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 */
163 #ifdef ADD_TO_REGEXEC
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_POSIXL_SSC_TEST_ANY_SET(ssc)) {
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);
963 ANYOF_POSIXL_ZERO(ssc);
968 S_ssc_is_cp_posixl_init(pTHX_ const RExC_state_t *pRExC_state,
969 const regnode_ssc *ssc)
971 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
972 * to the list of code points matched, and locale posix classes; hence does
973 * not check its flags) */
978 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
980 assert(is_ANYOF_SYNTHETIC(ssc));
982 invlist_iterinit(ssc->invlist);
983 ret = invlist_iternext(ssc->invlist, &start, &end)
987 invlist_iterfinish(ssc->invlist);
993 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1001 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1002 const regnode_charclass* const node)
1004 /* Returns a mortal inversion list defining which code points are matched
1005 * by 'node', which is of type ANYOF. Handles complementing the result if
1006 * appropriate. If some code points aren't knowable at this time, the
1007 * returned list must, and will, contain every code point that is a
1010 SV* invlist = sv_2mortal(_new_invlist(0));
1011 SV* only_utf8_locale_invlist = NULL;
1013 const U32 n = ARG(node);
1014 bool new_node_has_latin1 = FALSE;
1016 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1018 /* Look at the data structure created by S_set_ANYOF_arg() */
1019 if (n != ANYOF_NONBITMAP_EMPTY) {
1020 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1021 AV * const av = MUTABLE_AV(SvRV(rv));
1022 SV **const ary = AvARRAY(av);
1023 assert(RExC_rxi->data->what[n] == 's');
1025 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1026 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1028 else if (ary[0] && ary[0] != &PL_sv_undef) {
1030 /* Here, no compile-time swash, and there are things that won't be
1031 * known until runtime -- we have to assume it could be anything */
1032 return _add_range_to_invlist(invlist, 0, UV_MAX);
1034 else if (ary[3] && ary[3] != &PL_sv_undef) {
1036 /* Here no compile-time swash, and no run-time only data. Use the
1037 * node's inversion list */
1038 invlist = sv_2mortal(invlist_clone(ary[3]));
1041 /* Get the code points valid only under UTF-8 locales */
1042 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1043 && ary[2] && ary[2] != &PL_sv_undef)
1045 only_utf8_locale_invlist = ary[2];
1049 /* An ANYOF node contains a bitmap for the first 256 code points, and an
1050 * inversion list for the others, but if there are code points that should
1051 * match only conditionally on the target string being UTF-8, those are
1052 * placed in the inversion list, and not the bitmap. Since there are
1053 * circumstances under which they could match, they are included in the
1054 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1055 * here, so that when we invert below, the end result actually does include
1056 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1057 * before we add the unconditionally matched code points */
1058 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1059 _invlist_intersection_complement_2nd(invlist,
1064 /* Add in the points from the bit map */
1065 for (i = 0; i < 256; i++) {
1066 if (ANYOF_BITMAP_TEST(node, i)) {
1067 invlist = add_cp_to_invlist(invlist, i);
1068 new_node_has_latin1 = TRUE;
1072 /* If this can match all upper Latin1 code points, have to add them
1074 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1075 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1078 /* Similarly for these */
1079 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1080 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1083 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1084 _invlist_invert(invlist);
1086 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1088 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1089 * locale. We can skip this if there are no 0-255 at all. */
1090 _invlist_union(invlist, PL_Latin1, &invlist);
1093 /* Similarly add the UTF-8 locale possible matches. These have to be
1094 * deferred until after the non-UTF-8 locale ones are taken care of just
1095 * above, or it leads to wrong results under ANYOF_INVERT */
1096 if (only_utf8_locale_invlist) {
1097 _invlist_union_maybe_complement_2nd(invlist,
1098 only_utf8_locale_invlist,
1099 ANYOF_FLAGS(node) & ANYOF_INVERT,
1106 /* These two functions currently do the exact same thing */
1107 #define ssc_init_zero ssc_init
1109 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1110 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1112 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1113 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1114 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1117 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1118 const regnode_charclass *and_with)
1120 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1121 * another SSC or a regular ANYOF class. Can create false positives. */
1126 PERL_ARGS_ASSERT_SSC_AND;
1128 assert(is_ANYOF_SYNTHETIC(ssc));
1130 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1131 * the code point inversion list and just the relevant flags */
1132 if (is_ANYOF_SYNTHETIC(and_with)) {
1133 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1134 anded_flags = ANYOF_FLAGS(and_with);
1136 /* XXX This is a kludge around what appears to be deficiencies in the
1137 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1138 * there are paths through the optimizer where it doesn't get weeded
1139 * out when it should. And if we don't make some extra provision for
1140 * it like the code just below, it doesn't get added when it should.
1141 * This solution is to add it only when AND'ing, which is here, and
1142 * only when what is being AND'ed is the pristine, original node
1143 * matching anything. Thus it is like adding it to ssc_anything() but
1144 * only when the result is to be AND'ed. Probably the same solution
1145 * could be adopted for the same problem we have with /l matching,
1146 * which is solved differently in S_ssc_init(), and that would lead to
1147 * fewer false positives than that solution has. But if this solution
1148 * creates bugs, the consequences are only that a warning isn't raised
1149 * that should be; while the consequences for having /l bugs is
1150 * incorrect matches */
1151 if (ssc_is_anything((regnode_ssc *)and_with)) {
1152 anded_flags |= ANYOF_WARN_SUPER;
1156 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1157 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1160 ANYOF_FLAGS(ssc) &= anded_flags;
1162 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1163 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1164 * 'and_with' may be inverted. When not inverted, we have the situation of
1166 * (C1 | P1) & (C2 | P2)
1167 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1168 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1169 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1170 * <= ((C1 & C2) | P1 | P2)
1171 * Alternatively, the last few steps could be:
1172 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1173 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1174 * <= (C1 | C2 | (P1 & P2))
1175 * We favor the second approach if either P1 or P2 is non-empty. This is
1176 * because these components are a barrier to doing optimizations, as what
1177 * they match cannot be known until the moment of matching as they are
1178 * dependent on the current locale, 'AND"ing them likely will reduce or
1180 * But we can do better if we know that C1,P1 are in their initial state (a
1181 * frequent occurrence), each matching everything:
1182 * (<everything>) & (C2 | P2) = C2 | P2
1183 * Similarly, if C2,P2 are in their initial state (again a frequent
1184 * occurrence), the result is a no-op
1185 * (C1 | P1) & (<everything>) = C1 | P1
1188 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1189 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1190 * <= (C1 & ~C2) | (P1 & ~P2)
1193 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1194 && ! is_ANYOF_SYNTHETIC(and_with))
1198 ssc_intersection(ssc,
1200 FALSE /* Has already been inverted */
1203 /* If either P1 or P2 is empty, the intersection will be also; can skip
1205 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1206 ANYOF_POSIXL_ZERO(ssc);
1208 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1210 /* Note that the Posix class component P from 'and_with' actually
1212 * P = Pa | Pb | ... | Pn
1213 * where each component is one posix class, such as in [\w\s].
1215 * ~P = ~(Pa | Pb | ... | Pn)
1216 * = ~Pa & ~Pb & ... & ~Pn
1217 * <= ~Pa | ~Pb | ... | ~Pn
1218 * The last is something we can easily calculate, but unfortunately
1219 * is likely to have many false positives. We could do better
1220 * in some (but certainly not all) instances if two classes in
1221 * P have known relationships. For example
1222 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1224 * :lower: & :print: = :lower:
1225 * And similarly for classes that must be disjoint. For example,
1226 * since \s and \w can have no elements in common based on rules in
1227 * the POSIX standard,
1228 * \w & ^\S = nothing
1229 * Unfortunately, some vendor locales do not meet the Posix
1230 * standard, in particular almost everything by Microsoft.
1231 * The loop below just changes e.g., \w into \W and vice versa */
1233 regnode_charclass_posixl temp;
1234 int add = 1; /* To calculate the index of the complement */
1236 ANYOF_POSIXL_ZERO(&temp);
1237 for (i = 0; i < ANYOF_MAX; i++) {
1239 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1240 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1242 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1243 ANYOF_POSIXL_SET(&temp, i + add);
1245 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1247 ANYOF_POSIXL_AND(&temp, ssc);
1249 } /* else ssc already has no posixes */
1250 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1251 in its initial state */
1252 else if (! is_ANYOF_SYNTHETIC(and_with)
1253 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1255 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1256 * copy it over 'ssc' */
1257 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1258 if (is_ANYOF_SYNTHETIC(and_with)) {
1259 StructCopy(and_with, ssc, regnode_ssc);
1262 ssc->invlist = anded_cp_list;
1263 ANYOF_POSIXL_ZERO(ssc);
1264 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1265 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1269 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1270 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1272 /* One or the other of P1, P2 is non-empty. */
1273 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1274 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1276 ssc_union(ssc, anded_cp_list, FALSE);
1278 else { /* P1 = P2 = empty */
1279 ssc_intersection(ssc, anded_cp_list, FALSE);
1285 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1286 const regnode_charclass *or_with)
1288 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1289 * another SSC or a regular ANYOF class. Can create false positives if
1290 * 'or_with' is to be inverted. */
1295 PERL_ARGS_ASSERT_SSC_OR;
1297 assert(is_ANYOF_SYNTHETIC(ssc));
1299 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1300 * the code point inversion list and just the relevant flags */
1301 if (is_ANYOF_SYNTHETIC(or_with)) {
1302 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1303 ored_flags = ANYOF_FLAGS(or_with);
1306 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1307 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1310 ANYOF_FLAGS(ssc) |= ored_flags;
1312 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1313 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1314 * 'or_with' may be inverted. When not inverted, we have the simple
1315 * situation of computing:
1316 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1317 * If P1|P2 yields a situation with both a class and its complement are
1318 * set, like having both \w and \W, this matches all code points, and we
1319 * can delete these from the P component of the ssc going forward. XXX We
1320 * might be able to delete all the P components, but I (khw) am not certain
1321 * about this, and it is better to be safe.
1324 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1325 * <= (C1 | P1) | ~C2
1326 * <= (C1 | ~C2) | P1
1327 * (which results in actually simpler code than the non-inverted case)
1330 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1331 && ! is_ANYOF_SYNTHETIC(or_with))
1333 /* We ignore P2, leaving P1 going forward */
1334 } /* else Not inverted */
1335 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
1336 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1337 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1339 for (i = 0; i < ANYOF_MAX; i += 2) {
1340 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1342 ssc_match_all_cp(ssc);
1343 ANYOF_POSIXL_CLEAR(ssc, i);
1344 ANYOF_POSIXL_CLEAR(ssc, i+1);
1352 FALSE /* Already has been inverted */
1356 PERL_STATIC_INLINE void
1357 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1359 PERL_ARGS_ASSERT_SSC_UNION;
1361 assert(is_ANYOF_SYNTHETIC(ssc));
1363 _invlist_union_maybe_complement_2nd(ssc->invlist,
1369 PERL_STATIC_INLINE void
1370 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1372 const bool invert2nd)
1374 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1376 assert(is_ANYOF_SYNTHETIC(ssc));
1378 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1384 PERL_STATIC_INLINE void
1385 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1387 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1389 assert(is_ANYOF_SYNTHETIC(ssc));
1391 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1394 PERL_STATIC_INLINE void
1395 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1397 /* AND just the single code point 'cp' into the SSC 'ssc' */
1399 SV* cp_list = _new_invlist(2);
1401 PERL_ARGS_ASSERT_SSC_CP_AND;
1403 assert(is_ANYOF_SYNTHETIC(ssc));
1405 cp_list = add_cp_to_invlist(cp_list, cp);
1406 ssc_intersection(ssc, cp_list,
1407 FALSE /* Not inverted */
1409 SvREFCNT_dec_NN(cp_list);
1412 PERL_STATIC_INLINE void
1413 S_ssc_clear_locale(pTHX_ regnode_ssc *ssc)
1415 /* Set the SSC 'ssc' to not match any locale things */
1417 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1419 assert(is_ANYOF_SYNTHETIC(ssc));
1421 ANYOF_POSIXL_ZERO(ssc);
1422 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1426 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1428 /* The inversion list in the SSC is marked mortal; now we need a more
1429 * permanent copy, which is stored the same way that is done in a regular
1430 * ANYOF node, with the first 256 code points in a bit map */
1432 SV* invlist = invlist_clone(ssc->invlist);
1434 PERL_ARGS_ASSERT_SSC_FINALIZE;
1436 assert(is_ANYOF_SYNTHETIC(ssc));
1438 /* The code in this file assumes that all but these flags aren't relevant
1439 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1440 * time we reach here */
1441 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1443 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1445 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1446 NULL, NULL, NULL, FALSE);
1448 /* Make sure is clone-safe */
1449 ssc->invlist = NULL;
1451 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1452 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1455 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1458 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1459 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1460 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1461 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1462 ? (TRIE_LIST_CUR( idx ) - 1) \
1468 dump_trie(trie,widecharmap,revcharmap)
1469 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1470 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1472 These routines dump out a trie in a somewhat readable format.
1473 The _interim_ variants are used for debugging the interim
1474 tables that are used to generate the final compressed
1475 representation which is what dump_trie expects.
1477 Part of the reason for their existence is to provide a form
1478 of documentation as to how the different representations function.
1483 Dumps the final compressed table form of the trie to Perl_debug_log.
1484 Used for debugging make_trie().
1488 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1489 AV *revcharmap, U32 depth)
1492 SV *sv=sv_newmortal();
1493 int colwidth= widecharmap ? 6 : 4;
1495 GET_RE_DEBUG_FLAGS_DECL;
1497 PERL_ARGS_ASSERT_DUMP_TRIE;
1499 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1500 (int)depth * 2 + 2,"",
1501 "Match","Base","Ofs" );
1503 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1504 SV ** const tmp = av_fetch( revcharmap, state, 0);
1506 PerlIO_printf( Perl_debug_log, "%*s",
1508 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1509 PL_colors[0], PL_colors[1],
1510 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1511 PERL_PV_ESCAPE_FIRSTCHAR
1516 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1517 (int)depth * 2 + 2,"");
1519 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1520 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1521 PerlIO_printf( Perl_debug_log, "\n");
1523 for( state = 1 ; state < trie->statecount ; state++ ) {
1524 const U32 base = trie->states[ state ].trans.base;
1526 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1527 (int)depth * 2 + 2,"", (UV)state);
1529 if ( trie->states[ state ].wordnum ) {
1530 PerlIO_printf( Perl_debug_log, " W%4X",
1531 trie->states[ state ].wordnum );
1533 PerlIO_printf( Perl_debug_log, "%6s", "" );
1536 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1541 while( ( base + ofs < trie->uniquecharcount ) ||
1542 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1543 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1547 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1549 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1550 if ( ( base + ofs >= trie->uniquecharcount )
1551 && ( base + ofs - trie->uniquecharcount
1553 && trie->trans[ base + ofs
1554 - trie->uniquecharcount ].check == state )
1556 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1558 (UV)trie->trans[ base + ofs
1559 - trie->uniquecharcount ].next );
1561 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1565 PerlIO_printf( Perl_debug_log, "]");
1568 PerlIO_printf( Perl_debug_log, "\n" );
1570 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1572 for (word=1; word <= trie->wordcount; word++) {
1573 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1574 (int)word, (int)(trie->wordinfo[word].prev),
1575 (int)(trie->wordinfo[word].len));
1577 PerlIO_printf(Perl_debug_log, "\n" );
1580 Dumps a fully constructed but uncompressed trie in list form.
1581 List tries normally only are used for construction when the number of
1582 possible chars (trie->uniquecharcount) is very high.
1583 Used for debugging make_trie().
1586 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1587 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1591 SV *sv=sv_newmortal();
1592 int colwidth= widecharmap ? 6 : 4;
1593 GET_RE_DEBUG_FLAGS_DECL;
1595 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1597 /* print out the table precompression. */
1598 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1599 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1600 "------:-----+-----------------\n" );
1602 for( state=1 ; state < next_alloc ; state ++ ) {
1605 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1606 (int)depth * 2 + 2,"", (UV)state );
1607 if ( ! trie->states[ state ].wordnum ) {
1608 PerlIO_printf( Perl_debug_log, "%5s| ","");
1610 PerlIO_printf( Perl_debug_log, "W%4x| ",
1611 trie->states[ state ].wordnum
1614 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1615 SV ** const tmp = av_fetch( revcharmap,
1616 TRIE_LIST_ITEM(state,charid).forid, 0);
1618 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1620 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1622 PL_colors[0], PL_colors[1],
1623 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1624 | PERL_PV_ESCAPE_FIRSTCHAR
1626 TRIE_LIST_ITEM(state,charid).forid,
1627 (UV)TRIE_LIST_ITEM(state,charid).newstate
1630 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1631 (int)((depth * 2) + 14), "");
1634 PerlIO_printf( Perl_debug_log, "\n");
1639 Dumps a fully constructed but uncompressed trie in table form.
1640 This is the normal DFA style state transition table, with a few
1641 twists to facilitate compression later.
1642 Used for debugging make_trie().
1645 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1646 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1651 SV *sv=sv_newmortal();
1652 int colwidth= widecharmap ? 6 : 4;
1653 GET_RE_DEBUG_FLAGS_DECL;
1655 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1658 print out the table precompression so that we can do a visual check
1659 that they are identical.
1662 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1664 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1665 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1667 PerlIO_printf( Perl_debug_log, "%*s",
1669 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1670 PL_colors[0], PL_colors[1],
1671 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1672 PERL_PV_ESCAPE_FIRSTCHAR
1678 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1680 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1681 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1684 PerlIO_printf( Perl_debug_log, "\n" );
1686 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1688 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1689 (int)depth * 2 + 2,"",
1690 (UV)TRIE_NODENUM( state ) );
1692 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1693 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1695 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1697 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1699 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1700 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1701 (UV)trie->trans[ state ].check );
1703 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1704 (UV)trie->trans[ state ].check,
1705 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1713 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1714 startbranch: the first branch in the whole branch sequence
1715 first : start branch of sequence of branch-exact nodes.
1716 May be the same as startbranch
1717 last : Thing following the last branch.
1718 May be the same as tail.
1719 tail : item following the branch sequence
1720 count : words in the sequence
1721 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1722 depth : indent depth
1724 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1726 A trie is an N'ary tree where the branches are determined by digital
1727 decomposition of the key. IE, at the root node you look up the 1st character and
1728 follow that branch repeat until you find the end of the branches. Nodes can be
1729 marked as "accepting" meaning they represent a complete word. Eg:
1733 would convert into the following structure. Numbers represent states, letters
1734 following numbers represent valid transitions on the letter from that state, if
1735 the number is in square brackets it represents an accepting state, otherwise it
1736 will be in parenthesis.
1738 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1742 (1) +-i->(6)-+-s->[7]
1744 +-s->(3)-+-h->(4)-+-e->[5]
1746 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1748 This shows that when matching against the string 'hers' we will begin at state 1
1749 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1750 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1751 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1752 single traverse. We store a mapping from accepting to state to which word was
1753 matched, and then when we have multiple possibilities we try to complete the
1754 rest of the regex in the order in which they occured in the alternation.
1756 The only prior NFA like behaviour that would be changed by the TRIE support is
1757 the silent ignoring of duplicate alternations which are of the form:
1759 / (DUPE|DUPE) X? (?{ ... }) Y /x
1761 Thus EVAL blocks following a trie may be called a different number of times with
1762 and without the optimisation. With the optimisations dupes will be silently
1763 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1764 the following demonstrates:
1766 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1768 which prints out 'word' three times, but
1770 'words'=~/(word|word|word)(?{ print $1 })S/
1772 which doesnt print it out at all. This is due to other optimisations kicking in.
1774 Example of what happens on a structural level:
1776 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1778 1: CURLYM[1] {1,32767}(18)
1789 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1790 and should turn into:
1792 1: CURLYM[1] {1,32767}(18)
1794 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1802 Cases where tail != last would be like /(?foo|bar)baz/:
1812 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1813 and would end up looking like:
1816 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1823 d = uvchr_to_utf8_flags(d, uv, 0);
1825 is the recommended Unicode-aware way of saying
1830 #define TRIE_STORE_REVCHAR(val) \
1833 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1834 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1835 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1836 SvCUR_set(zlopp, kapow - flrbbbbb); \
1839 av_push(revcharmap, zlopp); \
1841 char ooooff = (char)val; \
1842 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1846 /* This gets the next character from the input, folding it if not already
1848 #define TRIE_READ_CHAR STMT_START { \
1851 /* if it is UTF then it is either already folded, or does not need \
1853 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1855 else if (folder == PL_fold_latin1) { \
1856 /* This folder implies Unicode rules, which in the range expressible \
1857 * by not UTF is the lower case, with the two exceptions, one of \
1858 * which should have been taken care of before calling this */ \
1859 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1860 uvc = toLOWER_L1(*uc); \
1861 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1864 /* raw data, will be folded later if needed */ \
1872 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1873 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1874 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1875 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1877 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1878 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1879 TRIE_LIST_CUR( state )++; \
1882 #define TRIE_LIST_NEW(state) STMT_START { \
1883 Newxz( trie->states[ state ].trans.list, \
1884 4, reg_trie_trans_le ); \
1885 TRIE_LIST_CUR( state ) = 1; \
1886 TRIE_LIST_LEN( state ) = 4; \
1889 #define TRIE_HANDLE_WORD(state) STMT_START { \
1890 U16 dupe= trie->states[ state ].wordnum; \
1891 regnode * const noper_next = regnext( noper ); \
1894 /* store the word for dumping */ \
1896 if (OP(noper) != NOTHING) \
1897 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1899 tmp = newSVpvn_utf8( "", 0, UTF ); \
1900 av_push( trie_words, tmp ); \
1904 trie->wordinfo[curword].prev = 0; \
1905 trie->wordinfo[curword].len = wordlen; \
1906 trie->wordinfo[curword].accept = state; \
1908 if ( noper_next < tail ) { \
1910 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1912 trie->jump[curword] = (U16)(noper_next - convert); \
1914 jumper = noper_next; \
1916 nextbranch= regnext(cur); \
1920 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1921 /* chain, so that when the bits of chain are later */\
1922 /* linked together, the dups appear in the chain */\
1923 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1924 trie->wordinfo[dupe].prev = curword; \
1926 /* we haven't inserted this word yet. */ \
1927 trie->states[ state ].wordnum = curword; \
1932 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1933 ( ( base + charid >= ucharcount \
1934 && base + charid < ubound \
1935 && state == trie->trans[ base - ucharcount + charid ].check \
1936 && trie->trans[ base - ucharcount + charid ].next ) \
1937 ? trie->trans[ base - ucharcount + charid ].next \
1938 : ( state==1 ? special : 0 ) \
1942 #define MADE_JUMP_TRIE 2
1943 #define MADE_EXACT_TRIE 4
1946 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1947 regnode *first, regnode *last, regnode *tail,
1948 U32 word_count, U32 flags, U32 depth)
1951 /* first pass, loop through and scan words */
1952 reg_trie_data *trie;
1953 HV *widecharmap = NULL;
1954 AV *revcharmap = newAV();
1960 regnode *jumper = NULL;
1961 regnode *nextbranch = NULL;
1962 regnode *convert = NULL;
1963 U32 *prev_states; /* temp array mapping each state to previous one */
1964 /* we just use folder as a flag in utf8 */
1965 const U8 * folder = NULL;
1968 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1969 AV *trie_words = NULL;
1970 /* along with revcharmap, this only used during construction but both are
1971 * useful during debugging so we store them in the struct when debugging.
1974 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1975 STRLEN trie_charcount=0;
1977 SV *re_trie_maxbuff;
1978 GET_RE_DEBUG_FLAGS_DECL;
1980 PERL_ARGS_ASSERT_MAKE_TRIE;
1982 PERL_UNUSED_ARG(depth);
1989 case EXACTFU: folder = PL_fold_latin1; break;
1990 case EXACTF: folder = PL_fold; break;
1991 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1994 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1996 trie->startstate = 1;
1997 trie->wordcount = word_count;
1998 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1999 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2001 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2002 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2003 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2006 trie_words = newAV();
2009 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2010 if (!SvIOK(re_trie_maxbuff)) {
2011 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2013 DEBUG_TRIE_COMPILE_r({
2014 PerlIO_printf( Perl_debug_log,
2015 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2016 (int)depth * 2 + 2, "",
2017 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2018 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2021 /* Find the node we are going to overwrite */
2022 if ( first == startbranch && OP( last ) != BRANCH ) {
2023 /* whole branch chain */
2026 /* branch sub-chain */
2027 convert = NEXTOPER( first );
2030 /* -- First loop and Setup --
2032 We first traverse the branches and scan each word to determine if it
2033 contains widechars, and how many unique chars there are, this is
2034 important as we have to build a table with at least as many columns as we
2037 We use an array of integers to represent the character codes 0..255
2038 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2039 the native representation of the character value as the key and IV's for
2042 *TODO* If we keep track of how many times each character is used we can
2043 remap the columns so that the table compression later on is more
2044 efficient in terms of memory by ensuring the most common value is in the
2045 middle and the least common are on the outside. IMO this would be better
2046 than a most to least common mapping as theres a decent chance the most
2047 common letter will share a node with the least common, meaning the node
2048 will not be compressible. With a middle is most common approach the worst
2049 case is when we have the least common nodes twice.
2053 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2054 regnode *noper = NEXTOPER( cur );
2055 const U8 *uc = (U8*)STRING( noper );
2056 const U8 *e = uc + STR_LEN( noper );
2058 U32 wordlen = 0; /* required init */
2059 STRLEN minchars = 0;
2060 STRLEN maxchars = 0;
2061 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2064 if (OP(noper) == NOTHING) {
2065 regnode *noper_next= regnext(noper);
2066 if (noper_next != tail && OP(noper_next) == flags) {
2068 uc= (U8*)STRING(noper);
2069 e= uc + STR_LEN(noper);
2070 trie->minlen= STR_LEN(noper);
2077 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2078 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2079 regardless of encoding */
2080 if (OP( noper ) == EXACTFU_SS) {
2081 /* false positives are ok, so just set this */
2082 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2085 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2087 TRIE_CHARCOUNT(trie)++;
2090 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2091 * is in effect. Under /i, this character can match itself, or
2092 * anything that folds to it. If not under /i, it can match just
2093 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2094 * all fold to k, and all are single characters. But some folds
2095 * expand to more than one character, so for example LATIN SMALL
2096 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2097 * the string beginning at 'uc' is 'ffi', it could be matched by
2098 * three characters, or just by the one ligature character. (It
2099 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2100 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2101 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2102 * match.) The trie needs to know the minimum and maximum number
2103 * of characters that could match so that it can use size alone to
2104 * quickly reject many match attempts. The max is simple: it is
2105 * the number of folded characters in this branch (since a fold is
2106 * never shorter than what folds to it. */
2110 /* And the min is equal to the max if not under /i (indicated by
2111 * 'folder' being NULL), or there are no multi-character folds. If
2112 * there is a multi-character fold, the min is incremented just
2113 * once, for the character that folds to the sequence. Each
2114 * character in the sequence needs to be added to the list below of
2115 * characters in the trie, but we count only the first towards the
2116 * min number of characters needed. This is done through the
2117 * variable 'foldlen', which is returned by the macros that look
2118 * for these sequences as the number of bytes the sequence
2119 * occupies. Each time through the loop, we decrement 'foldlen' by
2120 * how many bytes the current char occupies. Only when it reaches
2121 * 0 do we increment 'minchars' or look for another multi-character
2123 if (folder == NULL) {
2126 else if (foldlen > 0) {
2127 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2132 /* See if *uc is the beginning of a multi-character fold. If
2133 * so, we decrement the length remaining to look at, to account
2134 * for the current character this iteration. (We can use 'uc'
2135 * instead of the fold returned by TRIE_READ_CHAR because for
2136 * non-UTF, the latin1_safe macro is smart enough to account
2137 * for all the unfolded characters, and because for UTF, the
2138 * string will already have been folded earlier in the
2139 * compilation process */
2141 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2142 foldlen -= UTF8SKIP(uc);
2145 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2150 /* The current character (and any potential folds) should be added
2151 * to the possible matching characters for this position in this
2155 U8 folded= folder[ (U8) uvc ];
2156 if ( !trie->charmap[ folded ] ) {
2157 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2158 TRIE_STORE_REVCHAR( folded );
2161 if ( !trie->charmap[ uvc ] ) {
2162 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2163 TRIE_STORE_REVCHAR( uvc );
2166 /* store the codepoint in the bitmap, and its folded
2168 TRIE_BITMAP_SET(trie, uvc);
2170 /* store the folded codepoint */
2171 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2174 /* store first byte of utf8 representation of
2175 variant codepoints */
2176 if (! UVCHR_IS_INVARIANT(uvc)) {
2177 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2180 set_bit = 0; /* We've done our bit :-) */
2184 /* XXX We could come up with the list of code points that fold
2185 * to this using PL_utf8_foldclosures, except not for
2186 * multi-char folds, as there may be multiple combinations
2187 * there that could work, which needs to wait until runtime to
2188 * resolve (The comment about LIGATURE FFI above is such an
2193 widecharmap = newHV();
2195 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2198 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2200 if ( !SvTRUE( *svpp ) ) {
2201 sv_setiv( *svpp, ++trie->uniquecharcount );
2202 TRIE_STORE_REVCHAR(uvc);
2205 } /* end loop through characters in this branch of the trie */
2207 /* We take the min and max for this branch and combine to find the min
2208 * and max for all branches processed so far */
2209 if( cur == first ) {
2210 trie->minlen = minchars;
2211 trie->maxlen = maxchars;
2212 } else if (minchars < trie->minlen) {
2213 trie->minlen = minchars;
2214 } else if (maxchars > trie->maxlen) {
2215 trie->maxlen = maxchars;
2217 } /* end first pass */
2218 DEBUG_TRIE_COMPILE_r(
2219 PerlIO_printf( Perl_debug_log,
2220 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2221 (int)depth * 2 + 2,"",
2222 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2223 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2224 (int)trie->minlen, (int)trie->maxlen )
2228 We now know what we are dealing with in terms of unique chars and
2229 string sizes so we can calculate how much memory a naive
2230 representation using a flat table will take. If it's over a reasonable
2231 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2232 conservative but potentially much slower representation using an array
2235 At the end we convert both representations into the same compressed
2236 form that will be used in regexec.c for matching with. The latter
2237 is a form that cannot be used to construct with but has memory
2238 properties similar to the list form and access properties similar
2239 to the table form making it both suitable for fast searches and
2240 small enough that its feasable to store for the duration of a program.
2242 See the comment in the code where the compressed table is produced
2243 inplace from the flat tabe representation for an explanation of how
2244 the compression works.
2249 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2252 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2253 > SvIV(re_trie_maxbuff) )
2256 Second Pass -- Array Of Lists Representation
2258 Each state will be represented by a list of charid:state records
2259 (reg_trie_trans_le) the first such element holds the CUR and LEN
2260 points of the allocated array. (See defines above).
2262 We build the initial structure using the lists, and then convert
2263 it into the compressed table form which allows faster lookups
2264 (but cant be modified once converted).
2267 STRLEN transcount = 1;
2269 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2270 "%*sCompiling trie using list compiler\n",
2271 (int)depth * 2 + 2, ""));
2273 trie->states = (reg_trie_state *)
2274 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2275 sizeof(reg_trie_state) );
2279 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2281 regnode *noper = NEXTOPER( cur );
2282 U8 *uc = (U8*)STRING( noper );
2283 const U8 *e = uc + STR_LEN( noper );
2284 U32 state = 1; /* required init */
2285 U16 charid = 0; /* sanity init */
2286 U32 wordlen = 0; /* required init */
2288 if (OP(noper) == NOTHING) {
2289 regnode *noper_next= regnext(noper);
2290 if (noper_next != tail && OP(noper_next) == flags) {
2292 uc= (U8*)STRING(noper);
2293 e= uc + STR_LEN(noper);
2297 if (OP(noper) != NOTHING) {
2298 for ( ; uc < e ; uc += len ) {
2303 charid = trie->charmap[ uvc ];
2305 SV** const svpp = hv_fetch( widecharmap,
2312 charid=(U16)SvIV( *svpp );
2315 /* charid is now 0 if we dont know the char read, or
2316 * nonzero if we do */
2323 if ( !trie->states[ state ].trans.list ) {
2324 TRIE_LIST_NEW( state );
2327 check <= TRIE_LIST_USED( state );
2330 if ( TRIE_LIST_ITEM( state, check ).forid
2333 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2338 newstate = next_alloc++;
2339 prev_states[newstate] = state;
2340 TRIE_LIST_PUSH( state, charid, newstate );
2345 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2349 TRIE_HANDLE_WORD(state);
2351 } /* end second pass */
2353 /* next alloc is the NEXT state to be allocated */
2354 trie->statecount = next_alloc;
2355 trie->states = (reg_trie_state *)
2356 PerlMemShared_realloc( trie->states,
2358 * sizeof(reg_trie_state) );
2360 /* and now dump it out before we compress it */
2361 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2362 revcharmap, next_alloc,
2366 trie->trans = (reg_trie_trans *)
2367 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2374 for( state=1 ; state < next_alloc ; state ++ ) {
2378 DEBUG_TRIE_COMPILE_MORE_r(
2379 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2383 if (trie->states[state].trans.list) {
2384 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2388 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2389 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2390 if ( forid < minid ) {
2392 } else if ( forid > maxid ) {
2396 if ( transcount < tp + maxid - minid + 1) {
2398 trie->trans = (reg_trie_trans *)
2399 PerlMemShared_realloc( trie->trans,
2401 * sizeof(reg_trie_trans) );
2402 Zero( trie->trans + (transcount / 2),
2406 base = trie->uniquecharcount + tp - minid;
2407 if ( maxid == minid ) {
2409 for ( ; zp < tp ; zp++ ) {
2410 if ( ! trie->trans[ zp ].next ) {
2411 base = trie->uniquecharcount + zp - minid;
2412 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2414 trie->trans[ zp ].check = state;
2420 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2422 trie->trans[ tp ].check = state;
2427 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2428 const U32 tid = base
2429 - trie->uniquecharcount
2430 + TRIE_LIST_ITEM( state, idx ).forid;
2431 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2433 trie->trans[ tid ].check = state;
2435 tp += ( maxid - minid + 1 );
2437 Safefree(trie->states[ state ].trans.list);
2440 DEBUG_TRIE_COMPILE_MORE_r(
2441 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2444 trie->states[ state ].trans.base=base;
2446 trie->lasttrans = tp + 1;
2450 Second Pass -- Flat Table Representation.
2452 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2453 each. We know that we will need Charcount+1 trans at most to store
2454 the data (one row per char at worst case) So we preallocate both
2455 structures assuming worst case.
2457 We then construct the trie using only the .next slots of the entry
2460 We use the .check field of the first entry of the node temporarily
2461 to make compression both faster and easier by keeping track of how
2462 many non zero fields are in the node.
2464 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2467 There are two terms at use here: state as a TRIE_NODEIDX() which is
2468 a number representing the first entry of the node, and state as a
2469 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2470 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2471 if there are 2 entrys per node. eg:
2479 The table is internally in the right hand, idx form. However as we
2480 also have to deal with the states array which is indexed by nodenum
2481 we have to use TRIE_NODENUM() to convert.
2484 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2485 "%*sCompiling trie using table compiler\n",
2486 (int)depth * 2 + 2, ""));
2488 trie->trans = (reg_trie_trans *)
2489 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2490 * trie->uniquecharcount + 1,
2491 sizeof(reg_trie_trans) );
2492 trie->states = (reg_trie_state *)
2493 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2494 sizeof(reg_trie_state) );
2495 next_alloc = trie->uniquecharcount + 1;
2498 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2500 regnode *noper = NEXTOPER( cur );
2501 const U8 *uc = (U8*)STRING( noper );
2502 const U8 *e = uc + STR_LEN( noper );
2504 U32 state = 1; /* required init */
2506 U16 charid = 0; /* sanity init */
2507 U32 accept_state = 0; /* sanity init */
2509 U32 wordlen = 0; /* required init */
2511 if (OP(noper) == NOTHING) {
2512 regnode *noper_next= regnext(noper);
2513 if (noper_next != tail && OP(noper_next) == flags) {
2515 uc= (U8*)STRING(noper);
2516 e= uc + STR_LEN(noper);
2520 if ( OP(noper) != NOTHING ) {
2521 for ( ; uc < e ; uc += len ) {
2526 charid = trie->charmap[ uvc ];
2528 SV* const * const svpp = hv_fetch( widecharmap,
2532 charid = svpp ? (U16)SvIV(*svpp) : 0;
2536 if ( !trie->trans[ state + charid ].next ) {
2537 trie->trans[ state + charid ].next = next_alloc;
2538 trie->trans[ state ].check++;
2539 prev_states[TRIE_NODENUM(next_alloc)]
2540 = TRIE_NODENUM(state);
2541 next_alloc += trie->uniquecharcount;
2543 state = trie->trans[ state + charid ].next;
2545 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2547 /* charid is now 0 if we dont know the char read, or
2548 * nonzero if we do */
2551 accept_state = TRIE_NODENUM( state );
2552 TRIE_HANDLE_WORD(accept_state);
2554 } /* end second pass */
2556 /* and now dump it out before we compress it */
2557 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2559 next_alloc, depth+1));
2563 * Inplace compress the table.*
2565 For sparse data sets the table constructed by the trie algorithm will
2566 be mostly 0/FAIL transitions or to put it another way mostly empty.
2567 (Note that leaf nodes will not contain any transitions.)
2569 This algorithm compresses the tables by eliminating most such
2570 transitions, at the cost of a modest bit of extra work during lookup:
2572 - Each states[] entry contains a .base field which indicates the
2573 index in the state[] array wheres its transition data is stored.
2575 - If .base is 0 there are no valid transitions from that node.
2577 - If .base is nonzero then charid is added to it to find an entry in
2580 -If trans[states[state].base+charid].check!=state then the
2581 transition is taken to be a 0/Fail transition. Thus if there are fail
2582 transitions at the front of the node then the .base offset will point
2583 somewhere inside the previous nodes data (or maybe even into a node
2584 even earlier), but the .check field determines if the transition is
2588 The following process inplace converts the table to the compressed
2589 table: We first do not compress the root node 1,and mark all its
2590 .check pointers as 1 and set its .base pointer as 1 as well. This
2591 allows us to do a DFA construction from the compressed table later,
2592 and ensures that any .base pointers we calculate later are greater
2595 - We set 'pos' to indicate the first entry of the second node.
2597 - We then iterate over the columns of the node, finding the first and
2598 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2599 and set the .check pointers accordingly, and advance pos
2600 appropriately and repreat for the next node. Note that when we copy
2601 the next pointers we have to convert them from the original
2602 NODEIDX form to NODENUM form as the former is not valid post
2605 - If a node has no transitions used we mark its base as 0 and do not
2606 advance the pos pointer.
2608 - If a node only has one transition we use a second pointer into the
2609 structure to fill in allocated fail transitions from other states.
2610 This pointer is independent of the main pointer and scans forward
2611 looking for null transitions that are allocated to a state. When it
2612 finds one it writes the single transition into the "hole". If the
2613 pointer doesnt find one the single transition is appended as normal.
2615 - Once compressed we can Renew/realloc the structures to release the
2618 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2619 specifically Fig 3.47 and the associated pseudocode.
2623 const U32 laststate = TRIE_NODENUM( next_alloc );
2626 trie->statecount = laststate;
2628 for ( state = 1 ; state < laststate ; state++ ) {
2630 const U32 stateidx = TRIE_NODEIDX( state );
2631 const U32 o_used = trie->trans[ stateidx ].check;
2632 U32 used = trie->trans[ stateidx ].check;
2633 trie->trans[ stateidx ].check = 0;
2636 used && charid < trie->uniquecharcount;
2639 if ( flag || trie->trans[ stateidx + charid ].next ) {
2640 if ( trie->trans[ stateidx + charid ].next ) {
2642 for ( ; zp < pos ; zp++ ) {
2643 if ( ! trie->trans[ zp ].next ) {
2647 trie->states[ state ].trans.base
2649 + trie->uniquecharcount
2651 trie->trans[ zp ].next
2652 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2654 trie->trans[ zp ].check = state;
2655 if ( ++zp > pos ) pos = zp;
2662 trie->states[ state ].trans.base
2663 = pos + trie->uniquecharcount - charid ;
2665 trie->trans[ pos ].next
2666 = SAFE_TRIE_NODENUM(
2667 trie->trans[ stateidx + charid ].next );
2668 trie->trans[ pos ].check = state;
2673 trie->lasttrans = pos + 1;
2674 trie->states = (reg_trie_state *)
2675 PerlMemShared_realloc( trie->states, laststate
2676 * sizeof(reg_trie_state) );
2677 DEBUG_TRIE_COMPILE_MORE_r(
2678 PerlIO_printf( Perl_debug_log,
2679 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2680 (int)depth * 2 + 2,"",
2681 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2685 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2688 } /* end table compress */
2690 DEBUG_TRIE_COMPILE_MORE_r(
2691 PerlIO_printf(Perl_debug_log,
2692 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2693 (int)depth * 2 + 2, "",
2694 (UV)trie->statecount,
2695 (UV)trie->lasttrans)
2697 /* resize the trans array to remove unused space */
2698 trie->trans = (reg_trie_trans *)
2699 PerlMemShared_realloc( trie->trans, trie->lasttrans
2700 * sizeof(reg_trie_trans) );
2702 { /* Modify the program and insert the new TRIE node */
2703 U8 nodetype =(U8)(flags & 0xFF);
2707 regnode *optimize = NULL;
2708 #ifdef RE_TRACK_PATTERN_OFFSETS
2711 U32 mjd_nodelen = 0;
2712 #endif /* RE_TRACK_PATTERN_OFFSETS */
2713 #endif /* DEBUGGING */
2715 This means we convert either the first branch or the first Exact,
2716 depending on whether the thing following (in 'last') is a branch
2717 or not and whther first is the startbranch (ie is it a sub part of
2718 the alternation or is it the whole thing.)
2719 Assuming its a sub part we convert the EXACT otherwise we convert
2720 the whole branch sequence, including the first.
2722 /* Find the node we are going to overwrite */
2723 if ( first != startbranch || OP( last ) == BRANCH ) {
2724 /* branch sub-chain */
2725 NEXT_OFF( first ) = (U16)(last - first);
2726 #ifdef RE_TRACK_PATTERN_OFFSETS
2728 mjd_offset= Node_Offset((convert));
2729 mjd_nodelen= Node_Length((convert));
2732 /* whole branch chain */
2734 #ifdef RE_TRACK_PATTERN_OFFSETS
2737 const regnode *nop = NEXTOPER( convert );
2738 mjd_offset= Node_Offset((nop));
2739 mjd_nodelen= Node_Length((nop));
2743 PerlIO_printf(Perl_debug_log,
2744 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2745 (int)depth * 2 + 2, "",
2746 (UV)mjd_offset, (UV)mjd_nodelen)
2749 /* But first we check to see if there is a common prefix we can
2750 split out as an EXACT and put in front of the TRIE node. */
2751 trie->startstate= 1;
2752 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2754 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2758 const U32 base = trie->states[ state ].trans.base;
2760 if ( trie->states[state].wordnum )
2763 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2764 if ( ( base + ofs >= trie->uniquecharcount ) &&
2765 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2766 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2768 if ( ++count > 1 ) {
2769 SV **tmp = av_fetch( revcharmap, ofs, 0);
2770 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2771 if ( state == 1 ) break;
2773 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2775 PerlIO_printf(Perl_debug_log,
2776 "%*sNew Start State=%"UVuf" Class: [",
2777 (int)depth * 2 + 2, "",
2780 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2781 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2783 TRIE_BITMAP_SET(trie,*ch);
2785 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2787 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2791 TRIE_BITMAP_SET(trie,*ch);
2793 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2794 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2800 SV **tmp = av_fetch( revcharmap, idx, 0);
2802 char *ch = SvPV( *tmp, len );
2804 SV *sv=sv_newmortal();
2805 PerlIO_printf( Perl_debug_log,
2806 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2807 (int)depth * 2 + 2, "",
2809 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2810 PL_colors[0], PL_colors[1],
2811 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2812 PERL_PV_ESCAPE_FIRSTCHAR
2817 OP( convert ) = nodetype;
2818 str=STRING(convert);
2821 STR_LEN(convert) += len;
2827 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2832 trie->prefixlen = (state-1);
2834 regnode *n = convert+NODE_SZ_STR(convert);
2835 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2836 trie->startstate = state;
2837 trie->minlen -= (state - 1);
2838 trie->maxlen -= (state - 1);
2840 /* At least the UNICOS C compiler choked on this
2841 * being argument to DEBUG_r(), so let's just have
2844 #ifdef PERL_EXT_RE_BUILD
2850 regnode *fix = convert;
2851 U32 word = trie->wordcount;
2853 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2854 while( ++fix < n ) {
2855 Set_Node_Offset_Length(fix, 0, 0);
2858 SV ** const tmp = av_fetch( trie_words, word, 0 );
2860 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2861 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2863 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2871 NEXT_OFF(convert) = (U16)(tail - convert);
2872 DEBUG_r(optimize= n);
2878 if ( trie->maxlen ) {
2879 NEXT_OFF( convert ) = (U16)(tail - convert);
2880 ARG_SET( convert, data_slot );
2881 /* Store the offset to the first unabsorbed branch in
2882 jump[0], which is otherwise unused by the jump logic.
2883 We use this when dumping a trie and during optimisation. */
2885 trie->jump[0] = (U16)(nextbranch - convert);
2887 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2888 * and there is a bitmap
2889 * and the first "jump target" node we found leaves enough room
2890 * then convert the TRIE node into a TRIEC node, with the bitmap
2891 * embedded inline in the opcode - this is hypothetically faster.
2893 if ( !trie->states[trie->startstate].wordnum
2895 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2897 OP( convert ) = TRIEC;
2898 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2899 PerlMemShared_free(trie->bitmap);
2902 OP( convert ) = TRIE;
2904 /* store the type in the flags */
2905 convert->flags = nodetype;
2909 + regarglen[ OP( convert ) ];
2911 /* XXX We really should free up the resource in trie now,
2912 as we won't use them - (which resources?) dmq */
2914 /* needed for dumping*/
2915 DEBUG_r(if (optimize) {
2916 regnode *opt = convert;
2918 while ( ++opt < optimize) {
2919 Set_Node_Offset_Length(opt,0,0);
2922 Try to clean up some of the debris left after the
2925 while( optimize < jumper ) {
2926 mjd_nodelen += Node_Length((optimize));
2927 OP( optimize ) = OPTIMIZED;
2928 Set_Node_Offset_Length(optimize,0,0);
2931 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2933 } /* end node insert */
2934 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
2936 /* Finish populating the prev field of the wordinfo array. Walk back
2937 * from each accept state until we find another accept state, and if
2938 * so, point the first word's .prev field at the second word. If the
2939 * second already has a .prev field set, stop now. This will be the
2940 * case either if we've already processed that word's accept state,
2941 * or that state had multiple words, and the overspill words were
2942 * already linked up earlier.
2949 for (word=1; word <= trie->wordcount; word++) {
2951 if (trie->wordinfo[word].prev)
2953 state = trie->wordinfo[word].accept;
2955 state = prev_states[state];
2958 prev = trie->states[state].wordnum;
2962 trie->wordinfo[word].prev = prev;
2964 Safefree(prev_states);
2968 /* and now dump out the compressed format */
2969 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2971 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2973 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2974 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2976 SvREFCNT_dec_NN(revcharmap);
2980 : trie->startstate>1
2986 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2988 /* The Trie is constructed and compressed now so we can build a fail array if
2991 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2993 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2997 We find the fail state for each state in the trie, this state is the longest
2998 proper suffix of the current state's 'word' that is also a proper prefix of
2999 another word in our trie. State 1 represents the word '' and is thus the
3000 default fail state. This allows the DFA not to have to restart after its
3001 tried and failed a word at a given point, it simply continues as though it
3002 had been matching the other word in the first place.
3004 'abcdgu'=~/abcdefg|cdgu/
3005 When we get to 'd' we are still matching the first word, we would encounter
3006 'g' which would fail, which would bring us to the state representing 'd' in
3007 the second word where we would try 'g' and succeed, proceeding to match
3010 /* add a fail transition */
3011 const U32 trie_offset = ARG(source);
3012 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3014 const U32 ucharcount = trie->uniquecharcount;
3015 const U32 numstates = trie->statecount;
3016 const U32 ubound = trie->lasttrans + ucharcount;
3020 U32 base = trie->states[ 1 ].trans.base;
3023 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3024 GET_RE_DEBUG_FLAGS_DECL;
3026 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
3028 PERL_UNUSED_ARG(depth);
3032 ARG_SET( stclass, data_slot );
3033 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3034 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3035 aho->trie=trie_offset;
3036 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3037 Copy( trie->states, aho->states, numstates, reg_trie_state );
3038 Newxz( q, numstates, U32);
3039 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3042 /* initialize fail[0..1] to be 1 so that we always have
3043 a valid final fail state */
3044 fail[ 0 ] = fail[ 1 ] = 1;
3046 for ( charid = 0; charid < ucharcount ; charid++ ) {
3047 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3049 q[ q_write ] = newstate;
3050 /* set to point at the root */
3051 fail[ q[ q_write++ ] ]=1;
3054 while ( q_read < q_write) {
3055 const U32 cur = q[ q_read++ % numstates ];
3056 base = trie->states[ cur ].trans.base;
3058 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3059 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3061 U32 fail_state = cur;
3064 fail_state = fail[ fail_state ];
3065 fail_base = aho->states[ fail_state ].trans.base;
3066 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3068 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3069 fail[ ch_state ] = fail_state;
3070 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3072 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3074 q[ q_write++ % numstates] = ch_state;
3078 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3079 when we fail in state 1, this allows us to use the
3080 charclass scan to find a valid start char. This is based on the principle
3081 that theres a good chance the string being searched contains lots of stuff
3082 that cant be a start char.
3084 fail[ 0 ] = fail[ 1 ] = 0;
3085 DEBUG_TRIE_COMPILE_r({
3086 PerlIO_printf(Perl_debug_log,
3087 "%*sStclass Failtable (%"UVuf" states): 0",
3088 (int)(depth * 2), "", (UV)numstates
3090 for( q_read=1; q_read<numstates; q_read++ ) {
3091 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3093 PerlIO_printf(Perl_debug_log, "\n");
3096 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3100 #define DEBUG_PEEP(str,scan,depth) \
3101 DEBUG_OPTIMISE_r({if (scan){ \
3102 SV * const mysv=sv_newmortal(); \
3103 regnode *Next = regnext(scan); \
3104 regprop(RExC_rx, mysv, scan, NULL); \
3105 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3106 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3107 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3111 /* The below joins as many adjacent EXACTish nodes as possible into a single
3112 * one. The regop may be changed if the node(s) contain certain sequences that
3113 * require special handling. The joining is only done if:
3114 * 1) there is room in the current conglomerated node to entirely contain the
3116 * 2) they are the exact same node type
3118 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3119 * these get optimized out
3121 * If a node is to match under /i (folded), the number of characters it matches
3122 * can be different than its character length if it contains a multi-character
3123 * fold. *min_subtract is set to the total delta number of characters of the
3126 * And *unfolded_multi_char is set to indicate whether or not the node contains
3127 * an unfolded multi-char fold. This happens when whether the fold is valid or
3128 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3129 * SMALL LETTER SHARP S, as only if the target string being matched against
3130 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3131 * folding rules depend on the locale in force at runtime. (Multi-char folds
3132 * whose components are all above the Latin1 range are not run-time locale
3133 * dependent, and have already been folded by the time this function is
3136 * This is as good a place as any to discuss the design of handling these
3137 * multi-character fold sequences. It's been wrong in Perl for a very long
3138 * time. There are three code points in Unicode whose multi-character folds
3139 * were long ago discovered to mess things up. The previous designs for
3140 * dealing with these involved assigning a special node for them. This
3141 * approach doesn't always work, as evidenced by this example:
3142 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3143 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3144 * would match just the \xDF, it won't be able to handle the case where a
3145 * successful match would have to cross the node's boundary. The new approach
3146 * that hopefully generally solves the problem generates an EXACTFU_SS node
3147 * that is "sss" in this case.
3149 * It turns out that there are problems with all multi-character folds, and not
3150 * just these three. Now the code is general, for all such cases. The
3151 * approach taken is:
3152 * 1) This routine examines each EXACTFish node that could contain multi-
3153 * character folded sequences. Since a single character can fold into
3154 * such a sequence, the minimum match length for this node is less than
3155 * the number of characters in the node. This routine returns in
3156 * *min_subtract how many characters to subtract from the the actual
3157 * length of the string to get a real minimum match length; it is 0 if
3158 * there are no multi-char foldeds. This delta is used by the caller to
3159 * adjust the min length of the match, and the delta between min and max,
3160 * so that the optimizer doesn't reject these possibilities based on size
3162 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3163 * is used for an EXACTFU node that contains at least one "ss" sequence in
3164 * it. For non-UTF-8 patterns and strings, this is the only case where
3165 * there is a possible fold length change. That means that a regular
3166 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3167 * with length changes, and so can be processed faster. regexec.c takes
3168 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3169 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3170 * known until runtime). This saves effort in regex matching. However,
3171 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3172 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3173 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3174 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3175 * possibilities for the non-UTF8 patterns are quite simple, except for
3176 * the sharp s. All the ones that don't involve a UTF-8 target string are
3177 * members of a fold-pair, and arrays are set up for all of them so that
3178 * the other member of the pair can be found quickly. Code elsewhere in
3179 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3180 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3181 * described in the next item.
3182 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3183 * validity of the fold won't be known until runtime, and so must remain
3184 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3185 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3186 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3187 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3188 * The reason this is a problem is that the optimizer part of regexec.c
3189 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3190 * that a character in the pattern corresponds to at most a single
3191 * character in the target string. (And I do mean character, and not byte
3192 * here, unlike other parts of the documentation that have never been
3193 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3194 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3195 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3196 * nodes, violate the assumption, and they are the only instances where it
3197 * is violated. I'm reluctant to try to change the assumption, as the
3198 * code involved is impenetrable to me (khw), so instead the code here
3199 * punts. This routine examines EXACTFL nodes, and (when the pattern
3200 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3201 * boolean indicating whether or not the node contains such a fold. When
3202 * it is true, the caller sets a flag that later causes the optimizer in
3203 * this file to not set values for the floating and fixed string lengths,
3204 * and thus avoids the optimizer code in regexec.c that makes the invalid
3205 * assumption. Thus, there is no optimization based on string lengths for
3206 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3207 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3208 * assumption is wrong only in these cases is that all other non-UTF-8
3209 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3210 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3211 * EXACTF nodes because we don't know at compile time if it actually
3212 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3213 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3214 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3215 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3216 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3217 * string would require the pattern to be forced into UTF-8, the overhead
3218 * of which we want to avoid. Similarly the unfolded multi-char folds in
3219 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3222 * Similarly, the code that generates tries doesn't currently handle
3223 * not-already-folded multi-char folds, and it looks like a pain to change
3224 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3225 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3226 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3227 * using /iaa matching will be doing so almost entirely with ASCII
3228 * strings, so this should rarely be encountered in practice */
3230 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3231 if (PL_regkind[OP(scan)] == EXACT) \
3232 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3235 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3236 UV *min_subtract, bool *unfolded_multi_char,
3237 U32 flags,regnode *val, U32 depth)
3239 /* Merge several consecutive EXACTish nodes into one. */
3240 regnode *n = regnext(scan);
3242 regnode *next = scan + NODE_SZ_STR(scan);
3246 regnode *stop = scan;
3247 GET_RE_DEBUG_FLAGS_DECL;
3249 PERL_UNUSED_ARG(depth);
3252 PERL_ARGS_ASSERT_JOIN_EXACT;
3253 #ifndef EXPERIMENTAL_INPLACESCAN
3254 PERL_UNUSED_ARG(flags);
3255 PERL_UNUSED_ARG(val);
3257 DEBUG_PEEP("join",scan,depth);
3259 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3260 * EXACT ones that are mergeable to the current one. */
3262 && (PL_regkind[OP(n)] == NOTHING
3263 || (stringok && OP(n) == OP(scan)))
3265 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3268 if (OP(n) == TAIL || n > next)
3270 if (PL_regkind[OP(n)] == NOTHING) {
3271 DEBUG_PEEP("skip:",n,depth);
3272 NEXT_OFF(scan) += NEXT_OFF(n);
3273 next = n + NODE_STEP_REGNODE;
3280 else if (stringok) {
3281 const unsigned int oldl = STR_LEN(scan);
3282 regnode * const nnext = regnext(n);
3284 /* XXX I (khw) kind of doubt that this works on platforms (should
3285 * Perl ever run on one) where U8_MAX is above 255 because of lots
3286 * of other assumptions */
3287 /* Don't join if the sum can't fit into a single node */
3288 if (oldl + STR_LEN(n) > U8_MAX)
3291 DEBUG_PEEP("merg",n,depth);
3294 NEXT_OFF(scan) += NEXT_OFF(n);
3295 STR_LEN(scan) += STR_LEN(n);
3296 next = n + NODE_SZ_STR(n);
3297 /* Now we can overwrite *n : */
3298 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3306 #ifdef EXPERIMENTAL_INPLACESCAN
3307 if (flags && !NEXT_OFF(n)) {
3308 DEBUG_PEEP("atch", val, depth);
3309 if (reg_off_by_arg[OP(n)]) {
3310 ARG_SET(n, val - n);
3313 NEXT_OFF(n) = val - n;
3321 *unfolded_multi_char = FALSE;
3323 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3324 * can now analyze for sequences of problematic code points. (Prior to
3325 * this final joining, sequences could have been split over boundaries, and
3326 * hence missed). The sequences only happen in folding, hence for any
3327 * non-EXACT EXACTish node */
3328 if (OP(scan) != EXACT) {
3329 U8* s0 = (U8*) STRING(scan);
3331 U8* s_end = s0 + STR_LEN(scan);
3333 int total_count_delta = 0; /* Total delta number of characters that
3334 multi-char folds expand to */
3336 /* One pass is made over the node's string looking for all the
3337 * possibilities. To avoid some tests in the loop, there are two main
3338 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3343 if (OP(scan) == EXACTFL) {
3346 /* An EXACTFL node would already have been changed to another
3347 * node type unless there is at least one character in it that
3348 * is problematic; likely a character whose fold definition
3349 * won't be known until runtime, and so has yet to be folded.
3350 * For all but the UTF-8 locale, folds are 1-1 in length, but
3351 * to handle the UTF-8 case, we need to create a temporary
3352 * folded copy using UTF-8 locale rules in order to analyze it.
3353 * This is because our macros that look to see if a sequence is
3354 * a multi-char fold assume everything is folded (otherwise the
3355 * tests in those macros would be too complicated and slow).
3356 * Note that here, the non-problematic folds will have already
3357 * been done, so we can just copy such characters. We actually
3358 * don't completely fold the EXACTFL string. We skip the
3359 * unfolded multi-char folds, as that would just create work
3360 * below to figure out the size they already are */
3362 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3365 STRLEN s_len = UTF8SKIP(s);
3366 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3367 Copy(s, d, s_len, U8);
3370 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3371 *unfolded_multi_char = TRUE;
3372 Copy(s, d, s_len, U8);
3375 else if (isASCII(*s)) {
3376 *(d++) = toFOLD(*s);
3380 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3386 /* Point the remainder of the routine to look at our temporary
3390 } /* End of creating folded copy of EXACTFL string */
3392 /* Examine the string for a multi-character fold sequence. UTF-8
3393 * patterns have all characters pre-folded by the time this code is
3395 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3396 length sequence we are looking for is 2 */
3398 int count = 0; /* How many characters in a multi-char fold */
3399 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3400 if (! len) { /* Not a multi-char fold: get next char */
3405 /* Nodes with 'ss' require special handling, except for
3406 * EXACTFA-ish for which there is no multi-char fold to this */
3407 if (len == 2 && *s == 's' && *(s+1) == 's'
3408 && OP(scan) != EXACTFA
3409 && OP(scan) != EXACTFA_NO_TRIE)
3412 if (OP(scan) != EXACTFL) {
3413 OP(scan) = EXACTFU_SS;
3417 else { /* Here is a generic multi-char fold. */
3418 U8* multi_end = s + len;
3420 /* Count how many characters in it. In the case of /aa, no
3421 * folds which contain ASCII code points are allowed, so
3422 * check for those, and skip if found. */
3423 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3424 count = utf8_length(s, multi_end);
3428 while (s < multi_end) {
3431 goto next_iteration;
3441 /* The delta is how long the sequence is minus 1 (1 is how long
3442 * the character that folds to the sequence is) */
3443 total_count_delta += count - 1;
3447 /* We created a temporary folded copy of the string in EXACTFL
3448 * nodes. Therefore we need to be sure it doesn't go below zero,
3449 * as the real string could be shorter */
3450 if (OP(scan) == EXACTFL) {
3451 int total_chars = utf8_length((U8*) STRING(scan),
3452 (U8*) STRING(scan) + STR_LEN(scan));
3453 if (total_count_delta > total_chars) {
3454 total_count_delta = total_chars;
3458 *min_subtract += total_count_delta;
3461 else if (OP(scan) == EXACTFA) {
3463 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3464 * fold to the ASCII range (and there are no existing ones in the
3465 * upper latin1 range). But, as outlined in the comments preceding
3466 * this function, we need to flag any occurrences of the sharp s.
3467 * This character forbids trie formation (because of added
3470 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3471 OP(scan) = EXACTFA_NO_TRIE;
3472 *unfolded_multi_char = TRUE;
3481 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3482 * folds that are all Latin1. As explained in the comments
3483 * preceding this function, we look also for the sharp s in EXACTF
3484 * and EXACTFL nodes; it can be in the final position. Otherwise
3485 * we can stop looking 1 byte earlier because have to find at least
3486 * two characters for a multi-fold */
3487 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3492 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3493 if (! len) { /* Not a multi-char fold. */
3494 if (*s == LATIN_SMALL_LETTER_SHARP_S
3495 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3497 *unfolded_multi_char = TRUE;
3504 && isARG2_lower_or_UPPER_ARG1('s', *s)
3505 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3508 /* EXACTF nodes need to know that the minimum length
3509 * changed so that a sharp s in the string can match this
3510 * ss in the pattern, but they remain EXACTF nodes, as they
3511 * won't match this unless the target string is is UTF-8,
3512 * which we don't know until runtime. EXACTFL nodes can't
3513 * transform into EXACTFU nodes */
3514 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3515 OP(scan) = EXACTFU_SS;
3519 *min_subtract += len - 1;
3526 /* Allow dumping but overwriting the collection of skipped
3527 * ops and/or strings with fake optimized ops */
3528 n = scan + NODE_SZ_STR(scan);
3536 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3540 /* REx optimizer. Converts nodes into quicker variants "in place".
3541 Finds fixed substrings. */
3543 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3544 to the position after last scanned or to NULL. */
3546 #define INIT_AND_WITHP \
3547 assert(!and_withp); \
3548 Newx(and_withp,1, regnode_ssc); \
3549 SAVEFREEPV(and_withp)
3551 /* this is a chain of data about sub patterns we are processing that
3552 need to be handled separately/specially in study_chunk. Its so
3553 we can simulate recursion without losing state. */
3555 typedef struct scan_frame {
3556 regnode *last; /* last node to process in this frame */
3557 regnode *next; /* next node to process when last is reached */
3558 struct scan_frame *prev; /*previous frame*/
3559 U32 prev_recursed_depth;
3560 I32 stop; /* what stopparen do we use */
3565 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3566 SSize_t *minlenp, SSize_t *deltap,
3571 regnode_ssc *and_withp,
3572 U32 flags, U32 depth)
3573 /* scanp: Start here (read-write). */
3574 /* deltap: Write maxlen-minlen here. */
3575 /* last: Stop before this one. */
3576 /* data: string data about the pattern */
3577 /* stopparen: treat close N as END */
3578 /* recursed: which subroutines have we recursed into */
3579 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3582 /* There must be at least this number of characters to match */
3585 regnode *scan = *scanp, *next;
3587 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3588 int is_inf_internal = 0; /* The studied chunk is infinite */
3589 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3590 scan_data_t data_fake;
3591 SV *re_trie_maxbuff = NULL;
3592 regnode *first_non_open = scan;
3593 SSize_t stopmin = SSize_t_MAX;
3594 scan_frame *frame = NULL;
3595 GET_RE_DEBUG_FLAGS_DECL;
3597 PERL_ARGS_ASSERT_STUDY_CHUNK;
3600 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3603 while (first_non_open && OP(first_non_open) == OPEN)
3604 first_non_open=regnext(first_non_open);
3609 while ( scan && OP(scan) != END && scan < last ){
3610 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3611 node length to get a real minimum (because
3612 the folded version may be shorter) */
3613 bool unfolded_multi_char = FALSE;
3614 /* Peephole optimizer: */
3615 DEBUG_OPTIMISE_MORE_r(
3617 PerlIO_printf(Perl_debug_log,
3618 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3619 ((int) depth*2), "", (long)stopparen,
3620 (unsigned long)depth, (unsigned long)recursed_depth);
3621 if (recursed_depth) {
3624 for ( j = 0 ; j < recursed_depth ; j++ ) {
3625 PerlIO_printf(Perl_debug_log,"[");
3626 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3627 PerlIO_printf(Perl_debug_log,"%d",
3628 PAREN_TEST(RExC_study_chunk_recursed +
3629 (j * RExC_study_chunk_recursed_bytes), i)
3632 PerlIO_printf(Perl_debug_log,"]");
3635 PerlIO_printf(Perl_debug_log,"\n");
3638 DEBUG_STUDYDATA("Peep:", data, depth);
3639 DEBUG_PEEP("Peep", scan, depth);
3642 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3643 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3644 * by a different invocation of reg() -- Yves
3646 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3648 /* Follow the next-chain of the current node and optimize
3649 away all the NOTHINGs from it. */
3650 if (OP(scan) != CURLYX) {
3651 const int max = (reg_off_by_arg[OP(scan)]
3653 /* I32 may be smaller than U16 on CRAYs! */
3654 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3655 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3659 /* Skip NOTHING and LONGJMP. */
3660 while ((n = regnext(n))
3661 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3662 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3663 && off + noff < max)
3665 if (reg_off_by_arg[OP(scan)])
3668 NEXT_OFF(scan) = off;
3673 /* The principal pseudo-switch. Cannot be a switch, since we
3674 look into several different things. */
3675 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3676 || OP(scan) == IFTHEN) {
3677 next = regnext(scan);
3679 /* demq: the op(next)==code check is to see if we have
3680 * "branch-branch" AFAICT */
3682 if (OP(next) == code || code == IFTHEN) {
3683 /* NOTE - There is similar code to this block below for
3684 * handling TRIE nodes on a re-study. If you change stuff here
3685 * check there too. */
3686 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3688 regnode * const startbranch=scan;
3690 if (flags & SCF_DO_SUBSTR) {
3691 /* Cannot merge strings after this. */
3692 scan_commit(pRExC_state, data, minlenp, is_inf);
3695 if (flags & SCF_DO_STCLASS)
3696 ssc_init_zero(pRExC_state, &accum);
3698 while (OP(scan) == code) {
3699 SSize_t deltanext, minnext, fake;
3701 regnode_ssc this_class;
3704 data_fake.flags = 0;
3706 data_fake.whilem_c = data->whilem_c;
3707 data_fake.last_closep = data->last_closep;
3710 data_fake.last_closep = &fake;
3712 data_fake.pos_delta = delta;
3713 next = regnext(scan);
3714 scan = NEXTOPER(scan);
3716 scan = NEXTOPER(scan);
3717 if (flags & SCF_DO_STCLASS) {
3718 ssc_init(pRExC_state, &this_class);
3719 data_fake.start_class = &this_class;
3720 f = SCF_DO_STCLASS_AND;
3722 if (flags & SCF_WHILEM_VISITED_POS)
3723 f |= SCF_WHILEM_VISITED_POS;
3725 /* we suppose the run is continuous, last=next...*/
3726 minnext = study_chunk(pRExC_state, &scan, minlenp,
3727 &deltanext, next, &data_fake, stopparen,
3728 recursed_depth, NULL, f,depth+1);
3731 if (deltanext == SSize_t_MAX) {
3732 is_inf = is_inf_internal = 1;
3734 } else if (max1 < minnext + deltanext)
3735 max1 = minnext + deltanext;
3737 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3739 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3740 if ( stopmin > minnext)
3741 stopmin = min + min1;
3742 flags &= ~SCF_DO_SUBSTR;
3744 data->flags |= SCF_SEEN_ACCEPT;
3747 if (data_fake.flags & SF_HAS_EVAL)
3748 data->flags |= SF_HAS_EVAL;
3749 data->whilem_c = data_fake.whilem_c;
3751 if (flags & SCF_DO_STCLASS)
3752 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3754 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3756 if (flags & SCF_DO_SUBSTR) {
3757 data->pos_min += min1;
3758 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3759 data->pos_delta = SSize_t_MAX;
3761 data->pos_delta += max1 - min1;
3762 if (max1 != min1 || is_inf)
3763 data->longest = &(data->longest_float);
3766 if (delta == SSize_t_MAX
3767 || SSize_t_MAX - delta - (max1 - min1) < 0)
3768 delta = SSize_t_MAX;
3770 delta += max1 - min1;
3771 if (flags & SCF_DO_STCLASS_OR) {
3772 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3774 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3775 flags &= ~SCF_DO_STCLASS;
3778 else if (flags & SCF_DO_STCLASS_AND) {
3780 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3781 flags &= ~SCF_DO_STCLASS;
3784 /* Switch to OR mode: cache the old value of
3785 * data->start_class */
3787 StructCopy(data->start_class, and_withp, regnode_ssc);
3788 flags &= ~SCF_DO_STCLASS_AND;
3789 StructCopy(&accum, data->start_class, regnode_ssc);
3790 flags |= SCF_DO_STCLASS_OR;
3794 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3795 OP( startbranch ) == BRANCH )
3799 Assuming this was/is a branch we are dealing with: 'scan'
3800 now points at the item that follows the branch sequence,
3801 whatever it is. We now start at the beginning of the
3802 sequence and look for subsequences of
3808 which would be constructed from a pattern like
3811 If we can find such a subsequence we need to turn the first
3812 element into a trie and then add the subsequent branch exact
3813 strings to the trie.
3817 1. patterns where the whole set of branches can be
3820 2. patterns where only a subset can be converted.
3822 In case 1 we can replace the whole set with a single regop
3823 for the trie. In case 2 we need to keep the start and end
3826 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3827 becomes BRANCH TRIE; BRANCH X;
3829 There is an additional case, that being where there is a
3830 common prefix, which gets split out into an EXACT like node
3831 preceding the TRIE node.
3833 If x(1..n)==tail then we can do a simple trie, if not we make
3834 a "jump" trie, such that when we match the appropriate word
3835 we "jump" to the appropriate tail node. Essentially we turn
3836 a nested if into a case structure of sorts.
3841 if (!re_trie_maxbuff) {
3842 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3843 if (!SvIOK(re_trie_maxbuff))
3844 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3846 if ( SvIV(re_trie_maxbuff)>=0 ) {
3848 regnode *first = (regnode *)NULL;
3849 regnode *last = (regnode *)NULL;
3850 regnode *tail = scan;
3855 SV * const mysv = sv_newmortal(); /* for dumping */
3857 /* var tail is used because there may be a TAIL
3858 regop in the way. Ie, the exacts will point to the
3859 thing following the TAIL, but the last branch will
3860 point at the TAIL. So we advance tail. If we
3861 have nested (?:) we may have to move through several
3865 while ( OP( tail ) == TAIL ) {
3866 /* this is the TAIL generated by (?:) */
3867 tail = regnext( tail );
3871 DEBUG_TRIE_COMPILE_r({
3872 regprop(RExC_rx, mysv, tail, NULL);
3873 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3874 (int)depth * 2 + 2, "",
3875 "Looking for TRIE'able sequences. Tail node is: ",
3876 SvPV_nolen_const( mysv )
3882 Step through the branches
3883 cur represents each branch,
3884 noper is the first thing to be matched as part
3886 noper_next is the regnext() of that node.
3888 We normally handle a case like this
3889 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3890 support building with NOJUMPTRIE, which restricts
3891 the trie logic to structures like /FOO|BAR/.
3893 If noper is a trieable nodetype then the branch is
3894 a possible optimization target. If we are building
3895 under NOJUMPTRIE then we require that noper_next is
3896 the same as scan (our current position in the regex
3899 Once we have two or more consecutive such branches
3900 we can create a trie of the EXACT's contents and
3901 stitch it in place into the program.
3903 If the sequence represents all of the branches in
3904 the alternation we replace the entire thing with a
3907 Otherwise when it is a subsequence we need to
3908 stitch it in place and replace only the relevant
3909 branches. This means the first branch has to remain
3910 as it is used by the alternation logic, and its
3911 next pointer, and needs to be repointed at the item
3912 on the branch chain following the last branch we
3913 have optimized away.
3915 This could be either a BRANCH, in which case the
3916 subsequence is internal, or it could be the item
3917 following the branch sequence in which case the
3918 subsequence is at the end (which does not
3919 necessarily mean the first node is the start of the
3922 TRIE_TYPE(X) is a define which maps the optype to a
3926 ----------------+-----------
3930 EXACTFU_SS | EXACTFU
3935 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3936 ( EXACT == (X) ) ? EXACT : \
3937 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3938 ( EXACTFA == (X) ) ? EXACTFA : \
3941 /* dont use tail as the end marker for this traverse */
3942 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3943 regnode * const noper = NEXTOPER( cur );
3944 U8 noper_type = OP( noper );
3945 U8 noper_trietype = TRIE_TYPE( noper_type );
3946 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3947 regnode * const noper_next = regnext( noper );
3948 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3949 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3952 DEBUG_TRIE_COMPILE_r({
3953 regprop(RExC_rx, mysv, cur, NULL);
3954 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3955 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3957 regprop(RExC_rx, mysv, noper, NULL);
3958 PerlIO_printf( Perl_debug_log, " -> %s",
3959 SvPV_nolen_const(mysv));
3962 regprop(RExC_rx, mysv, noper_next, NULL);
3963 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3964 SvPV_nolen_const(mysv));
3966 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3967 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3968 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3972 /* Is noper a trieable nodetype that can be merged
3973 * with the current trie (if there is one)? */
3977 ( noper_trietype == NOTHING)
3978 || ( trietype == NOTHING )
3979 || ( trietype == noper_trietype )
3982 && noper_next == tail
3986 /* Handle mergable triable node Either we are
3987 * the first node in a new trieable sequence,
3988 * in which case we do some bookkeeping,
3989 * otherwise we update the end pointer. */
3992 if ( noper_trietype == NOTHING ) {
3993 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3994 regnode * const noper_next = regnext( noper );
3995 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3996 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3999 if ( noper_next_trietype ) {
4000 trietype = noper_next_trietype;
4001 } else if (noper_next_type) {
4002 /* a NOTHING regop is 1 regop wide.
4003 * We need at least two for a trie
4004 * so we can't merge this in */
4008 trietype = noper_trietype;
4011 if ( trietype == NOTHING )
4012 trietype = noper_trietype;
4017 } /* end handle mergable triable node */
4019 /* handle unmergable node -
4020 * noper may either be a triable node which can
4021 * not be tried together with the current trie,
4022 * or a non triable node */
4024 /* If last is set and trietype is not
4025 * NOTHING then we have found at least two
4026 * triable branch sequences in a row of a
4027 * similar trietype so we can turn them
4028 * into a trie. If/when we allow NOTHING to
4029 * start a trie sequence this condition
4030 * will be required, and it isn't expensive
4031 * so we leave it in for now. */
4032 if ( trietype && trietype != NOTHING )
4033 make_trie( pRExC_state,
4034 startbranch, first, cur, tail,
4035 count, trietype, depth+1 );
4036 last = NULL; /* note: we clear/update
4037 first, trietype etc below,
4038 so we dont do it here */
4042 && noper_next == tail
4045 /* noper is triable, so we can start a new
4049 trietype = noper_trietype;
4051 /* if we already saw a first but the
4052 * current node is not triable then we have
4053 * to reset the first information. */
4058 } /* end handle unmergable node */
4059 } /* loop over branches */
4060 DEBUG_TRIE_COMPILE_r({
4061 regprop(RExC_rx, mysv, cur, NULL);
4062 PerlIO_printf( Perl_debug_log,
4063 "%*s- %s (%d) <SCAN FINISHED>\n",
4065 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4068 if ( last && trietype ) {
4069 if ( trietype != NOTHING ) {
4070 /* the last branch of the sequence was part of
4071 * a trie, so we have to construct it here
4072 * outside of the loop */
4073 made= make_trie( pRExC_state, startbranch,
4074 first, scan, tail, count,
4075 trietype, depth+1 );
4076 #ifdef TRIE_STUDY_OPT
4077 if ( ((made == MADE_EXACT_TRIE &&
4078 startbranch == first)
4079 || ( first_non_open == first )) &&
4081 flags |= SCF_TRIE_RESTUDY;
4082 if ( startbranch == first
4085 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4090 /* at this point we know whatever we have is a
4091 * NOTHING sequence/branch AND if 'startbranch'
4092 * is 'first' then we can turn the whole thing
4095 if ( startbranch == first ) {
4097 /* the entire thing is a NOTHING sequence,
4098 * something like this: (?:|) So we can
4099 * turn it into a plain NOTHING op. */
4100 DEBUG_TRIE_COMPILE_r({
4101 regprop(RExC_rx, mysv, cur, NULL);
4102 PerlIO_printf( Perl_debug_log,
4103 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4104 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4107 OP(startbranch)= NOTHING;
4108 NEXT_OFF(startbranch)= tail - startbranch;
4109 for ( opt= startbranch + 1; opt < tail ; opt++ )
4113 } /* end if ( last) */
4114 } /* TRIE_MAXBUF is non zero */
4119 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4120 scan = NEXTOPER(NEXTOPER(scan));
4121 } else /* single branch is optimized. */
4122 scan = NEXTOPER(scan);
4124 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4125 scan_frame *newframe = NULL;
4129 U32 my_recursed_depth= recursed_depth;
4131 if (OP(scan) != SUSPEND) {
4132 /* set the pointer */
4133 if (OP(scan) == GOSUB) {
4135 RExC_recurse[ARG2L(scan)] = scan;
4136 start = RExC_open_parens[paren-1];
4137 end = RExC_close_parens[paren-1];
4140 start = RExC_rxi->program + 1;
4145 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4147 if (!recursed_depth) {
4148 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4150 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4151 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4152 RExC_study_chunk_recursed_bytes, U8);
4154 /* we havent recursed into this paren yet, so recurse into it */
4155 DEBUG_STUDYDATA("set:", data,depth);
4156 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4157 my_recursed_depth= recursed_depth + 1;
4158 Newx(newframe,1,scan_frame);
4160 DEBUG_STUDYDATA("inf:", data,depth);
4161 /* some form of infinite recursion, assume infinite length
4163 if (flags & SCF_DO_SUBSTR) {
4164 scan_commit(pRExC_state, data, minlenp, is_inf);
4165 data->longest = &(data->longest_float);
4167 is_inf = is_inf_internal = 1;
4168 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4169 ssc_anything(data->start_class);
4170 flags &= ~SCF_DO_STCLASS;
4173 Newx(newframe,1,scan_frame);
4176 end = regnext(scan);
4181 SAVEFREEPV(newframe);
4182 newframe->next = regnext(scan);
4183 newframe->last = last;
4184 newframe->stop = stopparen;
4185 newframe->prev = frame;
4186 newframe->prev_recursed_depth = recursed_depth;
4188 DEBUG_STUDYDATA("frame-new:",data,depth);
4189 DEBUG_PEEP("fnew", scan, depth);
4196 recursed_depth= my_recursed_depth;
4201 else if (OP(scan) == EXACT) {
4202 SSize_t l = STR_LEN(scan);
4205 const U8 * const s = (U8*)STRING(scan);
4206 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4207 l = utf8_length(s, s + l);
4209 uc = *((U8*)STRING(scan));
4212 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4213 /* The code below prefers earlier match for fixed
4214 offset, later match for variable offset. */
4215 if (data->last_end == -1) { /* Update the start info. */
4216 data->last_start_min = data->pos_min;
4217 data->last_start_max = is_inf
4218 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4220 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4222 SvUTF8_on(data->last_found);
4224 SV * const sv = data->last_found;
4225 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4226 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4227 if (mg && mg->mg_len >= 0)
4228 mg->mg_len += utf8_length((U8*)STRING(scan),
4229 (U8*)STRING(scan)+STR_LEN(scan));
4231 data->last_end = data->pos_min + l;
4232 data->pos_min += l; /* As in the first entry. */
4233 data->flags &= ~SF_BEFORE_EOL;
4236 /* ANDing the code point leaves at most it, and not in locale, and
4237 * can't match null string */
4238 if (flags & SCF_DO_STCLASS_AND) {
4239 ssc_cp_and(data->start_class, uc);
4240 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4241 ssc_clear_locale(data->start_class);
4243 else if (flags & SCF_DO_STCLASS_OR) {
4244 ssc_add_cp(data->start_class, uc);
4245 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4247 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4248 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4250 flags &= ~SCF_DO_STCLASS;
4252 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
4253 SSize_t l = STR_LEN(scan);
4254 UV uc = *((U8*)STRING(scan));
4255 SV* EXACTF_invlist = _new_invlist(4); /* Start out big enough for 2
4256 separate code points */
4258 /* Search for fixed substrings supports EXACT only. */
4259 if (flags & SCF_DO_SUBSTR) {
4261 scan_commit(pRExC_state, data, minlenp, is_inf);
4264 const U8 * const s = (U8 *)STRING(scan);
4265 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4266 l = utf8_length(s, s + l);
4268 if (unfolded_multi_char) {
4269 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4271 min += l - min_subtract;
4273 delta += min_subtract;
4274 if (flags & SCF_DO_SUBSTR) {
4275 data->pos_min += l - min_subtract;
4276 if (data->pos_min < 0) {
4279 data->pos_delta += min_subtract;
4281 data->longest = &(data->longest_float);
4284 if (OP(scan) == EXACTFL) {
4286 /* We don't know what the folds are; it could be anything. XXX
4287 * Actually, we only support UTF-8 encoding for code points
4288 * above Latin1, so we could know what those folds are. */
4289 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4293 else { /* Non-locale EXACTFish */
4294 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4295 if (flags & SCF_DO_STCLASS_AND) {
4296 ssc_clear_locale(data->start_class);
4298 if (uc < 256) { /* We know what the Latin1 folds are ... */
4299 if (IS_IN_SOME_FOLD_L1(uc)) { /* For instance, we
4300 know if anything folds
4302 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist,
4303 PL_fold_latin1[uc]);
4304 if (OP(scan) != EXACTFA) { /* The folds below aren't
4306 if (isARG2_lower_or_UPPER_ARG1('s', uc)) {
4308 = add_cp_to_invlist(EXACTF_invlist,
4309 LATIN_SMALL_LETTER_SHARP_S);
4311 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
4313 = add_cp_to_invlist(EXACTF_invlist, 's');
4315 = add_cp_to_invlist(EXACTF_invlist, 'S');
4319 /* We also know if there are above-Latin1 code points
4320 * that fold to this (none legal for ASCII and /iaa) */
4321 if ((! isASCII(uc) || OP(scan) != EXACTFA)
4322 && HAS_NONLATIN1_FOLD_CLOSURE(uc))
4324 /* XXX We could know exactly what does fold to this
4325 * if the reverse folds are loaded, as currently in
4327 _invlist_union(EXACTF_invlist,
4333 else { /* Non-locale, above Latin1. XXX We don't currently
4334 know what participates in folds with this, so have
4335 to assume anything could */
4337 /* XXX We could know exactly what does fold to this if the
4338 * reverse folds are loaded, as currently in S_regclass().
4339 * But we do know that under /iaa nothing in the ASCII
4340 * range can participate */
4341 if (OP(scan) == EXACTFA) {
4342 _invlist_union_complement_2nd(EXACTF_invlist,
4343 PL_XPosix_ptrs[_CC_ASCII],
4347 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4352 if (flags & SCF_DO_STCLASS_AND) {
4353 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4354 ANYOF_POSIXL_ZERO(data->start_class);
4355 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4357 else if (flags & SCF_DO_STCLASS_OR) {
4358 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4359 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4361 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4362 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4364 flags &= ~SCF_DO_STCLASS;
4365 SvREFCNT_dec(EXACTF_invlist);
4367 else if (REGNODE_VARIES(OP(scan))) {
4368 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4369 I32 fl = 0, f = flags;
4370 regnode * const oscan = scan;
4371 regnode_ssc this_class;
4372 regnode_ssc *oclass = NULL;
4373 I32 next_is_eval = 0;
4375 switch (PL_regkind[OP(scan)]) {
4376 case WHILEM: /* End of (?:...)* . */
4377 scan = NEXTOPER(scan);
4380 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4381 next = NEXTOPER(scan);
4382 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
4384 maxcount = REG_INFTY;
4385 next = regnext(scan);
4386 scan = NEXTOPER(scan);
4390 if (flags & SCF_DO_SUBSTR)
4395 if (flags & SCF_DO_STCLASS) {
4397 maxcount = REG_INFTY;
4398 next = regnext(scan);
4399 scan = NEXTOPER(scan);
4402 if (flags & SCF_DO_SUBSTR) {
4403 scan_commit(pRExC_state, data, minlenp, is_inf);
4404 /* Cannot extend fixed substrings */
4405 data->longest = &(data->longest_float);
4407 is_inf = is_inf_internal = 1;
4408 scan = regnext(scan);
4409 goto optimize_curly_tail;
4411 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4412 && (scan->flags == stopparen))
4417 mincount = ARG1(scan);
4418 maxcount = ARG2(scan);
4420 next = regnext(scan);
4421 if (OP(scan) == CURLYX) {
4422 I32 lp = (data ? *(data->last_closep) : 0);
4423 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4425 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4426 next_is_eval = (OP(scan) == EVAL);
4428 if (flags & SCF_DO_SUBSTR) {
4430 scan_commit(pRExC_state, data, minlenp, is_inf);
4431 /* Cannot extend fixed substrings */
4432 pos_before = data->pos_min;
4436 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4438 data->flags |= SF_IS_INF;
4440 if (flags & SCF_DO_STCLASS) {
4441 ssc_init(pRExC_state, &this_class);
4442 oclass = data->start_class;
4443 data->start_class = &this_class;
4444 f |= SCF_DO_STCLASS_AND;
4445 f &= ~SCF_DO_STCLASS_OR;
4447 /* Exclude from super-linear cache processing any {n,m}
4448 regops for which the combination of input pos and regex
4449 pos is not enough information to determine if a match
4452 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4453 regex pos at the \s*, the prospects for a match depend not
4454 only on the input position but also on how many (bar\s*)
4455 repeats into the {4,8} we are. */
4456 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4457 f &= ~SCF_WHILEM_VISITED_POS;
4459 /* This will finish on WHILEM, setting scan, or on NULL: */
4460 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4461 last, data, stopparen, recursed_depth, NULL,
4463 ? (f & ~SCF_DO_SUBSTR)
4467 if (flags & SCF_DO_STCLASS)
4468 data->start_class = oclass;
4469 if (mincount == 0 || minnext == 0) {
4470 if (flags & SCF_DO_STCLASS_OR) {
4471 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4473 else if (flags & SCF_DO_STCLASS_AND) {
4474 /* Switch to OR mode: cache the old value of
4475 * data->start_class */
4477 StructCopy(data->start_class, and_withp, regnode_ssc);
4478 flags &= ~SCF_DO_STCLASS_AND;
4479 StructCopy(&this_class, data->start_class, regnode_ssc);
4480 flags |= SCF_DO_STCLASS_OR;
4481 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
4483 } else { /* Non-zero len */
4484 if (flags & SCF_DO_STCLASS_OR) {
4485 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4486 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4488 else if (flags & SCF_DO_STCLASS_AND)
4489 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4490 flags &= ~SCF_DO_STCLASS;
4492 if (!scan) /* It was not CURLYX, but CURLY. */
4494 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4495 /* ? quantifier ok, except for (?{ ... }) */
4496 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4497 && (minnext == 0) && (deltanext == 0)
4498 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4499 && maxcount <= REG_INFTY/3) /* Complement check for big
4502 /* Fatal warnings may leak the regexp without this: */
4503 SAVEFREESV(RExC_rx_sv);
4504 ckWARNreg(RExC_parse,
4505 "Quantifier unexpected on zero-length expression");
4506 (void)ReREFCNT_inc(RExC_rx_sv);
4509 min += minnext * mincount;
4510 is_inf_internal |= deltanext == SSize_t_MAX
4511 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4512 is_inf |= is_inf_internal;
4514 delta = SSize_t_MAX;
4516 delta += (minnext + deltanext) * maxcount
4517 - minnext * mincount;
4519 /* Try powerful optimization CURLYX => CURLYN. */
4520 if ( OP(oscan) == CURLYX && data
4521 && data->flags & SF_IN_PAR
4522 && !(data->flags & SF_HAS_EVAL)
4523 && !deltanext && minnext == 1 ) {
4524 /* Try to optimize to CURLYN. */
4525 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4526 regnode * const nxt1 = nxt;
4533 if (!REGNODE_SIMPLE(OP(nxt))
4534 && !(PL_regkind[OP(nxt)] == EXACT
4535 && STR_LEN(nxt) == 1))
4541 if (OP(nxt) != CLOSE)
4543 if (RExC_open_parens) {
4544 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4545 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4547 /* Now we know that nxt2 is the only contents: */
4548 oscan->flags = (U8)ARG(nxt);
4550 OP(nxt1) = NOTHING; /* was OPEN. */
4553 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4554 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4555 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4556 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4557 OP(nxt + 1) = OPTIMIZED; /* was count. */
4558 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4563 /* Try optimization CURLYX => CURLYM. */
4564 if ( OP(oscan) == CURLYX && data
4565 && !(data->flags & SF_HAS_PAR)
4566 && !(data->flags & SF_HAS_EVAL)
4567 && !deltanext /* atom is fixed width */
4568 && minnext != 0 /* CURLYM can't handle zero width */
4570 /* Nor characters whose fold at run-time may be
4571 * multi-character */
4572 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4574 /* XXXX How to optimize if data == 0? */
4575 /* Optimize to a simpler form. */
4576 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4580 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4581 && (OP(nxt2) != WHILEM))
4583 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4584 /* Need to optimize away parenths. */
4585 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4586 /* Set the parenth number. */
4587 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4589 oscan->flags = (U8)ARG(nxt);
4590 if (RExC_open_parens) {
4591 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4592 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4594 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4595 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4598 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4599 OP(nxt + 1) = OPTIMIZED; /* was count. */
4600 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4601 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4604 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4605 regnode *nnxt = regnext(nxt1);
4607 if (reg_off_by_arg[OP(nxt1)])
4608 ARG_SET(nxt1, nxt2 - nxt1);
4609 else if (nxt2 - nxt1 < U16_MAX)
4610 NEXT_OFF(nxt1) = nxt2 - nxt1;
4612 OP(nxt) = NOTHING; /* Cannot beautify */
4617 /* Optimize again: */
4618 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4619 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4624 else if ((OP(oscan) == CURLYX)
4625 && (flags & SCF_WHILEM_VISITED_POS)
4626 /* See the comment on a similar expression above.
4627 However, this time it's not a subexpression
4628 we care about, but the expression itself. */
4629 && (maxcount == REG_INFTY)
4630 && data && ++data->whilem_c < 16) {
4631 /* This stays as CURLYX, we can put the count/of pair. */
4632 /* Find WHILEM (as in regexec.c) */
4633 regnode *nxt = oscan + NEXT_OFF(oscan);
4635 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4637 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4638 | (RExC_whilem_seen << 4)); /* On WHILEM */
4640 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4642 if (flags & SCF_DO_SUBSTR) {
4643 SV *last_str = NULL;
4644 STRLEN last_chrs = 0;
4645 int counted = mincount != 0;
4647 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4649 SSize_t b = pos_before >= data->last_start_min
4650 ? pos_before : data->last_start_min;
4652 const char * const s = SvPV_const(data->last_found, l);
4653 SSize_t old = b - data->last_start_min;
4656 old = utf8_hop((U8*)s, old) - (U8*)s;
4658 /* Get the added string: */
4659 last_str = newSVpvn_utf8(s + old, l, UTF);
4660 last_chrs = UTF ? utf8_length((U8*)(s + old),
4661 (U8*)(s + old + l)) : l;
4662 if (deltanext == 0 && pos_before == b) {
4663 /* What was added is a constant string */
4666 SvGROW(last_str, (mincount * l) + 1);
4667 repeatcpy(SvPVX(last_str) + l,
4668 SvPVX_const(last_str), l,
4670 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4671 /* Add additional parts. */
4672 SvCUR_set(data->last_found,
4673 SvCUR(data->last_found) - l);
4674 sv_catsv(data->last_found, last_str);
4676 SV * sv = data->last_found;
4678 SvUTF8(sv) && SvMAGICAL(sv) ?
4679 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4680 if (mg && mg->mg_len >= 0)
4681 mg->mg_len += last_chrs * (mincount-1);
4683 last_chrs *= mincount;
4684 data->last_end += l * (mincount - 1);
4687 /* start offset must point into the last copy */
4688 data->last_start_min += minnext * (mincount - 1);
4689 data->last_start_max += is_inf ? SSize_t_MAX
4690 : (maxcount - 1) * (minnext + data->pos_delta);
4693 /* It is counted once already... */
4694 data->pos_min += minnext * (mincount - counted);
4696 PerlIO_printf(Perl_debug_log, "counted=%"UVdf" deltanext=%"UVdf
4697 " SSize_t_MAX=%"UVdf" minnext=%"UVdf
4698 " maxcount=%"UVdf" mincount=%"UVdf"\n",
4699 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4701 if (deltanext != SSize_t_MAX)
4702 PerlIO_printf(Perl_debug_log, "LHS=%"UVdf" RHS=%"UVdf"\n",
4703 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4704 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4706 if (deltanext == SSize_t_MAX
4707 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4708 data->pos_delta = SSize_t_MAX;
4710 data->pos_delta += - counted * deltanext +
4711 (minnext + deltanext) * maxcount - minnext * mincount;
4712 if (mincount != maxcount) {
4713 /* Cannot extend fixed substrings found inside
4715 scan_commit(pRExC_state, data, minlenp, is_inf);
4716 if (mincount && last_str) {
4717 SV * const sv = data->last_found;
4718 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4719 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4723 sv_setsv(sv, last_str);
4724 data->last_end = data->pos_min;
4725 data->last_start_min = data->pos_min - last_chrs;
4726 data->last_start_max = is_inf
4728 : data->pos_min + data->pos_delta - last_chrs;
4730 data->longest = &(data->longest_float);
4732 SvREFCNT_dec(last_str);
4734 if (data && (fl & SF_HAS_EVAL))
4735 data->flags |= SF_HAS_EVAL;
4736 optimize_curly_tail:
4737 if (OP(oscan) != CURLYX) {
4738 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4740 NEXT_OFF(oscan) += NEXT_OFF(next);
4746 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4751 if (flags & SCF_DO_SUBSTR) {
4752 /* Cannot expect anything... */
4753 scan_commit(pRExC_state, data, minlenp, is_inf);
4754 data->longest = &(data->longest_float);
4756 is_inf = is_inf_internal = 1;
4757 if (flags & SCF_DO_STCLASS_OR) {
4758 if (OP(scan) == CLUMP) {
4759 /* Actually is any start char, but very few code points
4760 * aren't start characters */
4761 ssc_match_all_cp(data->start_class);
4764 ssc_anything(data->start_class);
4767 flags &= ~SCF_DO_STCLASS;
4771 else if (OP(scan) == LNBREAK) {
4772 if (flags & SCF_DO_STCLASS) {
4773 if (flags & SCF_DO_STCLASS_AND) {
4774 ssc_intersection(data->start_class,
4775 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
4776 ssc_clear_locale(data->start_class);
4777 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4779 else if (flags & SCF_DO_STCLASS_OR) {
4780 ssc_union(data->start_class,
4781 PL_XPosix_ptrs[_CC_VERTSPACE],
4783 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4785 /* See commit msg for
4786 * 749e076fceedeb708a624933726e7989f2302f6a */
4787 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4789 flags &= ~SCF_DO_STCLASS;
4792 delta++; /* Because of the 2 char string cr-lf */
4793 if (flags & SCF_DO_SUBSTR) {
4794 /* Cannot expect anything... */
4795 scan_commit(pRExC_state, data, minlenp, is_inf);
4797 data->pos_delta += 1;
4798 data->longest = &(data->longest_float);
4801 else if (REGNODE_SIMPLE(OP(scan))) {
4803 if (flags & SCF_DO_SUBSTR) {
4804 scan_commit(pRExC_state, data, minlenp, is_inf);
4808 if (flags & SCF_DO_STCLASS) {
4810 SV* my_invlist = sv_2mortal(_new_invlist(0));
4813 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4814 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4816 /* Some of the logic below assumes that switching
4817 locale on will only add false positives. */
4822 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
4827 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4828 ssc_match_all_cp(data->start_class);
4833 SV* REG_ANY_invlist = _new_invlist(2);
4834 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
4836 if (flags & SCF_DO_STCLASS_OR) {
4837 ssc_union(data->start_class,
4839 TRUE /* TRUE => invert, hence all but \n
4843 else if (flags & SCF_DO_STCLASS_AND) {
4844 ssc_intersection(data->start_class,
4846 TRUE /* TRUE => invert */
4848 ssc_clear_locale(data->start_class);
4850 SvREFCNT_dec_NN(REG_ANY_invlist);
4855 if (flags & SCF_DO_STCLASS_AND)
4856 ssc_and(pRExC_state, data->start_class,
4857 (regnode_charclass *) scan);
4859 ssc_or(pRExC_state, data->start_class,
4860 (regnode_charclass *) scan);
4868 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
4869 if (flags & SCF_DO_STCLASS_AND) {
4870 bool was_there = cBOOL(
4871 ANYOF_POSIXL_TEST(data->start_class,
4873 ANYOF_POSIXL_ZERO(data->start_class);
4874 if (was_there) { /* Do an AND */
4875 ANYOF_POSIXL_SET(data->start_class, namedclass);
4877 /* No individual code points can now match */
4878 data->start_class->invlist
4879 = sv_2mortal(_new_invlist(0));
4882 int complement = namedclass + ((invert) ? -1 : 1);
4884 assert(flags & SCF_DO_STCLASS_OR);
4886 /* If the complement of this class was already there,
4887 * the result is that they match all code points,
4888 * (\d + \D == everything). Remove the classes from
4889 * future consideration. Locale is not relevant in
4891 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
4892 ssc_match_all_cp(data->start_class);
4893 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
4894 ANYOF_POSIXL_CLEAR(data->start_class, complement);
4896 else { /* The usual case; just add this class to the
4898 ANYOF_POSIXL_SET(data->start_class, namedclass);
4903 case NPOSIXA: /* For these, we always know the exact set of
4908 if (FLAGS(scan) == _CC_ASCII) {
4909 my_invlist = PL_XPosix_ptrs[_CC_ASCII];
4912 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
4913 PL_XPosix_ptrs[_CC_ASCII],
4924 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
4926 /* NPOSIXD matches all upper Latin1 code points unless the
4927 * target string being matched is UTF-8, which is
4928 * unknowable until match time. Since we are going to
4929 * invert, we want to get rid of all of them so that the
4930 * inversion will match all */
4931 if (OP(scan) == NPOSIXD) {
4932 _invlist_subtract(my_invlist, PL_UpperLatin1,
4938 if (flags & SCF_DO_STCLASS_AND) {
4939 ssc_intersection(data->start_class, my_invlist, invert);
4940 ssc_clear_locale(data->start_class);
4943 assert(flags & SCF_DO_STCLASS_OR);
4944 ssc_union(data->start_class, my_invlist, invert);
4947 if (flags & SCF_DO_STCLASS_OR)
4948 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4949 flags &= ~SCF_DO_STCLASS;
4952 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4953 data->flags |= (OP(scan) == MEOL
4956 scan_commit(pRExC_state, data, minlenp, is_inf);
4959 else if ( PL_regkind[OP(scan)] == BRANCHJ
4960 /* Lookbehind, or need to calculate parens/evals/stclass: */
4961 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4962 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4963 if ( OP(scan) == UNLESSM &&
4965 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4966 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4969 regnode *upto= regnext(scan);
4971 SV * const mysv_val=sv_newmortal();
4972 DEBUG_STUDYDATA("OPFAIL",data,depth);
4974 /*DEBUG_PARSE_MSG("opfail");*/
4975 regprop(RExC_rx, mysv_val, upto, NULL);
4976 PerlIO_printf(Perl_debug_log,
4977 "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4978 SvPV_nolen_const(mysv_val),
4979 (IV)REG_NODE_NUM(upto),
4984 NEXT_OFF(scan) = upto - scan;
4985 for (opt= scan + 1; opt < upto ; opt++)
4986 OP(opt) = OPTIMIZED;
4990 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4991 || OP(scan) == UNLESSM )
4993 /* Negative Lookahead/lookbehind
4994 In this case we can't do fixed string optimisation.
4997 SSize_t deltanext, minnext, fake = 0;
5002 data_fake.flags = 0;
5004 data_fake.whilem_c = data->whilem_c;
5005 data_fake.last_closep = data->last_closep;
5008 data_fake.last_closep = &fake;
5009 data_fake.pos_delta = delta;
5010 if ( flags & SCF_DO_STCLASS && !scan->flags
5011 && OP(scan) == IFMATCH ) { /* Lookahead */
5012 ssc_init(pRExC_state, &intrnl);
5013 data_fake.start_class = &intrnl;
5014 f |= SCF_DO_STCLASS_AND;
5016 if (flags & SCF_WHILEM_VISITED_POS)
5017 f |= SCF_WHILEM_VISITED_POS;
5018 next = regnext(scan);
5019 nscan = NEXTOPER(NEXTOPER(scan));
5020 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5021 last, &data_fake, stopparen,
5022 recursed_depth, NULL, f, depth+1);
5025 FAIL("Variable length lookbehind not implemented");
5027 else if (minnext > (I32)U8_MAX) {
5028 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5031 scan->flags = (U8)minnext;
5034 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5036 if (data_fake.flags & SF_HAS_EVAL)
5037 data->flags |= SF_HAS_EVAL;
5038 data->whilem_c = data_fake.whilem_c;
5040 if (f & SCF_DO_STCLASS_AND) {
5041 if (flags & SCF_DO_STCLASS_OR) {
5042 /* OR before, AND after: ideally we would recurse with
5043 * data_fake to get the AND applied by study of the
5044 * remainder of the pattern, and then derecurse;
5045 * *** HACK *** for now just treat as "no information".
5046 * See [perl #56690].
5048 ssc_init(pRExC_state, data->start_class);
5050 /* AND before and after: combine and continue */
5051 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5055 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5057 /* Positive Lookahead/lookbehind
5058 In this case we can do fixed string optimisation,
5059 but we must be careful about it. Note in the case of
5060 lookbehind the positions will be offset by the minimum
5061 length of the pattern, something we won't know about
5062 until after the recurse.
5064 SSize_t deltanext, fake = 0;
5068 /* We use SAVEFREEPV so that when the full compile
5069 is finished perl will clean up the allocated
5070 minlens when it's all done. This way we don't
5071 have to worry about freeing them when we know
5072 they wont be used, which would be a pain.
5075 Newx( minnextp, 1, SSize_t );
5076 SAVEFREEPV(minnextp);
5079 StructCopy(data, &data_fake, scan_data_t);
5080 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5083 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5084 data_fake.last_found=newSVsv(data->last_found);
5088 data_fake.last_closep = &fake;
5089 data_fake.flags = 0;
5090 data_fake.pos_delta = delta;
5092 data_fake.flags |= SF_IS_INF;
5093 if ( flags & SCF_DO_STCLASS && !scan->flags
5094 && OP(scan) == IFMATCH ) { /* Lookahead */
5095 ssc_init(pRExC_state, &intrnl);
5096 data_fake.start_class = &intrnl;
5097 f |= SCF_DO_STCLASS_AND;
5099 if (flags & SCF_WHILEM_VISITED_POS)
5100 f |= SCF_WHILEM_VISITED_POS;
5101 next = regnext(scan);
5102 nscan = NEXTOPER(NEXTOPER(scan));
5104 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5105 &deltanext, last, &data_fake,
5106 stopparen, recursed_depth, NULL,
5110 FAIL("Variable length lookbehind not implemented");
5112 else if (*minnextp > (I32)U8_MAX) {
5113 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5116 scan->flags = (U8)*minnextp;
5121 if (f & SCF_DO_STCLASS_AND) {
5122 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5125 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5127 if (data_fake.flags & SF_HAS_EVAL)
5128 data->flags |= SF_HAS_EVAL;
5129 data->whilem_c = data_fake.whilem_c;
5130 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5131 if (RExC_rx->minlen<*minnextp)
5132 RExC_rx->minlen=*minnextp;
5133 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5134 SvREFCNT_dec_NN(data_fake.last_found);
5136 if ( data_fake.minlen_fixed != minlenp )
5138 data->offset_fixed= data_fake.offset_fixed;
5139 data->minlen_fixed= data_fake.minlen_fixed;
5140 data->lookbehind_fixed+= scan->flags;
5142 if ( data_fake.minlen_float != minlenp )
5144 data->minlen_float= data_fake.minlen_float;
5145 data->offset_float_min=data_fake.offset_float_min;
5146 data->offset_float_max=data_fake.offset_float_max;
5147 data->lookbehind_float+= scan->flags;
5154 else if (OP(scan) == OPEN) {
5155 if (stopparen != (I32)ARG(scan))
5158 else if (OP(scan) == CLOSE) {
5159 if (stopparen == (I32)ARG(scan)) {
5162 if ((I32)ARG(scan) == is_par) {
5163 next = regnext(scan);
5165 if ( next && (OP(next) != WHILEM) && next < last)
5166 is_par = 0; /* Disable optimization */
5169 *(data->last_closep) = ARG(scan);
5171 else if (OP(scan) == EVAL) {
5173 data->flags |= SF_HAS_EVAL;
5175 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5176 if (flags & SCF_DO_SUBSTR) {
5177 scan_commit(pRExC_state, data, minlenp, is_inf);
5178 flags &= ~SCF_DO_SUBSTR;
5180 if (data && OP(scan)==ACCEPT) {
5181 data->flags |= SCF_SEEN_ACCEPT;
5186 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5188 if (flags & SCF_DO_SUBSTR) {
5189 scan_commit(pRExC_state, data, minlenp, is_inf);
5190 data->longest = &(data->longest_float);
5192 is_inf = is_inf_internal = 1;
5193 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5194 ssc_anything(data->start_class);
5195 flags &= ~SCF_DO_STCLASS;
5197 else if (OP(scan) == GPOS) {
5198 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5199 !(delta || is_inf || (data && data->pos_delta)))
5201 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5202 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5203 if (RExC_rx->gofs < (STRLEN)min)
5204 RExC_rx->gofs = min;
5206 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5210 #ifdef TRIE_STUDY_OPT
5211 #ifdef FULL_TRIE_STUDY
5212 else if (PL_regkind[OP(scan)] == TRIE) {
5213 /* NOTE - There is similar code to this block above for handling
5214 BRANCH nodes on the initial study. If you change stuff here
5216 regnode *trie_node= scan;
5217 regnode *tail= regnext(scan);
5218 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5219 SSize_t max1 = 0, min1 = SSize_t_MAX;
5222 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5223 /* Cannot merge strings after this. */
5224 scan_commit(pRExC_state, data, minlenp, is_inf);
5226 if (flags & SCF_DO_STCLASS)
5227 ssc_init_zero(pRExC_state, &accum);
5233 const regnode *nextbranch= NULL;
5236 for ( word=1 ; word <= trie->wordcount ; word++)
5238 SSize_t deltanext=0, minnext=0, f = 0, fake;
5239 regnode_ssc this_class;
5241 data_fake.flags = 0;
5243 data_fake.whilem_c = data->whilem_c;
5244 data_fake.last_closep = data->last_closep;
5247 data_fake.last_closep = &fake;
5248 data_fake.pos_delta = delta;
5249 if (flags & SCF_DO_STCLASS) {
5250 ssc_init(pRExC_state, &this_class);
5251 data_fake.start_class = &this_class;
5252 f = SCF_DO_STCLASS_AND;
5254 if (flags & SCF_WHILEM_VISITED_POS)
5255 f |= SCF_WHILEM_VISITED_POS;
5257 if (trie->jump[word]) {
5259 nextbranch = trie_node + trie->jump[0];
5260 scan= trie_node + trie->jump[word];
5261 /* We go from the jump point to the branch that follows
5262 it. Note this means we need the vestigal unused
5263 branches even though they arent otherwise used. */
5264 minnext = study_chunk(pRExC_state, &scan, minlenp,
5265 &deltanext, (regnode *)nextbranch, &data_fake,
5266 stopparen, recursed_depth, NULL, f,depth+1);
5268 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5269 nextbranch= regnext((regnode*)nextbranch);
5271 if (min1 > (SSize_t)(minnext + trie->minlen))
5272 min1 = minnext + trie->minlen;
5273 if (deltanext == SSize_t_MAX) {
5274 is_inf = is_inf_internal = 1;
5276 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5277 max1 = minnext + deltanext + trie->maxlen;
5279 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5281 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5282 if ( stopmin > min + min1)
5283 stopmin = min + min1;
5284 flags &= ~SCF_DO_SUBSTR;
5286 data->flags |= SCF_SEEN_ACCEPT;
5289 if (data_fake.flags & SF_HAS_EVAL)
5290 data->flags |= SF_HAS_EVAL;
5291 data->whilem_c = data_fake.whilem_c;
5293 if (flags & SCF_DO_STCLASS)
5294 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5297 if (flags & SCF_DO_SUBSTR) {
5298 data->pos_min += min1;
5299 data->pos_delta += max1 - min1;
5300 if (max1 != min1 || is_inf)
5301 data->longest = &(data->longest_float);
5304 delta += max1 - min1;
5305 if (flags & SCF_DO_STCLASS_OR) {
5306 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5308 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5309 flags &= ~SCF_DO_STCLASS;
5312 else if (flags & SCF_DO_STCLASS_AND) {
5314 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5315 flags &= ~SCF_DO_STCLASS;
5318 /* Switch to OR mode: cache the old value of
5319 * data->start_class */
5321 StructCopy(data->start_class, and_withp, regnode_ssc);
5322 flags &= ~SCF_DO_STCLASS_AND;
5323 StructCopy(&accum, data->start_class, regnode_ssc);
5324 flags |= SCF_DO_STCLASS_OR;
5331 else if (PL_regkind[OP(scan)] == TRIE) {
5332 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5335 min += trie->minlen;
5336 delta += (trie->maxlen - trie->minlen);
5337 flags &= ~SCF_DO_STCLASS; /* xxx */
5338 if (flags & SCF_DO_SUBSTR) {
5339 /* Cannot expect anything... */
5340 scan_commit(pRExC_state, data, minlenp, is_inf);
5341 data->pos_min += trie->minlen;
5342 data->pos_delta += (trie->maxlen - trie->minlen);
5343 if (trie->maxlen != trie->minlen)
5344 data->longest = &(data->longest_float);
5346 if (trie->jump) /* no more substrings -- for now /grr*/
5347 flags &= ~SCF_DO_SUBSTR;
5349 #endif /* old or new */
5350 #endif /* TRIE_STUDY_OPT */
5352 /* Else: zero-length, ignore. */
5353 scan = regnext(scan);
5355 /* If we are exiting a recursion we can unset its recursed bit
5356 * and allow ourselves to enter it again - no danger of an
5357 * infinite loop there.
5358 if (stopparen > -1 && recursed) {
5359 DEBUG_STUDYDATA("unset:", data,depth);
5360 PAREN_UNSET( recursed, stopparen);
5364 DEBUG_STUDYDATA("frame-end:",data,depth);
5365 DEBUG_PEEP("fend", scan, depth);
5366 /* restore previous context */
5369 stopparen = frame->stop;
5370 recursed_depth = frame->prev_recursed_depth;
5373 frame = frame->prev;
5374 goto fake_study_recurse;
5379 DEBUG_STUDYDATA("pre-fin:",data,depth);
5382 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5384 if (flags & SCF_DO_SUBSTR && is_inf)
5385 data->pos_delta = SSize_t_MAX - data->pos_min;
5386 if (is_par > (I32)U8_MAX)
5388 if (is_par && pars==1 && data) {
5389 data->flags |= SF_IN_PAR;
5390 data->flags &= ~SF_HAS_PAR;
5392 else if (pars && data) {
5393 data->flags |= SF_HAS_PAR;
5394 data->flags &= ~SF_IN_PAR;
5396 if (flags & SCF_DO_STCLASS_OR)
5397 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5398 if (flags & SCF_TRIE_RESTUDY)
5399 data->flags |= SCF_TRIE_RESTUDY;
5401 DEBUG_STUDYDATA("post-fin:",data,depth);
5404 SSize_t final_minlen= min < stopmin ? min : stopmin;
5406 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) && (RExC_maxlen < final_minlen + delta)) {
5407 RExC_maxlen = final_minlen + delta;
5409 return final_minlen;
5415 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5417 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5419 PERL_ARGS_ASSERT_ADD_DATA;
5421 Renewc(RExC_rxi->data,
5422 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5423 char, struct reg_data);
5425 Renew(RExC_rxi->data->what, count + n, U8);
5427 Newx(RExC_rxi->data->what, n, U8);
5428 RExC_rxi->data->count = count + n;
5429 Copy(s, RExC_rxi->data->what + count, n, U8);
5433 /*XXX: todo make this not included in a non debugging perl */
5434 #ifndef PERL_IN_XSUB_RE
5436 Perl_reginitcolors(pTHX)
5439 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5441 char *t = savepv(s);
5445 t = strchr(t, '\t');
5451 PL_colors[i] = t = (char *)"";
5456 PL_colors[i++] = (char *)"";
5463 #ifdef TRIE_STUDY_OPT
5464 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5467 (data.flags & SCF_TRIE_RESTUDY) \
5475 #define CHECK_RESTUDY_GOTO_butfirst
5479 * pregcomp - compile a regular expression into internal code
5481 * Decides which engine's compiler to call based on the hint currently in
5485 #ifndef PERL_IN_XSUB_RE
5487 /* return the currently in-scope regex engine (or the default if none) */
5489 regexp_engine const *
5490 Perl_current_re_engine(pTHX)
5494 if (IN_PERL_COMPILETIME) {
5495 HV * const table = GvHV(PL_hintgv);
5498 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5499 return &reh_regexp_engine;
5500 ptr = hv_fetchs(table, "regcomp", FALSE);
5501 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5502 return &reh_regexp_engine;
5503 return INT2PTR(regexp_engine*,SvIV(*ptr));
5507 if (!PL_curcop->cop_hints_hash)
5508 return &reh_regexp_engine;
5509 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5510 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5511 return &reh_regexp_engine;
5512 return INT2PTR(regexp_engine*,SvIV(ptr));
5518 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5521 regexp_engine const *eng = current_re_engine();
5522 GET_RE_DEBUG_FLAGS_DECL;
5524 PERL_ARGS_ASSERT_PREGCOMP;
5526 /* Dispatch a request to compile a regexp to correct regexp engine. */
5528 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5531 return CALLREGCOMP_ENG(eng, pattern, flags);
5535 /* public(ish) entry point for the perl core's own regex compiling code.
5536 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5537 * pattern rather than a list of OPs, and uses the internal engine rather
5538 * than the current one */
5541 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5543 SV *pat = pattern; /* defeat constness! */
5544 PERL_ARGS_ASSERT_RE_COMPILE;
5545 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5546 #ifdef PERL_IN_XSUB_RE
5551 NULL, NULL, rx_flags, 0);
5555 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5556 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5557 * point to the realloced string and length.
5559 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5563 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5564 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5566 U8 *const src = (U8*)*pat_p;
5569 STRLEN s = 0, d = 0;
5571 GET_RE_DEBUG_FLAGS_DECL;
5573 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5574 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5576 Newx(dst, *plen_p * 2 + 1, U8);
5578 while (s < *plen_p) {
5579 if (NATIVE_BYTE_IS_INVARIANT(src[s]))
5582 dst[d++] = UTF8_EIGHT_BIT_HI(src[s]);
5583 dst[d] = UTF8_EIGHT_BIT_LO(src[s]);
5585 if (n < num_code_blocks) {
5586 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5587 pRExC_state->code_blocks[n].start = d;
5588 assert(dst[d] == '(');
5591 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5592 pRExC_state->code_blocks[n].end = d;
5593 assert(dst[d] == ')');
5603 *pat_p = (char*) dst;
5605 RExC_orig_utf8 = RExC_utf8 = 1;
5610 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5611 * while recording any code block indices, and handling overloading,
5612 * nested qr// objects etc. If pat is null, it will allocate a new
5613 * string, or just return the first arg, if there's only one.
5615 * Returns the malloced/updated pat.
5616 * patternp and pat_count is the array of SVs to be concatted;
5617 * oplist is the optional list of ops that generated the SVs;
5618 * recompile_p is a pointer to a boolean that will be set if
5619 * the regex will need to be recompiled.
5620 * delim, if non-null is an SV that will be inserted between each element
5624 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5625 SV *pat, SV ** const patternp, int pat_count,
5626 OP *oplist, bool *recompile_p, SV *delim)
5630 bool use_delim = FALSE;
5631 bool alloced = FALSE;
5633 /* if we know we have at least two args, create an empty string,
5634 * then concatenate args to that. For no args, return an empty string */
5635 if (!pat && pat_count != 1) {
5636 pat = newSVpvn("", 0);
5641 for (svp = patternp; svp < patternp + pat_count; svp++) {
5644 STRLEN orig_patlen = 0;
5646 SV *msv = use_delim ? delim : *svp;
5647 if (!msv) msv = &PL_sv_undef;
5649 /* if we've got a delimiter, we go round the loop twice for each
5650 * svp slot (except the last), using the delimiter the second
5659 if (SvTYPE(msv) == SVt_PVAV) {
5660 /* we've encountered an interpolated array within
5661 * the pattern, e.g. /...@a..../. Expand the list of elements,
5662 * then recursively append elements.
5663 * The code in this block is based on S_pushav() */
5665 AV *const av = (AV*)msv;
5666 const SSize_t maxarg = AvFILL(av) + 1;
5670 assert(oplist->op_type == OP_PADAV
5671 || oplist->op_type == OP_RV2AV);
5672 oplist = oplist->op_sibling;;
5675 if (SvRMAGICAL(av)) {
5678 Newx(array, maxarg, SV*);
5680 for (i=0; i < maxarg; i++) {
5681 SV ** const svp = av_fetch(av, i, FALSE);
5682 array[i] = svp ? *svp : &PL_sv_undef;
5686 array = AvARRAY(av);
5688 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5689 array, maxarg, NULL, recompile_p,
5691 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5697 /* we make the assumption here that each op in the list of
5698 * op_siblings maps to one SV pushed onto the stack,
5699 * except for code blocks, with have both an OP_NULL and
5701 * This allows us to match up the list of SVs against the
5702 * list of OPs to find the next code block.
5704 * Note that PUSHMARK PADSV PADSV ..
5706 * PADRANGE PADSV PADSV ..
5707 * so the alignment still works. */
5710 if (oplist->op_type == OP_NULL
5711 && (oplist->op_flags & OPf_SPECIAL))
5713 assert(n < pRExC_state->num_code_blocks);
5714 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5715 pRExC_state->code_blocks[n].block = oplist;
5716 pRExC_state->code_blocks[n].src_regex = NULL;
5719 oplist = oplist->op_sibling; /* skip CONST */
5722 oplist = oplist->op_sibling;;
5725 /* apply magic and QR overloading to arg */
5728 if (SvROK(msv) && SvAMAGIC(msv)) {
5729 SV *sv = AMG_CALLunary(msv, regexp_amg);
5733 if (SvTYPE(sv) != SVt_REGEXP)
5734 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5739 /* try concatenation overload ... */
5740 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5741 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5744 /* overloading involved: all bets are off over literal
5745 * code. Pretend we haven't seen it */
5746 pRExC_state->num_code_blocks -= n;
5750 /* ... or failing that, try "" overload */
5751 while (SvAMAGIC(msv)
5752 && (sv = AMG_CALLunary(msv, string_amg))
5756 && SvRV(msv) == SvRV(sv))
5761 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5765 /* this is a partially unrolled
5766 * sv_catsv_nomg(pat, msv);
5767 * that allows us to adjust code block indices if
5770 char *dst = SvPV_force_nomg(pat, dlen);
5772 if (SvUTF8(msv) && !SvUTF8(pat)) {
5773 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5774 sv_setpvn(pat, dst, dlen);
5777 sv_catsv_nomg(pat, msv);
5784 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5787 /* extract any code blocks within any embedded qr//'s */
5788 if (rx && SvTYPE(rx) == SVt_REGEXP
5789 && RX_ENGINE((REGEXP*)rx)->op_comp)
5792 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5793 if (ri->num_code_blocks) {
5795 /* the presence of an embedded qr// with code means
5796 * we should always recompile: the text of the
5797 * qr// may not have changed, but it may be a
5798 * different closure than last time */
5800 Renew(pRExC_state->code_blocks,
5801 pRExC_state->num_code_blocks + ri->num_code_blocks,
5802 struct reg_code_block);
5803 pRExC_state->num_code_blocks += ri->num_code_blocks;
5805 for (i=0; i < ri->num_code_blocks; i++) {
5806 struct reg_code_block *src, *dst;
5807 STRLEN offset = orig_patlen
5808 + ReANY((REGEXP *)rx)->pre_prefix;
5809 assert(n < pRExC_state->num_code_blocks);
5810 src = &ri->code_blocks[i];
5811 dst = &pRExC_state->code_blocks[n];
5812 dst->start = src->start + offset;
5813 dst->end = src->end + offset;
5814 dst->block = src->block;
5815 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5824 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5833 /* see if there are any run-time code blocks in the pattern.
5834 * False positives are allowed */
5837 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5838 char *pat, STRLEN plen)
5843 for (s = 0; s < plen; s++) {
5844 if (n < pRExC_state->num_code_blocks
5845 && s == pRExC_state->code_blocks[n].start)
5847 s = pRExC_state->code_blocks[n].end;
5851 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5853 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5855 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5862 /* Handle run-time code blocks. We will already have compiled any direct
5863 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5864 * copy of it, but with any literal code blocks blanked out and
5865 * appropriate chars escaped; then feed it into
5867 * eval "qr'modified_pattern'"
5871 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5875 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5877 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5878 * and merge them with any code blocks of the original regexp.
5880 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5881 * instead, just save the qr and return FALSE; this tells our caller that
5882 * the original pattern needs upgrading to utf8.
5886 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5887 char *pat, STRLEN plen)
5891 GET_RE_DEBUG_FLAGS_DECL;
5893 if (pRExC_state->runtime_code_qr) {
5894 /* this is the second time we've been called; this should
5895 * only happen if the main pattern got upgraded to utf8
5896 * during compilation; re-use the qr we compiled first time
5897 * round (which should be utf8 too)
5899 qr = pRExC_state->runtime_code_qr;
5900 pRExC_state->runtime_code_qr = NULL;
5901 assert(RExC_utf8 && SvUTF8(qr));
5907 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5911 /* determine how many extra chars we need for ' and \ escaping */
5912 for (s = 0; s < plen; s++) {
5913 if (pat[s] == '\'' || pat[s] == '\\')
5917 Newx(newpat, newlen, char);
5919 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5921 for (s = 0; s < plen; s++) {
5922 if (n < pRExC_state->num_code_blocks
5923 && s == pRExC_state->code_blocks[n].start)
5925 /* blank out literal code block */
5926 assert(pat[s] == '(');
5927 while (s <= pRExC_state->code_blocks[n].end) {
5935 if (pat[s] == '\'' || pat[s] == '\\')
5940 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5944 PerlIO_printf(Perl_debug_log,
5945 "%sre-parsing pattern for runtime code:%s %s\n",
5946 PL_colors[4],PL_colors[5],newpat);
5949 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5955 PUSHSTACKi(PERLSI_REQUIRE);
5956 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5957 * parsing qr''; normally only q'' does this. It also alters
5959 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5960 SvREFCNT_dec_NN(sv);
5965 SV * const errsv = ERRSV;
5966 if (SvTRUE_NN(errsv))
5968 Safefree(pRExC_state->code_blocks);
5969 /* use croak_sv ? */
5970 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
5973 assert(SvROK(qr_ref));
5975 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5976 /* the leaving below frees the tmp qr_ref.
5977 * Give qr a life of its own */
5985 if (!RExC_utf8 && SvUTF8(qr)) {
5986 /* first time through; the pattern got upgraded; save the
5987 * qr for the next time through */
5988 assert(!pRExC_state->runtime_code_qr);
5989 pRExC_state->runtime_code_qr = qr;
5994 /* extract any code blocks within the returned qr// */
5997 /* merge the main (r1) and run-time (r2) code blocks into one */
5999 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6000 struct reg_code_block *new_block, *dst;
6001 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6004 if (!r2->num_code_blocks) /* we guessed wrong */
6006 SvREFCNT_dec_NN(qr);
6011 r1->num_code_blocks + r2->num_code_blocks,
6012 struct reg_code_block);
6015 while ( i1 < r1->num_code_blocks
6016 || i2 < r2->num_code_blocks)
6018 struct reg_code_block *src;
6021 if (i1 == r1->num_code_blocks) {
6022 src = &r2->code_blocks[i2++];
6025 else if (i2 == r2->num_code_blocks)
6026 src = &r1->code_blocks[i1++];
6027 else if ( r1->code_blocks[i1].start
6028 < r2->code_blocks[i2].start)
6030 src = &r1->code_blocks[i1++];
6031 assert(src->end < r2->code_blocks[i2].start);
6034 assert( r1->code_blocks[i1].start
6035 > r2->code_blocks[i2].start);
6036 src = &r2->code_blocks[i2++];
6038 assert(src->end < r1->code_blocks[i1].start);
6041 assert(pat[src->start] == '(');
6042 assert(pat[src->end] == ')');
6043 dst->start = src->start;
6044 dst->end = src->end;
6045 dst->block = src->block;
6046 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6050 r1->num_code_blocks += r2->num_code_blocks;
6051 Safefree(r1->code_blocks);
6052 r1->code_blocks = new_block;
6055 SvREFCNT_dec_NN(qr);
6061 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6062 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6063 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6064 STRLEN longest_length, bool eol, bool meol)
6066 /* This is the common code for setting up the floating and fixed length
6067 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6068 * as to whether succeeded or not */
6073 if (! (longest_length
6074 || (eol /* Can't have SEOL and MULTI */
6075 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6077 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6078 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6083 /* copy the information about the longest from the reg_scan_data
6084 over to the program. */
6085 if (SvUTF8(sv_longest)) {
6086 *rx_utf8 = sv_longest;
6089 *rx_substr = sv_longest;
6092 /* end_shift is how many chars that must be matched that
6093 follow this item. We calculate it ahead of time as once the
6094 lookbehind offset is added in we lose the ability to correctly
6096 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6097 *rx_end_shift = ml - offset
6098 - longest_length + (SvTAIL(sv_longest) != 0)
6101 t = (eol/* Can't have SEOL and MULTI */
6102 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6103 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6109 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6110 * regular expression into internal code.
6111 * The pattern may be passed either as:
6112 * a list of SVs (patternp plus pat_count)
6113 * a list of OPs (expr)
6114 * If both are passed, the SV list is used, but the OP list indicates
6115 * which SVs are actually pre-compiled code blocks
6117 * The SVs in the list have magic and qr overloading applied to them (and
6118 * the list may be modified in-place with replacement SVs in the latter
6121 * If the pattern hasn't changed from old_re, then old_re will be
6124 * eng is the current engine. If that engine has an op_comp method, then
6125 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6126 * do the initial concatenation of arguments and pass on to the external
6129 * If is_bare_re is not null, set it to a boolean indicating whether the
6130 * arg list reduced (after overloading) to a single bare regex which has
6131 * been returned (i.e. /$qr/).
6133 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6135 * pm_flags contains the PMf_* flags, typically based on those from the
6136 * pm_flags field of the related PMOP. Currently we're only interested in
6137 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6139 * We can't allocate space until we know how big the compiled form will be,
6140 * but we can't compile it (and thus know how big it is) until we've got a
6141 * place to put the code. So we cheat: we compile it twice, once with code
6142 * generation turned off and size counting turned on, and once "for real".
6143 * This also means that we don't allocate space until we are sure that the
6144 * thing really will compile successfully, and we never have to move the
6145 * code and thus invalidate pointers into it. (Note that it has to be in
6146 * one piece because free() must be able to free it all.) [NB: not true in perl]
6148 * Beware that the optimization-preparation code in here knows about some
6149 * of the structure of the compiled regexp. [I'll say.]
6153 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6154 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6155 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6160 regexp_internal *ri;
6168 SV *code_blocksv = NULL;
6169 SV** new_patternp = patternp;
6171 /* these are all flags - maybe they should be turned
6172 * into a single int with different bit masks */
6173 I32 sawlookahead = 0;
6178 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6180 bool runtime_code = 0;
6182 RExC_state_t RExC_state;
6183 RExC_state_t * const pRExC_state = &RExC_state;
6184 #ifdef TRIE_STUDY_OPT
6186 RExC_state_t copyRExC_state;
6188 GET_RE_DEBUG_FLAGS_DECL;
6190 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6192 DEBUG_r(if (!PL_colorset) reginitcolors());
6194 #ifndef PERL_IN_XSUB_RE
6195 /* Initialize these here instead of as-needed, as is quick and avoids
6196 * having to test them each time otherwise */
6197 if (! PL_AboveLatin1) {
6198 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6199 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6200 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6201 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6202 PL_HasMultiCharFold =
6203 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6207 pRExC_state->code_blocks = NULL;
6208 pRExC_state->num_code_blocks = 0;
6211 *is_bare_re = FALSE;
6213 if (expr && (expr->op_type == OP_LIST ||
6214 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6215 /* allocate code_blocks if needed */
6219 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
6220 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6221 ncode++; /* count of DO blocks */
6223 pRExC_state->num_code_blocks = ncode;
6224 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6229 /* compile-time pattern with just OP_CONSTs and DO blocks */
6234 /* find how many CONSTs there are */
6237 if (expr->op_type == OP_CONST)
6240 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6241 if (o->op_type == OP_CONST)
6245 /* fake up an SV array */
6247 assert(!new_patternp);
6248 Newx(new_patternp, n, SV*);
6249 SAVEFREEPV(new_patternp);
6253 if (expr->op_type == OP_CONST)
6254 new_patternp[n] = cSVOPx_sv(expr);
6256 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6257 if (o->op_type == OP_CONST)
6258 new_patternp[n++] = cSVOPo_sv;
6263 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6264 "Assembling pattern from %d elements%s\n", pat_count,
6265 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6267 /* set expr to the first arg op */
6269 if (pRExC_state->num_code_blocks
6270 && expr->op_type != OP_CONST)
6272 expr = cLISTOPx(expr)->op_first;
6273 assert( expr->op_type == OP_PUSHMARK
6274 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6275 || expr->op_type == OP_PADRANGE);
6276 expr = expr->op_sibling;
6279 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6280 expr, &recompile, NULL);
6282 /* handle bare (possibly after overloading) regex: foo =~ $re */
6287 if (SvTYPE(re) == SVt_REGEXP) {
6291 Safefree(pRExC_state->code_blocks);
6292 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6293 "Precompiled pattern%s\n",
6294 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6300 exp = SvPV_nomg(pat, plen);
6302 if (!eng->op_comp) {
6303 if ((SvUTF8(pat) && IN_BYTES)
6304 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6306 /* make a temporary copy; either to convert to bytes,
6307 * or to avoid repeating get-magic / overloaded stringify */
6308 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6309 (IN_BYTES ? 0 : SvUTF8(pat)));
6311 Safefree(pRExC_state->code_blocks);
6312 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6315 /* ignore the utf8ness if the pattern is 0 length */
6316 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6317 RExC_uni_semantics = 0;
6318 RExC_contains_locale = 0;
6319 RExC_contains_i = 0;
6320 pRExC_state->runtime_code_qr = NULL;
6323 SV *dsv= sv_newmortal();
6324 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6325 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6326 PL_colors[4],PL_colors[5],s);
6330 /* we jump here if we upgrade the pattern to utf8 and have to
6333 if ((pm_flags & PMf_USE_RE_EVAL)
6334 /* this second condition covers the non-regex literal case,
6335 * i.e. $foo =~ '(?{})'. */
6336 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6338 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6340 /* return old regex if pattern hasn't changed */
6341 /* XXX: note in the below we have to check the flags as well as the
6344 * Things get a touch tricky as we have to compare the utf8 flag
6345 * independently from the compile flags. */
6349 && !!RX_UTF8(old_re) == !!RExC_utf8
6350 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6351 && RX_PRECOMP(old_re)
6352 && RX_PRELEN(old_re) == plen
6353 && memEQ(RX_PRECOMP(old_re), exp, plen)
6354 && !runtime_code /* with runtime code, always recompile */ )
6356 Safefree(pRExC_state->code_blocks);
6360 rx_flags = orig_rx_flags;
6362 if (rx_flags & PMf_FOLD) {
6363 RExC_contains_i = 1;
6365 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6367 /* Set to use unicode semantics if the pattern is in utf8 and has the
6368 * 'depends' charset specified, as it means unicode when utf8 */
6369 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6373 RExC_flags = rx_flags;
6374 RExC_pm_flags = pm_flags;
6377 if (TAINTING_get && TAINT_get)
6378 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6380 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6381 /* whoops, we have a non-utf8 pattern, whilst run-time code
6382 * got compiled as utf8. Try again with a utf8 pattern */
6383 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6384 pRExC_state->num_code_blocks);
6385 goto redo_first_pass;
6388 assert(!pRExC_state->runtime_code_qr);
6394 RExC_in_lookbehind = 0;
6395 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6397 RExC_override_recoding = 0;
6398 RExC_in_multi_char_class = 0;
6400 /* First pass: determine size, legality. */
6403 RExC_end = exp + plen;
6408 RExC_emit = (regnode *) &RExC_emit_dummy;
6409 RExC_whilem_seen = 0;
6410 RExC_open_parens = NULL;
6411 RExC_close_parens = NULL;
6413 RExC_paren_names = NULL;
6415 RExC_paren_name_list = NULL;
6417 RExC_recurse = NULL;
6418 RExC_study_chunk_recursed = NULL;
6419 RExC_study_chunk_recursed_bytes= 0;
6420 RExC_recurse_count = 0;
6421 pRExC_state->code_index = 0;
6423 #if 0 /* REGC() is (currently) a NOP at the first pass.
6424 * Clever compilers notice this and complain. --jhi */
6425 REGC((U8)REG_MAGIC, (char*)RExC_emit);
6428 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6430 RExC_lastparse=NULL;
6432 /* reg may croak on us, not giving us a chance to free
6433 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6434 need it to survive as long as the regexp (qr/(?{})/).
6435 We must check that code_blocksv is not already set, because we may
6436 have jumped back to restart the sizing pass. */
6437 if (pRExC_state->code_blocks && !code_blocksv) {
6438 code_blocksv = newSV_type(SVt_PV);
6439 SAVEFREESV(code_blocksv);
6440 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6441 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6443 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6444 /* It's possible to write a regexp in ascii that represents Unicode
6445 codepoints outside of the byte range, such as via \x{100}. If we
6446 detect such a sequence we have to convert the entire pattern to utf8
6447 and then recompile, as our sizing calculation will have been based
6448 on 1 byte == 1 character, but we will need to use utf8 to encode
6449 at least some part of the pattern, and therefore must convert the whole
6452 if (flags & RESTART_UTF8) {
6453 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6454 pRExC_state->num_code_blocks);
6455 goto redo_first_pass;
6457 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6460 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6463 PerlIO_printf(Perl_debug_log,
6464 "Required size %"IVdf" nodes\n"
6465 "Starting second pass (creation)\n",
6468 RExC_lastparse=NULL;
6471 /* The first pass could have found things that force Unicode semantics */
6472 if ((RExC_utf8 || RExC_uni_semantics)
6473 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6475 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6478 /* Small enough for pointer-storage convention?
6479 If extralen==0, this means that we will not need long jumps. */
6480 if (RExC_size >= 0x10000L && RExC_extralen)
6481 RExC_size += RExC_extralen;
6484 if (RExC_whilem_seen > 15)
6485 RExC_whilem_seen = 15;
6487 /* Allocate space and zero-initialize. Note, the two step process
6488 of zeroing when in debug mode, thus anything assigned has to
6489 happen after that */
6490 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6492 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6493 char, regexp_internal);
6494 if ( r == NULL || ri == NULL )
6495 FAIL("Regexp out of space");
6497 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6498 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6501 /* bulk initialize base fields with 0. */
6502 Zero(ri, sizeof(regexp_internal), char);
6505 /* non-zero initialization begins here */
6508 r->extflags = rx_flags;
6509 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6511 if (pm_flags & PMf_IS_QR) {
6512 ri->code_blocks = pRExC_state->code_blocks;
6513 ri->num_code_blocks = pRExC_state->num_code_blocks;
6518 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6519 if (pRExC_state->code_blocks[n].src_regex)
6520 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6521 SAVEFREEPV(pRExC_state->code_blocks);
6525 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6526 bool has_charset = (get_regex_charset(r->extflags)
6527 != REGEX_DEPENDS_CHARSET);
6529 /* The caret is output if there are any defaults: if not all the STD
6530 * flags are set, or if no character set specifier is needed */
6532 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6534 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6535 == REG_RUN_ON_COMMENT_SEEN);
6536 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6537 >> RXf_PMf_STD_PMMOD_SHIFT);
6538 const char *fptr = STD_PAT_MODS; /*"msix"*/
6540 /* Allocate for the worst case, which is all the std flags are turned
6541 * on. If more precision is desired, we could do a population count of
6542 * the flags set. This could be done with a small lookup table, or by
6543 * shifting, masking and adding, or even, when available, assembly
6544 * language for a machine-language population count.
6545 * We never output a minus, as all those are defaults, so are
6546 * covered by the caret */
6547 const STRLEN wraplen = plen + has_p + has_runon
6548 + has_default /* If needs a caret */
6550 /* If needs a character set specifier */
6551 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6552 + (sizeof(STD_PAT_MODS) - 1)
6553 + (sizeof("(?:)") - 1);
6555 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6556 r->xpv_len_u.xpvlenu_pv = p;
6558 SvFLAGS(rx) |= SVf_UTF8;
6561 /* If a default, cover it using the caret */
6563 *p++= DEFAULT_PAT_MOD;
6567 const char* const name = get_regex_charset_name(r->extflags, &len);
6568 Copy(name, p, len, char);
6572 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6575 while((ch = *fptr++)) {
6583 Copy(RExC_precomp, p, plen, char);
6584 assert ((RX_WRAPPED(rx) - p) < 16);
6585 r->pre_prefix = p - RX_WRAPPED(rx);
6591 SvCUR_set(rx, p - RX_WRAPPED(rx));
6595 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6597 /* setup various meta data about recursion, this all requires
6598 * RExC_npar to be correctly set, and a bit later on we clear it */
6599 if (RExC_seen & REG_RECURSE_SEEN) {
6600 Newxz(RExC_open_parens, RExC_npar,regnode *);
6601 SAVEFREEPV(RExC_open_parens);
6602 Newxz(RExC_close_parens,RExC_npar,regnode *);
6603 SAVEFREEPV(RExC_close_parens);
6605 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6606 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6607 * So its 1 if there are no parens. */
6608 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6609 ((RExC_npar & 0x07) != 0);
6610 Newx(RExC_study_chunk_recursed,
6611 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6612 SAVEFREEPV(RExC_study_chunk_recursed);
6615 /* Useful during FAIL. */
6616 #ifdef RE_TRACK_PATTERN_OFFSETS
6617 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6618 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6619 "%s %"UVuf" bytes for offset annotations.\n",
6620 ri->u.offsets ? "Got" : "Couldn't get",
6621 (UV)((2*RExC_size+1) * sizeof(U32))));
6623 SetProgLen(ri,RExC_size);
6627 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
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, NULL);
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->substrs->check_ix = 0;
6960 r->check_end_shift = r->anchored_end_shift;
6961 r->check_substr = r->anchored_substr;
6962 r->check_utf8 = r->anchored_utf8;
6963 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6964 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
6965 r->intflags |= PREGf_NOSCAN;
6968 r->substrs->check_ix = 1;
6969 r->check_end_shift = r->float_end_shift;
6970 r->check_substr = r->float_substr;
6971 r->check_utf8 = r->float_utf8;
6972 r->check_offset_min = r->float_min_offset;
6973 r->check_offset_max = r->float_max_offset;
6975 if ((r->check_substr || r->check_utf8) ) {
6976 r->extflags |= RXf_USE_INTUIT;
6977 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6978 r->extflags |= RXf_INTUIT_TAIL;
6980 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
6982 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6983 if ( (STRLEN)minlen < longest_float_length )
6984 minlen= longest_float_length;
6985 if ( (STRLEN)minlen < longest_fixed_length )
6986 minlen= longest_fixed_length;
6990 /* Several toplevels. Best we can is to set minlen. */
6992 regnode_ssc ch_class;
6993 SSize_t last_close = 0;
6995 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6997 scan = ri->program + 1;
6998 ssc_init(pRExC_state, &ch_class);
6999 data.start_class = &ch_class;
7000 data.last_closep = &last_close;
7003 minlen = study_chunk(pRExC_state,
7004 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7005 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7006 ? SCF_TRIE_DOING_RESTUDY
7010 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7012 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7013 = r->float_substr = r->float_utf8 = NULL;
7015 if (! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
7016 && ! ssc_is_anything(data.start_class))
7018 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7020 ssc_finalize(pRExC_state, data.start_class);
7022 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7023 StructCopy(data.start_class,
7024 (regnode_ssc*)RExC_rxi->data->data[n],
7026 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7027 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7028 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7029 regprop(r, sv, (regnode*)data.start_class, NULL);
7030 PerlIO_printf(Perl_debug_log,
7031 "synthetic stclass \"%s\".\n",
7032 SvPVX_const(sv));});
7033 data.start_class = NULL;
7037 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7038 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7039 r->maxlen = REG_INFTY;
7042 r->maxlen = RExC_maxlen;
7045 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7046 the "real" pattern. */
7048 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%ld\n",
7049 (IV)minlen, (IV)r->minlen, RExC_maxlen);
7051 r->minlenret = minlen;
7052 if (r->minlen < minlen)
7055 if (RExC_seen & REG_GPOS_SEEN)
7056 r->intflags |= PREGf_GPOS_SEEN;
7057 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7058 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7060 if (pRExC_state->num_code_blocks)
7061 r->extflags |= RXf_EVAL_SEEN;
7062 if (RExC_seen & REG_CANY_SEEN)
7063 r->intflags |= PREGf_CANY_SEEN;
7064 if (RExC_seen & REG_VERBARG_SEEN)
7066 r->intflags |= PREGf_VERBARG_SEEN;
7067 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7069 if (RExC_seen & REG_CUTGROUP_SEEN)
7070 r->intflags |= PREGf_CUTGROUP_SEEN;
7071 if (pm_flags & PMf_USE_RE_EVAL)
7072 r->intflags |= PREGf_USE_RE_EVAL;
7073 if (RExC_paren_names)
7074 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7076 RXp_PAREN_NAMES(r) = NULL;
7078 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7079 * so it can be used in pp.c */
7080 if (r->intflags & PREGf_ANCH)
7081 r->extflags |= RXf_IS_ANCHORED;
7085 /* this is used to identify "special" patterns that might result
7086 * in Perl NOT calling the regex engine and instead doing the match "itself",
7087 * particularly special cases in split//. By having the regex compiler
7088 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7089 * we avoid weird issues with equivalent patterns resulting in different behavior,
7090 * AND we allow non Perl engines to get the same optimizations by the setting the
7091 * flags appropriately - Yves */
7092 regnode *first = ri->program + 1;
7094 regnode *next = NEXTOPER(first);
7097 if (PL_regkind[fop] == NOTHING && nop == END)
7098 r->extflags |= RXf_NULL;
7099 else if (PL_regkind[fop] == BOL && nop == END)
7100 r->extflags |= RXf_START_ONLY;
7101 else if (fop == PLUS
7102 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7103 && OP(regnext(first)) == END)
7104 r->extflags |= RXf_WHITE;
7105 else if ( r->extflags & RXf_SPLIT
7107 && STR_LEN(first) == 1
7108 && *(STRING(first)) == ' '
7109 && OP(regnext(first)) == END )
7110 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7114 if (RExC_contains_locale) {
7115 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7119 if (RExC_paren_names) {
7120 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7121 ri->data->data[ri->name_list_idx]
7122 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7125 ri->name_list_idx = 0;
7127 if (RExC_recurse_count) {
7128 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7129 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7130 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7133 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7134 /* assume we don't need to swap parens around before we match */
7138 PerlIO_printf(Perl_debug_log,"Final program:\n");
7141 #ifdef RE_TRACK_PATTERN_OFFSETS
7142 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7143 const STRLEN len = ri->u.offsets[0];
7145 GET_RE_DEBUG_FLAGS_DECL;
7146 PerlIO_printf(Perl_debug_log,
7147 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7148 for (i = 1; i <= len; i++) {
7149 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7150 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7151 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7153 PerlIO_printf(Perl_debug_log, "\n");
7158 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7159 * by setting the regexp SV to readonly-only instead. If the
7160 * pattern's been recompiled, the USEDness should remain. */
7161 if (old_re && SvREADONLY(old_re))
7169 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7172 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7174 PERL_UNUSED_ARG(value);
7176 if (flags & RXapif_FETCH) {
7177 return reg_named_buff_fetch(rx, key, flags);
7178 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7179 Perl_croak_no_modify();
7181 } else if (flags & RXapif_EXISTS) {
7182 return reg_named_buff_exists(rx, key, flags)
7185 } else if (flags & RXapif_REGNAMES) {
7186 return reg_named_buff_all(rx, flags);
7187 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7188 return reg_named_buff_scalar(rx, flags);
7190 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7196 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7199 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7200 PERL_UNUSED_ARG(lastkey);
7202 if (flags & RXapif_FIRSTKEY)
7203 return reg_named_buff_firstkey(rx, flags);
7204 else if (flags & RXapif_NEXTKEY)
7205 return reg_named_buff_nextkey(rx, flags);
7207 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7214 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7217 AV *retarray = NULL;
7219 struct regexp *const rx = ReANY(r);
7221 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7223 if (flags & RXapif_ALL)
7226 if (rx && RXp_PAREN_NAMES(rx)) {
7227 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7230 SV* sv_dat=HeVAL(he_str);
7231 I32 *nums=(I32*)SvPVX(sv_dat);
7232 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7233 if ((I32)(rx->nparens) >= nums[i]
7234 && rx->offs[nums[i]].start != -1
7235 && rx->offs[nums[i]].end != -1)
7238 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7243 ret = newSVsv(&PL_sv_undef);
7246 av_push(retarray, ret);
7249 return newRV_noinc(MUTABLE_SV(retarray));
7256 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7259 struct regexp *const rx = ReANY(r);
7261 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7263 if (rx && RXp_PAREN_NAMES(rx)) {
7264 if (flags & RXapif_ALL) {
7265 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7267 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7269 SvREFCNT_dec_NN(sv);
7281 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7283 struct regexp *const rx = ReANY(r);
7285 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7287 if ( rx && RXp_PAREN_NAMES(rx) ) {
7288 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7290 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7297 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7299 struct regexp *const rx = ReANY(r);
7300 GET_RE_DEBUG_FLAGS_DECL;
7302 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7304 if (rx && RXp_PAREN_NAMES(rx)) {
7305 HV *hv = RXp_PAREN_NAMES(rx);
7307 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7310 SV* sv_dat = HeVAL(temphe);
7311 I32 *nums = (I32*)SvPVX(sv_dat);
7312 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7313 if ((I32)(rx->lastparen) >= nums[i] &&
7314 rx->offs[nums[i]].start != -1 &&
7315 rx->offs[nums[i]].end != -1)
7321 if (parno || flags & RXapif_ALL) {
7322 return newSVhek(HeKEY_hek(temphe));
7330 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7335 struct regexp *const rx = ReANY(r);
7337 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7339 if (rx && RXp_PAREN_NAMES(rx)) {
7340 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7341 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7342 } else if (flags & RXapif_ONE) {
7343 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7344 av = MUTABLE_AV(SvRV(ret));
7345 length = av_tindex(av);
7346 SvREFCNT_dec_NN(ret);
7347 return newSViv(length + 1);
7349 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7354 return &PL_sv_undef;
7358 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7360 struct regexp *const rx = ReANY(r);
7363 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7365 if (rx && RXp_PAREN_NAMES(rx)) {
7366 HV *hv= RXp_PAREN_NAMES(rx);
7368 (void)hv_iterinit(hv);
7369 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7372 SV* sv_dat = HeVAL(temphe);
7373 I32 *nums = (I32*)SvPVX(sv_dat);
7374 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7375 if ((I32)(rx->lastparen) >= nums[i] &&
7376 rx->offs[nums[i]].start != -1 &&
7377 rx->offs[nums[i]].end != -1)
7383 if (parno || flags & RXapif_ALL) {
7384 av_push(av, newSVhek(HeKEY_hek(temphe)));
7389 return newRV_noinc(MUTABLE_SV(av));
7393 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7396 struct regexp *const rx = ReANY(r);
7402 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7404 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7405 || n == RX_BUFF_IDX_CARET_FULLMATCH
7406 || n == RX_BUFF_IDX_CARET_POSTMATCH
7409 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7411 /* on something like
7414 * the KEEPCOPY is set on the PMOP rather than the regex */
7415 if (PL_curpm && r == PM_GETRE(PL_curpm))
7416 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7425 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7426 /* no need to distinguish between them any more */
7427 n = RX_BUFF_IDX_FULLMATCH;
7429 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7430 && rx->offs[0].start != -1)
7432 /* $`, ${^PREMATCH} */
7433 i = rx->offs[0].start;
7437 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7438 && rx->offs[0].end != -1)
7440 /* $', ${^POSTMATCH} */
7441 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7442 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7445 if ( 0 <= n && n <= (I32)rx->nparens &&
7446 (s1 = rx->offs[n].start) != -1 &&
7447 (t1 = rx->offs[n].end) != -1)
7449 /* $&, ${^MATCH}, $1 ... */
7451 s = rx->subbeg + s1 - rx->suboffset;
7456 assert(s >= rx->subbeg);
7457 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7459 #ifdef NO_TAINT_SUPPORT
7460 sv_setpvn(sv, s, i);
7462 const int oldtainted = TAINT_get;
7464 sv_setpvn(sv, s, i);
7465 TAINT_set(oldtainted);
7467 if ( (rx->intflags & PREGf_CANY_SEEN)
7468 ? (RXp_MATCH_UTF8(rx)
7469 && (!i || is_utf8_string((U8*)s, i)))
7470 : (RXp_MATCH_UTF8(rx)) )
7477 if (RXp_MATCH_TAINTED(rx)) {
7478 if (SvTYPE(sv) >= SVt_PVMG) {
7479 MAGIC* const mg = SvMAGIC(sv);
7482 SvMAGIC_set(sv, mg->mg_moremagic);
7484 if ((mgt = SvMAGIC(sv))) {
7485 mg->mg_moremagic = mgt;
7486 SvMAGIC_set(sv, mg);
7497 sv_setsv(sv,&PL_sv_undef);
7503 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7504 SV const * const value)
7506 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7508 PERL_UNUSED_ARG(rx);
7509 PERL_UNUSED_ARG(paren);
7510 PERL_UNUSED_ARG(value);
7513 Perl_croak_no_modify();
7517 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7520 struct regexp *const rx = ReANY(r);
7524 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7526 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7527 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7528 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7531 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7533 /* on something like
7536 * the KEEPCOPY is set on the PMOP rather than the regex */
7537 if (PL_curpm && r == PM_GETRE(PL_curpm))
7538 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7544 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7546 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7547 case RX_BUFF_IDX_PREMATCH: /* $` */
7548 if (rx->offs[0].start != -1) {
7549 i = rx->offs[0].start;
7558 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7559 case RX_BUFF_IDX_POSTMATCH: /* $' */
7560 if (rx->offs[0].end != -1) {
7561 i = rx->sublen - rx->offs[0].end;
7563 s1 = rx->offs[0].end;
7570 default: /* $& / ${^MATCH}, $1, $2, ... */
7571 if (paren <= (I32)rx->nparens &&
7572 (s1 = rx->offs[paren].start) != -1 &&
7573 (t1 = rx->offs[paren].end) != -1)
7579 if (ckWARN(WARN_UNINITIALIZED))
7580 report_uninit((const SV *)sv);
7585 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7586 const char * const s = rx->subbeg - rx->suboffset + s1;
7591 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7598 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7600 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7601 PERL_UNUSED_ARG(rx);
7605 return newSVpvs("Regexp");
7608 /* Scans the name of a named buffer from the pattern.
7609 * If flags is REG_RSN_RETURN_NULL returns null.
7610 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7611 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7612 * to the parsed name as looked up in the RExC_paren_names hash.
7613 * If there is an error throws a vFAIL().. type exception.
7616 #define REG_RSN_RETURN_NULL 0
7617 #define REG_RSN_RETURN_NAME 1
7618 #define REG_RSN_RETURN_DATA 2
7621 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7623 char *name_start = RExC_parse;
7625 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7627 assert (RExC_parse <= RExC_end);
7628 if (RExC_parse == RExC_end) NOOP;
7629 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7630 /* skip IDFIRST by using do...while */
7633 RExC_parse += UTF8SKIP(RExC_parse);
7634 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7638 } while (isWORDCHAR(*RExC_parse));
7640 RExC_parse++; /* so the <- from the vFAIL is after the offending
7642 vFAIL("Group name must start with a non-digit word character");
7646 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7647 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7648 if ( flags == REG_RSN_RETURN_NAME)
7650 else if (flags==REG_RSN_RETURN_DATA) {
7653 if ( ! sv_name ) /* should not happen*/
7654 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7655 if (RExC_paren_names)
7656 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7658 sv_dat = HeVAL(he_str);
7660 vFAIL("Reference to nonexistent named group");
7664 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7665 (unsigned long) flags);
7667 assert(0); /* NOT REACHED */
7672 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7673 int rem=(int)(RExC_end - RExC_parse); \
7682 if (RExC_lastparse!=RExC_parse) \
7683 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7686 iscut ? "..." : "<" \
7689 PerlIO_printf(Perl_debug_log,"%16s",""); \
7692 num = RExC_size + 1; \
7694 num=REG_NODE_NUM(RExC_emit); \
7695 if (RExC_lastnum!=num) \
7696 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7698 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7699 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7700 (int)((depth*2)), "", \
7704 RExC_lastparse=RExC_parse; \
7709 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7710 DEBUG_PARSE_MSG((funcname)); \
7711 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7713 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7714 DEBUG_PARSE_MSG((funcname)); \
7715 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7718 /* This section of code defines the inversion list object and its methods. The
7719 * interfaces are highly subject to change, so as much as possible is static to
7720 * this file. An inversion list is here implemented as a malloc'd C UV array
7721 * as an SVt_INVLIST scalar.
7723 * An inversion list for Unicode is an array of code points, sorted by ordinal
7724 * number. The zeroth element is the first code point in the list. The 1th
7725 * element is the first element beyond that not in the list. In other words,
7726 * the first range is
7727 * invlist[0]..(invlist[1]-1)
7728 * The other ranges follow. Thus every element whose index is divisible by two
7729 * marks the beginning of a range that is in the list, and every element not
7730 * divisible by two marks the beginning of a range not in the list. A single
7731 * element inversion list that contains the single code point N generally
7732 * consists of two elements
7735 * (The exception is when N is the highest representable value on the
7736 * machine, in which case the list containing just it would be a single
7737 * element, itself. By extension, if the last range in the list extends to
7738 * infinity, then the first element of that range will be in the inversion list
7739 * at a position that is divisible by two, and is the final element in the
7741 * Taking the complement (inverting) an inversion list is quite simple, if the
7742 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7743 * This implementation reserves an element at the beginning of each inversion
7744 * list to always contain 0; there is an additional flag in the header which
7745 * indicates if the list begins at the 0, or is offset to begin at the next
7748 * More about inversion lists can be found in "Unicode Demystified"
7749 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7750 * More will be coming when functionality is added later.
7752 * The inversion list data structure is currently implemented as an SV pointing
7753 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7754 * array of UV whose memory management is automatically handled by the existing
7755 * facilities for SV's.
7757 * Some of the methods should always be private to the implementation, and some
7758 * should eventually be made public */
7760 /* The header definitions are in F<inline_invlist.c> */
7762 PERL_STATIC_INLINE UV*
7763 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7765 /* Returns a pointer to the first element in the inversion list's array.
7766 * This is called upon initialization of an inversion list. Where the
7767 * array begins depends on whether the list has the code point U+0000 in it
7768 * or not. The other parameter tells it whether the code that follows this
7769 * call is about to put a 0 in the inversion list or not. The first
7770 * element is either the element reserved for 0, if TRUE, or the element
7771 * after it, if FALSE */
7773 bool* offset = get_invlist_offset_addr(invlist);
7774 UV* zero_addr = (UV *) SvPVX(invlist);
7776 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7779 assert(! _invlist_len(invlist));
7783 /* 1^1 = 0; 1^0 = 1 */
7784 *offset = 1 ^ will_have_0;
7785 return zero_addr + *offset;
7788 PERL_STATIC_INLINE UV*
7789 S_invlist_array(pTHX_ SV* const invlist)
7791 /* Returns the pointer to the inversion list's array. Every time the
7792 * length changes, this needs to be called in case malloc or realloc moved
7795 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7797 /* Must not be empty. If these fail, you probably didn't check for <len>
7798 * being non-zero before trying to get the array */
7799 assert(_invlist_len(invlist));
7801 /* The very first element always contains zero, The array begins either
7802 * there, or if the inversion list is offset, at the element after it.
7803 * The offset header field determines which; it contains 0 or 1 to indicate
7804 * how much additionally to add */
7805 assert(0 == *(SvPVX(invlist)));
7806 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7809 PERL_STATIC_INLINE void
7810 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7812 /* Sets the current number of elements stored in the inversion list.
7813 * Updates SvCUR correspondingly */
7815 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7817 assert(SvTYPE(invlist) == SVt_INVLIST);
7822 : TO_INTERNAL_SIZE(len + offset));
7823 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7826 PERL_STATIC_INLINE IV*
7827 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7829 /* Return the address of the IV that is reserved to hold the cached index
7832 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7834 assert(SvTYPE(invlist) == SVt_INVLIST);
7836 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7839 PERL_STATIC_INLINE IV
7840 S_invlist_previous_index(pTHX_ SV* const invlist)
7842 /* Returns cached index of previous search */
7844 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7846 return *get_invlist_previous_index_addr(invlist);
7849 PERL_STATIC_INLINE void
7850 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7852 /* Caches <index> for later retrieval */
7854 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7856 assert(index == 0 || index < (int) _invlist_len(invlist));
7858 *get_invlist_previous_index_addr(invlist) = index;
7861 PERL_STATIC_INLINE UV
7862 S_invlist_max(pTHX_ SV* const invlist)
7864 /* Returns the maximum number of elements storable in the inversion list's
7865 * array, without having to realloc() */
7867 PERL_ARGS_ASSERT_INVLIST_MAX;
7869 assert(SvTYPE(invlist) == SVt_INVLIST);
7871 /* Assumes worst case, in which the 0 element is not counted in the
7872 * inversion list, so subtracts 1 for that */
7873 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7874 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7875 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7878 #ifndef PERL_IN_XSUB_RE
7880 Perl__new_invlist(pTHX_ IV initial_size)
7883 /* Return a pointer to a newly constructed inversion list, with enough
7884 * space to store 'initial_size' elements. If that number is negative, a
7885 * system default is used instead */
7889 if (initial_size < 0) {
7893 /* Allocate the initial space */
7894 new_list = newSV_type(SVt_INVLIST);
7896 /* First 1 is in case the zero element isn't in the list; second 1 is for
7898 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7899 invlist_set_len(new_list, 0, 0);
7901 /* Force iterinit() to be used to get iteration to work */
7902 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7904 *get_invlist_previous_index_addr(new_list) = 0;
7910 Perl__new_invlist_C_array(pTHX_ const UV* const list)
7912 /* Return a pointer to a newly constructed inversion list, initialized to
7913 * point to <list>, which has to be in the exact correct inversion list
7914 * form, including internal fields. Thus this is a dangerous routine that
7915 * should not be used in the wrong hands. The passed in 'list' contains
7916 * several header fields at the beginning that are not part of the
7917 * inversion list body proper */
7919 const STRLEN length = (STRLEN) list[0];
7920 const UV version_id = list[1];
7921 const bool offset = cBOOL(list[2]);
7922 #define HEADER_LENGTH 3
7923 /* If any of the above changes in any way, you must change HEADER_LENGTH
7924 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7925 * perl -E 'say int(rand 2**31-1)'
7927 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7928 data structure type, so that one being
7929 passed in can be validated to be an
7930 inversion list of the correct vintage.
7933 SV* invlist = newSV_type(SVt_INVLIST);
7935 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7937 if (version_id != INVLIST_VERSION_ID) {
7938 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7941 /* The generated array passed in includes header elements that aren't part
7942 * of the list proper, so start it just after them */
7943 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7945 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7946 shouldn't touch it */
7948 *(get_invlist_offset_addr(invlist)) = offset;
7950 /* The 'length' passed to us is the physical number of elements in the
7951 * inversion list. But if there is an offset the logical number is one
7953 invlist_set_len(invlist, length - offset, offset);
7955 invlist_set_previous_index(invlist, 0);
7957 /* Initialize the iteration pointer. */
7958 invlist_iterfinish(invlist);
7960 SvREADONLY_on(invlist);
7964 #endif /* ifndef PERL_IN_XSUB_RE */
7967 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7969 /* Grow the maximum size of an inversion list */
7971 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7973 assert(SvTYPE(invlist) == SVt_INVLIST);
7975 /* Add one to account for the zero element at the beginning which may not
7976 * be counted by the calling parameters */
7977 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7980 PERL_STATIC_INLINE void
7981 S_invlist_trim(pTHX_ SV* const invlist)
7983 PERL_ARGS_ASSERT_INVLIST_TRIM;
7985 assert(SvTYPE(invlist) == SVt_INVLIST);
7987 /* Change the length of the inversion list to how many entries it currently
7989 SvPV_shrink_to_cur((SV *) invlist);
7993 S__append_range_to_invlist(pTHX_ SV* const invlist,
7994 const UV start, const UV end)
7996 /* Subject to change or removal. Append the range from 'start' to 'end' at
7997 * the end of the inversion list. The range must be above any existing
8001 UV max = invlist_max(invlist);
8002 UV len = _invlist_len(invlist);
8005 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8007 if (len == 0) { /* Empty lists must be initialized */
8008 offset = start != 0;
8009 array = _invlist_array_init(invlist, ! offset);
8012 /* Here, the existing list is non-empty. The current max entry in the
8013 * list is generally the first value not in the set, except when the
8014 * set extends to the end of permissible values, in which case it is
8015 * the first entry in that final set, and so this call is an attempt to
8016 * append out-of-order */
8018 UV final_element = len - 1;
8019 array = invlist_array(invlist);
8020 if (array[final_element] > start
8021 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8023 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",
8024 array[final_element], start,
8025 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8028 /* Here, it is a legal append. If the new range begins with the first
8029 * value not in the set, it is extending the set, so the new first
8030 * value not in the set is one greater than the newly extended range.
8032 offset = *get_invlist_offset_addr(invlist);
8033 if (array[final_element] == start) {
8034 if (end != UV_MAX) {
8035 array[final_element] = end + 1;
8038 /* But if the end is the maximum representable on the machine,
8039 * just let the range that this would extend to have no end */
8040 invlist_set_len(invlist, len - 1, offset);
8046 /* Here the new range doesn't extend any existing set. Add it */
8048 len += 2; /* Includes an element each for the start and end of range */
8050 /* If wll overflow the existing space, extend, which may cause the array to
8053 invlist_extend(invlist, len);
8055 /* Have to set len here to avoid assert failure in invlist_array() */
8056 invlist_set_len(invlist, len, offset);
8058 array = invlist_array(invlist);
8061 invlist_set_len(invlist, len, offset);
8064 /* The next item on the list starts the range, the one after that is
8065 * one past the new range. */
8066 array[len - 2] = start;
8067 if (end != UV_MAX) {
8068 array[len - 1] = end + 1;
8071 /* But if the end is the maximum representable on the machine, just let
8072 * the range have no end */
8073 invlist_set_len(invlist, len - 1, offset);
8077 #ifndef PERL_IN_XSUB_RE
8080 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
8082 /* Searches the inversion list for the entry that contains the input code
8083 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8084 * return value is the index into the list's array of the range that
8089 IV high = _invlist_len(invlist);
8090 const IV highest_element = high - 1;
8093 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8095 /* If list is empty, return failure. */
8100 /* (We can't get the array unless we know the list is non-empty) */
8101 array = invlist_array(invlist);
8103 mid = invlist_previous_index(invlist);
8104 assert(mid >=0 && mid <= highest_element);
8106 /* <mid> contains the cache of the result of the previous call to this
8107 * function (0 the first time). See if this call is for the same result,
8108 * or if it is for mid-1. This is under the theory that calls to this
8109 * function will often be for related code points that are near each other.
8110 * And benchmarks show that caching gives better results. We also test
8111 * here if the code point is within the bounds of the list. These tests
8112 * replace others that would have had to be made anyway to make sure that
8113 * the array bounds were not exceeded, and these give us extra information
8114 * at the same time */
8115 if (cp >= array[mid]) {
8116 if (cp >= array[highest_element]) {
8117 return highest_element;
8120 /* Here, array[mid] <= cp < array[highest_element]. This means that
8121 * the final element is not the answer, so can exclude it; it also
8122 * means that <mid> is not the final element, so can refer to 'mid + 1'
8124 if (cp < array[mid + 1]) {
8130 else { /* cp < aray[mid] */
8131 if (cp < array[0]) { /* Fail if outside the array */
8135 if (cp >= array[mid - 1]) {
8140 /* Binary search. What we are looking for is <i> such that
8141 * array[i] <= cp < array[i+1]
8142 * The loop below converges on the i+1. Note that there may not be an
8143 * (i+1)th element in the array, and things work nonetheless */
8144 while (low < high) {
8145 mid = (low + high) / 2;
8146 assert(mid <= highest_element);
8147 if (array[mid] <= cp) { /* cp >= array[mid] */
8150 /* We could do this extra test to exit the loop early.
8151 if (cp < array[low]) {
8156 else { /* cp < array[mid] */
8163 invlist_set_previous_index(invlist, high);
8168 Perl__invlist_populate_swatch(pTHX_ SV* const invlist,
8169 const UV start, const UV end, U8* swatch)
8171 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8172 * but is used when the swash has an inversion list. This makes this much
8173 * faster, as it uses a binary search instead of a linear one. This is
8174 * intimately tied to that function, and perhaps should be in utf8.c,
8175 * except it is intimately tied to inversion lists as well. It assumes
8176 * that <swatch> is all 0's on input */
8179 const IV len = _invlist_len(invlist);
8183 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8185 if (len == 0) { /* Empty inversion list */
8189 array = invlist_array(invlist);
8191 /* Find which element it is */
8192 i = _invlist_search(invlist, start);
8194 /* We populate from <start> to <end> */
8195 while (current < end) {
8198 /* The inversion list gives the results for every possible code point
8199 * after the first one in the list. Only those ranges whose index is
8200 * even are ones that the inversion list matches. For the odd ones,
8201 * and if the initial code point is not in the list, we have to skip
8202 * forward to the next element */
8203 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8205 if (i >= len) { /* Finished if beyond the end of the array */
8209 if (current >= end) { /* Finished if beyond the end of what we
8211 if (LIKELY(end < UV_MAX)) {
8215 /* We get here when the upper bound is the maximum
8216 * representable on the machine, and we are looking for just
8217 * that code point. Have to special case it */
8219 goto join_end_of_list;
8222 assert(current >= start);
8224 /* The current range ends one below the next one, except don't go past
8227 upper = (i < len && array[i] < end) ? array[i] : end;
8229 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8230 * for each code point in it */
8231 for (; current < upper; current++) {
8232 const STRLEN offset = (STRLEN)(current - start);
8233 swatch[offset >> 3] |= 1 << (offset & 7);
8238 /* Quit if at the end of the list */
8241 /* But first, have to deal with the highest possible code point on
8242 * the platform. The previous code assumes that <end> is one
8243 * beyond where we want to populate, but that is impossible at the
8244 * platform's infinity, so have to handle it specially */
8245 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8247 const STRLEN offset = (STRLEN)(end - start);
8248 swatch[offset >> 3] |= 1 << (offset & 7);
8253 /* Advance to the next range, which will be for code points not in the
8262 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8263 const bool complement_b, SV** output)
8265 /* Take the union of two inversion lists and point <output> to it. *output
8266 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8267 * the reference count to that list will be decremented if not already a
8268 * temporary (mortal); otherwise *output will be made correspondingly
8269 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8270 * second list is returned. If <complement_b> is TRUE, the union is taken
8271 * of the complement (inversion) of <b> instead of b itself.
8273 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8274 * Richard Gillam, published by Addison-Wesley, and explained at some
8275 * length there. The preface says to incorporate its examples into your
8276 * code at your own risk.
8278 * The algorithm is like a merge sort.
8280 * XXX A potential performance improvement is to keep track as we go along
8281 * if only one of the inputs contributes to the result, meaning the other
8282 * is a subset of that one. In that case, we can skip the final copy and
8283 * return the larger of the input lists, but then outside code might need
8284 * to keep track of whether to free the input list or not */
8286 const UV* array_a; /* a's array */
8288 UV len_a; /* length of a's array */
8291 SV* u; /* the resulting union */
8295 UV i_a = 0; /* current index into a's array */
8299 /* running count, as explained in the algorithm source book; items are
8300 * stopped accumulating and are output when the count changes to/from 0.
8301 * The count is incremented when we start a range that's in the set, and
8302 * decremented when we start a range that's not in the set. So its range
8303 * is 0 to 2. Only when the count is zero is something not in the set.
8307 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8310 /* If either one is empty, the union is the other one */
8311 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8312 bool make_temp = FALSE; /* Should we mortalize the result? */
8316 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8322 *output = invlist_clone(b);
8324 _invlist_invert(*output);
8326 } /* else *output already = b; */
8329 sv_2mortal(*output);
8333 else if ((len_b = _invlist_len(b)) == 0) {
8334 bool make_temp = FALSE;
8336 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8341 /* The complement of an empty list is a list that has everything in it,
8342 * so the union with <a> includes everything too */
8345 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8349 *output = _new_invlist(1);
8350 _append_range_to_invlist(*output, 0, UV_MAX);
8352 else if (*output != a) {
8353 *output = invlist_clone(a);
8355 /* else *output already = a; */
8358 sv_2mortal(*output);
8363 /* Here both lists exist and are non-empty */
8364 array_a = invlist_array(a);
8365 array_b = invlist_array(b);
8367 /* If are to take the union of 'a' with the complement of b, set it
8368 * up so are looking at b's complement. */
8371 /* To complement, we invert: if the first element is 0, remove it. To
8372 * do this, we just pretend the array starts one later */
8373 if (array_b[0] == 0) {
8379 /* But if the first element is not zero, we pretend the list starts
8380 * at the 0 that is always stored immediately before the array. */
8386 /* Size the union for the worst case: that the sets are completely
8388 u = _new_invlist(len_a + len_b);
8390 /* Will contain U+0000 if either component does */
8391 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8392 || (len_b > 0 && array_b[0] == 0));
8394 /* Go through each list item by item, stopping when exhausted one of
8396 while (i_a < len_a && i_b < len_b) {
8397 UV cp; /* The element to potentially add to the union's array */
8398 bool cp_in_set; /* is it in the the input list's set or not */
8400 /* We need to take one or the other of the two inputs for the union.
8401 * Since we are merging two sorted lists, we take the smaller of the
8402 * next items. In case of a tie, we take the one that is in its set
8403 * first. If we took one not in the set first, it would decrement the
8404 * count, possibly to 0 which would cause it to be output as ending the
8405 * range, and the next time through we would take the same number, and
8406 * output it again as beginning the next range. By doing it the
8407 * opposite way, there is no possibility that the count will be
8408 * momentarily decremented to 0, and thus the two adjoining ranges will
8409 * be seamlessly merged. (In a tie and both are in the set or both not
8410 * in the set, it doesn't matter which we take first.) */
8411 if (array_a[i_a] < array_b[i_b]
8412 || (array_a[i_a] == array_b[i_b]
8413 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8415 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8419 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8420 cp = array_b[i_b++];
8423 /* Here, have chosen which of the two inputs to look at. Only output
8424 * if the running count changes to/from 0, which marks the
8425 * beginning/end of a range in that's in the set */
8428 array_u[i_u++] = cp;
8435 array_u[i_u++] = cp;
8440 /* Here, we are finished going through at least one of the lists, which
8441 * means there is something remaining in at most one. We check if the list
8442 * that hasn't been exhausted is positioned such that we are in the middle
8443 * of a range in its set or not. (i_a and i_b point to the element beyond
8444 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8445 * is potentially more to output.
8446 * There are four cases:
8447 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8448 * in the union is entirely from the non-exhausted set.
8449 * 2) Both were in their sets, count is 2. Nothing further should
8450 * be output, as everything that remains will be in the exhausted
8451 * list's set, hence in the union; decrementing to 1 but not 0 insures
8453 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8454 * Nothing further should be output because the union includes
8455 * everything from the exhausted set. Not decrementing ensures that.
8456 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8457 * decrementing to 0 insures that we look at the remainder of the
8458 * non-exhausted set */
8459 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8460 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8465 /* The final length is what we've output so far, plus what else is about to
8466 * be output. (If 'count' is non-zero, then the input list we exhausted
8467 * has everything remaining up to the machine's limit in its set, and hence
8468 * in the union, so there will be no further output. */
8471 /* At most one of the subexpressions will be non-zero */
8472 len_u += (len_a - i_a) + (len_b - i_b);
8475 /* Set result to final length, which can change the pointer to array_u, so
8477 if (len_u != _invlist_len(u)) {
8478 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8480 array_u = invlist_array(u);
8483 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8484 * the other) ended with everything above it not in its set. That means
8485 * that the remaining part of the union is precisely the same as the
8486 * non-exhausted list, so can just copy it unchanged. (If both list were
8487 * exhausted at the same time, then the operations below will be both 0.)
8490 IV copy_count; /* At most one will have a non-zero copy count */
8491 if ((copy_count = len_a - i_a) > 0) {
8492 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8494 else if ((copy_count = len_b - i_b) > 0) {
8495 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8499 /* We may be removing a reference to one of the inputs. If so, the output
8500 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8501 * count decremented) */
8502 if (a == *output || b == *output) {
8503 assert(! invlist_is_iterating(*output));
8504 if ((SvTEMP(*output))) {
8508 SvREFCNT_dec_NN(*output);
8518 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8519 const bool complement_b, SV** i)
8521 /* Take the intersection of two inversion lists and point <i> to it. *i
8522 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8523 * the reference count to that list will be decremented if not already a
8524 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8525 * The first list, <a>, may be NULL, in which case an empty list is
8526 * returned. If <complement_b> is TRUE, the result will be the
8527 * intersection of <a> and the complement (or inversion) of <b> instead of
8530 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8531 * Richard Gillam, published by Addison-Wesley, and explained at some
8532 * length there. The preface says to incorporate its examples into your
8533 * code at your own risk. In fact, it had bugs
8535 * The algorithm is like a merge sort, and is essentially the same as the
8539 const UV* array_a; /* a's array */
8541 UV len_a; /* length of a's array */
8544 SV* r; /* the resulting intersection */
8548 UV i_a = 0; /* current index into a's array */
8552 /* running count, as explained in the algorithm source book; items are
8553 * stopped accumulating and are output when the count changes to/from 2.
8554 * The count is incremented when we start a range that's in the set, and
8555 * decremented when we start a range that's not in the set. So its range
8556 * is 0 to 2. Only when the count is 2 is something in the intersection.
8560 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8563 /* Special case if either one is empty */
8564 len_a = (a == NULL) ? 0 : _invlist_len(a);
8565 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8566 bool make_temp = FALSE;
8568 if (len_a != 0 && complement_b) {
8570 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8571 * be empty. Here, also we are using 'b's complement, which hence
8572 * must be every possible code point. Thus the intersection is
8576 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8581 *i = invlist_clone(a);
8583 /* else *i is already 'a' */
8591 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8592 * intersection must be empty */
8594 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8599 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8603 *i = _new_invlist(0);
8611 /* Here both lists exist and are non-empty */
8612 array_a = invlist_array(a);
8613 array_b = invlist_array(b);
8615 /* If are to take the intersection of 'a' with the complement of b, set it
8616 * up so are looking at b's complement. */
8619 /* To complement, we invert: if the first element is 0, remove it. To
8620 * do this, we just pretend the array starts one later */
8621 if (array_b[0] == 0) {
8627 /* But if the first element is not zero, we pretend the list starts
8628 * at the 0 that is always stored immediately before the array. */
8634 /* Size the intersection for the worst case: that the intersection ends up
8635 * fragmenting everything to be completely disjoint */
8636 r= _new_invlist(len_a + len_b);
8638 /* Will contain U+0000 iff both components do */
8639 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8640 && len_b > 0 && array_b[0] == 0);
8642 /* Go through each list item by item, stopping when exhausted one of
8644 while (i_a < len_a && i_b < len_b) {
8645 UV cp; /* The element to potentially add to the intersection's
8647 bool cp_in_set; /* Is it in the input list's set or not */
8649 /* We need to take one or the other of the two inputs for the
8650 * intersection. Since we are merging two sorted lists, we take the
8651 * smaller of the next items. In case of a tie, we take the one that
8652 * is not in its set first (a difference from the union algorithm). If
8653 * we took one in the set first, it would increment the count, possibly
8654 * to 2 which would cause it to be output as starting a range in the
8655 * intersection, and the next time through we would take that same
8656 * number, and output it again as ending the set. By doing it the
8657 * opposite of this, there is no possibility that the count will be
8658 * momentarily incremented to 2. (In a tie and both are in the set or
8659 * both not in the set, it doesn't matter which we take first.) */
8660 if (array_a[i_a] < array_b[i_b]
8661 || (array_a[i_a] == array_b[i_b]
8662 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8664 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8668 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8672 /* Here, have chosen which of the two inputs to look at. Only output
8673 * if the running count changes to/from 2, which marks the
8674 * beginning/end of a range that's in the intersection */
8678 array_r[i_r++] = cp;
8683 array_r[i_r++] = cp;
8689 /* Here, we are finished going through at least one of the lists, which
8690 * means there is something remaining in at most one. We check if the list
8691 * that has been exhausted is positioned such that we are in the middle
8692 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8693 * the ones we care about.) There are four cases:
8694 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8695 * nothing left in the intersection.
8696 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8697 * above 2. What should be output is exactly that which is in the
8698 * non-exhausted set, as everything it has is also in the intersection
8699 * set, and everything it doesn't have can't be in the intersection
8700 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8701 * gets incremented to 2. Like the previous case, the intersection is
8702 * everything that remains in the non-exhausted set.
8703 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8704 * remains 1. And the intersection has nothing more. */
8705 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8706 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8711 /* The final length is what we've output so far plus what else is in the
8712 * intersection. At most one of the subexpressions below will be non-zero
8716 len_r += (len_a - i_a) + (len_b - i_b);
8719 /* Set result to final length, which can change the pointer to array_r, so
8721 if (len_r != _invlist_len(r)) {
8722 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8724 array_r = invlist_array(r);
8727 /* Finish outputting any remaining */
8728 if (count >= 2) { /* At most one will have a non-zero copy count */
8730 if ((copy_count = len_a - i_a) > 0) {
8731 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8733 else if ((copy_count = len_b - i_b) > 0) {
8734 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8738 /* We may be removing a reference to one of the inputs. If so, the output
8739 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8740 * count decremented) */
8741 if (a == *i || b == *i) {
8742 assert(! invlist_is_iterating(*i));
8747 SvREFCNT_dec_NN(*i);
8757 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8759 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8760 * set. A pointer to the inversion list is returned. This may actually be
8761 * a new list, in which case the passed in one has been destroyed. The
8762 * passed in inversion list can be NULL, in which case a new one is created
8763 * with just the one range in it */
8768 if (invlist == NULL) {
8769 invlist = _new_invlist(2);
8773 len = _invlist_len(invlist);
8776 /* If comes after the final entry actually in the list, can just append it
8779 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8780 && start >= invlist_array(invlist)[len - 1]))
8782 _append_range_to_invlist(invlist, start, end);
8786 /* Here, can't just append things, create and return a new inversion list
8787 * which is the union of this range and the existing inversion list */
8788 range_invlist = _new_invlist(2);
8789 _append_range_to_invlist(range_invlist, start, end);
8791 _invlist_union(invlist, range_invlist, &invlist);
8793 /* The temporary can be freed */
8794 SvREFCNT_dec_NN(range_invlist);
8800 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
8801 UV** other_elements_ptr)
8803 /* Create and return an inversion list whose contents are to be populated
8804 * by the caller. The caller gives the number of elements (in 'size') and
8805 * the very first element ('element0'). This function will set
8806 * '*other_elements_ptr' to an array of UVs, where the remaining elements
8809 * Obviously there is some trust involved that the caller will properly
8810 * fill in the other elements of the array.
8812 * (The first element needs to be passed in, as the underlying code does
8813 * things differently depending on whether it is zero or non-zero) */
8815 SV* invlist = _new_invlist(size);
8818 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
8820 _append_range_to_invlist(invlist, element0, element0);
8821 offset = *get_invlist_offset_addr(invlist);
8823 invlist_set_len(invlist, size, offset);
8824 *other_elements_ptr = invlist_array(invlist) + 1;
8830 PERL_STATIC_INLINE SV*
8831 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8832 return _add_range_to_invlist(invlist, cp, cp);
8835 #ifndef PERL_IN_XSUB_RE
8837 Perl__invlist_invert(pTHX_ SV* const invlist)
8839 /* Complement the input inversion list. This adds a 0 if the list didn't
8840 * have a zero; removes it otherwise. As described above, the data
8841 * structure is set up so that this is very efficient */
8843 PERL_ARGS_ASSERT__INVLIST_INVERT;
8845 assert(! invlist_is_iterating(invlist));
8847 /* The inverse of matching nothing is matching everything */
8848 if (_invlist_len(invlist) == 0) {
8849 _append_range_to_invlist(invlist, 0, UV_MAX);
8853 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8858 PERL_STATIC_INLINE SV*
8859 S_invlist_clone(pTHX_ SV* const invlist)
8862 /* Return a new inversion list that is a copy of the input one, which is
8863 * unchanged. The new list will not be mortal even if the old one was. */
8865 /* Need to allocate extra space to accommodate Perl's addition of a
8866 * trailing NUL to SvPV's, since it thinks they are always strings */
8867 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8868 STRLEN physical_length = SvCUR(invlist);
8869 bool offset = *(get_invlist_offset_addr(invlist));
8871 PERL_ARGS_ASSERT_INVLIST_CLONE;
8873 *(get_invlist_offset_addr(new_invlist)) = offset;
8874 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8875 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8880 PERL_STATIC_INLINE STRLEN*
8881 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8883 /* Return the address of the UV that contains the current iteration
8886 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8888 assert(SvTYPE(invlist) == SVt_INVLIST);
8890 return &(((XINVLIST*) SvANY(invlist))->iterator);
8893 PERL_STATIC_INLINE void
8894 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8896 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8898 *get_invlist_iter_addr(invlist) = 0;
8901 PERL_STATIC_INLINE void
8902 S_invlist_iterfinish(pTHX_ SV* invlist)
8904 /* Terminate iterator for invlist. This is to catch development errors.
8905 * Any iteration that is interrupted before completed should call this
8906 * function. Functions that add code points anywhere else but to the end
8907 * of an inversion list assert that they are not in the middle of an
8908 * iteration. If they were, the addition would make the iteration
8909 * problematical: if the iteration hadn't reached the place where things
8910 * were being added, it would be ok */
8912 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8914 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8918 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8920 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8921 * This call sets in <*start> and <*end>, the next range in <invlist>.
8922 * Returns <TRUE> if successful and the next call will return the next
8923 * range; <FALSE> if was already at the end of the list. If the latter,
8924 * <*start> and <*end> are unchanged, and the next call to this function
8925 * will start over at the beginning of the list */
8927 STRLEN* pos = get_invlist_iter_addr(invlist);
8928 UV len = _invlist_len(invlist);
8931 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8934 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8938 array = invlist_array(invlist);
8940 *start = array[(*pos)++];
8946 *end = array[(*pos)++] - 1;
8952 PERL_STATIC_INLINE bool
8953 S_invlist_is_iterating(pTHX_ SV* const invlist)
8955 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8957 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8960 PERL_STATIC_INLINE UV
8961 S_invlist_highest(pTHX_ SV* const invlist)
8963 /* Returns the highest code point that matches an inversion list. This API
8964 * has an ambiguity, as it returns 0 under either the highest is actually
8965 * 0, or if the list is empty. If this distinction matters to you, check
8966 * for emptiness before calling this function */
8968 UV len = _invlist_len(invlist);
8971 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8977 array = invlist_array(invlist);
8979 /* The last element in the array in the inversion list always starts a
8980 * range that goes to infinity. That range may be for code points that are
8981 * matched in the inversion list, or it may be for ones that aren't
8982 * matched. In the latter case, the highest code point in the set is one
8983 * less than the beginning of this range; otherwise it is the final element
8984 * of this range: infinity */
8985 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8987 : array[len - 1] - 1;
8990 #ifndef PERL_IN_XSUB_RE
8992 Perl__invlist_contents(pTHX_ SV* const invlist)
8994 /* Get the contents of an inversion list into a string SV so that they can
8995 * be printed out. It uses the format traditionally done for debug tracing
8999 SV* output = newSVpvs("\n");
9001 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9003 assert(! invlist_is_iterating(invlist));
9005 invlist_iterinit(invlist);
9006 while (invlist_iternext(invlist, &start, &end)) {
9007 if (end == UV_MAX) {
9008 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9010 else if (end != start) {
9011 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9015 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9023 #ifndef PERL_IN_XSUB_RE
9025 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9026 const char * const indent, SV* const invlist)
9028 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9029 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9030 * the string 'indent'. The output looks like this:
9031 [0] 0x000A .. 0x000D
9033 [4] 0x2028 .. 0x2029
9034 [6] 0x3104 .. INFINITY
9035 * This means that the first range of code points matched by the list are
9036 * 0xA through 0xD; the second range contains only the single code point
9037 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9038 * are used to define each range (except if the final range extends to
9039 * infinity, only a single element is needed). The array index of the
9040 * first element for the corresponding range is given in brackets. */
9045 PERL_ARGS_ASSERT__INVLIST_DUMP;
9047 if (invlist_is_iterating(invlist)) {
9048 Perl_dump_indent(aTHX_ level, file,
9049 "%sCan't dump inversion list because is in middle of iterating\n",
9054 invlist_iterinit(invlist);
9055 while (invlist_iternext(invlist, &start, &end)) {
9056 if (end == UV_MAX) {
9057 Perl_dump_indent(aTHX_ level, file,
9058 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9059 indent, (UV)count, start);
9061 else if (end != start) {
9062 Perl_dump_indent(aTHX_ level, file,
9063 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9064 indent, (UV)count, start, end);
9067 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9068 indent, (UV)count, start);
9075 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9077 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9079 /* Return a boolean as to if the two passed in inversion lists are
9080 * identical. The final argument, if TRUE, says to take the complement of
9081 * the second inversion list before doing the comparison */
9083 const UV* array_a = invlist_array(a);
9084 const UV* array_b = invlist_array(b);
9085 UV len_a = _invlist_len(a);
9086 UV len_b = _invlist_len(b);
9088 UV i = 0; /* current index into the arrays */
9089 bool retval = TRUE; /* Assume are identical until proven otherwise */
9091 PERL_ARGS_ASSERT__INVLISTEQ;
9093 /* If are to compare 'a' with the complement of b, set it
9094 * up so are looking at b's complement. */
9097 /* The complement of nothing is everything, so <a> would have to have
9098 * just one element, starting at zero (ending at infinity) */
9100 return (len_a == 1 && array_a[0] == 0);
9102 else if (array_b[0] == 0) {
9104 /* Otherwise, to complement, we invert. Here, the first element is
9105 * 0, just remove it. To do this, we just pretend the array starts
9113 /* But if the first element is not zero, we pretend the list starts
9114 * at the 0 that is always stored immediately before the array. */
9120 /* Make sure that the lengths are the same, as well as the final element
9121 * before looping through the remainder. (Thus we test the length, final,
9122 * and first elements right off the bat) */
9123 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9126 else for (i = 0; i < len_a - 1; i++) {
9127 if (array_a[i] != array_b[i]) {
9137 #undef HEADER_LENGTH
9138 #undef TO_INTERNAL_SIZE
9139 #undef FROM_INTERNAL_SIZE
9140 #undef INVLIST_VERSION_ID
9142 /* End of inversion list object */
9145 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9147 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9148 * constructs, and updates RExC_flags with them. On input, RExC_parse
9149 * should point to the first flag; it is updated on output to point to the
9150 * final ')' or ':'. There needs to be at least one flag, or this will
9153 /* for (?g), (?gc), and (?o) warnings; warning
9154 about (?c) will warn about (?g) -- japhy */
9156 #define WASTED_O 0x01
9157 #define WASTED_G 0x02
9158 #define WASTED_C 0x04
9159 #define WASTED_GC (WASTED_G|WASTED_C)
9160 I32 wastedflags = 0x00;
9161 U32 posflags = 0, negflags = 0;
9162 U32 *flagsp = &posflags;
9163 char has_charset_modifier = '\0';
9165 bool has_use_defaults = FALSE;
9166 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9168 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9170 /* '^' as an initial flag sets certain defaults */
9171 if (UCHARAT(RExC_parse) == '^') {
9173 has_use_defaults = TRUE;
9174 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9175 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9176 ? REGEX_UNICODE_CHARSET
9177 : REGEX_DEPENDS_CHARSET);
9180 cs = get_regex_charset(RExC_flags);
9181 if (cs == REGEX_DEPENDS_CHARSET
9182 && (RExC_utf8 || RExC_uni_semantics))
9184 cs = REGEX_UNICODE_CHARSET;
9187 while (*RExC_parse) {
9188 /* && strchr("iogcmsx", *RExC_parse) */
9189 /* (?g), (?gc) and (?o) are useless here
9190 and must be globally applied -- japhy */
9191 switch (*RExC_parse) {
9193 /* Code for the imsx flags */
9194 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
9196 case LOCALE_PAT_MOD:
9197 if (has_charset_modifier) {
9198 goto excess_modifier;
9200 else if (flagsp == &negflags) {
9203 cs = REGEX_LOCALE_CHARSET;
9204 has_charset_modifier = LOCALE_PAT_MOD;
9206 case UNICODE_PAT_MOD:
9207 if (has_charset_modifier) {
9208 goto excess_modifier;
9210 else if (flagsp == &negflags) {
9213 cs = REGEX_UNICODE_CHARSET;
9214 has_charset_modifier = UNICODE_PAT_MOD;
9216 case ASCII_RESTRICT_PAT_MOD:
9217 if (flagsp == &negflags) {
9220 if (has_charset_modifier) {
9221 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9222 goto excess_modifier;
9224 /* Doubled modifier implies more restricted */
9225 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9228 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9230 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9232 case DEPENDS_PAT_MOD:
9233 if (has_use_defaults) {
9234 goto fail_modifiers;
9236 else if (flagsp == &negflags) {
9239 else if (has_charset_modifier) {
9240 goto excess_modifier;
9243 /* The dual charset means unicode semantics if the
9244 * pattern (or target, not known until runtime) are
9245 * utf8, or something in the pattern indicates unicode
9247 cs = (RExC_utf8 || RExC_uni_semantics)
9248 ? REGEX_UNICODE_CHARSET
9249 : REGEX_DEPENDS_CHARSET;
9250 has_charset_modifier = DEPENDS_PAT_MOD;
9254 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9255 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9257 else if (has_charset_modifier == *(RExC_parse - 1)) {
9258 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9262 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9267 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9270 case ONCE_PAT_MOD: /* 'o' */
9271 case GLOBAL_PAT_MOD: /* 'g' */
9272 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9273 const I32 wflagbit = *RExC_parse == 'o'
9276 if (! (wastedflags & wflagbit) ) {
9277 wastedflags |= wflagbit;
9278 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9281 "Useless (%s%c) - %suse /%c modifier",
9282 flagsp == &negflags ? "?-" : "?",
9284 flagsp == &negflags ? "don't " : "",
9291 case CONTINUE_PAT_MOD: /* 'c' */
9292 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9293 if (! (wastedflags & WASTED_C) ) {
9294 wastedflags |= WASTED_GC;
9295 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9298 "Useless (%sc) - %suse /gc modifier",
9299 flagsp == &negflags ? "?-" : "?",
9300 flagsp == &negflags ? "don't " : ""
9305 case KEEPCOPY_PAT_MOD: /* 'p' */
9306 if (flagsp == &negflags) {
9308 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9310 *flagsp |= RXf_PMf_KEEPCOPY;
9314 /* A flag is a default iff it is following a minus, so
9315 * if there is a minus, it means will be trying to
9316 * re-specify a default which is an error */
9317 if (has_use_defaults || flagsp == &negflags) {
9318 goto fail_modifiers;
9321 wastedflags = 0; /* reset so (?g-c) warns twice */
9325 RExC_flags |= posflags;
9326 RExC_flags &= ~negflags;
9327 set_regex_charset(&RExC_flags, cs);
9328 if (RExC_flags & RXf_PMf_FOLD) {
9329 RExC_contains_i = 1;
9335 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9336 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9337 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9338 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9347 - reg - regular expression, i.e. main body or parenthesized thing
9349 * Caller must absorb opening parenthesis.
9351 * Combining parenthesis handling with the base level of regular expression
9352 * is a trifle forced, but the need to tie the tails of the branches to what
9353 * follows makes it hard to avoid.
9355 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9357 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9359 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9362 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9363 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9364 needs to be restarted.
9365 Otherwise would only return NULL if regbranch() returns NULL, which
9368 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9369 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9370 * 2 is like 1, but indicates that nextchar() has been called to advance
9371 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9372 * this flag alerts us to the need to check for that */
9375 regnode *ret; /* Will be the head of the group. */
9378 regnode *ender = NULL;
9381 U32 oregflags = RExC_flags;
9382 bool have_branch = 0;
9384 I32 freeze_paren = 0;
9385 I32 after_freeze = 0;
9387 char * parse_start = RExC_parse; /* MJD */
9388 char * const oregcomp_parse = RExC_parse;
9390 GET_RE_DEBUG_FLAGS_DECL;
9392 PERL_ARGS_ASSERT_REG;
9393 DEBUG_PARSE("reg ");
9395 *flagp = 0; /* Tentatively. */
9398 /* Make an OPEN node, if parenthesized. */
9401 /* Under /x, space and comments can be gobbled up between the '(' and
9402 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9403 * intervening space, as the sequence is a token, and a token should be
9405 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9407 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9408 char *start_verb = RExC_parse;
9409 STRLEN verb_len = 0;
9410 char *start_arg = NULL;
9411 unsigned char op = 0;
9413 int internal_argval = 0; /* internal_argval is only useful if
9416 if (has_intervening_patws && SIZE_ONLY) {
9417 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
9419 while ( *RExC_parse && *RExC_parse != ')' ) {
9420 if ( *RExC_parse == ':' ) {
9421 start_arg = RExC_parse + 1;
9427 verb_len = RExC_parse - start_verb;
9430 while ( *RExC_parse && *RExC_parse != ')' )
9432 if ( *RExC_parse != ')' )
9433 vFAIL("Unterminated verb pattern argument");
9434 if ( RExC_parse == start_arg )
9437 if ( *RExC_parse != ')' )
9438 vFAIL("Unterminated verb pattern");
9441 switch ( *start_verb ) {
9442 case 'A': /* (*ACCEPT) */
9443 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9445 internal_argval = RExC_nestroot;
9448 case 'C': /* (*COMMIT) */
9449 if ( memEQs(start_verb,verb_len,"COMMIT") )
9452 case 'F': /* (*FAIL) */
9453 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9458 case ':': /* (*:NAME) */
9459 case 'M': /* (*MARK:NAME) */
9460 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9465 case 'P': /* (*PRUNE) */
9466 if ( memEQs(start_verb,verb_len,"PRUNE") )
9469 case 'S': /* (*SKIP) */
9470 if ( memEQs(start_verb,verb_len,"SKIP") )
9473 case 'T': /* (*THEN) */
9474 /* [19:06] <TimToady> :: is then */
9475 if ( memEQs(start_verb,verb_len,"THEN") ) {
9477 RExC_seen |= REG_CUTGROUP_SEEN;
9482 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9484 "Unknown verb pattern '%"UTF8f"'",
9485 UTF8fARG(UTF, verb_len, start_verb));
9488 if ( start_arg && internal_argval ) {
9489 vFAIL3("Verb pattern '%.*s' may not have an argument",
9490 verb_len, start_verb);
9491 } else if ( argok < 0 && !start_arg ) {
9492 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9493 verb_len, start_verb);
9495 ret = reganode(pRExC_state, op, internal_argval);
9496 if ( ! internal_argval && ! SIZE_ONLY ) {
9498 SV *sv = newSVpvn( start_arg,
9499 RExC_parse - start_arg);
9500 ARG(ret) = add_data( pRExC_state,
9502 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9509 if (!internal_argval)
9510 RExC_seen |= REG_VERBARG_SEEN;
9511 } else if ( start_arg ) {
9512 vFAIL3("Verb pattern '%.*s' may not have an argument",
9513 verb_len, start_verb);
9515 ret = reg_node(pRExC_state, op);
9517 nextchar(pRExC_state);
9520 else if (*RExC_parse == '?') { /* (?...) */
9521 bool is_logical = 0;
9522 const char * const seqstart = RExC_parse;
9523 if (has_intervening_patws && SIZE_ONLY) {
9524 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
9528 paren = *RExC_parse++;
9529 ret = NULL; /* For look-ahead/behind. */
9532 case 'P': /* (?P...) variants for those used to PCRE/Python */
9533 paren = *RExC_parse++;
9534 if ( paren == '<') /* (?P<...>) named capture */
9536 else if (paren == '>') { /* (?P>name) named recursion */
9537 goto named_recursion;
9539 else if (paren == '=') { /* (?P=...) named backref */
9540 /* this pretty much dupes the code for \k<NAME> in
9541 * regatom(), if you change this make sure you change that
9543 char* name_start = RExC_parse;
9545 SV *sv_dat = reg_scan_name(pRExC_state,
9546 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9547 if (RExC_parse == name_start || *RExC_parse != ')')
9548 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9549 vFAIL2("Sequence %.3s... not terminated",parse_start);
9552 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9553 RExC_rxi->data->data[num]=(void*)sv_dat;
9554 SvREFCNT_inc_simple_void(sv_dat);
9557 ret = reganode(pRExC_state,
9560 : (ASCII_FOLD_RESTRICTED)
9562 : (AT_LEAST_UNI_SEMANTICS)
9570 Set_Node_Offset(ret, parse_start+1);
9571 Set_Node_Cur_Length(ret, parse_start);
9573 nextchar(pRExC_state);
9577 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9578 vFAIL3("Sequence (%.*s...) not recognized",
9579 RExC_parse-seqstart, seqstart);
9581 case '<': /* (?<...) */
9582 if (*RExC_parse == '!')
9584 else if (*RExC_parse != '=')
9590 case '\'': /* (?'...') */
9591 name_start= RExC_parse;
9592 svname = reg_scan_name(pRExC_state,
9593 SIZE_ONLY /* reverse test from the others */
9594 ? REG_RSN_RETURN_NAME
9595 : REG_RSN_RETURN_NULL);
9596 if (RExC_parse == name_start || *RExC_parse != paren)
9597 vFAIL2("Sequence (?%c... not terminated",
9598 paren=='>' ? '<' : paren);
9602 if (!svname) /* shouldn't happen */
9604 "panic: reg_scan_name returned NULL");
9605 if (!RExC_paren_names) {
9606 RExC_paren_names= newHV();
9607 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9609 RExC_paren_name_list= newAV();
9610 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
9613 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
9615 sv_dat = HeVAL(he_str);
9617 /* croak baby croak */
9619 "panic: paren_name hash element allocation failed");
9620 } else if ( SvPOK(sv_dat) ) {
9621 /* (?|...) can mean we have dupes so scan to check
9622 its already been stored. Maybe a flag indicating
9623 we are inside such a construct would be useful,
9624 but the arrays are likely to be quite small, so
9625 for now we punt -- dmq */
9626 IV count = SvIV(sv_dat);
9627 I32 *pv = (I32*)SvPVX(sv_dat);
9629 for ( i = 0 ; i < count ; i++ ) {
9630 if ( pv[i] == RExC_npar ) {
9636 pv = (I32*)SvGROW(sv_dat,
9637 SvCUR(sv_dat) + sizeof(I32)+1);
9638 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
9639 pv[count] = RExC_npar;
9640 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
9643 (void)SvUPGRADE(sv_dat,SVt_PVNV);
9644 sv_setpvn(sv_dat, (char *)&(RExC_npar),
9647 SvIV_set(sv_dat, 1);
9650 /* Yes this does cause a memory leak in debugging Perls
9652 if (!av_store(RExC_paren_name_list,
9653 RExC_npar, SvREFCNT_inc(svname)))
9654 SvREFCNT_dec_NN(svname);
9657 /*sv_dump(sv_dat);*/
9659 nextchar(pRExC_state);
9661 goto capturing_parens;
9663 RExC_seen |= REG_LOOKBEHIND_SEEN;
9664 RExC_in_lookbehind++;
9666 case '=': /* (?=...) */
9667 RExC_seen_zerolen++;
9669 case '!': /* (?!...) */
9670 RExC_seen_zerolen++;
9671 if (*RExC_parse == ')') {
9672 ret=reg_node(pRExC_state, OPFAIL);
9673 nextchar(pRExC_state);
9677 case '|': /* (?|...) */
9678 /* branch reset, behave like a (?:...) except that
9679 buffers in alternations share the same numbers */
9681 after_freeze = freeze_paren = RExC_npar;
9683 case ':': /* (?:...) */
9684 case '>': /* (?>...) */
9686 case '$': /* (?$...) */
9687 case '@': /* (?@...) */
9688 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
9690 case '#': /* (?#...) */
9691 /* XXX As soon as we disallow separating the '?' and '*' (by
9692 * spaces or (?#...) comment), it is believed that this case
9693 * will be unreachable and can be removed. See
9695 while (*RExC_parse && *RExC_parse != ')')
9697 if (*RExC_parse != ')')
9698 FAIL("Sequence (?#... not terminated");
9699 nextchar(pRExC_state);
9702 case '0' : /* (?0) */
9703 case 'R' : /* (?R) */
9704 if (*RExC_parse != ')')
9705 FAIL("Sequence (?R) not terminated");
9706 ret = reg_node(pRExC_state, GOSTART);
9707 RExC_seen |= REG_GOSTART_SEEN;
9708 *flagp |= POSTPONED;
9709 nextchar(pRExC_state);
9712 { /* named and numeric backreferences */
9714 case '&': /* (?&NAME) */
9715 parse_start = RExC_parse - 1;
9718 SV *sv_dat = reg_scan_name(pRExC_state,
9719 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9720 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9722 if (RExC_parse == RExC_end || *RExC_parse != ')')
9723 vFAIL("Sequence (?&... not terminated");
9724 goto gen_recurse_regop;
9725 assert(0); /* NOT REACHED */
9727 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9729 vFAIL("Illegal pattern");
9731 goto parse_recursion;
9733 case '-': /* (?-1) */
9734 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9735 RExC_parse--; /* rewind to let it be handled later */
9739 case '1': case '2': case '3': case '4': /* (?1) */
9740 case '5': case '6': case '7': case '8': case '9':
9743 num = atoi(RExC_parse);
9744 parse_start = RExC_parse - 1; /* MJD */
9745 if (*RExC_parse == '-')
9747 while (isDIGIT(*RExC_parse))
9749 if (*RExC_parse!=')')
9750 vFAIL("Expecting close bracket");
9753 if ( paren == '-' ) {
9755 Diagram of capture buffer numbering.
9756 Top line is the normal capture buffer numbers
9757 Bottom line is the negative indexing as from
9761 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9765 num = RExC_npar + num;
9768 vFAIL("Reference to nonexistent group");
9770 } else if ( paren == '+' ) {
9771 num = RExC_npar + num - 1;
9774 ret = reganode(pRExC_state, GOSUB, num);
9776 if (num > (I32)RExC_rx->nparens) {
9778 vFAIL("Reference to nonexistent group");
9780 ARG2L_SET( ret, RExC_recurse_count++);
9782 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9783 "Recurse #%"UVuf" to %"IVdf"\n",
9784 (UV)ARG(ret), (IV)ARG2L(ret)));
9788 RExC_seen |= REG_RECURSE_SEEN;
9789 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9790 Set_Node_Offset(ret, parse_start); /* MJD */
9792 *flagp |= POSTPONED;
9793 nextchar(pRExC_state);
9795 } /* named and numeric backreferences */
9796 assert(0); /* NOT REACHED */
9798 case '?': /* (??...) */
9800 if (*RExC_parse != '{') {
9802 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9804 "Sequence (%"UTF8f"...) not recognized",
9805 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9808 *flagp |= POSTPONED;
9809 paren = *RExC_parse++;
9811 case '{': /* (?{...}) */
9814 struct reg_code_block *cb;
9816 RExC_seen_zerolen++;
9818 if ( !pRExC_state->num_code_blocks
9819 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9820 || pRExC_state->code_blocks[pRExC_state->code_index].start
9821 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9824 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9825 FAIL("panic: Sequence (?{...}): no code block found\n");
9826 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9828 /* this is a pre-compiled code block (?{...}) */
9829 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9830 RExC_parse = RExC_start + cb->end;
9833 if (cb->src_regex) {
9834 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
9835 RExC_rxi->data->data[n] =
9836 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9837 RExC_rxi->data->data[n+1] = (void*)o;
9840 n = add_data(pRExC_state,
9841 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
9842 RExC_rxi->data->data[n] = (void*)o;
9845 pRExC_state->code_index++;
9846 nextchar(pRExC_state);
9850 ret = reg_node(pRExC_state, LOGICAL);
9851 eval = reganode(pRExC_state, EVAL, n);
9854 /* for later propagation into (??{}) return value */
9855 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9857 REGTAIL(pRExC_state, ret, eval);
9858 /* deal with the length of this later - MJD */
9861 ret = reganode(pRExC_state, EVAL, n);
9862 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9863 Set_Node_Offset(ret, parse_start);
9866 case '(': /* (?(?{...})...) and (?(?=...)...) */
9869 if (RExC_parse[0] == '?') { /* (?(?...)) */
9870 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9871 || RExC_parse[1] == '<'
9872 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9876 ret = reg_node(pRExC_state, LOGICAL);
9880 tail = reg(pRExC_state, 1, &flag, depth+1);
9881 if (flag & RESTART_UTF8) {
9882 *flagp = RESTART_UTF8;
9885 REGTAIL(pRExC_state, ret, tail);
9889 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9890 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9892 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9893 char *name_start= RExC_parse++;
9895 SV *sv_dat=reg_scan_name(pRExC_state,
9896 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9897 if (RExC_parse == name_start || *RExC_parse != ch)
9898 vFAIL2("Sequence (?(%c... not terminated",
9899 (ch == '>' ? '<' : ch));
9902 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9903 RExC_rxi->data->data[num]=(void*)sv_dat;
9904 SvREFCNT_inc_simple_void(sv_dat);
9906 ret = reganode(pRExC_state,NGROUPP,num);
9907 goto insert_if_check_paren;
9909 else if (RExC_parse[0] == 'D' &&
9910 RExC_parse[1] == 'E' &&
9911 RExC_parse[2] == 'F' &&
9912 RExC_parse[3] == 'I' &&
9913 RExC_parse[4] == 'N' &&
9914 RExC_parse[5] == 'E')
9916 ret = reganode(pRExC_state,DEFINEP,0);
9919 goto insert_if_check_paren;
9921 else if (RExC_parse[0] == 'R') {
9924 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9925 parno = atoi(RExC_parse++);
9926 while (isDIGIT(*RExC_parse))
9928 } else if (RExC_parse[0] == '&') {
9931 sv_dat = reg_scan_name(pRExC_state,
9933 ? REG_RSN_RETURN_NULL
9934 : REG_RSN_RETURN_DATA);
9935 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9937 ret = reganode(pRExC_state,INSUBP,parno);
9938 goto insert_if_check_paren;
9940 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9944 parno = atoi(RExC_parse++);
9946 while (isDIGIT(*RExC_parse))
9948 ret = reganode(pRExC_state, GROUPP, parno);
9950 insert_if_check_paren:
9951 if (*(tmp = nextchar(pRExC_state)) != ')') {
9952 /* nextchar also skips comments, so undo its work
9953 * and skip over the the next character.
9956 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9957 vFAIL("Switch condition not recognized");
9960 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9961 br = regbranch(pRExC_state, &flags, 1,depth+1);
9963 if (flags & RESTART_UTF8) {
9964 *flagp = RESTART_UTF8;
9967 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9970 REGTAIL(pRExC_state, br, reganode(pRExC_state,
9972 c = *nextchar(pRExC_state);
9977 vFAIL("(?(DEFINE)....) does not allow branches");
9979 /* Fake one for optimizer. */
9980 lastbr = reganode(pRExC_state, IFTHEN, 0);
9982 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9983 if (flags & RESTART_UTF8) {
9984 *flagp = RESTART_UTF8;
9987 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9990 REGTAIL(pRExC_state, ret, lastbr);
9993 c = *nextchar(pRExC_state);
9998 vFAIL("Switch (?(condition)... contains too many branches");
9999 ender = reg_node(pRExC_state, TAIL);
10000 REGTAIL(pRExC_state, br, ender);
10002 REGTAIL(pRExC_state, lastbr, ender);
10003 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10006 REGTAIL(pRExC_state, ret, ender);
10007 RExC_size++; /* XXX WHY do we need this?!!
10008 For large programs it seems to be required
10009 but I can't figure out why. -- dmq*/
10013 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10014 vFAIL("Unknown switch condition (?(...))");
10017 case '[': /* (?[ ... ]) */
10018 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10021 RExC_parse--; /* for vFAIL to print correctly */
10022 vFAIL("Sequence (? incomplete");
10024 default: /* e.g., (?i) */
10027 parse_lparen_question_flags(pRExC_state);
10028 if (UCHARAT(RExC_parse) != ':') {
10029 nextchar(pRExC_state);
10034 nextchar(pRExC_state);
10044 ret = reganode(pRExC_state, OPEN, parno);
10046 if (!RExC_nestroot)
10047 RExC_nestroot = parno;
10048 if (RExC_seen & REG_RECURSE_SEEN
10049 && !RExC_open_parens[parno-1])
10051 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10052 "Setting open paren #%"IVdf" to %d\n",
10053 (IV)parno, REG_NODE_NUM(ret)));
10054 RExC_open_parens[parno-1]= ret;
10057 Set_Node_Length(ret, 1); /* MJD */
10058 Set_Node_Offset(ret, RExC_parse); /* MJD */
10066 /* Pick up the branches, linking them together. */
10067 parse_start = RExC_parse; /* MJD */
10068 br = regbranch(pRExC_state, &flags, 1,depth+1);
10070 /* branch_len = (paren != 0); */
10073 if (flags & RESTART_UTF8) {
10074 *flagp = RESTART_UTF8;
10077 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10079 if (*RExC_parse == '|') {
10080 if (!SIZE_ONLY && RExC_extralen) {
10081 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10084 reginsert(pRExC_state, BRANCH, br, depth+1);
10085 Set_Node_Length(br, paren != 0);
10086 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10090 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10092 else if (paren == ':') {
10093 *flagp |= flags&SIMPLE;
10095 if (is_open) { /* Starts with OPEN. */
10096 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10098 else if (paren != '?') /* Not Conditional */
10100 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10102 while (*RExC_parse == '|') {
10103 if (!SIZE_ONLY && RExC_extralen) {
10104 ender = reganode(pRExC_state, LONGJMP,0);
10106 /* Append to the previous. */
10107 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10110 RExC_extralen += 2; /* Account for LONGJMP. */
10111 nextchar(pRExC_state);
10112 if (freeze_paren) {
10113 if (RExC_npar > after_freeze)
10114 after_freeze = RExC_npar;
10115 RExC_npar = freeze_paren;
10117 br = regbranch(pRExC_state, &flags, 0, depth+1);
10120 if (flags & RESTART_UTF8) {
10121 *flagp = RESTART_UTF8;
10124 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10126 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10128 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10131 if (have_branch || paren != ':') {
10132 /* Make a closing node, and hook it on the end. */
10135 ender = reg_node(pRExC_state, TAIL);
10138 ender = reganode(pRExC_state, CLOSE, parno);
10139 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10140 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10141 "Setting close paren #%"IVdf" to %d\n",
10142 (IV)parno, REG_NODE_NUM(ender)));
10143 RExC_close_parens[parno-1]= ender;
10144 if (RExC_nestroot == parno)
10147 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10148 Set_Node_Length(ender,1); /* MJD */
10154 *flagp &= ~HASWIDTH;
10157 ender = reg_node(pRExC_state, SUCCEED);
10160 ender = reg_node(pRExC_state, END);
10162 assert(!RExC_opend); /* there can only be one! */
10163 RExC_opend = ender;
10167 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10168 SV * const mysv_val1=sv_newmortal();
10169 SV * const mysv_val2=sv_newmortal();
10170 DEBUG_PARSE_MSG("lsbr");
10171 regprop(RExC_rx, mysv_val1, lastbr, NULL);
10172 regprop(RExC_rx, mysv_val2, ender, NULL);
10173 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10174 SvPV_nolen_const(mysv_val1),
10175 (IV)REG_NODE_NUM(lastbr),
10176 SvPV_nolen_const(mysv_val2),
10177 (IV)REG_NODE_NUM(ender),
10178 (IV)(ender - lastbr)
10181 REGTAIL(pRExC_state, lastbr, ender);
10183 if (have_branch && !SIZE_ONLY) {
10184 char is_nothing= 1;
10186 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10188 /* Hook the tails of the branches to the closing node. */
10189 for (br = ret; br; br = regnext(br)) {
10190 const U8 op = PL_regkind[OP(br)];
10191 if (op == BRANCH) {
10192 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10193 if ( OP(NEXTOPER(br)) != NOTHING
10194 || regnext(NEXTOPER(br)) != ender)
10197 else if (op == BRANCHJ) {
10198 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10199 /* for now we always disable this optimisation * /
10200 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10201 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10207 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10208 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10209 SV * const mysv_val1=sv_newmortal();
10210 SV * const mysv_val2=sv_newmortal();
10211 DEBUG_PARSE_MSG("NADA");
10212 regprop(RExC_rx, mysv_val1, ret, NULL);
10213 regprop(RExC_rx, mysv_val2, ender, NULL);
10214 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10215 SvPV_nolen_const(mysv_val1),
10216 (IV)REG_NODE_NUM(ret),
10217 SvPV_nolen_const(mysv_val2),
10218 (IV)REG_NODE_NUM(ender),
10223 if (OP(ender) == TAIL) {
10228 for ( opt= br + 1; opt < ender ; opt++ )
10229 OP(opt)= OPTIMIZED;
10230 NEXT_OFF(br)= ender - br;
10238 static const char parens[] = "=!<,>";
10240 if (paren && (p = strchr(parens, paren))) {
10241 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10242 int flag = (p - parens) > 1;
10245 node = SUSPEND, flag = 0;
10246 reginsert(pRExC_state, node,ret, depth+1);
10247 Set_Node_Cur_Length(ret, parse_start);
10248 Set_Node_Offset(ret, parse_start + 1);
10250 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10254 /* Check for proper termination. */
10256 /* restore original flags, but keep (?p) */
10257 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10258 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10259 RExC_parse = oregcomp_parse;
10260 vFAIL("Unmatched (");
10263 else if (!paren && RExC_parse < RExC_end) {
10264 if (*RExC_parse == ')') {
10266 vFAIL("Unmatched )");
10269 FAIL("Junk on end of regexp"); /* "Can't happen". */
10270 assert(0); /* NOTREACHED */
10273 if (RExC_in_lookbehind) {
10274 RExC_in_lookbehind--;
10276 if (after_freeze > RExC_npar)
10277 RExC_npar = after_freeze;
10282 - regbranch - one alternative of an | operator
10284 * Implements the concatenation operator.
10286 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10290 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10294 regnode *chain = NULL;
10296 I32 flags = 0, c = 0;
10297 GET_RE_DEBUG_FLAGS_DECL;
10299 PERL_ARGS_ASSERT_REGBRANCH;
10301 DEBUG_PARSE("brnc");
10306 if (!SIZE_ONLY && RExC_extralen)
10307 ret = reganode(pRExC_state, BRANCHJ,0);
10309 ret = reg_node(pRExC_state, BRANCH);
10310 Set_Node_Length(ret, 1);
10314 if (!first && SIZE_ONLY)
10315 RExC_extralen += 1; /* BRANCHJ */
10317 *flagp = WORST; /* Tentatively. */
10320 nextchar(pRExC_state);
10321 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10322 flags &= ~TRYAGAIN;
10323 latest = regpiece(pRExC_state, &flags,depth+1);
10324 if (latest == NULL) {
10325 if (flags & TRYAGAIN)
10327 if (flags & RESTART_UTF8) {
10328 *flagp = RESTART_UTF8;
10331 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10333 else if (ret == NULL)
10335 *flagp |= flags&(HASWIDTH|POSTPONED);
10336 if (chain == NULL) /* First piece. */
10337 *flagp |= flags&SPSTART;
10340 REGTAIL(pRExC_state, chain, latest);
10345 if (chain == NULL) { /* Loop ran zero times. */
10346 chain = reg_node(pRExC_state, NOTHING);
10351 *flagp |= flags&SIMPLE;
10358 - regpiece - something followed by possible [*+?]
10360 * Note that the branching code sequences used for ? and the general cases
10361 * of * and + are somewhat optimized: they use the same NOTHING node as
10362 * both the endmarker for their branch list and the body of the last branch.
10363 * It might seem that this node could be dispensed with entirely, but the
10364 * endmarker role is not redundant.
10366 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10368 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10372 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10379 const char * const origparse = RExC_parse;
10381 I32 max = REG_INFTY;
10382 #ifdef RE_TRACK_PATTERN_OFFSETS
10385 const char *maxpos = NULL;
10387 /* Save the original in case we change the emitted regop to a FAIL. */
10388 regnode * const orig_emit = RExC_emit;
10390 GET_RE_DEBUG_FLAGS_DECL;
10392 PERL_ARGS_ASSERT_REGPIECE;
10394 DEBUG_PARSE("piec");
10396 ret = regatom(pRExC_state, &flags,depth+1);
10398 if (flags & (TRYAGAIN|RESTART_UTF8))
10399 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10401 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10407 if (op == '{' && regcurly(RExC_parse, FALSE)) {
10409 #ifdef RE_TRACK_PATTERN_OFFSETS
10410 parse_start = RExC_parse; /* MJD */
10412 next = RExC_parse + 1;
10413 while (isDIGIT(*next) || *next == ',') {
10414 if (*next == ',') {
10422 if (*next == '}') { /* got one */
10426 min = atoi(RExC_parse);
10427 if (*maxpos == ',')
10430 maxpos = RExC_parse;
10431 max = atoi(maxpos);
10432 if (!max && *maxpos != '0')
10433 max = REG_INFTY; /* meaning "infinity" */
10434 else if (max >= REG_INFTY)
10435 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10437 nextchar(pRExC_state);
10438 if (max < min) { /* If can't match, warn and optimize to fail
10441 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10443 /* We can't back off the size because we have to reserve
10444 * enough space for all the things we are about to throw
10445 * away, but we can shrink it by the ammount we are about
10446 * to re-use here */
10447 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10450 RExC_emit = orig_emit;
10452 ret = reg_node(pRExC_state, OPFAIL);
10455 else if (min == max
10456 && RExC_parse < RExC_end
10457 && (*RExC_parse == '?' || *RExC_parse == '+'))
10460 ckWARN2reg(RExC_parse + 1,
10461 "Useless use of greediness modifier '%c'",
10464 /* Absorb the modifier, so later code doesn't see nor use
10466 nextchar(pRExC_state);
10470 if ((flags&SIMPLE)) {
10471 RExC_naughty += 2 + RExC_naughty / 2;
10472 reginsert(pRExC_state, CURLY, ret, depth+1);
10473 Set_Node_Offset(ret, parse_start+1); /* MJD */
10474 Set_Node_Cur_Length(ret, parse_start);
10477 regnode * const w = reg_node(pRExC_state, WHILEM);
10480 REGTAIL(pRExC_state, ret, w);
10481 if (!SIZE_ONLY && RExC_extralen) {
10482 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10483 reginsert(pRExC_state, NOTHING,ret, depth+1);
10484 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10486 reginsert(pRExC_state, CURLYX,ret, depth+1);
10488 Set_Node_Offset(ret, parse_start+1);
10489 Set_Node_Length(ret,
10490 op == '{' ? (RExC_parse - parse_start) : 1);
10492 if (!SIZE_ONLY && RExC_extralen)
10493 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10494 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10496 RExC_whilem_seen++, RExC_extralen += 3;
10497 RExC_naughty += 4 + RExC_naughty; /* compound interest */
10504 *flagp |= HASWIDTH;
10506 ARG1_SET(ret, (U16)min);
10507 ARG2_SET(ret, (U16)max);
10509 if (max == REG_INFTY)
10510 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10516 if (!ISMULT1(op)) {
10521 #if 0 /* Now runtime fix should be reliable. */
10523 /* if this is reinstated, don't forget to put this back into perldiag:
10525 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10527 (F) The part of the regexp subject to either the * or + quantifier
10528 could match an empty string. The {#} shows in the regular
10529 expression about where the problem was discovered.
10533 if (!(flags&HASWIDTH) && op != '?')
10534 vFAIL("Regexp *+ operand could be empty");
10537 #ifdef RE_TRACK_PATTERN_OFFSETS
10538 parse_start = RExC_parse;
10540 nextchar(pRExC_state);
10542 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10544 if (op == '*' && (flags&SIMPLE)) {
10545 reginsert(pRExC_state, STAR, ret, depth+1);
10548 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10550 else if (op == '*') {
10554 else if (op == '+' && (flags&SIMPLE)) {
10555 reginsert(pRExC_state, PLUS, ret, depth+1);
10558 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10560 else if (op == '+') {
10564 else if (op == '?') {
10569 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10570 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10571 ckWARN2reg(RExC_parse,
10572 "%"UTF8f" matches null string many times",
10573 UTF8fARG(UTF, (RExC_parse >= origparse
10574 ? RExC_parse - origparse
10577 (void)ReREFCNT_inc(RExC_rx_sv);
10580 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10581 nextchar(pRExC_state);
10582 reginsert(pRExC_state, MINMOD, ret, depth+1);
10583 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10586 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10588 nextchar(pRExC_state);
10589 ender = reg_node(pRExC_state, SUCCEED);
10590 REGTAIL(pRExC_state, ret, ender);
10591 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10593 ender = reg_node(pRExC_state, TAIL);
10594 REGTAIL(pRExC_state, ret, ender);
10597 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10599 vFAIL("Nested quantifiers");
10606 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
10607 UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
10608 const bool strict /* Apply stricter parsing rules? */
10612 /* This is expected to be called by a parser routine that has recognized '\N'
10613 and needs to handle the rest. RExC_parse is expected to point at the first
10614 char following the N at the time of the call. On successful return,
10615 RExC_parse has been updated to point to just after the sequence identified
10616 by this routine, and <*flagp> has been updated.
10618 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
10621 \N may begin either a named sequence, or if outside a character class, mean
10622 to match a non-newline. For non single-quoted regexes, the tokenizer has
10623 attempted to decide which, and in the case of a named sequence, converted it
10624 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
10625 where c1... are the characters in the sequence. For single-quoted regexes,
10626 the tokenizer passes the \N sequence through unchanged; this code will not
10627 attempt to determine this nor expand those, instead raising a syntax error.
10628 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
10629 or there is no '}', it signals that this \N occurrence means to match a
10632 Only the \N{U+...} form should occur in a character class, for the same
10633 reason that '.' inside a character class means to just match a period: it
10634 just doesn't make sense.
10636 The function raises an error (via vFAIL), and doesn't return for various
10637 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
10638 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
10639 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
10640 only possible if node_p is non-NULL.
10643 If <valuep> is non-null, it means the caller can accept an input sequence
10644 consisting of a just a single code point; <*valuep> is set to that value
10645 if the input is such.
10647 If <node_p> is non-null it signifies that the caller can accept any other
10648 legal sequence (i.e., one that isn't just a single code point). <*node_p>
10650 1) \N means not-a-NL: points to a newly created REG_ANY node;
10651 2) \N{}: points to a new NOTHING node;
10652 3) otherwise: points to a new EXACT node containing the resolved
10654 Note that FALSE is returned for single code point sequences if <valuep> is
10658 char * endbrace; /* '}' following the name */
10660 char *endchar; /* Points to '.' or '}' ending cur char in the input
10662 bool has_multiple_chars; /* true if the input stream contains a sequence of
10663 more than one character */
10665 GET_RE_DEBUG_FLAGS_DECL;
10667 PERL_ARGS_ASSERT_GROK_BSLASH_N;
10669 GET_RE_DEBUG_FLAGS;
10671 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
10673 /* The [^\n] meaning of \N ignores spaces and comments under the /x
10674 * modifier. The other meaning does not, so use a temporary until we find
10675 * out which we are being called with */
10676 p = (RExC_flags & RXf_PMf_EXTENDED)
10677 ? regwhite( pRExC_state, RExC_parse )
10680 /* Disambiguate between \N meaning a named character versus \N meaning
10681 * [^\n]. The former is assumed when it can't be the latter. */
10682 if (*p != '{' || regcurly(p, FALSE)) {
10685 /* no bare \N allowed in a charclass */
10686 if (in_char_class) {
10687 vFAIL("\\N in a character class must be a named character: \\N{...}");
10691 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
10693 nextchar(pRExC_state);
10694 *node_p = reg_node(pRExC_state, REG_ANY);
10695 *flagp |= HASWIDTH|SIMPLE;
10697 Set_Node_Length(*node_p, 1); /* MJD */
10701 /* Here, we have decided it should be a named character or sequence */
10703 /* The test above made sure that the next real character is a '{', but
10704 * under the /x modifier, it could be separated by space (or a comment and
10705 * \n) and this is not allowed (for consistency with \x{...} and the
10706 * tokenizer handling of \N{NAME}). */
10707 if (*RExC_parse != '{') {
10708 vFAIL("Missing braces on \\N{}");
10711 RExC_parse++; /* Skip past the '{' */
10713 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
10714 || ! (endbrace == RExC_parse /* nothing between the {} */
10715 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below
10717 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg)
10720 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
10721 vFAIL("\\N{NAME} must be resolved by the lexer");
10724 if (endbrace == RExC_parse) { /* empty: \N{} */
10727 *node_p = reg_node(pRExC_state,NOTHING);
10729 else if (in_char_class) {
10730 if (SIZE_ONLY && in_char_class) {
10732 RExC_parse++; /* Position after the "}" */
10733 vFAIL("Zero length \\N{}");
10736 ckWARNreg(RExC_parse,
10737 "Ignoring zero length \\N{} in character class");
10745 nextchar(pRExC_state);
10749 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
10750 RExC_parse += 2; /* Skip past the 'U+' */
10752 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10754 /* Code points are separated by dots. If none, there is only one code
10755 * point, and is terminated by the brace */
10756 has_multiple_chars = (endchar < endbrace);
10758 if (valuep && (! has_multiple_chars || in_char_class)) {
10759 /* We only pay attention to the first char of
10760 multichar strings being returned in char classes. I kinda wonder
10761 if this makes sense as it does change the behaviour
10762 from earlier versions, OTOH that behaviour was broken
10763 as well. XXX Solution is to recharacterize as
10764 [rest-of-class]|multi1|multi2... */
10766 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10767 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10768 | PERL_SCAN_DISALLOW_PREFIX
10769 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10771 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10773 /* The tokenizer should have guaranteed validity, but it's possible to
10774 * bypass it by using single quoting, so check */
10775 if (length_of_hex == 0
10776 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10778 RExC_parse += length_of_hex; /* Includes all the valid */
10779 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10780 ? UTF8SKIP(RExC_parse)
10782 /* Guard against malformed utf8 */
10783 if (RExC_parse >= endchar) {
10784 RExC_parse = endchar;
10786 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10789 if (in_char_class && has_multiple_chars) {
10791 RExC_parse = endbrace;
10792 vFAIL("\\N{} in character class restricted to one character");
10795 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10799 RExC_parse = endbrace + 1;
10801 else if (! node_p || ! has_multiple_chars) {
10803 /* Here, the input is legal, but not according to the caller's
10804 * options. We fail without advancing the parse, so that the
10805 * caller can try again */
10811 /* What is done here is to convert this to a sub-pattern of the form
10812 * (?:\x{char1}\x{char2}...)
10813 * and then call reg recursively. That way, it retains its atomicness,
10814 * while not having to worry about special handling that some code
10815 * points may have. toke.c has converted the original Unicode values
10816 * to native, so that we can just pass on the hex values unchanged. We
10817 * do have to set a flag to keep recoding from happening in the
10820 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10822 char *orig_end = RExC_end;
10825 while (RExC_parse < endbrace) {
10827 /* Convert to notation the rest of the code understands */
10828 sv_catpv(substitute_parse, "\\x{");
10829 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10830 sv_catpv(substitute_parse, "}");
10832 /* Point to the beginning of the next character in the sequence. */
10833 RExC_parse = endchar + 1;
10834 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10836 sv_catpv(substitute_parse, ")");
10838 RExC_parse = SvPV(substitute_parse, len);
10840 /* Don't allow empty number */
10842 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10844 RExC_end = RExC_parse + len;
10846 /* The values are Unicode, and therefore not subject to recoding */
10847 RExC_override_recoding = 1;
10849 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10850 if (flags & RESTART_UTF8) {
10851 *flagp = RESTART_UTF8;
10854 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10857 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10859 RExC_parse = endbrace;
10860 RExC_end = orig_end;
10861 RExC_override_recoding = 0;
10863 nextchar(pRExC_state);
10873 * It returns the code point in utf8 for the value in *encp.
10874 * value: a code value in the source encoding
10875 * encp: a pointer to an Encode object
10877 * If the result from Encode is not a single character,
10878 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10881 S_reg_recode(pTHX_ const char value, SV **encp)
10884 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10885 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10886 const STRLEN newlen = SvCUR(sv);
10887 UV uv = UNICODE_REPLACEMENT;
10889 PERL_ARGS_ASSERT_REG_RECODE;
10893 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10896 if (!newlen || numlen != newlen) {
10897 uv = UNICODE_REPLACEMENT;
10903 PERL_STATIC_INLINE U8
10904 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10908 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10914 op = get_regex_charset(RExC_flags);
10915 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10916 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10917 been, so there is no hole */
10920 return op + EXACTF;
10923 PERL_STATIC_INLINE void
10924 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
10925 regnode *node, I32* flagp, STRLEN len, UV code_point,
10928 /* This knows the details about sizing an EXACTish node, setting flags for
10929 * it (by setting <*flagp>, and potentially populating it with a single
10932 * If <len> (the length in bytes) is non-zero, this function assumes that
10933 * the node has already been populated, and just does the sizing. In this
10934 * case <code_point> should be the final code point that has already been
10935 * placed into the node. This value will be ignored except that under some
10936 * circumstances <*flagp> is set based on it.
10938 * If <len> is zero, the function assumes that the node is to contain only
10939 * the single character given by <code_point> and calculates what <len>
10940 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10941 * additionally will populate the node's STRING with <code_point> or its
10944 * In both cases <*flagp> is appropriately set
10946 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10947 * 255, must be folded (the former only when the rules indicate it can
10950 * When it does the populating, it looks at the flag 'downgradable'. If
10951 * true with a node that folds, it checks if the single code point
10952 * participates in a fold, and if not downgrades the node to an EXACT.
10953 * This helps the optimizer */
10955 bool len_passed_in = cBOOL(len != 0);
10956 U8 character[UTF8_MAXBYTES_CASE+1];
10958 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10960 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
10961 * sizing difference, and is extra work that is thrown away */
10962 if (downgradable && ! PASS2) {
10963 downgradable = FALSE;
10966 if (! len_passed_in) {
10968 if (UNI_IS_INVARIANT(code_point)) {
10969 if (LOC || ! FOLD) { /* /l defers folding until runtime */
10970 *character = (U8) code_point;
10972 else { /* Here is /i and not /l (toFOLD() is defined on just
10973 ASCII, which isn't the same thing as INVARIANT on
10974 EBCDIC, but it works there, as the extra invariants
10975 fold to themselves) */
10976 *character = toFOLD((U8) code_point);
10978 && *character == code_point
10979 && ! HAS_NONLATIN1_FOLD_CLOSURE(code_point))
10986 else if (FOLD && (! LOC
10987 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
10988 { /* Folding, and ok to do so now */
10989 UV folded = _to_uni_fold_flags(
10993 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
10994 ? FOLD_FLAGS_NOMIX_ASCII
10997 && folded == code_point
10998 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11003 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11005 /* Not folding this cp, and can output it directly */
11006 *character = UTF8_TWO_BYTE_HI(code_point);
11007 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11011 uvchr_to_utf8( character, code_point);
11012 len = UTF8SKIP(character);
11014 } /* Else pattern isn't UTF8. */
11016 *character = (U8) code_point;
11018 } /* Else is folded non-UTF8 */
11019 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11021 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11022 * comments at join_exact()); */
11023 *character = (U8) code_point;
11026 /* Can turn into an EXACT node if we know the fold at compile time,
11027 * and it folds to itself and doesn't particpate in other folds */
11030 && PL_fold_latin1[code_point] == code_point
11031 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11032 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11036 } /* else is Sharp s. May need to fold it */
11037 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11039 *(character + 1) = 's';
11043 *character = LATIN_SMALL_LETTER_SHARP_S;
11049 RExC_size += STR_SZ(len);
11052 RExC_emit += STR_SZ(len);
11053 STR_LEN(node) = len;
11054 if (! len_passed_in) {
11055 Copy((char *) character, STRING(node), len, char);
11059 *flagp |= HASWIDTH;
11061 /* A single character node is SIMPLE, except for the special-cased SHARP S
11063 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11064 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11065 || ! FOLD || ! DEPENDS_SEMANTICS))
11070 /* The OP may not be well defined in PASS1 */
11071 if (PASS2 && OP(node) == EXACTFL) {
11072 RExC_contains_locale = 1;
11077 /* return atoi(p), unless it's too big to sensibly be a backref,
11078 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11081 S_backref_value(char *p)
11085 for (;isDIGIT(*q); q++); /* calculate length of num */
11086 if (q - p == 0 || q - p > 9)
11093 - regatom - the lowest level
11095 Try to identify anything special at the start of the pattern. If there
11096 is, then handle it as required. This may involve generating a single regop,
11097 such as for an assertion; or it may involve recursing, such as to
11098 handle a () structure.
11100 If the string doesn't start with something special then we gobble up
11101 as much literal text as we can.
11103 Once we have been able to handle whatever type of thing started the
11104 sequence, we return.
11106 Note: we have to be careful with escapes, as they can be both literal
11107 and special, and in the case of \10 and friends, context determines which.
11109 A summary of the code structure is:
11111 switch (first_byte) {
11112 cases for each special:
11113 handle this special;
11116 switch (2nd byte) {
11117 cases for each unambiguous special:
11118 handle this special;
11120 cases for each ambigous special/literal:
11122 if (special) handle here
11124 default: // unambiguously literal:
11127 default: // is a literal char
11130 create EXACTish node for literal;
11131 while (more input and node isn't full) {
11132 switch (input_byte) {
11133 cases for each special;
11134 make sure parse pointer is set so that the next call to
11135 regatom will see this special first
11136 goto loopdone; // EXACTish node terminated by prev. char
11138 append char to EXACTISH node;
11140 get next input byte;
11144 return the generated node;
11146 Specifically there are two separate switches for handling
11147 escape sequences, with the one for handling literal escapes requiring
11148 a dummy entry for all of the special escapes that are actually handled
11151 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11153 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11155 Otherwise does not return NULL.
11159 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11162 regnode *ret = NULL;
11164 char *parse_start = RExC_parse;
11168 GET_RE_DEBUG_FLAGS_DECL;
11170 *flagp = WORST; /* Tentatively. */
11172 DEBUG_PARSE("atom");
11174 PERL_ARGS_ASSERT_REGATOM;
11177 switch ((U8)*RExC_parse) {
11179 RExC_seen_zerolen++;
11180 nextchar(pRExC_state);
11181 if (RExC_flags & RXf_PMf_MULTILINE)
11182 ret = reg_node(pRExC_state, MBOL);
11183 else if (RExC_flags & RXf_PMf_SINGLELINE)
11184 ret = reg_node(pRExC_state, SBOL);
11186 ret = reg_node(pRExC_state, BOL);
11187 Set_Node_Length(ret, 1); /* MJD */
11190 nextchar(pRExC_state);
11192 RExC_seen_zerolen++;
11193 if (RExC_flags & RXf_PMf_MULTILINE)
11194 ret = reg_node(pRExC_state, MEOL);
11195 else if (RExC_flags & RXf_PMf_SINGLELINE)
11196 ret = reg_node(pRExC_state, SEOL);
11198 ret = reg_node(pRExC_state, EOL);
11199 Set_Node_Length(ret, 1); /* MJD */
11202 nextchar(pRExC_state);
11203 if (RExC_flags & RXf_PMf_SINGLELINE)
11204 ret = reg_node(pRExC_state, SANY);
11206 ret = reg_node(pRExC_state, REG_ANY);
11207 *flagp |= HASWIDTH|SIMPLE;
11209 Set_Node_Length(ret, 1); /* MJD */
11213 char * const oregcomp_parse = ++RExC_parse;
11214 ret = regclass(pRExC_state, flagp,depth+1,
11215 FALSE, /* means parse the whole char class */
11216 TRUE, /* allow multi-char folds */
11217 FALSE, /* don't silence non-portable warnings. */
11219 if (*RExC_parse != ']') {
11220 RExC_parse = oregcomp_parse;
11221 vFAIL("Unmatched [");
11224 if (*flagp & RESTART_UTF8)
11226 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11229 nextchar(pRExC_state);
11230 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11234 nextchar(pRExC_state);
11235 ret = reg(pRExC_state, 2, &flags,depth+1);
11237 if (flags & TRYAGAIN) {
11238 if (RExC_parse == RExC_end) {
11239 /* Make parent create an empty node if needed. */
11240 *flagp |= TRYAGAIN;
11245 if (flags & RESTART_UTF8) {
11246 *flagp = RESTART_UTF8;
11249 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11252 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11256 if (flags & TRYAGAIN) {
11257 *flagp |= TRYAGAIN;
11260 vFAIL("Internal urp");
11261 /* Supposed to be caught earlier. */
11264 if (!regcurly(RExC_parse, FALSE)) {
11273 vFAIL("Quantifier follows nothing");
11278 This switch handles escape sequences that resolve to some kind
11279 of special regop and not to literal text. Escape sequnces that
11280 resolve to literal text are handled below in the switch marked
11283 Every entry in this switch *must* have a corresponding entry
11284 in the literal escape switch. However, the opposite is not
11285 required, as the default for this switch is to jump to the
11286 literal text handling code.
11288 switch ((U8)*++RExC_parse) {
11290 /* Special Escapes */
11292 RExC_seen_zerolen++;
11293 ret = reg_node(pRExC_state, SBOL);
11295 goto finish_meta_pat;
11297 ret = reg_node(pRExC_state, GPOS);
11298 RExC_seen |= REG_GPOS_SEEN;
11300 goto finish_meta_pat;
11302 RExC_seen_zerolen++;
11303 ret = reg_node(pRExC_state, KEEPS);
11305 /* XXX:dmq : disabling in-place substitution seems to
11306 * be necessary here to avoid cases of memory corruption, as
11307 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11309 RExC_seen |= REG_LOOKBEHIND_SEEN;
11310 goto finish_meta_pat;
11312 ret = reg_node(pRExC_state, SEOL);
11314 RExC_seen_zerolen++; /* Do not optimize RE away */
11315 goto finish_meta_pat;
11317 ret = reg_node(pRExC_state, EOS);
11319 RExC_seen_zerolen++; /* Do not optimize RE away */
11320 goto finish_meta_pat;
11322 ret = reg_node(pRExC_state, CANY);
11323 RExC_seen |= REG_CANY_SEEN;
11324 *flagp |= HASWIDTH|SIMPLE;
11325 goto finish_meta_pat;
11327 ret = reg_node(pRExC_state, CLUMP);
11328 *flagp |= HASWIDTH;
11329 goto finish_meta_pat;
11335 arg = ANYOF_WORDCHAR;
11339 RExC_seen_zerolen++;
11340 RExC_seen |= REG_LOOKBEHIND_SEEN;
11341 op = BOUND + get_regex_charset(RExC_flags);
11342 if (op > BOUNDA) { /* /aa is same as /a */
11345 else if (op == BOUNDL) {
11346 RExC_contains_locale = 1;
11348 ret = reg_node(pRExC_state, op);
11349 FLAGS(ret) = get_regex_charset(RExC_flags);
11351 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11352 /* diag_listed_as: Use "%s" instead of "%s" */
11353 vFAIL("Use \"\\b\\{\" instead of \"\\b{\"");
11355 goto finish_meta_pat;
11357 RExC_seen_zerolen++;
11358 RExC_seen |= REG_LOOKBEHIND_SEEN;
11359 op = NBOUND + get_regex_charset(RExC_flags);
11360 if (op > NBOUNDA) { /* /aa is same as /a */
11363 else if (op == NBOUNDL) {
11364 RExC_contains_locale = 1;
11366 ret = reg_node(pRExC_state, op);
11367 FLAGS(ret) = get_regex_charset(RExC_flags);
11369 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11370 /* diag_listed_as: Use "%s" instead of "%s" */
11371 vFAIL("Use \"\\B\\{\" instead of \"\\B{\"");
11373 goto finish_meta_pat;
11383 ret = reg_node(pRExC_state, LNBREAK);
11384 *flagp |= HASWIDTH|SIMPLE;
11385 goto finish_meta_pat;
11393 goto join_posix_op_known;
11399 arg = ANYOF_VERTWS;
11401 goto join_posix_op_known;
11411 op = POSIXD + get_regex_charset(RExC_flags);
11412 if (op > POSIXA) { /* /aa is same as /a */
11415 else if (op == POSIXL) {
11416 RExC_contains_locale = 1;
11419 join_posix_op_known:
11422 op += NPOSIXD - POSIXD;
11425 ret = reg_node(pRExC_state, op);
11427 FLAGS(ret) = namedclass_to_classnum(arg);
11430 *flagp |= HASWIDTH|SIMPLE;
11434 nextchar(pRExC_state);
11435 Set_Node_Length(ret, 2); /* MJD */
11441 char* parse_start = RExC_parse - 2;
11446 ret = regclass(pRExC_state, flagp,depth+1,
11447 TRUE, /* means just parse this element */
11448 FALSE, /* don't allow multi-char folds */
11449 FALSE, /* don't silence non-portable warnings.
11450 It would be a bug if these returned
11453 /* regclass() can only return RESTART_UTF8 if multi-char folds
11456 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11461 Set_Node_Offset(ret, parse_start + 2);
11462 Set_Node_Cur_Length(ret, parse_start);
11463 nextchar(pRExC_state);
11467 /* Handle \N and \N{NAME} with multiple code points here and not
11468 * below because it can be multicharacter. join_exact() will join
11469 * them up later on. Also this makes sure that things like
11470 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11471 * The options to the grok function call causes it to fail if the
11472 * sequence is just a single code point. We then go treat it as
11473 * just another character in the current EXACT node, and hence it
11474 * gets uniform treatment with all the other characters. The
11475 * special treatment for quantifiers is not needed for such single
11476 * character sequences */
11478 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
11479 FALSE /* not strict */ )) {
11480 if (*flagp & RESTART_UTF8)
11486 case 'k': /* Handle \k<NAME> and \k'NAME' */
11489 char ch= RExC_parse[1];
11490 if (ch != '<' && ch != '\'' && ch != '{') {
11492 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11493 vFAIL2("Sequence %.2s... not terminated",parse_start);
11495 /* this pretty much dupes the code for (?P=...) in reg(), if
11496 you change this make sure you change that */
11497 char* name_start = (RExC_parse += 2);
11499 SV *sv_dat = reg_scan_name(pRExC_state,
11500 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11501 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11502 if (RExC_parse == name_start || *RExC_parse != ch)
11503 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11504 vFAIL2("Sequence %.3s... not terminated",parse_start);
11507 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11508 RExC_rxi->data->data[num]=(void*)sv_dat;
11509 SvREFCNT_inc_simple_void(sv_dat);
11513 ret = reganode(pRExC_state,
11516 : (ASCII_FOLD_RESTRICTED)
11518 : (AT_LEAST_UNI_SEMANTICS)
11524 *flagp |= HASWIDTH;
11526 /* override incorrect value set in reganode MJD */
11527 Set_Node_Offset(ret, parse_start+1);
11528 Set_Node_Cur_Length(ret, parse_start);
11529 nextchar(pRExC_state);
11535 case '1': case '2': case '3': case '4':
11536 case '5': case '6': case '7': case '8': case '9':
11541 if (*RExC_parse == 'g') {
11545 if (*RExC_parse == '{') {
11549 if (*RExC_parse == '-') {
11553 if (hasbrace && !isDIGIT(*RExC_parse)) {
11554 if (isrel) RExC_parse--;
11556 goto parse_named_seq;
11559 num = S_backref_value(RExC_parse);
11561 vFAIL("Reference to invalid group 0");
11562 else if (num == I32_MAX) {
11563 if (isDIGIT(*RExC_parse))
11564 vFAIL("Reference to nonexistent group");
11566 vFAIL("Unterminated \\g... pattern");
11570 num = RExC_npar - num;
11572 vFAIL("Reference to nonexistent or unclosed group");
11576 num = S_backref_value(RExC_parse);
11577 /* bare \NNN might be backref or octal - if it is larger than or equal
11578 * RExC_npar then it is assumed to be and octal escape.
11579 * Note RExC_npar is +1 from the actual number of parens*/
11580 if (num == I32_MAX || (num > 9 && num >= RExC_npar
11581 && *RExC_parse != '8' && *RExC_parse != '9'))
11583 /* Probably a character specified in octal, e.g. \35 */
11588 /* at this point RExC_parse definitely points to a backref
11591 #ifdef RE_TRACK_PATTERN_OFFSETS
11592 char * const parse_start = RExC_parse - 1; /* MJD */
11594 while (isDIGIT(*RExC_parse))
11597 if (*RExC_parse != '}')
11598 vFAIL("Unterminated \\g{...} pattern");
11602 if (num > (I32)RExC_rx->nparens)
11603 vFAIL("Reference to nonexistent group");
11606 ret = reganode(pRExC_state,
11609 : (ASCII_FOLD_RESTRICTED)
11611 : (AT_LEAST_UNI_SEMANTICS)
11617 *flagp |= HASWIDTH;
11619 /* override incorrect value set in reganode MJD */
11620 Set_Node_Offset(ret, parse_start+1);
11621 Set_Node_Cur_Length(ret, parse_start);
11623 nextchar(pRExC_state);
11628 if (RExC_parse >= RExC_end)
11629 FAIL("Trailing \\");
11632 /* Do not generate "unrecognized" warnings here, we fall
11633 back into the quick-grab loop below */
11640 if (RExC_flags & RXf_PMf_EXTENDED) {
11641 if ( reg_skipcomment( pRExC_state ) )
11648 parse_start = RExC_parse - 1;
11657 #define MAX_NODE_STRING_SIZE 127
11658 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
11660 U8 upper_parse = MAX_NODE_STRING_SIZE;
11661 U8 node_type = compute_EXACTish(pRExC_state);
11662 bool next_is_quantifier;
11663 char * oldp = NULL;
11665 /* We can convert EXACTF nodes to EXACTFU if they contain only
11666 * characters that match identically regardless of the target
11667 * string's UTF8ness. The reason to do this is that EXACTF is not
11668 * trie-able, EXACTFU is.
11670 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
11671 * contain only above-Latin1 characters (hence must be in UTF8),
11672 * which don't participate in folds with Latin1-range characters,
11673 * as the latter's folds aren't known until runtime. (We don't
11674 * need to figure this out until pass 2) */
11675 bool maybe_exactfu = PASS2
11676 && (node_type == EXACTF || node_type == EXACTFL);
11678 /* If a folding node contains only code points that don't
11679 * participate in folds, it can be changed into an EXACT node,
11680 * which allows the optimizer more things to look for */
11683 ret = reg_node(pRExC_state, node_type);
11685 /* In pass1, folded, we use a temporary buffer instead of the
11686 * actual node, as the node doesn't exist yet */
11687 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
11693 /* We do the EXACTFish to EXACT node only if folding. (And we
11694 * don't need to figure this out until pass 2) */
11695 maybe_exact = FOLD && PASS2;
11697 /* XXX The node can hold up to 255 bytes, yet this only goes to
11698 * 127. I (khw) do not know why. Keeping it somewhat less than
11699 * 255 allows us to not have to worry about overflow due to
11700 * converting to utf8 and fold expansion, but that value is
11701 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
11702 * split up by this limit into a single one using the real max of
11703 * 255. Even at 127, this breaks under rare circumstances. If
11704 * folding, we do not want to split a node at a character that is a
11705 * non-final in a multi-char fold, as an input string could just
11706 * happen to want to match across the node boundary. The join
11707 * would solve that problem if the join actually happens. But a
11708 * series of more than two nodes in a row each of 127 would cause
11709 * the first join to succeed to get to 254, but then there wouldn't
11710 * be room for the next one, which could at be one of those split
11711 * multi-char folds. I don't know of any fool-proof solution. One
11712 * could back off to end with only a code point that isn't such a
11713 * non-final, but it is possible for there not to be any in the
11715 for (p = RExC_parse - 1;
11716 len < upper_parse && p < RExC_end;
11721 if (RExC_flags & RXf_PMf_EXTENDED)
11722 p = regwhite( pRExC_state, p );
11733 /* Literal Escapes Switch
11735 This switch is meant to handle escape sequences that
11736 resolve to a literal character.
11738 Every escape sequence that represents something
11739 else, like an assertion or a char class, is handled
11740 in the switch marked 'Special Escapes' above in this
11741 routine, but also has an entry here as anything that
11742 isn't explicitly mentioned here will be treated as
11743 an unescaped equivalent literal.
11746 switch ((U8)*++p) {
11747 /* These are all the special escapes. */
11748 case 'A': /* Start assertion */
11749 case 'b': case 'B': /* Word-boundary assertion*/
11750 case 'C': /* Single char !DANGEROUS! */
11751 case 'd': case 'D': /* digit class */
11752 case 'g': case 'G': /* generic-backref, pos assertion */
11753 case 'h': case 'H': /* HORIZWS */
11754 case 'k': case 'K': /* named backref, keep marker */
11755 case 'p': case 'P': /* Unicode property */
11756 case 'R': /* LNBREAK */
11757 case 's': case 'S': /* space class */
11758 case 'v': case 'V': /* VERTWS */
11759 case 'w': case 'W': /* word class */
11760 case 'X': /* eXtended Unicode "combining
11761 character sequence" */
11762 case 'z': case 'Z': /* End of line/string assertion */
11766 /* Anything after here is an escape that resolves to a
11767 literal. (Except digits, which may or may not)
11773 case 'N': /* Handle a single-code point named character. */
11774 /* The options cause it to fail if a multiple code
11775 * point sequence. Handle those in the switch() above
11777 RExC_parse = p + 1;
11778 if (! grok_bslash_N(pRExC_state, NULL, &ender,
11779 flagp, depth, FALSE,
11780 FALSE /* not strict */ ))
11782 if (*flagp & RESTART_UTF8)
11783 FAIL("panic: grok_bslash_N set RESTART_UTF8");
11784 RExC_parse = p = oldp;
11788 if (ender > 0xff) {
11805 ender = ASCII_TO_NATIVE('\033');
11815 const char* error_msg;
11817 bool valid = grok_bslash_o(&p,
11820 TRUE, /* out warnings */
11821 FALSE, /* not strict */
11822 TRUE, /* Output warnings
11827 RExC_parse = p; /* going to die anyway; point
11828 to exact spot of failure */
11832 if (PL_encoding && ender < 0x100) {
11833 goto recode_encoding;
11835 if (ender > 0xff) {
11842 UV result = UV_MAX; /* initialize to erroneous
11844 const char* error_msg;
11846 bool valid = grok_bslash_x(&p,
11849 TRUE, /* out warnings */
11850 FALSE, /* not strict */
11851 TRUE, /* Output warnings
11856 RExC_parse = p; /* going to die anyway; point
11857 to exact spot of failure */
11862 if (PL_encoding && ender < 0x100) {
11863 goto recode_encoding;
11865 if (ender > 0xff) {
11872 ender = grok_bslash_c(*p++, SIZE_ONLY);
11874 case '8': case '9': /* must be a backreference */
11877 case '1': case '2': case '3':case '4':
11878 case '5': case '6': case '7':
11879 /* When we parse backslash escapes there is ambiguity
11880 * between backreferences and octal escapes. Any escape
11881 * from \1 - \9 is a backreference, any multi-digit
11882 * escape which does not start with 0 and which when
11883 * evaluated as decimal could refer to an already
11884 * parsed capture buffer is a backslash. Anything else
11887 * Note this implies that \118 could be interpreted as
11888 * 118 OR as "\11" . "8" depending on whether there
11889 * were 118 capture buffers defined already in the
11892 /* NOTE, RExC_npar is 1 more than the actual number of
11893 * parens we have seen so far, hence the < RExC_npar below. */
11895 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
11896 { /* Not to be treated as an octal constant, go
11903 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11905 ender = grok_oct(p, &numlen, &flags, NULL);
11906 if (ender > 0xff) {
11910 if (SIZE_ONLY /* like \08, \178 */
11913 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11915 reg_warn_non_literal_string(
11917 form_short_octal_warning(p, numlen));
11920 if (PL_encoding && ender < 0x100)
11921 goto recode_encoding;
11924 if (! RExC_override_recoding) {
11925 SV* enc = PL_encoding;
11926 ender = reg_recode((const char)(U8)ender, &enc);
11927 if (!enc && SIZE_ONLY)
11928 ckWARNreg(p, "Invalid escape in the specified encoding");
11934 FAIL("Trailing \\");
11937 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11938 /* Include any { following the alpha to emphasize
11939 * that it could be part of an escape at some point
11941 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11942 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11944 goto normal_default;
11945 } /* End of switch on '\' */
11947 default: /* A literal character */
11950 && RExC_flags & RXf_PMf_EXTENDED
11951 && ckWARN_d(WARN_DEPRECATED)
11952 && is_PATWS_non_low_safe(p, RExC_end, UTF))
11954 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11955 "Escape literal pattern white space under /x");
11959 if (UTF8_IS_START(*p) && UTF) {
11961 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11962 &numlen, UTF8_ALLOW_DEFAULT);
11968 } /* End of switch on the literal */
11970 /* Here, have looked at the literal character and <ender>
11971 * contains its ordinal, <p> points to the character after it
11974 if ( RExC_flags & RXf_PMf_EXTENDED)
11975 p = regwhite( pRExC_state, p );
11977 /* If the next thing is a quantifier, it applies to this
11978 * character only, which means that this character has to be in
11979 * its own node and can't just be appended to the string in an
11980 * existing node, so if there are already other characters in
11981 * the node, close the node with just them, and set up to do
11982 * this character again next time through, when it will be the
11983 * only thing in its new node */
11984 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11990 if (! FOLD /* The simple case, just append the literal */
11991 || (LOC /* Also don't fold for tricky chars under /l */
11992 && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)))
11995 const STRLEN unilen = reguni(pRExC_state, ender, s);
12001 /* The loop increments <len> each time, as all but this
12002 * path (and one other) through it add a single byte to
12003 * the EXACTish node. But this one has changed len to
12004 * be the correct final value, so subtract one to
12005 * cancel out the increment that follows */
12009 REGC((char)ender, s++);
12012 /* Can get here if folding only if is one of the /l
12013 * characters whose fold depends on the locale. The
12014 * occurrence of any of these indicate that we can't
12015 * simplify things */
12017 maybe_exact = FALSE;
12018 maybe_exactfu = FALSE;
12023 /* See comments for join_exact() as to why we fold this
12024 * non-UTF at compile time */
12025 || (node_type == EXACTFU
12026 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12028 /* Here, are folding and are not UTF-8 encoded; therefore
12029 * the character must be in the range 0-255, and is not /l
12030 * (Not /l because we already handled these under /l in
12031 * is_PROBLEMATIC_LOCALE_FOLD_cp */
12032 if (IS_IN_SOME_FOLD_L1(ender)) {
12033 maybe_exact = FALSE;
12035 /* See if the character's fold differs between /d and
12036 * /u. This includes the multi-char fold SHARP S to
12039 && (PL_fold[ender] != PL_fold_latin1[ender]
12040 || ender == LATIN_SMALL_LETTER_SHARP_S
12042 && isARG2_lower_or_UPPER_ARG1('s', ender)
12043 && isARG2_lower_or_UPPER_ARG1('s',
12046 maybe_exactfu = FALSE;
12050 /* Even when folding, we store just the input character, as
12051 * we have an array that finds its fold quickly */
12052 *(s++) = (char) ender;
12054 else { /* FOLD and UTF */
12055 /* Unlike the non-fold case, we do actually have to
12056 * calculate the results here in pass 1. This is for two
12057 * reasons, the folded length may be longer than the
12058 * unfolded, and we have to calculate how many EXACTish
12059 * nodes it will take; and we may run out of room in a node
12060 * in the middle of a potential multi-char fold, and have
12061 * to back off accordingly. (Hence we can't use REGC for
12062 * the simple case just below.) */
12065 if (isASCII(ender)) {
12066 folded = toFOLD(ender);
12067 *(s)++ = (U8) folded;
12072 folded = _to_uni_fold_flags(
12076 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12077 ? FOLD_FLAGS_NOMIX_ASCII
12081 /* The loop increments <len> each time, as all but this
12082 * path (and one other) through it add a single byte to
12083 * the EXACTish node. But this one has changed len to
12084 * be the correct final value, so subtract one to
12085 * cancel out the increment that follows */
12086 len += foldlen - 1;
12088 /* If this node only contains non-folding code points so
12089 * far, see if this new one is also non-folding */
12091 if (folded != ender) {
12092 maybe_exact = FALSE;
12095 /* Here the fold is the original; we have to check
12096 * further to see if anything folds to it */
12097 if (_invlist_contains_cp(PL_utf8_foldable,
12100 maybe_exact = FALSE;
12107 if (next_is_quantifier) {
12109 /* Here, the next input is a quantifier, and to get here,
12110 * the current character is the only one in the node.
12111 * Also, here <len> doesn't include the final byte for this
12117 } /* End of loop through literal characters */
12119 /* Here we have either exhausted the input or ran out of room in
12120 * the node. (If we encountered a character that can't be in the
12121 * node, transfer is made directly to <loopdone>, and so we
12122 * wouldn't have fallen off the end of the loop.) In the latter
12123 * case, we artificially have to split the node into two, because
12124 * we just don't have enough space to hold everything. This
12125 * creates a problem if the final character participates in a
12126 * multi-character fold in the non-final position, as a match that
12127 * should have occurred won't, due to the way nodes are matched,
12128 * and our artificial boundary. So back off until we find a non-
12129 * problematic character -- one that isn't at the beginning or
12130 * middle of such a fold. (Either it doesn't participate in any
12131 * folds, or appears only in the final position of all the folds it
12132 * does participate in.) A better solution with far fewer false
12133 * positives, and that would fill the nodes more completely, would
12134 * be to actually have available all the multi-character folds to
12135 * test against, and to back-off only far enough to be sure that
12136 * this node isn't ending with a partial one. <upper_parse> is set
12137 * further below (if we need to reparse the node) to include just
12138 * up through that final non-problematic character that this code
12139 * identifies, so when it is set to less than the full node, we can
12140 * skip the rest of this */
12141 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12143 const STRLEN full_len = len;
12145 assert(len >= MAX_NODE_STRING_SIZE);
12147 /* Here, <s> points to the final byte of the final character.
12148 * Look backwards through the string until find a non-
12149 * problematic character */
12153 /* This has no multi-char folds to non-UTF characters */
12154 if (ASCII_FOLD_RESTRICTED) {
12158 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12162 if (! PL_NonL1NonFinalFold) {
12163 PL_NonL1NonFinalFold = _new_invlist_C_array(
12164 NonL1_Perl_Non_Final_Folds_invlist);
12167 /* Point to the first byte of the final character */
12168 s = (char *) utf8_hop((U8 *) s, -1);
12170 while (s >= s0) { /* Search backwards until find
12171 non-problematic char */
12172 if (UTF8_IS_INVARIANT(*s)) {
12174 /* There are no ascii characters that participate
12175 * in multi-char folds under /aa. In EBCDIC, the
12176 * non-ascii invariants are all control characters,
12177 * so don't ever participate in any folds. */
12178 if (ASCII_FOLD_RESTRICTED
12179 || ! IS_NON_FINAL_FOLD(*s))
12184 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12185 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12191 else if (! _invlist_contains_cp(
12192 PL_NonL1NonFinalFold,
12193 valid_utf8_to_uvchr((U8 *) s, NULL)))
12198 /* Here, the current character is problematic in that
12199 * it does occur in the non-final position of some
12200 * fold, so try the character before it, but have to
12201 * special case the very first byte in the string, so
12202 * we don't read outside the string */
12203 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12204 } /* End of loop backwards through the string */
12206 /* If there were only problematic characters in the string,
12207 * <s> will point to before s0, in which case the length
12208 * should be 0, otherwise include the length of the
12209 * non-problematic character just found */
12210 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12213 /* Here, have found the final character, if any, that is
12214 * non-problematic as far as ending the node without splitting
12215 * it across a potential multi-char fold. <len> contains the
12216 * number of bytes in the node up-to and including that
12217 * character, or is 0 if there is no such character, meaning
12218 * the whole node contains only problematic characters. In
12219 * this case, give up and just take the node as-is. We can't
12224 /* If the node ends in an 's' we make sure it stays EXACTF,
12225 * as if it turns into an EXACTFU, it could later get
12226 * joined with another 's' that would then wrongly match
12228 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
12230 maybe_exactfu = FALSE;
12234 /* Here, the node does contain some characters that aren't
12235 * problematic. If one such is the final character in the
12236 * node, we are done */
12237 if (len == full_len) {
12240 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12242 /* If the final character is problematic, but the
12243 * penultimate is not, back-off that last character to
12244 * later start a new node with it */
12249 /* Here, the final non-problematic character is earlier
12250 * in the input than the penultimate character. What we do
12251 * is reparse from the beginning, going up only as far as
12252 * this final ok one, thus guaranteeing that the node ends
12253 * in an acceptable character. The reason we reparse is
12254 * that we know how far in the character is, but we don't
12255 * know how to correlate its position with the input parse.
12256 * An alternate implementation would be to build that
12257 * correlation as we go along during the original parse,
12258 * but that would entail extra work for every node, whereas
12259 * this code gets executed only when the string is too
12260 * large for the node, and the final two characters are
12261 * problematic, an infrequent occurrence. Yet another
12262 * possible strategy would be to save the tail of the
12263 * string, and the next time regatom is called, initialize
12264 * with that. The problem with this is that unless you
12265 * back off one more character, you won't be guaranteed
12266 * regatom will get called again, unless regbranch,
12267 * regpiece ... are also changed. If you do back off that
12268 * extra character, so that there is input guaranteed to
12269 * force calling regatom, you can't handle the case where
12270 * just the first character in the node is acceptable. I
12271 * (khw) decided to try this method which doesn't have that
12272 * pitfall; if performance issues are found, we can do a
12273 * combination of the current approach plus that one */
12279 } /* End of verifying node ends with an appropriate char */
12281 loopdone: /* Jumped to when encounters something that shouldn't be in
12284 /* I (khw) don't know if you can get here with zero length, but the
12285 * old code handled this situation by creating a zero-length EXACT
12286 * node. Might as well be NOTHING instead */
12292 /* If 'maybe_exact' is still set here, means there are no
12293 * code points in the node that participate in folds;
12294 * similarly for 'maybe_exactfu' and code points that match
12295 * differently depending on UTF8ness of the target string
12296 * (for /u), or depending on locale for /l */
12300 else if (maybe_exactfu) {
12304 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12305 FALSE /* Don't look to see if could
12306 be turned into an EXACT
12307 node, as we have already
12312 RExC_parse = p - 1;
12313 Set_Node_Cur_Length(ret, parse_start);
12314 nextchar(pRExC_state);
12316 /* len is STRLEN which is unsigned, need to copy to signed */
12319 vFAIL("Internal disaster");
12322 } /* End of label 'defchar:' */
12324 } /* End of giant switch on input character */
12330 S_regwhite( RExC_state_t *pRExC_state, char *p )
12332 const char *e = RExC_end;
12334 PERL_ARGS_ASSERT_REGWHITE;
12339 else if (*p == '#') {
12342 if (*p++ == '\n') {
12348 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12357 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12359 /* Returns the next non-pattern-white space, non-comment character (the
12360 * latter only if 'recognize_comment is true) in the string p, which is
12361 * ended by RExC_end. If there is no line break ending a comment,
12362 * RExC_seen has added the REG_RUN_ON_COMMENT_SEEN flag; */
12363 const char *e = RExC_end;
12365 PERL_ARGS_ASSERT_REGPATWS;
12369 if ((len = is_PATWS_safe(p, e, UTF))) {
12372 else if (recognize_comment && *p == '#') {
12376 if (is_LNBREAK_safe(p, e, UTF)) {
12382 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12391 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12393 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12394 * sets up the bitmap and any flags, removing those code points from the
12395 * inversion list, setting it to NULL should it become completely empty */
12397 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12398 assert(PL_regkind[OP(node)] == ANYOF);
12400 ANYOF_BITMAP_ZERO(node);
12401 if (*invlist_ptr) {
12403 /* This gets set if we actually need to modify things */
12404 bool change_invlist = FALSE;
12408 /* Start looking through *invlist_ptr */
12409 invlist_iterinit(*invlist_ptr);
12410 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12414 if (end == UV_MAX && start <= 256) {
12415 ANYOF_FLAGS(node) |= ANYOF_ABOVE_LATIN1_ALL;
12417 else if (end >= 256) {
12418 ANYOF_FLAGS(node) |= ANYOF_UTF8;
12421 /* Quit if are above what we should change */
12426 change_invlist = TRUE;
12428 /* Set all the bits in the range, up to the max that we are doing */
12429 high = (end < 255) ? end : 255;
12430 for (i = start; i <= (int) high; i++) {
12431 if (! ANYOF_BITMAP_TEST(node, i)) {
12432 ANYOF_BITMAP_SET(node, i);
12436 invlist_iterfinish(*invlist_ptr);
12438 /* Done with loop; remove any code points that are in the bitmap from
12439 * *invlist_ptr; similarly for code points above latin1 if we have a
12440 * flag to match all of them anyways */
12441 if (change_invlist) {
12442 _invlist_subtract(*invlist_ptr, PL_Latin1, invlist_ptr);
12444 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
12445 _invlist_intersection(*invlist_ptr, PL_Latin1, invlist_ptr);
12448 /* If have completely emptied it, remove it completely */
12449 if (_invlist_len(*invlist_ptr) == 0) {
12450 SvREFCNT_dec_NN(*invlist_ptr);
12451 *invlist_ptr = NULL;
12456 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12457 Character classes ([:foo:]) can also be negated ([:^foo:]).
12458 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12459 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12460 but trigger failures because they are currently unimplemented. */
12462 #define POSIXCC_DONE(c) ((c) == ':')
12463 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12464 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12466 PERL_STATIC_INLINE I32
12467 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12470 I32 namedclass = OOB_NAMEDCLASS;
12472 PERL_ARGS_ASSERT_REGPPOSIXCC;
12474 if (value == '[' && RExC_parse + 1 < RExC_end &&
12475 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12476 POSIXCC(UCHARAT(RExC_parse)))
12478 const char c = UCHARAT(RExC_parse);
12479 char* const s = RExC_parse++;
12481 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12483 if (RExC_parse == RExC_end) {
12486 /* Try to give a better location for the error (than the end of
12487 * the string) by looking for the matching ']' */
12489 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12492 vFAIL2("Unmatched '%c' in POSIX class", c);
12494 /* Grandfather lone [:, [=, [. */
12498 const char* const t = RExC_parse++; /* skip over the c */
12501 if (UCHARAT(RExC_parse) == ']') {
12502 const char *posixcc = s + 1;
12503 RExC_parse++; /* skip over the ending ] */
12506 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12507 const I32 skip = t - posixcc;
12509 /* Initially switch on the length of the name. */
12512 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12513 this is the Perl \w
12515 namedclass = ANYOF_WORDCHAR;
12518 /* Names all of length 5. */
12519 /* alnum alpha ascii blank cntrl digit graph lower
12520 print punct space upper */
12521 /* Offset 4 gives the best switch position. */
12522 switch (posixcc[4]) {
12524 if (memEQ(posixcc, "alph", 4)) /* alpha */
12525 namedclass = ANYOF_ALPHA;
12528 if (memEQ(posixcc, "spac", 4)) /* space */
12529 namedclass = ANYOF_PSXSPC;
12532 if (memEQ(posixcc, "grap", 4)) /* graph */
12533 namedclass = ANYOF_GRAPH;
12536 if (memEQ(posixcc, "asci", 4)) /* ascii */
12537 namedclass = ANYOF_ASCII;
12540 if (memEQ(posixcc, "blan", 4)) /* blank */
12541 namedclass = ANYOF_BLANK;
12544 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12545 namedclass = ANYOF_CNTRL;
12548 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12549 namedclass = ANYOF_ALPHANUMERIC;
12552 if (memEQ(posixcc, "lowe", 4)) /* lower */
12553 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12554 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12555 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12558 if (memEQ(posixcc, "digi", 4)) /* digit */
12559 namedclass = ANYOF_DIGIT;
12560 else if (memEQ(posixcc, "prin", 4)) /* print */
12561 namedclass = ANYOF_PRINT;
12562 else if (memEQ(posixcc, "punc", 4)) /* punct */
12563 namedclass = ANYOF_PUNCT;
12568 if (memEQ(posixcc, "xdigit", 6))
12569 namedclass = ANYOF_XDIGIT;
12573 if (namedclass == OOB_NAMEDCLASS)
12575 "POSIX class [:%"UTF8f":] unknown",
12576 UTF8fARG(UTF, t - s - 1, s + 1));
12578 /* The #defines are structured so each complement is +1 to
12579 * the normal one */
12583 assert (posixcc[skip] == ':');
12584 assert (posixcc[skip+1] == ']');
12585 } else if (!SIZE_ONLY) {
12586 /* [[=foo=]] and [[.foo.]] are still future. */
12588 /* adjust RExC_parse so the warning shows after
12589 the class closes */
12590 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
12592 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
12595 /* Maternal grandfather:
12596 * "[:" ending in ":" but not in ":]" */
12598 vFAIL("Unmatched '[' in POSIX class");
12601 /* Grandfather lone [:, [=, [. */
12611 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
12613 /* This applies some heuristics at the current parse position (which should
12614 * be at a '[') to see if what follows might be intended to be a [:posix:]
12615 * class. It returns true if it really is a posix class, of course, but it
12616 * also can return true if it thinks that what was intended was a posix
12617 * class that didn't quite make it.
12619 * It will return true for
12621 * [:alphanumerics] (as long as the ] isn't followed immediately by a
12622 * ')' indicating the end of the (?[
12623 * [:any garbage including %^&$ punctuation:]
12625 * This is designed to be called only from S_handle_regex_sets; it could be
12626 * easily adapted to be called from the spot at the beginning of regclass()
12627 * that checks to see in a normal bracketed class if the surrounding []
12628 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
12629 * change long-standing behavior, so I (khw) didn't do that */
12630 char* p = RExC_parse + 1;
12631 char first_char = *p;
12633 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
12635 assert(*(p - 1) == '[');
12637 if (! POSIXCC(first_char)) {
12642 while (p < RExC_end && isWORDCHAR(*p)) p++;
12644 if (p >= RExC_end) {
12648 if (p - RExC_parse > 2 /* Got at least 1 word character */
12649 && (*p == first_char
12650 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
12655 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
12658 && p - RExC_parse > 2 /* [:] evaluates to colon;
12659 [::] is a bad posix class. */
12660 && first_char == *(p - 1));
12664 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
12665 I32 *flagp, U32 depth,
12666 char * const oregcomp_parse)
12668 /* Handle the (?[...]) construct to do set operations */
12671 UV start, end; /* End points of code point ranges */
12673 char *save_end, *save_parse;
12678 const bool save_fold = FOLD;
12680 GET_RE_DEBUG_FLAGS_DECL;
12682 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
12685 vFAIL("(?[...]) not valid in locale");
12687 RExC_uni_semantics = 1;
12689 /* This will return only an ANYOF regnode, or (unlikely) something smaller
12690 * (such as EXACT). Thus we can skip most everything if just sizing. We
12691 * call regclass to handle '[]' so as to not have to reinvent its parsing
12692 * rules here (throwing away the size it computes each time). And, we exit
12693 * upon an unescaped ']' that isn't one ending a regclass. To do both
12694 * these things, we need to realize that something preceded by a backslash
12695 * is escaped, so we have to keep track of backslashes */
12697 UV depth = 0; /* how many nested (?[...]) constructs */
12699 Perl_ck_warner_d(aTHX_
12700 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
12701 "The regex_sets feature is experimental" REPORT_LOCATION,
12702 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
12704 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
12705 RExC_precomp + (RExC_parse - RExC_precomp)));
12707 while (RExC_parse < RExC_end) {
12708 SV* current = NULL;
12709 RExC_parse = regpatws(pRExC_state, RExC_parse,
12710 TRUE); /* means recognize comments */
12711 switch (*RExC_parse) {
12713 if (RExC_parse[1] == '[') depth++, RExC_parse++;
12718 /* Skip the next byte (which could cause us to end up in
12719 * the middle of a UTF-8 character, but since none of those
12720 * are confusable with anything we currently handle in this
12721 * switch (invariants all), it's safe. We'll just hit the
12722 * default: case next time and keep on incrementing until
12723 * we find one of the invariants we do handle. */
12728 /* If this looks like it is a [:posix:] class, leave the
12729 * parse pointer at the '[' to fool regclass() into
12730 * thinking it is part of a '[[:posix:]]'. That function
12731 * will use strict checking to force a syntax error if it
12732 * doesn't work out to a legitimate class */
12733 bool is_posix_class
12734 = could_it_be_a_POSIX_class(pRExC_state);
12735 if (! is_posix_class) {
12739 /* regclass() can only return RESTART_UTF8 if multi-char
12740 folds are allowed. */
12741 if (!regclass(pRExC_state, flagp,depth+1,
12742 is_posix_class, /* parse the whole char
12743 class only if not a
12745 FALSE, /* don't allow multi-char folds */
12746 TRUE, /* silence non-portable warnings. */
12748 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12751 /* function call leaves parse pointing to the ']', except
12752 * if we faked it */
12753 if (is_posix_class) {
12757 SvREFCNT_dec(current); /* In case it returned something */
12762 if (depth--) break;
12764 if (RExC_parse < RExC_end
12765 && *RExC_parse == ')')
12767 node = reganode(pRExC_state, ANYOF, 0);
12768 RExC_size += ANYOF_SKIP;
12769 nextchar(pRExC_state);
12770 Set_Node_Length(node,
12771 RExC_parse - oregcomp_parse + 1); /* MJD */
12780 FAIL("Syntax error in (?[...])");
12783 /* Pass 2 only after this. Everything in this construct is a
12784 * metacharacter. Operands begin with either a '\' (for an escape
12785 * sequence), or a '[' for a bracketed character class. Any other
12786 * character should be an operator, or parenthesis for grouping. Both
12787 * types of operands are handled by calling regclass() to parse them. It
12788 * is called with a parameter to indicate to return the computed inversion
12789 * list. The parsing here is implemented via a stack. Each entry on the
12790 * stack is a single character representing one of the operators, or the
12791 * '('; or else a pointer to an operand inversion list. */
12793 #define IS_OPERAND(a) (! SvIOK(a))
12795 /* The stack starts empty. It is a syntax error if the first thing parsed
12796 * is a binary operator; everything else is pushed on the stack. When an
12797 * operand is parsed, the top of the stack is examined. If it is a binary
12798 * operator, the item before it should be an operand, and both are replaced
12799 * by the result of doing that operation on the new operand and the one on
12800 * the stack. Thus a sequence of binary operands is reduced to a single
12801 * one before the next one is parsed.
12803 * A unary operator may immediately follow a binary in the input, for
12806 * When an operand is parsed and the top of the stack is a unary operator,
12807 * the operation is performed, and then the stack is rechecked to see if
12808 * this new operand is part of a binary operation; if so, it is handled as
12811 * A '(' is simply pushed on the stack; it is valid only if the stack is
12812 * empty, or the top element of the stack is an operator or another '('
12813 * (for which the parenthesized expression will become an operand). By the
12814 * time the corresponding ')' is parsed everything in between should have
12815 * been parsed and evaluated to a single operand (or else is a syntax
12816 * error), and is handled as a regular operand */
12818 sv_2mortal((SV *)(stack = newAV()));
12820 while (RExC_parse < RExC_end) {
12821 I32 top_index = av_tindex(stack);
12823 SV* current = NULL;
12825 /* Skip white space */
12826 RExC_parse = regpatws(pRExC_state, RExC_parse,
12827 TRUE); /* means recognize comments */
12828 if (RExC_parse >= RExC_end) {
12829 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
12831 if ((curchar = UCHARAT(RExC_parse)) == ']') {
12838 if (av_tindex(stack) >= 0 /* This makes sure that we can
12839 safely subtract 1 from
12840 RExC_parse in the next clause.
12841 If we have something on the
12842 stack, we have parsed something
12844 && UCHARAT(RExC_parse - 1) == '('
12845 && RExC_parse < RExC_end)
12847 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
12848 * This happens when we have some thing like
12850 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
12852 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
12854 * Here we would be handling the interpolated
12855 * '$thai_or_lao'. We handle this by a recursive call to
12856 * ourselves which returns the inversion list the
12857 * interpolated expression evaluates to. We use the flags
12858 * from the interpolated pattern. */
12859 U32 save_flags = RExC_flags;
12860 const char * const save_parse = ++RExC_parse;
12862 parse_lparen_question_flags(pRExC_state);
12864 if (RExC_parse == save_parse /* Makes sure there was at
12865 least one flag (or this
12866 embedding wasn't compiled)
12868 || RExC_parse >= RExC_end - 4
12869 || UCHARAT(RExC_parse) != ':'
12870 || UCHARAT(++RExC_parse) != '('
12871 || UCHARAT(++RExC_parse) != '?'
12872 || UCHARAT(++RExC_parse) != '[')
12875 /* In combination with the above, this moves the
12876 * pointer to the point just after the first erroneous
12877 * character (or if there are no flags, to where they
12878 * should have been) */
12879 if (RExC_parse >= RExC_end - 4) {
12880 RExC_parse = RExC_end;
12882 else if (RExC_parse != save_parse) {
12883 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12885 vFAIL("Expecting '(?flags:(?[...'");
12888 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
12889 depth+1, oregcomp_parse);
12891 /* Here, 'current' contains the embedded expression's
12892 * inversion list, and RExC_parse points to the trailing
12893 * ']'; the next character should be the ')' which will be
12894 * paired with the '(' that has been put on the stack, so
12895 * the whole embedded expression reduces to '(operand)' */
12898 RExC_flags = save_flags;
12899 goto handle_operand;
12904 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12905 vFAIL("Unexpected character");
12908 /* regclass() can only return RESTART_UTF8 if multi-char
12909 folds are allowed. */
12910 if (!regclass(pRExC_state, flagp,depth+1,
12911 TRUE, /* means parse just the next thing */
12912 FALSE, /* don't allow multi-char folds */
12913 FALSE, /* don't silence non-portable warnings. */
12915 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12917 /* regclass() will return with parsing just the \ sequence,
12918 * leaving the parse pointer at the next thing to parse */
12920 goto handle_operand;
12922 case '[': /* Is a bracketed character class */
12924 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
12926 if (! is_posix_class) {
12930 /* regclass() can only return RESTART_UTF8 if multi-char
12931 folds are allowed. */
12932 if(!regclass(pRExC_state, flagp,depth+1,
12933 is_posix_class, /* parse the whole char class
12934 only if not a posix class */
12935 FALSE, /* don't allow multi-char folds */
12936 FALSE, /* don't silence non-portable warnings. */
12938 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12940 /* function call leaves parse pointing to the ']', except if we
12942 if (is_posix_class) {
12946 goto handle_operand;
12955 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
12956 || ! IS_OPERAND(*top_ptr))
12959 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
12961 av_push(stack, newSVuv(curchar));
12965 av_push(stack, newSVuv(curchar));
12969 if (top_index >= 0) {
12970 top_ptr = av_fetch(stack, top_index, FALSE);
12972 if (IS_OPERAND(*top_ptr)) {
12974 vFAIL("Unexpected '(' with no preceding operator");
12977 av_push(stack, newSVuv(curchar));
12984 || ! (current = av_pop(stack))
12985 || ! IS_OPERAND(current)
12986 || ! (lparen = av_pop(stack))
12987 || IS_OPERAND(lparen)
12988 || SvUV(lparen) != '(')
12990 SvREFCNT_dec(current);
12992 vFAIL("Unexpected ')'");
12995 SvREFCNT_dec_NN(lparen);
13002 /* Here, we have an operand to process, in 'current' */
13004 if (top_index < 0) { /* Just push if stack is empty */
13005 av_push(stack, current);
13008 SV* top = av_pop(stack);
13010 char current_operator;
13012 if (IS_OPERAND(top)) {
13013 SvREFCNT_dec_NN(top);
13014 SvREFCNT_dec_NN(current);
13015 vFAIL("Operand with no preceding operator");
13017 current_operator = (char) SvUV(top);
13018 switch (current_operator) {
13019 case '(': /* Push the '(' back on followed by the new
13021 av_push(stack, top);
13022 av_push(stack, current);
13023 SvREFCNT_inc(top); /* Counters the '_dec' done
13024 just after the 'break', so
13025 it doesn't get wrongly freed
13030 _invlist_invert(current);
13032 /* Unlike binary operators, the top of the stack,
13033 * now that this unary one has been popped off, may
13034 * legally be an operator, and we now have operand
13037 SvREFCNT_dec_NN(top);
13038 goto handle_operand;
13041 prev = av_pop(stack);
13042 _invlist_intersection(prev,
13045 av_push(stack, current);
13050 prev = av_pop(stack);
13051 _invlist_union(prev, current, ¤t);
13052 av_push(stack, current);
13056 prev = av_pop(stack);;
13057 _invlist_subtract(prev, current, ¤t);
13058 av_push(stack, current);
13061 case '^': /* The union minus the intersection */
13067 prev = av_pop(stack);
13068 _invlist_union(prev, current, &u);
13069 _invlist_intersection(prev, current, &i);
13070 /* _invlist_subtract will overwrite current
13071 without freeing what it already contains */
13073 _invlist_subtract(u, i, ¤t);
13074 av_push(stack, current);
13075 SvREFCNT_dec_NN(i);
13076 SvREFCNT_dec_NN(u);
13077 SvREFCNT_dec_NN(element);
13082 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13084 SvREFCNT_dec_NN(top);
13085 SvREFCNT_dec(prev);
13089 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13092 if (av_tindex(stack) < 0 /* Was empty */
13093 || ((final = av_pop(stack)) == NULL)
13094 || ! IS_OPERAND(final)
13095 || av_tindex(stack) >= 0) /* More left on stack */
13097 vFAIL("Incomplete expression within '(?[ ])'");
13100 /* Here, 'final' is the resultant inversion list from evaluating the
13101 * expression. Return it if so requested */
13102 if (return_invlist) {
13103 *return_invlist = final;
13107 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13108 * expecting a string of ranges and individual code points */
13109 invlist_iterinit(final);
13110 result_string = newSVpvs("");
13111 while (invlist_iternext(final, &start, &end)) {
13112 if (start == end) {
13113 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13116 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13121 save_parse = RExC_parse;
13122 RExC_parse = SvPV(result_string, len);
13123 save_end = RExC_end;
13124 RExC_end = RExC_parse + len;
13126 /* We turn off folding around the call, as the class we have constructed
13127 * already has all folding taken into consideration, and we don't want
13128 * regclass() to add to that */
13129 RExC_flags &= ~RXf_PMf_FOLD;
13130 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13132 node = regclass(pRExC_state, flagp,depth+1,
13133 FALSE, /* means parse the whole char class */
13134 FALSE, /* don't allow multi-char folds */
13135 TRUE, /* silence non-portable warnings. The above may very
13136 well have generated non-portable code points, but
13137 they're valid on this machine */
13140 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13143 RExC_flags |= RXf_PMf_FOLD;
13145 RExC_parse = save_parse + 1;
13146 RExC_end = save_end;
13147 SvREFCNT_dec_NN(final);
13148 SvREFCNT_dec_NN(result_string);
13150 nextchar(pRExC_state);
13151 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13156 /* The names of properties whose definitions are not known at compile time are
13157 * stored in this SV, after a constant heading. So if the length has been
13158 * changed since initialization, then there is a run-time definition. */
13159 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13160 (SvCUR(listsv) != initial_listsv_len)
13163 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13164 const bool stop_at_1, /* Just parse the next thing, don't
13165 look for a full character class */
13166 bool allow_multi_folds,
13167 const bool silence_non_portable, /* Don't output warnings
13170 SV** ret_invlist) /* Return an inversion list, not a node */
13172 /* parse a bracketed class specification. Most of these will produce an
13173 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13174 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13175 * under /i with multi-character folds: it will be rewritten following the
13176 * paradigm of this example, where the <multi-fold>s are characters which
13177 * fold to multiple character sequences:
13178 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13179 * gets effectively rewritten as:
13180 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13181 * reg() gets called (recursively) on the rewritten version, and this
13182 * function will return what it constructs. (Actually the <multi-fold>s
13183 * aren't physically removed from the [abcdefghi], it's just that they are
13184 * ignored in the recursion by means of a flag:
13185 * <RExC_in_multi_char_class>.)
13187 * ANYOF nodes contain a bit map for the first 256 characters, with the
13188 * corresponding bit set if that character is in the list. For characters
13189 * above 255, a range list or swash is used. There are extra bits for \w,
13190 * etc. in locale ANYOFs, as what these match is not determinable at
13193 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13194 * to be restarted. This can only happen if ret_invlist is non-NULL.
13198 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13200 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13203 IV namedclass = OOB_NAMEDCLASS;
13204 char *rangebegin = NULL;
13205 bool need_class = 0;
13207 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13208 than just initialized. */
13209 SV* properties = NULL; /* Code points that match \p{} \P{} */
13210 SV* posixes = NULL; /* Code points that match classes like [:word:],
13211 extended beyond the Latin1 range. These have to
13212 be kept separate from other code points for much
13213 of this function because their handling is
13214 different under /i, and for most classes under
13216 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13217 separate for a while from the non-complemented
13218 versions because of complications with /d
13220 UV element_count = 0; /* Number of distinct elements in the class.
13221 Optimizations may be possible if this is tiny */
13222 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13223 character; used under /i */
13225 char * stop_ptr = RExC_end; /* where to stop parsing */
13226 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13228 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13230 /* Unicode properties are stored in a swash; this holds the current one
13231 * being parsed. If this swash is the only above-latin1 component of the
13232 * character class, an optimization is to pass it directly on to the
13233 * execution engine. Otherwise, it is set to NULL to indicate that there
13234 * are other things in the class that have to be dealt with at execution
13236 SV* swash = NULL; /* Code points that match \p{} \P{} */
13238 /* Set if a component of this character class is user-defined; just passed
13239 * on to the engine */
13240 bool has_user_defined_property = FALSE;
13242 /* inversion list of code points this node matches only when the target
13243 * string is in UTF-8. (Because is under /d) */
13244 SV* depends_list = NULL;
13246 /* Inversion list of code points this node matches regardless of things
13247 * like locale, folding, utf8ness of the target string */
13248 SV* cp_list = NULL;
13250 /* Like cp_list, but code points on this list need to be checked for things
13251 * that fold to/from them under /i */
13252 SV* cp_foldable_list = NULL;
13254 /* Like cp_list, but code points on this list are valid only when the
13255 * runtime locale is UTF-8 */
13256 SV* only_utf8_locale_list = NULL;
13259 /* In a range, counts how many 0-2 of the ends of it came from literals,
13260 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13261 UV literal_endpoint = 0;
13263 bool invert = FALSE; /* Is this class to be complemented */
13265 bool warn_super = ALWAYS_WARN_SUPER;
13267 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13268 case we need to change the emitted regop to an EXACT. */
13269 const char * orig_parse = RExC_parse;
13270 const SSize_t orig_size = RExC_size;
13271 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13272 GET_RE_DEBUG_FLAGS_DECL;
13274 PERL_ARGS_ASSERT_REGCLASS;
13276 PERL_UNUSED_ARG(depth);
13279 DEBUG_PARSE("clas");
13281 /* Assume we are going to generate an ANYOF node. */
13282 ret = reganode(pRExC_state, ANYOF, 0);
13285 RExC_size += ANYOF_SKIP;
13286 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13289 ANYOF_FLAGS(ret) = 0;
13291 RExC_emit += ANYOF_SKIP;
13292 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13293 initial_listsv_len = SvCUR(listsv);
13294 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13298 RExC_parse = regpatws(pRExC_state, RExC_parse,
13299 FALSE /* means don't recognize comments */);
13302 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13305 allow_multi_folds = FALSE;
13308 RExC_parse = regpatws(pRExC_state, RExC_parse,
13309 FALSE /* means don't recognize comments */);
13313 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13314 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13315 const char *s = RExC_parse;
13316 const char c = *s++;
13318 while (isWORDCHAR(*s))
13320 if (*s && c == *s && s[1] == ']') {
13321 SAVEFREESV(RExC_rx_sv);
13323 "POSIX syntax [%c %c] belongs inside character classes",
13325 (void)ReREFCNT_inc(RExC_rx_sv);
13329 /* If the caller wants us to just parse a single element, accomplish this
13330 * by faking the loop ending condition */
13331 if (stop_at_1 && RExC_end > RExC_parse) {
13332 stop_ptr = RExC_parse + 1;
13335 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13336 if (UCHARAT(RExC_parse) == ']')
13337 goto charclassloop;
13341 if (RExC_parse >= stop_ptr) {
13346 RExC_parse = regpatws(pRExC_state, RExC_parse,
13347 FALSE /* means don't recognize comments */);
13350 if (UCHARAT(RExC_parse) == ']') {
13356 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13357 save_value = value;
13358 save_prevvalue = prevvalue;
13361 rangebegin = RExC_parse;
13365 value = utf8n_to_uvchr((U8*)RExC_parse,
13366 RExC_end - RExC_parse,
13367 &numlen, UTF8_ALLOW_DEFAULT);
13368 RExC_parse += numlen;
13371 value = UCHARAT(RExC_parse++);
13374 && RExC_parse < RExC_end
13375 && POSIXCC(UCHARAT(RExC_parse)))
13377 namedclass = regpposixcc(pRExC_state, value, strict);
13379 else if (value == '\\') {
13381 value = utf8n_to_uvchr((U8*)RExC_parse,
13382 RExC_end - RExC_parse,
13383 &numlen, UTF8_ALLOW_DEFAULT);
13384 RExC_parse += numlen;
13387 value = UCHARAT(RExC_parse++);
13389 /* Some compilers cannot handle switching on 64-bit integer
13390 * values, therefore value cannot be an UV. Yes, this will
13391 * be a problem later if we want switch on Unicode.
13392 * A similar issue a little bit later when switching on
13393 * namedclass. --jhi */
13395 /* If the \ is escaping white space when white space is being
13396 * skipped, it means that that white space is wanted literally, and
13397 * is already in 'value'. Otherwise, need to translate the escape
13398 * into what it signifies. */
13399 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13401 case 'w': namedclass = ANYOF_WORDCHAR; break;
13402 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13403 case 's': namedclass = ANYOF_SPACE; break;
13404 case 'S': namedclass = ANYOF_NSPACE; break;
13405 case 'd': namedclass = ANYOF_DIGIT; break;
13406 case 'D': namedclass = ANYOF_NDIGIT; break;
13407 case 'v': namedclass = ANYOF_VERTWS; break;
13408 case 'V': namedclass = ANYOF_NVERTWS; break;
13409 case 'h': namedclass = ANYOF_HORIZWS; break;
13410 case 'H': namedclass = ANYOF_NHORIZWS; break;
13411 case 'N': /* Handle \N{NAME} in class */
13413 /* We only pay attention to the first char of
13414 multichar strings being returned. I kinda wonder
13415 if this makes sense as it does change the behaviour
13416 from earlier versions, OTOH that behaviour was broken
13418 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
13419 TRUE, /* => charclass */
13422 if (*flagp & RESTART_UTF8)
13423 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13433 /* We will handle any undefined properties ourselves */
13434 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
13435 /* And we actually would prefer to get
13436 * the straight inversion list of the
13437 * swash, since we will be accessing it
13438 * anyway, to save a little time */
13439 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
13441 if (RExC_parse >= RExC_end)
13442 vFAIL2("Empty \\%c{}", (U8)value);
13443 if (*RExC_parse == '{') {
13444 const U8 c = (U8)value;
13445 e = strchr(RExC_parse++, '}');
13447 vFAIL2("Missing right brace on \\%c{}", c);
13448 while (isSPACE(UCHARAT(RExC_parse)))
13450 if (e == RExC_parse)
13451 vFAIL2("Empty \\%c{}", c);
13452 n = e - RExC_parse;
13453 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
13465 if (UCHARAT(RExC_parse) == '^') {
13468 /* toggle. (The rhs xor gets the single bit that
13469 * differs between P and p; the other xor inverts just
13471 value ^= 'P' ^ 'p';
13473 while (isSPACE(UCHARAT(RExC_parse))) {
13478 /* Try to get the definition of the property into
13479 * <invlist>. If /i is in effect, the effective property
13480 * will have its name be <__NAME_i>. The design is
13481 * discussed in commit
13482 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
13483 formatted = Perl_form(aTHX_
13485 (FOLD) ? "__" : "",
13490 name = savepvn(formatted, strlen(formatted));
13492 /* Look up the property name, and get its swash and
13493 * inversion list, if the property is found */
13495 SvREFCNT_dec_NN(swash);
13497 swash = _core_swash_init("utf8", name, &PL_sv_undef,
13500 NULL, /* No inversion list */
13503 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
13505 SvREFCNT_dec_NN(swash);
13509 /* Here didn't find it. It could be a user-defined
13510 * property that will be available at run-time. If we
13511 * accept only compile-time properties, is an error;
13512 * otherwise add it to the list for run-time look up */
13514 RExC_parse = e + 1;
13516 "Property '%"UTF8f"' is unknown",
13517 UTF8fARG(UTF, n, name));
13519 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
13520 (value == 'p' ? '+' : '!'),
13521 UTF8fARG(UTF, n, name));
13522 has_user_defined_property = TRUE;
13524 /* We don't know yet, so have to assume that the
13525 * property could match something in the Latin1 range,
13526 * hence something that isn't utf8. Note that this
13527 * would cause things in <depends_list> to match
13528 * inappropriately, except that any \p{}, including
13529 * this one forces Unicode semantics, which means there
13530 * is no <depends_list> */
13531 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
13535 /* Here, did get the swash and its inversion list. If
13536 * the swash is from a user-defined property, then this
13537 * whole character class should be regarded as such */
13538 if (swash_init_flags
13539 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
13541 has_user_defined_property = TRUE;
13544 /* We warn on matching an above-Unicode code point
13545 * if the match would return true, except don't
13546 * warn for \p{All}, which has exactly one element
13548 (_invlist_contains_cp(invlist, 0x110000)
13549 && (! (_invlist_len(invlist) == 1
13550 && *invlist_array(invlist) == 0)))
13556 /* Invert if asking for the complement */
13557 if (value == 'P') {
13558 _invlist_union_complement_2nd(properties,
13562 /* The swash can't be used as-is, because we've
13563 * inverted things; delay removing it to here after
13564 * have copied its invlist above */
13565 SvREFCNT_dec_NN(swash);
13569 _invlist_union(properties, invlist, &properties);
13574 RExC_parse = e + 1;
13575 namedclass = ANYOF_UNIPROP; /* no official name, but it's
13578 /* \p means they want Unicode semantics */
13579 RExC_uni_semantics = 1;
13582 case 'n': value = '\n'; break;
13583 case 'r': value = '\r'; break;
13584 case 't': value = '\t'; break;
13585 case 'f': value = '\f'; break;
13586 case 'b': value = '\b'; break;
13587 case 'e': value = ASCII_TO_NATIVE('\033');break;
13588 case 'a': value = '\a'; break;
13590 RExC_parse--; /* function expects to be pointed at the 'o' */
13592 const char* error_msg;
13593 bool valid = grok_bslash_o(&RExC_parse,
13596 SIZE_ONLY, /* warnings in pass
13599 silence_non_portable,
13605 if (PL_encoding && value < 0x100) {
13606 goto recode_encoding;
13610 RExC_parse--; /* function expects to be pointed at the 'x' */
13612 const char* error_msg;
13613 bool valid = grok_bslash_x(&RExC_parse,
13616 TRUE, /* Output warnings */
13618 silence_non_portable,
13624 if (PL_encoding && value < 0x100)
13625 goto recode_encoding;
13628 value = grok_bslash_c(*RExC_parse++, SIZE_ONLY);
13630 case '0': case '1': case '2': case '3': case '4':
13631 case '5': case '6': case '7':
13633 /* Take 1-3 octal digits */
13634 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13635 numlen = (strict) ? 4 : 3;
13636 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
13637 RExC_parse += numlen;
13640 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13641 vFAIL("Need exactly 3 octal digits");
13643 else if (! SIZE_ONLY /* like \08, \178 */
13645 && RExC_parse < RExC_end
13646 && isDIGIT(*RExC_parse)
13647 && ckWARN(WARN_REGEXP))
13649 SAVEFREESV(RExC_rx_sv);
13650 reg_warn_non_literal_string(
13652 form_short_octal_warning(RExC_parse, numlen));
13653 (void)ReREFCNT_inc(RExC_rx_sv);
13656 if (PL_encoding && value < 0x100)
13657 goto recode_encoding;
13661 if (! RExC_override_recoding) {
13662 SV* enc = PL_encoding;
13663 value = reg_recode((const char)(U8)value, &enc);
13666 vFAIL("Invalid escape in the specified encoding");
13668 else if (SIZE_ONLY) {
13669 ckWARNreg(RExC_parse,
13670 "Invalid escape in the specified encoding");
13676 /* Allow \_ to not give an error */
13677 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
13679 vFAIL2("Unrecognized escape \\%c in character class",
13683 SAVEFREESV(RExC_rx_sv);
13684 ckWARN2reg(RExC_parse,
13685 "Unrecognized escape \\%c in character class passed through",
13687 (void)ReREFCNT_inc(RExC_rx_sv);
13691 } /* End of switch on char following backslash */
13692 } /* end of handling backslash escape sequences */
13695 literal_endpoint++;
13698 /* Here, we have the current token in 'value' */
13700 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
13703 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
13704 * literal, as is the character that began the false range, i.e.
13705 * the 'a' in the examples */
13708 const int w = (RExC_parse >= rangebegin)
13709 ? RExC_parse - rangebegin
13713 "False [] range \"%"UTF8f"\"",
13714 UTF8fARG(UTF, w, rangebegin));
13717 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
13718 ckWARN2reg(RExC_parse,
13719 "False [] range \"%"UTF8f"\"",
13720 UTF8fARG(UTF, w, rangebegin));
13721 (void)ReREFCNT_inc(RExC_rx_sv);
13722 cp_list = add_cp_to_invlist(cp_list, '-');
13723 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
13728 range = 0; /* this was not a true range */
13729 element_count += 2; /* So counts for three values */
13732 classnum = namedclass_to_classnum(namedclass);
13734 if (LOC && namedclass < ANYOF_POSIXL_MAX
13735 #ifndef HAS_ISASCII
13736 && classnum != _CC_ASCII
13739 /* What the Posix classes (like \w, [:space:]) match in locale
13740 * isn't knowable under locale until actual match time. Room
13741 * must be reserved (one time per outer bracketed class) to
13742 * store such classes. The space will contain a bit for each
13743 * named class that is to be matched against. This isn't
13744 * needed for \p{} and pseudo-classes, as they are not affected
13745 * by locale, and hence are dealt with separately */
13746 if (! need_class) {
13749 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13752 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13754 ANYOF_FLAGS(ret) |= ANYOF_POSIXL;
13755 ANYOF_POSIXL_ZERO(ret);
13758 /* See if it already matches the complement of this POSIX
13760 if ((ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13761 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
13765 posixl_matches_all = TRUE;
13766 break; /* No need to continue. Since it matches both
13767 e.g., \w and \W, it matches everything, and the
13768 bracketed class can be optimized into qr/./s */
13771 /* Add this class to those that should be checked at runtime */
13772 ANYOF_POSIXL_SET(ret, namedclass);
13774 /* The above-Latin1 characters are not subject to locale rules.
13775 * Just add them, in the second pass, to the
13776 * unconditionally-matched list */
13778 SV* scratch_list = NULL;
13780 /* Get the list of the above-Latin1 code points this
13782 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
13783 PL_XPosix_ptrs[classnum],
13785 /* Odd numbers are complements, like
13786 * NDIGIT, NASCII, ... */
13787 namedclass % 2 != 0,
13789 /* Checking if 'cp_list' is NULL first saves an extra
13790 * clone. Its reference count will be decremented at the
13791 * next union, etc, or if this is the only instance, at the
13792 * end of the routine */
13794 cp_list = scratch_list;
13797 _invlist_union(cp_list, scratch_list, &cp_list);
13798 SvREFCNT_dec_NN(scratch_list);
13800 continue; /* Go get next character */
13803 else if (! SIZE_ONLY) {
13805 /* Here, not in pass1 (in that pass we skip calculating the
13806 * contents of this class), and is /l, or is a POSIX class for
13807 * which /l doesn't matter (or is a Unicode property, which is
13808 * skipped here). */
13809 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
13810 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
13812 /* Here, should be \h, \H, \v, or \V. None of /d, /i
13813 * nor /l make a difference in what these match,
13814 * therefore we just add what they match to cp_list. */
13815 if (classnum != _CC_VERTSPACE) {
13816 assert( namedclass == ANYOF_HORIZWS
13817 || namedclass == ANYOF_NHORIZWS);
13819 /* It turns out that \h is just a synonym for
13821 classnum = _CC_BLANK;
13824 _invlist_union_maybe_complement_2nd(
13826 PL_XPosix_ptrs[classnum],
13827 namedclass % 2 != 0, /* Complement if odd
13828 (NHORIZWS, NVERTWS)
13833 else { /* Garden variety class. If is NASCII, NDIGIT, ...
13834 complement and use nposixes */
13835 SV** posixes_ptr = namedclass % 2 == 0
13838 SV** source_ptr = &PL_XPosix_ptrs[classnum];
13839 _invlist_union_maybe_complement_2nd(
13842 namedclass % 2 != 0,
13845 continue; /* Go get next character */
13847 } /* end of namedclass \blah */
13849 /* Here, we have a single value. If 'range' is set, it is the ending
13850 * of a range--check its validity. Later, we will handle each
13851 * individual code point in the range. If 'range' isn't set, this
13852 * could be the beginning of a range, so check for that by looking
13853 * ahead to see if the next real character to be processed is the range
13854 * indicator--the minus sign */
13857 RExC_parse = regpatws(pRExC_state, RExC_parse,
13858 FALSE /* means don't recognize comments */);
13862 if (prevvalue > value) /* b-a */ {
13863 const int w = RExC_parse - rangebegin;
13865 "Invalid [] range \"%"UTF8f"\"",
13866 UTF8fARG(UTF, w, rangebegin));
13867 range = 0; /* not a valid range */
13871 prevvalue = value; /* save the beginning of the potential range */
13872 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13873 && *RExC_parse == '-')
13875 char* next_char_ptr = RExC_parse + 1;
13876 if (skip_white) { /* Get the next real char after the '-' */
13877 next_char_ptr = regpatws(pRExC_state,
13879 FALSE); /* means don't recognize
13883 /* If the '-' is at the end of the class (just before the ']',
13884 * it is a literal minus; otherwise it is a range */
13885 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13886 RExC_parse = next_char_ptr;
13888 /* a bad range like \w-, [:word:]- ? */
13889 if (namedclass > OOB_NAMEDCLASS) {
13890 if (strict || ckWARN(WARN_REGEXP)) {
13892 RExC_parse >= rangebegin ?
13893 RExC_parse - rangebegin : 0;
13895 vFAIL4("False [] range \"%*.*s\"",
13900 "False [] range \"%*.*s\"",
13905 cp_list = add_cp_to_invlist(cp_list, '-');
13909 range = 1; /* yeah, it's a range! */
13910 continue; /* but do it the next time */
13915 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13918 /* non-Latin1 code point implies unicode semantics. Must be set in
13919 * pass1 so is there for the whole of pass 2 */
13921 RExC_uni_semantics = 1;
13924 /* Ready to process either the single value, or the completed range.
13925 * For single-valued non-inverted ranges, we consider the possibility
13926 * of multi-char folds. (We made a conscious decision to not do this
13927 * for the other cases because it can often lead to non-intuitive
13928 * results. For example, you have the peculiar case that:
13929 * "s s" =~ /^[^\xDF]+$/i => Y
13930 * "ss" =~ /^[^\xDF]+$/i => N
13932 * See [perl #89750] */
13933 if (FOLD && allow_multi_folds && value == prevvalue) {
13934 if (value == LATIN_SMALL_LETTER_SHARP_S
13935 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13938 /* Here <value> is indeed a multi-char fold. Get what it is */
13940 U8 foldbuf[UTF8_MAXBYTES_CASE];
13943 UV folded = _to_uni_fold_flags(
13947 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
13948 ? FOLD_FLAGS_NOMIX_ASCII
13952 /* Here, <folded> should be the first character of the
13953 * multi-char fold of <value>, with <foldbuf> containing the
13954 * whole thing. But, if this fold is not allowed (because of
13955 * the flags), <fold> will be the same as <value>, and should
13956 * be processed like any other character, so skip the special
13958 if (folded != value) {
13960 /* Skip if we are recursed, currently parsing the class
13961 * again. Otherwise add this character to the list of
13962 * multi-char folds. */
13963 if (! RExC_in_multi_char_class) {
13964 AV** this_array_ptr;
13966 STRLEN cp_count = utf8_length(foldbuf,
13967 foldbuf + foldlen);
13968 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13970 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13973 if (! multi_char_matches) {
13974 multi_char_matches = newAV();
13977 /* <multi_char_matches> is actually an array of arrays.
13978 * There will be one or two top-level elements: [2],
13979 * and/or [3]. The [2] element is an array, each
13980 * element thereof is a character which folds to TWO
13981 * characters; [3] is for folds to THREE characters.
13982 * (Unicode guarantees a maximum of 3 characters in any
13983 * fold.) When we rewrite the character class below,
13984 * we will do so such that the longest folds are
13985 * written first, so that it prefers the longest
13986 * matching strings first. This is done even if it
13987 * turns out that any quantifier is non-greedy, out of
13988 * programmer laziness. Tom Christiansen has agreed
13989 * that this is ok. This makes the test for the
13990 * ligature 'ffi' come before the test for 'ff' */
13991 if (av_exists(multi_char_matches, cp_count)) {
13992 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13994 this_array = *this_array_ptr;
13997 this_array = newAV();
13998 av_store(multi_char_matches, cp_count,
14001 av_push(this_array, multi_fold);
14004 /* This element should not be processed further in this
14007 value = save_value;
14008 prevvalue = save_prevvalue;
14014 /* Deal with this element of the class */
14017 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14020 SV* this_range = _new_invlist(1);
14021 _append_range_to_invlist(this_range, prevvalue, value);
14023 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
14024 * If this range was specified using something like 'i-j', we want
14025 * to include only the 'i' and the 'j', and not anything in
14026 * between, so exclude non-ASCII, non-alphabetics from it.
14027 * However, if the range was specified with something like
14028 * [\x89-\x91] or [\x89-j], all code points within it should be
14029 * included. literal_endpoint==2 means both ends of the range used
14030 * a literal character, not \x{foo} */
14031 if (literal_endpoint == 2
14032 && ((prevvalue >= 'a' && value <= 'z')
14033 || (prevvalue >= 'A' && value <= 'Z')))
14035 _invlist_intersection(this_range, PL_ASCII,
14038 /* Since this above only contains ascii, the intersection of it
14039 * with anything will still yield only ascii */
14040 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ALPHA],
14043 _invlist_union(cp_foldable_list, this_range, &cp_foldable_list);
14044 literal_endpoint = 0;
14048 range = 0; /* this range (if it was one) is done now */
14049 } /* End of loop through all the text within the brackets */
14051 /* If anything in the class expands to more than one character, we have to
14052 * deal with them by building up a substitute parse string, and recursively
14053 * calling reg() on it, instead of proceeding */
14054 if (multi_char_matches) {
14055 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14058 char *save_end = RExC_end;
14059 char *save_parse = RExC_parse;
14060 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14065 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14066 because too confusing */
14068 sv_catpv(substitute_parse, "(?:");
14072 /* Look at the longest folds first */
14073 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14075 if (av_exists(multi_char_matches, cp_count)) {
14076 AV** this_array_ptr;
14079 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14081 while ((this_sequence = av_pop(*this_array_ptr)) !=
14084 if (! first_time) {
14085 sv_catpv(substitute_parse, "|");
14087 first_time = FALSE;
14089 sv_catpv(substitute_parse, SvPVX(this_sequence));
14094 /* If the character class contains anything else besides these
14095 * multi-character folds, have to include it in recursive parsing */
14096 if (element_count) {
14097 sv_catpv(substitute_parse, "|[");
14098 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14099 sv_catpv(substitute_parse, "]");
14102 sv_catpv(substitute_parse, ")");
14105 /* This is a way to get the parse to skip forward a whole named
14106 * sequence instead of matching the 2nd character when it fails the
14108 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14112 RExC_parse = SvPV(substitute_parse, len);
14113 RExC_end = RExC_parse + len;
14114 RExC_in_multi_char_class = 1;
14115 RExC_emit = (regnode *)orig_emit;
14117 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14119 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14121 RExC_parse = save_parse;
14122 RExC_end = save_end;
14123 RExC_in_multi_char_class = 0;
14124 SvREFCNT_dec_NN(multi_char_matches);
14128 /* Here, we've gone through the entire class and dealt with multi-char
14129 * folds. We are now in a position that we can do some checks to see if we
14130 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14131 * Currently we only do two checks:
14132 * 1) is in the unlikely event that the user has specified both, eg. \w and
14133 * \W under /l, then the class matches everything. (This optimization
14134 * is done only to make the optimizer code run later work.)
14135 * 2) if the character class contains only a single element (including a
14136 * single range), we see if there is an equivalent node for it.
14137 * Other checks are possible */
14138 if (! ret_invlist /* Can't optimize if returning the constructed
14140 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14145 if (UNLIKELY(posixl_matches_all)) {
14148 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14149 \w or [:digit:] or \p{foo}
14152 /* All named classes are mapped into POSIXish nodes, with its FLAG
14153 * argument giving which class it is */
14154 switch ((I32)namedclass) {
14155 case ANYOF_UNIPROP:
14158 /* These don't depend on the charset modifiers. They always
14159 * match under /u rules */
14160 case ANYOF_NHORIZWS:
14161 case ANYOF_HORIZWS:
14162 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14165 case ANYOF_NVERTWS:
14170 /* The actual POSIXish node for all the rest depends on the
14171 * charset modifier. The ones in the first set depend only on
14172 * ASCII or, if available on this platform, locale */
14176 op = (LOC) ? POSIXL : POSIXA;
14187 /* under /a could be alpha */
14189 if (ASCII_RESTRICTED) {
14190 namedclass = ANYOF_ALPHA + (namedclass % 2);
14198 /* The rest have more possibilities depending on the charset.
14199 * We take advantage of the enum ordering of the charset
14200 * modifiers to get the exact node type, */
14202 op = POSIXD + get_regex_charset(RExC_flags);
14203 if (op > POSIXA) { /* /aa is same as /a */
14208 /* The odd numbered ones are the complements of the
14209 * next-lower even number one */
14210 if (namedclass % 2 == 1) {
14214 arg = namedclass_to_classnum(namedclass);
14218 else if (value == prevvalue) {
14220 /* Here, the class consists of just a single code point */
14223 if (! LOC && value == '\n') {
14224 op = REG_ANY; /* Optimize [^\n] */
14225 *flagp |= HASWIDTH|SIMPLE;
14229 else if (value < 256 || UTF) {
14231 /* Optimize a single value into an EXACTish node, but not if it
14232 * would require converting the pattern to UTF-8. */
14233 op = compute_EXACTish(pRExC_state);
14235 } /* Otherwise is a range */
14236 else if (! LOC) { /* locale could vary these */
14237 if (prevvalue == '0') {
14238 if (value == '9') {
14245 /* Here, we have changed <op> away from its initial value iff we found
14246 * an optimization */
14249 /* Throw away this ANYOF regnode, and emit the calculated one,
14250 * which should correspond to the beginning, not current, state of
14252 const char * cur_parse = RExC_parse;
14253 RExC_parse = (char *)orig_parse;
14257 /* To get locale nodes to not use the full ANYOF size would
14258 * require moving the code above that writes the portions
14259 * of it that aren't in other nodes to after this point.
14260 * e.g. ANYOF_POSIXL_SET */
14261 RExC_size = orig_size;
14265 RExC_emit = (regnode *)orig_emit;
14266 if (PL_regkind[op] == POSIXD) {
14267 if (op == POSIXL) {
14268 RExC_contains_locale = 1;
14271 op += NPOSIXD - POSIXD;
14276 ret = reg_node(pRExC_state, op);
14278 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14282 *flagp |= HASWIDTH|SIMPLE;
14284 else if (PL_regkind[op] == EXACT) {
14285 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14286 TRUE /* downgradable to EXACT */
14290 RExC_parse = (char *) cur_parse;
14292 SvREFCNT_dec(posixes);
14293 SvREFCNT_dec(nposixes);
14294 SvREFCNT_dec(cp_list);
14295 SvREFCNT_dec(cp_foldable_list);
14302 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14304 /* If folding, we calculate all characters that could fold to or from the
14305 * ones already on the list */
14306 if (cp_foldable_list) {
14308 UV start, end; /* End points of code point ranges */
14310 SV* fold_intersection = NULL;
14313 /* Our calculated list will be for Unicode rules. For locale
14314 * matching, we have to keep a separate list that is consulted at
14315 * runtime only when the locale indicates Unicode rules. For
14316 * non-locale, we just use to the general list */
14318 use_list = &only_utf8_locale_list;
14321 use_list = &cp_list;
14324 /* Only the characters in this class that participate in folds need
14325 * be checked. Get the intersection of this class and all the
14326 * possible characters that are foldable. This can quickly narrow
14327 * down a large class */
14328 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14329 &fold_intersection);
14331 /* The folds for all the Latin1 characters are hard-coded into this
14332 * program, but we have to go out to disk to get the others. */
14333 if (invlist_highest(cp_foldable_list) >= 256) {
14335 /* This is a hash that for a particular fold gives all
14336 * characters that are involved in it */
14337 if (! PL_utf8_foldclosures) {
14339 /* If the folds haven't been read in, call a fold function
14341 if (! PL_utf8_tofold) {
14342 U8 dummy[UTF8_MAXBYTES_CASE+1];
14344 /* This string is just a short named one above \xff */
14345 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
14346 assert(PL_utf8_tofold); /* Verify that worked */
14348 PL_utf8_foldclosures
14349 = _swash_inversion_hash(PL_utf8_tofold);
14353 /* Now look at the foldable characters in this class individually */
14354 invlist_iterinit(fold_intersection);
14355 while (invlist_iternext(fold_intersection, &start, &end)) {
14358 /* Look at every character in the range */
14359 for (j = start; j <= end; j++) {
14360 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14366 /* We have the latin1 folding rules hard-coded here so
14367 * that an innocent-looking character class, like
14368 * /[ks]/i won't have to go out to disk to find the
14369 * possible matches. XXX It would be better to
14370 * generate these via regen, in case a new version of
14371 * the Unicode standard adds new mappings, though that
14372 * is not really likely, and may be caught by the
14373 * default: case of the switch below. */
14375 if (IS_IN_SOME_FOLD_L1(j)) {
14377 /* ASCII is always matched; non-ASCII is matched
14378 * only under Unicode rules (which could happen
14379 * under /l if the locale is a UTF-8 one */
14380 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
14381 *use_list = add_cp_to_invlist(*use_list,
14382 PL_fold_latin1[j]);
14386 add_cp_to_invlist(depends_list,
14387 PL_fold_latin1[j]);
14391 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
14392 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
14394 /* Certain Latin1 characters have matches outside
14395 * Latin1. To get here, <j> is one of those
14396 * characters. None of these matches is valid for
14397 * ASCII characters under /aa, which is why the 'if'
14398 * just above excludes those. These matches only
14399 * happen when the target string is utf8. The code
14400 * below adds the single fold closures for <j> to the
14401 * inversion list. */
14407 add_cp_to_invlist(*use_list, KELVIN_SIGN);
14411 *use_list = add_cp_to_invlist(*use_list,
14412 LATIN_SMALL_LETTER_LONG_S);
14415 *use_list = add_cp_to_invlist(*use_list,
14416 GREEK_CAPITAL_LETTER_MU);
14417 *use_list = add_cp_to_invlist(*use_list,
14418 GREEK_SMALL_LETTER_MU);
14420 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14421 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14423 add_cp_to_invlist(*use_list, ANGSTROM_SIGN);
14425 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14426 *use_list = add_cp_to_invlist(*use_list,
14427 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14429 case LATIN_SMALL_LETTER_SHARP_S:
14430 *use_list = add_cp_to_invlist(*use_list,
14431 LATIN_CAPITAL_LETTER_SHARP_S);
14433 case 'F': case 'f':
14434 case 'I': case 'i':
14435 case 'L': case 'l':
14436 case 'T': case 't':
14437 case 'A': case 'a':
14438 case 'H': case 'h':
14439 case 'J': case 'j':
14440 case 'N': case 'n':
14441 case 'W': case 'w':
14442 case 'Y': case 'y':
14443 /* These all are targets of multi-character
14444 * folds from code points that require UTF8
14445 * to express, so they can't match unless
14446 * the target string is in UTF-8, so no
14447 * action here is necessary, as regexec.c
14448 * properly handles the general case for
14449 * UTF-8 matching and multi-char folds */
14452 /* Use deprecated warning to increase the
14453 * chances of this being output */
14454 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
14461 /* Here is an above Latin1 character. We don't have the
14462 * rules hard-coded for it. First, get its fold. This is
14463 * the simple fold, as the multi-character folds have been
14464 * handled earlier and separated out */
14465 _to_uni_fold_flags(j, foldbuf, &foldlen,
14466 (ASCII_FOLD_RESTRICTED)
14467 ? FOLD_FLAGS_NOMIX_ASCII
14470 /* Single character fold of above Latin1. Add everything in
14471 * its fold closure to the list that this node should match.
14472 * The fold closures data structure is a hash with the keys
14473 * being the UTF-8 of every character that is folded to, like
14474 * 'k', and the values each an array of all code points that
14475 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
14476 * Multi-character folds are not included */
14477 if ((listp = hv_fetch(PL_utf8_foldclosures,
14478 (char *) foldbuf, foldlen, FALSE)))
14480 AV* list = (AV*) *listp;
14482 for (k = 0; k <= av_tindex(list); k++) {
14483 SV** c_p = av_fetch(list, k, FALSE);
14486 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
14490 /* /aa doesn't allow folds between ASCII and non- */
14491 if ((ASCII_FOLD_RESTRICTED
14492 && (isASCII(c) != isASCII(j))))
14497 /* Folds under /l which cross the 255/256 boundary
14498 * are added to a separate list. (These are valid
14499 * only when the locale is UTF-8.) */
14500 if (c < 256 && LOC) {
14501 *use_list = add_cp_to_invlist(*use_list, c);
14505 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
14507 cp_list = add_cp_to_invlist(cp_list, c);
14510 /* Similarly folds involving non-ascii Latin1
14511 * characters under /d are added to their list */
14512 depends_list = add_cp_to_invlist(depends_list,
14519 SvREFCNT_dec_NN(fold_intersection);
14522 /* Now that we have finished adding all the folds, there is no reason
14523 * to keep the foldable list separate */
14524 _invlist_union(cp_list, cp_foldable_list, &cp_list);
14525 SvREFCNT_dec_NN(cp_foldable_list);
14528 /* And combine the result (if any) with any inversion list from posix
14529 * classes. The lists are kept separate up to now because we don't want to
14530 * fold the classes (folding of those is automatically handled by the swash
14531 * fetching code) */
14532 if (posixes || nposixes) {
14533 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
14534 /* Under /a and /aa, nothing above ASCII matches these */
14535 _invlist_intersection(posixes,
14536 PL_XPosix_ptrs[_CC_ASCII],
14540 if (DEPENDS_SEMANTICS) {
14541 /* Under /d, everything in the upper half of the Latin1 range
14542 * matches these complements */
14543 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_NON_ASCII_ALL;
14545 else if (AT_LEAST_ASCII_RESTRICTED) {
14546 /* Under /a and /aa, everything above ASCII matches these
14548 _invlist_union_complement_2nd(nposixes,
14549 PL_XPosix_ptrs[_CC_ASCII],
14553 _invlist_union(posixes, nposixes, &posixes);
14554 SvREFCNT_dec_NN(nposixes);
14557 posixes = nposixes;
14560 if (! DEPENDS_SEMANTICS) {
14562 _invlist_union(cp_list, posixes, &cp_list);
14563 SvREFCNT_dec_NN(posixes);
14570 /* Under /d, we put into a separate list the Latin1 things that
14571 * match only when the target string is utf8 */
14572 SV* nonascii_but_latin1_properties = NULL;
14573 _invlist_intersection(posixes, PL_UpperLatin1,
14574 &nonascii_but_latin1_properties);
14575 _invlist_subtract(posixes, nonascii_but_latin1_properties,
14578 _invlist_union(cp_list, posixes, &cp_list);
14579 SvREFCNT_dec_NN(posixes);
14585 if (depends_list) {
14586 _invlist_union(depends_list, nonascii_but_latin1_properties,
14588 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
14591 depends_list = nonascii_but_latin1_properties;
14596 /* And combine the result (if any) with any inversion list from properties.
14597 * The lists are kept separate up to now so that we can distinguish the two
14598 * in regards to matching above-Unicode. A run-time warning is generated
14599 * if a Unicode property is matched against a non-Unicode code point. But,
14600 * we allow user-defined properties to match anything, without any warning,
14601 * and we also suppress the warning if there is a portion of the character
14602 * class that isn't a Unicode property, and which matches above Unicode, \W
14603 * or [\x{110000}] for example.
14604 * (Note that in this case, unlike the Posix one above, there is no
14605 * <depends_list>, because having a Unicode property forces Unicode
14610 /* If it matters to the final outcome, see if a non-property
14611 * component of the class matches above Unicode. If so, the
14612 * warning gets suppressed. This is true even if just a single
14613 * such code point is specified, as though not strictly correct if
14614 * another such code point is matched against, the fact that they
14615 * are using above-Unicode code points indicates they should know
14616 * the issues involved */
14618 warn_super = ! (invert
14619 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
14622 _invlist_union(properties, cp_list, &cp_list);
14623 SvREFCNT_dec_NN(properties);
14626 cp_list = properties;
14630 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
14634 /* Here, we have calculated what code points should be in the character
14637 * Now we can see about various optimizations. Fold calculation (which we
14638 * did above) needs to take place before inversion. Otherwise /[^k]/i
14639 * would invert to include K, which under /i would match k, which it
14640 * shouldn't. Therefore we can't invert folded locale now, as it won't be
14641 * folded until runtime */
14643 /* If we didn't do folding, it's because some information isn't available
14644 * until runtime; set the run-time fold flag for these. (We don't have to
14645 * worry about properties folding, as that is taken care of by the swash
14646 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
14647 * locales, or the class matches at least one 0-255 range code point */
14649 if (only_utf8_locale_list) {
14650 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14652 else if (cp_list) { /* Look to see if there a 0-255 code point is in
14655 invlist_iterinit(cp_list);
14656 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
14657 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14659 invlist_iterfinish(cp_list);
14663 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
14664 * at compile time. Besides not inverting folded locale now, we can't
14665 * invert if there are things such as \w, which aren't known until runtime
14668 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14670 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14672 _invlist_invert(cp_list);
14674 /* Any swash can't be used as-is, because we've inverted things */
14676 SvREFCNT_dec_NN(swash);
14680 /* Clear the invert flag since have just done it here */
14685 *ret_invlist = cp_list;
14686 SvREFCNT_dec(swash);
14688 /* Discard the generated node */
14690 RExC_size = orig_size;
14693 RExC_emit = orig_emit;
14698 /* Some character classes are equivalent to other nodes. Such nodes take
14699 * up less room and generally fewer operations to execute than ANYOF nodes.
14700 * Above, we checked for and optimized into some such equivalents for
14701 * certain common classes that are easy to test. Getting to this point in
14702 * the code means that the class didn't get optimized there. Since this
14703 * code is only executed in Pass 2, it is too late to save space--it has
14704 * been allocated in Pass 1, and currently isn't given back. But turning
14705 * things into an EXACTish node can allow the optimizer to join it to any
14706 * adjacent such nodes. And if the class is equivalent to things like /./,
14707 * expensive run-time swashes can be avoided. Now that we have more
14708 * complete information, we can find things necessarily missed by the
14709 * earlier code. I (khw) am not sure how much to look for here. It would
14710 * be easy, but perhaps too slow, to check any candidates against all the
14711 * node types they could possibly match using _invlistEQ(). */
14716 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14717 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14719 /* We don't optimize if we are supposed to make sure all non-Unicode
14720 * code points raise a warning, as only ANYOF nodes have this check.
14722 && ! ((ANYOF_FLAGS(ret) | ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
14725 U8 op = END; /* The optimzation node-type */
14726 const char * cur_parse= RExC_parse;
14728 invlist_iterinit(cp_list);
14729 if (! invlist_iternext(cp_list, &start, &end)) {
14731 /* Here, the list is empty. This happens, for example, when a
14732 * Unicode property is the only thing in the character class, and
14733 * it doesn't match anything. (perluniprops.pod notes such
14736 *flagp |= HASWIDTH|SIMPLE;
14738 else if (start == end) { /* The range is a single code point */
14739 if (! invlist_iternext(cp_list, &start, &end)
14741 /* Don't do this optimization if it would require changing
14742 * the pattern to UTF-8 */
14743 && (start < 256 || UTF))
14745 /* Here, the list contains a single code point. Can optimize
14746 * into an EXACTish node */
14755 /* A locale node under folding with one code point can be
14756 * an EXACTFL, as its fold won't be calculated until
14762 /* Here, we are generally folding, but there is only one
14763 * code point to match. If we have to, we use an EXACT
14764 * node, but it would be better for joining with adjacent
14765 * nodes in the optimization pass if we used the same
14766 * EXACTFish node that any such are likely to be. We can
14767 * do this iff the code point doesn't participate in any
14768 * folds. For example, an EXACTF of a colon is the same as
14769 * an EXACT one, since nothing folds to or from a colon. */
14771 if (IS_IN_SOME_FOLD_L1(value)) {
14776 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
14781 /* If we haven't found the node type, above, it means we
14782 * can use the prevailing one */
14784 op = compute_EXACTish(pRExC_state);
14789 else if (start == 0) {
14790 if (end == UV_MAX) {
14792 *flagp |= HASWIDTH|SIMPLE;
14795 else if (end == '\n' - 1
14796 && invlist_iternext(cp_list, &start, &end)
14797 && start == '\n' + 1 && end == UV_MAX)
14800 *flagp |= HASWIDTH|SIMPLE;
14804 invlist_iterfinish(cp_list);
14807 RExC_parse = (char *)orig_parse;
14808 RExC_emit = (regnode *)orig_emit;
14810 ret = reg_node(pRExC_state, op);
14812 RExC_parse = (char *)cur_parse;
14814 if (PL_regkind[op] == EXACT) {
14815 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14816 TRUE /* downgradable to EXACT */
14820 SvREFCNT_dec_NN(cp_list);
14825 /* Here, <cp_list> contains all the code points we can determine at
14826 * compile time that match under all conditions. Go through it, and
14827 * for things that belong in the bitmap, put them there, and delete from
14828 * <cp_list>. While we are at it, see if everything above 255 is in the
14829 * list, and if so, set a flag to speed up execution */
14831 populate_ANYOF_from_invlist(ret, &cp_list);
14834 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
14837 /* Here, the bitmap has been populated with all the Latin1 code points that
14838 * always match. Can now add to the overall list those that match only
14839 * when the target string is UTF-8 (<depends_list>). */
14840 if (depends_list) {
14842 _invlist_union(cp_list, depends_list, &cp_list);
14843 SvREFCNT_dec_NN(depends_list);
14846 cp_list = depends_list;
14848 ANYOF_FLAGS(ret) |= ANYOF_UTF8;
14851 /* If there is a swash and more than one element, we can't use the swash in
14852 * the optimization below. */
14853 if (swash && element_count > 1) {
14854 SvREFCNT_dec_NN(swash);
14858 set_ANYOF_arg(pRExC_state, ret, cp_list,
14859 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14861 only_utf8_locale_list,
14862 swash, has_user_defined_property);
14864 *flagp |= HASWIDTH|SIMPLE;
14866 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
14867 RExC_contains_locale = 1;
14873 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14876 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
14877 regnode* const node,
14879 SV* const runtime_defns,
14880 SV* const only_utf8_locale_list,
14882 const bool has_user_defined_property)
14884 /* Sets the arg field of an ANYOF-type node 'node', using information about
14885 * the node passed-in. If there is nothing outside the node's bitmap, the
14886 * arg is set to ANYOF_NONBITMAP_EMPTY. Otherwise, it sets the argument to
14887 * the count returned by add_data(), having allocated and stored an array,
14888 * av, that that count references, as follows:
14889 * av[0] stores the character class description in its textual form.
14890 * This is used later (regexec.c:Perl_regclass_swash()) to
14891 * initialize the appropriate swash, and is also useful for dumping
14892 * the regnode. This is set to &PL_sv_undef if the textual
14893 * description is not needed at run-time (as happens if the other
14894 * elements completely define the class)
14895 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
14896 * computed from av[0]. But if no further computation need be done,
14897 * the swash is stored here now (and av[0] is &PL_sv_undef).
14898 * av[2] stores the inversion list of code points that match only if the
14899 * current locale is UTF-8
14900 * av[3] stores the cp_list inversion list for use in addition or instead
14901 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
14902 * (Otherwise everything needed is already in av[0] and av[1])
14903 * av[4] is set if any component of the class is from a user-defined
14904 * property; used only if av[3] exists */
14908 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
14910 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
14911 assert(! (ANYOF_FLAGS(node)
14912 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8)));
14913 ARG_SET(node, ANYOF_NONBITMAP_EMPTY);
14916 AV * const av = newAV();
14919 assert(ANYOF_FLAGS(node)
14920 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8|ANYOF_LOC_FOLD));
14922 av_store(av, 0, (runtime_defns)
14923 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
14925 av_store(av, 1, swash);
14926 SvREFCNT_dec_NN(cp_list);
14929 av_store(av, 1, &PL_sv_undef);
14931 av_store(av, 3, cp_list);
14932 av_store(av, 4, newSVuv(has_user_defined_property));
14936 if (only_utf8_locale_list) {
14937 av_store(av, 2, only_utf8_locale_list);
14940 av_store(av, 2, &PL_sv_undef);
14943 rv = newRV_noinc(MUTABLE_SV(av));
14944 n = add_data(pRExC_state, STR_WITH_LEN("s"));
14945 RExC_rxi->data->data[n] = (void*)rv;
14951 /* reg_skipcomment()
14953 Absorbs an /x style # comments from the input stream.
14954 Returns true if there is more text remaining in the stream.
14955 Will set the REG_RUN_ON_COMMENT_SEEN flag if the comment
14956 terminates the pattern without including a newline.
14958 Note its the callers responsibility to ensure that we are
14959 actually in /x mode
14964 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14968 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14970 while (RExC_parse < RExC_end)
14971 if (*RExC_parse++ == '\n') {
14976 /* we ran off the end of the pattern without ending
14977 the comment, so we have to add an \n when wrapping */
14978 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
14986 Advances the parse position, and optionally absorbs
14987 "whitespace" from the inputstream.
14989 Without /x "whitespace" means (?#...) style comments only,
14990 with /x this means (?#...) and # comments and whitespace proper.
14992 Returns the RExC_parse point from BEFORE the scan occurs.
14994 This is the /x friendly way of saying RExC_parse++.
14998 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
15000 char* const retval = RExC_parse++;
15002 PERL_ARGS_ASSERT_NEXTCHAR;
15005 if (RExC_end - RExC_parse >= 3
15006 && *RExC_parse == '('
15007 && RExC_parse[1] == '?'
15008 && RExC_parse[2] == '#')
15010 while (*RExC_parse != ')') {
15011 if (RExC_parse == RExC_end)
15012 FAIL("Sequence (?#... not terminated");
15018 if (RExC_flags & RXf_PMf_EXTENDED) {
15019 if (isSPACE(*RExC_parse)) {
15023 else if (*RExC_parse == '#') {
15024 if ( reg_skipcomment( pRExC_state ) )
15033 - reg_node - emit a node
15035 STATIC regnode * /* Location. */
15036 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
15040 regnode * const ret = RExC_emit;
15041 GET_RE_DEBUG_FLAGS_DECL;
15043 PERL_ARGS_ASSERT_REG_NODE;
15046 SIZE_ALIGN(RExC_size);
15050 if (RExC_emit >= RExC_emit_bound)
15051 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15052 op, RExC_emit, RExC_emit_bound);
15054 NODE_ALIGN_FILL(ret);
15056 FILL_ADVANCE_NODE(ptr, op);
15057 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
15058 #ifdef RE_TRACK_PATTERN_OFFSETS
15059 if (RExC_offsets) { /* MJD */
15061 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
15062 "reg_node", __LINE__,
15064 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
15065 ? "Overwriting end of array!\n" : "OK",
15066 (UV)(RExC_emit - RExC_emit_start),
15067 (UV)(RExC_parse - RExC_start),
15068 (UV)RExC_offsets[0]));
15069 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
15077 - reganode - emit a node with an argument
15079 STATIC regnode * /* Location. */
15080 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
15084 regnode * const ret = RExC_emit;
15085 GET_RE_DEBUG_FLAGS_DECL;
15087 PERL_ARGS_ASSERT_REGANODE;
15090 SIZE_ALIGN(RExC_size);
15095 assert(2==regarglen[op]+1);
15097 Anything larger than this has to allocate the extra amount.
15098 If we changed this to be:
15100 RExC_size += (1 + regarglen[op]);
15102 then it wouldn't matter. Its not clear what side effect
15103 might come from that so its not done so far.
15108 if (RExC_emit >= RExC_emit_bound)
15109 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15110 op, RExC_emit, RExC_emit_bound);
15112 NODE_ALIGN_FILL(ret);
15114 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15115 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
15116 #ifdef RE_TRACK_PATTERN_OFFSETS
15117 if (RExC_offsets) { /* MJD */
15119 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15123 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
15124 "Overwriting end of array!\n" : "OK",
15125 (UV)(RExC_emit - RExC_emit_start),
15126 (UV)(RExC_parse - RExC_start),
15127 (UV)RExC_offsets[0]));
15128 Set_Cur_Node_Offset;
15136 - reguni - emit (if appropriate) a Unicode character
15138 PERL_STATIC_INLINE STRLEN
15139 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
15143 PERL_ARGS_ASSERT_REGUNI;
15145 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
15149 - reginsert - insert an operator in front of already-emitted operand
15151 * Means relocating the operand.
15154 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15160 const int offset = regarglen[(U8)op];
15161 const int size = NODE_STEP_REGNODE + offset;
15162 GET_RE_DEBUG_FLAGS_DECL;
15164 PERL_ARGS_ASSERT_REGINSERT;
15165 PERL_UNUSED_ARG(depth);
15166 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15167 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15176 if (RExC_open_parens) {
15178 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15179 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15180 if ( RExC_open_parens[paren] >= opnd ) {
15181 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15182 RExC_open_parens[paren] += size;
15184 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15186 if ( RExC_close_parens[paren] >= opnd ) {
15187 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15188 RExC_close_parens[paren] += size;
15190 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15195 while (src > opnd) {
15196 StructCopy(--src, --dst, regnode);
15197 #ifdef RE_TRACK_PATTERN_OFFSETS
15198 if (RExC_offsets) { /* MJD 20010112 */
15200 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15204 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15205 ? "Overwriting end of array!\n" : "OK",
15206 (UV)(src - RExC_emit_start),
15207 (UV)(dst - RExC_emit_start),
15208 (UV)RExC_offsets[0]));
15209 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15210 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15216 place = opnd; /* Op node, where operand used to be. */
15217 #ifdef RE_TRACK_PATTERN_OFFSETS
15218 if (RExC_offsets) { /* MJD */
15220 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15224 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15225 ? "Overwriting end of array!\n" : "OK",
15226 (UV)(place - RExC_emit_start),
15227 (UV)(RExC_parse - RExC_start),
15228 (UV)RExC_offsets[0]));
15229 Set_Node_Offset(place, RExC_parse);
15230 Set_Node_Length(place, 1);
15233 src = NEXTOPER(place);
15234 FILL_ADVANCE_NODE(place, op);
15235 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
15236 Zero(src, offset, regnode);
15240 - regtail - set the next-pointer at the end of a node chain of p to val.
15241 - SEE ALSO: regtail_study
15243 /* TODO: All three parms should be const */
15245 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15246 const regnode *val,U32 depth)
15250 GET_RE_DEBUG_FLAGS_DECL;
15252 PERL_ARGS_ASSERT_REGTAIL;
15254 PERL_UNUSED_ARG(depth);
15260 /* Find last node. */
15263 regnode * const temp = regnext(scan);
15265 SV * const mysv=sv_newmortal();
15266 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15267 regprop(RExC_rx, mysv, scan, NULL);
15268 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15269 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
15270 (temp == NULL ? "->" : ""),
15271 (temp == NULL ? PL_reg_name[OP(val)] : "")
15279 if (reg_off_by_arg[OP(scan)]) {
15280 ARG_SET(scan, val - scan);
15283 NEXT_OFF(scan) = val - scan;
15289 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15290 - Look for optimizable sequences at the same time.
15291 - currently only looks for EXACT chains.
15293 This is experimental code. The idea is to use this routine to perform
15294 in place optimizations on branches and groups as they are constructed,
15295 with the long term intention of removing optimization from study_chunk so
15296 that it is purely analytical.
15298 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15299 to control which is which.
15302 /* TODO: All four parms should be const */
15305 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15306 const regnode *val,U32 depth)
15311 #ifdef EXPERIMENTAL_INPLACESCAN
15314 GET_RE_DEBUG_FLAGS_DECL;
15316 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15322 /* Find last node. */
15326 regnode * const temp = regnext(scan);
15327 #ifdef EXPERIMENTAL_INPLACESCAN
15328 if (PL_regkind[OP(scan)] == EXACT) {
15329 bool unfolded_multi_char; /* Unexamined in this routine */
15330 if (join_exact(pRExC_state, scan, &min,
15331 &unfolded_multi_char, 1, val, depth+1))
15336 switch (OP(scan)) {
15339 case EXACTFA_NO_TRIE:
15344 if( exact == PSEUDO )
15346 else if ( exact != OP(scan) )
15355 SV * const mysv=sv_newmortal();
15356 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
15357 regprop(RExC_rx, mysv, scan, NULL);
15358 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
15359 SvPV_nolen_const(mysv),
15360 REG_NODE_NUM(scan),
15361 PL_reg_name[exact]);
15368 SV * const mysv_val=sv_newmortal();
15369 DEBUG_PARSE_MSG("");
15370 regprop(RExC_rx, mysv_val, val, NULL);
15371 PerlIO_printf(Perl_debug_log,
15372 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
15373 SvPV_nolen_const(mysv_val),
15374 (IV)REG_NODE_NUM(val),
15378 if (reg_off_by_arg[OP(scan)]) {
15379 ARG_SET(scan, val - scan);
15382 NEXT_OFF(scan) = val - scan;
15390 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
15395 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
15400 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15402 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
15403 if (flags & (1<<bit)) {
15404 if (!set++ && lead)
15405 PerlIO_printf(Perl_debug_log, "%s",lead);
15406 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
15411 PerlIO_printf(Perl_debug_log, "\n");
15413 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15418 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
15424 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15426 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
15427 if (flags & (1<<bit)) {
15428 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
15431 if (!set++ && lead)
15432 PerlIO_printf(Perl_debug_log, "%s",lead);
15433 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
15436 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
15437 if (!set++ && lead) {
15438 PerlIO_printf(Perl_debug_log, "%s",lead);
15441 case REGEX_UNICODE_CHARSET:
15442 PerlIO_printf(Perl_debug_log, "UNICODE");
15444 case REGEX_LOCALE_CHARSET:
15445 PerlIO_printf(Perl_debug_log, "LOCALE");
15447 case REGEX_ASCII_RESTRICTED_CHARSET:
15448 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
15450 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
15451 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
15454 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
15460 PerlIO_printf(Perl_debug_log, "\n");
15462 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15468 Perl_regdump(pTHX_ const regexp *r)
15472 SV * const sv = sv_newmortal();
15473 SV *dsv= sv_newmortal();
15474 RXi_GET_DECL(r,ri);
15475 GET_RE_DEBUG_FLAGS_DECL;
15477 PERL_ARGS_ASSERT_REGDUMP;
15479 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
15481 /* Header fields of interest. */
15482 if (r->anchored_substr) {
15483 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
15484 RE_SV_DUMPLEN(r->anchored_substr), 30);
15485 PerlIO_printf(Perl_debug_log,
15486 "anchored %s%s at %"IVdf" ",
15487 s, RE_SV_TAIL(r->anchored_substr),
15488 (IV)r->anchored_offset);
15489 } else if (r->anchored_utf8) {
15490 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
15491 RE_SV_DUMPLEN(r->anchored_utf8), 30);
15492 PerlIO_printf(Perl_debug_log,
15493 "anchored utf8 %s%s at %"IVdf" ",
15494 s, RE_SV_TAIL(r->anchored_utf8),
15495 (IV)r->anchored_offset);
15497 if (r->float_substr) {
15498 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
15499 RE_SV_DUMPLEN(r->float_substr), 30);
15500 PerlIO_printf(Perl_debug_log,
15501 "floating %s%s at %"IVdf"..%"UVuf" ",
15502 s, RE_SV_TAIL(r->float_substr),
15503 (IV)r->float_min_offset, (UV)r->float_max_offset);
15504 } else if (r->float_utf8) {
15505 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
15506 RE_SV_DUMPLEN(r->float_utf8), 30);
15507 PerlIO_printf(Perl_debug_log,
15508 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
15509 s, RE_SV_TAIL(r->float_utf8),
15510 (IV)r->float_min_offset, (UV)r->float_max_offset);
15512 if (r->check_substr || r->check_utf8)
15513 PerlIO_printf(Perl_debug_log,
15515 (r->check_substr == r->float_substr
15516 && r->check_utf8 == r->float_utf8
15517 ? "(checking floating" : "(checking anchored"));
15518 if (r->intflags & PREGf_NOSCAN)
15519 PerlIO_printf(Perl_debug_log, " noscan");
15520 if (r->extflags & RXf_CHECK_ALL)
15521 PerlIO_printf(Perl_debug_log, " isall");
15522 if (r->check_substr || r->check_utf8)
15523 PerlIO_printf(Perl_debug_log, ") ");
15525 if (ri->regstclass) {
15526 regprop(r, sv, ri->regstclass, NULL);
15527 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
15529 if (r->intflags & PREGf_ANCH) {
15530 PerlIO_printf(Perl_debug_log, "anchored");
15531 if (r->intflags & PREGf_ANCH_BOL)
15532 PerlIO_printf(Perl_debug_log, "(BOL)");
15533 if (r->intflags & PREGf_ANCH_MBOL)
15534 PerlIO_printf(Perl_debug_log, "(MBOL)");
15535 if (r->intflags & PREGf_ANCH_SBOL)
15536 PerlIO_printf(Perl_debug_log, "(SBOL)");
15537 if (r->intflags & PREGf_ANCH_GPOS)
15538 PerlIO_printf(Perl_debug_log, "(GPOS)");
15539 PerlIO_putc(Perl_debug_log, ' ');
15541 if (r->intflags & PREGf_GPOS_SEEN)
15542 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
15543 if (r->intflags & PREGf_SKIP)
15544 PerlIO_printf(Perl_debug_log, "plus ");
15545 if (r->intflags & PREGf_IMPLICIT)
15546 PerlIO_printf(Perl_debug_log, "implicit ");
15547 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
15548 if (r->extflags & RXf_EVAL_SEEN)
15549 PerlIO_printf(Perl_debug_log, "with eval ");
15550 PerlIO_printf(Perl_debug_log, "\n");
15552 regdump_extflags("r->extflags: ",r->extflags);
15553 regdump_intflags("r->intflags: ",r->intflags);
15556 PERL_ARGS_ASSERT_REGDUMP;
15557 PERL_UNUSED_CONTEXT;
15558 PERL_UNUSED_ARG(r);
15559 #endif /* DEBUGGING */
15563 - regprop - printable representation of opcode, with run time support
15567 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo)
15573 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
15574 static const char * const anyofs[] = {
15575 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
15576 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
15577 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
15578 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
15579 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
15580 || _CC_VERTSPACE != 16
15581 #error Need to adjust order of anyofs[]
15618 RXi_GET_DECL(prog,progi);
15619 GET_RE_DEBUG_FLAGS_DECL;
15621 PERL_ARGS_ASSERT_REGPROP;
15625 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
15626 /* It would be nice to FAIL() here, but this may be called from
15627 regexec.c, and it would be hard to supply pRExC_state. */
15628 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
15629 (int)OP(o), (int)REGNODE_MAX);
15630 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
15632 k = PL_regkind[OP(o)];
15635 sv_catpvs(sv, " ");
15636 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
15637 * is a crude hack but it may be the best for now since
15638 * we have no flag "this EXACTish node was UTF-8"
15640 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
15641 PERL_PV_ESCAPE_UNI_DETECT |
15642 PERL_PV_ESCAPE_NONASCII |
15643 PERL_PV_PRETTY_ELLIPSES |
15644 PERL_PV_PRETTY_LTGT |
15645 PERL_PV_PRETTY_NOCLEAR
15647 } else if (k == TRIE) {
15648 /* print the details of the trie in dumpuntil instead, as
15649 * progi->data isn't available here */
15650 const char op = OP(o);
15651 const U32 n = ARG(o);
15652 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
15653 (reg_ac_data *)progi->data->data[n] :
15655 const reg_trie_data * const trie
15656 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
15658 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
15659 DEBUG_TRIE_COMPILE_r(
15660 Perl_sv_catpvf(aTHX_ sv,
15661 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
15662 (UV)trie->startstate,
15663 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
15664 (UV)trie->wordcount,
15667 (UV)TRIE_CHARCOUNT(trie),
15668 (UV)trie->uniquecharcount
15671 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
15672 sv_catpvs(sv, "[");
15673 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
15675 : TRIE_BITMAP(trie));
15676 sv_catpvs(sv, "]");
15679 } else if (k == CURLY) {
15680 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
15681 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
15682 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
15684 else if (k == WHILEM && o->flags) /* Ordinal/of */
15685 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
15686 else if (k == REF || k == OPEN || k == CLOSE
15687 || k == GROUPP || OP(o)==ACCEPT)
15689 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
15690 if ( RXp_PAREN_NAMES(prog) ) {
15691 if ( k != REF || (OP(o) < NREF)) {
15692 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
15693 SV **name= av_fetch(list, ARG(o), 0 );
15695 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15698 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
15699 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
15700 I32 *nums=(I32*)SvPVX(sv_dat);
15701 SV **name= av_fetch(list, nums[0], 0 );
15704 for ( n=0; n<SvIVX(sv_dat); n++ ) {
15705 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
15706 (n ? "," : ""), (IV)nums[n]);
15708 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15712 if ( k == REF && reginfo) {
15713 U32 n = ARG(o); /* which paren pair */
15714 I32 ln = prog->offs[n].start;
15715 if (prog->lastparen < n || ln == -1)
15716 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
15717 else if (ln == prog->offs[n].end)
15718 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
15720 const char *s = reginfo->strbeg + ln;
15721 Perl_sv_catpvf(aTHX_ sv, ": ");
15722 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
15723 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
15726 } else if (k == GOSUB)
15727 /* Paren and offset */
15728 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
15729 else if (k == VERB) {
15731 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
15732 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
15733 } else if (k == LOGICAL)
15734 /* 2: embedded, otherwise 1 */
15735 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
15736 else if (k == ANYOF) {
15737 const U8 flags = ANYOF_FLAGS(o);
15741 if (flags & ANYOF_LOCALE_FLAGS)
15742 sv_catpvs(sv, "{loc}");
15743 if (flags & ANYOF_LOC_FOLD)
15744 sv_catpvs(sv, "{i}");
15745 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
15746 if (flags & ANYOF_INVERT)
15747 sv_catpvs(sv, "^");
15749 /* output what the standard cp 0-255 bitmap matches */
15750 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
15752 /* output any special charclass tests (used entirely under use
15754 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
15756 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
15757 if (ANYOF_POSIXL_TEST(o,i)) {
15758 sv_catpv(sv, anyofs[i]);
15764 if ((flags & (ANYOF_ABOVE_LATIN1_ALL
15766 |ANYOF_NONBITMAP_NON_UTF8
15770 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
15771 if (flags & ANYOF_INVERT)
15772 /*make sure the invert info is in each */
15773 sv_catpvs(sv, "^");
15776 if (flags & ANYOF_NON_UTF8_NON_ASCII_ALL) {
15777 sv_catpvs(sv, "{non-utf8-latin1-all}");
15780 /* output information about the unicode matching */
15781 if (flags & ANYOF_ABOVE_LATIN1_ALL)
15782 sv_catpvs(sv, "{unicode_all}");
15783 else if (ARG(o) != ANYOF_NONBITMAP_EMPTY) {
15784 SV *lv; /* Set if there is something outside the bit map. */
15785 bool byte_output = FALSE; /* If something in the bitmap has
15787 SV *only_utf8_locale;
15789 /* Get the stuff that wasn't in the bitmap */
15790 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
15791 &lv, &only_utf8_locale);
15792 if (lv && lv != &PL_sv_undef) {
15793 char *s = savesvpv(lv);
15794 char * const origs = s;
15796 while (*s && *s != '\n')
15800 const char * const t = ++s;
15802 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
15803 sv_catpvs(sv, "{outside bitmap}");
15806 sv_catpvs(sv, "{utf8}");
15810 sv_catpvs(sv, " ");
15816 /* Truncate very long output */
15817 if (s - origs > 256) {
15818 Perl_sv_catpvf(aTHX_ sv,
15820 (int) (s - origs - 1),
15826 else if (*s == '\t') {
15840 SvREFCNT_dec_NN(lv);
15843 if ((flags & ANYOF_LOC_FOLD)
15844 && only_utf8_locale
15845 && only_utf8_locale != &PL_sv_undef)
15848 int max_entries = 256;
15850 sv_catpvs(sv, "{utf8 locale}");
15851 invlist_iterinit(only_utf8_locale);
15852 while (invlist_iternext(only_utf8_locale,
15854 put_range(sv, start, end);
15856 if (max_entries < 0) {
15857 sv_catpvs(sv, "...");
15861 invlist_iterfinish(only_utf8_locale);
15866 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
15868 else if (k == POSIXD || k == NPOSIXD) {
15869 U8 index = FLAGS(o) * 2;
15870 if (index < C_ARRAY_LENGTH(anyofs)) {
15871 if (*anyofs[index] != '[') {
15874 sv_catpv(sv, anyofs[index]);
15875 if (*anyofs[index] != '[') {
15880 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
15883 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
15884 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
15886 PERL_UNUSED_CONTEXT;
15887 PERL_UNUSED_ARG(sv);
15888 PERL_UNUSED_ARG(o);
15889 PERL_UNUSED_ARG(prog);
15890 PERL_UNUSED_ARG(reginfo);
15891 #endif /* DEBUGGING */
15897 Perl_re_intuit_string(pTHX_ REGEXP * const r)
15898 { /* Assume that RE_INTUIT is set */
15900 struct regexp *const prog = ReANY(r);
15901 GET_RE_DEBUG_FLAGS_DECL;
15903 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
15904 PERL_UNUSED_CONTEXT;
15908 const char * const s = SvPV_nolen_const(prog->check_substr
15909 ? prog->check_substr : prog->check_utf8);
15911 if (!PL_colorset) reginitcolors();
15912 PerlIO_printf(Perl_debug_log,
15913 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
15915 prog->check_substr ? "" : "utf8 ",
15916 PL_colors[5],PL_colors[0],
15919 (strlen(s) > 60 ? "..." : ""));
15922 return prog->check_substr ? prog->check_substr : prog->check_utf8;
15928 handles refcounting and freeing the perl core regexp structure. When
15929 it is necessary to actually free the structure the first thing it
15930 does is call the 'free' method of the regexp_engine associated to
15931 the regexp, allowing the handling of the void *pprivate; member
15932 first. (This routine is not overridable by extensions, which is why
15933 the extensions free is called first.)
15935 See regdupe and regdupe_internal if you change anything here.
15937 #ifndef PERL_IN_XSUB_RE
15939 Perl_pregfree(pTHX_ REGEXP *r)
15945 Perl_pregfree2(pTHX_ REGEXP *rx)
15948 struct regexp *const r = ReANY(rx);
15949 GET_RE_DEBUG_FLAGS_DECL;
15951 PERL_ARGS_ASSERT_PREGFREE2;
15953 if (r->mother_re) {
15954 ReREFCNT_dec(r->mother_re);
15956 CALLREGFREE_PVT(rx); /* free the private data */
15957 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15958 Safefree(r->xpv_len_u.xpvlenu_pv);
15961 SvREFCNT_dec(r->anchored_substr);
15962 SvREFCNT_dec(r->anchored_utf8);
15963 SvREFCNT_dec(r->float_substr);
15964 SvREFCNT_dec(r->float_utf8);
15965 Safefree(r->substrs);
15967 RX_MATCH_COPY_FREE(rx);
15968 #ifdef PERL_ANY_COW
15969 SvREFCNT_dec(r->saved_copy);
15972 SvREFCNT_dec(r->qr_anoncv);
15973 rx->sv_u.svu_rx = 0;
15978 This is a hacky workaround to the structural issue of match results
15979 being stored in the regexp structure which is in turn stored in
15980 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15981 could be PL_curpm in multiple contexts, and could require multiple
15982 result sets being associated with the pattern simultaneously, such
15983 as when doing a recursive match with (??{$qr})
15985 The solution is to make a lightweight copy of the regexp structure
15986 when a qr// is returned from the code executed by (??{$qr}) this
15987 lightweight copy doesn't actually own any of its data except for
15988 the starp/end and the actual regexp structure itself.
15994 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15996 struct regexp *ret;
15997 struct regexp *const r = ReANY(rx);
15998 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
16000 PERL_ARGS_ASSERT_REG_TEMP_COPY;
16003 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
16005 SvOK_off((SV *)ret_x);
16007 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
16008 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
16009 made both spots point to the same regexp body.) */
16010 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
16011 assert(!SvPVX(ret_x));
16012 ret_x->sv_u.svu_rx = temp->sv_any;
16013 temp->sv_any = NULL;
16014 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
16015 SvREFCNT_dec_NN(temp);
16016 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
16017 ing below will not set it. */
16018 SvCUR_set(ret_x, SvCUR(rx));
16021 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
16022 sv_force_normal(sv) is called. */
16024 ret = ReANY(ret_x);
16026 SvFLAGS(ret_x) |= SvUTF8(rx);
16027 /* We share the same string buffer as the original regexp, on which we
16028 hold a reference count, incremented when mother_re is set below.
16029 The string pointer is copied here, being part of the regexp struct.
16031 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
16032 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
16034 const I32 npar = r->nparens+1;
16035 Newx(ret->offs, npar, regexp_paren_pair);
16036 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16039 Newx(ret->substrs, 1, struct reg_substr_data);
16040 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16042 SvREFCNT_inc_void(ret->anchored_substr);
16043 SvREFCNT_inc_void(ret->anchored_utf8);
16044 SvREFCNT_inc_void(ret->float_substr);
16045 SvREFCNT_inc_void(ret->float_utf8);
16047 /* check_substr and check_utf8, if non-NULL, point to either their
16048 anchored or float namesakes, and don't hold a second reference. */
16050 RX_MATCH_COPIED_off(ret_x);
16051 #ifdef PERL_ANY_COW
16052 ret->saved_copy = NULL;
16054 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
16055 SvREFCNT_inc_void(ret->qr_anoncv);
16061 /* regfree_internal()
16063 Free the private data in a regexp. This is overloadable by
16064 extensions. Perl takes care of the regexp structure in pregfree(),
16065 this covers the *pprivate pointer which technically perl doesn't
16066 know about, however of course we have to handle the
16067 regexp_internal structure when no extension is in use.
16069 Note this is called before freeing anything in the regexp
16074 Perl_regfree_internal(pTHX_ REGEXP * const rx)
16077 struct regexp *const r = ReANY(rx);
16078 RXi_GET_DECL(r,ri);
16079 GET_RE_DEBUG_FLAGS_DECL;
16081 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
16087 SV *dsv= sv_newmortal();
16088 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
16089 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
16090 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
16091 PL_colors[4],PL_colors[5],s);
16094 #ifdef RE_TRACK_PATTERN_OFFSETS
16096 Safefree(ri->u.offsets); /* 20010421 MJD */
16098 if (ri->code_blocks) {
16100 for (n = 0; n < ri->num_code_blocks; n++)
16101 SvREFCNT_dec(ri->code_blocks[n].src_regex);
16102 Safefree(ri->code_blocks);
16106 int n = ri->data->count;
16109 /* If you add a ->what type here, update the comment in regcomp.h */
16110 switch (ri->data->what[n]) {
16116 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
16119 Safefree(ri->data->data[n]);
16125 { /* Aho Corasick add-on structure for a trie node.
16126 Used in stclass optimization only */
16128 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16130 refcount = --aho->refcount;
16133 PerlMemShared_free(aho->states);
16134 PerlMemShared_free(aho->fail);
16135 /* do this last!!!! */
16136 PerlMemShared_free(ri->data->data[n]);
16137 PerlMemShared_free(ri->regstclass);
16143 /* trie structure. */
16145 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16147 refcount = --trie->refcount;
16150 PerlMemShared_free(trie->charmap);
16151 PerlMemShared_free(trie->states);
16152 PerlMemShared_free(trie->trans);
16154 PerlMemShared_free(trie->bitmap);
16156 PerlMemShared_free(trie->jump);
16157 PerlMemShared_free(trie->wordinfo);
16158 /* do this last!!!! */
16159 PerlMemShared_free(ri->data->data[n]);
16164 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16165 ri->data->what[n]);
16168 Safefree(ri->data->what);
16169 Safefree(ri->data);
16175 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16176 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16177 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16180 re_dup - duplicate a regexp.
16182 This routine is expected to clone a given regexp structure. It is only
16183 compiled under USE_ITHREADS.
16185 After all of the core data stored in struct regexp is duplicated
16186 the regexp_engine.dupe method is used to copy any private data
16187 stored in the *pprivate pointer. This allows extensions to handle
16188 any duplication it needs to do.
16190 See pregfree() and regfree_internal() if you change anything here.
16192 #if defined(USE_ITHREADS)
16193 #ifndef PERL_IN_XSUB_RE
16195 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16199 const struct regexp *r = ReANY(sstr);
16200 struct regexp *ret = ReANY(dstr);
16202 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16204 npar = r->nparens+1;
16205 Newx(ret->offs, npar, regexp_paren_pair);
16206 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16208 if (ret->substrs) {
16209 /* Do it this way to avoid reading from *r after the StructCopy().
16210 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16211 cache, it doesn't matter. */
16212 const bool anchored = r->check_substr
16213 ? r->check_substr == r->anchored_substr
16214 : r->check_utf8 == r->anchored_utf8;
16215 Newx(ret->substrs, 1, struct reg_substr_data);
16216 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16218 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16219 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16220 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16221 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16223 /* check_substr and check_utf8, if non-NULL, point to either their
16224 anchored or float namesakes, and don't hold a second reference. */
16226 if (ret->check_substr) {
16228 assert(r->check_utf8 == r->anchored_utf8);
16229 ret->check_substr = ret->anchored_substr;
16230 ret->check_utf8 = ret->anchored_utf8;
16232 assert(r->check_substr == r->float_substr);
16233 assert(r->check_utf8 == r->float_utf8);
16234 ret->check_substr = ret->float_substr;
16235 ret->check_utf8 = ret->float_utf8;
16237 } else if (ret->check_utf8) {
16239 ret->check_utf8 = ret->anchored_utf8;
16241 ret->check_utf8 = ret->float_utf8;
16246 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16247 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16250 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16252 if (RX_MATCH_COPIED(dstr))
16253 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16255 ret->subbeg = NULL;
16256 #ifdef PERL_ANY_COW
16257 ret->saved_copy = NULL;
16260 /* Whether mother_re be set or no, we need to copy the string. We
16261 cannot refrain from copying it when the storage points directly to
16262 our mother regexp, because that's
16263 1: a buffer in a different thread
16264 2: something we no longer hold a reference on
16265 so we need to copy it locally. */
16266 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16267 ret->mother_re = NULL;
16269 #endif /* PERL_IN_XSUB_RE */
16274 This is the internal complement to regdupe() which is used to copy
16275 the structure pointed to by the *pprivate pointer in the regexp.
16276 This is the core version of the extension overridable cloning hook.
16277 The regexp structure being duplicated will be copied by perl prior
16278 to this and will be provided as the regexp *r argument, however
16279 with the /old/ structures pprivate pointer value. Thus this routine
16280 may override any copying normally done by perl.
16282 It returns a pointer to the new regexp_internal structure.
16286 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
16289 struct regexp *const r = ReANY(rx);
16290 regexp_internal *reti;
16292 RXi_GET_DECL(r,ri);
16294 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
16298 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
16299 char, regexp_internal);
16300 Copy(ri->program, reti->program, len+1, regnode);
16302 reti->num_code_blocks = ri->num_code_blocks;
16303 if (ri->code_blocks) {
16305 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
16306 struct reg_code_block);
16307 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
16308 struct reg_code_block);
16309 for (n = 0; n < ri->num_code_blocks; n++)
16310 reti->code_blocks[n].src_regex = (REGEXP*)
16311 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
16314 reti->code_blocks = NULL;
16316 reti->regstclass = NULL;
16319 struct reg_data *d;
16320 const int count = ri->data->count;
16323 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
16324 char, struct reg_data);
16325 Newx(d->what, count, U8);
16328 for (i = 0; i < count; i++) {
16329 d->what[i] = ri->data->what[i];
16330 switch (d->what[i]) {
16331 /* see also regcomp.h and regfree_internal() */
16332 case 'a': /* actually an AV, but the dup function is identical. */
16336 case 'u': /* actually an HV, but the dup function is identical. */
16337 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
16340 /* This is cheating. */
16341 Newx(d->data[i], 1, regnode_ssc);
16342 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
16343 reti->regstclass = (regnode*)d->data[i];
16346 /* Trie stclasses are readonly and can thus be shared
16347 * without duplication. We free the stclass in pregfree
16348 * when the corresponding reg_ac_data struct is freed.
16350 reti->regstclass= ri->regstclass;
16354 ((reg_trie_data*)ri->data->data[i])->refcount++;
16359 d->data[i] = ri->data->data[i];
16362 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
16363 ri->data->what[i]);
16372 reti->name_list_idx = ri->name_list_idx;
16374 #ifdef RE_TRACK_PATTERN_OFFSETS
16375 if (ri->u.offsets) {
16376 Newx(reti->u.offsets, 2*len+1, U32);
16377 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
16380 SetProgLen(reti,len);
16383 return (void*)reti;
16386 #endif /* USE_ITHREADS */
16388 #ifndef PERL_IN_XSUB_RE
16391 - regnext - dig the "next" pointer out of a node
16394 Perl_regnext(pTHX_ regnode *p)
16402 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
16403 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16404 (int)OP(p), (int)REGNODE_MAX);
16407 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
16416 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
16419 STRLEN l1 = strlen(pat1);
16420 STRLEN l2 = strlen(pat2);
16423 const char *message;
16425 PERL_ARGS_ASSERT_RE_CROAK2;
16431 Copy(pat1, buf, l1 , char);
16432 Copy(pat2, buf + l1, l2 , char);
16433 buf[l1 + l2] = '\n';
16434 buf[l1 + l2 + 1] = '\0';
16435 va_start(args, pat2);
16436 msv = vmess(buf, &args);
16438 message = SvPV_const(msv,l1);
16441 Copy(message, buf, l1 , char);
16442 /* l1-1 to avoid \n */
16443 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
16446 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
16448 #ifndef PERL_IN_XSUB_RE
16450 Perl_save_re_context(pTHX)
16454 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
16456 const REGEXP * const rx = PM_GETRE(PL_curpm);
16459 for (i = 1; i <= RX_NPARENS(rx); i++) {
16460 char digits[TYPE_CHARS(long)];
16461 const STRLEN len = my_snprintf(digits, sizeof(digits),
16463 GV *const *const gvp
16464 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
16467 GV * const gv = *gvp;
16468 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
16480 S_put_byte(pTHX_ SV *sv, int c)
16482 PERL_ARGS_ASSERT_PUT_BYTE;
16486 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
16487 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
16488 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
16489 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
16490 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
16493 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
16498 const char string = c;
16499 if (c == '-' || c == ']' || c == '\\' || c == '^')
16500 sv_catpvs(sv, "\\");
16501 sv_catpvn(sv, &string, 1);
16506 S_put_range(pTHX_ SV *sv, UV start, UV end)
16509 /* Appends to 'sv' a displayable version of the range of code points from
16510 * 'start' to 'end' */
16512 assert(start <= end);
16514 PERL_ARGS_ASSERT_PUT_RANGE;
16516 if (end - start < 3) { /* Individual chars in short ranges */
16517 for (; start <= end; start++)
16518 put_byte(sv, start);
16520 else if ( end > 255
16521 || ! isALPHANUMERIC(start)
16522 || ! isALPHANUMERIC(end)
16523 || isDIGIT(start) != isDIGIT(end)
16524 || isUPPER(start) != isUPPER(end)
16525 || isLOWER(start) != isLOWER(end)
16527 /* This final test should get optimized out except on EBCDIC
16528 * platforms, where it causes ranges that cross discontinuities
16529 * like i/j to be shown as hex instead of the misleading,
16530 * e.g. H-K (since that range includes more than H, I, J, K).
16532 || (end - start) != NATIVE_TO_ASCII(end) - NATIVE_TO_ASCII(start))
16534 Perl_sv_catpvf(aTHX_ sv, "\\x{%02" UVXf "}-\\x{%02" UVXf "}",
16536 (end < 256) ? end : 255);
16538 else { /* Here, the ends of the range are both digits, or both uppercase,
16539 or both lowercase; and there's no discontinuity in the range
16540 (which could happen on EBCDIC platforms) */
16541 put_byte(sv, start);
16542 sv_catpvs(sv, "-");
16548 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
16550 /* Appends to 'sv' a displayable version of the innards of the bracketed
16551 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
16552 * output anything */
16555 bool has_output_anything = FALSE;
16557 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
16559 for (i = 0; i < 256; i++) {
16560 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
16562 /* The character at index i should be output. Find the next
16563 * character that should NOT be output */
16565 for (j = i + 1; j <= 256; j++) {
16566 if (! BITMAP_TEST((U8 *) bitmap, j)) {
16571 /* Everything between them is a single range that should be output
16573 put_range(sv, i, j - 1);
16574 has_output_anything = TRUE;
16579 return has_output_anything;
16582 #define CLEAR_OPTSTART \
16583 if (optstart) STMT_START { \
16584 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
16585 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
16589 #define DUMPUNTIL(b,e) \
16591 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
16593 STATIC const regnode *
16594 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
16595 const regnode *last, const regnode *plast,
16596 SV* sv, I32 indent, U32 depth)
16599 U8 op = PSEUDO; /* Arbitrary non-END op. */
16600 const regnode *next;
16601 const regnode *optstart= NULL;
16603 RXi_GET_DECL(r,ri);
16604 GET_RE_DEBUG_FLAGS_DECL;
16606 PERL_ARGS_ASSERT_DUMPUNTIL;
16608 #ifdef DEBUG_DUMPUNTIL
16609 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
16610 last ? last-start : 0,plast ? plast-start : 0);
16613 if (plast && plast < last)
16616 while (PL_regkind[op] != END && (!last || node < last)) {
16617 /* While that wasn't END last time... */
16620 if (op == CLOSE || op == WHILEM)
16622 next = regnext((regnode *)node);
16625 if (OP(node) == OPTIMIZED) {
16626 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
16633 regprop(r, sv, node, NULL);
16634 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
16635 (int)(2*indent + 1), "", SvPVX_const(sv));
16637 if (OP(node) != OPTIMIZED) {
16638 if (next == NULL) /* Next ptr. */
16639 PerlIO_printf(Perl_debug_log, " (0)");
16640 else if (PL_regkind[(U8)op] == BRANCH
16641 && PL_regkind[OP(next)] != BRANCH )
16642 PerlIO_printf(Perl_debug_log, " (FAIL)");
16644 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
16645 (void)PerlIO_putc(Perl_debug_log, '\n');
16649 if (PL_regkind[(U8)op] == BRANCHJ) {
16652 const regnode *nnode = (OP(next) == LONGJMP
16653 ? regnext((regnode *)next)
16655 if (last && nnode > last)
16657 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
16660 else if (PL_regkind[(U8)op] == BRANCH) {
16662 DUMPUNTIL(NEXTOPER(node), next);
16664 else if ( PL_regkind[(U8)op] == TRIE ) {
16665 const regnode *this_trie = node;
16666 const char op = OP(node);
16667 const U32 n = ARG(node);
16668 const reg_ac_data * const ac = op>=AHOCORASICK ?
16669 (reg_ac_data *)ri->data->data[n] :
16671 const reg_trie_data * const trie =
16672 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
16674 AV *const trie_words
16675 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
16677 const regnode *nextbranch= NULL;
16680 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
16681 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
16683 PerlIO_printf(Perl_debug_log, "%*s%s ",
16684 (int)(2*(indent+3)), "",
16686 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
16687 SvCUR(*elem_ptr), 60,
16688 PL_colors[0], PL_colors[1],
16690 ? PERL_PV_ESCAPE_UNI
16692 | PERL_PV_PRETTY_ELLIPSES
16693 | PERL_PV_PRETTY_LTGT
16698 U16 dist= trie->jump[word_idx+1];
16699 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
16700 (UV)((dist ? this_trie + dist : next) - start));
16703 nextbranch= this_trie + trie->jump[0];
16704 DUMPUNTIL(this_trie + dist, nextbranch);
16706 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
16707 nextbranch= regnext((regnode *)nextbranch);
16709 PerlIO_printf(Perl_debug_log, "\n");
16712 if (last && next > last)
16717 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
16718 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
16719 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
16721 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
16723 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
16725 else if ( op == PLUS || op == STAR) {
16726 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
16728 else if (PL_regkind[(U8)op] == ANYOF) {
16729 /* arglen 1 + class block */
16730 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_POSIXL)
16731 ? ANYOF_POSIXL_SKIP
16733 node = NEXTOPER(node);
16735 else if (PL_regkind[(U8)op] == EXACT) {
16736 /* Literal string, where present. */
16737 node += NODE_SZ_STR(node) - 1;
16738 node = NEXTOPER(node);
16741 node = NEXTOPER(node);
16742 node += regarglen[(U8)op];
16744 if (op == CURLYX || op == OPEN)
16748 #ifdef DEBUG_DUMPUNTIL
16749 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
16754 #endif /* DEBUGGING */
16758 * c-indentation-style: bsd
16759 * c-basic-offset: 4
16760 * indent-tabs-mode: nil
16763 * ex: set ts=8 sts=4 sw=4 et: