes can be
1741 marked as "accepting" meaning they represent a complete word. Eg:
1745 would convert into the following structure. Numbers represent states, letters
1746 following numbers represent valid transitions on the letter from that state, if
1747 the number is in square brackets it represents an accepting state, otherwise it
1748 will be in parenthesis.
1750 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1754 (1) +-i->(6)-+-s->[7]
1756 +-s->(3)-+-h->(4)-+-e->[5]
1758 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1760 This shows that when matching against the string 'hers' we will begin at state 1
1761 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1762 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1763 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1764 single traverse. We store a mapping from accepting to state to which word was
1765 matched, and then when we have multiple possibilities we try to complete the
1766 rest of the regex in the order in which they occured in the alternation.
1768 The only prior NFA like behaviour that would be changed by the TRIE support is
1769 the silent ignoring of duplicate alternations which are of the form:
1771 / (DUPE|DUPE) X? (?{ ... }) Y /x
1773 Thus EVAL blocks following a trie may be called a different number of times with
1774 and without the optimisation. With the optimisations dupes will be silently
1775 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1776 the following demonstrates:
1778 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1780 which prints out 'word' three times, but
1782 'words'=~/(word|word|word)(?{ print $1 })S/
1784 which doesnt print it out at all. This is due to other optimisations kicking in.
1786 Example of what happens on a structural level:
1788 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1790 1: CURLYM[1] {1,32767}(18)
1801 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1802 and should turn into:
1804 1: CURLYM[1] {1,32767}(18)
1806 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1814 Cases where tail != last would be like /(?foo|bar)baz/:
1824 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1825 and would end up looking like:
1828 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1835 d = uvchr_to_utf8_flags(d, uv, 0);
1837 is the recommended Unicode-aware way of saying
1842 #define TRIE_STORE_REVCHAR(val) \
1845 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1846 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1847 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1848 SvCUR_set(zlopp, kapow - flrbbbbb); \
1851 av_push(revcharmap, zlopp); \
1853 char ooooff = (char)val; \
1854 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1858 /* This gets the next character from the input, folding it if not already
1860 #define TRIE_READ_CHAR STMT_START { \
1863 /* if it is UTF then it is either already folded, or does not need \
1865 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1867 else if (folder == PL_fold_latin1) { \
1868 /* This folder implies Unicode rules, which in the range expressible \
1869 * by not UTF is the lower case, with the two exceptions, one of \
1870 * which should have been taken care of before calling this */ \
1871 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1872 uvc = toLOWER_L1(*uc); \
1873 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1876 /* raw data, will be folded later if needed */ \
1884 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1885 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1886 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1887 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1889 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1890 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1891 TRIE_LIST_CUR( state )++; \
1894 #define TRIE_LIST_NEW(state) STMT_START { \
1895 Newxz( trie->states[ state ].trans.list, \
1896 4, reg_trie_trans_le ); \
1897 TRIE_LIST_CUR( state ) = 1; \
1898 TRIE_LIST_LEN( state ) = 4; \
1901 #define TRIE_HANDLE_WORD(state) STMT_START { \
1902 U16 dupe= trie->states[ state ].wordnum; \
1903 regnode * const noper_next = regnext( noper ); \
1906 /* store the word for dumping */ \
1908 if (OP(noper) != NOTHING) \
1909 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1911 tmp = newSVpvn_utf8( "", 0, UTF ); \
1912 av_push( trie_words, tmp ); \
1916 trie->wordinfo[curword].prev = 0; \
1917 trie->wordinfo[curword].len = wordlen; \
1918 trie->wordinfo[curword].accept = state; \
1920 if ( noper_next < tail ) { \
1922 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1924 trie->jump[curword] = (U16)(noper_next - convert); \
1926 jumper = noper_next; \
1928 nextbranch= regnext(cur); \
1932 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1933 /* chain, so that when the bits of chain are later */\
1934 /* linked together, the dups appear in the chain */\
1935 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1936 trie->wordinfo[dupe].prev = curword; \
1938 /* we haven't inserted this word yet. */ \
1939 trie->states[ state ].wordnum = curword; \
1944 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1945 ( ( base + charid >= ucharcount \
1946 && base + charid < ubound \
1947 && state == trie->trans[ base - ucharcount + charid ].check \
1948 && trie->trans[ base - ucharcount + charid ].next ) \
1949 ? trie->trans[ base - ucharcount + charid ].next \
1950 : ( state==1 ? special : 0 ) \
1954 #define MADE_JUMP_TRIE 2
1955 #define MADE_EXACT_TRIE 4
1958 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1959 regnode *first, regnode *last, regnode *tail,
1960 U32 word_count, U32 flags, U32 depth)
1963 /* first pass, loop through and scan words */
1964 reg_trie_data *trie;
1965 HV *widecharmap = NULL;
1966 AV *revcharmap = newAV();
1972 regnode *jumper = NULL;
1973 regnode *nextbranch = NULL;
1974 regnode *convert = NULL;
1975 U32 *prev_states; /* temp array mapping each state to previous one */
1976 /* we just use folder as a flag in utf8 */
1977 const U8 * folder = NULL;
1980 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1981 AV *trie_words = NULL;
1982 /* along with revcharmap, this only used during construction but both are
1983 * useful during debugging so we store them in the struct when debugging.
1986 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1987 STRLEN trie_charcount=0;
1989 SV *re_trie_maxbuff;
1990 GET_RE_DEBUG_FLAGS_DECL;
1992 PERL_ARGS_ASSERT_MAKE_TRIE;
1994 PERL_UNUSED_ARG(depth);
2001 case EXACTFU: folder = PL_fold_latin1; break;
2002 case EXACTF: folder = PL_fold; break;
2003 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2006 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2008 trie->startstate = 1;
2009 trie->wordcount = word_count;
2010 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2011 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2013 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2014 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2015 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2018 trie_words = newAV();
2021 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2022 if (!SvIOK(re_trie_maxbuff)) {
2023 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2025 DEBUG_TRIE_COMPILE_r({
2026 PerlIO_printf( Perl_debug_log,
2027 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2028 (int)depth * 2 + 2, "",
2029 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2030 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2033 /* Find the node we are going to overwrite */
2034 if ( first == startbranch && OP( last ) != BRANCH ) {
2035 /* whole branch chain */
2038 /* branch sub-chain */
2039 convert = NEXTOPER( first );
2042 /* -- First loop and Setup --
2044 We first traverse the branches and scan each word to determine if it
2045 contains widechars, and how many unique chars there are, this is
2046 important as we have to build a table with at least as many columns as we
2049 We use an array of integers to represent the character codes 0..255
2050 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2051 the native representation of the character value as the key and IV's for
2054 *TODO* If we keep track of how many times each character is used we can
2055 remap the columns so that the table compression later on is more
2056 efficient in terms of memory by ensuring the most common value is in the
2057 middle and the least common are on the outside. IMO this would be better
2058 than a most to least common mapping as theres a decent chance the most
2059 common letter will share a node with the least common, meaning the node
2060 will not be compressible. With a middle is most common approach the worst
2061 case is when we have the least common nodes twice.
2065 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2066 regnode *noper = NEXTOPER( cur );
2067 const U8 *uc = (U8*)STRING( noper );
2068 const U8 *e = uc + STR_LEN( noper );
2070 U32 wordlen = 0; /* required init */
2071 STRLEN minbytes = 0;
2072 STRLEN maxbytes = 0;
2073 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2076 if (OP(noper) == NOTHING) {
2077 regnode *noper_next= regnext(noper);
2078 if (noper_next != tail && OP(noper_next) == flags) {
2080 uc= (U8*)STRING(noper);
2081 e= uc + STR_LEN(noper);
2082 trie->minlen= STR_LEN(noper);
2089 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2090 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2091 regardless of encoding */
2092 if (OP( noper ) == EXACTFU_SS) {
2093 /* false positives are ok, so just set this */
2094 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2097 for ( ; uc < e ; uc += len ) {
2098 TRIE_CHARCOUNT(trie)++;
2101 /* Acummulate to the current values, the range in the number of
2102 * bytes that this character could match. The max is presumed to
2103 * be the same as the folded input (which TRIE_READ_CHAR returns),
2104 * except that when this is not in UTF-8, it could be matched
2105 * against a string which is UTF-8, and the variant characters
2106 * could be 2 bytes instead of the 1 here. Likewise, for the
2107 * minimum number of bytes when not folded. When folding, the min
2108 * is assumed to be 1 byte could fold to match the single character
2109 * here, or in the case of a multi-char fold, 1 byte can fold to
2110 * the whole sequence. 'foldlen' is used to denote whether we are
2111 * in such a sequence, skipping the min setting if so. XXX TODO
2112 * Use the exact list of what folds to each character, from
2113 * PL_utf8_foldclosures */
2115 maxbytes += UTF8SKIP(uc);
2117 /* A non-UTF-8 string could be 1 byte to match our 2 */
2118 minbytes += (UTF8_IS_DOWNGRADEABLE_START(*uc))
2124 foldlen -= UTF8SKIP(uc);
2127 foldlen = is_MULTI_CHAR_FOLD_utf8(uc);
2133 maxbytes += (UNI_IS_INVARIANT(*uc))
2144 foldlen = is_MULTI_CHAR_FOLD_latin1(uc);
2151 U8 folded= folder[ (U8) uvc ];
2152 if ( !trie->charmap[ folded ] ) {
2153 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2154 TRIE_STORE_REVCHAR( folded );
2157 if ( !trie->charmap[ uvc ] ) {
2158 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2159 TRIE_STORE_REVCHAR( uvc );
2162 /* store the codepoint in the bitmap, and its folded
2164 TRIE_BITMAP_SET(trie, uvc);
2166 /* store the folded codepoint */
2167 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2170 /* store first byte of utf8 representation of
2171 variant codepoints */
2172 if (! UVCHR_IS_INVARIANT(uvc)) {
2173 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2176 set_bit = 0; /* We've done our bit :-) */
2181 widecharmap = newHV();
2183 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2186 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2188 if ( !SvTRUE( *svpp ) ) {
2189 sv_setiv( *svpp, ++trie->uniquecharcount );
2190 TRIE_STORE_REVCHAR(uvc);
2194 if( cur == first ) {
2195 trie->minlen = minbytes;
2196 trie->maxlen = maxbytes;
2197 } else if (minbytes < trie->minlen) {
2198 trie->minlen = minbytes;
2199 } else if (maxbytes > trie->maxlen) {
2200 trie->maxlen = maxbytes;
2202 } /* end first pass */
2203 DEBUG_TRIE_COMPILE_r(
2204 PerlIO_printf( Perl_debug_log,
2205 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2206 (int)depth * 2 + 2,"",
2207 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2208 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2209 (int)trie->minlen, (int)trie->maxlen )
2213 We now know what we are dealing with in terms of unique chars and
2214 string sizes so we can calculate how much memory a naive
2215 representation using a flat table will take. If it's over a reasonable
2216 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2217 conservative but potentially much slower representation using an array
2220 At the end we convert both representations into the same compressed
2221 form that will be used in regexec.c for matching with. The latter
2222 is a form that cannot be used to construct with but has memory
2223 properties similar to the list form and access properties similar
2224 to the table form making it both suitable for fast searches and
2225 small enough that its feasable to store for the duration of a program.
2227 See the comment in the code where the compressed table is produced
2228 inplace from the flat tabe representation for an explanation of how
2229 the compression works.
2234 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2237 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2238 > SvIV(re_trie_maxbuff) )
2241 Second Pass -- Array Of Lists Representation
2243 Each state will be represented by a list of charid:state records
2244 (reg_trie_trans_le) the first such element holds the CUR and LEN
2245 points of the allocated array. (See defines above).
2247 We build the initial structure using the lists, and then convert
2248 it into the compressed table form which allows faster lookups
2249 (but cant be modified once converted).
2252 STRLEN transcount = 1;
2254 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2255 "%*sCompiling trie using list compiler\n",
2256 (int)depth * 2 + 2, ""));
2258 trie->states = (reg_trie_state *)
2259 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2260 sizeof(reg_trie_state) );
2264 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2266 regnode *noper = NEXTOPER( cur );
2267 U8 *uc = (U8*)STRING( noper );
2268 const U8 *e = uc + STR_LEN( noper );
2269 U32 state = 1; /* required init */
2270 U16 charid = 0; /* sanity init */
2271 U32 wordlen = 0; /* required init */
2273 if (OP(noper) == NOTHING) {
2274 regnode *noper_next= regnext(noper);
2275 if (noper_next != tail && OP(noper_next) == flags) {
2277 uc= (U8*)STRING(noper);
2278 e= uc + STR_LEN(noper);
2282 if (OP(noper) != NOTHING) {
2283 for ( ; uc < e ; uc += len ) {
2288 charid = trie->charmap[ uvc ];
2290 SV** const svpp = hv_fetch( widecharmap,
2297 charid=(U16)SvIV( *svpp );
2300 /* charid is now 0 if we dont know the char read, or
2301 * nonzero if we do */
2308 if ( !trie->states[ state ].trans.list ) {
2309 TRIE_LIST_NEW( state );
2312 check <= TRIE_LIST_USED( state );
2315 if ( TRIE_LIST_ITEM( state, check ).forid
2318 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2323 newstate = next_alloc++;
2324 prev_states[newstate] = state;
2325 TRIE_LIST_PUSH( state, charid, newstate );
2330 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2334 TRIE_HANDLE_WORD(state);
2336 } /* end second pass */
2338 /* next alloc is the NEXT state to be allocated */
2339 trie->statecount = next_alloc;
2340 trie->states = (reg_trie_state *)
2341 PerlMemShared_realloc( trie->states,
2343 * sizeof(reg_trie_state) );
2345 /* and now dump it out before we compress it */
2346 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2347 revcharmap, next_alloc,
2351 trie->trans = (reg_trie_trans *)
2352 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2359 for( state=1 ; state < next_alloc ; state ++ ) {
2363 DEBUG_TRIE_COMPILE_MORE_r(
2364 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2368 if (trie->states[state].trans.list) {
2369 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2373 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2374 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2375 if ( forid < minid ) {
2377 } else if ( forid > maxid ) {
2381 if ( transcount < tp + maxid - minid + 1) {
2383 trie->trans = (reg_trie_trans *)
2384 PerlMemShared_realloc( trie->trans,
2386 * sizeof(reg_trie_trans) );
2387 Zero( trie->trans + (transcount / 2),
2391 base = trie->uniquecharcount + tp - minid;
2392 if ( maxid == minid ) {
2394 for ( ; zp < tp ; zp++ ) {
2395 if ( ! trie->trans[ zp ].next ) {
2396 base = trie->uniquecharcount + zp - minid;
2397 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2399 trie->trans[ zp ].check = state;
2405 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2407 trie->trans[ tp ].check = state;
2412 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2413 const U32 tid = base
2414 - trie->uniquecharcount
2415 + TRIE_LIST_ITEM( state, idx ).forid;
2416 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2418 trie->trans[ tid ].check = state;
2420 tp += ( maxid - minid + 1 );
2422 Safefree(trie->states[ state ].trans.list);
2425 DEBUG_TRIE_COMPILE_MORE_r(
2426 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2429 trie->states[ state ].trans.base=base;
2431 trie->lasttrans = tp + 1;
2435 Second Pass -- Flat Table Representation.
2437 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2438 each. We know that we will need Charcount+1 trans at most to store
2439 the data (one row per char at worst case) So we preallocate both
2440 structures assuming worst case.
2442 We then construct the trie using only the .next slots of the entry
2445 We use the .check field of the first entry of the node temporarily
2446 to make compression both faster and easier by keeping track of how
2447 many non zero fields are in the node.
2449 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2452 There are two terms at use here: state as a TRIE_NODEIDX() which is
2453 a number representing the first entry of the node, and state as a
2454 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2455 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2456 if there are 2 entrys per node. eg:
2464 The table is internally in the right hand, idx form. However as we
2465 also have to deal with the states array which is indexed by nodenum
2466 we have to use TRIE_NODENUM() to convert.
2469 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2470 "%*sCompiling trie using table compiler\n",
2471 (int)depth * 2 + 2, ""));
2473 trie->trans = (reg_trie_trans *)
2474 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2475 * trie->uniquecharcount + 1,
2476 sizeof(reg_trie_trans) );
2477 trie->states = (reg_trie_state *)
2478 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2479 sizeof(reg_trie_state) );
2480 next_alloc = trie->uniquecharcount + 1;
2483 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2485 regnode *noper = NEXTOPER( cur );
2486 const U8 *uc = (U8*)STRING( noper );
2487 const U8 *e = uc + STR_LEN( noper );
2489 U32 state = 1; /* required init */
2491 U16 charid = 0; /* sanity init */
2492 U32 accept_state = 0; /* sanity init */
2494 U32 wordlen = 0; /* required init */
2496 if (OP(noper) == NOTHING) {
2497 regnode *noper_next= regnext(noper);
2498 if (noper_next != tail && OP(noper_next) == flags) {
2500 uc= (U8*)STRING(noper);
2501 e= uc + STR_LEN(noper);
2505 if ( OP(noper) != NOTHING ) {
2506 for ( ; uc < e ; uc += len ) {
2511 charid = trie->charmap[ uvc ];
2513 SV* const * const svpp = hv_fetch( widecharmap,
2517 charid = svpp ? (U16)SvIV(*svpp) : 0;
2521 if ( !trie->trans[ state + charid ].next ) {
2522 trie->trans[ state + charid ].next = next_alloc;
2523 trie->trans[ state ].check++;
2524 prev_states[TRIE_NODENUM(next_alloc)]
2525 = TRIE_NODENUM(state);
2526 next_alloc += trie->uniquecharcount;
2528 state = trie->trans[ state + charid ].next;
2530 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2532 /* charid is now 0 if we dont know the char read, or
2533 * nonzero if we do */
2536 accept_state = TRIE_NODENUM( state );
2537 TRIE_HANDLE_WORD(accept_state);
2539 } /* end second pass */
2541 /* and now dump it out before we compress it */
2542 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2544 next_alloc, depth+1));
2548 * Inplace compress the table.*
2550 For sparse data sets the table constructed by the trie algorithm will
2551 be mostly 0/FAIL transitions or to put it another way mostly empty.
2552 (Note that leaf nodes will not contain any transitions.)
2554 This algorithm compresses the tables by eliminating most such
2555 transitions, at the cost of a modest bit of extra work during lookup:
2557 - Each states[] entry contains a .base field which indicates the
2558 index in the state[] array wheres its transition data is stored.
2560 - If .base is 0 there are no valid transitions from that node.
2562 - If .base is nonzero then charid is added to it to find an entry in
2565 -If trans[states[state].base+charid].check!=state then the
2566 transition is taken to be a 0/Fail transition. Thus if there are fail
2567 transitions at the front of the node then the .base offset will point
2568 somewhere inside the previous nodes data (or maybe even into a node
2569 even earlier), but the .check field determines if the transition is
2573 The following process inplace converts the table to the compressed
2574 table: We first do not compress the root node 1,and mark all its
2575 .check pointers as 1 and set its .base pointer as 1 as well. This
2576 allows us to do a DFA construction from the compressed table later,
2577 and ensures that any .base pointers we calculate later are greater
2580 - We set 'pos' to indicate the first entry of the second node.
2582 - We then iterate over the columns of the node, finding the first and
2583 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2584 and set the .check pointers accordingly, and advance pos
2585 appropriately and repreat for the next node. Note that when we copy
2586 the next pointers we have to convert them from the original
2587 NODEIDX form to NODENUM form as the former is not valid post
2590 - If a node has no transitions used we mark its base as 0 and do not
2591 advance the pos pointer.
2593 - If a node only has one transition we use a second pointer into the
2594 structure to fill in allocated fail transitions from other states.
2595 This pointer is independent of the main pointer and scans forward
2596 looking for null transitions that are allocated to a state. When it
2597 finds one it writes the single transition into the "hole". If the
2598 pointer doesnt find one the single transition is appended as normal.
2600 - Once compressed we can Renew/realloc the structures to release the
2603 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2604 specifically Fig 3.47 and the associated pseudocode.
2608 const U32 laststate = TRIE_NODENUM( next_alloc );
2611 trie->statecount = laststate;
2613 for ( state = 1 ; state < laststate ; state++ ) {
2615 const U32 stateidx = TRIE_NODEIDX( state );
2616 const U32 o_used = trie->trans[ stateidx ].check;
2617 U32 used = trie->trans[ stateidx ].check;
2618 trie->trans[ stateidx ].check = 0;
2621 used && charid < trie->uniquecharcount;
2624 if ( flag || trie->trans[ stateidx + charid ].next ) {
2625 if ( trie->trans[ stateidx + charid ].next ) {
2627 for ( ; zp < pos ; zp++ ) {
2628 if ( ! trie->trans[ zp ].next ) {
2632 trie->states[ state ].trans.base
2634 + trie->uniquecharcount
2636 trie->trans[ zp ].next
2637 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2639 trie->trans[ zp ].check = state;
2640 if ( ++zp > pos ) pos = zp;
2647 trie->states[ state ].trans.base
2648 = pos + trie->uniquecharcount - charid ;
2650 trie->trans[ pos ].next
2651 = SAFE_TRIE_NODENUM(
2652 trie->trans[ stateidx + charid ].next );
2653 trie->trans[ pos ].check = state;
2658 trie->lasttrans = pos + 1;
2659 trie->states = (reg_trie_state *)
2660 PerlMemShared_realloc( trie->states, laststate
2661 * sizeof(reg_trie_state) );
2662 DEBUG_TRIE_COMPILE_MORE_r(
2663 PerlIO_printf( Perl_debug_log,
2664 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2665 (int)depth * 2 + 2,"",
2666 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2670 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2673 } /* end table compress */
2675 DEBUG_TRIE_COMPILE_MORE_r(
2676 PerlIO_printf(Perl_debug_log,
2677 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2678 (int)depth * 2 + 2, "",
2679 (UV)trie->statecount,
2680 (UV)trie->lasttrans)
2682 /* resize the trans array to remove unused space */
2683 trie->trans = (reg_trie_trans *)
2684 PerlMemShared_realloc( trie->trans, trie->lasttrans
2685 * sizeof(reg_trie_trans) );
2687 { /* Modify the program and insert the new TRIE node */
2688 U8 nodetype =(U8)(flags & 0xFF);
2692 regnode *optimize = NULL;
2693 #ifdef RE_TRACK_PATTERN_OFFSETS
2696 U32 mjd_nodelen = 0;
2697 #endif /* RE_TRACK_PATTERN_OFFSETS */
2698 #endif /* DEBUGGING */
2700 This means we convert either the first branch or the first Exact,
2701 depending on whether the thing following (in 'last') is a branch
2702 or not and whther first is the startbranch (ie is it a sub part of
2703 the alternation or is it the whole thing.)
2704 Assuming its a sub part we convert the EXACT otherwise we convert
2705 the whole branch sequence, including the first.
2707 /* Find the node we are going to overwrite */
2708 if ( first != startbranch || OP( last ) == BRANCH ) {
2709 /* branch sub-chain */
2710 NEXT_OFF( first ) = (U16)(last - first);
2711 #ifdef RE_TRACK_PATTERN_OFFSETS
2713 mjd_offset= Node_Offset((convert));
2714 mjd_nodelen= Node_Length((convert));
2717 /* whole branch chain */
2719 #ifdef RE_TRACK_PATTERN_OFFSETS
2722 const regnode *nop = NEXTOPER( convert );
2723 mjd_offset= Node_Offset((nop));
2724 mjd_nodelen= Node_Length((nop));
2728 PerlIO_printf(Perl_debug_log,
2729 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2730 (int)depth * 2 + 2, "",
2731 (UV)mjd_offset, (UV)mjd_nodelen)
2734 /* But first we check to see if there is a common prefix we can
2735 split out as an EXACT and put in front of the TRIE node. */
2736 trie->startstate= 1;
2737 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2739 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2743 const U32 base = trie->states[ state ].trans.base;
2745 if ( trie->states[state].wordnum )
2748 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2749 if ( ( base + ofs >= trie->uniquecharcount ) &&
2750 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2751 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2753 if ( ++count > 1 ) {
2754 SV **tmp = av_fetch( revcharmap, ofs, 0);
2755 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2756 if ( state == 1 ) break;
2758 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2760 PerlIO_printf(Perl_debug_log,
2761 "%*sNew Start State=%"UVuf" Class: [",
2762 (int)depth * 2 + 2, "",
2765 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2766 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2768 TRIE_BITMAP_SET(trie,*ch);
2770 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2772 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2776 TRIE_BITMAP_SET(trie,*ch);
2778 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2779 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2785 SV **tmp = av_fetch( revcharmap, idx, 0);
2787 char *ch = SvPV( *tmp, len );
2789 SV *sv=sv_newmortal();
2790 PerlIO_printf( Perl_debug_log,
2791 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2792 (int)depth * 2 + 2, "",
2794 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2795 PL_colors[0], PL_colors[1],
2796 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2797 PERL_PV_ESCAPE_FIRSTCHAR
2802 OP( convert ) = nodetype;
2803 str=STRING(convert);
2806 STR_LEN(convert) += len;
2812 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2817 trie->prefixlen = (state-1);
2819 regnode *n = convert+NODE_SZ_STR(convert);
2820 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2821 trie->startstate = state;
2822 trie->minlen -= (state - 1);
2823 trie->maxlen -= (state - 1);
2825 /* At least the UNICOS C compiler choked on this
2826 * being argument to DEBUG_r(), so let's just have
2829 #ifdef PERL_EXT_RE_BUILD
2835 regnode *fix = convert;
2836 U32 word = trie->wordcount;
2838 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2839 while( ++fix < n ) {
2840 Set_Node_Offset_Length(fix, 0, 0);
2843 SV ** const tmp = av_fetch( trie_words, word, 0 );
2845 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2846 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2848 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2856 NEXT_OFF(convert) = (U16)(tail - convert);
2857 DEBUG_r(optimize= n);
2863 if ( trie->maxlen ) {
2864 NEXT_OFF( convert ) = (U16)(tail - convert);
2865 ARG_SET( convert, data_slot );
2866 /* Store the offset to the first unabsorbed branch in
2867 jump[0], which is otherwise unused by the jump logic.
2868 We use this when dumping a trie and during optimisation. */
2870 trie->jump[0] = (U16)(nextbranch - convert);
2872 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2873 * and there is a bitmap
2874 * and the first "jump target" node we found leaves enough room
2875 * then convert the TRIE node into a TRIEC node, with the bitmap
2876 * embedded inline in the opcode - this is hypothetically faster.
2878 if ( !trie->states[trie->startstate].wordnum
2880 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2882 OP( convert ) = TRIEC;
2883 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2884 PerlMemShared_free(trie->bitmap);
2887 OP( convert ) = TRIE;
2889 /* store the type in the flags */
2890 convert->flags = nodetype;
2894 + regarglen[ OP( convert ) ];
2896 /* XXX We really should free up the resource in trie now,
2897 as we won't use them - (which resources?) dmq */
2899 /* needed for dumping*/
2900 DEBUG_r(if (optimize) {
2901 regnode *opt = convert;
2903 while ( ++opt < optimize) {
2904 Set_Node_Offset_Length(opt,0,0);
2907 Try to clean up some of the debris left after the
2910 while( optimize < jumper ) {
2911 mjd_nodelen += Node_Length((optimize));
2912 OP( optimize ) = OPTIMIZED;
2913 Set_Node_Offset_Length(optimize,0,0);
2916 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2918 } /* end node insert */
2920 /* Finish populating the prev field of the wordinfo array. Walk back
2921 * from each accept state until we find another accept state, and if
2922 * so, point the first word's .prev field at the second word. If the
2923 * second already has a .prev field set, stop now. This will be the
2924 * case either if we've already processed that word's accept state,
2925 * or that state had multiple words, and the overspill words were
2926 * already linked up earlier.
2933 for (word=1; word <= trie->wordcount; word++) {
2935 if (trie->wordinfo[word].prev)
2937 state = trie->wordinfo[word].accept;
2939 state = prev_states[state];
2942 prev = trie->states[state].wordnum;
2946 trie->wordinfo[word].prev = prev;
2948 Safefree(prev_states);
2952 /* and now dump out the compressed format */
2953 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2955 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2957 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2958 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2960 SvREFCNT_dec_NN(revcharmap);
2964 : trie->startstate>1
2970 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2972 /* The Trie is constructed and compressed now so we can build a fail array if
2975 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2977 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2981 We find the fail state for each state in the trie, this state is the longest
2982 proper suffix of the current state's 'word' that is also a proper prefix of
2983 another word in our trie. State 1 represents the word '' and is thus the
2984 default fail state. This allows the DFA not to have to restart after its
2985 tried and failed a word at a given point, it simply continues as though it
2986 had been matching the other word in the first place.
2988 'abcdgu'=~/abcdefg|cdgu/
2989 When we get to 'd' we are still matching the first word, we would encounter
2990 'g' which would fail, which would bring us to the state representing 'd' in
2991 the second word where we would try 'g' and succeed, proceeding to match
2994 /* add a fail transition */
2995 const U32 trie_offset = ARG(source);
2996 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2998 const U32 ucharcount = trie->uniquecharcount;
2999 const U32 numstates = trie->statecount;
3000 const U32 ubound = trie->lasttrans + ucharcount;
3004 U32 base = trie->states[ 1 ].trans.base;
3007 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3008 GET_RE_DEBUG_FLAGS_DECL;
3010 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
3012 PERL_UNUSED_ARG(depth);
3016 ARG_SET( stclass, data_slot );
3017 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3018 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3019 aho->trie=trie_offset;
3020 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3021 Copy( trie->states, aho->states, numstates, reg_trie_state );
3022 Newxz( q, numstates, U32);
3023 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3026 /* initialize fail[0..1] to be 1 so that we always have
3027 a valid final fail state */
3028 fail[ 0 ] = fail[ 1 ] = 1;
3030 for ( charid = 0; charid < ucharcount ; charid++ ) {
3031 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3033 q[ q_write ] = newstate;
3034 /* set to point at the root */
3035 fail[ q[ q_write++ ] ]=1;
3038 while ( q_read < q_write) {
3039 const U32 cur = q[ q_read++ % numstates ];
3040 base = trie->states[ cur ].trans.base;
3042 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3043 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3045 U32 fail_state = cur;
3048 fail_state = fail[ fail_state ];
3049 fail_base = aho->states[ fail_state ].trans.base;
3050 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3052 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3053 fail[ ch_state ] = fail_state;
3054 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3056 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3058 q[ q_write++ % numstates] = ch_state;
3062 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3063 when we fail in state 1, this allows us to use the
3064 charclass scan to find a valid start char. This is based on the principle
3065 that theres a good chance the string being searched contains lots of stuff
3066 that cant be a start char.
3068 fail[ 0 ] = fail[ 1 ] = 0;
3069 DEBUG_TRIE_COMPILE_r({
3070 PerlIO_printf(Perl_debug_log,
3071 "%*sStclass Failtable (%"UVuf" states): 0",
3072 (int)(depth * 2), "", (UV)numstates
3074 for( q_read=1; q_read<numstates; q_read++ ) {
3075 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3077 PerlIO_printf(Perl_debug_log, "\n");
3080 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3084 #define DEBUG_PEEP(str,scan,depth) \
3085 DEBUG_OPTIMISE_r({if (scan){ \
3086 SV * const mysv=sv_newmortal(); \
3087 regnode *Next = regnext(scan); \
3088 regprop(RExC_rx, mysv, scan); \
3089 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3090 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3091 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3095 /* The below joins as many adjacent EXACTish nodes as possible into a single
3096 * one. The regop may be changed if the node(s) contain certain sequences that
3097 * require special handling. The joining is only done if:
3098 * 1) there is room in the current conglomerated node to entirely contain the
3100 * 2) they are the exact same node type
3102 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3103 * these get optimized out
3105 * If a node is to match under /i (folded), the number of characters it matches
3106 * can be different than its character length if it contains a multi-character
3107 * fold. *min_subtract is set to the total delta number of characters of the
3110 * And *unfolded_multi_char is set to indicate whether or not the node contains
3111 * an unfolded multi-char fold. This happens when whether the fold is valid or
3112 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3113 * SMALL LETTER SHARP S, as only if the target string being matched against
3114 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3115 * folding rules depend on the locale in force at runtime. (Multi-char folds
3116 * whose components are all above the Latin1 range are not run-time locale
3117 * dependent, and have already been folded by the time this function is
3120 * This is as good a place as any to discuss the design of handling these
3121 * multi-character fold sequences. It's been wrong in Perl for a very long
3122 * time. There are three code points in Unicode whose multi-character folds
3123 * were long ago discovered to mess things up. The previous designs for
3124 * dealing with these involved assigning a special node for them. This
3125 * approach doesn't always work, as evidenced by this example:
3126 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3127 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3128 * would match just the \xDF, it won't be able to handle the case where a
3129 * successful match would have to cross the node's boundary. The new approach
3130 * that hopefully generally solves the problem generates an EXACTFU_SS node
3131 * that is "sss" in this case.
3133 * It turns out that there are problems with all multi-character folds, and not
3134 * just these three. Now the code is general, for all such cases. The
3135 * approach taken is:
3136 * 1) This routine examines each EXACTFish node that could contain multi-
3137 * character folded sequences. Since a single character can fold into
3138 * such a sequence, the minimum match length for this node is less than
3139 * the number of characters in the node. This routine returns in
3140 * *min_subtract how many characters to subtract from the the actual
3141 * length of the string to get a real minimum match length; it is 0 if
3142 * there are no multi-char foldeds. This delta is used by the caller to
3143 * adjust the min length of the match, and the delta between min and max,
3144 * so that the optimizer doesn't reject these possibilities based on size
3146 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3147 * is used for an EXACTFU node that contains at least one "ss" sequence in
3148 * it. For non-UTF-8 patterns and strings, this is the only case where
3149 * there is a possible fold length change. That means that a regular
3150 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3151 * with length changes, and so can be processed faster. regexec.c takes
3152 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3153 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3154 * known until runtime). This saves effort in regex matching. However,
3155 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3156 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3157 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3158 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3159 * possibilities for the non-UTF8 patterns are quite simple, except for
3160 * the sharp s. All the ones that don't involve a UTF-8 target string are
3161 * members of a fold-pair, and arrays are set up for all of them so that
3162 * the other member of the pair can be found quickly. Code elsewhere in
3163 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3164 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3165 * described in the next item.
3166 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3167 * validity of the fold won't be known until runtime, and so must remain
3168 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3169 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3170 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3171 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3172 * The reason this is a problem is that the optimizer part of regexec.c
3173 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3174 * that a character in the pattern corresponds to at most a single
3175 * character in the target string. (And I do mean character, and not byte
3176 * here, unlike other parts of the documentation that have never been
3177 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3178 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3179 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3180 * nodes, violate the assumption, and they are the only instances where it
3181 * is violated. I'm reluctant to try to change the assumption, as the
3182 * code involved is impenetrable to me (khw), so instead the code here
3183 * punts. This routine examines EXACTFL nodes, and (when the pattern
3184 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3185 * boolean indicating whether or not the node contains such a fold. When
3186 * it is true, the caller sets a flag that later causes the optimizer in
3187 * this file to not set values for the floating and fixed string lengths,
3188 * and thus avoids the optimizer code in regexec.c that makes the invalid
3189 * assumption. Thus, there is no optimization based on string lengths for
3190 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3191 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3192 * assumption is wrong only in these cases is that all other non-UTF-8
3193 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3194 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3195 * EXACTF nodes because we don't know at compile time if it actually
3196 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3197 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3198 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3199 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3200 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3201 * string would require the pattern to be forced into UTF-8, the overhead
3202 * of which we want to avoid. Similarly the unfolded multi-char folds in
3203 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3206 * Similarly, the code that generates tries doesn't currently handle
3207 * not-already-folded multi-char folds, and it looks like a pain to change
3208 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3209 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3210 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3211 * using /iaa matching will be doing so almost entirely with ASCII
3212 * strings, so this should rarely be encountered in practice */
3214 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3215 if (PL_regkind[OP(scan)] == EXACT) \
3216 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3219 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3220 UV *min_subtract, bool *unfolded_multi_char,
3221 U32 flags,regnode *val, U32 depth)
3223 /* Merge several consecutive EXACTish nodes into one. */
3224 regnode *n = regnext(scan);
3226 regnode *next = scan + NODE_SZ_STR(scan);
3230 regnode *stop = scan;
3231 GET_RE_DEBUG_FLAGS_DECL;
3233 PERL_UNUSED_ARG(depth);
3236 PERL_ARGS_ASSERT_JOIN_EXACT;
3237 #ifndef EXPERIMENTAL_INPLACESCAN
3238 PERL_UNUSED_ARG(flags);
3239 PERL_UNUSED_ARG(val);
3241 DEBUG_PEEP("join",scan,depth);
3243 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3244 * EXACT ones that are mergeable to the current one. */
3246 && (PL_regkind[OP(n)] == NOTHING
3247 || (stringok && OP(n) == OP(scan)))
3249 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3252 if (OP(n) == TAIL || n > next)
3254 if (PL_regkind[OP(n)] == NOTHING) {
3255 DEBUG_PEEP("skip:",n,depth);
3256 NEXT_OFF(scan) += NEXT_OFF(n);
3257 next = n + NODE_STEP_REGNODE;
3264 else if (stringok) {
3265 const unsigned int oldl = STR_LEN(scan);
3266 regnode * const nnext = regnext(n);
3268 /* XXX I (khw) kind of doubt that this works on platforms (should
3269 * Perl ever run on one) where U8_MAX is above 255 because of lots
3270 * of other assumptions */
3271 /* Don't join if the sum can't fit into a single node */
3272 if (oldl + STR_LEN(n) > U8_MAX)
3275 DEBUG_PEEP("merg",n,depth);
3278 NEXT_OFF(scan) += NEXT_OFF(n);
3279 STR_LEN(scan) += STR_LEN(n);
3280 next = n + NODE_SZ_STR(n);
3281 /* Now we can overwrite *n : */
3282 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3290 #ifdef EXPERIMENTAL_INPLACESCAN
3291 if (flags && !NEXT_OFF(n)) {
3292 DEBUG_PEEP("atch", val, depth);
3293 if (reg_off_by_arg[OP(n)]) {
3294 ARG_SET(n, val - n);
3297 NEXT_OFF(n) = val - n;
3305 *unfolded_multi_char = FALSE;
3307 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3308 * can now analyze for sequences of problematic code points. (Prior to
3309 * this final joining, sequences could have been split over boundaries, and
3310 * hence missed). The sequences only happen in folding, hence for any
3311 * non-EXACT EXACTish node */
3312 if (OP(scan) != EXACT) {
3313 U8* s0 = (U8*) STRING(scan);
3315 U8* s_end = s0 + STR_LEN(scan);
3317 int total_count_delta = 0; /* Total delta number of characters that
3318 multi-char folds expand to */
3320 /* One pass is made over the node's string looking for all the
3321 * possibilities. To avoid some tests in the loop, there are two main
3322 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3327 if (OP(scan) == EXACTFL) {
3330 /* An EXACTFL node would already have been changed to another
3331 * node type unless there is at least one character in it that
3332 * is problematic; likely a character whose fold definition
3333 * won't be known until runtime, and so has yet to be folded.
3334 * For all but the UTF-8 locale, folds are 1-1 in length, but
3335 * to handle the UTF-8 case, we need to create a temporary
3336 * folded copy using UTF-8 locale rules in order to analyze it.
3337 * This is because our macros that look to see if a sequence is
3338 * a multi-char fold assume everything is folded (otherwise the
3339 * tests in those macros would be too complicated and slow).
3340 * Note that here, the non-problematic folds will have already
3341 * been done, so we can just copy such characters. We actually
3342 * don't completely fold the EXACTFL string. We skip the
3343 * unfolded multi-char folds, as that would just create work
3344 * below to figure out the size they already are */
3346 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3349 STRLEN s_len = UTF8SKIP(s);
3350 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3351 Copy(s, d, s_len, U8);
3354 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3355 *unfolded_multi_char = TRUE;
3356 Copy(s, d, s_len, U8);
3359 else if (isASCII(*s)) {
3360 *(d++) = toFOLD(*s);
3364 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3370 /* Point the remainder of the routine to look at our temporary
3374 } /* End of creating folded copy of EXACTFL string */
3376 /* Examine the string for a multi-character fold sequence. UTF-8
3377 * patterns have all characters pre-folded by the time this code is
3379 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3380 length sequence we are looking for is 2 */
3382 int count = 0; /* How many characters in a multi-char fold */
3383 int len = is_MULTI_CHAR_FOLD_utf8(s);
3384 if (! len) { /* Not a multi-char fold: get next char */
3389 /* Nodes with 'ss' require special handling, except for
3390 * EXACTFA-ish for which there is no multi-char fold to this */
3391 if (len == 2 && *s == 's' && *(s+1) == 's'
3392 && OP(scan) != EXACTFA
3393 && OP(scan) != EXACTFA_NO_TRIE)
3396 if (OP(scan) != EXACTFL) {
3397 OP(scan) = EXACTFU_SS;
3401 else { /* Here is a generic multi-char fold. */
3402 U8* multi_end = s + len;
3404 /* Count how many characters in it. In the case of /aa, no
3405 * folds which contain ASCII code points are allowed, so
3406 * check for those, and skip if found. */
3407 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3408 count = utf8_length(s, multi_end);
3412 while (s < multi_end) {
3415 goto next_iteration;
3425 /* The delta is how long the sequence is minus 1 (1 is how long
3426 * the character that folds to the sequence is) */
3427 total_count_delta += count - 1;
3431 /* We created a temporary folded copy of the string in EXACTFL
3432 * nodes. Therefore we need to be sure it doesn't go below zero,
3433 * as the real string could be shorter */
3434 if (OP(scan) == EXACTFL) {
3435 int total_chars = utf8_length((U8*) STRING(scan),
3436 (U8*) STRING(scan) + STR_LEN(scan));
3437 if (total_count_delta > total_chars) {
3438 total_count_delta = total_chars;
3442 *min_subtract += total_count_delta;
3445 else if (OP(scan) == EXACTFA) {
3447 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3448 * fold to the ASCII range (and there are no existing ones in the
3449 * upper latin1 range). But, as outlined in the comments preceding
3450 * this function, we need to flag any occurrences of the sharp s.
3451 * This character forbids trie formation (because of added
3454 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3455 OP(scan) = EXACTFA_NO_TRIE;
3456 *unfolded_multi_char = TRUE;
3465 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3466 * folds that are all Latin1. As explained in the comments
3467 * preceding this function, we look also for the sharp s in EXACTF
3468 * and EXACTFL nodes; it can be in the final position. Otherwise
3469 * we can stop looking 1 byte earlier because have to find at least
3470 * two characters for a multi-fold */
3471 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3476 int len = is_MULTI_CHAR_FOLD_latin1(s);
3477 if (! len) { /* Not a multi-char fold. */
3478 if (*s == LATIN_SMALL_LETTER_SHARP_S
3479 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3481 *unfolded_multi_char = TRUE;
3488 && isARG2_lower_or_UPPER_ARG1('s', *s)
3489 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3492 /* EXACTF nodes need to know that the minimum length
3493 * changed so that a sharp s in the string can match this
3494 * ss in the pattern, but they remain EXACTF nodes, as they
3495 * won't match this unless the target string is is UTF-8,
3496 * which we don't know until runtime. EXACTFL nodes can't
3497 * transform into EXACTFU nodes */
3498 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3499 OP(scan) = EXACTFU_SS;
3503 *min_subtract += len - 1;
3510 /* Allow dumping but overwriting the collection of skipped
3511 * ops and/or strings with fake optimized ops */
3512 n = scan + NODE_SZ_STR(scan);
3520 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3524 /* REx optimizer. Converts nodes into quicker variants "in place".
3525 Finds fixed substrings. */
3527 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3528 to the position after last scanned or to NULL. */
3530 #define INIT_AND_WITHP \
3531 assert(!and_withp); \
3532 Newx(and_withp,1, regnode_ssc); \
3533 SAVEFREEPV(and_withp)
3535 /* this is a chain of data about sub patterns we are processing that
3536 need to be handled separately/specially in study_chunk. Its so
3537 we can simulate recursion without losing state. */
3539 typedef struct scan_frame {
3540 regnode *last; /* last node to process in this frame */
3541 regnode *next; /* next node to process when last is reached */
3542 struct scan_frame *prev; /*previous frame*/
3543 U32 prev_recursed_depth;
3544 I32 stop; /* what stopparen do we use */
3548 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3551 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3552 SSize_t *minlenp, SSize_t *deltap,
3557 regnode_ssc *and_withp,
3558 U32 flags, U32 depth)
3559 /* scanp: Start here (read-write). */
3560 /* deltap: Write maxlen-minlen here. */
3561 /* last: Stop before this one. */
3562 /* data: string data about the pattern */
3563 /* stopparen: treat close N as END */
3564 /* recursed: which subroutines have we recursed into */
3565 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3568 /* There must be at least this number of characters to match */
3571 regnode *scan = *scanp, *next;
3573 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3574 int is_inf_internal = 0; /* The studied chunk is infinite */
3575 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3576 scan_data_t data_fake;
3577 SV *re_trie_maxbuff = NULL;
3578 regnode *first_non_open = scan;
3579 SSize_t stopmin = SSize_t_MAX;
3580 scan_frame *frame = NULL;
3581 GET_RE_DEBUG_FLAGS_DECL;
3583 PERL_ARGS_ASSERT_STUDY_CHUNK;
3586 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3589 while (first_non_open && OP(first_non_open) == OPEN)
3590 first_non_open=regnext(first_non_open);
3595 while ( scan && OP(scan) != END && scan < last ){
3596 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3597 node length to get a real minimum (because
3598 the folded version may be shorter) */
3599 bool unfolded_multi_char = FALSE;
3600 /* Peephole optimizer: */
3601 DEBUG_OPTIMISE_MORE_r(
3603 PerlIO_printf(Perl_debug_log,
3604 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3605 ((int) depth*2), "", (long)stopparen,
3606 (unsigned long)depth, (unsigned long)recursed_depth);
3607 if (recursed_depth) {
3610 for ( j = 0 ; j < recursed_depth ; j++ ) {
3611 PerlIO_printf(Perl_debug_log,"[");
3612 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3613 PerlIO_printf(Perl_debug_log,"%d",
3614 PAREN_TEST(RExC_study_chunk_recursed +
3615 (j * RExC_study_chunk_recursed_bytes), i)
3618 PerlIO_printf(Perl_debug_log,"]");
3621 PerlIO_printf(Perl_debug_log,"\n");
3624 DEBUG_STUDYDATA("Peep:", data, depth);
3625 DEBUG_PEEP("Peep", scan, depth);
3628 /* Its not clear to khw or hv why this is done here, and not in the
3629 * clauses that deal with EXACT nodes. khw's guess is that it's
3630 * because of a previous design */
3631 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3633 /* Follow the next-chain of the current node and optimize
3634 away all the NOTHINGs from it. */
3635 if (OP(scan) != CURLYX) {
3636 const int max = (reg_off_by_arg[OP(scan)]
3638 /* I32 may be smaller than U16 on CRAYs! */
3639 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3640 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3644 /* Skip NOTHING and LONGJMP. */
3645 while ((n = regnext(n))
3646 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3647 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3648 && off + noff < max)
3650 if (reg_off_by_arg[OP(scan)])
3653 NEXT_OFF(scan) = off;
3658 /* The principal pseudo-switch. Cannot be a switch, since we
3659 look into several different things. */
3660 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3661 || OP(scan) == IFTHEN) {
3662 next = regnext(scan);
3664 /* demq: the op(next)==code check is to see if we have
3665 * "branch-branch" AFAICT */
3667 if (OP(next) == code || code == IFTHEN) {
3668 /* NOTE - There is similar code to this block below for
3669 * handling TRIE nodes on a re-study. If you change stuff here
3670 * check there too. */
3671 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3673 regnode * const startbranch=scan;
3675 if (flags & SCF_DO_SUBSTR)
3676 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge
3679 if (flags & SCF_DO_STCLASS)
3680 ssc_init_zero(pRExC_state, &accum);
3682 while (OP(scan) == code) {
3683 SSize_t deltanext, minnext, fake;
3685 regnode_ssc this_class;
3688 data_fake.flags = 0;
3690 data_fake.whilem_c = data->whilem_c;
3691 data_fake.last_closep = data->last_closep;
3694 data_fake.last_closep = &fake;
3696 data_fake.pos_delta = delta;
3697 next = regnext(scan);
3698 scan = NEXTOPER(scan);
3700 scan = NEXTOPER(scan);
3701 if (flags & SCF_DO_STCLASS) {
3702 ssc_init(pRExC_state, &this_class);
3703 data_fake.start_class = &this_class;
3704 f = SCF_DO_STCLASS_AND;
3706 if (flags & SCF_WHILEM_VISITED_POS)
3707 f |= SCF_WHILEM_VISITED_POS;
3709 /* we suppose the run is continuous, last=next...*/
3710 minnext = study_chunk(pRExC_state, &scan, minlenp,
3711 &deltanext, next, &data_fake, stopparen,
3712 recursed_depth, NULL, f,depth+1);
3715 if (deltanext == SSize_t_MAX) {
3716 is_inf = is_inf_internal = 1;
3718 } else if (max1 < minnext + deltanext)
3719 max1 = minnext + deltanext;
3721 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3723 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3724 if ( stopmin > minnext)
3725 stopmin = min + min1;
3726 flags &= ~SCF_DO_SUBSTR;
3728 data->flags |= SCF_SEEN_ACCEPT;
3731 if (data_fake.flags & SF_HAS_EVAL)
3732 data->flags |= SF_HAS_EVAL;
3733 data->whilem_c = data_fake.whilem_c;
3735 if (flags & SCF_DO_STCLASS)
3736 ssc_or(pRExC_state, &accum, &this_class);
3738 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3740 if (flags & SCF_DO_SUBSTR) {
3741 data->pos_min += min1;
3742 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3743 data->pos_delta = SSize_t_MAX;
3745 data->pos_delta += max1 - min1;
3746 if (max1 != min1 || is_inf)
3747 data->longest = &(data->longest_float);
3750 if (delta == SSize_t_MAX
3751 || SSize_t_MAX - delta - (max1 - min1) < 0)
3752 delta = SSize_t_MAX;
3754 delta += max1 - min1;
3755 if (flags & SCF_DO_STCLASS_OR) {
3756 ssc_or(pRExC_state, data->start_class, &accum);
3758 ssc_and(pRExC_state, data->start_class, and_withp);
3759 flags &= ~SCF_DO_STCLASS;
3762 else if (flags & SCF_DO_STCLASS_AND) {
3764 ssc_and(pRExC_state, data->start_class, &accum);
3765 flags &= ~SCF_DO_STCLASS;
3768 /* Switch to OR mode: cache the old value of
3769 * data->start_class */
3771 StructCopy(data->start_class, and_withp, regnode_ssc);
3772 flags &= ~SCF_DO_STCLASS_AND;
3773 StructCopy(&accum, data->start_class, regnode_ssc);
3774 flags |= SCF_DO_STCLASS_OR;
3778 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch )
3783 Assuming this was/is a branch we are dealing with: 'scan'
3784 now points at the item that follows the branch sequence,
3785 whatever it is. We now start at the beginning of the
3786 sequence and look for subsequences of
3792 which would be constructed from a pattern like
3795 If we can find such a subsequence we need to turn the first
3796 element into a trie and then add the subsequent branch exact
3797 strings to the trie.
3801 1. patterns where the whole set of branches can be
3804 2. patterns where only a subset can be converted.
3806 In case 1 we can replace the whole set with a single regop
3807 for the trie. In case 2 we need to keep the start and end
3810 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3811 becomes BRANCH TRIE; BRANCH X;
3813 There is an additional case, that being where there is a
3814 common prefix, which gets split out into an EXACT like node
3815 preceding the TRIE node.
3817 If x(1..n)==tail then we can do a simple trie, if not we make
3818 a "jump" trie, such that when we match the appropriate word
3819 we "jump" to the appropriate tail node. Essentially we turn
3820 a nested if into a case structure of sorts.
3825 if (!re_trie_maxbuff) {
3826 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3827 if (!SvIOK(re_trie_maxbuff))
3828 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3830 if ( SvIV(re_trie_maxbuff)>=0 ) {
3832 regnode *first = (regnode *)NULL;
3833 regnode *last = (regnode *)NULL;
3834 regnode *tail = scan;
3839 SV * const mysv = sv_newmortal(); /* for dumping */
3841 /* var tail is used because there may be a TAIL
3842 regop in the way. Ie, the exacts will point to the
3843 thing following the TAIL, but the last branch will
3844 point at the TAIL. So we advance tail. If we
3845 have nested (?:) we may have to move through several
3849 while ( OP( tail ) == TAIL ) {
3850 /* this is the TAIL generated by (?:) */
3851 tail = regnext( tail );
3855 DEBUG_TRIE_COMPILE_r({
3856 regprop(RExC_rx, mysv, tail );
3857 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3858 (int)depth * 2 + 2, "",
3859 "Looking for TRIE'able sequences. Tail node is: ",
3860 SvPV_nolen_const( mysv )
3866 Step through the branches
3867 cur represents each branch,
3868 noper is the first thing to be matched as part
3870 noper_next is the regnext() of that node.
3872 We normally handle a case like this
3873 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3874 support building with NOJUMPTRIE, which restricts
3875 the trie logic to structures like /FOO|BAR/.
3877 If noper is a trieable nodetype then the branch is
3878 a possible optimization target. If we are building
3879 under NOJUMPTRIE then we require that noper_next is
3880 the same as scan (our current position in the regex
3883 Once we have two or more consecutive such branches
3884 we can create a trie of the EXACT's contents and
3885 stitch it in place into the program.
3887 If the sequence represents all of the branches in
3888 the alternation we replace the entire thing with a
3891 Otherwise when it is a subsequence we need to
3892 stitch it in place and replace only the relevant
3893 branches. This means the first branch has to remain
3894 as it is used by the alternation logic, and its
3895 next pointer, and needs to be repointed at the item
3896 on the branch chain following the last branch we
3897 have optimized away.
3899 This could be either a BRANCH, in which case the
3900 subsequence is internal, or it could be the item
3901 following the branch sequence in which case the
3902 subsequence is at the end (which does not
3903 necessarily mean the first node is the start of the
3906 TRIE_TYPE(X) is a define which maps the optype to a
3910 ----------------+-----------
3914 EXACTFU_SS | EXACTFU
3919 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3920 ( EXACT == (X) ) ? EXACT : \
3921 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3922 ( EXACTFA == (X) ) ? EXACTFA : \
3925 /* dont use tail as the end marker for this traverse */
3926 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3927 regnode * const noper = NEXTOPER( cur );
3928 U8 noper_type = OP( noper );
3929 U8 noper_trietype = TRIE_TYPE( noper_type );
3930 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3931 regnode * const noper_next = regnext( noper );
3932 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3933 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3936 DEBUG_TRIE_COMPILE_r({
3937 regprop(RExC_rx, mysv, cur);
3938 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3939 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3941 regprop(RExC_rx, mysv, noper);
3942 PerlIO_printf( Perl_debug_log, " -> %s",
3943 SvPV_nolen_const(mysv));
3946 regprop(RExC_rx, mysv, noper_next );
3947 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3948 SvPV_nolen_const(mysv));
3950 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3951 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3952 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3956 /* Is noper a trieable nodetype that can be merged
3957 * with the current trie (if there is one)? */
3961 ( noper_trietype == NOTHING)
3962 || ( trietype == NOTHING )
3963 || ( trietype == noper_trietype )
3966 && noper_next == tail
3970 /* Handle mergable triable node Either we are
3971 * the first node in a new trieable sequence,
3972 * in which case we do some bookkeeping,
3973 * otherwise we update the end pointer. */
3976 if ( noper_trietype == NOTHING ) {
3977 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3978 regnode * const noper_next = regnext( noper );
3979 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3980 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3983 if ( noper_next_trietype ) {
3984 trietype = noper_next_trietype;
3985 } else if (noper_next_type) {
3986 /* a NOTHING regop is 1 regop wide.
3987 * We need at least two for a trie
3988 * so we can't merge this in */
3992 trietype = noper_trietype;
3995 if ( trietype == NOTHING )
3996 trietype = noper_trietype;
4001 } /* end handle mergable triable node */
4003 /* handle unmergable node -
4004 * noper may either be a triable node which can
4005 * not be tried together with the current trie,
4006 * or a non triable node */
4008 /* If last is set and trietype is not
4009 * NOTHING then we have found at least two
4010 * triable branch sequences in a row of a
4011 * similar trietype so we can turn them
4012 * into a trie. If/when we allow NOTHING to
4013 * start a trie sequence this condition
4014 * will be required, and it isn't expensive
4015 * so we leave it in for now. */
4016 if ( trietype && trietype != NOTHING )
4017 make_trie( pRExC_state,
4018 startbranch, first, cur, tail,
4019 count, trietype, depth+1 );
4020 last = NULL; /* note: we clear/update
4021 first, trietype etc below,
4022 so we dont do it here */
4026 && noper_next == tail
4029 /* noper is triable, so we can start a new
4033 trietype = noper_trietype;
4035 /* if we already saw a first but the
4036 * current node is not triable then we have
4037 * to reset the first information. */
4042 } /* end handle unmergable node */
4043 } /* loop over branches */
4044 DEBUG_TRIE_COMPILE_r({
4045 regprop(RExC_rx, mysv, cur);
4046 PerlIO_printf( Perl_debug_log,
4047 "%*s- %s (%d) <SCAN FINISHED>\n",
4049 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4052 if ( last && trietype ) {
4053 if ( trietype != NOTHING ) {
4054 /* the last branch of the sequence was part of
4055 * a trie, so we have to construct it here
4056 * outside of the loop */
4057 made= make_trie( pRExC_state, startbranch,
4058 first, scan, tail, count,
4059 trietype, depth+1 );
4060 #ifdef TRIE_STUDY_OPT
4061 if ( ((made == MADE_EXACT_TRIE &&
4062 startbranch == first)
4063 || ( first_non_open == first )) &&
4065 flags |= SCF_TRIE_RESTUDY;
4066 if ( startbranch == first
4069 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4074 /* at this point we know whatever we have is a
4075 * NOTHING sequence/branch AND if 'startbranch'
4076 * is 'first' then we can turn the whole thing
4079 if ( startbranch == first ) {
4081 /* the entire thing is a NOTHING sequence,
4082 * something like this: (?:|) So we can
4083 * turn it into a plain NOTHING op. */
4084 DEBUG_TRIE_COMPILE_r({
4085 regprop(RExC_rx, mysv, cur);
4086 PerlIO_printf( Perl_debug_log,
4087 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4088 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4091 OP(startbranch)= NOTHING;
4092 NEXT_OFF(startbranch)= tail - startbranch;
4093 for ( opt= startbranch + 1; opt < tail ; opt++ )
4097 } /* end if ( last) */
4098 } /* TRIE_MAXBUF is non zero */
4103 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4104 scan = NEXTOPER(NEXTOPER(scan));
4105 } else /* single branch is optimized. */
4106 scan = NEXTOPER(scan);
4108 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4109 scan_frame *newframe = NULL;
4113 U32 my_recursed_depth= recursed_depth;
4115 if (OP(scan) != SUSPEND) {
4116 /* set the pointer */
4117 if (OP(scan) == GOSUB) {
4119 RExC_recurse[ARG2L(scan)] = scan;
4120 start = RExC_open_parens[paren-1];
4121 end = RExC_close_parens[paren-1];
4124 start = RExC_rxi->program + 1;
4129 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4131 if (!recursed_depth) {
4132 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4134 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4135 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4136 RExC_study_chunk_recursed_bytes, U8);
4138 /* we havent recursed into this paren yet, so recurse into it */
4139 DEBUG_STUDYDATA("set:", data,depth);
4140 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4141 my_recursed_depth= recursed_depth + 1;
4142 Newx(newframe,1,scan_frame);
4144 DEBUG_STUDYDATA("inf:", data,depth);
4145 /* some form of infinite recursion, assume infinite length
4147 if (flags & SCF_DO_SUBSTR) {
4148 SCAN_COMMIT(pRExC_state,data,minlenp);
4149 data->longest = &(data->longest_float);
4151 is_inf = is_inf_internal = 1;
4152 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4153 ssc_anything(data->start_class);
4154 flags &= ~SCF_DO_STCLASS;
4157 Newx(newframe,1,scan_frame);
4160 end = regnext(scan);
4165 SAVEFREEPV(newframe);
4166 newframe->next = regnext(scan);
4167 newframe->last = last;
4168 newframe->stop = stopparen;
4169 newframe->prev = frame;
4170 newframe->prev_recursed_depth = recursed_depth;
4172 DEBUG_STUDYDATA("frame-new:",data,depth);
4173 DEBUG_PEEP("fnew", scan, depth);
4180 recursed_depth= my_recursed_depth;
4185 else if (OP(scan) == EXACT) {
4186 SSize_t l = STR_LEN(scan);
4189 const U8 * const s = (U8*)STRING(scan);
4190 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4191 l = utf8_length(s, s + l);
4193 uc = *((U8*)STRING(scan));
4196 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4197 /* The code below prefers earlier match for fixed
4198 offset, later match for variable offset. */
4199 if (data->last_end == -1) { /* Update the start info. */
4200 data->last_start_min = data->pos_min;
4201 data->last_start_max = is_inf
4202 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4204 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4206 SvUTF8_on(data->last_found);
4208 SV * const sv = data->last_found;
4209 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4210 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4211 if (mg && mg->mg_len >= 0)
4212 mg->mg_len += utf8_length((U8*)STRING(scan),
4213 (U8*)STRING(scan)+STR_LEN(scan));
4215 data->last_end = data->pos_min + l;
4216 data->pos_min += l; /* As in the first entry. */
4217 data->flags &= ~SF_BEFORE_EOL;
4220 /* ANDing the code point leaves at most it, and not in locale, and
4221 * can't match null string */
4222 if (flags & SCF_DO_STCLASS_AND) {
4223 ssc_cp_and(data->start_class, uc);
4224 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4225 ssc_clear_locale(data->start_class);
4227 else if (flags & SCF_DO_STCLASS_OR) {
4228 ssc_add_cp(data->start_class, uc);
4229 ssc_and(pRExC_state, data->start_class, and_withp);
4231 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4232 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4234 flags &= ~SCF_DO_STCLASS;
4236 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
4237 SSize_t l = STR_LEN(scan);
4238 UV uc = *((U8*)STRING(scan));
4239 SV* EXACTF_invlist = _new_invlist(4); /* Start out big enough for 2
4240 separate code points */
4242 /* Search for fixed substrings supports EXACT only. */
4243 if (flags & SCF_DO_SUBSTR) {
4245 SCAN_COMMIT(pRExC_state, data, minlenp);
4248 const U8 * const s = (U8 *)STRING(scan);
4249 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4250 l = utf8_length(s, s + l);
4252 if (unfolded_multi_char) {
4253 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4255 min += l - min_subtract;
4257 delta += min_subtract;
4258 if (flags & SCF_DO_SUBSTR) {
4259 data->pos_min += l - min_subtract;
4260 if (data->pos_min < 0) {
4263 data->pos_delta += min_subtract;
4265 data->longest = &(data->longest_float);
4268 if (OP(scan) == EXACTFL) {
4269 if (flags & SCF_DO_STCLASS_AND) {
4270 ssc_flags_and(data->start_class, ANYOF_LOCALE);
4272 else if (flags & SCF_DO_STCLASS_OR) {
4273 ANYOF_FLAGS(data->start_class) |= ANYOF_LOCALE;
4276 /* We don't know what the folds are; it could be anything. XXX
4277 * Actually, we only support UTF-8 encoding for code points
4278 * above Latin1, so we could know what those folds are. */
4279 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4283 else { /* Non-locale EXACTFish */
4284 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4285 if (flags & SCF_DO_STCLASS_AND) {
4286 ssc_clear_locale(data->start_class);
4288 if (uc < 256) { /* We know what the Latin1 folds are ... */
4289 if (IS_IN_SOME_FOLD_L1(uc)) { /* For instance, we
4290 know if anything folds
4292 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist,
4293 PL_fold_latin1[uc]);
4294 if (OP(scan) != EXACTFA) { /* The folds below aren't
4296 if (isARG2_lower_or_UPPER_ARG1('s', uc)) {
4298 = add_cp_to_invlist(EXACTF_invlist,
4299 LATIN_SMALL_LETTER_SHARP_S);
4301 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
4303 = add_cp_to_invlist(EXACTF_invlist, 's');
4305 = add_cp_to_invlist(EXACTF_invlist, 'S');
4309 /* We also know if there are above-Latin1 code points
4310 * that fold to this (none legal for ASCII and /iaa) */
4311 if ((! isASCII(uc) || OP(scan) != EXACTFA)
4312 && HAS_NONLATIN1_FOLD_CLOSURE(uc))
4314 /* XXX We could know exactly what does fold to this
4315 * if the reverse folds are loaded, as currently in
4317 _invlist_union(EXACTF_invlist,
4323 else { /* Non-locale, above Latin1. XXX We don't currently
4324 know what participates in folds with this, so have
4325 to assume anything could */
4327 /* XXX We could know exactly what does fold to this if the
4328 * reverse folds are loaded, as currently in S_regclass().
4329 * But we do know that under /iaa nothing in the ASCII
4330 * range can participate */
4331 if (OP(scan) == EXACTFA) {
4332 _invlist_union_complement_2nd(EXACTF_invlist,
4333 PL_XPosix_ptrs[_CC_ASCII],
4337 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4342 if (flags & SCF_DO_STCLASS_AND) {
4343 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4344 ANYOF_POSIXL_ZERO(data->start_class);
4345 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4347 else if (flags & SCF_DO_STCLASS_OR) {
4348 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4349 ssc_and(pRExC_state, data->start_class, and_withp);
4351 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4352 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4354 flags &= ~SCF_DO_STCLASS;
4355 SvREFCNT_dec(EXACTF_invlist);
4357 else if (REGNODE_VARIES(OP(scan))) {
4358 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4359 I32 fl = 0, f = flags;
4360 regnode * const oscan = scan;
4361 regnode_ssc this_class;
4362 regnode_ssc *oclass = NULL;
4363 I32 next_is_eval = 0;
4365 switch (PL_regkind[OP(scan)]) {
4366 case WHILEM: /* End of (?:...)* . */
4367 scan = NEXTOPER(scan);
4370 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4371 next = NEXTOPER(scan);
4372 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
4374 maxcount = REG_INFTY;
4375 next = regnext(scan);
4376 scan = NEXTOPER(scan);
4380 if (flags & SCF_DO_SUBSTR)
4385 if (flags & SCF_DO_STCLASS) {
4387 maxcount = REG_INFTY;
4388 next = regnext(scan);
4389 scan = NEXTOPER(scan);
4392 is_inf = is_inf_internal = 1;
4393 scan = regnext(scan);
4394 if (flags & SCF_DO_SUBSTR) {
4395 SCAN_COMMIT(pRExC_state, data, minlenp);
4396 /* Cannot extend fixed substrings */
4397 data->longest = &(data->longest_float);
4399 goto optimize_curly_tail;
4401 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4402 && (scan->flags == stopparen))
4407 mincount = ARG1(scan);
4408 maxcount = ARG2(scan);
4410 next = regnext(scan);
4411 if (OP(scan) == CURLYX) {
4412 I32 lp = (data ? *(data->last_closep) : 0);
4413 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4415 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4416 next_is_eval = (OP(scan) == EVAL);
4418 if (flags & SCF_DO_SUBSTR) {
4419 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp);
4420 /* Cannot extend fixed substrings */
4421 pos_before = data->pos_min;
4425 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4427 data->flags |= SF_IS_INF;
4429 if (flags & SCF_DO_STCLASS) {
4430 ssc_init(pRExC_state, &this_class);
4431 oclass = data->start_class;
4432 data->start_class = &this_class;
4433 f |= SCF_DO_STCLASS_AND;
4434 f &= ~SCF_DO_STCLASS_OR;
4436 /* Exclude from super-linear cache processing any {n,m}
4437 regops for which the combination of input pos and regex
4438 pos is not enough information to determine if a match
4441 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4442 regex pos at the \s*, the prospects for a match depend not
4443 only on the input position but also on how many (bar\s*)
4444 repeats into the {4,8} we are. */
4445 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4446 f &= ~SCF_WHILEM_VISITED_POS;
4448 /* This will finish on WHILEM, setting scan, or on NULL: */
4449 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4450 last, data, stopparen, recursed_depth, NULL,
4452 ? (f & ~SCF_DO_SUBSTR)
4456 if (flags & SCF_DO_STCLASS)
4457 data->start_class = oclass;
4458 if (mincount == 0 || minnext == 0) {
4459 if (flags & SCF_DO_STCLASS_OR) {
4460 ssc_or(pRExC_state, data->start_class, &this_class);
4462 else if (flags & SCF_DO_STCLASS_AND) {
4463 /* Switch to OR mode: cache the old value of
4464 * data->start_class */
4466 StructCopy(data->start_class, and_withp, regnode_ssc);
4467 flags &= ~SCF_DO_STCLASS_AND;
4468 StructCopy(&this_class, data->start_class, regnode_ssc);
4469 flags |= SCF_DO_STCLASS_OR;
4470 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
4472 } else { /* Non-zero len */
4473 if (flags & SCF_DO_STCLASS_OR) {
4474 ssc_or(pRExC_state, data->start_class, &this_class);
4475 ssc_and(pRExC_state, data->start_class, and_withp);
4477 else if (flags & SCF_DO_STCLASS_AND)
4478 ssc_and(pRExC_state, data->start_class, &this_class);
4479 flags &= ~SCF_DO_STCLASS;
4481 if (!scan) /* It was not CURLYX, but CURLY. */
4483 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4484 /* ? quantifier ok, except for (?{ ... }) */
4485 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4486 && (minnext == 0) && (deltanext == 0)
4487 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4488 && maxcount <= REG_INFTY/3) /* Complement check for big
4491 /* Fatal warnings may leak the regexp without this: */
4492 SAVEFREESV(RExC_rx_sv);
4493 ckWARNreg(RExC_parse,
4494 "Quantifier unexpected on zero-length expression");
4495 (void)ReREFCNT_inc(RExC_rx_sv);
4498 min += minnext * mincount;
4499 is_inf_internal |= deltanext == SSize_t_MAX
4500 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4501 is_inf |= is_inf_internal;
4503 delta = SSize_t_MAX;
4505 delta += (minnext + deltanext) * maxcount
4506 - minnext * mincount;
4508 /* Try powerful optimization CURLYX => CURLYN. */
4509 if ( OP(oscan) == CURLYX && data
4510 && data->flags & SF_IN_PAR
4511 && !(data->flags & SF_HAS_EVAL)
4512 && !deltanext && minnext == 1 ) {
4513 /* Try to optimize to CURLYN. */
4514 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4515 regnode * const nxt1 = nxt;
4522 if (!REGNODE_SIMPLE(OP(nxt))
4523 && !(PL_regkind[OP(nxt)] == EXACT
4524 && STR_LEN(nxt) == 1))
4530 if (OP(nxt) != CLOSE)
4532 if (RExC_open_parens) {
4533 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4534 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4536 /* Now we know that nxt2 is the only contents: */
4537 oscan->flags = (U8)ARG(nxt);
4539 OP(nxt1) = NOTHING; /* was OPEN. */
4542 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4543 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4544 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4545 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4546 OP(nxt + 1) = OPTIMIZED; /* was count. */
4547 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4552 /* Try optimization CURLYX => CURLYM. */
4553 if ( OP(oscan) == CURLYX && data
4554 && !(data->flags & SF_HAS_PAR)
4555 && !(data->flags & SF_HAS_EVAL)
4556 && !deltanext /* atom is fixed width */
4557 && minnext != 0 /* CURLYM can't handle zero width */
4559 /* Nor characters whose fold at run-time may be
4560 * multi-character */
4561 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4563 /* XXXX How to optimize if data == 0? */
4564 /* Optimize to a simpler form. */
4565 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4569 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4570 && (OP(nxt2) != WHILEM))
4572 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4573 /* Need to optimize away parenths. */
4574 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4575 /* Set the parenth number. */
4576 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4578 oscan->flags = (U8)ARG(nxt);
4579 if (RExC_open_parens) {
4580 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4581 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4583 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4584 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4587 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4588 OP(nxt + 1) = OPTIMIZED; /* was count. */
4589 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4590 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4593 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4594 regnode *nnxt = regnext(nxt1);
4596 if (reg_off_by_arg[OP(nxt1)])
4597 ARG_SET(nxt1, nxt2 - nxt1);
4598 else if (nxt2 - nxt1 < U16_MAX)
4599 NEXT_OFF(nxt1) = nxt2 - nxt1;
4601 OP(nxt) = NOTHING; /* Cannot beautify */
4606 /* Optimize again: */
4607 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4608 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4613 else if ((OP(oscan) == CURLYX)
4614 && (flags & SCF_WHILEM_VISITED_POS)
4615 /* See the comment on a similar expression above.
4616 However, this time it's not a subexpression
4617 we care about, but the expression itself. */
4618 && (maxcount == REG_INFTY)
4619 && data && ++data->whilem_c < 16) {
4620 /* This stays as CURLYX, we can put the count/of pair. */
4621 /* Find WHILEM (as in regexec.c) */
4622 regnode *nxt = oscan + NEXT_OFF(oscan);
4624 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4626 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4627 | (RExC_whilem_seen << 4)); /* On WHILEM */
4629 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4631 if (flags & SCF_DO_SUBSTR) {
4632 SV *last_str = NULL;
4633 int counted = mincount != 0;
4635 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4637 SSize_t b = pos_before >= data->last_start_min
4638 ? pos_before : data->last_start_min;
4640 const char * const s = SvPV_const(data->last_found, l);
4641 SSize_t old = b - data->last_start_min;
4644 old = utf8_hop((U8*)s, old) - (U8*)s;
4646 /* Get the added string: */
4647 last_str = newSVpvn_utf8(s + old, l, UTF);
4648 if (deltanext == 0 && pos_before == b) {
4649 /* What was added is a constant string */
4651 SvGROW(last_str, (mincount * l) + 1);
4652 repeatcpy(SvPVX(last_str) + l,
4653 SvPVX_const(last_str), l,
4655 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4656 /* Add additional parts. */
4657 SvCUR_set(data->last_found,
4658 SvCUR(data->last_found) - l);
4659 sv_catsv(data->last_found, last_str);
4661 SV * sv = data->last_found;
4663 SvUTF8(sv) && SvMAGICAL(sv) ?
4664 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4665 if (mg && mg->mg_len >= 0)
4666 mg->mg_len += CHR_SVLEN(last_str) - l;
4668 data->last_end += l * (mincount - 1);
4671 /* start offset must point into the last copy */
4672 data->last_start_min += minnext * (mincount - 1);
4673 data->last_start_max += is_inf ? SSize_t_MAX
4674 : (maxcount - 1) * (minnext + data->pos_delta);
4677 /* It is counted once already... */
4678 data->pos_min += minnext * (mincount - counted);
4680 PerlIO_printf(Perl_debug_log, "counted=%"UVdf" deltanext=%"UVdf
4681 " SSize_t_MAX=%"UVdf" minnext=%"UVdf
4682 " maxcount=%"UVdf" mincount=%"UVdf"\n",
4683 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4685 if (deltanext != SSize_t_MAX)
4686 PerlIO_printf(Perl_debug_log, "LHS=%"UVdf" RHS=%"UVdf"\n",
4687 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4688 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4690 if (deltanext == SSize_t_MAX
4691 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4692 data->pos_delta = SSize_t_MAX;
4694 data->pos_delta += - counted * deltanext +
4695 (minnext + deltanext) * maxcount - minnext * mincount;
4696 if (mincount != maxcount) {
4697 /* Cannot extend fixed substrings found inside
4699 SCAN_COMMIT(pRExC_state,data,minlenp);
4700 if (mincount && last_str) {
4701 SV * const sv = data->last_found;
4702 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4703 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4707 sv_setsv(sv, last_str);
4708 data->last_end = data->pos_min;
4709 data->last_start_min =
4710 data->pos_min - CHR_SVLEN(last_str);
4711 data->last_start_max = is_inf
4713 : data->pos_min + data->pos_delta
4714 - CHR_SVLEN(last_str);
4716 data->longest = &(data->longest_float);
4718 SvREFCNT_dec(last_str);
4720 if (data && (fl & SF_HAS_EVAL))
4721 data->flags |= SF_HAS_EVAL;
4722 optimize_curly_tail:
4723 if (OP(oscan) != CURLYX) {
4724 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4726 NEXT_OFF(oscan) += NEXT_OFF(next);
4732 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4737 if (flags & SCF_DO_SUBSTR) {
4738 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect
4740 data->longest = &(data->longest_float);
4742 is_inf = is_inf_internal = 1;
4743 if (flags & SCF_DO_STCLASS_OR) {
4744 if (OP(scan) == CLUMP) {
4745 /* Actually is any start char, but very few code points
4746 * aren't start characters */
4747 ssc_match_all_cp(data->start_class);
4750 ssc_anything(data->start_class);
4753 flags &= ~SCF_DO_STCLASS;
4757 else if (OP(scan) == LNBREAK) {
4758 if (flags & SCF_DO_STCLASS) {
4759 if (flags & SCF_DO_STCLASS_AND) {
4760 ssc_intersection(data->start_class,
4761 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
4762 ssc_clear_locale(data->start_class);
4763 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4765 else if (flags & SCF_DO_STCLASS_OR) {
4766 ssc_union(data->start_class,
4767 PL_XPosix_ptrs[_CC_VERTSPACE],
4769 ssc_and(pRExC_state, data->start_class, and_withp);
4771 /* See commit msg for
4772 * 749e076fceedeb708a624933726e7989f2302f6a */
4773 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4775 flags &= ~SCF_DO_STCLASS;
4778 delta++; /* Because of the 2 char string cr-lf */
4779 if (flags & SCF_DO_SUBSTR) {
4780 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect
4783 data->pos_delta += 1;
4784 data->longest = &(data->longest_float);
4787 else if (REGNODE_SIMPLE(OP(scan))) {
4789 if (flags & SCF_DO_SUBSTR) {
4790 SCAN_COMMIT(pRExC_state,data,minlenp);
4794 if (flags & SCF_DO_STCLASS) {
4796 SV* my_invlist = sv_2mortal(_new_invlist(0));
4799 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4800 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4802 /* Some of the logic below assumes that switching
4803 locale on will only add false positives. */
4808 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
4813 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4814 ssc_match_all_cp(data->start_class);
4819 SV* REG_ANY_invlist = _new_invlist(2);
4820 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
4822 if (flags & SCF_DO_STCLASS_OR) {
4823 ssc_union(data->start_class,
4825 TRUE /* TRUE => invert, hence all but \n
4829 else if (flags & SCF_DO_STCLASS_AND) {
4830 ssc_intersection(data->start_class,
4832 TRUE /* TRUE => invert */
4834 ssc_clear_locale(data->start_class);
4836 SvREFCNT_dec_NN(REG_ANY_invlist);
4841 if (flags & SCF_DO_STCLASS_AND)
4842 ssc_and(pRExC_state, data->start_class,
4843 (regnode_ssc*) scan);
4845 ssc_or(pRExC_state, data->start_class,
4846 (regnode_ssc*)scan);
4854 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
4855 if (flags & SCF_DO_STCLASS_AND) {
4856 bool was_there = cBOOL(
4857 ANYOF_POSIXL_TEST(data->start_class,
4859 ANYOF_POSIXL_ZERO(data->start_class);
4860 if (was_there) { /* Do an AND */
4861 ANYOF_POSIXL_SET(data->start_class, namedclass);
4863 /* No individual code points can now match */
4864 data->start_class->invlist
4865 = sv_2mortal(_new_invlist(0));
4868 int complement = namedclass + ((invert) ? -1 : 1);
4870 assert(flags & SCF_DO_STCLASS_OR);
4872 /* If the complement of this class was already there,
4873 * the result is that they match all code points,
4874 * (\d + \D == everything). Remove the classes from
4875 * future consideration. Locale is not relevant in
4877 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
4878 ssc_match_all_cp(data->start_class);
4879 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
4880 ANYOF_POSIXL_CLEAR(data->start_class, complement);
4881 if (! ANYOF_POSIXL_TEST_ANY_SET(data->start_class))
4883 ANYOF_FLAGS(data->start_class) &= ~ANYOF_POSIXL;
4886 else { /* The usual case; just add this class to the
4888 ANYOF_POSIXL_SET(data->start_class, namedclass);
4889 ANYOF_FLAGS(data->start_class)
4890 |= ANYOF_LOCALE|ANYOF_POSIXL;
4895 case NPOSIXA: /* For these, we always know the exact set of
4900 if (FLAGS(scan) == _CC_ASCII) {
4901 my_invlist = PL_XPosix_ptrs[_CC_ASCII];
4904 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
4905 PL_XPosix_ptrs[_CC_ASCII],
4916 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
4918 /* NPOSIXD matches all upper Latin1 code points unless the
4919 * target string being matched is UTF-8, which is
4920 * unknowable until match time. Since we are going to
4921 * invert, we want to get rid of all of them so that the
4922 * inversion will match all */
4923 if (OP(scan) == NPOSIXD) {
4924 _invlist_subtract(my_invlist, PL_UpperLatin1,
4930 if (flags & SCF_DO_STCLASS_AND) {
4931 ssc_intersection(data->start_class, my_invlist, invert);
4932 ssc_clear_locale(data->start_class);
4935 assert(flags & SCF_DO_STCLASS_OR);
4936 ssc_union(data->start_class, my_invlist, invert);
4939 if (flags & SCF_DO_STCLASS_OR)
4940 ssc_and(pRExC_state, data->start_class, and_withp);
4941 flags &= ~SCF_DO_STCLASS;
4944 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4945 data->flags |= (OP(scan) == MEOL
4948 SCAN_COMMIT(pRExC_state, data, minlenp);
4951 else if ( PL_regkind[OP(scan)] == BRANCHJ
4952 /* Lookbehind, or need to calculate parens/evals/stclass: */
4953 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4954 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4955 if ( OP(scan) == UNLESSM &&
4957 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4958 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4961 regnode *upto= regnext(scan);
4963 SV * const mysv_val=sv_newmortal();
4964 DEBUG_STUDYDATA("OPFAIL",data,depth);
4966 /*DEBUG_PARSE_MSG("opfail");*/
4967 regprop(RExC_rx, mysv_val, upto);
4968 PerlIO_printf(Perl_debug_log,
4969 "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4970 SvPV_nolen_const(mysv_val),
4971 (IV)REG_NODE_NUM(upto),
4976 NEXT_OFF(scan) = upto - scan;
4977 for (opt= scan + 1; opt < upto ; opt++)
4978 OP(opt) = OPTIMIZED;
4982 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4983 || OP(scan) == UNLESSM )
4985 /* Negative Lookahead/lookbehind
4986 In this case we can't do fixed string optimisation.
4989 SSize_t deltanext, minnext, fake = 0;
4994 data_fake.flags = 0;
4996 data_fake.whilem_c = data->whilem_c;
4997 data_fake.last_closep = data->last_closep;
5000 data_fake.last_closep = &fake;
5001 data_fake.pos_delta = delta;
5002 if ( flags & SCF_DO_STCLASS && !scan->flags
5003 && OP(scan) == IFMATCH ) { /* Lookahead */
5004 ssc_init(pRExC_state, &intrnl);
5005 data_fake.start_class = &intrnl;
5006 f |= SCF_DO_STCLASS_AND;
5008 if (flags & SCF_WHILEM_VISITED_POS)
5009 f |= SCF_WHILEM_VISITED_POS;
5010 next = regnext(scan);
5011 nscan = NEXTOPER(NEXTOPER(scan));
5012 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5013 last, &data_fake, stopparen,
5014 recursed_depth, NULL, f, depth+1);
5017 FAIL("Variable length lookbehind not implemented");
5019 else if (minnext > (I32)U8_MAX) {
5020 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5023 scan->flags = (U8)minnext;
5026 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5028 if (data_fake.flags & SF_HAS_EVAL)
5029 data->flags |= SF_HAS_EVAL;
5030 data->whilem_c = data_fake.whilem_c;
5032 if (f & SCF_DO_STCLASS_AND) {
5033 if (flags & SCF_DO_STCLASS_OR) {
5034 /* OR before, AND after: ideally we would recurse with
5035 * data_fake to get the AND applied by study of the
5036 * remainder of the pattern, and then derecurse;
5037 * *** HACK *** for now just treat as "no information".
5038 * See [perl #56690].
5040 ssc_init(pRExC_state, data->start_class);
5042 /* AND before and after: combine and continue */
5043 ssc_and(pRExC_state, data->start_class, &intrnl);
5047 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5049 /* Positive Lookahead/lookbehind
5050 In this case we can do fixed string optimisation,
5051 but we must be careful about it. Note in the case of
5052 lookbehind the positions will be offset by the minimum
5053 length of the pattern, something we won't know about
5054 until after the recurse.
5056 SSize_t deltanext, fake = 0;
5060 /* We use SAVEFREEPV so that when the full compile
5061 is finished perl will clean up the allocated
5062 minlens when it's all done. This way we don't
5063 have to worry about freeing them when we know
5064 they wont be used, which would be a pain.
5067 Newx( minnextp, 1, SSize_t );
5068 SAVEFREEPV(minnextp);
5071 StructCopy(data, &data_fake, scan_data_t);
5072 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5075 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
5076 data_fake.last_found=newSVsv(data->last_found);
5080 data_fake.last_closep = &fake;
5081 data_fake.flags = 0;
5082 data_fake.pos_delta = delta;
5084 data_fake.flags |= SF_IS_INF;
5085 if ( flags & SCF_DO_STCLASS && !scan->flags
5086 && OP(scan) == IFMATCH ) { /* Lookahead */
5087 ssc_init(pRExC_state, &intrnl);
5088 data_fake.start_class = &intrnl;
5089 f |= SCF_DO_STCLASS_AND;
5091 if (flags & SCF_WHILEM_VISITED_POS)
5092 f |= SCF_WHILEM_VISITED_POS;
5093 next = regnext(scan);
5094 nscan = NEXTOPER(NEXTOPER(scan));
5096 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5097 &deltanext, last, &data_fake,
5098 stopparen, recursed_depth, NULL,
5102 FAIL("Variable length lookbehind not implemented");
5104 else if (*minnextp > (I32)U8_MAX) {
5105 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5108 scan->flags = (U8)*minnextp;
5113 if (f & SCF_DO_STCLASS_AND) {
5114 ssc_and(pRExC_state, data->start_class, &intrnl);
5117 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5119 if (data_fake.flags & SF_HAS_EVAL)
5120 data->flags |= SF_HAS_EVAL;
5121 data->whilem_c = data_fake.whilem_c;
5122 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5123 if (RExC_rx->minlen<*minnextp)
5124 RExC_rx->minlen=*minnextp;
5125 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
5126 SvREFCNT_dec_NN(data_fake.last_found);
5128 if ( data_fake.minlen_fixed != minlenp )
5130 data->offset_fixed= data_fake.offset_fixed;
5131 data->minlen_fixed= data_fake.minlen_fixed;
5132 data->lookbehind_fixed+= scan->flags;
5134 if ( data_fake.minlen_float != minlenp )
5136 data->minlen_float= data_fake.minlen_float;
5137 data->offset_float_min=data_fake.offset_float_min;
5138 data->offset_float_max=data_fake.offset_float_max;
5139 data->lookbehind_float+= scan->flags;
5146 else if (OP(scan) == OPEN) {
5147 if (stopparen != (I32)ARG(scan))
5150 else if (OP(scan) == CLOSE) {
5151 if (stopparen == (I32)ARG(scan)) {
5154 if ((I32)ARG(scan) == is_par) {
5155 next = regnext(scan);
5157 if ( next && (OP(next) != WHILEM) && next < last)
5158 is_par = 0; /* Disable optimization */
5161 *(data->last_closep) = ARG(scan);
5163 else if (OP(scan) == EVAL) {
5165 data->flags |= SF_HAS_EVAL;
5167 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5168 if (flags & SCF_DO_SUBSTR) {
5169 SCAN_COMMIT(pRExC_state,data,minlenp);
5170 flags &= ~SCF_DO_SUBSTR;
5172 if (data && OP(scan)==ACCEPT) {
5173 data->flags |= SCF_SEEN_ACCEPT;
5178 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5180 if (flags & SCF_DO_SUBSTR) {
5181 SCAN_COMMIT(pRExC_state,data,minlenp);
5182 data->longest = &(data->longest_float);
5184 is_inf = is_inf_internal = 1;
5185 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5186 ssc_anything(data->start_class);
5187 flags &= ~SCF_DO_STCLASS;
5189 else if (OP(scan) == GPOS) {
5190 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5191 !(delta || is_inf || (data && data->pos_delta)))
5193 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5194 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5195 if (RExC_rx->gofs < (STRLEN)min)
5196 RExC_rx->gofs = min;
5198 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5202 #ifdef TRIE_STUDY_OPT
5203 #ifdef FULL_TRIE_STUDY
5204 else if (PL_regkind[OP(scan)] == TRIE) {
5205 /* NOTE - There is similar code to this block above for handling
5206 BRANCH nodes on the initial study. If you change stuff here
5208 regnode *trie_node= scan;
5209 regnode *tail= regnext(scan);
5210 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5211 SSize_t max1 = 0, min1 = SSize_t_MAX;
5214 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
5215 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings
5217 if (flags & SCF_DO_STCLASS)
5218 ssc_init_zero(pRExC_state, &accum);
5224 const regnode *nextbranch= NULL;
5227 for ( word=1 ; word <= trie->wordcount ; word++)
5229 SSize_t deltanext=0, minnext=0, f = 0, fake;
5230 regnode_ssc this_class;
5232 data_fake.flags = 0;
5234 data_fake.whilem_c = data->whilem_c;
5235 data_fake.last_closep = data->last_closep;
5238 data_fake.last_closep = &fake;
5239 data_fake.pos_delta = delta;
5240 if (flags & SCF_DO_STCLASS) {
5241 ssc_init(pRExC_state, &this_class);
5242 data_fake.start_class = &this_class;
5243 f = SCF_DO_STCLASS_AND;
5245 if (flags & SCF_WHILEM_VISITED_POS)
5246 f |= SCF_WHILEM_VISITED_POS;
5248 if (trie->jump[word]) {
5250 nextbranch = trie_node + trie->jump[0];
5251 scan= trie_node + trie->jump[word];
5252 /* We go from the jump point to the branch that follows
5253 it. Note this means we need the vestigal unused
5254 branches even though they arent otherwise used. */
5255 minnext = study_chunk(pRExC_state, &scan, minlenp,
5256 &deltanext, (regnode *)nextbranch, &data_fake,
5257 stopparen, recursed_depth, NULL, f,depth+1);
5259 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5260 nextbranch= regnext((regnode*)nextbranch);
5262 if (min1 > (SSize_t)(minnext + trie->minlen))
5263 min1 = minnext + trie->minlen;
5264 if (deltanext == SSize_t_MAX) {
5265 is_inf = is_inf_internal = 1;
5267 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5268 max1 = minnext + deltanext + trie->maxlen;
5270 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5272 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5273 if ( stopmin > min + min1)
5274 stopmin = min + min1;
5275 flags &= ~SCF_DO_SUBSTR;
5277 data->flags |= SCF_SEEN_ACCEPT;
5280 if (data_fake.flags & SF_HAS_EVAL)
5281 data->flags |= SF_HAS_EVAL;
5282 data->whilem_c = data_fake.whilem_c;
5284 if (flags & SCF_DO_STCLASS)
5285 ssc_or(pRExC_state, &accum, &this_class);
5288 if (flags & SCF_DO_SUBSTR) {
5289 data->pos_min += min1;
5290 data->pos_delta += max1 - min1;
5291 if (max1 != min1 || is_inf)
5292 data->longest = &(data->longest_float);
5295 delta += max1 - min1;
5296 if (flags & SCF_DO_STCLASS_OR) {
5297 ssc_or(pRExC_state, data->start_class, &accum);
5299 ssc_and(pRExC_state, data->start_class, and_withp);
5300 flags &= ~SCF_DO_STCLASS;
5303 else if (flags & SCF_DO_STCLASS_AND) {
5305 ssc_and(pRExC_state, data->start_class, &accum);
5306 flags &= ~SCF_DO_STCLASS;
5309 /* Switch to OR mode: cache the old value of
5310 * data->start_class */
5312 StructCopy(data->start_class, and_withp, regnode_ssc);
5313 flags &= ~SCF_DO_STCLASS_AND;
5314 StructCopy(&accum, data->start_class, regnode_ssc);
5315 flags |= SCF_DO_STCLASS_OR;
5322 else if (PL_regkind[OP(scan)] == TRIE) {
5323 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5326 min += trie->minlen;
5327 delta += (trie->maxlen - trie->minlen);
5328 flags &= ~SCF_DO_STCLASS; /* xxx */
5329 if (flags & SCF_DO_SUBSTR) {
5330 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect
5332 data->pos_min += trie->minlen;
5333 data->pos_delta += (trie->maxlen - trie->minlen);
5334 if (trie->maxlen != trie->minlen)
5335 data->longest = &(data->longest_float);
5337 if (trie->jump) /* no more substrings -- for now /grr*/
5338 flags &= ~SCF_DO_SUBSTR;
5340 #endif /* old or new */
5341 #endif /* TRIE_STUDY_OPT */
5343 /* Else: zero-length, ignore. */
5344 scan = regnext(scan);
5346 /* If we are exiting a recursion we can unset its recursed bit
5347 * and allow ourselves to enter it again - no danger of an
5348 * infinite loop there.
5349 if (stopparen > -1 && recursed) {
5350 DEBUG_STUDYDATA("unset:", data,depth);
5351 PAREN_UNSET( recursed, stopparen);
5355 DEBUG_STUDYDATA("frame-end:",data,depth);
5356 DEBUG_PEEP("fend", scan, depth);
5357 /* restore previous context */
5360 stopparen = frame->stop;
5361 recursed_depth = frame->prev_recursed_depth;
5364 frame = frame->prev;
5365 goto fake_study_recurse;
5370 DEBUG_STUDYDATA("pre-fin:",data,depth);
5373 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5375 if (flags & SCF_DO_SUBSTR && is_inf)
5376 data->pos_delta = SSize_t_MAX - data->pos_min;
5377 if (is_par > (I32)U8_MAX)
5379 if (is_par && pars==1 && data) {
5380 data->flags |= SF_IN_PAR;
5381 data->flags &= ~SF_HAS_PAR;
5383 else if (pars && data) {
5384 data->flags |= SF_HAS_PAR;
5385 data->flags &= ~SF_IN_PAR;
5387 if (flags & SCF_DO_STCLASS_OR)
5388 ssc_and(pRExC_state, data->start_class, and_withp);
5389 if (flags & SCF_TRIE_RESTUDY)
5390 data->flags |= SCF_TRIE_RESTUDY;
5392 DEBUG_STUDYDATA("post-fin:",data,depth);
5395 SSize_t final_minlen= min < stopmin ? min : stopmin;
5397 if (RExC_maxlen < final_minlen + delta) {
5398 RExC_maxlen = final_minlen + delta;
5401 return final_minlen;
5407 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5409 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5411 PERL_ARGS_ASSERT_ADD_DATA;
5413 Renewc(RExC_rxi->data,
5414 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5415 char, struct reg_data);
5417 Renew(RExC_rxi->data->what, count + n, U8);
5419 Newx(RExC_rxi->data->what, n, U8);
5420 RExC_rxi->data->count = count + n;
5421 Copy(s, RExC_rxi->data->what + count, n, U8);
5425 /*XXX: todo make this not included in a non debugging perl */
5426 #ifndef PERL_IN_XSUB_RE
5428 Perl_reginitcolors(pTHX)
5431 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5433 char *t = savepv(s);
5437 t = strchr(t, '\t');
5443 PL_colors[i] = t = (char *)"";
5448 PL_colors[i++] = (char *)"";
5455 #ifdef TRIE_STUDY_OPT
5456 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5459 (data.flags & SCF_TRIE_RESTUDY) \
5467 #define CHECK_RESTUDY_GOTO_butfirst
5471 * pregcomp - compile a regular expression into internal code
5473 * Decides which engine's compiler to call based on the hint currently in
5477 #ifndef PERL_IN_XSUB_RE
5479 /* return the currently in-scope regex engine (or the default if none) */
5481 regexp_engine const *
5482 Perl_current_re_engine(pTHX)
5486 if (IN_PERL_COMPILETIME) {
5487 HV * const table = GvHV(PL_hintgv);
5490 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5491 return &PL_core_reg_engine;
5492 ptr = hv_fetchs(table, "regcomp", FALSE);
5493 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5494 return &PL_core_reg_engine;
5495 return INT2PTR(regexp_engine*,SvIV(*ptr));
5499 if (!PL_curcop->cop_hints_hash)
5500 return &PL_core_reg_engine;
5501 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5502 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5503 return &PL_core_reg_engine;
5504 return INT2PTR(regexp_engine*,SvIV(ptr));
5510 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5513 regexp_engine const *eng = current_re_engine();
5514 GET_RE_DEBUG_FLAGS_DECL;
5516 PERL_ARGS_ASSERT_PREGCOMP;
5518 /* Dispatch a request to compile a regexp to correct regexp engine. */
5520 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5523 return CALLREGCOMP_ENG(eng, pattern, flags);
5527 /* public(ish) entry point for the perl core's own regex compiling code.
5528 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5529 * pattern rather than a list of OPs, and uses the internal engine rather
5530 * than the current one */
5533 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5535 SV *pat = pattern; /* defeat constness! */
5536 PERL_ARGS_ASSERT_RE_COMPILE;
5537 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5538 #ifdef PERL_IN_XSUB_RE
5541 &PL_core_reg_engine,
5543 NULL, NULL, rx_flags, 0);
5547 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5548 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5549 * point to the realloced string and length.
5551 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5555 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5556 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5558 U8 *const src = (U8*)*pat_p;
5561 STRLEN s = 0, d = 0;
5563 GET_RE_DEBUG_FLAGS_DECL;
5565 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5566 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5568 Newx(dst, *plen_p * 2 + 1, U8);
5570 while (s < *plen_p) {
5571 if (NATIVE_BYTE_IS_INVARIANT(src[s]))
5574 dst[d++] = UTF8_EIGHT_BIT_HI(src[s]);
5575 dst[d] = UTF8_EIGHT_BIT_LO(src[s]);
5577 if (n < num_code_blocks) {
5578 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5579 pRExC_state->code_blocks[n].start = d;
5580 assert(dst[d] == '(');
5583 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5584 pRExC_state->code_blocks[n].end = d;
5585 assert(dst[d] == ')');
5595 *pat_p = (char*) dst;
5597 RExC_orig_utf8 = RExC_utf8 = 1;
5602 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5603 * while recording any code block indices, and handling overloading,
5604 * nested qr// objects etc. If pat is null, it will allocate a new
5605 * string, or just return the first arg, if there's only one.
5607 * Returns the malloced/updated pat.
5608 * patternp and pat_count is the array of SVs to be concatted;
5609 * oplist is the optional list of ops that generated the SVs;
5610 * recompile_p is a pointer to a boolean that will be set if
5611 * the regex will need to be recompiled.
5612 * delim, if non-null is an SV that will be inserted between each element
5616 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5617 SV *pat, SV ** const patternp, int pat_count,
5618 OP *oplist, bool *recompile_p, SV *delim)
5622 bool use_delim = FALSE;
5623 bool alloced = FALSE;
5625 /* if we know we have at least two args, create an empty string,
5626 * then concatenate args to that. For no args, return an empty string */
5627 if (!pat && pat_count != 1) {
5628 pat = newSVpvn("", 0);
5633 for (svp = patternp; svp < patternp + pat_count; svp++) {
5636 STRLEN orig_patlen = 0;
5638 SV *msv = use_delim ? delim : *svp;
5639 if (!msv) msv = &PL_sv_undef;
5641 /* if we've got a delimiter, we go round the loop twice for each
5642 * svp slot (except the last), using the delimiter the second
5651 if (SvTYPE(msv) == SVt_PVAV) {
5652 /* we've encountered an interpolated array within
5653 * the pattern, e.g. /...@a..../. Expand the list of elements,
5654 * then recursively append elements.
5655 * The code in this block is based on S_pushav() */
5657 AV *const av = (AV*)msv;
5658 const SSize_t maxarg = AvFILL(av) + 1;
5662 assert(oplist->op_type == OP_PADAV
5663 || oplist->op_type == OP_RV2AV);
5664 oplist = oplist->op_sibling;;
5667 if (SvRMAGICAL(av)) {
5670 Newx(array, maxarg, SV*);
5672 for (i=0; i < maxarg; i++) {
5673 SV ** const svp = av_fetch(av, i, FALSE);
5674 array[i] = svp ? *svp : &PL_sv_undef;
5678 array = AvARRAY(av);
5680 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5681 array, maxarg, NULL, recompile_p,
5683 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5689 /* we make the assumption here that each op in the list of
5690 * op_siblings maps to one SV pushed onto the stack,
5691 * except for code blocks, with have both an OP_NULL and
5693 * This allows us to match up the list of SVs against the
5694 * list of OPs to find the next code block.
5696 * Note that PUSHMARK PADSV PADSV ..
5698 * PADRANGE PADSV PADSV ..
5699 * so the alignment still works. */
5702 if (oplist->op_type == OP_NULL
5703 && (oplist->op_flags & OPf_SPECIAL))
5705 assert(n < pRExC_state->num_code_blocks);
5706 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5707 pRExC_state->code_blocks[n].block = oplist;
5708 pRExC_state->code_blocks[n].src_regex = NULL;
5711 oplist = oplist->op_sibling; /* skip CONST */
5714 oplist = oplist->op_sibling;;
5717 /* apply magic and QR overloading to arg */
5720 if (SvROK(msv) && SvAMAGIC(msv)) {
5721 SV *sv = AMG_CALLunary(msv, regexp_amg);
5725 if (SvTYPE(sv) != SVt_REGEXP)
5726 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5731 /* try concatenation overload ... */
5732 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5733 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5736 /* overloading involved: all bets are off over literal
5737 * code. Pretend we haven't seen it */
5738 pRExC_state->num_code_blocks -= n;
5742 /* ... or failing that, try "" overload */
5743 while (SvAMAGIC(msv)
5744 && (sv = AMG_CALLunary(msv, string_amg))
5748 && SvRV(msv) == SvRV(sv))
5753 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5757 /* this is a partially unrolled
5758 * sv_catsv_nomg(pat, msv);
5759 * that allows us to adjust code block indices if
5762 char *dst = SvPV_force_nomg(pat, dlen);
5764 if (SvUTF8(msv) && !SvUTF8(pat)) {
5765 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5766 sv_setpvn(pat, dst, dlen);
5769 sv_catsv_nomg(pat, msv);
5776 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5779 /* extract any code blocks within any embedded qr//'s */
5780 if (rx && SvTYPE(rx) == SVt_REGEXP
5781 && RX_ENGINE((REGEXP*)rx)->op_comp)
5784 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5785 if (ri->num_code_blocks) {
5787 /* the presence of an embedded qr// with code means
5788 * we should always recompile: the text of the
5789 * qr// may not have changed, but it may be a
5790 * different closure than last time */
5792 Renew(pRExC_state->code_blocks,
5793 pRExC_state->num_code_blocks + ri->num_code_blocks,
5794 struct reg_code_block);
5795 pRExC_state->num_code_blocks += ri->num_code_blocks;
5797 for (i=0; i < ri->num_code_blocks; i++) {
5798 struct reg_code_block *src, *dst;
5799 STRLEN offset = orig_patlen
5800 + ReANY((REGEXP *)rx)->pre_prefix;
5801 assert(n < pRExC_state->num_code_blocks);
5802 src = &ri->code_blocks[i];
5803 dst = &pRExC_state->code_blocks[n];
5804 dst->start = src->start + offset;
5805 dst->end = src->end + offset;
5806 dst->block = src->block;
5807 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5816 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5825 /* see if there are any run-time code blocks in the pattern.
5826 * False positives are allowed */
5829 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5830 char *pat, STRLEN plen)
5835 for (s = 0; s < plen; s++) {
5836 if (n < pRExC_state->num_code_blocks
5837 && s == pRExC_state->code_blocks[n].start)
5839 s = pRExC_state->code_blocks[n].end;
5843 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5845 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5847 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5854 /* Handle run-time code blocks. We will already have compiled any direct
5855 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5856 * copy of it, but with any literal code blocks blanked out and
5857 * appropriate chars escaped; then feed it into
5859 * eval "qr'modified_pattern'"
5863 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5867 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5869 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5870 * and merge them with any code blocks of the original regexp.
5872 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5873 * instead, just save the qr and return FALSE; this tells our caller that
5874 * the original pattern needs upgrading to utf8.
5878 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5879 char *pat, STRLEN plen)
5883 GET_RE_DEBUG_FLAGS_DECL;
5885 if (pRExC_state->runtime_code_qr) {
5886 /* this is the second time we've been called; this should
5887 * only happen if the main pattern got upgraded to utf8
5888 * during compilation; re-use the qr we compiled first time
5889 * round (which should be utf8 too)
5891 qr = pRExC_state->runtime_code_qr;
5892 pRExC_state->runtime_code_qr = NULL;
5893 assert(RExC_utf8 && SvUTF8(qr));
5899 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5903 /* determine how many extra chars we need for ' and \ escaping */
5904 for (s = 0; s < plen; s++) {
5905 if (pat[s] == '\'' || pat[s] == '\\')
5909 Newx(newpat, newlen, char);
5911 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5913 for (s = 0; s < plen; s++) {
5914 if (n < pRExC_state->num_code_blocks
5915 && s == pRExC_state->code_blocks[n].start)
5917 /* blank out literal code block */
5918 assert(pat[s] == '(');
5919 while (s <= pRExC_state->code_blocks[n].end) {
5927 if (pat[s] == '\'' || pat[s] == '\\')
5932 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5936 PerlIO_printf(Perl_debug_log,
5937 "%sre-parsing pattern for runtime code:%s %s\n",
5938 PL_colors[4],PL_colors[5],newpat);
5941 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5947 PUSHSTACKi(PERLSI_REQUIRE);
5948 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5949 * parsing qr''; normally only q'' does this. It also alters
5951 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5952 SvREFCNT_dec_NN(sv);
5957 SV * const errsv = ERRSV;
5958 if (SvTRUE_NN(errsv))
5960 Safefree(pRExC_state->code_blocks);
5961 /* use croak_sv ? */
5962 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
5965 assert(SvROK(qr_ref));
5967 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5968 /* the leaving below frees the tmp qr_ref.
5969 * Give qr a life of its own */
5977 if (!RExC_utf8 && SvUTF8(qr)) {
5978 /* first time through; the pattern got upgraded; save the
5979 * qr for the next time through */
5980 assert(!pRExC_state->runtime_code_qr);
5981 pRExC_state->runtime_code_qr = qr;
5986 /* extract any code blocks within the returned qr// */
5989 /* merge the main (r1) and run-time (r2) code blocks into one */
5991 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5992 struct reg_code_block *new_block, *dst;
5993 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5996 if (!r2->num_code_blocks) /* we guessed wrong */
5998 SvREFCNT_dec_NN(qr);
6003 r1->num_code_blocks + r2->num_code_blocks,
6004 struct reg_code_block);
6007 while ( i1 < r1->num_code_blocks
6008 || i2 < r2->num_code_blocks)
6010 struct reg_code_block *src;
6013 if (i1 == r1->num_code_blocks) {
6014 src = &r2->code_blocks[i2++];
6017 else if (i2 == r2->num_code_blocks)
6018 src = &r1->code_blocks[i1++];
6019 else if ( r1->code_blocks[i1].start
6020 < r2->code_blocks[i2].start)
6022 src = &r1->code_blocks[i1++];
6023 assert(src->end < r2->code_blocks[i2].start);
6026 assert( r1->code_blocks[i1].start
6027 > r2->code_blocks[i2].start);
6028 src = &r2->code_blocks[i2++];
6030 assert(src->end < r1->code_blocks[i1].start);
6033 assert(pat[src->start] == '(');
6034 assert(pat[src->end] == ')');
6035 dst->start = src->start;
6036 dst->end = src->end;
6037 dst->block = src->block;
6038 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6042 r1->num_code_blocks += r2->num_code_blocks;
6043 Safefree(r1->code_blocks);
6044 r1->code_blocks = new_block;
6047 SvREFCNT_dec_NN(qr);
6053 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6054 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6055 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6056 STRLEN longest_length, bool eol, bool meol)
6058 /* This is the common code for setting up the floating and fixed length
6059 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6060 * as to whether succeeded or not */
6065 if (! (longest_length
6066 || (eol /* Can't have SEOL and MULTI */
6067 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6069 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6070 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6075 /* copy the information about the longest from the reg_scan_data
6076 over to the program. */
6077 if (SvUTF8(sv_longest)) {
6078 *rx_utf8 = sv_longest;
6081 *rx_substr = sv_longest;
6084 /* end_shift is how many chars that must be matched that
6085 follow this item. We calculate it ahead of time as once the
6086 lookbehind offset is added in we lose the ability to correctly
6088 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6089 *rx_end_shift = ml - offset
6090 - longest_length + (SvTAIL(sv_longest) != 0)
6093 t = (eol/* Can't have SEOL and MULTI */
6094 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6095 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6101 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6102 * regular expression into internal code.
6103 * The pattern may be passed either as:
6104 * a list of SVs (patternp plus pat_count)
6105 * a list of OPs (expr)
6106 * If both are passed, the SV list is used, but the OP list indicates
6107 * which SVs are actually pre-compiled code blocks
6109 * The SVs in the list have magic and qr overloading applied to them (and
6110 * the list may be modified in-place with replacement SVs in the latter
6113 * If the pattern hasn't changed from old_re, then old_re will be
6116 * eng is the current engine. If that engine has an op_comp method, then
6117 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6118 * do the initial concatenation of arguments and pass on to the external
6121 * If is_bare_re is not null, set it to a boolean indicating whether the
6122 * arg list reduced (after overloading) to a single bare regex which has
6123 * been returned (i.e. /$qr/).
6125 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6127 * pm_flags contains the PMf_* flags, typically based on those from the
6128 * pm_flags field of the related PMOP. Currently we're only interested in
6129 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6131 * We can't allocate space until we know how big the compiled form will be,
6132 * but we can't compile it (and thus know how big it is) until we've got a
6133 * place to put the code. So we cheat: we compile it twice, once with code
6134 * generation turned off and size counting turned on, and once "for real".
6135 * This also means that we don't allocate space until we are sure that the
6136 * thing really will compile successfully, and we never have to move the
6137 * code and thus invalidate pointers into it. (Note that it has to be in
6138 * one piece because free() must be able to free it all.) [NB: not true in perl]
6140 * Beware that the optimization-preparation code in here knows about some
6141 * of the structure of the compiled regexp. [I'll say.]
6145 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6146 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6147 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6152 regexp_internal *ri;
6160 SV *code_blocksv = NULL;
6161 SV** new_patternp = patternp;
6163 /* these are all flags - maybe they should be turned
6164 * into a single int with different bit masks */
6165 I32 sawlookahead = 0;
6170 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6172 bool runtime_code = 0;
6174 RExC_state_t RExC_state;
6175 RExC_state_t * const pRExC_state = &RExC_state;
6176 #ifdef TRIE_STUDY_OPT
6178 RExC_state_t copyRExC_state;
6180 GET_RE_DEBUG_FLAGS_DECL;
6182 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6184 DEBUG_r(if (!PL_colorset) reginitcolors());
6186 #ifndef PERL_IN_XSUB_RE
6187 /* Initialize these here instead of as-needed, as is quick and avoids
6188 * having to test them each time otherwise */
6189 if (! PL_AboveLatin1) {
6190 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6191 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6192 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6193 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6194 PL_HasMultiCharFold =
6195 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6199 pRExC_state->code_blocks = NULL;
6200 pRExC_state->num_code_blocks = 0;
6203 *is_bare_re = FALSE;
6205 if (expr && (expr->op_type == OP_LIST ||
6206 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6207 /* allocate code_blocks if needed */
6211 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
6212 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6213 ncode++; /* count of DO blocks */
6215 pRExC_state->num_code_blocks = ncode;
6216 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6221 /* compile-time pattern with just OP_CONSTs and DO blocks */
6226 /* find how many CONSTs there are */
6229 if (expr->op_type == OP_CONST)
6232 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6233 if (o->op_type == OP_CONST)
6237 /* fake up an SV array */
6239 assert(!new_patternp);
6240 Newx(new_patternp, n, SV*);
6241 SAVEFREEPV(new_patternp);
6245 if (expr->op_type == OP_CONST)
6246 new_patternp[n] = cSVOPx_sv(expr);
6248 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6249 if (o->op_type == OP_CONST)
6250 new_patternp[n++] = cSVOPo_sv;
6255 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6256 "Assembling pattern from %d elements%s\n", pat_count,
6257 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6259 /* set expr to the first arg op */
6261 if (pRExC_state->num_code_blocks
6262 && expr->op_type != OP_CONST)
6264 expr = cLISTOPx(expr)->op_first;
6265 assert( expr->op_type == OP_PUSHMARK
6266 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6267 || expr->op_type == OP_PADRANGE);
6268 expr = expr->op_sibling;
6271 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6272 expr, &recompile, NULL);
6274 /* handle bare (possibly after overloading) regex: foo =~ $re */
6279 if (SvTYPE(re) == SVt_REGEXP) {
6283 Safefree(pRExC_state->code_blocks);
6284 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6285 "Precompiled pattern%s\n",
6286 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6292 exp = SvPV_nomg(pat, plen);
6294 if (!eng->op_comp) {
6295 if ((SvUTF8(pat) && IN_BYTES)
6296 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6298 /* make a temporary copy; either to convert to bytes,
6299 * or to avoid repeating get-magic / overloaded stringify */
6300 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6301 (IN_BYTES ? 0 : SvUTF8(pat)));
6303 Safefree(pRExC_state->code_blocks);
6304 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6307 /* ignore the utf8ness if the pattern is 0 length */
6308 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6310 /* 'use utf8' in the program indicates Unicode rules are wanted */
6311 RExC_uni_semantics = (PL_hints & HINT_UTF8);
6313 RExC_contains_locale = 0;
6314 RExC_contains_i = 0;
6315 pRExC_state->runtime_code_qr = NULL;
6318 SV *dsv= sv_newmortal();
6319 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6320 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6321 PL_colors[4],PL_colors[5],s);
6325 /* we jump here if we upgrade the pattern to utf8 and have to
6328 if ((pm_flags & PMf_USE_RE_EVAL)
6329 /* this second condition covers the non-regex literal case,
6330 * i.e. $foo =~ '(?{})'. */
6331 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6333 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6335 /* return old regex if pattern hasn't changed */
6336 /* XXX: note in the below we have to check the flags as well as the
6339 * Things get a touch tricky as we have to compare the utf8 flag
6340 * independently from the compile flags. */
6344 && !!RX_UTF8(old_re) == !!RExC_utf8
6345 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6346 && RX_PRECOMP(old_re)
6347 && RX_PRELEN(old_re) == plen
6348 && memEQ(RX_PRECOMP(old_re), exp, plen)
6349 && !runtime_code /* with runtime code, always recompile */ )
6351 Safefree(pRExC_state->code_blocks);
6355 rx_flags = orig_rx_flags;
6357 if (rx_flags & PMf_FOLD) {
6358 RExC_contains_i = 1;
6360 if (initial_charset == REGEX_LOCALE_CHARSET) {
6361 RExC_contains_locale = 1;
6363 else if ((RExC_utf8 || RExC_uni_semantics)
6364 && initial_charset == REGEX_DEPENDS_CHARSET)
6367 /* Set to use unicode semantics if has the 'depends' charset specified,
6368 * and either the pattern is in utf8 (as it means unicode when utf8),
6369 * or we already know we want unicode rules */
6370 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6374 RExC_flags = rx_flags;
6375 RExC_pm_flags = pm_flags;
6378 if (TAINTING_get && TAINT_get)
6379 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6381 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6382 /* whoops, we have a non-utf8 pattern, whilst run-time code
6383 * got compiled as utf8. Try again with a utf8 pattern */
6384 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6385 pRExC_state->num_code_blocks);
6386 goto redo_first_pass;
6389 assert(!pRExC_state->runtime_code_qr);
6395 RExC_in_lookbehind = 0;
6396 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6398 RExC_override_recoding = 0;
6399 RExC_in_multi_char_class = 0;
6401 /* First pass: determine size, legality. */
6404 RExC_end = exp + plen;
6409 RExC_emit = (regnode *) &RExC_emit_dummy;
6410 RExC_whilem_seen = 0;
6411 RExC_open_parens = NULL;
6412 RExC_close_parens = NULL;
6414 RExC_paren_names = NULL;
6416 RExC_paren_name_list = NULL;
6418 RExC_recurse = NULL;
6419 RExC_study_chunk_recursed = NULL;
6420 RExC_study_chunk_recursed_bytes= 0;
6421 RExC_recurse_count = 0;
6422 pRExC_state->code_index = 0;
6424 #if 0 /* REGC() is (currently) a NOP at the first pass.
6425 * Clever compilers notice this and complain. --jhi */
6426 REGC((U8)REG_MAGIC, (char*)RExC_emit);
6429 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6431 RExC_lastparse=NULL;
6433 /* reg may croak on us, not giving us a chance to free
6434 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6435 need it to survive as long as the regexp (qr/(?{})/).
6436 We must check that code_blocksv is not already set, because we may
6437 have jumped back to restart the sizing pass. */
6438 if (pRExC_state->code_blocks && !code_blocksv) {
6439 code_blocksv = newSV_type(SVt_PV);
6440 SAVEFREESV(code_blocksv);
6441 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6442 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6444 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6445 /* It's possible to write a regexp in ascii that represents Unicode
6446 codepoints outside of the byte range, such as via \x{100}. If we
6447 detect such a sequence we have to convert the entire pattern to utf8
6448 and then recompile, as our sizing calculation will have been based
6449 on 1 byte == 1 character, but we will need to use utf8 to encode
6450 at least some part of the pattern, and therefore must convert the whole
6453 if (flags & RESTART_UTF8) {
6454 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6455 pRExC_state->num_code_blocks);
6456 goto redo_first_pass;
6458 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6461 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6464 PerlIO_printf(Perl_debug_log,
6465 "Required size %"IVdf" nodes\n"
6466 "Starting second pass (creation)\n",
6469 RExC_lastparse=NULL;
6472 /* The first pass could have found things that force Unicode semantics */
6473 if ((RExC_utf8 || RExC_uni_semantics)
6474 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6476 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6479 /* Small enough for pointer-storage convention?
6480 If extralen==0, this means that we will not need long jumps. */
6481 if (RExC_size >= 0x10000L && RExC_extralen)
6482 RExC_size += RExC_extralen;
6485 if (RExC_whilem_seen > 15)
6486 RExC_whilem_seen = 15;
6488 /* Allocate space and zero-initialize. Note, the two step process
6489 of zeroing when in debug mode, thus anything assigned has to
6490 happen after that */
6491 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6493 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6494 char, regexp_internal);
6495 if ( r == NULL || ri == NULL )
6496 FAIL("Regexp out of space");
6498 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6499 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6502 /* bulk initialize base fields with 0. */
6503 Zero(ri, sizeof(regexp_internal), char);
6506 /* non-zero initialization begins here */
6509 r->extflags = rx_flags;
6510 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6512 if (pm_flags & PMf_IS_QR) {
6513 ri->code_blocks = pRExC_state->code_blocks;
6514 ri->num_code_blocks = pRExC_state->num_code_blocks;
6519 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6520 if (pRExC_state->code_blocks[n].src_regex)
6521 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6522 SAVEFREEPV(pRExC_state->code_blocks);
6526 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6527 bool has_charset = (get_regex_charset(r->extflags)
6528 != REGEX_DEPENDS_CHARSET);
6530 /* The caret is output if there are any defaults: if not all the STD
6531 * flags are set, or if no character set specifier is needed */
6533 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6535 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6536 == REG_RUN_ON_COMMENT_SEEN);
6537 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6538 >> RXf_PMf_STD_PMMOD_SHIFT);
6539 const char *fptr = STD_PAT_MODS; /*"msix"*/
6541 /* Allocate for the worst case, which is all the std flags are turned
6542 * on. If more precision is desired, we could do a population count of
6543 * the flags set. This could be done with a small lookup table, or by
6544 * shifting, masking and adding, or even, when available, assembly
6545 * language for a machine-language population count.
6546 * We never output a minus, as all those are defaults, so are
6547 * covered by the caret */
6548 const STRLEN wraplen = plen + has_p + has_runon
6549 + has_default /* If needs a caret */
6551 /* If needs a character set specifier */
6552 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6553 + (sizeof(STD_PAT_MODS) - 1)
6554 + (sizeof("(?:)") - 1);
6556 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6557 r->xpv_len_u.xpvlenu_pv = p;
6559 SvFLAGS(rx) |= SVf_UTF8;
6562 /* If a default, cover it using the caret */
6564 *p++= DEFAULT_PAT_MOD;
6568 const char* const name = get_regex_charset_name(r->extflags, &len);
6569 Copy(name, p, len, char);
6573 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6576 while((ch = *fptr++)) {
6584 Copy(RExC_precomp, p, plen, char);
6585 assert ((RX_WRAPPED(rx) - p) < 16);
6586 r->pre_prefix = p - RX_WRAPPED(rx);
6592 SvCUR_set(rx, p - RX_WRAPPED(rx));
6596 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6598 /* setup various meta data about recursion, this all requires
6599 * RExC_npar to be correctly set, and a bit later on we clear it */
6600 if (RExC_seen & REG_RECURSE_SEEN) {
6601 Newxz(RExC_open_parens, RExC_npar,regnode *);
6602 SAVEFREEPV(RExC_open_parens);
6603 Newxz(RExC_close_parens,RExC_npar,regnode *);
6604 SAVEFREEPV(RExC_close_parens);
6606 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6607 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6608 * So its 1 if there are no parens. */
6609 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6610 ((RExC_npar & 0x07) != 0);
6611 Newx(RExC_study_chunk_recursed,
6612 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6613 SAVEFREEPV(RExC_study_chunk_recursed);
6616 /* Useful during FAIL. */
6617 #ifdef RE_TRACK_PATTERN_OFFSETS
6618 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6619 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6620 "%s %"UVuf" bytes for offset annotations.\n",
6621 ri->u.offsets ? "Got" : "Couldn't get",
6622 (UV)((2*RExC_size+1) * sizeof(U32))));
6624 SetProgLen(ri,RExC_size);
6629 /* Second pass: emit code. */
6630 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6631 RExC_pm_flags = pm_flags;
6633 RExC_end = exp + plen;
6636 RExC_emit_start = ri->program;
6637 RExC_emit = ri->program;
6638 RExC_emit_bound = ri->program + RExC_size + 1;
6639 pRExC_state->code_index = 0;
6641 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6642 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6644 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6646 /* XXXX To minimize changes to RE engine we always allocate
6647 3-units-long substrs field. */
6648 Newx(r->substrs, 1, struct reg_substr_data);
6649 if (RExC_recurse_count) {
6650 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6651 SAVEFREEPV(RExC_recurse);
6655 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6656 Zero(r->substrs, 1, struct reg_substr_data);
6657 if (RExC_study_chunk_recursed)
6658 Zero(RExC_study_chunk_recursed,
6659 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6661 #ifdef TRIE_STUDY_OPT
6663 StructCopy(&zero_scan_data, &data, scan_data_t);
6664 copyRExC_state = RExC_state;
6667 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6669 RExC_state = copyRExC_state;
6670 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6671 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6673 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6674 StructCopy(&zero_scan_data, &data, scan_data_t);
6677 StructCopy(&zero_scan_data, &data, scan_data_t);
6680 /* Dig out information for optimizations. */
6681 r->extflags = RExC_flags; /* was pm_op */
6682 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6685 SvUTF8_on(rx); /* Unicode in it? */
6686 ri->regstclass = NULL;
6687 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6688 r->intflags |= PREGf_NAUGHTY;
6689 scan = ri->program + 1; /* First BRANCH. */
6691 /* testing for BRANCH here tells us whether there is "must appear"
6692 data in the pattern. If there is then we can use it for optimisations */
6693 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6696 STRLEN longest_float_length, longest_fixed_length;
6697 regnode_ssc ch_class; /* pointed to by data */
6699 SSize_t last_close = 0; /* pointed to by data */
6700 regnode *first= scan;
6701 regnode *first_next= regnext(first);
6703 * Skip introductions and multiplicators >= 1
6704 * so that we can extract the 'meat' of the pattern that must
6705 * match in the large if() sequence following.
6706 * NOTE that EXACT is NOT covered here, as it is normally
6707 * picked up by the optimiser separately.
6709 * This is unfortunate as the optimiser isnt handling lookahead
6710 * properly currently.
6713 while ((OP(first) == OPEN && (sawopen = 1)) ||
6714 /* An OR of *one* alternative - should not happen now. */
6715 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6716 /* for now we can't handle lookbehind IFMATCH*/
6717 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6718 (OP(first) == PLUS) ||
6719 (OP(first) == MINMOD) ||
6720 /* An {n,m} with n>0 */
6721 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6722 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6725 * the only op that could be a regnode is PLUS, all the rest
6726 * will be regnode_1 or regnode_2.
6728 * (yves doesn't think this is true)
6730 if (OP(first) == PLUS)
6733 if (OP(first) == MINMOD)
6735 first += regarglen[OP(first)];
6737 first = NEXTOPER(first);
6738 first_next= regnext(first);
6741 /* Starting-point info. */
6743 DEBUG_PEEP("first:",first,0);
6744 /* Ignore EXACT as we deal with it later. */
6745 if (PL_regkind[OP(first)] == EXACT) {
6746 if (OP(first) == EXACT)
6747 NOOP; /* Empty, get anchored substr later. */
6749 ri->regstclass = first;
6752 else if (PL_regkind[OP(first)] == TRIE &&
6753 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6756 /* this can happen only on restudy */
6757 if ( OP(first) == TRIE ) {
6758 struct regnode_1 *trieop = (struct regnode_1 *)
6759 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6760 StructCopy(first,trieop,struct regnode_1);
6761 trie_op=(regnode *)trieop;
6763 struct regnode_charclass *trieop = (struct regnode_charclass *)
6764 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6765 StructCopy(first,trieop,struct regnode_charclass);
6766 trie_op=(regnode *)trieop;
6769 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6770 ri->regstclass = trie_op;
6773 else if (REGNODE_SIMPLE(OP(first)))
6774 ri->regstclass = first;
6775 else if (PL_regkind[OP(first)] == BOUND ||
6776 PL_regkind[OP(first)] == NBOUND)
6777 ri->regstclass = first;
6778 else if (PL_regkind[OP(first)] == BOL) {
6779 r->intflags |= (OP(first) == MBOL
6781 : (OP(first) == SBOL
6784 first = NEXTOPER(first);
6787 else if (OP(first) == GPOS) {
6788 r->intflags |= PREGf_ANCH_GPOS;
6789 first = NEXTOPER(first);
6792 else if ((!sawopen || !RExC_sawback) &&
6793 (OP(first) == STAR &&
6794 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6795 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
6797 /* turn .* into ^.* with an implied $*=1 */
6799 (OP(NEXTOPER(first)) == REG_ANY)
6802 r->intflags |= (type | PREGf_IMPLICIT);
6803 first = NEXTOPER(first);
6806 if (sawplus && !sawminmod && !sawlookahead
6807 && (!sawopen || !RExC_sawback)
6808 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6809 /* x+ must match at the 1st pos of run of x's */
6810 r->intflags |= PREGf_SKIP;
6812 /* Scan is after the zeroth branch, first is atomic matcher. */
6813 #ifdef TRIE_STUDY_OPT
6816 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6817 (IV)(first - scan + 1))
6821 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6822 (IV)(first - scan + 1))
6828 * If there's something expensive in the r.e., find the
6829 * longest literal string that must appear and make it the
6830 * regmust. Resolve ties in favor of later strings, since
6831 * the regstart check works with the beginning of the r.e.
6832 * and avoiding duplication strengthens checking. Not a
6833 * strong reason, but sufficient in the absence of others.
6834 * [Now we resolve ties in favor of the earlier string if
6835 * it happens that c_offset_min has been invalidated, since the
6836 * earlier string may buy us something the later one won't.]
6839 data.longest_fixed = newSVpvs("");
6840 data.longest_float = newSVpvs("");
6841 data.last_found = newSVpvs("");
6842 data.longest = &(data.longest_fixed);
6843 ENTER_with_name("study_chunk");
6844 SAVEFREESV(data.longest_fixed);
6845 SAVEFREESV(data.longest_float);
6846 SAVEFREESV(data.last_found);
6848 if (!ri->regstclass) {
6849 ssc_init(pRExC_state, &ch_class);
6850 data.start_class = &ch_class;
6851 stclass_flag = SCF_DO_STCLASS_AND;
6852 } else /* XXXX Check for BOUND? */
6854 data.last_closep = &last_close;
6857 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
6858 scan + RExC_size, /* Up to end */
6860 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6861 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6865 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6868 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6869 && data.last_start_min == 0 && data.last_end > 0
6870 && !RExC_seen_zerolen
6871 && !(RExC_seen & REG_VERBARG_SEEN)
6872 && !(RExC_seen & REG_GPOS_SEEN)
6874 r->extflags |= RXf_CHECK_ALL;
6876 scan_commit(pRExC_state, &data,&minlen,0);
6878 longest_float_length = CHR_SVLEN(data.longest_float);
6880 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6881 && data.offset_fixed == data.offset_float_min
6882 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6883 && S_setup_longest (aTHX_ pRExC_state,
6887 &(r->float_end_shift),
6888 data.lookbehind_float,
6889 data.offset_float_min,
6891 longest_float_length,
6892 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6893 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6895 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6896 r->float_max_offset = data.offset_float_max;
6897 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
6898 r->float_max_offset -= data.lookbehind_float;
6899 SvREFCNT_inc_simple_void_NN(data.longest_float);
6902 r->float_substr = r->float_utf8 = NULL;
6903 longest_float_length = 0;
6906 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6908 if (S_setup_longest (aTHX_ pRExC_state,
6910 &(r->anchored_utf8),
6911 &(r->anchored_substr),
6912 &(r->anchored_end_shift),
6913 data.lookbehind_fixed,
6916 longest_fixed_length,
6917 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6918 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6920 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6921 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6924 r->anchored_substr = r->anchored_utf8 = NULL;
6925 longest_fixed_length = 0;
6927 LEAVE_with_name("study_chunk");
6930 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6931 ri->regstclass = NULL;
6933 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6935 && ! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
6936 && !ssc_is_anything(data.start_class))
6938 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
6940 ssc_finalize(pRExC_state, data.start_class);
6942 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
6943 StructCopy(data.start_class,
6944 (regnode_ssc*)RExC_rxi->data->data[n],
6946 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6947 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6948 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6949 regprop(r, sv, (regnode*)data.start_class);
6950 PerlIO_printf(Perl_debug_log,
6951 "synthetic stclass \"%s\".\n",
6952 SvPVX_const(sv));});
6953 data.start_class = NULL;
6956 /* A temporary algorithm prefers floated substr to fixed one to dig
6958 if (longest_fixed_length > longest_float_length) {
6959 r->check_end_shift = r->anchored_end_shift;
6960 r->check_substr = r->anchored_substr;
6961 r->check_utf8 = r->anchored_utf8;
6962 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6963 if (r->intflags & PREGf_ANCH_SINGLE)
6964 r->intflags |= PREGf_NOSCAN;
6967 r->check_end_shift = r->float_end_shift;
6968 r->check_substr = r->float_substr;
6969 r->check_utf8 = r->float_utf8;
6970 r->check_offset_min = r->float_min_offset;
6971 r->check_offset_max = r->float_max_offset;
6973 if ((r->check_substr || r->check_utf8) ) {
6974 r->extflags |= RXf_USE_INTUIT;
6975 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6976 r->extflags |= RXf_INTUIT_TAIL;
6978 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6979 if ( (STRLEN)minlen < longest_float_length )
6980 minlen= longest_float_length;
6981 if ( (STRLEN)minlen < longest_fixed_length )
6982 minlen= longest_fixed_length;
6986 /* Several toplevels. Best we can is to set minlen. */
6988 regnode_ssc ch_class;
6989 SSize_t last_close = 0;
6991 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6993 scan = ri->program + 1;
6994 ssc_init(pRExC_state, &ch_class);
6995 data.start_class = &ch_class;
6996 data.last_closep = &last_close;
6999 minlen = study_chunk(pRExC_state,
7000 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7001 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7002 ? SCF_TRIE_DOING_RESTUDY
7006 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7008 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7009 = r->float_substr = r->float_utf8 = NULL;
7011 if (! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
7012 && ! ssc_is_anything(data.start_class))
7014 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7016 ssc_finalize(pRExC_state, data.start_class);
7018 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7019 StructCopy(data.start_class,
7020 (regnode_ssc*)RExC_rxi->data->data[n],
7022 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7023 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7024 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7025 regprop(r, sv, (regnode*)data.start_class);
7026 PerlIO_printf(Perl_debug_log,
7027 "synthetic stclass \"%s\".\n",
7028 SvPVX_const(sv));});
7029 data.start_class = NULL;
7033 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7034 the "real" pattern. */
7036 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%ld\n",
7037 (IV)minlen, (IV)r->minlen, RExC_maxlen);
7039 r->minlenret = minlen;
7040 if (r->minlen < minlen)
7045 if (RExC_seen & REG_GPOS_SEEN)
7046 r->intflags |= PREGf_GPOS_SEEN;
7047 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7048 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7050 if (pRExC_state->num_code_blocks)
7051 r->extflags |= RXf_EVAL_SEEN;
7052 if (RExC_seen & REG_CANY_SEEN)
7053 r->intflags |= PREGf_CANY_SEEN;
7054 if (RExC_seen & REG_VERBARG_SEEN)
7056 r->intflags |= PREGf_VERBARG_SEEN;
7057 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7059 if (RExC_seen & REG_CUTGROUP_SEEN)
7060 r->intflags |= PREGf_CUTGROUP_SEEN;
7061 if (pm_flags & PMf_USE_RE_EVAL)
7062 r->intflags |= PREGf_USE_RE_EVAL;
7063 if (RExC_paren_names)
7064 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7066 RXp_PAREN_NAMES(r) = NULL;
7068 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)
7069 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7071 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7072 * so it can be used in pp.c */
7073 if (r->intflags & PREGf_ANCH)
7074 r->extflags |= RXf_IS_ANCHORED;
7077 regnode *first = ri->program + 1;
7079 regnode *next = NEXTOPER(first);
7082 if (PL_regkind[fop] == NOTHING && nop == END)
7083 r->extflags |= RXf_NULL;
7084 else if (PL_regkind[fop] == BOL && nop == END)
7085 r->extflags |= RXf_START_ONLY;
7086 else if (fop == PLUS
7087 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7088 && OP(regnext(first)) == END)
7089 r->extflags |= RXf_WHITE;
7090 else if ( r->extflags & RXf_SPLIT
7092 && STR_LEN(first) == 1
7093 && *(STRING(first)) == ' '
7094 && OP(regnext(first)) == END )
7095 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7099 if (RExC_paren_names) {
7100 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7101 ri->data->data[ri->name_list_idx]
7102 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7105 ri->name_list_idx = 0;
7107 if (RExC_recurse_count) {
7108 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7109 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7110 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7113 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7114 /* assume we don't need to swap parens around before we match */
7118 PerlIO_printf(Perl_debug_log,"Final program:\n");
7121 #ifdef RE_TRACK_PATTERN_OFFSETS
7122 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7123 const STRLEN len = ri->u.offsets[0];
7125 GET_RE_DEBUG_FLAGS_DECL;
7126 PerlIO_printf(Perl_debug_log,
7127 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7128 for (i = 1; i <= len; i++) {
7129 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7130 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7131 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7133 PerlIO_printf(Perl_debug_log, "\n");
7138 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7139 * by setting the regexp SV to readonly-only instead. If the
7140 * pattern's been recompiled, the USEDness should remain. */
7141 if (old_re && SvREADONLY(old_re))
7149 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7152 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7154 PERL_UNUSED_ARG(value);
7156 if (flags & RXapif_FETCH) {
7157 return reg_named_buff_fetch(rx, key, flags);
7158 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7159 Perl_croak_no_modify();
7161 } else if (flags & RXapif_EXISTS) {
7162 return reg_named_buff_exists(rx, key, flags)
7165 } else if (flags & RXapif_REGNAMES) {
7166 return reg_named_buff_all(rx, flags);
7167 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7168 return reg_named_buff_scalar(rx, flags);
7170 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7176 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7179 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7180 PERL_UNUSED_ARG(lastkey);
7182 if (flags & RXapif_FIRSTKEY)
7183 return reg_named_buff_firstkey(rx, flags);
7184 else if (flags & RXapif_NEXTKEY)
7185 return reg_named_buff_nextkey(rx, flags);
7187 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7194 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7197 AV *retarray = NULL;
7199 struct regexp *const rx = ReANY(r);
7201 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7203 if (flags & RXapif_ALL)
7206 if (rx && RXp_PAREN_NAMES(rx)) {
7207 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7210 SV* sv_dat=HeVAL(he_str);
7211 I32 *nums=(I32*)SvPVX(sv_dat);
7212 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7213 if ((I32)(rx->nparens) >= nums[i]
7214 && rx->offs[nums[i]].start != -1
7215 && rx->offs[nums[i]].end != -1)
7218 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7223 ret = newSVsv(&PL_sv_undef);
7226 av_push(retarray, ret);
7229 return newRV_noinc(MUTABLE_SV(retarray));
7236 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7239 struct regexp *const rx = ReANY(r);
7241 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7243 if (rx && RXp_PAREN_NAMES(rx)) {
7244 if (flags & RXapif_ALL) {
7245 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7247 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7249 SvREFCNT_dec_NN(sv);
7261 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7263 struct regexp *const rx = ReANY(r);
7265 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7267 if ( rx && RXp_PAREN_NAMES(rx) ) {
7268 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7270 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7277 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7279 struct regexp *const rx = ReANY(r);
7280 GET_RE_DEBUG_FLAGS_DECL;
7282 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7284 if (rx && RXp_PAREN_NAMES(rx)) {
7285 HV *hv = RXp_PAREN_NAMES(rx);
7287 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7290 SV* sv_dat = HeVAL(temphe);
7291 I32 *nums = (I32*)SvPVX(sv_dat);
7292 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7293 if ((I32)(rx->lastparen) >= nums[i] &&
7294 rx->offs[nums[i]].start != -1 &&
7295 rx->offs[nums[i]].end != -1)
7301 if (parno || flags & RXapif_ALL) {
7302 return newSVhek(HeKEY_hek(temphe));
7310 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7315 struct regexp *const rx = ReANY(r);
7317 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7319 if (rx && RXp_PAREN_NAMES(rx)) {
7320 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7321 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7322 } else if (flags & RXapif_ONE) {
7323 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7324 av = MUTABLE_AV(SvRV(ret));
7325 length = av_len(av);
7326 SvREFCNT_dec_NN(ret);
7327 return newSViv(length + 1);
7329 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7334 return &PL_sv_undef;
7338 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7340 struct regexp *const rx = ReANY(r);
7343 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7345 if (rx && RXp_PAREN_NAMES(rx)) {
7346 HV *hv= RXp_PAREN_NAMES(rx);
7348 (void)hv_iterinit(hv);
7349 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7352 SV* sv_dat = HeVAL(temphe);
7353 I32 *nums = (I32*)SvPVX(sv_dat);
7354 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7355 if ((I32)(rx->lastparen) >= nums[i] &&
7356 rx->offs[nums[i]].start != -1 &&
7357 rx->offs[nums[i]].end != -1)
7363 if (parno || flags & RXapif_ALL) {
7364 av_push(av, newSVhek(HeKEY_hek(temphe)));
7369 return newRV_noinc(MUTABLE_SV(av));
7373 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7376 struct regexp *const rx = ReANY(r);
7382 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7384 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7385 || n == RX_BUFF_IDX_CARET_FULLMATCH
7386 || n == RX_BUFF_IDX_CARET_POSTMATCH
7389 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7391 /* on something like
7394 * the KEEPCOPY is set on the PMOP rather than the regex */
7395 if (PL_curpm && r == PM_GETRE(PL_curpm))
7396 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7405 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7406 /* no need to distinguish between them any more */
7407 n = RX_BUFF_IDX_FULLMATCH;
7409 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7410 && rx->offs[0].start != -1)
7412 /* $`, ${^PREMATCH} */
7413 i = rx->offs[0].start;
7417 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7418 && rx->offs[0].end != -1)
7420 /* $', ${^POSTMATCH} */
7421 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7422 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7425 if ( 0 <= n && n <= (I32)rx->nparens &&
7426 (s1 = rx->offs[n].start) != -1 &&
7427 (t1 = rx->offs[n].end) != -1)
7429 /* $&, ${^MATCH}, $1 ... */
7431 s = rx->subbeg + s1 - rx->suboffset;
7436 assert(s >= rx->subbeg);
7437 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7439 #if NO_TAINT_SUPPORT
7440 sv_setpvn(sv, s, i);
7442 const int oldtainted = TAINT_get;
7444 sv_setpvn(sv, s, i);
7445 TAINT_set(oldtainted);
7447 if ( (rx->intflags & PREGf_CANY_SEEN)
7448 ? (RXp_MATCH_UTF8(rx)
7449 && (!i || is_utf8_string((U8*)s, i)))
7450 : (RXp_MATCH_UTF8(rx)) )
7457 if (RXp_MATCH_TAINTED(rx)) {
7458 if (SvTYPE(sv) >= SVt_PVMG) {
7459 MAGIC* const mg = SvMAGIC(sv);
7462 SvMAGIC_set(sv, mg->mg_moremagic);
7464 if ((mgt = SvMAGIC(sv))) {
7465 mg->mg_moremagic = mgt;
7466 SvMAGIC_set(sv, mg);
7477 sv_setsv(sv,&PL_sv_undef);
7483 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7484 SV const * const value)
7486 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7488 PERL_UNUSED_ARG(rx);
7489 PERL_UNUSED_ARG(paren);
7490 PERL_UNUSED_ARG(value);
7493 Perl_croak_no_modify();
7497 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7500 struct regexp *const rx = ReANY(r);
7504 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7506 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7507 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7508 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7511 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7513 /* on something like
7516 * the KEEPCOPY is set on the PMOP rather than the regex */
7517 if (PL_curpm && r == PM_GETRE(PL_curpm))
7518 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7524 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7526 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7527 case RX_BUFF_IDX_PREMATCH: /* $` */
7528 if (rx->offs[0].start != -1) {
7529 i = rx->offs[0].start;
7538 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7539 case RX_BUFF_IDX_POSTMATCH: /* $' */
7540 if (rx->offs[0].end != -1) {
7541 i = rx->sublen - rx->offs[0].end;
7543 s1 = rx->offs[0].end;
7550 default: /* $& / ${^MATCH}, $1, $2, ... */
7551 if (paren <= (I32)rx->nparens &&
7552 (s1 = rx->offs[paren].start) != -1 &&
7553 (t1 = rx->offs[paren].end) != -1)
7559 if (ckWARN(WARN_UNINITIALIZED))
7560 report_uninit((const SV *)sv);
7565 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7566 const char * const s = rx->subbeg - rx->suboffset + s1;
7571 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7578 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7580 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7581 PERL_UNUSED_ARG(rx);
7585 return newSVpvs("Regexp");
7588 /* Scans the name of a named buffer from the pattern.
7589 * If flags is REG_RSN_RETURN_NULL returns null.
7590 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7591 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7592 * to the parsed name as looked up in the RExC_paren_names hash.
7593 * If there is an error throws a vFAIL().. type exception.
7596 #define REG_RSN_RETURN_NULL 0
7597 #define REG_RSN_RETURN_NAME 1
7598 #define REG_RSN_RETURN_DATA 2
7601 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7603 char *name_start = RExC_parse;
7605 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7607 assert (RExC_parse <= RExC_end);
7608 if (RExC_parse == RExC_end) NOOP;
7609 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7610 /* skip IDFIRST by using do...while */
7613 RExC_parse += UTF8SKIP(RExC_parse);
7614 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7618 } while (isWORDCHAR(*RExC_parse));
7620 RExC_parse++; /* so the <- from the vFAIL is after the offending
7622 vFAIL("Group name must start with a non-digit word character");
7626 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7627 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7628 if ( flags == REG_RSN_RETURN_NAME)
7630 else if (flags==REG_RSN_RETURN_DATA) {
7633 if ( ! sv_name ) /* should not happen*/
7634 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7635 if (RExC_paren_names)
7636 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7638 sv_dat = HeVAL(he_str);
7640 vFAIL("Reference to nonexistent named group");
7644 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7645 (unsigned long) flags);
7647 assert(0); /* NOT REACHED */
7652 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7653 int rem=(int)(RExC_end - RExC_parse); \
7662 if (RExC_lastparse!=RExC_parse) \
7663 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7666 iscut ? "..." : "<" \
7669 PerlIO_printf(Perl_debug_log,"%16s",""); \
7672 num = RExC_size + 1; \
7674 num=REG_NODE_NUM(RExC_emit); \
7675 if (RExC_lastnum!=num) \
7676 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7678 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7679 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7680 (int)((depth*2)), "", \
7684 RExC_lastparse=RExC_parse; \
7689 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7690 DEBUG_PARSE_MSG((funcname)); \
7691 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7693 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7694 DEBUG_PARSE_MSG((funcname)); \
7695 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7698 /* This section of code defines the inversion list object and its methods. The
7699 * interfaces are highly subject to change, so as much as possible is static to
7700 * this file. An inversion list is here implemented as a malloc'd C UV array
7701 * as an SVt_INVLIST scalar.
7703 * An inversion list for Unicode is an array of code points, sorted by ordinal
7704 * number. The zeroth element is the first code point in the list. The 1th
7705 * element is the first element beyond that not in the list. In other words,
7706 * the first range is
7707 * invlist[0]..(invlist[1]-1)
7708 * The other ranges follow. Thus every element whose index is divisible by two
7709 * marks the beginning of a range that is in the list, and every element not
7710 * divisible by two marks the beginning of a range not in the list. A single
7711 * element inversion list that contains the single code point N generally
7712 * consists of two elements
7715 * (The exception is when N is the highest representable value on the
7716 * machine, in which case the list containing just it would be a single
7717 * element, itself. By extension, if the last range in the list extends to
7718 * infinity, then the first element of that range will be in the inversion list
7719 * at a position that is divisible by two, and is the final element in the
7721 * Taking the complement (inverting) an inversion list is quite simple, if the
7722 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7723 * This implementation reserves an element at the beginning of each inversion
7724 * list to always contain 0; there is an additional flag in the header which
7725 * indicates if the list begins at the 0, or is offset to begin at the next
7728 * More about inversion lists can be found in "Unicode Demystified"
7729 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7730 * More will be coming when functionality is added later.
7732 * The inversion list data structure is currently implemented as an SV pointing
7733 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7734 * array of UV whose memory management is automatically handled by the existing
7735 * facilities for SV's.
7737 * Some of the methods should always be private to the implementation, and some
7738 * should eventually be made public */
7740 /* The header definitions are in F<inline_invlist.c> */
7742 PERL_STATIC_INLINE UV*
7743 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7745 /* Returns a pointer to the first element in the inversion list's array.
7746 * This is called upon initialization of an inversion list. Where the
7747 * array begins depends on whether the list has the code point U+0000 in it
7748 * or not. The other parameter tells it whether the code that follows this
7749 * call is about to put a 0 in the inversion list or not. The first
7750 * element is either the element reserved for 0, if TRUE, or the element
7751 * after it, if FALSE */
7753 bool* offset = get_invlist_offset_addr(invlist);
7754 UV* zero_addr = (UV *) SvPVX(invlist);
7756 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7759 assert(! _invlist_len(invlist));
7763 /* 1^1 = 0; 1^0 = 1 */
7764 *offset = 1 ^ will_have_0;
7765 return zero_addr + *offset;
7768 PERL_STATIC_INLINE UV*
7769 S_invlist_array(pTHX_ SV* const invlist)
7771 /* Returns the pointer to the inversion list's array. Every time the
7772 * length changes, this needs to be called in case malloc or realloc moved
7775 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7777 /* Must not be empty. If these fail, you probably didn't check for <len>
7778 * being non-zero before trying to get the array */
7779 assert(_invlist_len(invlist));
7781 /* The very first element always contains zero, The array begins either
7782 * there, or if the inversion list is offset, at the element after it.
7783 * The offset header field determines which; it contains 0 or 1 to indicate
7784 * how much additionally to add */
7785 assert(0 == *(SvPVX(invlist)));
7786 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7789 PERL_STATIC_INLINE void
7790 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7792 /* Sets the current number of elements stored in the inversion list.
7793 * Updates SvCUR correspondingly */
7795 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7797 assert(SvTYPE(invlist) == SVt_INVLIST);
7802 : TO_INTERNAL_SIZE(len + offset));
7803 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7806 PERL_STATIC_INLINE IV*
7807 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7809 /* Return the address of the IV that is reserved to hold the cached index
7812 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7814 assert(SvTYPE(invlist) == SVt_INVLIST);
7816 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7819 PERL_STATIC_INLINE IV
7820 S_invlist_previous_index(pTHX_ SV* const invlist)
7822 /* Returns cached index of previous search */
7824 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7826 return *get_invlist_previous_index_addr(invlist);
7829 PERL_STATIC_INLINE void
7830 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7832 /* Caches <index> for later retrieval */
7834 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7836 assert(index == 0 || index < (int) _invlist_len(invlist));
7838 *get_invlist_previous_index_addr(invlist) = index;
7841 PERL_STATIC_INLINE UV
7842 S_invlist_max(pTHX_ SV* const invlist)
7844 /* Returns the maximum number of elements storable in the inversion list's
7845 * array, without having to realloc() */
7847 PERL_ARGS_ASSERT_INVLIST_MAX;
7849 assert(SvTYPE(invlist) == SVt_INVLIST);
7851 /* Assumes worst case, in which the 0 element is not counted in the
7852 * inversion list, so subtracts 1 for that */
7853 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7854 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7855 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7858 #ifndef PERL_IN_XSUB_RE
7860 Perl__new_invlist(pTHX_ IV initial_size)
7863 /* Return a pointer to a newly constructed inversion list, with enough
7864 * space to store 'initial_size' elements. If that number is negative, a
7865 * system default is used instead */
7869 if (initial_size < 0) {
7873 /* Allocate the initial space */
7874 new_list = newSV_type(SVt_INVLIST);
7876 /* First 1 is in case the zero element isn't in the list; second 1 is for
7878 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7879 invlist_set_len(new_list, 0, 0);
7881 /* Force iterinit() to be used to get iteration to work */
7882 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7884 *get_invlist_previous_index_addr(new_list) = 0;
7890 Perl__new_invlist_C_array(pTHX_ const UV* const list)
7892 /* Return a pointer to a newly constructed inversion list, initialized to
7893 * point to <list>, which has to be in the exact correct inversion list
7894 * form, including internal fields. Thus this is a dangerous routine that
7895 * should not be used in the wrong hands. The passed in 'list' contains
7896 * several header fields at the beginning that are not part of the
7897 * inversion list body proper */
7899 const STRLEN length = (STRLEN) list[0];
7900 const UV version_id = list[1];
7901 const bool offset = cBOOL(list[2]);
7902 #define HEADER_LENGTH 3
7903 /* If any of the above changes in any way, you must change HEADER_LENGTH
7904 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7905 * perl -E 'say int(rand 2**31-1)'
7907 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7908 data structure type, so that one being
7909 passed in can be validated to be an
7910 inversion list of the correct vintage.
7913 SV* invlist = newSV_type(SVt_INVLIST);
7915 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7917 if (version_id != INVLIST_VERSION_ID) {
7918 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7921 /* The generated array passed in includes header elements that aren't part
7922 * of the list proper, so start it just after them */
7923 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7925 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7926 shouldn't touch it */
7928 *(get_invlist_offset_addr(invlist)) = offset;
7930 /* The 'length' passed to us is the physical number of elements in the
7931 * inversion list. But if there is an offset the logical number is one
7933 invlist_set_len(invlist, length - offset, offset);
7935 invlist_set_previous_index(invlist, 0);
7937 /* Initialize the iteration pointer. */
7938 invlist_iterfinish(invlist);
7940 SvREADONLY_on(invlist);
7944 #endif /* ifndef PERL_IN_XSUB_RE */
7947 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7949 /* Grow the maximum size of an inversion list */
7951 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7953 assert(SvTYPE(invlist) == SVt_INVLIST);
7955 /* Add one to account for the zero element at the beginning which may not
7956 * be counted by the calling parameters */
7957 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7960 PERL_STATIC_INLINE void
7961 S_invlist_trim(pTHX_ SV* const invlist)
7963 PERL_ARGS_ASSERT_INVLIST_TRIM;
7965 assert(SvTYPE(invlist) == SVt_INVLIST);
7967 /* Change the length of the inversion list to how many entries it currently
7969 SvPV_shrink_to_cur((SV *) invlist);
7973 S__append_range_to_invlist(pTHX_ SV* const invlist,
7974 const UV start, const UV end)
7976 /* Subject to change or removal. Append the range from 'start' to 'end' at
7977 * the end of the inversion list. The range must be above any existing
7981 UV max = invlist_max(invlist);
7982 UV len = _invlist_len(invlist);
7985 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7987 if (len == 0) { /* Empty lists must be initialized */
7988 offset = start != 0;
7989 array = _invlist_array_init(invlist, ! offset);
7992 /* Here, the existing list is non-empty. The current max entry in the
7993 * list is generally the first value not in the set, except when the
7994 * set extends to the end of permissible values, in which case it is
7995 * the first entry in that final set, and so this call is an attempt to
7996 * append out-of-order */
7998 UV final_element = len - 1;
7999 array = invlist_array(invlist);
8000 if (array[final_element] > start
8001 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8003 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",
8004 array[final_element], start,
8005 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8008 /* Here, it is a legal append. If the new range begins with the first
8009 * value not in the set, it is extending the set, so the new first
8010 * value not in the set is one greater than the newly extended range.
8012 offset = *get_invlist_offset_addr(invlist);
8013 if (array[final_element] == start) {
8014 if (end != UV_MAX) {
8015 array[final_element] = end + 1;
8018 /* But if the end is the maximum representable on the machine,
8019 * just let the range that this would extend to have no end */
8020 invlist_set_len(invlist, len - 1, offset);
8026 /* Here the new range doesn't extend any existing set. Add it */
8028 len += 2; /* Includes an element each for the start and end of range */
8030 /* If wll overflow the existing space, extend, which may cause the array to
8033 invlist_extend(invlist, len);
8035 /* Have to set len here to avoid assert failure in invlist_array() */
8036 invlist_set_len(invlist, len, offset);
8038 array = invlist_array(invlist);
8041 invlist_set_len(invlist, len, offset);
8044 /* The next item on the list starts the range, the one after that is
8045 * one past the new range. */
8046 array[len - 2] = start;
8047 if (end != UV_MAX) {
8048 array[len - 1] = end + 1;
8051 /* But if the end is the maximum representable on the machine, just let
8052 * the range have no end */
8053 invlist_set_len(invlist, len - 1, offset);
8057 #ifndef PERL_IN_XSUB_RE
8060 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
8062 /* Searches the inversion list for the entry that contains the input code
8063 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8064 * return value is the index into the list's array of the range that
8069 IV high = _invlist_len(invlist);
8070 const IV highest_element = high - 1;
8073 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8075 /* If list is empty, return failure. */
8080 /* (We can't get the array unless we know the list is non-empty) */
8081 array = invlist_array(invlist);
8083 mid = invlist_previous_index(invlist);
8084 assert(mid >=0 && mid <= highest_element);
8086 /* <mid> contains the cache of the result of the previous call to this
8087 * function (0 the first time). See if this call is for the same result,
8088 * or if it is for mid-1. This is under the theory that calls to this
8089 * function will often be for related code points that are near each other.
8090 * And benchmarks show that caching gives better results. We also test
8091 * here if the code point is within the bounds of the list. These tests
8092 * replace others that would have had to be made anyway to make sure that
8093 * the array bounds were not exceeded, and these give us extra information
8094 * at the same time */
8095 if (cp >= array[mid]) {
8096 if (cp >= array[highest_element]) {
8097 return highest_element;
8100 /* Here, array[mid] <= cp < array[highest_element]. This means that
8101 * the final element is not the answer, so can exclude it; it also
8102 * means that <mid> is not the final element, so can refer to 'mid + 1'
8104 if (cp < array[mid + 1]) {
8110 else { /* cp < aray[mid] */
8111 if (cp < array[0]) { /* Fail if outside the array */
8115 if (cp >= array[mid - 1]) {
8120 /* Binary search. What we are looking for is <i> such that
8121 * array[i] <= cp < array[i+1]
8122 * The loop below converges on the i+1. Note that there may not be an
8123 * (i+1)th element in the array, and things work nonetheless */
8124 while (low < high) {
8125 mid = (low + high) / 2;
8126 assert(mid <= highest_element);
8127 if (array[mid] <= cp) { /* cp >= array[mid] */
8130 /* We could do this extra test to exit the loop early.
8131 if (cp < array[low]) {
8136 else { /* cp < array[mid] */
8143 invlist_set_previous_index(invlist, high);
8148 Perl__invlist_populate_swatch(pTHX_ SV* const invlist,
8149 const UV start, const UV end, U8* swatch)
8151 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8152 * but is used when the swash has an inversion list. This makes this much
8153 * faster, as it uses a binary search instead of a linear one. This is
8154 * intimately tied to that function, and perhaps should be in utf8.c,
8155 * except it is intimately tied to inversion lists as well. It assumes
8156 * that <swatch> is all 0's on input */
8159 const IV len = _invlist_len(invlist);
8163 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8165 if (len == 0) { /* Empty inversion list */
8169 array = invlist_array(invlist);
8171 /* Find which element it is */
8172 i = _invlist_search(invlist, start);
8174 /* We populate from <start> to <end> */
8175 while (current < end) {
8178 /* The inversion list gives the results for every possible code point
8179 * after the first one in the list. Only those ranges whose index is
8180 * even are ones that the inversion list matches. For the odd ones,
8181 * and if the initial code point is not in the list, we have to skip
8182 * forward to the next element */
8183 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8185 if (i >= len) { /* Finished if beyond the end of the array */
8189 if (current >= end) { /* Finished if beyond the end of what we
8191 if (LIKELY(end < UV_MAX)) {
8195 /* We get here when the upper bound is the maximum
8196 * representable on the machine, and we are looking for just
8197 * that code point. Have to special case it */
8199 goto join_end_of_list;
8202 assert(current >= start);
8204 /* The current range ends one below the next one, except don't go past
8207 upper = (i < len && array[i] < end) ? array[i] : end;
8209 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8210 * for each code point in it */
8211 for (; current < upper; current++) {
8212 const STRLEN offset = (STRLEN)(current - start);
8213 swatch[offset >> 3] |= 1 << (offset & 7);
8218 /* Quit if at the end of the list */
8221 /* But first, have to deal with the highest possible code point on
8222 * the platform. The previous code assumes that <end> is one
8223 * beyond where we want to populate, but that is impossible at the
8224 * platform's infinity, so have to handle it specially */
8225 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8227 const STRLEN offset = (STRLEN)(end - start);
8228 swatch[offset >> 3] |= 1 << (offset & 7);
8233 /* Advance to the next range, which will be for code points not in the
8242 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8243 const bool complement_b, SV** output)
8245 /* Take the union of two inversion lists and point <output> to it. *output
8246 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8247 * the reference count to that list will be decremented if not already a
8248 * temporary (mortal); otherwise *output will be made correspondingly
8249 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8250 * second list is returned. If <complement_b> is TRUE, the union is taken
8251 * of the complement (inversion) of <b> instead of b itself.
8253 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8254 * Richard Gillam, published by Addison-Wesley, and explained at some
8255 * length there. The preface says to incorporate its examples into your
8256 * code at your own risk.
8258 * The algorithm is like a merge sort.
8260 * XXX A potential performance improvement is to keep track as we go along
8261 * if only one of the inputs contributes to the result, meaning the other
8262 * is a subset of that one. In that case, we can skip the final copy and
8263 * return the larger of the input lists, but then outside code might need
8264 * to keep track of whether to free the input list or not */
8266 const UV* array_a; /* a's array */
8268 UV len_a; /* length of a's array */
8271 SV* u; /* the resulting union */
8275 UV i_a = 0; /* current index into a's array */
8279 /* running count, as explained in the algorithm source book; items are
8280 * stopped accumulating and are output when the count changes to/from 0.
8281 * The count is incremented when we start a range that's in the set, and
8282 * decremented when we start a range that's not in the set. So its range
8283 * is 0 to 2. Only when the count is zero is something not in the set.
8287 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8290 /* If either one is empty, the union is the other one */
8291 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8292 bool make_temp = FALSE; /* Should we mortalize the result? */
8296 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8302 *output = invlist_clone(b);
8304 _invlist_invert(*output);
8306 } /* else *output already = b; */
8309 sv_2mortal(*output);
8313 else if ((len_b = _invlist_len(b)) == 0) {
8314 bool make_temp = FALSE;
8316 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8321 /* The complement of an empty list is a list that has everything in it,
8322 * so the union with <a> includes everything too */
8325 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8329 *output = _new_invlist(1);
8330 _append_range_to_invlist(*output, 0, UV_MAX);
8332 else if (*output != a) {
8333 *output = invlist_clone(a);
8335 /* else *output already = a; */
8338 sv_2mortal(*output);
8343 /* Here both lists exist and are non-empty */
8344 array_a = invlist_array(a);
8345 array_b = invlist_array(b);
8347 /* If are to take the union of 'a' with the complement of b, set it
8348 * up so are looking at b's complement. */
8351 /* To complement, we invert: if the first element is 0, remove it. To
8352 * do this, we just pretend the array starts one later */
8353 if (array_b[0] == 0) {
8359 /* But if the first element is not zero, we pretend the list starts
8360 * at the 0 that is always stored immediately before the array. */
8366 /* Size the union for the worst case: that the sets are completely
8368 u = _new_invlist(len_a + len_b);
8370 /* Will contain U+0000 if either component does */
8371 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8372 || (len_b > 0 && array_b[0] == 0));
8374 /* Go through each list item by item, stopping when exhausted one of
8376 while (i_a < len_a && i_b < len_b) {
8377 UV cp; /* The element to potentially add to the union's array */
8378 bool cp_in_set; /* is it in the the input list's set or not */
8380 /* We need to take one or the other of the two inputs for the union.
8381 * Since we are merging two sorted lists, we take the smaller of the
8382 * next items. In case of a tie, we take the one that is in its set
8383 * first. If we took one not in the set first, it would decrement the
8384 * count, possibly to 0 which would cause it to be output as ending the
8385 * range, and the next time through we would take the same number, and
8386 * output it again as beginning the next range. By doing it the
8387 * opposite way, there is no possibility that the count will be
8388 * momentarily decremented to 0, and thus the two adjoining ranges will
8389 * be seamlessly merged. (In a tie and both are in the set or both not
8390 * in the set, it doesn't matter which we take first.) */
8391 if (array_a[i_a] < array_b[i_b]
8392 || (array_a[i_a] == array_b[i_b]
8393 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8395 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8399 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8400 cp = array_b[i_b++];
8403 /* Here, have chosen which of the two inputs to look at. Only output
8404 * if the running count changes to/from 0, which marks the
8405 * beginning/end of a range in that's in the set */
8408 array_u[i_u++] = cp;
8415 array_u[i_u++] = cp;
8420 /* Here, we are finished going through at least one of the lists, which
8421 * means there is something remaining in at most one. We check if the list
8422 * that hasn't been exhausted is positioned such that we are in the middle
8423 * of a range in its set or not. (i_a and i_b point to the element beyond
8424 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8425 * is potentially more to output.
8426 * There are four cases:
8427 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8428 * in the union is entirely from the non-exhausted set.
8429 * 2) Both were in their sets, count is 2. Nothing further should
8430 * be output, as everything that remains will be in the exhausted
8431 * list's set, hence in the union; decrementing to 1 but not 0 insures
8433 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8434 * Nothing further should be output because the union includes
8435 * everything from the exhausted set. Not decrementing ensures that.
8436 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8437 * decrementing to 0 insures that we look at the remainder of the
8438 * non-exhausted set */
8439 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8440 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8445 /* The final length is what we've output so far, plus what else is about to
8446 * be output. (If 'count' is non-zero, then the input list we exhausted
8447 * has everything remaining up to the machine's limit in its set, and hence
8448 * in the union, so there will be no further output. */
8451 /* At most one of the subexpressions will be non-zero */
8452 len_u += (len_a - i_a) + (len_b - i_b);
8455 /* Set result to final length, which can change the pointer to array_u, so
8457 if (len_u != _invlist_len(u)) {
8458 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8460 array_u = invlist_array(u);
8463 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8464 * the other) ended with everything above it not in its set. That means
8465 * that the remaining part of the union is precisely the same as the
8466 * non-exhausted list, so can just copy it unchanged. (If both list were
8467 * exhausted at the same time, then the operations below will be both 0.)
8470 IV copy_count; /* At most one will have a non-zero copy count */
8471 if ((copy_count = len_a - i_a) > 0) {
8472 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8474 else if ((copy_count = len_b - i_b) > 0) {
8475 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8479 /* We may be removing a reference to one of the inputs. If so, the output
8480 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8481 * count decremented) */
8482 if (a == *output || b == *output) {
8483 assert(! invlist_is_iterating(*output));
8484 if ((SvTEMP(*output))) {
8488 SvREFCNT_dec_NN(*output);
8498 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8499 const bool complement_b, SV** i)
8501 /* Take the intersection of two inversion lists and point <i> to it. *i
8502 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8503 * the reference count to that list will be decremented if not already a
8504 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8505 * The first list, <a>, may be NULL, in which case an empty list is
8506 * returned. If <complement_b> is TRUE, the result will be the
8507 * intersection of <a> and the complement (or inversion) of <b> instead of
8510 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8511 * Richard Gillam, published by Addison-Wesley, and explained at some
8512 * length there. The preface says to incorporate its examples into your
8513 * code at your own risk. In fact, it had bugs
8515 * The algorithm is like a merge sort, and is essentially the same as the
8519 const UV* array_a; /* a's array */
8521 UV len_a; /* length of a's array */
8524 SV* r; /* the resulting intersection */
8528 UV i_a = 0; /* current index into a's array */
8532 /* running count, as explained in the algorithm source book; items are
8533 * stopped accumulating and are output when the count changes to/from 2.
8534 * The count is incremented when we start a range that's in the set, and
8535 * decremented when we start a range that's not in the set. So its range
8536 * is 0 to 2. Only when the count is 2 is something in the intersection.
8540 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8543 /* Special case if either one is empty */
8544 len_a = (a == NULL) ? 0 : _invlist_len(a);
8545 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8546 bool make_temp = FALSE;
8548 if (len_a != 0 && complement_b) {
8550 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8551 * be empty. Here, also we are using 'b's complement, which hence
8552 * must be every possible code point. Thus the intersection is
8556 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8561 *i = invlist_clone(a);
8563 /* else *i is already 'a' */
8571 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8572 * intersection must be empty */
8574 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8579 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8583 *i = _new_invlist(0);
8591 /* Here both lists exist and are non-empty */
8592 array_a = invlist_array(a);
8593 array_b = invlist_array(b);
8595 /* If are to take the intersection of 'a' with the complement of b, set it
8596 * up so are looking at b's complement. */
8599 /* To complement, we invert: if the first element is 0, remove it. To
8600 * do this, we just pretend the array starts one later */
8601 if (array_b[0] == 0) {
8607 /* But if the first element is not zero, we pretend the list starts
8608 * at the 0 that is always stored immediately before the array. */
8614 /* Size the intersection for the worst case: that the intersection ends up
8615 * fragmenting everything to be completely disjoint */
8616 r= _new_invlist(len_a + len_b);
8618 /* Will contain U+0000 iff both components do */
8619 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8620 && len_b > 0 && array_b[0] == 0);
8622 /* Go through each list item by item, stopping when exhausted one of
8624 while (i_a < len_a && i_b < len_b) {
8625 UV cp; /* The element to potentially add to the intersection's
8627 bool cp_in_set; /* Is it in the input list's set or not */
8629 /* We need to take one or the other of the two inputs for the
8630 * intersection. Since we are merging two sorted lists, we take the
8631 * smaller of the next items. In case of a tie, we take the one that
8632 * is not in its set first (a difference from the union algorithm). If
8633 * we took one in the set first, it would increment the count, possibly
8634 * to 2 which would cause it to be output as starting a range in the
8635 * intersection, and the next time through we would take that same
8636 * number, and output it again as ending the set. By doing it the
8637 * opposite of this, there is no possibility that the count will be
8638 * momentarily incremented to 2. (In a tie and both are in the set or
8639 * both not in the set, it doesn't matter which we take first.) */
8640 if (array_a[i_a] < array_b[i_b]
8641 || (array_a[i_a] == array_b[i_b]
8642 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8644 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8648 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8652 /* Here, have chosen which of the two inputs to look at. Only output
8653 * if the running count changes to/from 2, which marks the
8654 * beginning/end of a range that's in the intersection */
8658 array_r[i_r++] = cp;
8663 array_r[i_r++] = cp;
8669 /* Here, we are finished going through at least one of the lists, which
8670 * means there is something remaining in at most one. We check if the list
8671 * that has been exhausted is positioned such that we are in the middle
8672 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8673 * the ones we care about.) There are four cases:
8674 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8675 * nothing left in the intersection.
8676 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8677 * above 2. What should be output is exactly that which is in the
8678 * non-exhausted set, as everything it has is also in the intersection
8679 * set, and everything it doesn't have can't be in the intersection
8680 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8681 * gets incremented to 2. Like the previous case, the intersection is
8682 * everything that remains in the non-exhausted set.
8683 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8684 * remains 1. And the intersection has nothing more. */
8685 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8686 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8691 /* The final length is what we've output so far plus what else is in the
8692 * intersection. At most one of the subexpressions below will be non-zero
8696 len_r += (len_a - i_a) + (len_b - i_b);
8699 /* Set result to final length, which can change the pointer to array_r, so
8701 if (len_r != _invlist_len(r)) {
8702 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8704 array_r = invlist_array(r);
8707 /* Finish outputting any remaining */
8708 if (count >= 2) { /* At most one will have a non-zero copy count */
8710 if ((copy_count = len_a - i_a) > 0) {
8711 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8713 else if ((copy_count = len_b - i_b) > 0) {
8714 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8718 /* We may be removing a reference to one of the inputs. If so, the output
8719 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8720 * count decremented) */
8721 if (a == *i || b == *i) {
8722 assert(! invlist_is_iterating(*i));
8727 SvREFCNT_dec_NN(*i);
8737 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8739 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8740 * set. A pointer to the inversion list is returned. This may actually be
8741 * a new list, in which case the passed in one has been destroyed. The
8742 * passed in inversion list can be NULL, in which case a new one is created
8743 * with just the one range in it */
8748 if (invlist == NULL) {
8749 invlist = _new_invlist(2);
8753 len = _invlist_len(invlist);
8756 /* If comes after the final entry actually in the list, can just append it
8759 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8760 && start >= invlist_array(invlist)[len - 1]))
8762 _append_range_to_invlist(invlist, start, end);
8766 /* Here, can't just append things, create and return a new inversion list
8767 * which is the union of this range and the existing inversion list */
8768 range_invlist = _new_invlist(2);
8769 _append_range_to_invlist(range_invlist, start, end);
8771 _invlist_union(invlist, range_invlist, &invlist);
8773 /* The temporary can be freed */
8774 SvREFCNT_dec_NN(range_invlist);
8780 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
8781 UV** other_elements_ptr)
8783 /* Create and return an inversion list whose contents are to be populated
8784 * by the caller. The caller gives the number of elements (in 'size') and
8785 * the very first element ('element0'). This function will set
8786 * '*other_elements_ptr' to an array of UVs, where the remaining elements
8789 * Obviously there is some trust involved that the caller will properly
8790 * fill in the other elements of the array.
8792 * (The first element needs to be passed in, as the underlying code does
8793 * things differently depending on whether it is zero or non-zero) */
8795 SV* invlist = _new_invlist(size);
8798 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
8800 _append_range_to_invlist(invlist, element0, element0);
8801 offset = *get_invlist_offset_addr(invlist);
8803 invlist_set_len(invlist, size, offset);
8804 *other_elements_ptr = invlist_array(invlist) + 1;
8810 PERL_STATIC_INLINE SV*
8811 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8812 return _add_range_to_invlist(invlist, cp, cp);
8815 #ifndef PERL_IN_XSUB_RE
8817 Perl__invlist_invert(pTHX_ SV* const invlist)
8819 /* Complement the input inversion list. This adds a 0 if the list didn't
8820 * have a zero; removes it otherwise. As described above, the data
8821 * structure is set up so that this is very efficient */
8823 PERL_ARGS_ASSERT__INVLIST_INVERT;
8825 assert(! invlist_is_iterating(invlist));
8827 /* The inverse of matching nothing is matching everything */
8828 if (_invlist_len(invlist) == 0) {
8829 _append_range_to_invlist(invlist, 0, UV_MAX);
8833 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8838 PERL_STATIC_INLINE SV*
8839 S_invlist_clone(pTHX_ SV* const invlist)
8842 /* Return a new inversion list that is a copy of the input one, which is
8843 * unchanged. The new list will not be mortal even if the old one was. */
8845 /* Need to allocate extra space to accommodate Perl's addition of a
8846 * trailing NUL to SvPV's, since it thinks they are always strings */
8847 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8848 STRLEN physical_length = SvCUR(invlist);
8849 bool offset = *(get_invlist_offset_addr(invlist));
8851 PERL_ARGS_ASSERT_INVLIST_CLONE;
8853 *(get_invlist_offset_addr(new_invlist)) = offset;
8854 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8855 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8860 PERL_STATIC_INLINE STRLEN*
8861 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8863 /* Return the address of the UV that contains the current iteration
8866 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8868 assert(SvTYPE(invlist) == SVt_INVLIST);
8870 return &(((XINVLIST*) SvANY(invlist))->iterator);
8873 PERL_STATIC_INLINE void
8874 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8876 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8878 *get_invlist_iter_addr(invlist) = 0;
8881 PERL_STATIC_INLINE void
8882 S_invlist_iterfinish(pTHX_ SV* invlist)
8884 /* Terminate iterator for invlist. This is to catch development errors.
8885 * Any iteration that is interrupted before completed should call this
8886 * function. Functions that add code points anywhere else but to the end
8887 * of an inversion list assert that they are not in the middle of an
8888 * iteration. If they were, the addition would make the iteration
8889 * problematical: if the iteration hadn't reached the place where things
8890 * were being added, it would be ok */
8892 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8894 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8898 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8900 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8901 * This call sets in <*start> and <*end>, the next range in <invlist>.
8902 * Returns <TRUE> if successful and the next call will return the next
8903 * range; <FALSE> if was already at the end of the list. If the latter,
8904 * <*start> and <*end> are unchanged, and the next call to this function
8905 * will start over at the beginning of the list */
8907 STRLEN* pos = get_invlist_iter_addr(invlist);
8908 UV len = _invlist_len(invlist);
8911 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8914 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8918 array = invlist_array(invlist);
8920 *start = array[(*pos)++];
8926 *end = array[(*pos)++] - 1;
8932 PERL_STATIC_INLINE bool
8933 S_invlist_is_iterating(pTHX_ SV* const invlist)
8935 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8937 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8940 PERL_STATIC_INLINE UV
8941 S_invlist_highest(pTHX_ SV* const invlist)
8943 /* Returns the highest code point that matches an inversion list. This API
8944 * has an ambiguity, as it returns 0 under either the highest is actually
8945 * 0, or if the list is empty. If this distinction matters to you, check
8946 * for emptiness before calling this function */
8948 UV len = _invlist_len(invlist);
8951 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8957 array = invlist_array(invlist);
8959 /* The last element in the array in the inversion list always starts a
8960 * range that goes to infinity. That range may be for code points that are
8961 * matched in the inversion list, or it may be for ones that aren't
8962 * matched. In the latter case, the highest code point in the set is one
8963 * less than the beginning of this range; otherwise it is the final element
8964 * of this range: infinity */
8965 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8967 : array[len - 1] - 1;
8970 #ifndef PERL_IN_XSUB_RE
8972 Perl__invlist_contents(pTHX_ SV* const invlist)
8974 /* Get the contents of an inversion list into a string SV so that they can
8975 * be printed out. It uses the format traditionally done for debug tracing
8979 SV* output = newSVpvs("\n");
8981 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8983 assert(! invlist_is_iterating(invlist));
8985 invlist_iterinit(invlist);
8986 while (invlist_iternext(invlist, &start, &end)) {
8987 if (end == UV_MAX) {
8988 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8990 else if (end != start) {
8991 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8995 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9003 #ifndef PERL_IN_XSUB_RE
9005 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9006 const char * const indent, SV* const invlist)
9008 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9009 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9010 * the string 'indent'. The output looks like this:
9011 [0] 0x000A .. 0x000D
9013 [4] 0x2028 .. 0x2029
9014 [6] 0x3104 .. INFINITY
9015 * This means that the first range of code points matched by the list are
9016 * 0xA through 0xD; the second range contains only the single code point
9017 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9018 * are used to define each range (except if the final range extends to
9019 * infinity, only a single element is needed). The array index of the
9020 * first element for the corresponding range is given in brackets. */
9025 PERL_ARGS_ASSERT__INVLIST_DUMP;
9027 if (invlist_is_iterating(invlist)) {
9028 Perl_dump_indent(aTHX_ level, file,
9029 "%sCan't dump inversion list because is in middle of iterating\n",
9034 invlist_iterinit(invlist);
9035 while (invlist_iternext(invlist, &start, &end)) {
9036 if (end == UV_MAX) {
9037 Perl_dump_indent(aTHX_ level, file,
9038 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9039 indent, (UV)count, start);
9041 else if (end != start) {
9042 Perl_dump_indent(aTHX_ level, file,
9043 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9044 indent, (UV)count, start, end);
9047 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9048 indent, (UV)count, start);
9055 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9057 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9059 /* Return a boolean as to if the two passed in inversion lists are
9060 * identical. The final argument, if TRUE, says to take the complement of
9061 * the second inversion list before doing the comparison */
9063 const UV* array_a = invlist_array(a);
9064 const UV* array_b = invlist_array(b);
9065 UV len_a = _invlist_len(a);
9066 UV len_b = _invlist_len(b);
9068 UV i = 0; /* current index into the arrays */
9069 bool retval = TRUE; /* Assume are identical until proven otherwise */
9071 PERL_ARGS_ASSERT__INVLISTEQ;
9073 /* If are to compare 'a' with the complement of b, set it
9074 * up so are looking at b's complement. */
9077 /* The complement of nothing is everything, so <a> would have to have
9078 * just one element, starting at zero (ending at infinity) */
9080 return (len_a == 1 && array_a[0] == 0);
9082 else if (array_b[0] == 0) {
9084 /* Otherwise, to complement, we invert. Here, the first element is
9085 * 0, just remove it. To do this, we just pretend the array starts
9093 /* But if the first element is not zero, we pretend the list starts
9094 * at the 0 that is always stored immediately before the array. */
9100 /* Make sure that the lengths are the same, as well as the final element
9101 * before looping through the remainder. (Thus we test the length, final,
9102 * and first elements right off the bat) */
9103 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9106 else for (i = 0; i < len_a - 1; i++) {
9107 if (array_a[i] != array_b[i]) {
9117 #undef HEADER_LENGTH
9118 #undef TO_INTERNAL_SIZE
9119 #undef FROM_INTERNAL_SIZE
9120 #undef INVLIST_VERSION_ID
9122 /* End of inversion list object */
9125 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9127 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9128 * constructs, and updates RExC_flags with them. On input, RExC_parse
9129 * should point to the first flag; it is updated on output to point to the
9130 * final ')' or ':'. There needs to be at least one flag, or this will
9133 /* for (?g), (?gc), and (?o) warnings; warning
9134 about (?c) will warn about (?g) -- japhy */
9136 #define WASTED_O 0x01
9137 #define WASTED_G 0x02
9138 #define WASTED_C 0x04
9139 #define WASTED_GC (WASTED_G|WASTED_C)
9140 I32 wastedflags = 0x00;
9141 U32 posflags = 0, negflags = 0;
9142 U32 *flagsp = &posflags;
9143 char has_charset_modifier = '\0';
9145 bool has_use_defaults = FALSE;
9146 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9148 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9150 /* '^' as an initial flag sets certain defaults */
9151 if (UCHARAT(RExC_parse) == '^') {
9153 has_use_defaults = TRUE;
9154 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9155 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9156 ? REGEX_UNICODE_CHARSET
9157 : REGEX_DEPENDS_CHARSET);
9160 cs = get_regex_charset(RExC_flags);
9161 if (cs == REGEX_DEPENDS_CHARSET
9162 && (RExC_utf8 || RExC_uni_semantics))
9164 cs = REGEX_UNICODE_CHARSET;
9167 while (*RExC_parse) {
9168 /* && strchr("iogcmsx", *RExC_parse) */
9169 /* (?g), (?gc) and (?o) are useless here
9170 and must be globally applied -- japhy */
9171 switch (*RExC_parse) {
9173 /* Code for the imsx flags */
9174 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
9176 case LOCALE_PAT_MOD:
9177 if (has_charset_modifier) {
9178 goto excess_modifier;
9180 else if (flagsp == &negflags) {
9183 cs = REGEX_LOCALE_CHARSET;
9184 has_charset_modifier = LOCALE_PAT_MOD;
9185 RExC_contains_locale = 1;
9187 case UNICODE_PAT_MOD:
9188 if (has_charset_modifier) {
9189 goto excess_modifier;
9191 else if (flagsp == &negflags) {
9194 cs = REGEX_UNICODE_CHARSET;
9195 has_charset_modifier = UNICODE_PAT_MOD;
9197 case ASCII_RESTRICT_PAT_MOD:
9198 if (flagsp == &negflags) {
9201 if (has_charset_modifier) {
9202 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9203 goto excess_modifier;
9205 /* Doubled modifier implies more restricted */
9206 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9209 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9211 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9213 case DEPENDS_PAT_MOD:
9214 if (has_use_defaults) {
9215 goto fail_modifiers;
9217 else if (flagsp == &negflags) {
9220 else if (has_charset_modifier) {
9221 goto excess_modifier;
9224 /* The dual charset means unicode semantics if the
9225 * pattern (or target, not known until runtime) are
9226 * utf8, or something in the pattern indicates unicode
9228 cs = (RExC_utf8 || RExC_uni_semantics)
9229 ? REGEX_UNICODE_CHARSET
9230 : REGEX_DEPENDS_CHARSET;
9231 has_charset_modifier = DEPENDS_PAT_MOD;
9235 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9236 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9238 else if (has_charset_modifier == *(RExC_parse - 1)) {
9239 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9243 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9248 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9251 case ONCE_PAT_MOD: /* 'o' */
9252 case GLOBAL_PAT_MOD: /* 'g' */
9253 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9254 const I32 wflagbit = *RExC_parse == 'o'
9257 if (! (wastedflags & wflagbit) ) {
9258 wastedflags |= wflagbit;
9259 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9262 "Useless (%s%c) - %suse /%c modifier",
9263 flagsp == &negflags ? "?-" : "?",
9265 flagsp == &negflags ? "don't " : "",
9272 case CONTINUE_PAT_MOD: /* 'c' */
9273 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9274 if (! (wastedflags & WASTED_C) ) {
9275 wastedflags |= WASTED_GC;
9276 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9279 "Useless (%sc) - %suse /gc modifier",
9280 flagsp == &negflags ? "?-" : "?",
9281 flagsp == &negflags ? "don't " : ""
9286 case KEEPCOPY_PAT_MOD: /* 'p' */
9287 if (flagsp == &negflags) {
9289 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9291 *flagsp |= RXf_PMf_KEEPCOPY;
9295 /* A flag is a default iff it is following a minus, so
9296 * if there is a minus, it means will be trying to
9297 * re-specify a default which is an error */
9298 if (has_use_defaults || flagsp == &negflags) {
9299 goto fail_modifiers;
9302 wastedflags = 0; /* reset so (?g-c) warns twice */
9306 RExC_flags |= posflags;
9307 RExC_flags &= ~negflags;
9308 set_regex_charset(&RExC_flags, cs);
9309 if (RExC_flags & RXf_PMf_FOLD) {
9310 RExC_contains_i = 1;
9316 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9317 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9318 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9319 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9328 - reg - regular expression, i.e. main body or parenthesized thing
9330 * Caller must absorb opening parenthesis.
9332 * Combining parenthesis handling with the base level of regular expression
9333 * is a trifle forced, but the need to tie the tails of the branches to what
9334 * follows makes it hard to avoid.
9336 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9338 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9340 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9343 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9344 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9345 needs to be restarted.
9346 Otherwise would only return NULL if regbranch() returns NULL, which
9349 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9350 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9351 * 2 is like 1, but indicates that nextchar() has been called to advance
9352 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9353 * this flag alerts us to the need to check for that */
9356 regnode *ret; /* Will be the head of the group. */
9359 regnode *ender = NULL;
9362 U32 oregflags = RExC_flags;
9363 bool have_branch = 0;
9365 I32 freeze_paren = 0;
9366 I32 after_freeze = 0;
9368 char * parse_start = RExC_parse; /* MJD */
9369 char * const oregcomp_parse = RExC_parse;
9371 GET_RE_DEBUG_FLAGS_DECL;
9373 PERL_ARGS_ASSERT_REG;
9374 DEBUG_PARSE("reg ");
9376 *flagp = 0; /* Tentatively. */
9379 /* Make an OPEN node, if parenthesized. */
9382 /* Under /x, space and comments can be gobbled up between the '(' and
9383 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9384 * intervening space, as the sequence is a token, and a token should be
9386 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9388 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9389 char *start_verb = RExC_parse;
9390 STRLEN verb_len = 0;
9391 char *start_arg = NULL;
9392 unsigned char op = 0;
9394 int internal_argval = 0; /* internal_argval is only useful if
9397 if (has_intervening_patws && SIZE_ONLY) {
9398 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
9400 while ( *RExC_parse && *RExC_parse != ')' ) {
9401 if ( *RExC_parse == ':' ) {
9402 start_arg = RExC_parse + 1;
9408 verb_len = RExC_parse - start_verb;
9411 while ( *RExC_parse && *RExC_parse != ')' )
9413 if ( *RExC_parse != ')' )
9414 vFAIL("Unterminated verb pattern argument");
9415 if ( RExC_parse == start_arg )
9418 if ( *RExC_parse != ')' )
9419 vFAIL("Unterminated verb pattern");
9422 switch ( *start_verb ) {
9423 case 'A': /* (*ACCEPT) */
9424 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9426 internal_argval = RExC_nestroot;
9429 case 'C': /* (*COMMIT) */
9430 if ( memEQs(start_verb,verb_len,"COMMIT") )
9433 case 'F': /* (*FAIL) */
9434 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9439 case ':': /* (*:NAME) */
9440 case 'M': /* (*MARK:NAME) */
9441 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9446 case 'P': /* (*PRUNE) */
9447 if ( memEQs(start_verb,verb_len,"PRUNE") )
9450 case 'S': /* (*SKIP) */
9451 if ( memEQs(start_verb,verb_len,"SKIP") )
9454 case 'T': /* (*THEN) */
9455 /* [19:06] <TimToady> :: is then */
9456 if ( memEQs(start_verb,verb_len,"THEN") ) {
9458 RExC_seen |= REG_CUTGROUP_SEEN;
9463 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9465 "Unknown verb pattern '%"UTF8f"'",
9466 UTF8fARG(UTF, verb_len, start_verb));
9469 if ( start_arg && internal_argval ) {
9470 vFAIL3("Verb pattern '%.*s' may not have an argument",
9471 verb_len, start_verb);
9472 } else if ( argok < 0 && !start_arg ) {
9473 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9474 verb_len, start_verb);
9476 ret = reganode(pRExC_state, op, internal_argval);
9477 if ( ! internal_argval && ! SIZE_ONLY ) {
9479 SV *sv = newSVpvn( start_arg,
9480 RExC_parse - start_arg);
9481 ARG(ret) = add_data( pRExC_state,
9483 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9490 if (!internal_argval)
9491 RExC_seen |= REG_VERBARG_SEEN;
9492 } else if ( start_arg ) {
9493 vFAIL3("Verb pattern '%.*s' may not have an argument",
9494 verb_len, start_verb);
9496 ret = reg_node(pRExC_state, op);
9498 nextchar(pRExC_state);
9501 else if (*RExC_parse == '?') { /* (?...) */
9502 bool is_logical = 0;
9503 const char * const seqstart = RExC_parse;
9504 if (has_intervening_patws && SIZE_ONLY) {
9505 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
9509 paren = *RExC_parse++;
9510 ret = NULL; /* For look-ahead/behind. */
9513 case 'P': /* (?P...) variants for those used to PCRE/Python */
9514 paren = *RExC_parse++;
9515 if ( paren == '<') /* (?P<...>) named capture */
9517 else if (paren == '>') { /* (?P>name) named recursion */
9518 goto named_recursion;
9520 else if (paren == '=') { /* (?P=...) named backref */
9521 /* this pretty much dupes the code for \k<NAME> in
9522 * regatom(), if you change this make sure you change that
9524 char* name_start = RExC_parse;
9526 SV *sv_dat = reg_scan_name(pRExC_state,
9527 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9528 if (RExC_parse == name_start || *RExC_parse != ')')
9529 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9530 vFAIL2("Sequence %.3s... not terminated",parse_start);
9533 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9534 RExC_rxi->data->data[num]=(void*)sv_dat;
9535 SvREFCNT_inc_simple_void(sv_dat);
9538 ret = reganode(pRExC_state,
9541 : (ASCII_FOLD_RESTRICTED)
9543 : (AT_LEAST_UNI_SEMANTICS)
9551 Set_Node_Offset(ret, parse_start+1);
9552 Set_Node_Cur_Length(ret, parse_start);
9554 nextchar(pRExC_state);
9558 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9559 vFAIL3("Sequence (%.*s...) not recognized",
9560 RExC_parse-seqstart, seqstart);
9562 case '<': /* (?<...) */
9563 if (*RExC_parse == '!')
9565 else if (*RExC_parse != '=')
9571 case '\'': /* (?'...') */
9572 name_start= RExC_parse;
9573 svname = reg_scan_name(pRExC_state,
9574 SIZE_ONLY /* reverse test from the others */
9575 ? REG_RSN_RETURN_NAME
9576 : REG_RSN_RETURN_NULL);
9577 if (RExC_parse == name_start || *RExC_parse != paren)
9578 vFAIL2("Sequence (?%c... not terminated",
9579 paren=='>' ? '<' : paren);
9583 if (!svname) /* shouldn't happen */
9585 "panic: reg_scan_name returned NULL");
9586 if (!RExC_paren_names) {
9587 RExC_paren_names= newHV();
9588 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9590 RExC_paren_name_list= newAV();
9591 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
9594 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
9596 sv_dat = HeVAL(he_str);
9598 /* croak baby croak */
9600 "panic: paren_name hash element allocation failed");
9601 } else if ( SvPOK(sv_dat) ) {
9602 /* (?|...) can mean we have dupes so scan to check
9603 its already been stored. Maybe a flag indicating
9604 we are inside such a construct would be useful,
9605 but the arrays are likely to be quite small, so
9606 for now we punt -- dmq */
9607 IV count = SvIV(sv_dat);
9608 I32 *pv = (I32*)SvPVX(sv_dat);
9610 for ( i = 0 ; i < count ; i++ ) {
9611 if ( pv[i] == RExC_npar ) {
9617 pv = (I32*)SvGROW(sv_dat,
9618 SvCUR(sv_dat) + sizeof(I32)+1);
9619 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
9620 pv[count] = RExC_npar;
9621 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
9624 (void)SvUPGRADE(sv_dat,SVt_PVNV);
9625 sv_setpvn(sv_dat, (char *)&(RExC_npar),
9628 SvIV_set(sv_dat, 1);
9631 /* Yes this does cause a memory leak in debugging Perls
9633 if (!av_store(RExC_paren_name_list,
9634 RExC_npar, SvREFCNT_inc(svname)))
9635 SvREFCNT_dec_NN(svname);
9638 /*sv_dump(sv_dat);*/
9640 nextchar(pRExC_state);
9642 goto capturing_parens;
9644 RExC_seen |= REG_LOOKBEHIND_SEEN;
9645 RExC_in_lookbehind++;
9647 case '=': /* (?=...) */
9648 RExC_seen_zerolen++;
9650 case '!': /* (?!...) */
9651 RExC_seen_zerolen++;
9652 if (*RExC_parse == ')') {
9653 ret=reg_node(pRExC_state, OPFAIL);
9654 nextchar(pRExC_state);
9658 case '|': /* (?|...) */
9659 /* branch reset, behave like a (?:...) except that
9660 buffers in alternations share the same numbers */
9662 after_freeze = freeze_paren = RExC_npar;
9664 case ':': /* (?:...) */
9665 case '>': /* (?>...) */
9667 case '$': /* (?$...) */
9668 case '@': /* (?@...) */
9669 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
9671 case '#': /* (?#...) */
9672 /* XXX As soon as we disallow separating the '?' and '*' (by
9673 * spaces or (?#...) comment), it is believed that this case
9674 * will be unreachable and can be removed. See
9676 while (*RExC_parse && *RExC_parse != ')')
9678 if (*RExC_parse != ')')
9679 FAIL("Sequence (?#... not terminated");
9680 nextchar(pRExC_state);
9683 case '0' : /* (?0) */
9684 case 'R' : /* (?R) */
9685 if (*RExC_parse != ')')
9686 FAIL("Sequence (?R) not terminated");
9687 ret = reg_node(pRExC_state, GOSTART);
9688 RExC_seen |= REG_GOSTART_SEEN;
9689 *flagp |= POSTPONED;
9690 nextchar(pRExC_state);
9693 { /* named and numeric backreferences */
9695 case '&': /* (?&NAME) */
9696 parse_start = RExC_parse - 1;
9699 SV *sv_dat = reg_scan_name(pRExC_state,
9700 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9701 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9703 if (RExC_parse == RExC_end || *RExC_parse != ')')
9704 vFAIL("Sequence (?&... not terminated");
9705 goto gen_recurse_regop;
9706 assert(0); /* NOT REACHED */
9708 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9710 vFAIL("Illegal pattern");
9712 goto parse_recursion;
9714 case '-': /* (?-1) */
9715 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9716 RExC_parse--; /* rewind to let it be handled later */
9720 case '1': case '2': case '3': case '4': /* (?1) */
9721 case '5': case '6': case '7': case '8': case '9':
9724 num = atoi(RExC_parse);
9725 parse_start = RExC_parse - 1; /* MJD */
9726 if (*RExC_parse == '-')
9728 while (isDIGIT(*RExC_parse))
9730 if (*RExC_parse!=')')
9731 vFAIL("Expecting close bracket");
9734 if ( paren == '-' ) {
9736 Diagram of capture buffer numbering.
9737 Top line is the normal capture buffer numbers
9738 Bottom line is the negative indexing as from
9742 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9746 num = RExC_npar + num;
9749 vFAIL("Reference to nonexistent group");
9751 } else if ( paren == '+' ) {
9752 num = RExC_npar + num - 1;
9755 ret = reganode(pRExC_state, GOSUB, num);
9757 if (num > (I32)RExC_rx->nparens) {
9759 vFAIL("Reference to nonexistent group");
9761 ARG2L_SET( ret, RExC_recurse_count++);
9763 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9764 "Recurse #%"UVuf" to %"IVdf"\n",
9765 (UV)ARG(ret), (IV)ARG2L(ret)));
9769 RExC_seen |= REG_RECURSE_SEEN;
9770 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9771 Set_Node_Offset(ret, parse_start); /* MJD */
9773 *flagp |= POSTPONED;
9774 nextchar(pRExC_state);
9776 } /* named and numeric backreferences */
9777 assert(0); /* NOT REACHED */
9779 case '?': /* (??...) */
9781 if (*RExC_parse != '{') {
9783 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9785 "Sequence (%"UTF8f"...) not recognized",
9786 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9789 *flagp |= POSTPONED;
9790 paren = *RExC_parse++;
9792 case '{': /* (?{...}) */
9795 struct reg_code_block *cb;
9797 RExC_seen_zerolen++;
9799 if ( !pRExC_state->num_code_blocks
9800 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9801 || pRExC_state->code_blocks[pRExC_state->code_index].start
9802 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9805 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9806 FAIL("panic: Sequence (?{...}): no code block found\n");
9807 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9809 /* this is a pre-compiled code block (?{...}) */
9810 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9811 RExC_parse = RExC_start + cb->end;
9814 if (cb->src_regex) {
9815 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
9816 RExC_rxi->data->data[n] =
9817 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9818 RExC_rxi->data->data[n+1] = (void*)o;
9821 n = add_data(pRExC_state,
9822 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
9823 RExC_rxi->data->data[n] = (void*)o;
9826 pRExC_state->code_index++;
9827 nextchar(pRExC_state);
9831 ret = reg_node(pRExC_state, LOGICAL);
9832 eval = reganode(pRExC_state, EVAL, n);
9835 /* for later propagation into (??{}) return value */
9836 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9838 REGTAIL(pRExC_state, ret, eval);
9839 /* deal with the length of this later - MJD */
9842 ret = reganode(pRExC_state, EVAL, n);
9843 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9844 Set_Node_Offset(ret, parse_start);
9847 case '(': /* (?(?{...})...) and (?(?=...)...) */
9850 if (RExC_parse[0] == '?') { /* (?(?...)) */
9851 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9852 || RExC_parse[1] == '<'
9853 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9857 ret = reg_node(pRExC_state, LOGICAL);
9861 tail = reg(pRExC_state, 1, &flag, depth+1);
9862 if (flag & RESTART_UTF8) {
9863 *flagp = RESTART_UTF8;
9866 REGTAIL(pRExC_state, ret, tail);
9870 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9871 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9873 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9874 char *name_start= RExC_parse++;
9876 SV *sv_dat=reg_scan_name(pRExC_state,
9877 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9878 if (RExC_parse == name_start || *RExC_parse != ch)
9879 vFAIL2("Sequence (?(%c... not terminated",
9880 (ch == '>' ? '<' : ch));
9883 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9884 RExC_rxi->data->data[num]=(void*)sv_dat;
9885 SvREFCNT_inc_simple_void(sv_dat);
9887 ret = reganode(pRExC_state,NGROUPP,num);
9888 goto insert_if_check_paren;
9890 else if (RExC_parse[0] == 'D' &&
9891 RExC_parse[1] == 'E' &&
9892 RExC_parse[2] == 'F' &&
9893 RExC_parse[3] == 'I' &&
9894 RExC_parse[4] == 'N' &&
9895 RExC_parse[5] == 'E')
9897 ret = reganode(pRExC_state,DEFINEP,0);
9900 goto insert_if_check_paren;
9902 else if (RExC_parse[0] == 'R') {
9905 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9906 parno = atoi(RExC_parse++);
9907 while (isDIGIT(*RExC_parse))
9909 } else if (RExC_parse[0] == '&') {
9912 sv_dat = reg_scan_name(pRExC_state,
9914 ? REG_RSN_RETURN_NULL
9915 : REG_RSN_RETURN_DATA);
9916 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9918 ret = reganode(pRExC_state,INSUBP,parno);
9919 goto insert_if_check_paren;
9921 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9925 parno = atoi(RExC_parse++);
9927 while (isDIGIT(*RExC_parse))
9929 ret = reganode(pRExC_state, GROUPP, parno);
9931 insert_if_check_paren:
9932 if (*(tmp = nextchar(pRExC_state)) != ')') {
9933 /* nextchar also skips comments, so undo its work
9934 * and skip over the the next character.
9937 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9938 vFAIL("Switch condition not recognized");
9941 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9942 br = regbranch(pRExC_state, &flags, 1,depth+1);
9944 if (flags & RESTART_UTF8) {
9945 *flagp = RESTART_UTF8;
9948 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9951 REGTAIL(pRExC_state, br, reganode(pRExC_state,
9953 c = *nextchar(pRExC_state);
9958 vFAIL("(?(DEFINE)....) does not allow branches");
9960 /* Fake one for optimizer. */
9961 lastbr = reganode(pRExC_state, IFTHEN, 0);
9963 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9964 if (flags & RESTART_UTF8) {
9965 *flagp = RESTART_UTF8;
9968 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9971 REGTAIL(pRExC_state, ret, lastbr);
9974 c = *nextchar(pRExC_state);
9979 vFAIL("Switch (?(condition)... contains too many branches");
9980 ender = reg_node(pRExC_state, TAIL);
9981 REGTAIL(pRExC_state, br, ender);
9983 REGTAIL(pRExC_state, lastbr, ender);
9984 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9987 REGTAIL(pRExC_state, ret, ender);
9988 RExC_size++; /* XXX WHY do we need this?!!
9989 For large programs it seems to be required
9990 but I can't figure out why. -- dmq*/
9994 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9995 vFAIL("Unknown switch condition (?(...))");
9998 case '[': /* (?[ ... ]) */
9999 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10002 RExC_parse--; /* for vFAIL to print correctly */
10003 vFAIL("Sequence (? incomplete");
10005 default: /* e.g., (?i) */
10008 parse_lparen_question_flags(pRExC_state);
10009 if (UCHARAT(RExC_parse) != ':') {
10010 nextchar(pRExC_state);
10015 nextchar(pRExC_state);
10025 ret = reganode(pRExC_state, OPEN, parno);
10027 if (!RExC_nestroot)
10028 RExC_nestroot = parno;
10029 if (RExC_seen & REG_RECURSE_SEEN
10030 && !RExC_open_parens[parno-1])
10032 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10033 "Setting open paren #%"IVdf" to %d\n",
10034 (IV)parno, REG_NODE_NUM(ret)));
10035 RExC_open_parens[parno-1]= ret;
10038 Set_Node_Length(ret, 1); /* MJD */
10039 Set_Node_Offset(ret, RExC_parse); /* MJD */
10047 /* Pick up the branches, linking them together. */
10048 parse_start = RExC_parse; /* MJD */
10049 br = regbranch(pRExC_state, &flags, 1,depth+1);
10051 /* branch_len = (paren != 0); */
10054 if (flags & RESTART_UTF8) {
10055 *flagp = RESTART_UTF8;
10058 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10060 if (*RExC_parse == '|') {
10061 if (!SIZE_ONLY && RExC_extralen) {
10062 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10065 reginsert(pRExC_state, BRANCH, br, depth+1);
10066 Set_Node_Length(br, paren != 0);
10067 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10071 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10073 else if (paren == ':') {
10074 *flagp |= flags&SIMPLE;
10076 if (is_open) { /* Starts with OPEN. */
10077 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10079 else if (paren != '?') /* Not Conditional */
10081 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10083 while (*RExC_parse == '|') {
10084 if (!SIZE_ONLY && RExC_extralen) {
10085 ender = reganode(pRExC_state, LONGJMP,0);
10087 /* Append to the previous. */
10088 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10091 RExC_extralen += 2; /* Account for LONGJMP. */
10092 nextchar(pRExC_state);
10093 if (freeze_paren) {
10094 if (RExC_npar > after_freeze)
10095 after_freeze = RExC_npar;
10096 RExC_npar = freeze_paren;
10098 br = regbranch(pRExC_state, &flags, 0, depth+1);
10101 if (flags & RESTART_UTF8) {
10102 *flagp = RESTART_UTF8;
10105 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10107 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10109 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10112 if (have_branch || paren != ':') {
10113 /* Make a closing node, and hook it on the end. */
10116 ender = reg_node(pRExC_state, TAIL);
10119 ender = reganode(pRExC_state, CLOSE, parno);
10120 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10121 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10122 "Setting close paren #%"IVdf" to %d\n",
10123 (IV)parno, REG_NODE_NUM(ender)));
10124 RExC_close_parens[parno-1]= ender;
10125 if (RExC_nestroot == parno)
10128 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10129 Set_Node_Length(ender,1); /* MJD */
10135 *flagp &= ~HASWIDTH;
10138 ender = reg_node(pRExC_state, SUCCEED);
10141 ender = reg_node(pRExC_state, END);
10143 assert(!RExC_opend); /* there can only be one! */
10144 RExC_opend = ender;
10148 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10149 SV * const mysv_val1=sv_newmortal();
10150 SV * const mysv_val2=sv_newmortal();
10151 DEBUG_PARSE_MSG("lsbr");
10152 regprop(RExC_rx, mysv_val1, lastbr);
10153 regprop(RExC_rx, mysv_val2, ender);
10154 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10155 SvPV_nolen_const(mysv_val1),
10156 (IV)REG_NODE_NUM(lastbr),
10157 SvPV_nolen_const(mysv_val2),
10158 (IV)REG_NODE_NUM(ender),
10159 (IV)(ender - lastbr)
10162 REGTAIL(pRExC_state, lastbr, ender);
10164 if (have_branch && !SIZE_ONLY) {
10165 char is_nothing= 1;
10167 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10169 /* Hook the tails of the branches to the closing node. */
10170 for (br = ret; br; br = regnext(br)) {
10171 const U8 op = PL_regkind[OP(br)];
10172 if (op == BRANCH) {
10173 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10174 if ( OP(NEXTOPER(br)) != NOTHING
10175 || regnext(NEXTOPER(br)) != ender)
10178 else if (op == BRANCHJ) {
10179 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10180 /* for now we always disable this optimisation * /
10181 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10182 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10188 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10189 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10190 SV * const mysv_val1=sv_newmortal();
10191 SV * const mysv_val2=sv_newmortal();
10192 DEBUG_PARSE_MSG("NADA");
10193 regprop(RExC_rx, mysv_val1, ret);
10194 regprop(RExC_rx, mysv_val2, ender);
10195 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10196 SvPV_nolen_const(mysv_val1),
10197 (IV)REG_NODE_NUM(ret),
10198 SvPV_nolen_const(mysv_val2),
10199 (IV)REG_NODE_NUM(ender),
10204 if (OP(ender) == TAIL) {
10209 for ( opt= br + 1; opt < ender ; opt++ )
10210 OP(opt)= OPTIMIZED;
10211 NEXT_OFF(br)= ender - br;
10219 static const char parens[] = "=!<,>";
10221 if (paren && (p = strchr(parens, paren))) {
10222 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10223 int flag = (p - parens) > 1;
10226 node = SUSPEND, flag = 0;
10227 reginsert(pRExC_state, node,ret, depth+1);
10228 Set_Node_Cur_Length(ret, parse_start);
10229 Set_Node_Offset(ret, parse_start + 1);
10231 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10235 /* Check for proper termination. */
10237 /* restore original flags, but keep (?p) */
10238 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10239 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10240 RExC_parse = oregcomp_parse;
10241 vFAIL("Unmatched (");
10244 else if (!paren && RExC_parse < RExC_end) {
10245 if (*RExC_parse == ')') {
10247 vFAIL("Unmatched )");
10250 FAIL("Junk on end of regexp"); /* "Can't happen". */
10251 assert(0); /* NOTREACHED */
10254 if (RExC_in_lookbehind) {
10255 RExC_in_lookbehind--;
10257 if (after_freeze > RExC_npar)
10258 RExC_npar = after_freeze;
10263 - regbranch - one alternative of an | operator
10265 * Implements the concatenation operator.
10267 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10271 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10275 regnode *chain = NULL;
10277 I32 flags = 0, c = 0;
10278 GET_RE_DEBUG_FLAGS_DECL;
10280 PERL_ARGS_ASSERT_REGBRANCH;
10282 DEBUG_PARSE("brnc");
10287 if (!SIZE_ONLY && RExC_extralen)
10288 ret = reganode(pRExC_state, BRANCHJ,0);
10290 ret = reg_node(pRExC_state, BRANCH);
10291 Set_Node_Length(ret, 1);
10295 if (!first && SIZE_ONLY)
10296 RExC_extralen += 1; /* BRANCHJ */
10298 *flagp = WORST; /* Tentatively. */
10301 nextchar(pRExC_state);
10302 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10303 flags &= ~TRYAGAIN;
10304 latest = regpiece(pRExC_state, &flags,depth+1);
10305 if (latest == NULL) {
10306 if (flags & TRYAGAIN)
10308 if (flags & RESTART_UTF8) {
10309 *flagp = RESTART_UTF8;
10312 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10314 else if (ret == NULL)
10316 *flagp |= flags&(HASWIDTH|POSTPONED);
10317 if (chain == NULL) /* First piece. */
10318 *flagp |= flags&SPSTART;
10321 REGTAIL(pRExC_state, chain, latest);
10326 if (chain == NULL) { /* Loop ran zero times. */
10327 chain = reg_node(pRExC_state, NOTHING);
10332 *flagp |= flags&SIMPLE;
10339 - regpiece - something followed by possible [*+?]
10341 * Note that the branching code sequences used for ? and the general cases
10342 * of * and + are somewhat optimized: they use the same NOTHING node as
10343 * both the endmarker for their branch list and the body of the last branch.
10344 * It might seem that this node could be dispensed with entirely, but the
10345 * endmarker role is not redundant.
10347 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10349 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10353 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10360 const char * const origparse = RExC_parse;
10362 I32 max = REG_INFTY;
10363 #ifdef RE_TRACK_PATTERN_OFFSETS
10366 const char *maxpos = NULL;
10368 /* Save the original in case we change the emitted regop to a FAIL. */
10369 regnode * const orig_emit = RExC_emit;
10371 GET_RE_DEBUG_FLAGS_DECL;
10373 PERL_ARGS_ASSERT_REGPIECE;
10375 DEBUG_PARSE("piec");
10377 ret = regatom(pRExC_state, &flags,depth+1);
10379 if (flags & (TRYAGAIN|RESTART_UTF8))
10380 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10382 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10388 if (op == '{' && regcurly(RExC_parse, FALSE)) {
10390 #ifdef RE_TRACK_PATTERN_OFFSETS
10391 parse_start = RExC_parse; /* MJD */
10393 next = RExC_parse + 1;
10394 while (isDIGIT(*next) || *next == ',') {
10395 if (*next == ',') {
10403 if (*next == '}') { /* got one */
10407 min = atoi(RExC_parse);
10408 if (*maxpos == ',')
10411 maxpos = RExC_parse;
10412 max = atoi(maxpos);
10413 if (!max && *maxpos != '0')
10414 max = REG_INFTY; /* meaning "infinity" */
10415 else if (max >= REG_INFTY)
10416 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10418 nextchar(pRExC_state);
10419 if (max < min) { /* If can't match, warn and optimize to fail
10422 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10424 /* We can't back off the size because we have to reserve
10425 * enough space for all the things we are about to throw
10426 * away, but we can shrink it by the ammount we are about
10427 * to re-use here */
10428 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10431 RExC_emit = orig_emit;
10433 ret = reg_node(pRExC_state, OPFAIL);
10436 else if (min == max
10437 && RExC_parse < RExC_end
10438 && (*RExC_parse == '?' || *RExC_parse == '+'))
10441 ckWARN2reg(RExC_parse + 1,
10442 "Useless use of greediness modifier '%c'",
10445 /* Absorb the modifier, so later code doesn't see nor use
10447 nextchar(pRExC_state);
10451 if ((flags&SIMPLE)) {
10452 RExC_naughty += 2 + RExC_naughty / 2;
10453 reginsert(pRExC_state, CURLY, ret, depth+1);
10454 Set_Node_Offset(ret, parse_start+1); /* MJD */
10455 Set_Node_Cur_Length(ret, parse_start);
10458 regnode * const w = reg_node(pRExC_state, WHILEM);
10461 REGTAIL(pRExC_state, ret, w);
10462 if (!SIZE_ONLY && RExC_extralen) {
10463 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10464 reginsert(pRExC_state, NOTHING,ret, depth+1);
10465 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10467 reginsert(pRExC_state, CURLYX,ret, depth+1);
10469 Set_Node_Offset(ret, parse_start+1);
10470 Set_Node_Length(ret,
10471 op == '{' ? (RExC_parse - parse_start) : 1);
10473 if (!SIZE_ONLY && RExC_extralen)
10474 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10475 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10477 RExC_whilem_seen++, RExC_extralen += 3;
10478 RExC_naughty += 4 + RExC_naughty; /* compound interest */
10485 *flagp |= HASWIDTH;
10487 ARG1_SET(ret, (U16)min);
10488 ARG2_SET(ret, (U16)max);
10490 if (max == REG_INFTY)
10491 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10497 if (!ISMULT1(op)) {
10502 #if 0 /* Now runtime fix should be reliable. */
10504 /* if this is reinstated, don't forget to put this back into perldiag:
10506 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10508 (F) The part of the regexp subject to either the * or + quantifier
10509 could match an empty string. The {#} shows in the regular
10510 expression about where the problem was discovered.
10514 if (!(flags&HASWIDTH) && op != '?')
10515 vFAIL("Regexp *+ operand could be empty");
10518 #ifdef RE_TRACK_PATTERN_OFFSETS
10519 parse_start = RExC_parse;
10521 nextchar(pRExC_state);
10523 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10525 if (op == '*' && (flags&SIMPLE)) {
10526 reginsert(pRExC_state, STAR, ret, depth+1);
10529 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10531 else if (op == '*') {
10535 else if (op == '+' && (flags&SIMPLE)) {
10536 reginsert(pRExC_state, PLUS, ret, depth+1);
10539 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10541 else if (op == '+') {
10545 else if (op == '?') {
10550 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10551 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10552 ckWARN2reg(RExC_parse,
10553 "%"UTF8f" matches null string many times",
10554 UTF8fARG(UTF, (RExC_parse >= origparse
10555 ? RExC_parse - origparse
10558 (void)ReREFCNT_inc(RExC_rx_sv);
10561 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10562 nextchar(pRExC_state);
10563 reginsert(pRExC_state, MINMOD, ret, depth+1);
10564 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10567 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10569 nextchar(pRExC_state);
10570 ender = reg_node(pRExC_state, SUCCEED);
10571 REGTAIL(pRExC_state, ret, ender);
10572 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10574 ender = reg_node(pRExC_state, TAIL);
10575 REGTAIL(pRExC_state, ret, ender);
10578 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10580 vFAIL("Nested quantifiers");
10587 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
10588 UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
10589 const bool strict /* Apply stricter parsing rules? */
10593 /* This is expected to be called by a parser routine that has recognized '\N'
10594 and needs to handle the rest. RExC_parse is expected to point at the first
10595 char following the N at the time of the call. On successful return,
10596 RExC_parse has been updated to point to just after the sequence identified
10597 by this routine, and <*flagp> has been updated.
10599 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
10602 \N may begin either a named sequence, or if outside a character class, mean
10603 to match a non-newline. For non single-quoted regexes, the tokenizer has
10604 attempted to decide which, and in the case of a named sequence, converted it
10605 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
10606 where c1... are the characters in the sequence. For single-quoted regexes,
10607 the tokenizer passes the \N sequence through unchanged; this code will not
10608 attempt to determine this nor expand those, instead raising a syntax error.
10609 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
10610 or there is no '}', it signals that this \N occurrence means to match a
10613 Only the \N{U+...} form should occur in a character class, for the same
10614 reason that '.' inside a character class means to just match a period: it
10615 just doesn't make sense.
10617 The function raises an error (via vFAIL), and doesn't return for various
10618 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
10619 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
10620 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
10621 only possible if node_p is non-NULL.
10624 If <valuep> is non-null, it means the caller can accept an input sequence
10625 consisting of a just a single code point; <*valuep> is set to that value
10626 if the input is such.
10628 If <node_p> is non-null it signifies that the caller can accept any other
10629 legal sequence (i.e., one that isn't just a single code point). <*node_p>
10631 1) \N means not-a-NL: points to a newly created REG_ANY node;
10632 2) \N{}: points to a new NOTHING node;
10633 3) otherwise: points to a new EXACT node containing the resolved
10635 Note that FALSE is returned for single code point sequences if <valuep> is
10639 char * endbrace; /* '}' following the name */
10641 char *endchar; /* Points to '.' or '}' ending cur char in the input
10643 bool has_multiple_chars; /* true if the input stream contains a sequence of
10644 more than one character */
10646 GET_RE_DEBUG_FLAGS_DECL;
10648 PERL_ARGS_ASSERT_GROK_BSLASH_N;
10650 GET_RE_DEBUG_FLAGS;
10652 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
10654 /* The [^\n] meaning of \N ignores spaces and comments under the /x
10655 * modifier. The other meaning does not, so use a temporary until we find
10656 * out which we are being called with */
10657 p = (RExC_flags & RXf_PMf_EXTENDED)
10658 ? regwhite( pRExC_state, RExC_parse )
10661 /* Disambiguate between \N meaning a named character versus \N meaning
10662 * [^\n]. The former is assumed when it can't be the latter. */
10663 if (*p != '{' || regcurly(p, FALSE)) {
10666 /* no bare \N allowed in a charclass */
10667 if (in_char_class) {
10668 vFAIL("\\N in a character class must be a named character: \\N{...}");
10672 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
10674 nextchar(pRExC_state);
10675 *node_p = reg_node(pRExC_state, REG_ANY);
10676 *flagp |= HASWIDTH|SIMPLE;
10678 Set_Node_Length(*node_p, 1); /* MJD */
10682 /* Here, we have decided it should be a named character or sequence */
10684 /* The test above made sure that the next real character is a '{', but
10685 * under the /x modifier, it could be separated by space (or a comment and
10686 * \n) and this is not allowed (for consistency with \x{...} and the
10687 * tokenizer handling of \N{NAME}). */
10688 if (*RExC_parse != '{') {
10689 vFAIL("Missing braces on \\N{}");
10692 RExC_parse++; /* Skip past the '{' */
10694 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
10695 || ! (endbrace == RExC_parse /* nothing between the {} */
10696 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below
10698 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg)
10701 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
10702 vFAIL("\\N{NAME} must be resolved by the lexer");
10705 if (endbrace == RExC_parse) { /* empty: \N{} */
10708 *node_p = reg_node(pRExC_state,NOTHING);
10710 else if (in_char_class) {
10711 if (SIZE_ONLY && in_char_class) {
10713 RExC_parse++; /* Position after the "}" */
10714 vFAIL("Zero length \\N{}");
10717 ckWARNreg(RExC_parse,
10718 "Ignoring zero length \\N{} in character class");
10726 nextchar(pRExC_state);
10730 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
10731 RExC_parse += 2; /* Skip past the 'U+' */
10733 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10735 /* Code points are separated by dots. If none, there is only one code
10736 * point, and is terminated by the brace */
10737 has_multiple_chars = (endchar < endbrace);
10739 if (valuep && (! has_multiple_chars || in_char_class)) {
10740 /* We only pay attention to the first char of
10741 multichar strings being returned in char classes. I kinda wonder
10742 if this makes sense as it does change the behaviour
10743 from earlier versions, OTOH that behaviour was broken
10744 as well. XXX Solution is to recharacterize as
10745 [rest-of-class]|multi1|multi2... */
10747 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10748 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10749 | PERL_SCAN_DISALLOW_PREFIX
10750 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10752 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10754 /* The tokenizer should have guaranteed validity, but it's possible to
10755 * bypass it by using single quoting, so check */
10756 if (length_of_hex == 0
10757 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10759 RExC_parse += length_of_hex; /* Includes all the valid */
10760 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10761 ? UTF8SKIP(RExC_parse)
10763 /* Guard against malformed utf8 */
10764 if (RExC_parse >= endchar) {
10765 RExC_parse = endchar;
10767 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10770 if (in_char_class && has_multiple_chars) {
10772 RExC_parse = endbrace;
10773 vFAIL("\\N{} in character class restricted to one character");
10776 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10780 RExC_parse = endbrace + 1;
10782 else if (! node_p || ! has_multiple_chars) {
10784 /* Here, the input is legal, but not according to the caller's
10785 * options. We fail without advancing the parse, so that the
10786 * caller can try again */
10792 /* What is done here is to convert this to a sub-pattern of the form
10793 * (?:\x{char1}\x{char2}...)
10794 * and then call reg recursively. That way, it retains its atomicness,
10795 * while not having to worry about special handling that some code
10796 * points may have. toke.c has converted the original Unicode values
10797 * to native, so that we can just pass on the hex values unchanged. We
10798 * do have to set a flag to keep recoding from happening in the
10801 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10803 char *orig_end = RExC_end;
10806 while (RExC_parse < endbrace) {
10808 /* Convert to notation the rest of the code understands */
10809 sv_catpv(substitute_parse, "\\x{");
10810 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10811 sv_catpv(substitute_parse, "}");
10813 /* Point to the beginning of the next character in the sequence. */
10814 RExC_parse = endchar + 1;
10815 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10817 sv_catpv(substitute_parse, ")");
10819 RExC_parse = SvPV(substitute_parse, len);
10821 /* Don't allow empty number */
10823 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10825 RExC_end = RExC_parse + len;
10827 /* The values are Unicode, and therefore not subject to recoding */
10828 RExC_override_recoding = 1;
10830 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10831 if (flags & RESTART_UTF8) {
10832 *flagp = RESTART_UTF8;
10835 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10838 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10840 RExC_parse = endbrace;
10841 RExC_end = orig_end;
10842 RExC_override_recoding = 0;
10844 nextchar(pRExC_state);
10854 * It returns the code point in utf8 for the value in *encp.
10855 * value: a code value in the source encoding
10856 * encp: a pointer to an Encode object
10858 * If the result from Encode is not a single character,
10859 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10862 S_reg_recode(pTHX_ const char value, SV **encp)
10865 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10866 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10867 const STRLEN newlen = SvCUR(sv);
10868 UV uv = UNICODE_REPLACEMENT;
10870 PERL_ARGS_ASSERT_REG_RECODE;
10874 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10877 if (!newlen || numlen != newlen) {
10878 uv = UNICODE_REPLACEMENT;
10884 PERL_STATIC_INLINE U8
10885 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10889 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10895 op = get_regex_charset(RExC_flags);
10896 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10897 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10898 been, so there is no hole */
10901 return op + EXACTF;
10904 PERL_STATIC_INLINE void
10905 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
10906 regnode *node, I32* flagp, STRLEN len, UV code_point)
10908 /* This knows the details about sizing an EXACTish node, setting flags for
10909 * it (by setting <*flagp>, and potentially populating it with a single
10912 * If <len> (the length in bytes) is non-zero, this function assumes that
10913 * the node has already been populated, and just does the sizing. In this
10914 * case <code_point> should be the final code point that has already been
10915 * placed into the node. This value will be ignored except that under some
10916 * circumstances <*flagp> is set based on it.
10918 * If <len> is zero, the function assumes that the node is to contain only
10919 * the single character given by <code_point> and calculates what <len>
10920 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10921 * additionally will populate the node's STRING with <code_point> or its
10924 * In both cases <*flagp> is appropriately set
10926 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10927 * 255, must be folded (the former only when the rules indicate it can
10930 bool len_passed_in = cBOOL(len != 0);
10931 U8 character[UTF8_MAXBYTES_CASE+1];
10933 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10935 if (! len_passed_in) {
10937 if (UNI_IS_INVARIANT(code_point)) {
10938 if (LOC || ! FOLD) { /* /l defers folding until runtime */
10939 *character = (U8) code_point;
10941 else { /* Here is /i and not /l (toFOLD() is defined on just
10942 ASCII, which isn't the same thing as INVARIANT on
10943 EBCDIC, but it works there, as the extra invariants
10944 fold to themselves) */
10945 *character = toFOLD((U8) code_point);
10949 else if (FOLD && (! LOC
10950 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
10951 { /* Folding, and ok to do so now */
10952 _to_uni_fold_flags(code_point,
10955 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
10956 ? FOLD_FLAGS_NOMIX_ASCII
10959 else if (code_point <= MAX_UTF8_TWO_BYTE) {
10961 /* Not folding this cp, and can output it directly */
10962 *character = UTF8_TWO_BYTE_HI(code_point);
10963 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
10967 uvchr_to_utf8( character, code_point);
10968 len = UTF8SKIP(character);
10970 } /* Else pattern isn't UTF8. We only fold the sharp s, when
10972 else if (UNLIKELY(code_point == LATIN_SMALL_LETTER_SHARP_S)
10974 && AT_LEAST_UNI_SEMANTICS
10975 && ! ASCII_FOLD_RESTRICTED)
10978 *(character + 1) = 's';
10982 *character = (U8) code_point;
10988 RExC_size += STR_SZ(len);
10991 RExC_emit += STR_SZ(len);
10992 STR_LEN(node) = len;
10993 if (! len_passed_in) {
10994 Copy((char *) character, STRING(node), len, char);
10998 *flagp |= HASWIDTH;
11000 /* A single character node is SIMPLE, except for the special-cased SHARP S
11002 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11003 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11004 || ! FOLD || ! DEPENDS_SEMANTICS))
11011 /* return atoi(p), unless it's too big to sensibly be a backref,
11012 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11015 S_backref_value(char *p)
11019 for (;isDIGIT(*q); q++); /* calculate length of num */
11020 if (q - p == 0 || q - p > 9)
11027 - regatom - the lowest level
11029 Try to identify anything special at the start of the pattern. If there
11030 is, then handle it as required. This may involve generating a single regop,
11031 such as for an assertion; or it may involve recursing, such as to
11032 handle a () structure.
11034 If the string doesn't start with something special then we gobble up
11035 as much literal text as we can.
11037 Once we have been able to handle whatever type of thing started the
11038 sequence, we return.
11040 Note: we have to be careful with escapes, as they can be both literal
11041 and special, and in the case of \10 and friends, context determines which.
11043 A summary of the code structure is:
11045 switch (first_byte) {
11046 cases for each special:
11047 handle this special;
11050 switch (2nd byte) {
11051 cases for each unambiguous special:
11052 handle this special;
11054 cases for each ambigous special/literal:
11056 if (special) handle here
11058 default: // unambiguously literal:
11061 default: // is a literal char
11064 create EXACTish node for literal;
11065 while (more input and node isn't full) {
11066 switch (input_byte) {
11067 cases for each special;
11068 make sure parse pointer is set so that the next call to
11069 regatom will see this special first
11070 goto loopdone; // EXACTish node terminated by prev. char
11072 append char to EXACTISH node;
11074 get next input byte;
11078 return the generated node;
11080 Specifically there are two separate switches for handling
11081 escape sequences, with the one for handling literal escapes requiring
11082 a dummy entry for all of the special escapes that are actually handled
11085 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11087 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11089 Otherwise does not return NULL.
11093 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11096 regnode *ret = NULL;
11098 char *parse_start = RExC_parse;
11102 GET_RE_DEBUG_FLAGS_DECL;
11104 *flagp = WORST; /* Tentatively. */
11106 DEBUG_PARSE("atom");
11108 PERL_ARGS_ASSERT_REGATOM;
11111 switch ((U8)*RExC_parse) {
11113 RExC_seen_zerolen++;
11114 nextchar(pRExC_state);
11115 if (RExC_flags & RXf_PMf_MULTILINE)
11116 ret = reg_node(pRExC_state, MBOL);
11117 else if (RExC_flags & RXf_PMf_SINGLELINE)
11118 ret = reg_node(pRExC_state, SBOL);
11120 ret = reg_node(pRExC_state, BOL);
11121 Set_Node_Length(ret, 1); /* MJD */
11124 nextchar(pRExC_state);
11126 RExC_seen_zerolen++;
11127 if (RExC_flags & RXf_PMf_MULTILINE)
11128 ret = reg_node(pRExC_state, MEOL);
11129 else if (RExC_flags & RXf_PMf_SINGLELINE)
11130 ret = reg_node(pRExC_state, SEOL);
11132 ret = reg_node(pRExC_state, EOL);
11133 Set_Node_Length(ret, 1); /* MJD */
11136 nextchar(pRExC_state);
11137 if (RExC_flags & RXf_PMf_SINGLELINE)
11138 ret = reg_node(pRExC_state, SANY);
11140 ret = reg_node(pRExC_state, REG_ANY);
11141 *flagp |= HASWIDTH|SIMPLE;
11143 Set_Node_Length(ret, 1); /* MJD */
11147 char * const oregcomp_parse = ++RExC_parse;
11148 ret = regclass(pRExC_state, flagp,depth+1,
11149 FALSE, /* means parse the whole char class */
11150 TRUE, /* allow multi-char folds */
11151 FALSE, /* don't silence non-portable warnings. */
11153 if (*RExC_parse != ']') {
11154 RExC_parse = oregcomp_parse;
11155 vFAIL("Unmatched [");
11158 if (*flagp & RESTART_UTF8)
11160 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11163 nextchar(pRExC_state);
11164 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11168 nextchar(pRExC_state);
11169 ret = reg(pRExC_state, 2, &flags,depth+1);
11171 if (flags & TRYAGAIN) {
11172 if (RExC_parse == RExC_end) {
11173 /* Make parent create an empty node if needed. */
11174 *flagp |= TRYAGAIN;
11179 if (flags & RESTART_UTF8) {
11180 *flagp = RESTART_UTF8;
11183 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11186 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11190 if (flags & TRYAGAIN) {
11191 *flagp |= TRYAGAIN;
11194 vFAIL("Internal urp");
11195 /* Supposed to be caught earlier. */
11198 if (!regcurly(RExC_parse, FALSE)) {
11207 vFAIL("Quantifier follows nothing");
11212 This switch handles escape sequences that resolve to some kind
11213 of special regop and not to literal text. Escape sequnces that
11214 resolve to literal text are handled below in the switch marked
11217 Every entry in this switch *must* have a corresponding entry
11218 in the literal escape switch. However, the opposite is not
11219 required, as the default for this switch is to jump to the
11220 literal text handling code.
11222 switch ((U8)*++RExC_parse) {
11224 /* Special Escapes */
11226 RExC_seen_zerolen++;
11227 ret = reg_node(pRExC_state, SBOL);
11229 goto finish_meta_pat;
11231 ret = reg_node(pRExC_state, GPOS);
11232 RExC_seen |= REG_GPOS_SEEN;
11234 goto finish_meta_pat;
11236 RExC_seen_zerolen++;
11237 ret = reg_node(pRExC_state, KEEPS);
11239 /* XXX:dmq : disabling in-place substitution seems to
11240 * be necessary here to avoid cases of memory corruption, as
11241 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11243 RExC_seen |= REG_LOOKBEHIND_SEEN;
11244 goto finish_meta_pat;
11246 ret = reg_node(pRExC_state, SEOL);
11248 RExC_seen_zerolen++; /* Do not optimize RE away */
11249 goto finish_meta_pat;
11251 ret = reg_node(pRExC_state, EOS);
11253 RExC_seen_zerolen++; /* Do not optimize RE away */
11254 goto finish_meta_pat;
11256 ret = reg_node(pRExC_state, CANY);
11257 RExC_seen |= REG_CANY_SEEN;
11258 *flagp |= HASWIDTH|SIMPLE;
11259 goto finish_meta_pat;
11261 ret = reg_node(pRExC_state, CLUMP);
11262 *flagp |= HASWIDTH;
11263 goto finish_meta_pat;
11269 arg = ANYOF_WORDCHAR;
11273 RExC_seen_zerolen++;
11274 RExC_seen |= REG_LOOKBEHIND_SEEN;
11275 op = BOUND + get_regex_charset(RExC_flags);
11276 if (op > BOUNDA) { /* /aa is same as /a */
11279 ret = reg_node(pRExC_state, op);
11280 FLAGS(ret) = get_regex_charset(RExC_flags);
11282 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11283 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
11285 goto finish_meta_pat;
11287 RExC_seen_zerolen++;
11288 RExC_seen |= REG_LOOKBEHIND_SEEN;
11289 op = NBOUND + get_regex_charset(RExC_flags);
11290 if (op > NBOUNDA) { /* /aa is same as /a */
11293 ret = reg_node(pRExC_state, op);
11294 FLAGS(ret) = get_regex_charset(RExC_flags);
11296 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11297 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
11299 goto finish_meta_pat;
11309 ret = reg_node(pRExC_state, LNBREAK);
11310 *flagp |= HASWIDTH|SIMPLE;
11311 goto finish_meta_pat;
11319 goto join_posix_op_known;
11325 arg = ANYOF_VERTWS;
11327 goto join_posix_op_known;
11337 op = POSIXD + get_regex_charset(RExC_flags);
11338 if (op > POSIXA) { /* /aa is same as /a */
11342 join_posix_op_known:
11345 op += NPOSIXD - POSIXD;
11348 ret = reg_node(pRExC_state, op);
11350 FLAGS(ret) = namedclass_to_classnum(arg);
11353 *flagp |= HASWIDTH|SIMPLE;
11357 nextchar(pRExC_state);
11358 Set_Node_Length(ret, 2); /* MJD */
11364 char* parse_start = RExC_parse - 2;
11369 ret = regclass(pRExC_state, flagp,depth+1,
11370 TRUE, /* means just parse this element */
11371 FALSE, /* don't allow multi-char folds */
11372 FALSE, /* don't silence non-portable warnings.
11373 It would be a bug if these returned
11376 /* regclass() can only return RESTART_UTF8 if multi-char folds
11379 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11384 Set_Node_Offset(ret, parse_start + 2);
11385 Set_Node_Cur_Length(ret, parse_start);
11386 nextchar(pRExC_state);
11390 /* Handle \N and \N{NAME} with multiple code points here and not
11391 * below because it can be multicharacter. join_exact() will join
11392 * them up later on. Also this makes sure that things like
11393 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11394 * The options to the grok function call causes it to fail if the
11395 * sequence is just a single code point. We then go treat it as
11396 * just another character in the current EXACT node, and hence it
11397 * gets uniform treatment with all the other characters. The
11398 * special treatment for quantifiers is not needed for such single
11399 * character sequences */
11401 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
11402 FALSE /* not strict */ )) {
11403 if (*flagp & RESTART_UTF8)
11409 case 'k': /* Handle \k<NAME> and \k'NAME' */
11412 char ch= RExC_parse[1];
11413 if (ch != '<' && ch != '\'' && ch != '{') {
11415 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11416 vFAIL2("Sequence %.2s... not terminated",parse_start);
11418 /* this pretty much dupes the code for (?P=...) in reg(), if
11419 you change this make sure you change that */
11420 char* name_start = (RExC_parse += 2);
11422 SV *sv_dat = reg_scan_name(pRExC_state,
11423 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11424 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11425 if (RExC_parse == name_start || *RExC_parse != ch)
11426 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11427 vFAIL2("Sequence %.3s... not terminated",parse_start);
11430 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11431 RExC_rxi->data->data[num]=(void*)sv_dat;
11432 SvREFCNT_inc_simple_void(sv_dat);
11436 ret = reganode(pRExC_state,
11439 : (ASCII_FOLD_RESTRICTED)
11441 : (AT_LEAST_UNI_SEMANTICS)
11447 *flagp |= HASWIDTH;
11449 /* override incorrect value set in reganode MJD */
11450 Set_Node_Offset(ret, parse_start+1);
11451 Set_Node_Cur_Length(ret, parse_start);
11452 nextchar(pRExC_state);
11458 case '1': case '2': case '3': case '4':
11459 case '5': case '6': case '7': case '8': case '9':
11464 if (*RExC_parse == 'g') {
11468 if (*RExC_parse == '{') {
11472 if (*RExC_parse == '-') {
11476 if (hasbrace && !isDIGIT(*RExC_parse)) {
11477 if (isrel) RExC_parse--;
11479 goto parse_named_seq;
11482 num = S_backref_value(RExC_parse);
11484 vFAIL("Reference to invalid group 0");
11485 else if (num == I32_MAX) {
11486 if (isDIGIT(*RExC_parse))
11487 vFAIL("Reference to nonexistent group");
11489 vFAIL("Unterminated \\g... pattern");
11493 num = RExC_npar - num;
11495 vFAIL("Reference to nonexistent or unclosed group");
11499 num = S_backref_value(RExC_parse);
11500 /* bare \NNN might be backref or octal */
11501 if (num == I32_MAX || (num > 9 && num >= RExC_npar
11502 && *RExC_parse != '8' && *RExC_parse != '9'))
11503 /* Probably a character specified in octal, e.g. \35 */
11507 /* at this point RExC_parse definitely points to a backref
11510 #ifdef RE_TRACK_PATTERN_OFFSETS
11511 char * const parse_start = RExC_parse - 1; /* MJD */
11513 while (isDIGIT(*RExC_parse))
11516 if (*RExC_parse != '}')
11517 vFAIL("Unterminated \\g{...} pattern");
11521 if (num > (I32)RExC_rx->nparens)
11522 vFAIL("Reference to nonexistent group");
11525 ret = reganode(pRExC_state,
11528 : (ASCII_FOLD_RESTRICTED)
11530 : (AT_LEAST_UNI_SEMANTICS)
11536 *flagp |= HASWIDTH;
11538 /* override incorrect value set in reganode MJD */
11539 Set_Node_Offset(ret, parse_start+1);
11540 Set_Node_Cur_Length(ret, parse_start);
11542 nextchar(pRExC_state);
11547 if (RExC_parse >= RExC_end)
11548 FAIL("Trailing \\");
11551 /* Do not generate "unrecognized" warnings here, we fall
11552 back into the quick-grab loop below */
11559 if (RExC_flags & RXf_PMf_EXTENDED) {
11560 if ( reg_skipcomment( pRExC_state ) )
11567 parse_start = RExC_parse - 1;
11576 #define MAX_NODE_STRING_SIZE 127
11577 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
11579 U8 upper_parse = MAX_NODE_STRING_SIZE;
11580 U8 node_type = compute_EXACTish(pRExC_state);
11581 bool next_is_quantifier;
11582 char * oldp = NULL;
11584 /* We can convert EXACTF nodes to EXACTFU if they contain only
11585 * characters that match identically regardless of the target
11586 * string's UTF8ness. The reason to do this is that EXACTF is not
11587 * trie-able, EXACTFU is.
11589 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
11590 * contain only above-Latin1 characters (hence must be in UTF8),
11591 * which don't participate in folds with Latin1-range characters,
11592 * as the latter's folds aren't known until runtime. (We don't
11593 * need to figure this out until pass 2) */
11594 bool maybe_exactfu = PASS2
11595 && (node_type == EXACTF || node_type == EXACTFL);
11597 /* If a folding node contains only code points that don't
11598 * participate in folds, it can be changed into an EXACT node,
11599 * which allows the optimizer more things to look for */
11602 ret = reg_node(pRExC_state, node_type);
11604 /* In pass1, folded, we use a temporary buffer instead of the
11605 * actual node, as the node doesn't exist yet */
11606 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
11612 /* We do the EXACTFish to EXACT node only if folding. (And we
11613 * don't need to figure this out until pass 2) */
11614 maybe_exact = FOLD && PASS2;
11616 /* XXX The node can hold up to 255 bytes, yet this only goes to
11617 * 127. I (khw) do not know why. Keeping it somewhat less than
11618 * 255 allows us to not have to worry about overflow due to
11619 * converting to utf8 and fold expansion, but that value is
11620 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
11621 * split up by this limit into a single one using the real max of
11622 * 255. Even at 127, this breaks under rare circumstances. If
11623 * folding, we do not want to split a node at a character that is a
11624 * non-final in a multi-char fold, as an input string could just
11625 * happen to want to match across the node boundary. The join
11626 * would solve that problem if the join actually happens. But a
11627 * series of more than two nodes in a row each of 127 would cause
11628 * the first join to succeed to get to 254, but then there wouldn't
11629 * be room for the next one, which could at be one of those split
11630 * multi-char folds. I don't know of any fool-proof solution. One
11631 * could back off to end with only a code point that isn't such a
11632 * non-final, but it is possible for there not to be any in the
11634 for (p = RExC_parse - 1;
11635 len < upper_parse && p < RExC_end;
11640 if (RExC_flags & RXf_PMf_EXTENDED)
11641 p = regwhite( pRExC_state, p );
11652 /* Literal Escapes Switch
11654 This switch is meant to handle escape sequences that
11655 resolve to a literal character.
11657 Every escape sequence that represents something
11658 else, like an assertion or a char class, is handled
11659 in the switch marked 'Special Escapes' above in this
11660 routine, but also has an entry here as anything that
11661 isn't explicitly mentioned here will be treated as
11662 an unescaped equivalent literal.
11665 switch ((U8)*++p) {
11666 /* These are all the special escapes. */
11667 case 'A': /* Start assertion */
11668 case 'b': case 'B': /* Word-boundary assertion*/
11669 case 'C': /* Single char !DANGEROUS! */
11670 case 'd': case 'D': /* digit class */
11671 case 'g': case 'G': /* generic-backref, pos assertion */
11672 case 'h': case 'H': /* HORIZWS */
11673 case 'k': case 'K': /* named backref, keep marker */
11674 case 'p': case 'P': /* Unicode property */
11675 case 'R': /* LNBREAK */
11676 case 's': case 'S': /* space class */
11677 case 'v': case 'V': /* VERTWS */
11678 case 'w': case 'W': /* word class */
11679 case 'X': /* eXtended Unicode "combining
11680 character sequence" */
11681 case 'z': case 'Z': /* End of line/string assertion */
11685 /* Anything after here is an escape that resolves to a
11686 literal. (Except digits, which may or may not)
11692 case 'N': /* Handle a single-code point named character. */
11693 /* The options cause it to fail if a multiple code
11694 * point sequence. Handle those in the switch() above
11696 RExC_parse = p + 1;
11697 if (! grok_bslash_N(pRExC_state, NULL, &ender,
11698 flagp, depth, FALSE,
11699 FALSE /* not strict */ ))
11701 if (*flagp & RESTART_UTF8)
11702 FAIL("panic: grok_bslash_N set RESTART_UTF8");
11703 RExC_parse = p = oldp;
11707 if (ender > 0xff) {
11724 ender = ASCII_TO_NATIVE('\033');
11734 const char* error_msg;
11736 bool valid = grok_bslash_o(&p,
11739 TRUE, /* out warnings */
11740 FALSE, /* not strict */
11741 TRUE, /* Output warnings
11746 RExC_parse = p; /* going to die anyway; point
11747 to exact spot of failure */
11751 if (PL_encoding && ender < 0x100) {
11752 goto recode_encoding;
11754 if (ender > 0xff) {
11761 UV result = UV_MAX; /* initialize to erroneous
11763 const char* error_msg;
11765 bool valid = grok_bslash_x(&p,
11768 TRUE, /* out warnings */
11769 FALSE, /* not strict */
11770 TRUE, /* Output warnings
11775 RExC_parse = p; /* going to die anyway; point
11776 to exact spot of failure */
11781 if (PL_encoding && ender < 0x100) {
11782 goto recode_encoding;
11784 if (ender > 0xff) {
11791 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
11793 case '8': case '9': /* must be a backreference */
11796 case '1': case '2': case '3':case '4':
11797 case '5': case '6': case '7':
11798 /* When we parse backslash escapes there is ambiguity
11799 * between backreferences and octal escapes. Any escape
11800 * from \1 - \9 is a backreference, any multi-digit
11801 * escape which does not start with 0 and which when
11802 * evaluated as decimal could refer to an already
11803 * parsed capture buffer is a backslash. Anything else
11806 * Note this implies that \118 could be interpreted as
11807 * 118 OR as "\11" . "8" depending on whether there
11808 * were 118 capture buffers defined already in the
11810 if ( !isDIGIT(p[1]) || S_backref_value(p) <= RExC_npar)
11811 { /* Not to be treated as an octal constant, go
11818 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11820 ender = grok_oct(p, &numlen, &flags, NULL);
11821 if (ender > 0xff) {
11825 if (SIZE_ONLY /* like \08, \178 */
11828 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11830 reg_warn_non_literal_string(
11832 form_short_octal_warning(p, numlen));
11835 if (PL_encoding && ender < 0x100)
11836 goto recode_encoding;
11839 if (! RExC_override_recoding) {
11840 SV* enc = PL_encoding;
11841 ender = reg_recode((const char)(U8)ender, &enc);
11842 if (!enc && SIZE_ONLY)
11843 ckWARNreg(p, "Invalid escape in the specified encoding");
11849 FAIL("Trailing \\");
11852 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11853 /* Include any { following the alpha to emphasize
11854 * that it could be part of an escape at some point
11856 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11857 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11859 goto normal_default;
11860 } /* End of switch on '\' */
11862 default: /* A literal character */
11865 && RExC_flags & RXf_PMf_EXTENDED
11866 && ckWARN_d(WARN_DEPRECATED)
11867 && is_PATWS_non_low(p, UTF))
11869 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11870 "Escape literal pattern white space under /x");
11874 if (UTF8_IS_START(*p) && UTF) {
11876 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11877 &numlen, UTF8_ALLOW_DEFAULT);
11883 } /* End of switch on the literal */
11885 /* Here, have looked at the literal character and <ender>
11886 * contains its ordinal, <p> points to the character after it
11889 if ( RExC_flags & RXf_PMf_EXTENDED)
11890 p = regwhite( pRExC_state, p );
11892 /* If the next thing is a quantifier, it applies to this
11893 * character only, which means that this character has to be in
11894 * its own node and can't just be appended to the string in an
11895 * existing node, so if there are already other characters in
11896 * the node, close the node with just them, and set up to do
11897 * this character again next time through, when it will be the
11898 * only thing in its new node */
11899 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11905 if (! FOLD /* The simple case, just append the literal */
11906 || (LOC /* Also don't fold for tricky chars under /l */
11907 && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)))
11910 const STRLEN unilen = reguni(pRExC_state, ender, s);
11916 /* The loop increments <len> each time, as all but this
11917 * path (and one other) through it add a single byte to
11918 * the EXACTish node. But this one has changed len to
11919 * be the correct final value, so subtract one to
11920 * cancel out the increment that follows */
11924 REGC((char)ender, s++);
11927 /* Can get here if folding only if is one of the /l
11928 * characters whose fold depends on the locale. The
11929 * occurrence of any of these indicate that we can't
11930 * simplify things */
11932 maybe_exact = FALSE;
11933 maybe_exactfu = FALSE;
11938 /* See comments for join_exact() as to why we fold this
11939 * non-UTF at compile time */
11940 || (node_type == EXACTFU
11941 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11943 /* Here, are folding and are not UTF-8 encoded; therefore
11944 * the character must be in the range 0-255, and is not /l
11945 * (Not /l because we already handled these under /l in
11946 * is_PROBLEMATIC_LOCALE_FOLD_cp */
11947 if (IS_IN_SOME_FOLD_L1(ender)) {
11948 maybe_exact = FALSE;
11950 /* See if the character's fold differs between /d and
11951 * /u. This includes the multi-char fold SHARP S to
11954 && (PL_fold[ender] != PL_fold_latin1[ender]
11955 || ender == LATIN_SMALL_LETTER_SHARP_S
11957 && isARG2_lower_or_UPPER_ARG1('s', ender)
11958 && isARG2_lower_or_UPPER_ARG1('s',
11961 maybe_exactfu = FALSE;
11965 /* Even when folding, we store just the input character, as
11966 * we have an array that finds its fold quickly */
11967 *(s++) = (char) ender;
11969 else { /* FOLD and UTF */
11970 /* Unlike the non-fold case, we do actually have to
11971 * calculate the results here in pass 1. This is for two
11972 * reasons, the folded length may be longer than the
11973 * unfolded, and we have to calculate how many EXACTish
11974 * nodes it will take; and we may run out of room in a node
11975 * in the middle of a potential multi-char fold, and have
11976 * to back off accordingly. (Hence we can't use REGC for
11977 * the simple case just below.) */
11980 if (isASCII(ender)) {
11981 folded = toFOLD(ender);
11982 *(s)++ = (U8) folded;
11987 folded = _to_uni_fold_flags(
11991 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11992 ? FOLD_FLAGS_NOMIX_ASCII
11996 /* The loop increments <len> each time, as all but this
11997 * path (and one other) through it add a single byte to
11998 * the EXACTish node. But this one has changed len to
11999 * be the correct final value, so subtract one to
12000 * cancel out the increment that follows */
12001 len += foldlen - 1;
12003 /* If this node only contains non-folding code points so
12004 * far, see if this new one is also non-folding */
12006 if (folded != ender) {
12007 maybe_exact = FALSE;
12010 /* Here the fold is the original; we have to check
12011 * further to see if anything folds to it */
12012 if (_invlist_contains_cp(PL_utf8_foldable,
12015 maybe_exact = FALSE;
12022 if (next_is_quantifier) {
12024 /* Here, the next input is a quantifier, and to get here,
12025 * the current character is the only one in the node.
12026 * Also, here <len> doesn't include the final byte for this
12032 } /* End of loop through literal characters */
12034 /* Here we have either exhausted the input or ran out of room in
12035 * the node. (If we encountered a character that can't be in the
12036 * node, transfer is made directly to <loopdone>, and so we
12037 * wouldn't have fallen off the end of the loop.) In the latter
12038 * case, we artificially have to split the node into two, because
12039 * we just don't have enough space to hold everything. This
12040 * creates a problem if the final character participates in a
12041 * multi-character fold in the non-final position, as a match that
12042 * should have occurred won't, due to the way nodes are matched,
12043 * and our artificial boundary. So back off until we find a non-
12044 * problematic character -- one that isn't at the beginning or
12045 * middle of such a fold. (Either it doesn't participate in any
12046 * folds, or appears only in the final position of all the folds it
12047 * does participate in.) A better solution with far fewer false
12048 * positives, and that would fill the nodes more completely, would
12049 * be to actually have available all the multi-character folds to
12050 * test against, and to back-off only far enough to be sure that
12051 * this node isn't ending with a partial one. <upper_parse> is set
12052 * further below (if we need to reparse the node) to include just
12053 * up through that final non-problematic character that this code
12054 * identifies, so when it is set to less than the full node, we can
12055 * skip the rest of this */
12056 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12058 const STRLEN full_len = len;
12060 assert(len >= MAX_NODE_STRING_SIZE);
12062 /* Here, <s> points to the final byte of the final character.
12063 * Look backwards through the string until find a non-
12064 * problematic character */
12068 /* This has no multi-char folds to non-UTF characters */
12069 if (ASCII_FOLD_RESTRICTED) {
12073 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12077 if (! PL_NonL1NonFinalFold) {
12078 PL_NonL1NonFinalFold = _new_invlist_C_array(
12079 NonL1_Perl_Non_Final_Folds_invlist);
12082 /* Point to the first byte of the final character */
12083 s = (char *) utf8_hop((U8 *) s, -1);
12085 while (s >= s0) { /* Search backwards until find
12086 non-problematic char */
12087 if (UTF8_IS_INVARIANT(*s)) {
12089 /* There are no ascii characters that participate
12090 * in multi-char folds under /aa. In EBCDIC, the
12091 * non-ascii invariants are all control characters,
12092 * so don't ever participate in any folds. */
12093 if (ASCII_FOLD_RESTRICTED
12094 || ! IS_NON_FINAL_FOLD(*s))
12099 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12100 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12106 else if (! _invlist_contains_cp(
12107 PL_NonL1NonFinalFold,
12108 valid_utf8_to_uvchr((U8 *) s, NULL)))
12113 /* Here, the current character is problematic in that
12114 * it does occur in the non-final position of some
12115 * fold, so try the character before it, but have to
12116 * special case the very first byte in the string, so
12117 * we don't read outside the string */
12118 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12119 } /* End of loop backwards through the string */
12121 /* If there were only problematic characters in the string,
12122 * <s> will point to before s0, in which case the length
12123 * should be 0, otherwise include the length of the
12124 * non-problematic character just found */
12125 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12128 /* Here, have found the final character, if any, that is
12129 * non-problematic as far as ending the node without splitting
12130 * it across a potential multi-char fold. <len> contains the
12131 * number of bytes in the node up-to and including that
12132 * character, or is 0 if there is no such character, meaning
12133 * the whole node contains only problematic characters. In
12134 * this case, give up and just take the node as-is. We can't
12139 /* If the node ends in an 's' we make sure it stays EXACTF,
12140 * as if it turns into an EXACTFU, it could later get
12141 * joined with another 's' that would then wrongly match
12143 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
12145 maybe_exactfu = FALSE;
12149 /* Here, the node does contain some characters that aren't
12150 * problematic. If one such is the final character in the
12151 * node, we are done */
12152 if (len == full_len) {
12155 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12157 /* If the final character is problematic, but the
12158 * penultimate is not, back-off that last character to
12159 * later start a new node with it */
12164 /* Here, the final non-problematic character is earlier
12165 * in the input than the penultimate character. What we do
12166 * is reparse from the beginning, going up only as far as
12167 * this final ok one, thus guaranteeing that the node ends
12168 * in an acceptable character. The reason we reparse is
12169 * that we know how far in the character is, but we don't
12170 * know how to correlate its position with the input parse.
12171 * An alternate implementation would be to build that
12172 * correlation as we go along during the original parse,
12173 * but that would entail extra work for every node, whereas
12174 * this code gets executed only when the string is too
12175 * large for the node, and the final two characters are
12176 * problematic, an infrequent occurrence. Yet another
12177 * possible strategy would be to save the tail of the
12178 * string, and the next time regatom is called, initialize
12179 * with that. The problem with this is that unless you
12180 * back off one more character, you won't be guaranteed
12181 * regatom will get called again, unless regbranch,
12182 * regpiece ... are also changed. If you do back off that
12183 * extra character, so that there is input guaranteed to
12184 * force calling regatom, you can't handle the case where
12185 * just the first character in the node is acceptable. I
12186 * (khw) decided to try this method which doesn't have that
12187 * pitfall; if performance issues are found, we can do a
12188 * combination of the current approach plus that one */
12194 } /* End of verifying node ends with an appropriate char */
12196 loopdone: /* Jumped to when encounters something that shouldn't be in
12199 /* I (khw) don't know if you can get here with zero length, but the
12200 * old code handled this situation by creating a zero-length EXACT
12201 * node. Might as well be NOTHING instead */
12207 /* If 'maybe_exact' is still set here, means there are no
12208 * code points in the node that participate in folds;
12209 * similarly for 'maybe_exactfu' and code points that match
12210 * differently depending on UTF8ness of the target string
12211 * (for /u), or depending on locale for /l */
12215 else if (maybe_exactfu) {
12219 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
12222 RExC_parse = p - 1;
12223 Set_Node_Cur_Length(ret, parse_start);
12224 nextchar(pRExC_state);
12226 /* len is STRLEN which is unsigned, need to copy to signed */
12229 vFAIL("Internal disaster");
12232 } /* End of label 'defchar:' */
12234 } /* End of giant switch on input character */
12240 S_regwhite( RExC_state_t *pRExC_state, char *p )
12242 const char *e = RExC_end;
12244 PERL_ARGS_ASSERT_REGWHITE;
12249 else if (*p == '#') {
12252 if (*p++ == '\n') {
12258 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12267 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12269 /* Returns the next non-pattern-white space, non-comment character (the
12270 * latter only if 'recognize_comment is true) in the string p, which is
12271 * ended by RExC_end. If there is no line break ending a comment,
12272 * RExC_seen has added the REG_RUN_ON_COMMENT_SEEN flag; */
12273 const char *e = RExC_end;
12275 PERL_ARGS_ASSERT_REGPATWS;
12279 if ((len = is_PATWS_safe(p, e, UTF))) {
12282 else if (recognize_comment && *p == '#') {
12286 if (is_LNBREAK_safe(p, e, UTF)) {
12292 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12301 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12303 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12304 * sets up the bitmap and any flags, removing those code points from the
12305 * inversion list, setting it to NULL should it become completely empty */
12307 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12308 assert(PL_regkind[OP(node)] == ANYOF);
12310 ANYOF_BITMAP_ZERO(node);
12311 if (*invlist_ptr) {
12313 /* This gets set if we actually need to modify things */
12314 bool change_invlist = FALSE;
12318 /* Start looking through *invlist_ptr */
12319 invlist_iterinit(*invlist_ptr);
12320 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12324 if (end == UV_MAX && start <= 256) {
12325 ANYOF_FLAGS(node) |= ANYOF_ABOVE_LATIN1_ALL;
12328 /* Quit if are above what we should change */
12333 change_invlist = TRUE;
12335 /* Set all the bits in the range, up to the max that we are doing */
12336 high = (end < 255) ? end : 255;
12337 for (i = start; i <= (int) high; i++) {
12338 if (! ANYOF_BITMAP_TEST(node, i)) {
12339 ANYOF_BITMAP_SET(node, i);
12343 invlist_iterfinish(*invlist_ptr);
12345 /* Done with loop; remove any code points that are in the bitmap from
12346 * *invlist_ptr; similarly for code points above latin1 if we have a
12347 * flag to match all of them anyways */
12348 if (change_invlist) {
12349 _invlist_subtract(*invlist_ptr, PL_Latin1, invlist_ptr);
12351 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
12352 _invlist_intersection(*invlist_ptr, PL_Latin1, invlist_ptr);
12355 /* If have completely emptied it, remove it completely */
12356 if (_invlist_len(*invlist_ptr) == 0) {
12357 SvREFCNT_dec_NN(*invlist_ptr);
12358 *invlist_ptr = NULL;
12363 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12364 Character classes ([:foo:]) can also be negated ([:^foo:]).
12365 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12366 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12367 but trigger failures because they are currently unimplemented. */
12369 #define POSIXCC_DONE(c) ((c) == ':')
12370 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12371 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12373 PERL_STATIC_INLINE I32
12374 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12377 I32 namedclass = OOB_NAMEDCLASS;
12379 PERL_ARGS_ASSERT_REGPPOSIXCC;
12381 if (value == '[' && RExC_parse + 1 < RExC_end &&
12382 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12383 POSIXCC(UCHARAT(RExC_parse)))
12385 const char c = UCHARAT(RExC_parse);
12386 char* const s = RExC_parse++;
12388 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12390 if (RExC_parse == RExC_end) {
12393 /* Try to give a better location for the error (than the end of
12394 * the string) by looking for the matching ']' */
12396 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12399 vFAIL2("Unmatched '%c' in POSIX class", c);
12401 /* Grandfather lone [:, [=, [. */
12405 const char* const t = RExC_parse++; /* skip over the c */
12408 if (UCHARAT(RExC_parse) == ']') {
12409 const char *posixcc = s + 1;
12410 RExC_parse++; /* skip over the ending ] */
12413 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12414 const I32 skip = t - posixcc;
12416 /* Initially switch on the length of the name. */
12419 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12420 this is the Perl \w
12422 namedclass = ANYOF_WORDCHAR;
12425 /* Names all of length 5. */
12426 /* alnum alpha ascii blank cntrl digit graph lower
12427 print punct space upper */
12428 /* Offset 4 gives the best switch position. */
12429 switch (posixcc[4]) {
12431 if (memEQ(posixcc, "alph", 4)) /* alpha */
12432 namedclass = ANYOF_ALPHA;
12435 if (memEQ(posixcc, "spac", 4)) /* space */
12436 namedclass = ANYOF_PSXSPC;
12439 if (memEQ(posixcc, "grap", 4)) /* graph */
12440 namedclass = ANYOF_GRAPH;
12443 if (memEQ(posixcc, "asci", 4)) /* ascii */
12444 namedclass = ANYOF_ASCII;
12447 if (memEQ(posixcc, "blan", 4)) /* blank */
12448 namedclass = ANYOF_BLANK;
12451 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12452 namedclass = ANYOF_CNTRL;
12455 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12456 namedclass = ANYOF_ALPHANUMERIC;
12459 if (memEQ(posixcc, "lowe", 4)) /* lower */
12460 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12461 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12462 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12465 if (memEQ(posixcc, "digi", 4)) /* digit */
12466 namedclass = ANYOF_DIGIT;
12467 else if (memEQ(posixcc, "prin", 4)) /* print */
12468 namedclass = ANYOF_PRINT;
12469 else if (memEQ(posixcc, "punc", 4)) /* punct */
12470 namedclass = ANYOF_PUNCT;
12475 if (memEQ(posixcc, "xdigit", 6))
12476 namedclass = ANYOF_XDIGIT;
12480 if (namedclass == OOB_NAMEDCLASS)
12482 "POSIX class [:%"UTF8f":] unknown",
12483 UTF8fARG(UTF, t - s - 1, s + 1));
12485 /* The #defines are structured so each complement is +1 to
12486 * the normal one */
12490 assert (posixcc[skip] == ':');
12491 assert (posixcc[skip+1] == ']');
12492 } else if (!SIZE_ONLY) {
12493 /* [[=foo=]] and [[.foo.]] are still future. */
12495 /* adjust RExC_parse so the warning shows after
12496 the class closes */
12497 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
12499 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
12502 /* Maternal grandfather:
12503 * "[:" ending in ":" but not in ":]" */
12505 vFAIL("Unmatched '[' in POSIX class");
12508 /* Grandfather lone [:, [=, [. */
12518 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
12520 /* This applies some heuristics at the current parse position (which should
12521 * be at a '[') to see if what follows might be intended to be a [:posix:]
12522 * class. It returns true if it really is a posix class, of course, but it
12523 * also can return true if it thinks that what was intended was a posix
12524 * class that didn't quite make it.
12526 * It will return true for
12528 * [:alphanumerics] (as long as the ] isn't followed immediately by a
12529 * ')' indicating the end of the (?[
12530 * [:any garbage including %^&$ punctuation:]
12532 * This is designed to be called only from S_handle_regex_sets; it could be
12533 * easily adapted to be called from the spot at the beginning of regclass()
12534 * that checks to see in a normal bracketed class if the surrounding []
12535 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
12536 * change long-standing behavior, so I (khw) didn't do that */
12537 char* p = RExC_parse + 1;
12538 char first_char = *p;
12540 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
12542 assert(*(p - 1) == '[');
12544 if (! POSIXCC(first_char)) {
12549 while (p < RExC_end && isWORDCHAR(*p)) p++;
12551 if (p >= RExC_end) {
12555 if (p - RExC_parse > 2 /* Got at least 1 word character */
12556 && (*p == first_char
12557 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
12562 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
12565 && p - RExC_parse > 2 /* [:] evaluates to colon;
12566 [::] is a bad posix class. */
12567 && first_char == *(p - 1));
12571 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
12572 I32 *flagp, U32 depth,
12573 char * const oregcomp_parse)
12575 /* Handle the (?[...]) construct to do set operations */
12578 UV start, end; /* End points of code point ranges */
12580 char *save_end, *save_parse;
12585 const bool save_fold = FOLD;
12587 GET_RE_DEBUG_FLAGS_DECL;
12589 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
12592 vFAIL("(?[...]) not valid in locale");
12594 RExC_uni_semantics = 1;
12596 /* This will return only an ANYOF regnode, or (unlikely) something smaller
12597 * (such as EXACT). Thus we can skip most everything if just sizing. We
12598 * call regclass to handle '[]' so as to not have to reinvent its parsing
12599 * rules here (throwing away the size it computes each time). And, we exit
12600 * upon an unescaped ']' that isn't one ending a regclass. To do both
12601 * these things, we need to realize that something preceded by a backslash
12602 * is escaped, so we have to keep track of backslashes */
12604 UV depth = 0; /* how many nested (?[...]) constructs */
12606 Perl_ck_warner_d(aTHX_
12607 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
12608 "The regex_sets feature is experimental" REPORT_LOCATION,
12609 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
12611 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
12612 RExC_precomp + (RExC_parse - RExC_precomp)));
12614 while (RExC_parse < RExC_end) {
12615 SV* current = NULL;
12616 RExC_parse = regpatws(pRExC_state, RExC_parse,
12617 TRUE); /* means recognize comments */
12618 switch (*RExC_parse) {
12620 if (RExC_parse[1] == '[') depth++, RExC_parse++;
12625 /* Skip the next byte (which could cause us to end up in
12626 * the middle of a UTF-8 character, but since none of those
12627 * are confusable with anything we currently handle in this
12628 * switch (invariants all), it's safe. We'll just hit the
12629 * default: case next time and keep on incrementing until
12630 * we find one of the invariants we do handle. */
12635 /* If this looks like it is a [:posix:] class, leave the
12636 * parse pointer at the '[' to fool regclass() into
12637 * thinking it is part of a '[[:posix:]]'. That function
12638 * will use strict checking to force a syntax error if it
12639 * doesn't work out to a legitimate class */
12640 bool is_posix_class
12641 = could_it_be_a_POSIX_class(pRExC_state);
12642 if (! is_posix_class) {
12646 /* regclass() can only return RESTART_UTF8 if multi-char
12647 folds are allowed. */
12648 if (!regclass(pRExC_state, flagp,depth+1,
12649 is_posix_class, /* parse the whole char
12650 class only if not a
12652 FALSE, /* don't allow multi-char folds */
12653 TRUE, /* silence non-portable warnings. */
12655 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12658 /* function call leaves parse pointing to the ']', except
12659 * if we faked it */
12660 if (is_posix_class) {
12664 SvREFCNT_dec(current); /* In case it returned something */
12669 if (depth--) break;
12671 if (RExC_parse < RExC_end
12672 && *RExC_parse == ')')
12674 node = reganode(pRExC_state, ANYOF, 0);
12675 RExC_size += ANYOF_SKIP;
12676 nextchar(pRExC_state);
12677 Set_Node_Length(node,
12678 RExC_parse - oregcomp_parse + 1); /* MJD */
12687 FAIL("Syntax error in (?[...])");
12690 /* Pass 2 only after this. Everything in this construct is a
12691 * metacharacter. Operands begin with either a '\' (for an escape
12692 * sequence), or a '[' for a bracketed character class. Any other
12693 * character should be an operator, or parenthesis for grouping. Both
12694 * types of operands are handled by calling regclass() to parse them. It
12695 * is called with a parameter to indicate to return the computed inversion
12696 * list. The parsing here is implemented via a stack. Each entry on the
12697 * stack is a single character representing one of the operators, or the
12698 * '('; or else a pointer to an operand inversion list. */
12700 #define IS_OPERAND(a) (! SvIOK(a))
12702 /* The stack starts empty. It is a syntax error if the first thing parsed
12703 * is a binary operator; everything else is pushed on the stack. When an
12704 * operand is parsed, the top of the stack is examined. If it is a binary
12705 * operator, the item before it should be an operand, and both are replaced
12706 * by the result of doing that operation on the new operand and the one on
12707 * the stack. Thus a sequence of binary operands is reduced to a single
12708 * one before the next one is parsed.
12710 * A unary operator may immediately follow a binary in the input, for
12713 * When an operand is parsed and the top of the stack is a unary operator,
12714 * the operation is performed, and then the stack is rechecked to see if
12715 * this new operand is part of a binary operation; if so, it is handled as
12718 * A '(' is simply pushed on the stack; it is valid only if the stack is
12719 * empty, or the top element of the stack is an operator or another '('
12720 * (for which the parenthesized expression will become an operand). By the
12721 * time the corresponding ')' is parsed everything in between should have
12722 * been parsed and evaluated to a single operand (or else is a syntax
12723 * error), and is handled as a regular operand */
12725 sv_2mortal((SV *)(stack = newAV()));
12727 while (RExC_parse < RExC_end) {
12728 I32 top_index = av_tindex(stack);
12730 SV* current = NULL;
12732 /* Skip white space */
12733 RExC_parse = regpatws(pRExC_state, RExC_parse,
12734 TRUE); /* means recognize comments */
12735 if (RExC_parse >= RExC_end) {
12736 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
12738 if ((curchar = UCHARAT(RExC_parse)) == ']') {
12745 if (av_tindex(stack) >= 0 /* This makes sure that we can
12746 safely subtract 1 from
12747 RExC_parse in the next clause.
12748 If we have something on the
12749 stack, we have parsed something
12751 && UCHARAT(RExC_parse - 1) == '('
12752 && RExC_parse < RExC_end)
12754 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
12755 * This happens when we have some thing like
12757 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
12759 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
12761 * Here we would be handling the interpolated
12762 * '$thai_or_lao'. We handle this by a recursive call to
12763 * ourselves which returns the inversion list the
12764 * interpolated expression evaluates to. We use the flags
12765 * from the interpolated pattern. */
12766 U32 save_flags = RExC_flags;
12767 const char * const save_parse = ++RExC_parse;
12769 parse_lparen_question_flags(pRExC_state);
12771 if (RExC_parse == save_parse /* Makes sure there was at
12772 least one flag (or this
12773 embedding wasn't compiled)
12775 || RExC_parse >= RExC_end - 4
12776 || UCHARAT(RExC_parse) != ':'
12777 || UCHARAT(++RExC_parse) != '('
12778 || UCHARAT(++RExC_parse) != '?'
12779 || UCHARAT(++RExC_parse) != '[')
12782 /* In combination with the above, this moves the
12783 * pointer to the point just after the first erroneous
12784 * character (or if there are no flags, to where they
12785 * should have been) */
12786 if (RExC_parse >= RExC_end - 4) {
12787 RExC_parse = RExC_end;
12789 else if (RExC_parse != save_parse) {
12790 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12792 vFAIL("Expecting '(?flags:(?[...'");
12795 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
12796 depth+1, oregcomp_parse);
12798 /* Here, 'current' contains the embedded expression's
12799 * inversion list, and RExC_parse points to the trailing
12800 * ']'; the next character should be the ')' which will be
12801 * paired with the '(' that has been put on the stack, so
12802 * the whole embedded expression reduces to '(operand)' */
12805 RExC_flags = save_flags;
12806 goto handle_operand;
12811 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12812 vFAIL("Unexpected character");
12815 /* regclass() can only return RESTART_UTF8 if multi-char
12816 folds are allowed. */
12817 if (!regclass(pRExC_state, flagp,depth+1,
12818 TRUE, /* means parse just the next thing */
12819 FALSE, /* don't allow multi-char folds */
12820 FALSE, /* don't silence non-portable warnings. */
12822 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12824 /* regclass() will return with parsing just the \ sequence,
12825 * leaving the parse pointer at the next thing to parse */
12827 goto handle_operand;
12829 case '[': /* Is a bracketed character class */
12831 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
12833 if (! is_posix_class) {
12837 /* regclass() can only return RESTART_UTF8 if multi-char
12838 folds are allowed. */
12839 if(!regclass(pRExC_state, flagp,depth+1,
12840 is_posix_class, /* parse the whole char class
12841 only if not a posix class */
12842 FALSE, /* don't allow multi-char folds */
12843 FALSE, /* don't silence non-portable warnings. */
12845 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12847 /* function call leaves parse pointing to the ']', except if we
12849 if (is_posix_class) {
12853 goto handle_operand;
12862 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
12863 || ! IS_OPERAND(*top_ptr))
12866 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
12868 av_push(stack, newSVuv(curchar));
12872 av_push(stack, newSVuv(curchar));
12876 if (top_index >= 0) {
12877 top_ptr = av_fetch(stack, top_index, FALSE);
12879 if (IS_OPERAND(*top_ptr)) {
12881 vFAIL("Unexpected '(' with no preceding operator");
12884 av_push(stack, newSVuv(curchar));
12891 || ! (current = av_pop(stack))
12892 || ! IS_OPERAND(current)
12893 || ! (lparen = av_pop(stack))
12894 || IS_OPERAND(lparen)
12895 || SvUV(lparen) != '(')
12897 SvREFCNT_dec(current);
12899 vFAIL("Unexpected ')'");
12902 SvREFCNT_dec_NN(lparen);
12909 /* Here, we have an operand to process, in 'current' */
12911 if (top_index < 0) { /* Just push if stack is empty */
12912 av_push(stack, current);
12915 SV* top = av_pop(stack);
12917 char current_operator;
12919 if (IS_OPERAND(top)) {
12920 SvREFCNT_dec_NN(top);
12921 SvREFCNT_dec_NN(current);
12922 vFAIL("Operand with no preceding operator");
12924 current_operator = (char) SvUV(top);
12925 switch (current_operator) {
12926 case '(': /* Push the '(' back on followed by the new
12928 av_push(stack, top);
12929 av_push(stack, current);
12930 SvREFCNT_inc(top); /* Counters the '_dec' done
12931 just after the 'break', so
12932 it doesn't get wrongly freed
12937 _invlist_invert(current);
12939 /* Unlike binary operators, the top of the stack,
12940 * now that this unary one has been popped off, may
12941 * legally be an operator, and we now have operand
12944 SvREFCNT_dec_NN(top);
12945 goto handle_operand;
12948 prev = av_pop(stack);
12949 _invlist_intersection(prev,
12952 av_push(stack, current);
12957 prev = av_pop(stack);
12958 _invlist_union(prev, current, ¤t);
12959 av_push(stack, current);
12963 prev = av_pop(stack);;
12964 _invlist_subtract(prev, current, ¤t);
12965 av_push(stack, current);
12968 case '^': /* The union minus the intersection */
12974 prev = av_pop(stack);
12975 _invlist_union(prev, current, &u);
12976 _invlist_intersection(prev, current, &i);
12977 /* _invlist_subtract will overwrite current
12978 without freeing what it already contains */
12980 _invlist_subtract(u, i, ¤t);
12981 av_push(stack, current);
12982 SvREFCNT_dec_NN(i);
12983 SvREFCNT_dec_NN(u);
12984 SvREFCNT_dec_NN(element);
12989 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12991 SvREFCNT_dec_NN(top);
12992 SvREFCNT_dec(prev);
12996 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12999 if (av_tindex(stack) < 0 /* Was empty */
13000 || ((final = av_pop(stack)) == NULL)
13001 || ! IS_OPERAND(final)
13002 || av_tindex(stack) >= 0) /* More left on stack */
13004 vFAIL("Incomplete expression within '(?[ ])'");
13007 /* Here, 'final' is the resultant inversion list from evaluating the
13008 * expression. Return it if so requested */
13009 if (return_invlist) {
13010 *return_invlist = final;
13014 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13015 * expecting a string of ranges and individual code points */
13016 invlist_iterinit(final);
13017 result_string = newSVpvs("");
13018 while (invlist_iternext(final, &start, &end)) {
13019 if (start == end) {
13020 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13023 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13028 save_parse = RExC_parse;
13029 RExC_parse = SvPV(result_string, len);
13030 save_end = RExC_end;
13031 RExC_end = RExC_parse + len;
13033 /* We turn off folding around the call, as the class we have constructed
13034 * already has all folding taken into consideration, and we don't want
13035 * regclass() to add to that */
13036 RExC_flags &= ~RXf_PMf_FOLD;
13037 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13039 node = regclass(pRExC_state, flagp,depth+1,
13040 FALSE, /* means parse the whole char class */
13041 FALSE, /* don't allow multi-char folds */
13042 TRUE, /* silence non-portable warnings. The above may very
13043 well have generated non-portable code points, but
13044 they're valid on this machine */
13047 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13050 RExC_flags |= RXf_PMf_FOLD;
13052 RExC_parse = save_parse + 1;
13053 RExC_end = save_end;
13054 SvREFCNT_dec_NN(final);
13055 SvREFCNT_dec_NN(result_string);
13057 nextchar(pRExC_state);
13058 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13063 /* The names of properties whose definitions are not known at compile time are
13064 * stored in this SV, after a constant heading. So if the length has been
13065 * changed since initialization, then there is a run-time definition. */
13066 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13067 (SvCUR(listsv) != initial_listsv_len)
13070 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13071 const bool stop_at_1, /* Just parse the next thing, don't
13072 look for a full character class */
13073 bool allow_multi_folds,
13074 const bool silence_non_portable, /* Don't output warnings
13077 SV** ret_invlist) /* Return an inversion list, not a node */
13079 /* parse a bracketed class specification. Most of these will produce an
13080 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13081 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13082 * under /i with multi-character folds: it will be rewritten following the
13083 * paradigm of this example, where the <multi-fold>s are characters which
13084 * fold to multiple character sequences:
13085 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13086 * gets effectively rewritten as:
13087 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13088 * reg() gets called (recursively) on the rewritten version, and this
13089 * function will return what it constructs. (Actually the <multi-fold>s
13090 * aren't physically removed from the [abcdefghi], it's just that they are
13091 * ignored in the recursion by means of a flag:
13092 * <RExC_in_multi_char_class>.)
13094 * ANYOF nodes contain a bit map for the first 256 characters, with the
13095 * corresponding bit set if that character is in the list. For characters
13096 * above 255, a range list or swash is used. There are extra bits for \w,
13097 * etc. in locale ANYOFs, as what these match is not determinable at
13100 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13101 * to be restarted. This can only happen if ret_invlist is non-NULL.
13105 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13107 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13110 IV namedclass = OOB_NAMEDCLASS;
13111 char *rangebegin = NULL;
13112 bool need_class = 0;
13114 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13115 than just initialized. */
13116 SV* properties = NULL; /* Code points that match \p{} \P{} */
13117 SV* posixes = NULL; /* Code points that match classes like [:word:],
13118 extended beyond the Latin1 range. These have to
13119 be kept separate from other code points for much
13120 of this function because their handling is
13121 different under /i, and for most classes under
13123 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13124 separate for a while from the non-complemented
13125 versions because of complications with /d
13127 UV element_count = 0; /* Number of distinct elements in the class.
13128 Optimizations may be possible if this is tiny */
13129 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13130 character; used under /i */
13132 char * stop_ptr = RExC_end; /* where to stop parsing */
13133 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13135 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13137 /* Unicode properties are stored in a swash; this holds the current one
13138 * being parsed. If this swash is the only above-latin1 component of the
13139 * character class, an optimization is to pass it directly on to the
13140 * execution engine. Otherwise, it is set to NULL to indicate that there
13141 * are other things in the class that have to be dealt with at execution
13143 SV* swash = NULL; /* Code points that match \p{} \P{} */
13145 /* Set if a component of this character class is user-defined; just passed
13146 * on to the engine */
13147 bool has_user_defined_property = FALSE;
13149 /* inversion list of code points this node matches only when the target
13150 * string is in UTF-8. (Because is under /d) */
13151 SV* depends_list = NULL;
13153 /* Inversion list of code points this node matches regardless of things
13154 * like locale, folding, utf8ness of the target string */
13155 SV* cp_list = NULL;
13157 /* Like cp_list, but code points on this list need to be checked for things
13158 * that fold to/from them under /i */
13159 SV* cp_foldable_list = NULL;
13162 /* In a range, counts how many 0-2 of the ends of it came from literals,
13163 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13164 UV literal_endpoint = 0;
13166 bool invert = FALSE; /* Is this class to be complemented */
13168 bool warn_super = ALWAYS_WARN_SUPER;
13170 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13171 case we need to change the emitted regop to an EXACT. */
13172 const char * orig_parse = RExC_parse;
13173 const SSize_t orig_size = RExC_size;
13174 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13175 GET_RE_DEBUG_FLAGS_DECL;
13177 PERL_ARGS_ASSERT_REGCLASS;
13179 PERL_UNUSED_ARG(depth);
13182 DEBUG_PARSE("clas");
13184 /* Assume we are going to generate an ANYOF node. */
13185 ret = reganode(pRExC_state, ANYOF, 0);
13188 RExC_size += ANYOF_SKIP;
13189 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13192 ANYOF_FLAGS(ret) = 0;
13194 RExC_emit += ANYOF_SKIP;
13196 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
13198 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13199 initial_listsv_len = SvCUR(listsv);
13200 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13204 RExC_parse = regpatws(pRExC_state, RExC_parse,
13205 FALSE /* means don't recognize comments */);
13208 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13211 allow_multi_folds = FALSE;
13214 RExC_parse = regpatws(pRExC_state, RExC_parse,
13215 FALSE /* means don't recognize comments */);
13219 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13220 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13221 const char *s = RExC_parse;
13222 const char c = *s++;
13224 while (isWORDCHAR(*s))
13226 if (*s && c == *s && s[1] == ']') {
13227 SAVEFREESV(RExC_rx_sv);
13229 "POSIX syntax [%c %c] belongs inside character classes",
13231 (void)ReREFCNT_inc(RExC_rx_sv);
13235 /* If the caller wants us to just parse a single element, accomplish this
13236 * by faking the loop ending condition */
13237 if (stop_at_1 && RExC_end > RExC_parse) {
13238 stop_ptr = RExC_parse + 1;
13241 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13242 if (UCHARAT(RExC_parse) == ']')
13243 goto charclassloop;
13247 if (RExC_parse >= stop_ptr) {
13252 RExC_parse = regpatws(pRExC_state, RExC_parse,
13253 FALSE /* means don't recognize comments */);
13256 if (UCHARAT(RExC_parse) == ']') {
13262 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13263 save_value = value;
13264 save_prevvalue = prevvalue;
13267 rangebegin = RExC_parse;
13271 value = utf8n_to_uvchr((U8*)RExC_parse,
13272 RExC_end - RExC_parse,
13273 &numlen, UTF8_ALLOW_DEFAULT);
13274 RExC_parse += numlen;
13277 value = UCHARAT(RExC_parse++);
13280 && RExC_parse < RExC_end
13281 && POSIXCC(UCHARAT(RExC_parse)))
13283 namedclass = regpposixcc(pRExC_state, value, strict);
13285 else if (value == '\\') {
13287 value = utf8n_to_uvchr((U8*)RExC_parse,
13288 RExC_end - RExC_parse,
13289 &numlen, UTF8_ALLOW_DEFAULT);
13290 RExC_parse += numlen;
13293 value = UCHARAT(RExC_parse++);
13295 /* Some compilers cannot handle switching on 64-bit integer
13296 * values, therefore value cannot be an UV. Yes, this will
13297 * be a problem later if we want switch on Unicode.
13298 * A similar issue a little bit later when switching on
13299 * namedclass. --jhi */
13301 /* If the \ is escaping white space when white space is being
13302 * skipped, it means that that white space is wanted literally, and
13303 * is already in 'value'. Otherwise, need to translate the escape
13304 * into what it signifies. */
13305 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13307 case 'w': namedclass = ANYOF_WORDCHAR; break;
13308 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13309 case 's': namedclass = ANYOF_SPACE; break;
13310 case 'S': namedclass = ANYOF_NSPACE; break;
13311 case 'd': namedclass = ANYOF_DIGIT; break;
13312 case 'D': namedclass = ANYOF_NDIGIT; break;
13313 case 'v': namedclass = ANYOF_VERTWS; break;
13314 case 'V': namedclass = ANYOF_NVERTWS; break;
13315 case 'h': namedclass = ANYOF_HORIZWS; break;
13316 case 'H': namedclass = ANYOF_NHORIZWS; break;
13317 case 'N': /* Handle \N{NAME} in class */
13319 /* We only pay attention to the first char of
13320 multichar strings being returned. I kinda wonder
13321 if this makes sense as it does change the behaviour
13322 from earlier versions, OTOH that behaviour was broken
13324 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
13325 TRUE, /* => charclass */
13328 if (*flagp & RESTART_UTF8)
13329 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13339 /* We will handle any undefined properties ourselves */
13340 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
13341 /* And we actually would prefer to get
13342 * the straight inversion list of the
13343 * swash, since we will be accessing it
13344 * anyway, to save a little time */
13345 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
13347 if (RExC_parse >= RExC_end)
13348 vFAIL2("Empty \\%c{}", (U8)value);
13349 if (*RExC_parse == '{') {
13350 const U8 c = (U8)value;
13351 e = strchr(RExC_parse++, '}');
13353 vFAIL2("Missing right brace on \\%c{}", c);
13354 while (isSPACE(UCHARAT(RExC_parse)))
13356 if (e == RExC_parse)
13357 vFAIL2("Empty \\%c{}", c);
13358 n = e - RExC_parse;
13359 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
13371 if (UCHARAT(RExC_parse) == '^') {
13374 /* toggle. (The rhs xor gets the single bit that
13375 * differs between P and p; the other xor inverts just
13377 value ^= 'P' ^ 'p';
13379 while (isSPACE(UCHARAT(RExC_parse))) {
13384 /* Try to get the definition of the property into
13385 * <invlist>. If /i is in effect, the effective property
13386 * will have its name be <__NAME_i>. The design is
13387 * discussed in commit
13388 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
13389 formatted = Perl_form(aTHX_
13391 (FOLD) ? "__" : "",
13396 name = savepvn(formatted, strlen(formatted));
13398 /* Look up the property name, and get its swash and
13399 * inversion list, if the property is found */
13401 SvREFCNT_dec_NN(swash);
13403 swash = _core_swash_init("utf8", name, &PL_sv_undef,
13406 NULL, /* No inversion list */
13409 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
13411 SvREFCNT_dec_NN(swash);
13415 /* Here didn't find it. It could be a user-defined
13416 * property that will be available at run-time. If we
13417 * accept only compile-time properties, is an error;
13418 * otherwise add it to the list for run-time look up */
13420 RExC_parse = e + 1;
13422 "Property '%"UTF8f"' is unknown",
13423 UTF8fARG(UTF, n, name));
13425 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
13426 (value == 'p' ? '+' : '!'),
13427 UTF8fARG(UTF, n, name));
13428 has_user_defined_property = TRUE;
13430 /* We don't know yet, so have to assume that the
13431 * property could match something in the Latin1 range,
13432 * hence something that isn't utf8. Note that this
13433 * would cause things in <depends_list> to match
13434 * inappropriately, except that any \p{}, including
13435 * this one forces Unicode semantics, which means there
13436 * is no <depends_list> */
13437 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
13441 /* Here, did get the swash and its inversion list. If
13442 * the swash is from a user-defined property, then this
13443 * whole character class should be regarded as such */
13444 if (swash_init_flags
13445 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
13447 has_user_defined_property = TRUE;
13450 /* We warn on matching an above-Unicode code point
13451 * if the match would return true, except don't
13452 * warn for \p{All}, which has exactly one element
13454 (_invlist_contains_cp(invlist, 0x110000)
13455 && (! (_invlist_len(invlist) == 1
13456 && *invlist_array(invlist) == 0)))
13462 /* Invert if asking for the complement */
13463 if (value == 'P') {
13464 _invlist_union_complement_2nd(properties,
13468 /* The swash can't be used as-is, because we've
13469 * inverted things; delay removing it to here after
13470 * have copied its invlist above */
13471 SvREFCNT_dec_NN(swash);
13475 _invlist_union(properties, invlist, &properties);
13480 RExC_parse = e + 1;
13481 namedclass = ANYOF_UNIPROP; /* no official name, but it's
13484 /* \p means they want Unicode semantics */
13485 RExC_uni_semantics = 1;
13488 case 'n': value = '\n'; break;
13489 case 'r': value = '\r'; break;
13490 case 't': value = '\t'; break;
13491 case 'f': value = '\f'; break;
13492 case 'b': value = '\b'; break;
13493 case 'e': value = ASCII_TO_NATIVE('\033');break;
13494 case 'a': value = '\a'; break;
13496 RExC_parse--; /* function expects to be pointed at the 'o' */
13498 const char* error_msg;
13499 bool valid = grok_bslash_o(&RExC_parse,
13502 SIZE_ONLY, /* warnings in pass
13505 silence_non_portable,
13511 if (PL_encoding && value < 0x100) {
13512 goto recode_encoding;
13516 RExC_parse--; /* function expects to be pointed at the 'x' */
13518 const char* error_msg;
13519 bool valid = grok_bslash_x(&RExC_parse,
13522 TRUE, /* Output warnings */
13524 silence_non_portable,
13530 if (PL_encoding && value < 0x100)
13531 goto recode_encoding;
13534 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
13536 case '0': case '1': case '2': case '3': case '4':
13537 case '5': case '6': case '7':
13539 /* Take 1-3 octal digits */
13540 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13541 numlen = (strict) ? 4 : 3;
13542 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
13543 RExC_parse += numlen;
13546 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13547 vFAIL("Need exactly 3 octal digits");
13549 else if (! SIZE_ONLY /* like \08, \178 */
13551 && RExC_parse < RExC_end
13552 && isDIGIT(*RExC_parse)
13553 && ckWARN(WARN_REGEXP))
13555 SAVEFREESV(RExC_rx_sv);
13556 reg_warn_non_literal_string(
13558 form_short_octal_warning(RExC_parse, numlen));
13559 (void)ReREFCNT_inc(RExC_rx_sv);
13562 if (PL_encoding && value < 0x100)
13563 goto recode_encoding;
13567 if (! RExC_override_recoding) {
13568 SV* enc = PL_encoding;
13569 value = reg_recode((const char)(U8)value, &enc);
13572 vFAIL("Invalid escape in the specified encoding");
13574 else if (SIZE_ONLY) {
13575 ckWARNreg(RExC_parse,
13576 "Invalid escape in the specified encoding");
13582 /* Allow \_ to not give an error */
13583 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
13585 vFAIL2("Unrecognized escape \\%c in character class",
13589 SAVEFREESV(RExC_rx_sv);
13590 ckWARN2reg(RExC_parse,
13591 "Unrecognized escape \\%c in character class passed through",
13593 (void)ReREFCNT_inc(RExC_rx_sv);
13597 } /* End of switch on char following backslash */
13598 } /* end of handling backslash escape sequences */
13601 literal_endpoint++;
13604 /* Here, we have the current token in 'value' */
13606 /* What matches in a locale is not known until runtime. This includes
13607 * what the Posix classes (like \w, [:space:]) match. Room must be
13608 * reserved (one time per outer bracketed class) to store such classes,
13609 * either if Perl is compiled so that locale nodes always should have
13610 * this space, or if there is such posix class info to be stored. The
13611 * space will contain a bit for each named class that is to be matched
13612 * against. This isn't needed for \p{} and pseudo-classes, as they are
13613 * not affected by locale, and hence are dealt with separately */
13615 if (FOLD && ! need_class) {
13618 RExC_size += ANYOF_POSIXL_FOLD_SKIP - ANYOF_SKIP;
13621 RExC_emit += ANYOF_POSIXL_FOLD_SKIP - ANYOF_SKIP;
13624 /* We need to initialize this here because this node type has
13625 * this field, and will skip getting initialized when we get to
13626 * a posix class since are doing it here */
13627 ANYOF_POSIXL_ZERO(ret);
13629 if (ANYOF_LOCALE == ANYOF_POSIXL
13630 || (namedclass > OOB_NAMEDCLASS
13631 && namedclass < ANYOF_POSIXL_MAX))
13633 if (! need_class) {
13636 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13639 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13641 ANYOF_POSIXL_ZERO(ret);
13643 ANYOF_FLAGS(ret) |= ANYOF_POSIXL;
13647 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
13650 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
13651 * literal, as is the character that began the false range, i.e.
13652 * the 'a' in the examples */
13655 const int w = (RExC_parse >= rangebegin)
13656 ? RExC_parse - rangebegin
13660 "False [] range \"%"UTF8f"\"",
13661 UTF8fARG(UTF, w, rangebegin));
13664 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
13665 ckWARN2reg(RExC_parse,
13666 "False [] range \"%"UTF8f"\"",
13667 UTF8fARG(UTF, w, rangebegin));
13668 (void)ReREFCNT_inc(RExC_rx_sv);
13669 cp_list = add_cp_to_invlist(cp_list, '-');
13670 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
13675 range = 0; /* this was not a true range */
13676 element_count += 2; /* So counts for three values */
13679 classnum = namedclass_to_classnum(namedclass);
13681 if (LOC && namedclass < ANYOF_POSIXL_MAX
13682 #ifndef HAS_ISASCII
13683 && classnum != _CC_ASCII
13687 /* See if it already matches the complement of this POSIX
13689 if ((ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13690 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
13694 posixl_matches_all = TRUE;
13695 break; /* No need to continue. Since it matches both
13696 e.g., \w and \W, it matches everything, and the
13697 bracketed class can be optimized into qr/./s */
13700 /* Add this class to those that should be checked at runtime */
13701 ANYOF_POSIXL_SET(ret, namedclass);
13703 /* The above-Latin1 characters are not subject to locale rules.
13704 * Just add them, in the second pass, to the
13705 * unconditionally-matched list */
13707 SV* scratch_list = NULL;
13709 /* Get the list of the above-Latin1 code points this
13711 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
13712 PL_XPosix_ptrs[classnum],
13714 /* Odd numbers are complements, like
13715 * NDIGIT, NASCII, ... */
13716 namedclass % 2 != 0,
13718 /* Checking if 'cp_list' is NULL first saves an extra
13719 * clone. Its reference count will be decremented at the
13720 * next union, etc, or if this is the only instance, at the
13721 * end of the routine */
13723 cp_list = scratch_list;
13726 _invlist_union(cp_list, scratch_list, &cp_list);
13727 SvREFCNT_dec_NN(scratch_list);
13729 continue; /* Go get next character */
13732 else if (! SIZE_ONLY) {
13734 /* Here, not in pass1 (in that pass we skip calculating the
13735 * contents of this class), and is /l, or is a POSIX class for
13736 * which /l doesn't matter (or is a Unicode property, which is
13737 * skipped here). */
13738 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
13739 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
13741 /* Here, should be \h, \H, \v, or \V. None of /d, /i
13742 * nor /l make a difference in what these match,
13743 * therefore we just add what they match to cp_list. */
13744 if (classnum != _CC_VERTSPACE) {
13745 assert( namedclass == ANYOF_HORIZWS
13746 || namedclass == ANYOF_NHORIZWS);
13748 /* It turns out that \h is just a synonym for
13750 classnum = _CC_BLANK;
13753 _invlist_union_maybe_complement_2nd(
13755 PL_XPosix_ptrs[classnum],
13756 namedclass % 2 != 0, /* Complement if odd
13757 (NHORIZWS, NVERTWS)
13762 else { /* Garden variety class. If is NASCII, NDIGIT, ...
13763 complement and use nposixes */
13764 SV** posixes_ptr = namedclass % 2 == 0
13767 SV** source_ptr = &PL_XPosix_ptrs[classnum];
13768 _invlist_union_maybe_complement_2nd(
13771 namedclass % 2 != 0,
13774 continue; /* Go get next character */
13776 } /* end of namedclass \blah */
13778 /* Here, we have a single value. If 'range' is set, it is the ending
13779 * of a range--check its validity. Later, we will handle each
13780 * individual code point in the range. If 'range' isn't set, this
13781 * could be the beginning of a range, so check for that by looking
13782 * ahead to see if the next real character to be processed is the range
13783 * indicator--the minus sign */
13786 RExC_parse = regpatws(pRExC_state, RExC_parse,
13787 FALSE /* means don't recognize comments */);
13791 if (prevvalue > value) /* b-a */ {
13792 const int w = RExC_parse - rangebegin;
13794 "Invalid [] range \"%"UTF8f"\"",
13795 UTF8fARG(UTF, w, rangebegin));
13796 range = 0; /* not a valid range */
13800 prevvalue = value; /* save the beginning of the potential range */
13801 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13802 && *RExC_parse == '-')
13804 char* next_char_ptr = RExC_parse + 1;
13805 if (skip_white) { /* Get the next real char after the '-' */
13806 next_char_ptr = regpatws(pRExC_state,
13808 FALSE); /* means don't recognize
13812 /* If the '-' is at the end of the class (just before the ']',
13813 * it is a literal minus; otherwise it is a range */
13814 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13815 RExC_parse = next_char_ptr;
13817 /* a bad range like \w-, [:word:]- ? */
13818 if (namedclass > OOB_NAMEDCLASS) {
13819 if (strict || ckWARN(WARN_REGEXP)) {
13821 RExC_parse >= rangebegin ?
13822 RExC_parse - rangebegin : 0;
13824 vFAIL4("False [] range \"%*.*s\"",
13829 "False [] range \"%*.*s\"",
13834 cp_list = add_cp_to_invlist(cp_list, '-');
13838 range = 1; /* yeah, it's a range! */
13839 continue; /* but do it the next time */
13844 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13847 /* non-Latin1 code point implies unicode semantics. Must be set in
13848 * pass1 so is there for the whole of pass 2 */
13850 RExC_uni_semantics = 1;
13853 /* Ready to process either the single value, or the completed range.
13854 * For single-valued non-inverted ranges, we consider the possibility
13855 * of multi-char folds. (We made a conscious decision to not do this
13856 * for the other cases because it can often lead to non-intuitive
13857 * results. For example, you have the peculiar case that:
13858 * "s s" =~ /^[^\xDF]+$/i => Y
13859 * "ss" =~ /^[^\xDF]+$/i => N
13861 * See [perl #89750] */
13862 if (FOLD && allow_multi_folds && value == prevvalue) {
13863 if (value == LATIN_SMALL_LETTER_SHARP_S
13864 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13867 /* Here <value> is indeed a multi-char fold. Get what it is */
13869 U8 foldbuf[UTF8_MAXBYTES_CASE];
13872 UV folded = _to_uni_fold_flags(
13876 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
13877 ? FOLD_FLAGS_NOMIX_ASCII
13881 /* Here, <folded> should be the first character of the
13882 * multi-char fold of <value>, with <foldbuf> containing the
13883 * whole thing. But, if this fold is not allowed (because of
13884 * the flags), <fold> will be the same as <value>, and should
13885 * be processed like any other character, so skip the special
13887 if (folded != value) {
13889 /* Skip if we are recursed, currently parsing the class
13890 * again. Otherwise add this character to the list of
13891 * multi-char folds. */
13892 if (! RExC_in_multi_char_class) {
13893 AV** this_array_ptr;
13895 STRLEN cp_count = utf8_length(foldbuf,
13896 foldbuf + foldlen);
13897 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13899 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13902 if (! multi_char_matches) {
13903 multi_char_matches = newAV();
13906 /* <multi_char_matches> is actually an array of arrays.
13907 * There will be one or two top-level elements: [2],
13908 * and/or [3]. The [2] element is an array, each
13909 * element thereof is a character which folds to TWO
13910 * characters; [3] is for folds to THREE characters.
13911 * (Unicode guarantees a maximum of 3 characters in any
13912 * fold.) When we rewrite the character class below,
13913 * we will do so such that the longest folds are
13914 * written first, so that it prefers the longest
13915 * matching strings first. This is done even if it
13916 * turns out that any quantifier is non-greedy, out of
13917 * programmer laziness. Tom Christiansen has agreed
13918 * that this is ok. This makes the test for the
13919 * ligature 'ffi' come before the test for 'ff' */
13920 if (av_exists(multi_char_matches, cp_count)) {
13921 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13923 this_array = *this_array_ptr;
13926 this_array = newAV();
13927 av_store(multi_char_matches, cp_count,
13930 av_push(this_array, multi_fold);
13933 /* This element should not be processed further in this
13936 value = save_value;
13937 prevvalue = save_prevvalue;
13943 /* Deal with this element of the class */
13946 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
13949 SV* this_range = _new_invlist(1);
13950 _append_range_to_invlist(this_range, prevvalue, value);
13952 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13953 * If this range was specified using something like 'i-j', we want
13954 * to include only the 'i' and the 'j', and not anything in
13955 * between, so exclude non-ASCII, non-alphabetics from it.
13956 * However, if the range was specified with something like
13957 * [\x89-\x91] or [\x89-j], all code points within it should be
13958 * included. literal_endpoint==2 means both ends of the range used
13959 * a literal character, not \x{foo} */
13960 if (literal_endpoint == 2
13961 && ((prevvalue >= 'a' && value <= 'z')
13962 || (prevvalue >= 'A' && value <= 'Z')))
13964 _invlist_intersection(this_range, PL_ASCII,
13967 /* Since this above only contains ascii, the intersection of it
13968 * with anything will still yield only ascii */
13969 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ALPHA],
13972 _invlist_union(cp_foldable_list, this_range, &cp_foldable_list);
13973 literal_endpoint = 0;
13977 range = 0; /* this range (if it was one) is done now */
13978 } /* End of loop through all the text within the brackets */
13980 /* If anything in the class expands to more than one character, we have to
13981 * deal with them by building up a substitute parse string, and recursively
13982 * calling reg() on it, instead of proceeding */
13983 if (multi_char_matches) {
13984 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13987 char *save_end = RExC_end;
13988 char *save_parse = RExC_parse;
13989 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13994 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13995 because too confusing */
13997 sv_catpv(substitute_parse, "(?:");
14001 /* Look at the longest folds first */
14002 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
14004 if (av_exists(multi_char_matches, cp_count)) {
14005 AV** this_array_ptr;
14008 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14010 while ((this_sequence = av_pop(*this_array_ptr)) !=
14013 if (! first_time) {
14014 sv_catpv(substitute_parse, "|");
14016 first_time = FALSE;
14018 sv_catpv(substitute_parse, SvPVX(this_sequence));
14023 /* If the character class contains anything else besides these
14024 * multi-character folds, have to include it in recursive parsing */
14025 if (element_count) {
14026 sv_catpv(substitute_parse, "|[");
14027 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14028 sv_catpv(substitute_parse, "]");
14031 sv_catpv(substitute_parse, ")");
14034 /* This is a way to get the parse to skip forward a whole named
14035 * sequence instead of matching the 2nd character when it fails the
14037 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14041 RExC_parse = SvPV(substitute_parse, len);
14042 RExC_end = RExC_parse + len;
14043 RExC_in_multi_char_class = 1;
14044 RExC_emit = (regnode *)orig_emit;
14046 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14048 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14050 RExC_parse = save_parse;
14051 RExC_end = save_end;
14052 RExC_in_multi_char_class = 0;
14053 SvREFCNT_dec_NN(multi_char_matches);
14057 /* Here, we've gone through the entire class and dealt with multi-char
14058 * folds. We are now in a position that we can do some checks to see if we
14059 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14060 * Currently we only do two checks:
14061 * 1) is in the unlikely event that the user has specified both, eg. \w and
14062 * \W under /l, then the class matches everything. (This optimization
14063 * is done only to make the optimizer code run later work.)
14064 * 2) if the character class contains only a single element (including a
14065 * single range), we see if there is an equivalent node for it.
14066 * Other checks are possible */
14067 if (! ret_invlist /* Can't optimize if returning the constructed
14069 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14074 if (UNLIKELY(posixl_matches_all)) {
14077 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14078 \w or [:digit:] or \p{foo}
14081 /* All named classes are mapped into POSIXish nodes, with its FLAG
14082 * argument giving which class it is */
14083 switch ((I32)namedclass) {
14084 case ANYOF_UNIPROP:
14087 /* These don't depend on the charset modifiers. They always
14088 * match under /u rules */
14089 case ANYOF_NHORIZWS:
14090 case ANYOF_HORIZWS:
14091 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14094 case ANYOF_NVERTWS:
14099 /* The actual POSIXish node for all the rest depends on the
14100 * charset modifier. The ones in the first set depend only on
14101 * ASCII or, if available on this platform, locale */
14105 op = (LOC) ? POSIXL : POSIXA;
14116 /* under /a could be alpha */
14118 if (ASCII_RESTRICTED) {
14119 namedclass = ANYOF_ALPHA + (namedclass % 2);
14127 /* The rest have more possibilities depending on the charset.
14128 * We take advantage of the enum ordering of the charset
14129 * modifiers to get the exact node type, */
14131 op = POSIXD + get_regex_charset(RExC_flags);
14132 if (op > POSIXA) { /* /aa is same as /a */
14137 /* The odd numbered ones are the complements of the
14138 * next-lower even number one */
14139 if (namedclass % 2 == 1) {
14143 arg = namedclass_to_classnum(namedclass);
14147 else if (value == prevvalue) {
14149 /* Here, the class consists of just a single code point */
14152 if (! LOC && value == '\n') {
14153 op = REG_ANY; /* Optimize [^\n] */
14154 *flagp |= HASWIDTH|SIMPLE;
14158 else if (value < 256 || UTF) {
14160 /* Optimize a single value into an EXACTish node, but not if it
14161 * would require converting the pattern to UTF-8. */
14162 op = compute_EXACTish(pRExC_state);
14164 } /* Otherwise is a range */
14165 else if (! LOC) { /* locale could vary these */
14166 if (prevvalue == '0') {
14167 if (value == '9') {
14174 /* Here, we have changed <op> away from its initial value iff we found
14175 * an optimization */
14178 /* Throw away this ANYOF regnode, and emit the calculated one,
14179 * which should correspond to the beginning, not current, state of
14181 const char * cur_parse = RExC_parse;
14182 RExC_parse = (char *)orig_parse;
14186 /* To get locale nodes to not use the full ANYOF size would
14187 * require moving the code above that writes the portions
14188 * of it that aren't in other nodes to after this point.
14189 * e.g. ANYOF_POSIXL_SET */
14190 RExC_size = orig_size;
14194 RExC_emit = (regnode *)orig_emit;
14195 if (PL_regkind[op] == POSIXD) {
14197 op += NPOSIXD - POSIXD;
14202 ret = reg_node(pRExC_state, op);
14204 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14208 *flagp |= HASWIDTH|SIMPLE;
14210 else if (PL_regkind[op] == EXACT) {
14211 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
14214 RExC_parse = (char *) cur_parse;
14216 SvREFCNT_dec(posixes);
14217 SvREFCNT_dec(nposixes);
14218 SvREFCNT_dec(cp_list);
14219 SvREFCNT_dec(cp_foldable_list);
14226 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14228 /* If folding, we calculate all characters that could fold to or from the
14229 * ones already on the list */
14230 if (cp_foldable_list) {
14232 UV start, end; /* End points of code point ranges */
14234 SV* fold_intersection = NULL;
14237 /* Our calculated list will be for Unicode rules. For locale
14238 * matching, we have to keep a separate list that is consulted at
14239 * runtime only when the locale indicates Unicode rules. For
14240 * non-locale, we just use to the general list */
14242 use_list = &ANYOF_UTF8_LOCALE_INVLIST(ret);
14246 use_list = &cp_list;
14249 /* Only the characters in this class that participate in folds need
14250 * be checked. Get the intersection of this class and all the
14251 * possible characters that are foldable. This can quickly narrow
14252 * down a large class */
14253 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14254 &fold_intersection);
14256 /* The folds for all the Latin1 characters are hard-coded into this
14257 * program, but we have to go out to disk to get the others. */
14258 if (invlist_highest(cp_foldable_list) >= 256) {
14260 /* This is a hash that for a particular fold gives all
14261 * characters that are involved in it */
14262 if (! PL_utf8_foldclosures) {
14264 /* If the folds haven't been read in, call a fold function
14266 if (! PL_utf8_tofold) {
14267 U8 dummy[UTF8_MAXBYTES_CASE+1];
14269 /* This string is just a short named one above \xff */
14270 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
14271 assert(PL_utf8_tofold); /* Verify that worked */
14273 PL_utf8_foldclosures
14274 = _swash_inversion_hash(PL_utf8_tofold);
14278 /* Now look at the foldable characters in this class individually */
14279 invlist_iterinit(fold_intersection);
14280 while (invlist_iternext(fold_intersection, &start, &end)) {
14283 /* Look at every character in the range */
14284 for (j = start; j <= end; j++) {
14285 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14291 /* We have the latin1 folding rules hard-coded here so
14292 * that an innocent-looking character class, like
14293 * /[ks]/i won't have to go out to disk to find the
14294 * possible matches. XXX It would be better to
14295 * generate these via regen, in case a new version of
14296 * the Unicode standard adds new mappings, though that
14297 * is not really likely, and may be caught by the
14298 * default: case of the switch below. */
14300 if (IS_IN_SOME_FOLD_L1(j)) {
14302 /* ASCII is always matched; non-ASCII is matched
14303 * only under Unicode rules (which could happen
14304 * under /l if the locale is a UTF-8 one */
14305 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
14306 *use_list = add_cp_to_invlist(*use_list,
14307 PL_fold_latin1[j]);
14311 add_cp_to_invlist(depends_list,
14312 PL_fold_latin1[j]);
14316 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
14317 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
14319 /* Certain Latin1 characters have matches outside
14320 * Latin1. To get here, <j> is one of those
14321 * characters. None of these matches is valid for
14322 * ASCII characters under /aa, which is why the 'if'
14323 * just above excludes those. These matches only
14324 * happen when the target string is utf8. The code
14325 * below adds the single fold closures for <j> to the
14326 * inversion list. */
14332 add_cp_to_invlist(*use_list, KELVIN_SIGN);
14336 *use_list = add_cp_to_invlist(*use_list,
14337 LATIN_SMALL_LETTER_LONG_S);
14340 *use_list = add_cp_to_invlist(*use_list,
14341 GREEK_CAPITAL_LETTER_MU);
14342 *use_list = add_cp_to_invlist(*use_list,
14343 GREEK_SMALL_LETTER_MU);
14345 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14346 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14348 add_cp_to_invlist(*use_list, ANGSTROM_SIGN);
14350 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14351 *use_list = add_cp_to_invlist(*use_list,
14352 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14354 case LATIN_SMALL_LETTER_SHARP_S:
14355 *use_list = add_cp_to_invlist(*use_list,
14356 LATIN_CAPITAL_LETTER_SHARP_S);
14358 case 'F': case 'f':
14359 case 'I': case 'i':
14360 case 'L': case 'l':
14361 case 'T': case 't':
14362 case 'A': case 'a':
14363 case 'H': case 'h':
14364 case 'J': case 'j':
14365 case 'N': case 'n':
14366 case 'W': case 'w':
14367 case 'Y': case 'y':
14368 /* These all are targets of multi-character
14369 * folds from code points that require UTF8
14370 * to express, so they can't match unless
14371 * the target string is in UTF-8, so no
14372 * action here is necessary, as regexec.c
14373 * properly handles the general case for
14374 * UTF-8 matching and multi-char folds */
14377 /* Use deprecated warning to increase the
14378 * chances of this being output */
14379 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
14386 /* Here is an above Latin1 character. We don't have the
14387 * rules hard-coded for it. First, get its fold. This is
14388 * the simple fold, as the multi-character folds have been
14389 * handled earlier and separated out */
14390 _to_uni_fold_flags(j, foldbuf, &foldlen,
14391 (ASCII_FOLD_RESTRICTED)
14392 ? FOLD_FLAGS_NOMIX_ASCII
14395 /* Single character fold of above Latin1. Add everything in
14396 * its fold closure to the list that this node should match.
14397 * The fold closures data structure is a hash with the keys
14398 * being the UTF-8 of every character that is folded to, like
14399 * 'k', and the values each an array of all code points that
14400 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
14401 * Multi-character folds are not included */
14402 if ((listp = hv_fetch(PL_utf8_foldclosures,
14403 (char *) foldbuf, foldlen, FALSE)))
14405 AV* list = (AV*) *listp;
14407 for (k = 0; k <= av_len(list); k++) {
14408 SV** c_p = av_fetch(list, k, FALSE);
14411 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
14415 /* /aa doesn't allow folds between ASCII and non- */
14416 if ((ASCII_FOLD_RESTRICTED
14417 && (isASCII(c) != isASCII(j))))
14422 /* Folds under /l which cross the 255/256 boundary
14423 * are added to a separate list. (These are valid
14424 * only when the locale is UTF-8.) */
14425 if (c < 256 && LOC) {
14426 *use_list = add_cp_to_invlist(*use_list, c);
14430 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
14432 cp_list = add_cp_to_invlist(cp_list, c);
14435 /* Similarly folds involving non-ascii Latin1
14436 * characters under /d are added to their list */
14437 depends_list = add_cp_to_invlist(depends_list,
14444 SvREFCNT_dec_NN(fold_intersection);
14447 /* Now that we have finished adding all the folds, there is no reason
14448 * to keep the foldable list separate */
14449 _invlist_union(cp_list, cp_foldable_list, &cp_list);
14450 SvREFCNT_dec_NN(cp_foldable_list);
14453 /* And combine the result (if any) with any inversion list from posix
14454 * classes. The lists are kept separate up to now because we don't want to
14455 * fold the classes (folding of those is automatically handled by the swash
14456 * fetching code) */
14457 if (posixes || nposixes) {
14458 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
14459 /* Under /a and /aa, nothing above ASCII matches these */
14460 _invlist_intersection(posixes,
14461 PL_XPosix_ptrs[_CC_ASCII],
14465 if (DEPENDS_SEMANTICS) {
14466 /* Under /d, everything in the upper half of the Latin1 range
14467 * matches these complements */
14468 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_NON_ASCII_ALL;
14470 else if (AT_LEAST_ASCII_RESTRICTED) {
14471 /* Under /a and /aa, everything above ASCII matches these
14473 _invlist_union_complement_2nd(nposixes,
14474 PL_XPosix_ptrs[_CC_ASCII],
14478 _invlist_union(posixes, nposixes, &posixes);
14479 SvREFCNT_dec_NN(nposixes);
14482 posixes = nposixes;
14485 if (! DEPENDS_SEMANTICS) {
14487 _invlist_union(cp_list, posixes, &cp_list);
14488 SvREFCNT_dec_NN(posixes);
14495 /* Under /d, we put into a separate list the Latin1 things that
14496 * match only when the target string is utf8 */
14497 SV* nonascii_but_latin1_properties = NULL;
14498 _invlist_intersection(posixes, PL_UpperLatin1,
14499 &nonascii_but_latin1_properties);
14500 _invlist_subtract(posixes, nonascii_but_latin1_properties,
14503 _invlist_union(cp_list, posixes, &cp_list);
14504 SvREFCNT_dec_NN(posixes);
14510 if (depends_list) {
14511 _invlist_union(depends_list, nonascii_but_latin1_properties,
14513 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
14516 depends_list = nonascii_but_latin1_properties;
14521 /* And combine the result (if any) with any inversion list from properties.
14522 * The lists are kept separate up to now so that we can distinguish the two
14523 * in regards to matching above-Unicode. A run-time warning is generated
14524 * if a Unicode property is matched against a non-Unicode code point. But,
14525 * we allow user-defined properties to match anything, without any warning,
14526 * and we also suppress the warning if there is a portion of the character
14527 * class that isn't a Unicode property, and which matches above Unicode, \W
14528 * or [\x{110000}] for example.
14529 * (Note that in this case, unlike the Posix one above, there is no
14530 * <depends_list>, because having a Unicode property forces Unicode
14535 /* If it matters to the final outcome, see if a non-property
14536 * component of the class matches above Unicode. If so, the
14537 * warning gets suppressed. This is true even if just a single
14538 * such code point is specified, as though not strictly correct if
14539 * another such code point is matched against, the fact that they
14540 * are using above-Unicode code points indicates they should know
14541 * the issues involved */
14543 warn_super = ! (invert
14544 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
14547 _invlist_union(properties, cp_list, &cp_list);
14548 SvREFCNT_dec_NN(properties);
14551 cp_list = properties;
14555 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
14559 /* Here, we have calculated what code points should be in the character
14562 * Now we can see about various optimizations. Fold calculation (which we
14563 * did above) needs to take place before inversion. Otherwise /[^k]/i
14564 * would invert to include K, which under /i would match k, which it
14565 * shouldn't. Therefore we can't invert folded locale now, as it won't be
14566 * folded until runtime */
14568 /* If we didn't do folding, it's because some information isn't available
14569 * until runtime; set the run-time fold flag for these. (We don't have to
14570 * worry about properties folding, as that is taken care of by the swash
14571 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
14572 * locales, or the class matches at least one 0-255 range code point */
14574 if (ANYOF_UTF8_LOCALE_INVLIST(ret)) {
14575 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14577 else if (cp_list) { /* Look to see if there a 0-255 code point is in
14580 invlist_iterinit(cp_list);
14581 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
14582 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14584 invlist_iterfinish(cp_list);
14588 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
14589 * at compile time. Besides not inverting folded locale now, we can't
14590 * invert if there are things such as \w, which aren't known until runtime
14593 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOC_FOLD|ANYOF_POSIXL))
14595 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14597 _invlist_invert(cp_list);
14599 /* Any swash can't be used as-is, because we've inverted things */
14601 SvREFCNT_dec_NN(swash);
14605 /* Clear the invert flag since have just done it here */
14610 *ret_invlist = cp_list;
14611 SvREFCNT_dec(swash);
14613 /* Discard the generated node */
14615 RExC_size = orig_size;
14618 RExC_emit = orig_emit;
14623 /* Some character classes are equivalent to other nodes. Such nodes take
14624 * up less room and generally fewer operations to execute than ANYOF nodes.
14625 * Above, we checked for and optimized into some such equivalents for
14626 * certain common classes that are easy to test. Getting to this point in
14627 * the code means that the class didn't get optimized there. Since this
14628 * code is only executed in Pass 2, it is too late to save space--it has
14629 * been allocated in Pass 1, and currently isn't given back. But turning
14630 * things into an EXACTish node can allow the optimizer to join it to any
14631 * adjacent such nodes. And if the class is equivalent to things like /./,
14632 * expensive run-time swashes can be avoided. Now that we have more
14633 * complete information, we can find things necessarily missed by the
14634 * earlier code. I (khw) am not sure how much to look for here. It would
14635 * be easy, but perhaps too slow, to check any candidates against all the
14636 * node types they could possibly match using _invlistEQ(). */
14641 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOC_FOLD|ANYOF_POSIXL))
14642 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14644 /* We don't optimize if we are supposed to make sure all non-Unicode
14645 * code points raise a warning, as only ANYOF nodes have this check.
14647 && ! ((ANYOF_FLAGS(ret) | ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
14650 U8 op = END; /* The optimzation node-type */
14651 const char * cur_parse= RExC_parse;
14653 invlist_iterinit(cp_list);
14654 if (! invlist_iternext(cp_list, &start, &end)) {
14656 /* Here, the list is empty. This happens, for example, when a
14657 * Unicode property is the only thing in the character class, and
14658 * it doesn't match anything. (perluniprops.pod notes such
14661 *flagp |= HASWIDTH|SIMPLE;
14663 else if (start == end) { /* The range is a single code point */
14664 if (! invlist_iternext(cp_list, &start, &end)
14666 /* Don't do this optimization if it would require changing
14667 * the pattern to UTF-8 */
14668 && (start < 256 || UTF))
14670 /* Here, the list contains a single code point. Can optimize
14671 * into an EXACTish node */
14680 /* A locale node under folding with one code point can be
14681 * an EXACTFL, as its fold won't be calculated until
14687 /* Here, we are generally folding, but there is only one
14688 * code point to match. If we have to, we use an EXACT
14689 * node, but it would be better for joining with adjacent
14690 * nodes in the optimization pass if we used the same
14691 * EXACTFish node that any such are likely to be. We can
14692 * do this iff the code point doesn't participate in any
14693 * folds. For example, an EXACTF of a colon is the same as
14694 * an EXACT one, since nothing folds to or from a colon. */
14696 if (IS_IN_SOME_FOLD_L1(value)) {
14701 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
14706 /* If we haven't found the node type, above, it means we
14707 * can use the prevailing one */
14709 op = compute_EXACTish(pRExC_state);
14714 else if (start == 0) {
14715 if (end == UV_MAX) {
14717 *flagp |= HASWIDTH|SIMPLE;
14720 else if (end == '\n' - 1
14721 && invlist_iternext(cp_list, &start, &end)
14722 && start == '\n' + 1 && end == UV_MAX)
14725 *flagp |= HASWIDTH|SIMPLE;
14729 invlist_iterfinish(cp_list);
14732 RExC_parse = (char *)orig_parse;
14733 RExC_emit = (regnode *)orig_emit;
14735 ret = reg_node(pRExC_state, op);
14737 RExC_parse = (char *)cur_parse;
14739 if (PL_regkind[op] == EXACT) {
14740 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
14743 SvREFCNT_dec_NN(cp_list);
14748 /* Here, <cp_list> contains all the code points we can determine at
14749 * compile time that match under all conditions. Go through it, and
14750 * for things that belong in the bitmap, put them there, and delete from
14751 * <cp_list>. While we are at it, see if everything above 255 is in the
14752 * list, and if so, set a flag to speed up execution */
14754 populate_ANYOF_from_invlist(ret, &cp_list);
14757 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
14760 /* Here, the bitmap has been populated with all the Latin1 code points that
14761 * always match. Can now add to the overall list those that match only
14762 * when the target string is UTF-8 (<depends_list>). */
14763 if (depends_list) {
14765 _invlist_union(cp_list, depends_list, &cp_list);
14766 SvREFCNT_dec_NN(depends_list);
14769 cp_list = depends_list;
14773 /* If there is a swash and more than one element, we can't use the swash in
14774 * the optimization below. */
14775 if (swash && element_count > 1) {
14776 SvREFCNT_dec_NN(swash);
14780 set_ANYOF_arg(pRExC_state, ret, cp_list,
14781 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14783 swash, has_user_defined_property);
14785 *flagp |= HASWIDTH|SIMPLE;
14789 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14792 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
14793 regnode* const node,
14795 SV* const runtime_defns,
14797 const bool has_user_defined_property)
14799 /* Sets the arg field of an ANYOF-type node 'node', using information about
14800 * the node passed-in. If there is nothing outside the node's bitmap, the
14801 * arg is set to ANYOF_NONBITMAP_EMPTY. Otherwise, it sets the argument to
14802 * the count returned by add_data(), having allocated and stored an array,
14803 * av, that that count references, as follows:
14804 * av[0] stores the character class description in its textual form.
14805 * This is used later (regexec.c:Perl_regclass_swash()) to
14806 * initialize the appropriate swash, and is also useful for dumping
14807 * the regnode. This is set to &PL_sv_undef if the textual
14808 * description is not needed at run-time (as happens if the other
14809 * elements completely define the class)
14810 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
14811 * computed from av[0]. But if no further computation need be done,
14812 * the swash is stored here now (and av[0] is &PL_sv_undef).
14813 * av[2] stores the cp_list inversion list for use in addition or instead
14814 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
14815 * (Otherwise everything needed is already in av[0] and av[1])
14816 * av[3] is set if any component of the class is from a user-defined
14817 * property; used only if av[2] exists */
14821 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
14823 if (! cp_list && ! runtime_defns) {
14824 ARG_SET(node, ANYOF_NONBITMAP_EMPTY);
14827 AV * const av = newAV();
14830 av_store(av, 0, (runtime_defns)
14831 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
14833 av_store(av, 1, swash);
14834 SvREFCNT_dec_NN(cp_list);
14837 av_store(av, 1, &PL_sv_undef);
14839 av_store(av, 2, cp_list);
14840 av_store(av, 3, newSVuv(has_user_defined_property));
14844 rv = newRV_noinc(MUTABLE_SV(av));
14845 n = add_data(pRExC_state, STR_WITH_LEN("s"));
14846 RExC_rxi->data->data[n] = (void*)rv;
14852 /* reg_skipcomment()
14854 Absorbs an /x style # comments from the input stream.
14855 Returns true if there is more text remaining in the stream.
14856 Will set the REG_RUN_ON_COMMENT_SEEN flag if the comment
14857 terminates the pattern without including a newline.
14859 Note its the callers responsibility to ensure that we are
14860 actually in /x mode
14865 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14869 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14871 while (RExC_parse < RExC_end)
14872 if (*RExC_parse++ == '\n') {
14877 /* we ran off the end of the pattern without ending
14878 the comment, so we have to add an \n when wrapping */
14879 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
14887 Advances the parse position, and optionally absorbs
14888 "whitespace" from the inputstream.
14890 Without /x "whitespace" means (?#...) style comments only,
14891 with /x this means (?#...) and # comments and whitespace proper.
14893 Returns the RExC_parse point from BEFORE the scan occurs.
14895 This is the /x friendly way of saying RExC_parse++.
14899 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14901 char* const retval = RExC_parse++;
14903 PERL_ARGS_ASSERT_NEXTCHAR;
14906 if (RExC_end - RExC_parse >= 3
14907 && *RExC_parse == '('
14908 && RExC_parse[1] == '?'
14909 && RExC_parse[2] == '#')
14911 while (*RExC_parse != ')') {
14912 if (RExC_parse == RExC_end)
14913 FAIL("Sequence (?#... not terminated");
14919 if (RExC_flags & RXf_PMf_EXTENDED) {
14920 if (isSPACE(*RExC_parse)) {
14924 else if (*RExC_parse == '#') {
14925 if ( reg_skipcomment( pRExC_state ) )
14934 - reg_node - emit a node
14936 STATIC regnode * /* Location. */
14937 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14941 regnode * const ret = RExC_emit;
14942 GET_RE_DEBUG_FLAGS_DECL;
14944 PERL_ARGS_ASSERT_REG_NODE;
14947 SIZE_ALIGN(RExC_size);
14951 if (RExC_emit >= RExC_emit_bound)
14952 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14953 op, RExC_emit, RExC_emit_bound);
14955 NODE_ALIGN_FILL(ret);
14957 FILL_ADVANCE_NODE(ptr, op);
14958 #ifdef RE_TRACK_PATTERN_OFFSETS
14959 if (RExC_offsets) { /* MJD */
14961 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14962 "reg_node", __LINE__,
14964 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14965 ? "Overwriting end of array!\n" : "OK",
14966 (UV)(RExC_emit - RExC_emit_start),
14967 (UV)(RExC_parse - RExC_start),
14968 (UV)RExC_offsets[0]));
14969 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14977 - reganode - emit a node with an argument
14979 STATIC regnode * /* Location. */
14980 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14984 regnode * const ret = RExC_emit;
14985 GET_RE_DEBUG_FLAGS_DECL;
14987 PERL_ARGS_ASSERT_REGANODE;
14990 SIZE_ALIGN(RExC_size);
14995 assert(2==regarglen[op]+1);
14997 Anything larger than this has to allocate the extra amount.
14998 If we changed this to be:
15000 RExC_size += (1 + regarglen[op]);
15002 then it wouldn't matter. Its not clear what side effect
15003 might come from that so its not done so far.
15008 if (RExC_emit >= RExC_emit_bound)
15009 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15010 op, RExC_emit, RExC_emit_bound);
15012 NODE_ALIGN_FILL(ret);
15014 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15015 #ifdef RE_TRACK_PATTERN_OFFSETS
15016 if (RExC_offsets) { /* MJD */
15018 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15022 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
15023 "Overwriting end of array!\n" : "OK",
15024 (UV)(RExC_emit - RExC_emit_start),
15025 (UV)(RExC_parse - RExC_start),
15026 (UV)RExC_offsets[0]));
15027 Set_Cur_Node_Offset;
15035 - reguni - emit (if appropriate) a Unicode character
15037 PERL_STATIC_INLINE STRLEN
15038 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
15042 PERL_ARGS_ASSERT_REGUNI;
15044 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
15048 - reginsert - insert an operator in front of already-emitted operand
15050 * Means relocating the operand.
15053 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15059 const int offset = regarglen[(U8)op];
15060 const int size = NODE_STEP_REGNODE + offset;
15061 GET_RE_DEBUG_FLAGS_DECL;
15063 PERL_ARGS_ASSERT_REGINSERT;
15064 PERL_UNUSED_ARG(depth);
15065 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15066 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15075 if (RExC_open_parens) {
15077 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15078 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15079 if ( RExC_open_parens[paren] >= opnd ) {
15080 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15081 RExC_open_parens[paren] += size;
15083 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15085 if ( RExC_close_parens[paren] >= opnd ) {
15086 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15087 RExC_close_parens[paren] += size;
15089 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15094 while (src > opnd) {
15095 StructCopy(--src, --dst, regnode);
15096 #ifdef RE_TRACK_PATTERN_OFFSETS
15097 if (RExC_offsets) { /* MJD 20010112 */
15099 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15103 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15104 ? "Overwriting end of array!\n" : "OK",
15105 (UV)(src - RExC_emit_start),
15106 (UV)(dst - RExC_emit_start),
15107 (UV)RExC_offsets[0]));
15108 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15109 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15115 place = opnd; /* Op node, where operand used to be. */
15116 #ifdef RE_TRACK_PATTERN_OFFSETS
15117 if (RExC_offsets) { /* MJD */
15119 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15123 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15124 ? "Overwriting end of array!\n" : "OK",
15125 (UV)(place - RExC_emit_start),
15126 (UV)(RExC_parse - RExC_start),
15127 (UV)RExC_offsets[0]));
15128 Set_Node_Offset(place, RExC_parse);
15129 Set_Node_Length(place, 1);
15132 src = NEXTOPER(place);
15133 FILL_ADVANCE_NODE(place, op);
15134 Zero(src, offset, regnode);
15138 - regtail - set the next-pointer at the end of a node chain of p to val.
15139 - SEE ALSO: regtail_study
15141 /* TODO: All three parms should be const */
15143 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15144 const regnode *val,U32 depth)
15148 GET_RE_DEBUG_FLAGS_DECL;
15150 PERL_ARGS_ASSERT_REGTAIL;
15152 PERL_UNUSED_ARG(depth);
15158 /* Find last node. */
15161 regnode * const temp = regnext(scan);
15163 SV * const mysv=sv_newmortal();
15164 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15165 regprop(RExC_rx, mysv, scan);
15166 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15167 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
15168 (temp == NULL ? "->" : ""),
15169 (temp == NULL ? PL_reg_name[OP(val)] : "")
15177 if (reg_off_by_arg[OP(scan)]) {
15178 ARG_SET(scan, val - scan);
15181 NEXT_OFF(scan) = val - scan;
15187 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15188 - Look for optimizable sequences at the same time.
15189 - currently only looks for EXACT chains.
15191 This is experimental code. The idea is to use this routine to perform
15192 in place optimizations on branches and groups as they are constructed,
15193 with the long term intention of removing optimization from study_chunk so
15194 that it is purely analytical.
15196 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15197 to control which is which.
15200 /* TODO: All four parms should be const */
15203 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15204 const regnode *val,U32 depth)
15209 #ifdef EXPERIMENTAL_INPLACESCAN
15212 GET_RE_DEBUG_FLAGS_DECL;
15214 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15220 /* Find last node. */
15224 regnode * const temp = regnext(scan);
15225 #ifdef EXPERIMENTAL_INPLACESCAN
15226 if (PL_regkind[OP(scan)] == EXACT) {
15227 bool unfolded_multi_char; /* Unexamined in this routine */
15228 if (join_exact(pRExC_state, scan, &min,
15229 &unfolded_multi_char, 1, val, depth+1))
15234 switch (OP(scan)) {
15237 case EXACTFA_NO_TRIE:
15242 if( exact == PSEUDO )
15244 else if ( exact != OP(scan) )
15253 SV * const mysv=sv_newmortal();
15254 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
15255 regprop(RExC_rx, mysv, scan);
15256 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
15257 SvPV_nolen_const(mysv),
15258 REG_NODE_NUM(scan),
15259 PL_reg_name[exact]);
15266 SV * const mysv_val=sv_newmortal();
15267 DEBUG_PARSE_MSG("");
15268 regprop(RExC_rx, mysv_val, val);
15269 PerlIO_printf(Perl_debug_log,
15270 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
15271 SvPV_nolen_const(mysv_val),
15272 (IV)REG_NODE_NUM(val),
15276 if (reg_off_by_arg[OP(scan)]) {
15277 ARG_SET(scan, val - scan);
15280 NEXT_OFF(scan) = val - scan;
15288 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
15293 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
15298 for (bit=0; bit<32; bit++) {
15299 if (flags & (1<<bit)) {
15300 if (!set++ && lead)
15301 PerlIO_printf(Perl_debug_log, "%s",lead);
15302 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
15307 PerlIO_printf(Perl_debug_log, "\n");
15309 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15314 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
15320 for (bit=0; bit<32; bit++) {
15321 if (flags & (1<<bit)) {
15322 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
15325 if (!set++ && lead)
15326 PerlIO_printf(Perl_debug_log, "%s",lead);
15327 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
15330 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
15331 if (!set++ && lead) {
15332 PerlIO_printf(Perl_debug_log, "%s",lead);
15335 case REGEX_UNICODE_CHARSET:
15336 PerlIO_printf(Perl_debug_log, "UNICODE");
15338 case REGEX_LOCALE_CHARSET:
15339 PerlIO_printf(Perl_debug_log, "LOCALE");
15341 case REGEX_ASCII_RESTRICTED_CHARSET:
15342 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
15344 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
15345 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
15348 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
15354 PerlIO_printf(Perl_debug_log, "\n");
15356 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15362 Perl_regdump(pTHX_ const regexp *r)
15366 SV * const sv = sv_newmortal();
15367 SV *dsv= sv_newmortal();
15368 RXi_GET_DECL(r,ri);
15369 GET_RE_DEBUG_FLAGS_DECL;
15371 PERL_ARGS_ASSERT_REGDUMP;
15373 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
15375 /* Header fields of interest. */
15376 if (r->anchored_substr) {
15377 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
15378 RE_SV_DUMPLEN(r->anchored_substr), 30);
15379 PerlIO_printf(Perl_debug_log,
15380 "anchored %s%s at %"IVdf" ",
15381 s, RE_SV_TAIL(r->anchored_substr),
15382 (IV)r->anchored_offset);
15383 } else if (r->anchored_utf8) {
15384 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
15385 RE_SV_DUMPLEN(r->anchored_utf8), 30);
15386 PerlIO_printf(Perl_debug_log,
15387 "anchored utf8 %s%s at %"IVdf" ",
15388 s, RE_SV_TAIL(r->anchored_utf8),
15389 (IV)r->anchored_offset);
15391 if (r->float_substr) {
15392 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
15393 RE_SV_DUMPLEN(r->float_substr), 30);
15394 PerlIO_printf(Perl_debug_log,
15395 "floating %s%s at %"IVdf"..%"UVuf" ",
15396 s, RE_SV_TAIL(r->float_substr),
15397 (IV)r->float_min_offset, (UV)r->float_max_offset);
15398 } else if (r->float_utf8) {
15399 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
15400 RE_SV_DUMPLEN(r->float_utf8), 30);
15401 PerlIO_printf(Perl_debug_log,
15402 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
15403 s, RE_SV_TAIL(r->float_utf8),
15404 (IV)r->float_min_offset, (UV)r->float_max_offset);
15406 if (r->check_substr || r->check_utf8)
15407 PerlIO_printf(Perl_debug_log,
15409 (r->check_substr == r->float_substr
15410 && r->check_utf8 == r->float_utf8
15411 ? "(checking floating" : "(checking anchored"));
15412 if (r->intflags & PREGf_NOSCAN)
15413 PerlIO_printf(Perl_debug_log, " noscan");
15414 if (r->extflags & RXf_CHECK_ALL)
15415 PerlIO_printf(Perl_debug_log, " isall");
15416 if (r->check_substr || r->check_utf8)
15417 PerlIO_printf(Perl_debug_log, ") ");
15419 if (ri->regstclass) {
15420 regprop(r, sv, ri->regstclass);
15421 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
15423 if (r->intflags & PREGf_ANCH) {
15424 PerlIO_printf(Perl_debug_log, "anchored");
15425 if (r->intflags & PREGf_ANCH_BOL)
15426 PerlIO_printf(Perl_debug_log, "(BOL)");
15427 if (r->intflags & PREGf_ANCH_MBOL)
15428 PerlIO_printf(Perl_debug_log, "(MBOL)");
15429 if (r->intflags & PREGf_ANCH_SBOL)
15430 PerlIO_printf(Perl_debug_log, "(SBOL)");
15431 if (r->intflags & PREGf_ANCH_GPOS)
15432 PerlIO_printf(Perl_debug_log, "(GPOS)");
15433 PerlIO_putc(Perl_debug_log, ' ');
15435 if (r->intflags & PREGf_GPOS_SEEN)
15436 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
15437 if (r->intflags & PREGf_SKIP)
15438 PerlIO_printf(Perl_debug_log, "plus ");
15439 if (r->intflags & PREGf_IMPLICIT)
15440 PerlIO_printf(Perl_debug_log, "implicit ");
15441 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
15442 if (r->extflags & RXf_EVAL_SEEN)
15443 PerlIO_printf(Perl_debug_log, "with eval ");
15444 PerlIO_printf(Perl_debug_log, "\n");
15446 regdump_extflags("r->extflags: ",r->extflags);
15447 regdump_intflags("r->intflags: ",r->intflags);
15450 PERL_ARGS_ASSERT_REGDUMP;
15451 PERL_UNUSED_CONTEXT;
15452 PERL_UNUSED_ARG(r);
15453 #endif /* DEBUGGING */
15457 - regprop - printable representation of opcode
15461 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
15467 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
15468 static const char * const anyofs[] = {
15469 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
15470 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
15471 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
15472 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
15473 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
15474 || _CC_VERTSPACE != 16
15475 #error Need to adjust order of anyofs[]
15512 RXi_GET_DECL(prog,progi);
15513 GET_RE_DEBUG_FLAGS_DECL;
15515 PERL_ARGS_ASSERT_REGPROP;
15519 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
15520 /* It would be nice to FAIL() here, but this may be called from
15521 regexec.c, and it would be hard to supply pRExC_state. */
15522 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
15523 (int)OP(o), (int)REGNODE_MAX);
15524 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
15526 k = PL_regkind[OP(o)];
15529 sv_catpvs(sv, " ");
15530 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
15531 * is a crude hack but it may be the best for now since
15532 * we have no flag "this EXACTish node was UTF-8"
15534 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
15535 PERL_PV_ESCAPE_UNI_DETECT |
15536 PERL_PV_ESCAPE_NONASCII |
15537 PERL_PV_PRETTY_ELLIPSES |
15538 PERL_PV_PRETTY_LTGT |
15539 PERL_PV_PRETTY_NOCLEAR
15541 } else if (k == TRIE) {
15542 /* print the details of the trie in dumpuntil instead, as
15543 * progi->data isn't available here */
15544 const char op = OP(o);
15545 const U32 n = ARG(o);
15546 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
15547 (reg_ac_data *)progi->data->data[n] :
15549 const reg_trie_data * const trie
15550 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
15552 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
15553 DEBUG_TRIE_COMPILE_r(
15554 Perl_sv_catpvf(aTHX_ sv,
15555 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
15556 (UV)trie->startstate,
15557 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
15558 (UV)trie->wordcount,
15561 (UV)TRIE_CHARCOUNT(trie),
15562 (UV)trie->uniquecharcount
15565 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
15566 sv_catpvs(sv, "[");
15567 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
15569 : TRIE_BITMAP(trie));
15570 sv_catpvs(sv, "]");
15573 } else if (k == CURLY) {
15574 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
15575 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
15576 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
15578 else if (k == WHILEM && o->flags) /* Ordinal/of */
15579 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
15580 else if (k == REF || k == OPEN || k == CLOSE
15581 || k == GROUPP || OP(o)==ACCEPT)
15583 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
15584 if ( RXp_PAREN_NAMES(prog) ) {
15585 if ( k != REF || (OP(o) < NREF)) {
15586 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
15587 SV **name= av_fetch(list, ARG(o), 0 );
15589 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15592 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
15593 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
15594 I32 *nums=(I32*)SvPVX(sv_dat);
15595 SV **name= av_fetch(list, nums[0], 0 );
15598 for ( n=0; n<SvIVX(sv_dat); n++ ) {
15599 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
15600 (n ? "," : ""), (IV)nums[n]);
15602 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15606 } else if (k == GOSUB)
15607 /* Paren and offset */
15608 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
15609 else if (k == VERB) {
15611 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
15612 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
15613 } else if (k == LOGICAL)
15614 /* 2: embedded, otherwise 1 */
15615 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
15616 else if (k == ANYOF) {
15617 const U8 flags = ANYOF_FLAGS(o);
15621 if (flags & ANYOF_LOCALE)
15622 sv_catpvs(sv, "{loc}");
15623 if (flags & ANYOF_LOC_FOLD)
15624 sv_catpvs(sv, "{i}");
15625 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
15626 if (flags & ANYOF_INVERT)
15627 sv_catpvs(sv, "^");
15629 /* output what the standard cp 0-255 bitmap matches */
15630 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
15632 /* output any special charclass tests (used entirely under use
15634 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
15636 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
15637 if (ANYOF_POSIXL_TEST(o,i)) {
15638 sv_catpv(sv, anyofs[i]);
15644 if ((flags & ANYOF_ABOVE_LATIN1_ALL)
15645 || ANYOF_UTF8_LOCALE_INVLIST(o) || ANYOF_NONBITMAP(o))
15648 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
15649 if (flags & ANYOF_INVERT)
15650 /*make sure the invert info is in each */
15651 sv_catpvs(sv, "^");
15654 if (flags & ANYOF_NON_UTF8_NON_ASCII_ALL) {
15655 sv_catpvs(sv, "{non-utf8-latin1-all}");
15658 /* output information about the unicode matching */
15659 if (flags & ANYOF_ABOVE_LATIN1_ALL)
15660 sv_catpvs(sv, "{unicode_all}");
15661 else if (ANYOF_NONBITMAP(o)) {
15662 SV *lv; /* Set if there is something outside the bit map. */
15663 bool byte_output = FALSE; /* If something in the bitmap has
15666 /* Get the stuff that wasn't in the bitmap */
15667 (void) regclass_swash(prog, o, FALSE, &lv, NULL);
15668 if (lv && lv != &PL_sv_undef) {
15669 char *s = savesvpv(lv);
15670 char * const origs = s;
15672 while (*s && *s != '\n')
15676 const char * const t = ++s;
15678 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
15679 sv_catpvs(sv, "{outside bitmap}");
15682 sv_catpvs(sv, "{utf8}");
15686 sv_catpvs(sv, " ");
15692 /* Truncate very long output */
15693 if (s - origs > 256) {
15694 Perl_sv_catpvf(aTHX_ sv,
15696 (int) (s - origs - 1),
15702 else if (*s == '\t') {
15716 SvREFCNT_dec_NN(lv);
15720 /* Output any UTF-8 locale code points */
15721 if (flags & ANYOF_LOC_FOLD && ANYOF_UTF8_LOCALE_INVLIST(o)) {
15723 int max_entries = 256;
15725 sv_catpvs(sv, "{utf8 locale}");
15726 invlist_iterinit(ANYOF_UTF8_LOCALE_INVLIST(o));
15727 while (invlist_iternext(ANYOF_UTF8_LOCALE_INVLIST(o),
15729 put_range(sv, start, end);
15731 if (max_entries < 0) {
15732 sv_catpvs(sv, "...");
15736 invlist_iterfinish(ANYOF_UTF8_LOCALE_INVLIST(o));
15740 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
15742 else if (k == POSIXD || k == NPOSIXD) {
15743 U8 index = FLAGS(o) * 2;
15744 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
15745 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
15748 if (*anyofs[index] != '[') {
15751 sv_catpv(sv, anyofs[index]);
15752 if (*anyofs[index] != '[') {
15757 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
15758 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
15760 PERL_UNUSED_CONTEXT;
15761 PERL_UNUSED_ARG(sv);
15762 PERL_UNUSED_ARG(o);
15763 PERL_UNUSED_ARG(prog);
15764 #endif /* DEBUGGING */
15768 Perl_re_intuit_string(pTHX_ REGEXP * const r)
15769 { /* Assume that RE_INTUIT is set */
15771 struct regexp *const prog = ReANY(r);
15772 GET_RE_DEBUG_FLAGS_DECL;
15774 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
15775 PERL_UNUSED_CONTEXT;
15779 const char * const s = SvPV_nolen_const(prog->check_substr
15780 ? prog->check_substr : prog->check_utf8);
15782 if (!PL_colorset) reginitcolors();
15783 PerlIO_printf(Perl_debug_log,
15784 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
15786 prog->check_substr ? "" : "utf8 ",
15787 PL_colors[5],PL_colors[0],
15790 (strlen(s) > 60 ? "..." : ""));
15793 return prog->check_substr ? prog->check_substr : prog->check_utf8;
15799 handles refcounting and freeing the perl core regexp structure. When
15800 it is necessary to actually free the structure the first thing it
15801 does is call the 'free' method of the regexp_engine associated to
15802 the regexp, allowing the handling of the void *pprivate; member
15803 first. (This routine is not overridable by extensions, which is why
15804 the extensions free is called first.)
15806 See regdupe and regdupe_internal if you change anything here.
15808 #ifndef PERL_IN_XSUB_RE
15810 Perl_pregfree(pTHX_ REGEXP *r)
15816 Perl_pregfree2(pTHX_ REGEXP *rx)
15819 struct regexp *const r = ReANY(rx);
15820 GET_RE_DEBUG_FLAGS_DECL;
15822 PERL_ARGS_ASSERT_PREGFREE2;
15824 if (r->mother_re) {
15825 ReREFCNT_dec(r->mother_re);
15827 CALLREGFREE_PVT(rx); /* free the private data */
15828 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15829 Safefree(r->xpv_len_u.xpvlenu_pv);
15832 SvREFCNT_dec(r->anchored_substr);
15833 SvREFCNT_dec(r->anchored_utf8);
15834 SvREFCNT_dec(r->float_substr);
15835 SvREFCNT_dec(r->float_utf8);
15836 Safefree(r->substrs);
15838 RX_MATCH_COPY_FREE(rx);
15839 #ifdef PERL_ANY_COW
15840 SvREFCNT_dec(r->saved_copy);
15843 SvREFCNT_dec(r->qr_anoncv);
15844 rx->sv_u.svu_rx = 0;
15849 This is a hacky workaround to the structural issue of match results
15850 being stored in the regexp structure which is in turn stored in
15851 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15852 could be PL_curpm in multiple contexts, and could require multiple
15853 result sets being associated with the pattern simultaneously, such
15854 as when doing a recursive match with (??{$qr})
15856 The solution is to make a lightweight copy of the regexp structure
15857 when a qr// is returned from the code executed by (??{$qr}) this
15858 lightweight copy doesn't actually own any of its data except for
15859 the starp/end and the actual regexp structure itself.
15865 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15867 struct regexp *ret;
15868 struct regexp *const r = ReANY(rx);
15869 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15871 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15874 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15876 SvOK_off((SV *)ret_x);
15878 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15879 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15880 made both spots point to the same regexp body.) */
15881 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15882 assert(!SvPVX(ret_x));
15883 ret_x->sv_u.svu_rx = temp->sv_any;
15884 temp->sv_any = NULL;
15885 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15886 SvREFCNT_dec_NN(temp);
15887 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15888 ing below will not set it. */
15889 SvCUR_set(ret_x, SvCUR(rx));
15892 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15893 sv_force_normal(sv) is called. */
15895 ret = ReANY(ret_x);
15897 SvFLAGS(ret_x) |= SvUTF8(rx);
15898 /* We share the same string buffer as the original regexp, on which we
15899 hold a reference count, incremented when mother_re is set below.
15900 The string pointer is copied here, being part of the regexp struct.
15902 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15903 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15905 const I32 npar = r->nparens+1;
15906 Newx(ret->offs, npar, regexp_paren_pair);
15907 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15910 Newx(ret->substrs, 1, struct reg_substr_data);
15911 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15913 SvREFCNT_inc_void(ret->anchored_substr);
15914 SvREFCNT_inc_void(ret->anchored_utf8);
15915 SvREFCNT_inc_void(ret->float_substr);
15916 SvREFCNT_inc_void(ret->float_utf8);
15918 /* check_substr and check_utf8, if non-NULL, point to either their
15919 anchored or float namesakes, and don't hold a second reference. */
15921 RX_MATCH_COPIED_off(ret_x);
15922 #ifdef PERL_ANY_COW
15923 ret->saved_copy = NULL;
15925 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15926 SvREFCNT_inc_void(ret->qr_anoncv);
15932 /* regfree_internal()
15934 Free the private data in a regexp. This is overloadable by
15935 extensions. Perl takes care of the regexp structure in pregfree(),
15936 this covers the *pprivate pointer which technically perl doesn't
15937 know about, however of course we have to handle the
15938 regexp_internal structure when no extension is in use.
15940 Note this is called before freeing anything in the regexp
15945 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15948 struct regexp *const r = ReANY(rx);
15949 RXi_GET_DECL(r,ri);
15950 GET_RE_DEBUG_FLAGS_DECL;
15952 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15958 SV *dsv= sv_newmortal();
15959 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15960 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15961 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15962 PL_colors[4],PL_colors[5],s);
15965 #ifdef RE_TRACK_PATTERN_OFFSETS
15967 Safefree(ri->u.offsets); /* 20010421 MJD */
15969 if (ri->code_blocks) {
15971 for (n = 0; n < ri->num_code_blocks; n++)
15972 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15973 Safefree(ri->code_blocks);
15977 int n = ri->data->count;
15980 /* If you add a ->what type here, update the comment in regcomp.h */
15981 switch (ri->data->what[n]) {
15987 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15990 Safefree(ri->data->data[n]);
15996 { /* Aho Corasick add-on structure for a trie node.
15997 Used in stclass optimization only */
15999 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16001 refcount = --aho->refcount;
16004 PerlMemShared_free(aho->states);
16005 PerlMemShared_free(aho->fail);
16006 /* do this last!!!! */
16007 PerlMemShared_free(ri->data->data[n]);
16008 PerlMemShared_free(ri->regstclass);
16014 /* trie structure. */
16016 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16018 refcount = --trie->refcount;
16021 PerlMemShared_free(trie->charmap);
16022 PerlMemShared_free(trie->states);
16023 PerlMemShared_free(trie->trans);
16025 PerlMemShared_free(trie->bitmap);
16027 PerlMemShared_free(trie->jump);
16028 PerlMemShared_free(trie->wordinfo);
16029 /* do this last!!!! */
16030 PerlMemShared_free(ri->data->data[n]);
16035 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16036 ri->data->what[n]);
16039 Safefree(ri->data->what);
16040 Safefree(ri->data);
16046 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16047 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16048 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16051 re_dup - duplicate a regexp.
16053 This routine is expected to clone a given regexp structure. It is only
16054 compiled under USE_ITHREADS.
16056 After all of the core data stored in struct regexp is duplicated
16057 the regexp_engine.dupe method is used to copy any private data
16058 stored in the *pprivate pointer. This allows extensions to handle
16059 any duplication it needs to do.
16061 See pregfree() and regfree_internal() if you change anything here.
16063 #if defined(USE_ITHREADS)
16064 #ifndef PERL_IN_XSUB_RE
16066 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16070 const struct regexp *r = ReANY(sstr);
16071 struct regexp *ret = ReANY(dstr);
16073 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16075 npar = r->nparens+1;
16076 Newx(ret->offs, npar, regexp_paren_pair);
16077 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16079 if (ret->substrs) {
16080 /* Do it this way to avoid reading from *r after the StructCopy().
16081 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16082 cache, it doesn't matter. */
16083 const bool anchored = r->check_substr
16084 ? r->check_substr == r->anchored_substr
16085 : r->check_utf8 == r->anchored_utf8;
16086 Newx(ret->substrs, 1, struct reg_substr_data);
16087 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16089 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16090 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16091 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16092 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16094 /* check_substr and check_utf8, if non-NULL, point to either their
16095 anchored or float namesakes, and don't hold a second reference. */
16097 if (ret->check_substr) {
16099 assert(r->check_utf8 == r->anchored_utf8);
16100 ret->check_substr = ret->anchored_substr;
16101 ret->check_utf8 = ret->anchored_utf8;
16103 assert(r->check_substr == r->float_substr);
16104 assert(r->check_utf8 == r->float_utf8);
16105 ret->check_substr = ret->float_substr;
16106 ret->check_utf8 = ret->float_utf8;
16108 } else if (ret->check_utf8) {
16110 ret->check_utf8 = ret->anchored_utf8;
16112 ret->check_utf8 = ret->float_utf8;
16117 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16118 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16121 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16123 if (RX_MATCH_COPIED(dstr))
16124 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16126 ret->subbeg = NULL;
16127 #ifdef PERL_ANY_COW
16128 ret->saved_copy = NULL;
16131 /* Whether mother_re be set or no, we need to copy the string. We
16132 cannot refrain from copying it when the storage points directly to
16133 our mother regexp, because that's
16134 1: a buffer in a different thread
16135 2: something we no longer hold a reference on
16136 so we need to copy it locally. */
16137 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16138 ret->mother_re = NULL;
16140 #endif /* PERL_IN_XSUB_RE */
16145 This is the internal complement to regdupe() which is used to copy
16146 the structure pointed to by the *pprivate pointer in the regexp.
16147 This is the core version of the extension overridable cloning hook.
16148 The regexp structure being duplicated will be copied by perl prior
16149 to this and will be provided as the regexp *r argument, however
16150 with the /old/ structures pprivate pointer value. Thus this routine
16151 may override any copying normally done by perl.
16153 It returns a pointer to the new regexp_internal structure.
16157 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
16160 struct regexp *const r = ReANY(rx);
16161 regexp_internal *reti;
16163 RXi_GET_DECL(r,ri);
16165 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
16169 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
16170 char, regexp_internal);
16171 Copy(ri->program, reti->program, len+1, regnode);
16173 reti->num_code_blocks = ri->num_code_blocks;
16174 if (ri->code_blocks) {
16176 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
16177 struct reg_code_block);
16178 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
16179 struct reg_code_block);
16180 for (n = 0; n < ri->num_code_blocks; n++)
16181 reti->code_blocks[n].src_regex = (REGEXP*)
16182 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
16185 reti->code_blocks = NULL;
16187 reti->regstclass = NULL;
16190 struct reg_data *d;
16191 const int count = ri->data->count;
16194 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
16195 char, struct reg_data);
16196 Newx(d->what, count, U8);
16199 for (i = 0; i < count; i++) {
16200 d->what[i] = ri->data->what[i];
16201 switch (d->what[i]) {
16202 /* see also regcomp.h and regfree_internal() */
16203 case 'a': /* actually an AV, but the dup function is identical. */
16207 case 'u': /* actually an HV, but the dup function is identical. */
16208 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
16211 /* This is cheating. */
16212 Newx(d->data[i], 1, regnode_ssc);
16213 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
16214 reti->regstclass = (regnode*)d->data[i];
16217 /* Trie stclasses are readonly and can thus be shared
16218 * without duplication. We free the stclass in pregfree
16219 * when the corresponding reg_ac_data struct is freed.
16221 reti->regstclass= ri->regstclass;
16225 ((reg_trie_data*)ri->data->data[i])->refcount++;
16230 d->data[i] = ri->data->data[i];
16233 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
16234 ri->data->what[i]);
16243 reti->name_list_idx = ri->name_list_idx;
16245 #ifdef RE_TRACK_PATTERN_OFFSETS
16246 if (ri->u.offsets) {
16247 Newx(reti->u.offsets, 2*len+1, U32);
16248 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
16251 SetProgLen(reti,len);
16254 return (void*)reti;
16257 #endif /* USE_ITHREADS */
16259 #ifndef PERL_IN_XSUB_RE
16262 - regnext - dig the "next" pointer out of a node
16265 Perl_regnext(pTHX_ regnode *p)
16273 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
16274 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16275 (int)OP(p), (int)REGNODE_MAX);
16278 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
16287 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
16290 STRLEN l1 = strlen(pat1);
16291 STRLEN l2 = strlen(pat2);
16294 const char *message;
16296 PERL_ARGS_ASSERT_RE_CROAK2;
16302 Copy(pat1, buf, l1 , char);
16303 Copy(pat2, buf + l1, l2 , char);
16304 buf[l1 + l2] = '\n';
16305 buf[l1 + l2 + 1] = '\0';
16306 va_start(args, pat2);
16307 msv = vmess(buf, &args);
16309 message = SvPV_const(msv,l1);
16312 Copy(message, buf, l1 , char);
16313 /* l1-1 to avoid \n */
16314 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
16317 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
16319 #ifndef PERL_IN_XSUB_RE
16321 Perl_save_re_context(pTHX)
16325 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
16327 const REGEXP * const rx = PM_GETRE(PL_curpm);
16330 for (i = 1; i <= RX_NPARENS(rx); i++) {
16331 char digits[TYPE_CHARS(long)];
16332 const STRLEN len = my_snprintf(digits, sizeof(digits),
16334 GV *const *const gvp
16335 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
16338 GV * const gv = *gvp;
16339 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
16351 S_put_byte(pTHX_ SV *sv, int c)
16353 PERL_ARGS_ASSERT_PUT_BYTE;
16357 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
16358 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
16359 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
16360 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
16361 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
16364 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
16369 const char string = c;
16370 if (c == '-' || c == ']' || c == '\\' || c == '^')
16371 sv_catpvs(sv, "\\");
16372 sv_catpvn(sv, &string, 1);
16377 S_put_range(pTHX_ SV *sv, UV start, UV end)
16380 /* Appends to 'sv' a displayable version of the range of code points from
16381 * 'start' to 'end' */
16383 assert(start <= end);
16385 PERL_ARGS_ASSERT_PUT_RANGE;
16387 if (end - start < 3) { /* Individual chars in short ranges */
16388 for (; start <= end; start++)
16389 put_byte(sv, start);
16391 else if ( end > 255
16392 || ! isALPHANUMERIC(start)
16393 || ! isALPHANUMERIC(end)
16394 || isDIGIT(start) != isDIGIT(end)
16395 || isUPPER(start) != isUPPER(end)
16396 || isLOWER(start) != isLOWER(end)
16398 /* This final test should get optimized out except on EBCDIC
16399 * platforms, where it causes ranges that cross discontinuities
16400 * like i/j to be shown as hex instead of the misleading,
16401 * e.g. H-K (since that range includes more than H, I, J, K).
16403 || (end - start) != NATIVE_TO_ASCII(end) - NATIVE_TO_ASCII(start))
16405 Perl_sv_catpvf(aTHX_ sv, "\\x{%02" UVXf "}-\\x{%02" UVXf "}",
16407 (end < 256) ? end : 255);
16409 else { /* Here, the ends of the range are both digits, or both uppercase,
16410 or both lowercase; and there's no discontinuity in the range
16411 (which could happen on EBCDIC platforms) */
16412 put_byte(sv, start);
16413 sv_catpvs(sv, "-");
16419 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
16421 /* Appends to 'sv' a displayable version of the innards of the bracketed
16422 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
16423 * output anything */
16426 bool has_output_anything = FALSE;
16428 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
16430 for (i = 0; i < 256; i++) {
16431 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
16433 /* The character at index i should be output. Find the next
16434 * character that should NOT be output */
16436 for (j = i + 1; j <= 256; j++) {
16437 if (! BITMAP_TEST((U8 *) bitmap, j)) {
16442 /* Everything between them is a single range that should be output
16444 put_range(sv, i, j - 1);
16445 has_output_anything = TRUE;
16450 return has_output_anything;
16453 #define CLEAR_OPTSTART \
16454 if (optstart) STMT_START { \
16455 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
16456 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
16460 #define DUMPUNTIL(b,e) \
16462 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
16464 STATIC const regnode *
16465 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
16466 const regnode *last, const regnode *plast,
16467 SV* sv, I32 indent, U32 depth)
16470 U8 op = PSEUDO; /* Arbitrary non-END op. */
16471 const regnode *next;
16472 const regnode *optstart= NULL;
16474 RXi_GET_DECL(r,ri);
16475 GET_RE_DEBUG_FLAGS_DECL;
16477 PERL_ARGS_ASSERT_DUMPUNTIL;
16479 #ifdef DEBUG_DUMPUNTIL
16480 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
16481 last ? last-start : 0,plast ? plast-start : 0);
16484 if (plast && plast < last)
16487 while (PL_regkind[op] != END && (!last || node < last)) {
16488 /* While that wasn't END last time... */
16491 if (op == CLOSE || op == WHILEM)
16493 next = regnext((regnode *)node);
16496 if (OP(node) == OPTIMIZED) {
16497 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
16504 regprop(r, sv, node);
16505 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
16506 (int)(2*indent + 1), "", SvPVX_const(sv));
16508 if (OP(node) != OPTIMIZED) {
16509 if (next == NULL) /* Next ptr. */
16510 PerlIO_printf(Perl_debug_log, " (0)");
16511 else if (PL_regkind[(U8)op] == BRANCH
16512 && PL_regkind[OP(next)] != BRANCH )
16513 PerlIO_printf(Perl_debug_log, " (FAIL)");
16515 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
16516 (void)PerlIO_putc(Perl_debug_log, '\n');
16520 if (PL_regkind[(U8)op] == BRANCHJ) {
16523 const regnode *nnode = (OP(next) == LONGJMP
16524 ? regnext((regnode *)next)
16526 if (last && nnode > last)
16528 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
16531 else if (PL_regkind[(U8)op] == BRANCH) {
16533 DUMPUNTIL(NEXTOPER(node), next);
16535 else if ( PL_regkind[(U8)op] == TRIE ) {
16536 const regnode *this_trie = node;
16537 const char op = OP(node);
16538 const U32 n = ARG(node);
16539 const reg_ac_data * const ac = op>=AHOCORASICK ?
16540 (reg_ac_data *)ri->data->data[n] :
16542 const reg_trie_data * const trie =
16543 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
16545 AV *const trie_words
16546 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
16548 const regnode *nextbranch= NULL;
16551 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
16552 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
16554 PerlIO_printf(Perl_debug_log, "%*s%s ",
16555 (int)(2*(indent+3)), "",
16557 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
16558 SvCUR(*elem_ptr), 60,
16559 PL_colors[0], PL_colors[1],
16561 ? PERL_PV_ESCAPE_UNI
16563 | PERL_PV_PRETTY_ELLIPSES
16564 | PERL_PV_PRETTY_LTGT
16569 U16 dist= trie->jump[word_idx+1];
16570 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
16571 (UV)((dist ? this_trie + dist : next) - start));
16574 nextbranch= this_trie + trie->jump[0];
16575 DUMPUNTIL(this_trie + dist, nextbranch);
16577 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
16578 nextbranch= regnext((regnode *)nextbranch);
16580 PerlIO_printf(Perl_debug_log, "\n");
16583 if (last && next > last)
16588 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
16589 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
16590 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
16592 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
16594 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
16596 else if ( op == PLUS || op == STAR) {
16597 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
16599 else if (PL_regkind[(U8)op] == ANYOF) {
16600 /* arglen 1 + class block */
16601 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
16602 ? ANYOF_POSIXL_FOLD_SKIP
16603 : (ANYOF_FLAGS(node) & ANYOF_POSIXL)
16604 ? ANYOF_POSIXL_SKIP
16606 node = NEXTOPER(node);
16608 else if (PL_regkind[(U8)op] == EXACT) {
16609 /* Literal string, where present. */
16610 node += NODE_SZ_STR(node) - 1;
16611 node = NEXTOPER(node);
16614 node = NEXTOPER(node);
16615 node += regarglen[(U8)op];
16617 if (op == CURLYX || op == OPEN)
16621 #ifdef DEBUG_DUMPUNTIL
16622 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
16627 #endif /* DEBUGGING */
16631 * c-indentation-style: bsd
16632 * c-basic-offset: 4
16633 * indent-tabs-mode: nil
16636 * ex: set ts=8 sts=4 sw=4 et: