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 */
2935 /* Finish populating the prev field of the wordinfo array. Walk back
2936 * from each accept state until we find another accept state, and if
2937 * so, point the first word's .prev field at the second word. If the
2938 * second already has a .prev field set, stop now. This will be the
2939 * case either if we've already processed that word's accept state,
2940 * or that state had multiple words, and the overspill words were
2941 * already linked up earlier.
2948 for (word=1; word <= trie->wordcount; word++) {
2950 if (trie->wordinfo[word].prev)
2952 state = trie->wordinfo[word].accept;
2954 state = prev_states[state];
2957 prev = trie->states[state].wordnum;
2961 trie->wordinfo[word].prev = prev;
2963 Safefree(prev_states);
2967 /* and now dump out the compressed format */
2968 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2970 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2972 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2973 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2975 SvREFCNT_dec_NN(revcharmap);
2979 : trie->startstate>1
2985 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2987 /* The Trie is constructed and compressed now so we can build a fail array if
2990 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2992 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2996 We find the fail state for each state in the trie, this state is the longest
2997 proper suffix of the current state's 'word' that is also a proper prefix of
2998 another word in our trie. State 1 represents the word '' and is thus the
2999 default fail state. This allows the DFA not to have to restart after its
3000 tried and failed a word at a given point, it simply continues as though it
3001 had been matching the other word in the first place.
3003 'abcdgu'=~/abcdefg|cdgu/
3004 When we get to 'd' we are still matching the first word, we would encounter
3005 'g' which would fail, which would bring us to the state representing 'd' in
3006 the second word where we would try 'g' and succeed, proceeding to match
3009 /* add a fail transition */
3010 const U32 trie_offset = ARG(source);
3011 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3013 const U32 ucharcount = trie->uniquecharcount;
3014 const U32 numstates = trie->statecount;
3015 const U32 ubound = trie->lasttrans + ucharcount;
3019 U32 base = trie->states[ 1 ].trans.base;
3022 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3023 GET_RE_DEBUG_FLAGS_DECL;
3025 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
3027 PERL_UNUSED_ARG(depth);
3031 ARG_SET( stclass, data_slot );
3032 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3033 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3034 aho->trie=trie_offset;
3035 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3036 Copy( trie->states, aho->states, numstates, reg_trie_state );
3037 Newxz( q, numstates, U32);
3038 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3041 /* initialize fail[0..1] to be 1 so that we always have
3042 a valid final fail state */
3043 fail[ 0 ] = fail[ 1 ] = 1;
3045 for ( charid = 0; charid < ucharcount ; charid++ ) {
3046 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3048 q[ q_write ] = newstate;
3049 /* set to point at the root */
3050 fail[ q[ q_write++ ] ]=1;
3053 while ( q_read < q_write) {
3054 const U32 cur = q[ q_read++ % numstates ];
3055 base = trie->states[ cur ].trans.base;
3057 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3058 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3060 U32 fail_state = cur;
3063 fail_state = fail[ fail_state ];
3064 fail_base = aho->states[ fail_state ].trans.base;
3065 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3067 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3068 fail[ ch_state ] = fail_state;
3069 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3071 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3073 q[ q_write++ % numstates] = ch_state;
3077 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3078 when we fail in state 1, this allows us to use the
3079 charclass scan to find a valid start char. This is based on the principle
3080 that theres a good chance the string being searched contains lots of stuff
3081 that cant be a start char.
3083 fail[ 0 ] = fail[ 1 ] = 0;
3084 DEBUG_TRIE_COMPILE_r({
3085 PerlIO_printf(Perl_debug_log,
3086 "%*sStclass Failtable (%"UVuf" states): 0",
3087 (int)(depth * 2), "", (UV)numstates
3089 for( q_read=1; q_read<numstates; q_read++ ) {
3090 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3092 PerlIO_printf(Perl_debug_log, "\n");
3095 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3099 #define DEBUG_PEEP(str,scan,depth) \
3100 DEBUG_OPTIMISE_r({if (scan){ \
3101 SV * const mysv=sv_newmortal(); \
3102 regnode *Next = regnext(scan); \
3103 regprop(RExC_rx, mysv, scan, NULL); \
3104 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3105 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3106 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3110 /* The below joins as many adjacent EXACTish nodes as possible into a single
3111 * one. The regop may be changed if the node(s) contain certain sequences that
3112 * require special handling. The joining is only done if:
3113 * 1) there is room in the current conglomerated node to entirely contain the
3115 * 2) they are the exact same node type
3117 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3118 * these get optimized out
3120 * If a node is to match under /i (folded), the number of characters it matches
3121 * can be different than its character length if it contains a multi-character
3122 * fold. *min_subtract is set to the total delta number of characters of the
3125 * And *unfolded_multi_char is set to indicate whether or not the node contains
3126 * an unfolded multi-char fold. This happens when whether the fold is valid or
3127 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3128 * SMALL LETTER SHARP S, as only if the target string being matched against
3129 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3130 * folding rules depend on the locale in force at runtime. (Multi-char folds
3131 * whose components are all above the Latin1 range are not run-time locale
3132 * dependent, and have already been folded by the time this function is
3135 * This is as good a place as any to discuss the design of handling these
3136 * multi-character fold sequences. It's been wrong in Perl for a very long
3137 * time. There are three code points in Unicode whose multi-character folds
3138 * were long ago discovered to mess things up. The previous designs for
3139 * dealing with these involved assigning a special node for them. This
3140 * approach doesn't always work, as evidenced by this example:
3141 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3142 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3143 * would match just the \xDF, it won't be able to handle the case where a
3144 * successful match would have to cross the node's boundary. The new approach
3145 * that hopefully generally solves the problem generates an EXACTFU_SS node
3146 * that is "sss" in this case.
3148 * It turns out that there are problems with all multi-character folds, and not
3149 * just these three. Now the code is general, for all such cases. The
3150 * approach taken is:
3151 * 1) This routine examines each EXACTFish node that could contain multi-
3152 * character folded sequences. Since a single character can fold into
3153 * such a sequence, the minimum match length for this node is less than
3154 * the number of characters in the node. This routine returns in
3155 * *min_subtract how many characters to subtract from the the actual
3156 * length of the string to get a real minimum match length; it is 0 if
3157 * there are no multi-char foldeds. This delta is used by the caller to
3158 * adjust the min length of the match, and the delta between min and max,
3159 * so that the optimizer doesn't reject these possibilities based on size
3161 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3162 * is used for an EXACTFU node that contains at least one "ss" sequence in
3163 * it. For non-UTF-8 patterns and strings, this is the only case where
3164 * there is a possible fold length change. That means that a regular
3165 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3166 * with length changes, and so can be processed faster. regexec.c takes
3167 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3168 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3169 * known until runtime). This saves effort in regex matching. However,
3170 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3171 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3172 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3173 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3174 * possibilities for the non-UTF8 patterns are quite simple, except for
3175 * the sharp s. All the ones that don't involve a UTF-8 target string are
3176 * members of a fold-pair, and arrays are set up for all of them so that
3177 * the other member of the pair can be found quickly. Code elsewhere in
3178 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3179 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3180 * described in the next item.
3181 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3182 * validity of the fold won't be known until runtime, and so must remain
3183 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3184 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3185 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3186 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3187 * The reason this is a problem is that the optimizer part of regexec.c
3188 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3189 * that a character in the pattern corresponds to at most a single
3190 * character in the target string. (And I do mean character, and not byte
3191 * here, unlike other parts of the documentation that have never been
3192 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3193 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3194 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3195 * nodes, violate the assumption, and they are the only instances where it
3196 * is violated. I'm reluctant to try to change the assumption, as the
3197 * code involved is impenetrable to me (khw), so instead the code here
3198 * punts. This routine examines EXACTFL nodes, and (when the pattern
3199 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3200 * boolean indicating whether or not the node contains such a fold. When
3201 * it is true, the caller sets a flag that later causes the optimizer in
3202 * this file to not set values for the floating and fixed string lengths,
3203 * and thus avoids the optimizer code in regexec.c that makes the invalid
3204 * assumption. Thus, there is no optimization based on string lengths for
3205 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3206 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3207 * assumption is wrong only in these cases is that all other non-UTF-8
3208 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3209 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3210 * EXACTF nodes because we don't know at compile time if it actually
3211 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3212 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3213 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3214 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3215 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3216 * string would require the pattern to be forced into UTF-8, the overhead
3217 * of which we want to avoid. Similarly the unfolded multi-char folds in
3218 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3221 * Similarly, the code that generates tries doesn't currently handle
3222 * not-already-folded multi-char folds, and it looks like a pain to change
3223 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3224 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3225 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3226 * using /iaa matching will be doing so almost entirely with ASCII
3227 * strings, so this should rarely be encountered in practice */
3229 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3230 if (PL_regkind[OP(scan)] == EXACT) \
3231 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3234 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3235 UV *min_subtract, bool *unfolded_multi_char,
3236 U32 flags,regnode *val, U32 depth)
3238 /* Merge several consecutive EXACTish nodes into one. */
3239 regnode *n = regnext(scan);
3241 regnode *next = scan + NODE_SZ_STR(scan);
3245 regnode *stop = scan;
3246 GET_RE_DEBUG_FLAGS_DECL;
3248 PERL_UNUSED_ARG(depth);
3251 PERL_ARGS_ASSERT_JOIN_EXACT;
3252 #ifndef EXPERIMENTAL_INPLACESCAN
3253 PERL_UNUSED_ARG(flags);
3254 PERL_UNUSED_ARG(val);
3256 DEBUG_PEEP("join",scan,depth);
3258 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3259 * EXACT ones that are mergeable to the current one. */
3261 && (PL_regkind[OP(n)] == NOTHING
3262 || (stringok && OP(n) == OP(scan)))
3264 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3267 if (OP(n) == TAIL || n > next)
3269 if (PL_regkind[OP(n)] == NOTHING) {
3270 DEBUG_PEEP("skip:",n,depth);
3271 NEXT_OFF(scan) += NEXT_OFF(n);
3272 next = n + NODE_STEP_REGNODE;
3279 else if (stringok) {
3280 const unsigned int oldl = STR_LEN(scan);
3281 regnode * const nnext = regnext(n);
3283 /* XXX I (khw) kind of doubt that this works on platforms (should
3284 * Perl ever run on one) where U8_MAX is above 255 because of lots
3285 * of other assumptions */
3286 /* Don't join if the sum can't fit into a single node */
3287 if (oldl + STR_LEN(n) > U8_MAX)
3290 DEBUG_PEEP("merg",n,depth);
3293 NEXT_OFF(scan) += NEXT_OFF(n);
3294 STR_LEN(scan) += STR_LEN(n);
3295 next = n + NODE_SZ_STR(n);
3296 /* Now we can overwrite *n : */
3297 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3305 #ifdef EXPERIMENTAL_INPLACESCAN
3306 if (flags && !NEXT_OFF(n)) {
3307 DEBUG_PEEP("atch", val, depth);
3308 if (reg_off_by_arg[OP(n)]) {
3309 ARG_SET(n, val - n);
3312 NEXT_OFF(n) = val - n;
3320 *unfolded_multi_char = FALSE;
3322 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3323 * can now analyze for sequences of problematic code points. (Prior to
3324 * this final joining, sequences could have been split over boundaries, and
3325 * hence missed). The sequences only happen in folding, hence for any
3326 * non-EXACT EXACTish node */
3327 if (OP(scan) != EXACT) {
3328 U8* s0 = (U8*) STRING(scan);
3330 U8* s_end = s0 + STR_LEN(scan);
3332 int total_count_delta = 0; /* Total delta number of characters that
3333 multi-char folds expand to */
3335 /* One pass is made over the node's string looking for all the
3336 * possibilities. To avoid some tests in the loop, there are two main
3337 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3342 if (OP(scan) == EXACTFL) {
3345 /* An EXACTFL node would already have been changed to another
3346 * node type unless there is at least one character in it that
3347 * is problematic; likely a character whose fold definition
3348 * won't be known until runtime, and so has yet to be folded.
3349 * For all but the UTF-8 locale, folds are 1-1 in length, but
3350 * to handle the UTF-8 case, we need to create a temporary
3351 * folded copy using UTF-8 locale rules in order to analyze it.
3352 * This is because our macros that look to see if a sequence is
3353 * a multi-char fold assume everything is folded (otherwise the
3354 * tests in those macros would be too complicated and slow).
3355 * Note that here, the non-problematic folds will have already
3356 * been done, so we can just copy such characters. We actually
3357 * don't completely fold the EXACTFL string. We skip the
3358 * unfolded multi-char folds, as that would just create work
3359 * below to figure out the size they already are */
3361 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3364 STRLEN s_len = UTF8SKIP(s);
3365 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3366 Copy(s, d, s_len, U8);
3369 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3370 *unfolded_multi_char = TRUE;
3371 Copy(s, d, s_len, U8);
3374 else if (isASCII(*s)) {
3375 *(d++) = toFOLD(*s);
3379 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3385 /* Point the remainder of the routine to look at our temporary
3389 } /* End of creating folded copy of EXACTFL string */
3391 /* Examine the string for a multi-character fold sequence. UTF-8
3392 * patterns have all characters pre-folded by the time this code is
3394 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3395 length sequence we are looking for is 2 */
3397 int count = 0; /* How many characters in a multi-char fold */
3398 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3399 if (! len) { /* Not a multi-char fold: get next char */
3404 /* Nodes with 'ss' require special handling, except for
3405 * EXACTFA-ish for which there is no multi-char fold to this */
3406 if (len == 2 && *s == 's' && *(s+1) == 's'
3407 && OP(scan) != EXACTFA
3408 && OP(scan) != EXACTFA_NO_TRIE)
3411 if (OP(scan) != EXACTFL) {
3412 OP(scan) = EXACTFU_SS;
3416 else { /* Here is a generic multi-char fold. */
3417 U8* multi_end = s + len;
3419 /* Count how many characters in it. In the case of /aa, no
3420 * folds which contain ASCII code points are allowed, so
3421 * check for those, and skip if found. */
3422 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3423 count = utf8_length(s, multi_end);
3427 while (s < multi_end) {
3430 goto next_iteration;
3440 /* The delta is how long the sequence is minus 1 (1 is how long
3441 * the character that folds to the sequence is) */
3442 total_count_delta += count - 1;
3446 /* We created a temporary folded copy of the string in EXACTFL
3447 * nodes. Therefore we need to be sure it doesn't go below zero,
3448 * as the real string could be shorter */
3449 if (OP(scan) == EXACTFL) {
3450 int total_chars = utf8_length((U8*) STRING(scan),
3451 (U8*) STRING(scan) + STR_LEN(scan));
3452 if (total_count_delta > total_chars) {
3453 total_count_delta = total_chars;
3457 *min_subtract += total_count_delta;
3460 else if (OP(scan) == EXACTFA) {
3462 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3463 * fold to the ASCII range (and there are no existing ones in the
3464 * upper latin1 range). But, as outlined in the comments preceding
3465 * this function, we need to flag any occurrences of the sharp s.
3466 * This character forbids trie formation (because of added
3469 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3470 OP(scan) = EXACTFA_NO_TRIE;
3471 *unfolded_multi_char = TRUE;
3480 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3481 * folds that are all Latin1. As explained in the comments
3482 * preceding this function, we look also for the sharp s in EXACTF
3483 * and EXACTFL nodes; it can be in the final position. Otherwise
3484 * we can stop looking 1 byte earlier because have to find at least
3485 * two characters for a multi-fold */
3486 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3491 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3492 if (! len) { /* Not a multi-char fold. */
3493 if (*s == LATIN_SMALL_LETTER_SHARP_S
3494 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3496 *unfolded_multi_char = TRUE;
3503 && isARG2_lower_or_UPPER_ARG1('s', *s)
3504 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3507 /* EXACTF nodes need to know that the minimum length
3508 * changed so that a sharp s in the string can match this
3509 * ss in the pattern, but they remain EXACTF nodes, as they
3510 * won't match this unless the target string is is UTF-8,
3511 * which we don't know until runtime. EXACTFL nodes can't
3512 * transform into EXACTFU nodes */
3513 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3514 OP(scan) = EXACTFU_SS;
3518 *min_subtract += len - 1;
3525 /* Allow dumping but overwriting the collection of skipped
3526 * ops and/or strings with fake optimized ops */
3527 n = scan + NODE_SZ_STR(scan);
3535 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3539 /* REx optimizer. Converts nodes into quicker variants "in place".
3540 Finds fixed substrings. */
3542 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3543 to the position after last scanned or to NULL. */
3545 #define INIT_AND_WITHP \
3546 assert(!and_withp); \
3547 Newx(and_withp,1, regnode_ssc); \
3548 SAVEFREEPV(and_withp)
3550 /* this is a chain of data about sub patterns we are processing that
3551 need to be handled separately/specially in study_chunk. Its so
3552 we can simulate recursion without losing state. */
3554 typedef struct scan_frame {
3555 regnode *last; /* last node to process in this frame */
3556 regnode *next; /* next node to process when last is reached */
3557 struct scan_frame *prev; /*previous frame*/
3558 U32 prev_recursed_depth;
3559 I32 stop; /* what stopparen do we use */
3564 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3565 SSize_t *minlenp, SSize_t *deltap,
3570 regnode_ssc *and_withp,
3571 U32 flags, U32 depth)
3572 /* scanp: Start here (read-write). */
3573 /* deltap: Write maxlen-minlen here. */
3574 /* last: Stop before this one. */
3575 /* data: string data about the pattern */
3576 /* stopparen: treat close N as END */
3577 /* recursed: which subroutines have we recursed into */
3578 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3581 /* There must be at least this number of characters to match */
3584 regnode *scan = *scanp, *next;
3586 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3587 int is_inf_internal = 0; /* The studied chunk is infinite */
3588 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3589 scan_data_t data_fake;
3590 SV *re_trie_maxbuff = NULL;
3591 regnode *first_non_open = scan;
3592 SSize_t stopmin = SSize_t_MAX;
3593 scan_frame *frame = NULL;
3594 GET_RE_DEBUG_FLAGS_DECL;
3596 PERL_ARGS_ASSERT_STUDY_CHUNK;
3599 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3602 while (first_non_open && OP(first_non_open) == OPEN)
3603 first_non_open=regnext(first_non_open);
3608 while ( scan && OP(scan) != END && scan < last ){
3609 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3610 node length to get a real minimum (because
3611 the folded version may be shorter) */
3612 bool unfolded_multi_char = FALSE;
3613 /* Peephole optimizer: */
3614 DEBUG_OPTIMISE_MORE_r(
3616 PerlIO_printf(Perl_debug_log,
3617 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3618 ((int) depth*2), "", (long)stopparen,
3619 (unsigned long)depth, (unsigned long)recursed_depth);
3620 if (recursed_depth) {
3623 for ( j = 0 ; j < recursed_depth ; j++ ) {
3624 PerlIO_printf(Perl_debug_log,"[");
3625 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3626 PerlIO_printf(Perl_debug_log,"%d",
3627 PAREN_TEST(RExC_study_chunk_recursed +
3628 (j * RExC_study_chunk_recursed_bytes), i)
3631 PerlIO_printf(Perl_debug_log,"]");
3634 PerlIO_printf(Perl_debug_log,"\n");
3637 DEBUG_STUDYDATA("Peep:", data, depth);
3638 DEBUG_PEEP("Peep", scan, depth);
3641 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3642 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3643 * by a different invocation of reg() -- Yves
3645 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3647 /* Follow the next-chain of the current node and optimize
3648 away all the NOTHINGs from it. */
3649 if (OP(scan) != CURLYX) {
3650 const int max = (reg_off_by_arg[OP(scan)]
3652 /* I32 may be smaller than U16 on CRAYs! */
3653 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3654 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3658 /* Skip NOTHING and LONGJMP. */
3659 while ((n = regnext(n))
3660 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3661 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3662 && off + noff < max)
3664 if (reg_off_by_arg[OP(scan)])
3667 NEXT_OFF(scan) = off;
3672 /* The principal pseudo-switch. Cannot be a switch, since we
3673 look into several different things. */
3674 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3675 || OP(scan) == IFTHEN) {
3676 next = regnext(scan);
3678 /* demq: the op(next)==code check is to see if we have
3679 * "branch-branch" AFAICT */
3681 if (OP(next) == code || code == IFTHEN) {
3682 /* NOTE - There is similar code to this block below for
3683 * handling TRIE nodes on a re-study. If you change stuff here
3684 * check there too. */
3685 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3687 regnode * const startbranch=scan;
3689 if (flags & SCF_DO_SUBSTR) {
3690 /* Cannot merge strings after this. */
3691 scan_commit(pRExC_state, data, minlenp, is_inf);
3694 if (flags & SCF_DO_STCLASS)
3695 ssc_init_zero(pRExC_state, &accum);
3697 while (OP(scan) == code) {
3698 SSize_t deltanext, minnext, fake;
3700 regnode_ssc this_class;
3703 data_fake.flags = 0;
3705 data_fake.whilem_c = data->whilem_c;
3706 data_fake.last_closep = data->last_closep;
3709 data_fake.last_closep = &fake;
3711 data_fake.pos_delta = delta;
3712 next = regnext(scan);
3713 scan = NEXTOPER(scan);
3715 scan = NEXTOPER(scan);
3716 if (flags & SCF_DO_STCLASS) {
3717 ssc_init(pRExC_state, &this_class);
3718 data_fake.start_class = &this_class;
3719 f = SCF_DO_STCLASS_AND;
3721 if (flags & SCF_WHILEM_VISITED_POS)
3722 f |= SCF_WHILEM_VISITED_POS;
3724 /* we suppose the run is continuous, last=next...*/
3725 minnext = study_chunk(pRExC_state, &scan, minlenp,
3726 &deltanext, next, &data_fake, stopparen,
3727 recursed_depth, NULL, f,depth+1);
3730 if (deltanext == SSize_t_MAX) {
3731 is_inf = is_inf_internal = 1;
3733 } else if (max1 < minnext + deltanext)
3734 max1 = minnext + deltanext;
3736 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3738 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3739 if ( stopmin > minnext)
3740 stopmin = min + min1;
3741 flags &= ~SCF_DO_SUBSTR;
3743 data->flags |= SCF_SEEN_ACCEPT;
3746 if (data_fake.flags & SF_HAS_EVAL)
3747 data->flags |= SF_HAS_EVAL;
3748 data->whilem_c = data_fake.whilem_c;
3750 if (flags & SCF_DO_STCLASS)
3751 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3753 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3755 if (flags & SCF_DO_SUBSTR) {
3756 data->pos_min += min1;
3757 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3758 data->pos_delta = SSize_t_MAX;
3760 data->pos_delta += max1 - min1;
3761 if (max1 != min1 || is_inf)
3762 data->longest = &(data->longest_float);
3765 if (delta == SSize_t_MAX
3766 || SSize_t_MAX - delta - (max1 - min1) < 0)
3767 delta = SSize_t_MAX;
3769 delta += max1 - min1;
3770 if (flags & SCF_DO_STCLASS_OR) {
3771 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3773 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3774 flags &= ~SCF_DO_STCLASS;
3777 else if (flags & SCF_DO_STCLASS_AND) {
3779 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3780 flags &= ~SCF_DO_STCLASS;
3783 /* Switch to OR mode: cache the old value of
3784 * data->start_class */
3786 StructCopy(data->start_class, and_withp, regnode_ssc);
3787 flags &= ~SCF_DO_STCLASS_AND;
3788 StructCopy(&accum, data->start_class, regnode_ssc);
3789 flags |= SCF_DO_STCLASS_OR;
3793 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3794 OP( startbranch ) == BRANCH )
3798 Assuming this was/is a branch we are dealing with: 'scan'
3799 now points at the item that follows the branch sequence,
3800 whatever it is. We now start at the beginning of the
3801 sequence and look for subsequences of
3807 which would be constructed from a pattern like
3810 If we can find such a subsequence we need to turn the first
3811 element into a trie and then add the subsequent branch exact
3812 strings to the trie.
3816 1. patterns where the whole set of branches can be
3819 2. patterns where only a subset can be converted.
3821 In case 1 we can replace the whole set with a single regop
3822 for the trie. In case 2 we need to keep the start and end
3825 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3826 becomes BRANCH TRIE; BRANCH X;
3828 There is an additional case, that being where there is a
3829 common prefix, which gets split out into an EXACT like node
3830 preceding the TRIE node.
3832 If x(1..n)==tail then we can do a simple trie, if not we make
3833 a "jump" trie, such that when we match the appropriate word
3834 we "jump" to the appropriate tail node. Essentially we turn
3835 a nested if into a case structure of sorts.
3840 if (!re_trie_maxbuff) {
3841 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3842 if (!SvIOK(re_trie_maxbuff))
3843 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3845 if ( SvIV(re_trie_maxbuff)>=0 ) {
3847 regnode *first = (regnode *)NULL;
3848 regnode *last = (regnode *)NULL;
3849 regnode *tail = scan;
3854 SV * const mysv = sv_newmortal(); /* for dumping */
3856 /* var tail is used because there may be a TAIL
3857 regop in the way. Ie, the exacts will point to the
3858 thing following the TAIL, but the last branch will
3859 point at the TAIL. So we advance tail. If we
3860 have nested (?:) we may have to move through several
3864 while ( OP( tail ) == TAIL ) {
3865 /* this is the TAIL generated by (?:) */
3866 tail = regnext( tail );
3870 DEBUG_TRIE_COMPILE_r({
3871 regprop(RExC_rx, mysv, tail, NULL);
3872 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3873 (int)depth * 2 + 2, "",
3874 "Looking for TRIE'able sequences. Tail node is: ",
3875 SvPV_nolen_const( mysv )
3881 Step through the branches
3882 cur represents each branch,
3883 noper is the first thing to be matched as part
3885 noper_next is the regnext() of that node.
3887 We normally handle a case like this
3888 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3889 support building with NOJUMPTRIE, which restricts
3890 the trie logic to structures like /FOO|BAR/.
3892 If noper is a trieable nodetype then the branch is
3893 a possible optimization target. If we are building
3894 under NOJUMPTRIE then we require that noper_next is
3895 the same as scan (our current position in the regex
3898 Once we have two or more consecutive such branches
3899 we can create a trie of the EXACT's contents and
3900 stitch it in place into the program.
3902 If the sequence represents all of the branches in
3903 the alternation we replace the entire thing with a
3906 Otherwise when it is a subsequence we need to
3907 stitch it in place and replace only the relevant
3908 branches. This means the first branch has to remain
3909 as it is used by the alternation logic, and its
3910 next pointer, and needs to be repointed at the item
3911 on the branch chain following the last branch we
3912 have optimized away.
3914 This could be either a BRANCH, in which case the
3915 subsequence is internal, or it could be the item
3916 following the branch sequence in which case the
3917 subsequence is at the end (which does not
3918 necessarily mean the first node is the start of the
3921 TRIE_TYPE(X) is a define which maps the optype to a
3925 ----------------+-----------
3929 EXACTFU_SS | EXACTFU
3934 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3935 ( EXACT == (X) ) ? EXACT : \
3936 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3937 ( EXACTFA == (X) ) ? EXACTFA : \
3940 /* dont use tail as the end marker for this traverse */
3941 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3942 regnode * const noper = NEXTOPER( cur );
3943 U8 noper_type = OP( noper );
3944 U8 noper_trietype = TRIE_TYPE( noper_type );
3945 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3946 regnode * const noper_next = regnext( noper );
3947 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3948 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3951 DEBUG_TRIE_COMPILE_r({
3952 regprop(RExC_rx, mysv, cur, NULL);
3953 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3954 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3956 regprop(RExC_rx, mysv, noper, NULL);
3957 PerlIO_printf( Perl_debug_log, " -> %s",
3958 SvPV_nolen_const(mysv));
3961 regprop(RExC_rx, mysv, noper_next, NULL);
3962 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3963 SvPV_nolen_const(mysv));
3965 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3966 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3967 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3971 /* Is noper a trieable nodetype that can be merged
3972 * with the current trie (if there is one)? */
3976 ( noper_trietype == NOTHING)
3977 || ( trietype == NOTHING )
3978 || ( trietype == noper_trietype )
3981 && noper_next == tail
3985 /* Handle mergable triable node Either we are
3986 * the first node in a new trieable sequence,
3987 * in which case we do some bookkeeping,
3988 * otherwise we update the end pointer. */
3991 if ( noper_trietype == NOTHING ) {
3992 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3993 regnode * const noper_next = regnext( noper );
3994 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3995 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3998 if ( noper_next_trietype ) {
3999 trietype = noper_next_trietype;
4000 } else if (noper_next_type) {
4001 /* a NOTHING regop is 1 regop wide.
4002 * We need at least two for a trie
4003 * so we can't merge this in */
4007 trietype = noper_trietype;
4010 if ( trietype == NOTHING )
4011 trietype = noper_trietype;
4016 } /* end handle mergable triable node */
4018 /* handle unmergable node -
4019 * noper may either be a triable node which can
4020 * not be tried together with the current trie,
4021 * or a non triable node */
4023 /* If last is set and trietype is not
4024 * NOTHING then we have found at least two
4025 * triable branch sequences in a row of a
4026 * similar trietype so we can turn them
4027 * into a trie. If/when we allow NOTHING to
4028 * start a trie sequence this condition
4029 * will be required, and it isn't expensive
4030 * so we leave it in for now. */
4031 if ( trietype && trietype != NOTHING )
4032 make_trie( pRExC_state,
4033 startbranch, first, cur, tail,
4034 count, trietype, depth+1 );
4035 last = NULL; /* note: we clear/update
4036 first, trietype etc below,
4037 so we dont do it here */
4041 && noper_next == tail
4044 /* noper is triable, so we can start a new
4048 trietype = noper_trietype;
4050 /* if we already saw a first but the
4051 * current node is not triable then we have
4052 * to reset the first information. */
4057 } /* end handle unmergable node */
4058 } /* loop over branches */
4059 DEBUG_TRIE_COMPILE_r({
4060 regprop(RExC_rx, mysv, cur, NULL);
4061 PerlIO_printf( Perl_debug_log,
4062 "%*s- %s (%d) <SCAN FINISHED>\n",
4064 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4067 if ( last && trietype ) {
4068 if ( trietype != NOTHING ) {
4069 /* the last branch of the sequence was part of
4070 * a trie, so we have to construct it here
4071 * outside of the loop */
4072 made= make_trie( pRExC_state, startbranch,
4073 first, scan, tail, count,
4074 trietype, depth+1 );
4075 #ifdef TRIE_STUDY_OPT
4076 if ( ((made == MADE_EXACT_TRIE &&
4077 startbranch == first)
4078 || ( first_non_open == first )) &&
4080 flags |= SCF_TRIE_RESTUDY;
4081 if ( startbranch == first
4084 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4089 /* at this point we know whatever we have is a
4090 * NOTHING sequence/branch AND if 'startbranch'
4091 * is 'first' then we can turn the whole thing
4094 if ( startbranch == first ) {
4096 /* the entire thing is a NOTHING sequence,
4097 * something like this: (?:|) So we can
4098 * turn it into a plain NOTHING op. */
4099 DEBUG_TRIE_COMPILE_r({
4100 regprop(RExC_rx, mysv, cur, NULL);
4101 PerlIO_printf( Perl_debug_log,
4102 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4103 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4106 OP(startbranch)= NOTHING;
4107 NEXT_OFF(startbranch)= tail - startbranch;
4108 for ( opt= startbranch + 1; opt < tail ; opt++ )
4112 } /* end if ( last) */
4113 } /* TRIE_MAXBUF is non zero */
4118 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4119 scan = NEXTOPER(NEXTOPER(scan));
4120 } else /* single branch is optimized. */
4121 scan = NEXTOPER(scan);
4123 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4124 scan_frame *newframe = NULL;
4128 U32 my_recursed_depth= recursed_depth;
4130 if (OP(scan) != SUSPEND) {
4131 /* set the pointer */
4132 if (OP(scan) == GOSUB) {
4134 RExC_recurse[ARG2L(scan)] = scan;
4135 start = RExC_open_parens[paren-1];
4136 end = RExC_close_parens[paren-1];
4139 start = RExC_rxi->program + 1;
4144 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4146 if (!recursed_depth) {
4147 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4149 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4150 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4151 RExC_study_chunk_recursed_bytes, U8);
4153 /* we havent recursed into this paren yet, so recurse into it */
4154 DEBUG_STUDYDATA("set:", data,depth);
4155 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4156 my_recursed_depth= recursed_depth + 1;
4157 Newx(newframe,1,scan_frame);
4159 DEBUG_STUDYDATA("inf:", data,depth);
4160 /* some form of infinite recursion, assume infinite length
4162 if (flags & SCF_DO_SUBSTR) {
4163 scan_commit(pRExC_state, data, minlenp, is_inf);
4164 data->longest = &(data->longest_float);
4166 is_inf = is_inf_internal = 1;
4167 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4168 ssc_anything(data->start_class);
4169 flags &= ~SCF_DO_STCLASS;
4172 Newx(newframe,1,scan_frame);
4175 end = regnext(scan);
4180 SAVEFREEPV(newframe);
4181 newframe->next = regnext(scan);
4182 newframe->last = last;
4183 newframe->stop = stopparen;
4184 newframe->prev = frame;
4185 newframe->prev_recursed_depth = recursed_depth;
4187 DEBUG_STUDYDATA("frame-new:",data,depth);
4188 DEBUG_PEEP("fnew", scan, depth);
4195 recursed_depth= my_recursed_depth;
4200 else if (OP(scan) == EXACT) {
4201 SSize_t l = STR_LEN(scan);
4204 const U8 * const s = (U8*)STRING(scan);
4205 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4206 l = utf8_length(s, s + l);
4208 uc = *((U8*)STRING(scan));
4211 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4212 /* The code below prefers earlier match for fixed
4213 offset, later match for variable offset. */
4214 if (data->last_end == -1) { /* Update the start info. */
4215 data->last_start_min = data->pos_min;
4216 data->last_start_max = is_inf
4217 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4219 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4221 SvUTF8_on(data->last_found);
4223 SV * const sv = data->last_found;
4224 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4225 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4226 if (mg && mg->mg_len >= 0)
4227 mg->mg_len += utf8_length((U8*)STRING(scan),
4228 (U8*)STRING(scan)+STR_LEN(scan));
4230 data->last_end = data->pos_min + l;
4231 data->pos_min += l; /* As in the first entry. */
4232 data->flags &= ~SF_BEFORE_EOL;
4235 /* ANDing the code point leaves at most it, and not in locale, and
4236 * can't match null string */
4237 if (flags & SCF_DO_STCLASS_AND) {
4238 ssc_cp_and(data->start_class, uc);
4239 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4240 ssc_clear_locale(data->start_class);
4242 else if (flags & SCF_DO_STCLASS_OR) {
4243 ssc_add_cp(data->start_class, uc);
4244 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4246 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4247 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4249 flags &= ~SCF_DO_STCLASS;
4251 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
4252 SSize_t l = STR_LEN(scan);
4253 UV uc = *((U8*)STRING(scan));
4254 SV* EXACTF_invlist = _new_invlist(4); /* Start out big enough for 2
4255 separate code points */
4257 /* Search for fixed substrings supports EXACT only. */
4258 if (flags & SCF_DO_SUBSTR) {
4260 scan_commit(pRExC_state, data, minlenp, is_inf);
4263 const U8 * const s = (U8 *)STRING(scan);
4264 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4265 l = utf8_length(s, s + l);
4267 if (unfolded_multi_char) {
4268 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4270 min += l - min_subtract;
4272 delta += min_subtract;
4273 if (flags & SCF_DO_SUBSTR) {
4274 data->pos_min += l - min_subtract;
4275 if (data->pos_min < 0) {
4278 data->pos_delta += min_subtract;
4280 data->longest = &(data->longest_float);
4283 if (OP(scan) == EXACTFL) {
4285 /* We don't know what the folds are; it could be anything. XXX
4286 * Actually, we only support UTF-8 encoding for code points
4287 * above Latin1, so we could know what those folds are. */
4288 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4292 else { /* Non-locale EXACTFish */
4293 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4294 if (flags & SCF_DO_STCLASS_AND) {
4295 ssc_clear_locale(data->start_class);
4297 if (uc < 256) { /* We know what the Latin1 folds are ... */
4298 if (IS_IN_SOME_FOLD_L1(uc)) { /* For instance, we
4299 know if anything folds
4301 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist,
4302 PL_fold_latin1[uc]);
4303 if (OP(scan) != EXACTFA) { /* The folds below aren't
4305 if (isARG2_lower_or_UPPER_ARG1('s', uc)) {
4307 = add_cp_to_invlist(EXACTF_invlist,
4308 LATIN_SMALL_LETTER_SHARP_S);
4310 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
4312 = add_cp_to_invlist(EXACTF_invlist, 's');
4314 = add_cp_to_invlist(EXACTF_invlist, 'S');
4318 /* We also know if there are above-Latin1 code points
4319 * that fold to this (none legal for ASCII and /iaa) */
4320 if ((! isASCII(uc) || OP(scan) != EXACTFA)
4321 && HAS_NONLATIN1_FOLD_CLOSURE(uc))
4323 /* XXX We could know exactly what does fold to this
4324 * if the reverse folds are loaded, as currently in
4326 _invlist_union(EXACTF_invlist,
4332 else { /* Non-locale, above Latin1. XXX We don't currently
4333 know what participates in folds with this, so have
4334 to assume anything could */
4336 /* XXX We could know exactly what does fold to this if the
4337 * reverse folds are loaded, as currently in S_regclass().
4338 * But we do know that under /iaa nothing in the ASCII
4339 * range can participate */
4340 if (OP(scan) == EXACTFA) {
4341 _invlist_union_complement_2nd(EXACTF_invlist,
4342 PL_XPosix_ptrs[_CC_ASCII],
4346 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4351 if (flags & SCF_DO_STCLASS_AND) {
4352 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4353 ANYOF_POSIXL_ZERO(data->start_class);
4354 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4356 else if (flags & SCF_DO_STCLASS_OR) {
4357 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4358 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4360 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4361 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4363 flags &= ~SCF_DO_STCLASS;
4364 SvREFCNT_dec(EXACTF_invlist);
4366 else if (REGNODE_VARIES(OP(scan))) {
4367 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4368 I32 fl = 0, f = flags;
4369 regnode * const oscan = scan;
4370 regnode_ssc this_class;
4371 regnode_ssc *oclass = NULL;
4372 I32 next_is_eval = 0;
4374 switch (PL_regkind[OP(scan)]) {
4375 case WHILEM: /* End of (?:...)* . */
4376 scan = NEXTOPER(scan);
4379 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4380 next = NEXTOPER(scan);
4381 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
4383 maxcount = REG_INFTY;
4384 next = regnext(scan);
4385 scan = NEXTOPER(scan);
4389 if (flags & SCF_DO_SUBSTR)
4394 if (flags & SCF_DO_STCLASS) {
4396 maxcount = REG_INFTY;
4397 next = regnext(scan);
4398 scan = NEXTOPER(scan);
4401 if (flags & SCF_DO_SUBSTR) {
4402 scan_commit(pRExC_state, data, minlenp, is_inf);
4403 /* Cannot extend fixed substrings */
4404 data->longest = &(data->longest_float);
4406 is_inf = is_inf_internal = 1;
4407 scan = regnext(scan);
4408 goto optimize_curly_tail;
4410 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4411 && (scan->flags == stopparen))
4416 mincount = ARG1(scan);
4417 maxcount = ARG2(scan);
4419 next = regnext(scan);
4420 if (OP(scan) == CURLYX) {
4421 I32 lp = (data ? *(data->last_closep) : 0);
4422 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4424 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4425 next_is_eval = (OP(scan) == EVAL);
4427 if (flags & SCF_DO_SUBSTR) {
4429 scan_commit(pRExC_state, data, minlenp, is_inf);
4430 /* Cannot extend fixed substrings */
4431 pos_before = data->pos_min;
4435 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4437 data->flags |= SF_IS_INF;
4439 if (flags & SCF_DO_STCLASS) {
4440 ssc_init(pRExC_state, &this_class);
4441 oclass = data->start_class;
4442 data->start_class = &this_class;
4443 f |= SCF_DO_STCLASS_AND;
4444 f &= ~SCF_DO_STCLASS_OR;
4446 /* Exclude from super-linear cache processing any {n,m}
4447 regops for which the combination of input pos and regex
4448 pos is not enough information to determine if a match
4451 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4452 regex pos at the \s*, the prospects for a match depend not
4453 only on the input position but also on how many (bar\s*)
4454 repeats into the {4,8} we are. */
4455 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4456 f &= ~SCF_WHILEM_VISITED_POS;
4458 /* This will finish on WHILEM, setting scan, or on NULL: */
4459 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4460 last, data, stopparen, recursed_depth, NULL,
4462 ? (f & ~SCF_DO_SUBSTR)
4466 if (flags & SCF_DO_STCLASS)
4467 data->start_class = oclass;
4468 if (mincount == 0 || minnext == 0) {
4469 if (flags & SCF_DO_STCLASS_OR) {
4470 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4472 else if (flags & SCF_DO_STCLASS_AND) {
4473 /* Switch to OR mode: cache the old value of
4474 * data->start_class */
4476 StructCopy(data->start_class, and_withp, regnode_ssc);
4477 flags &= ~SCF_DO_STCLASS_AND;
4478 StructCopy(&this_class, data->start_class, regnode_ssc);
4479 flags |= SCF_DO_STCLASS_OR;
4480 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
4482 } else { /* Non-zero len */
4483 if (flags & SCF_DO_STCLASS_OR) {
4484 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4485 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4487 else if (flags & SCF_DO_STCLASS_AND)
4488 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4489 flags &= ~SCF_DO_STCLASS;
4491 if (!scan) /* It was not CURLYX, but CURLY. */
4493 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4494 /* ? quantifier ok, except for (?{ ... }) */
4495 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4496 && (minnext == 0) && (deltanext == 0)
4497 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4498 && maxcount <= REG_INFTY/3) /* Complement check for big
4501 /* Fatal warnings may leak the regexp without this: */
4502 SAVEFREESV(RExC_rx_sv);
4503 ckWARNreg(RExC_parse,
4504 "Quantifier unexpected on zero-length expression");
4505 (void)ReREFCNT_inc(RExC_rx_sv);
4508 min += minnext * mincount;
4509 is_inf_internal |= deltanext == SSize_t_MAX
4510 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4511 is_inf |= is_inf_internal;
4513 delta = SSize_t_MAX;
4515 delta += (minnext + deltanext) * maxcount
4516 - minnext * mincount;
4518 /* Try powerful optimization CURLYX => CURLYN. */
4519 if ( OP(oscan) == CURLYX && data
4520 && data->flags & SF_IN_PAR
4521 && !(data->flags & SF_HAS_EVAL)
4522 && !deltanext && minnext == 1 ) {
4523 /* Try to optimize to CURLYN. */
4524 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4525 regnode * const nxt1 = nxt;
4532 if (!REGNODE_SIMPLE(OP(nxt))
4533 && !(PL_regkind[OP(nxt)] == EXACT
4534 && STR_LEN(nxt) == 1))
4540 if (OP(nxt) != CLOSE)
4542 if (RExC_open_parens) {
4543 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4544 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4546 /* Now we know that nxt2 is the only contents: */
4547 oscan->flags = (U8)ARG(nxt);
4549 OP(nxt1) = NOTHING; /* was OPEN. */
4552 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4553 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4554 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4555 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4556 OP(nxt + 1) = OPTIMIZED; /* was count. */
4557 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4562 /* Try optimization CURLYX => CURLYM. */
4563 if ( OP(oscan) == CURLYX && data
4564 && !(data->flags & SF_HAS_PAR)
4565 && !(data->flags & SF_HAS_EVAL)
4566 && !deltanext /* atom is fixed width */
4567 && minnext != 0 /* CURLYM can't handle zero width */
4569 /* Nor characters whose fold at run-time may be
4570 * multi-character */
4571 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4573 /* XXXX How to optimize if data == 0? */
4574 /* Optimize to a simpler form. */
4575 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4579 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4580 && (OP(nxt2) != WHILEM))
4582 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4583 /* Need to optimize away parenths. */
4584 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4585 /* Set the parenth number. */
4586 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4588 oscan->flags = (U8)ARG(nxt);
4589 if (RExC_open_parens) {
4590 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4591 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4593 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4594 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4597 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4598 OP(nxt + 1) = OPTIMIZED; /* was count. */
4599 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4600 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4603 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4604 regnode *nnxt = regnext(nxt1);
4606 if (reg_off_by_arg[OP(nxt1)])
4607 ARG_SET(nxt1, nxt2 - nxt1);
4608 else if (nxt2 - nxt1 < U16_MAX)
4609 NEXT_OFF(nxt1) = nxt2 - nxt1;
4611 OP(nxt) = NOTHING; /* Cannot beautify */
4616 /* Optimize again: */
4617 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4618 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4623 else if ((OP(oscan) == CURLYX)
4624 && (flags & SCF_WHILEM_VISITED_POS)
4625 /* See the comment on a similar expression above.
4626 However, this time it's not a subexpression
4627 we care about, but the expression itself. */
4628 && (maxcount == REG_INFTY)
4629 && data && ++data->whilem_c < 16) {
4630 /* This stays as CURLYX, we can put the count/of pair. */
4631 /* Find WHILEM (as in regexec.c) */
4632 regnode *nxt = oscan + NEXT_OFF(oscan);
4634 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4636 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4637 | (RExC_whilem_seen << 4)); /* On WHILEM */
4639 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4641 if (flags & SCF_DO_SUBSTR) {
4642 SV *last_str = NULL;
4643 STRLEN last_chrs = 0;
4644 int counted = mincount != 0;
4646 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4648 SSize_t b = pos_before >= data->last_start_min
4649 ? pos_before : data->last_start_min;
4651 const char * const s = SvPV_const(data->last_found, l);
4652 SSize_t old = b - data->last_start_min;
4655 old = utf8_hop((U8*)s, old) - (U8*)s;
4657 /* Get the added string: */
4658 last_str = newSVpvn_utf8(s + old, l, UTF);
4659 last_chrs = UTF ? utf8_length((U8*)(s + old),
4660 (U8*)(s + old + l)) : l;
4661 if (deltanext == 0 && pos_before == b) {
4662 /* What was added is a constant string */
4665 SvGROW(last_str, (mincount * l) + 1);
4666 repeatcpy(SvPVX(last_str) + l,
4667 SvPVX_const(last_str), l,
4669 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4670 /* Add additional parts. */
4671 SvCUR_set(data->last_found,
4672 SvCUR(data->last_found) - l);
4673 sv_catsv(data->last_found, last_str);
4675 SV * sv = data->last_found;
4677 SvUTF8(sv) && SvMAGICAL(sv) ?
4678 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4679 if (mg && mg->mg_len >= 0)
4680 mg->mg_len += last_chrs * (mincount-1);
4682 last_chrs *= mincount;
4683 data->last_end += l * (mincount - 1);
4686 /* start offset must point into the last copy */
4687 data->last_start_min += minnext * (mincount - 1);
4688 data->last_start_max += is_inf ? SSize_t_MAX
4689 : (maxcount - 1) * (minnext + data->pos_delta);
4692 /* It is counted once already... */
4693 data->pos_min += minnext * (mincount - counted);
4695 PerlIO_printf(Perl_debug_log, "counted=%"UVdf" deltanext=%"UVdf
4696 " SSize_t_MAX=%"UVdf" minnext=%"UVdf
4697 " maxcount=%"UVdf" mincount=%"UVdf"\n",
4698 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4700 if (deltanext != SSize_t_MAX)
4701 PerlIO_printf(Perl_debug_log, "LHS=%"UVdf" RHS=%"UVdf"\n",
4702 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4703 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4705 if (deltanext == SSize_t_MAX
4706 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4707 data->pos_delta = SSize_t_MAX;
4709 data->pos_delta += - counted * deltanext +
4710 (minnext + deltanext) * maxcount - minnext * mincount;
4711 if (mincount != maxcount) {
4712 /* Cannot extend fixed substrings found inside
4714 scan_commit(pRExC_state, data, minlenp, is_inf);
4715 if (mincount && last_str) {
4716 SV * const sv = data->last_found;
4717 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4718 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4722 sv_setsv(sv, last_str);
4723 data->last_end = data->pos_min;
4724 data->last_start_min = data->pos_min - last_chrs;
4725 data->last_start_max = is_inf
4727 : data->pos_min + data->pos_delta - last_chrs;
4729 data->longest = &(data->longest_float);
4731 SvREFCNT_dec(last_str);
4733 if (data && (fl & SF_HAS_EVAL))
4734 data->flags |= SF_HAS_EVAL;
4735 optimize_curly_tail:
4736 if (OP(oscan) != CURLYX) {
4737 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4739 NEXT_OFF(oscan) += NEXT_OFF(next);
4745 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4750 if (flags & SCF_DO_SUBSTR) {
4751 /* Cannot expect anything... */
4752 scan_commit(pRExC_state, data, minlenp, is_inf);
4753 data->longest = &(data->longest_float);
4755 is_inf = is_inf_internal = 1;
4756 if (flags & SCF_DO_STCLASS_OR) {
4757 if (OP(scan) == CLUMP) {
4758 /* Actually is any start char, but very few code points
4759 * aren't start characters */
4760 ssc_match_all_cp(data->start_class);
4763 ssc_anything(data->start_class);
4766 flags &= ~SCF_DO_STCLASS;
4770 else if (OP(scan) == LNBREAK) {
4771 if (flags & SCF_DO_STCLASS) {
4772 if (flags & SCF_DO_STCLASS_AND) {
4773 ssc_intersection(data->start_class,
4774 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
4775 ssc_clear_locale(data->start_class);
4776 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4778 else if (flags & SCF_DO_STCLASS_OR) {
4779 ssc_union(data->start_class,
4780 PL_XPosix_ptrs[_CC_VERTSPACE],
4782 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4784 /* See commit msg for
4785 * 749e076fceedeb708a624933726e7989f2302f6a */
4786 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4788 flags &= ~SCF_DO_STCLASS;
4791 delta++; /* Because of the 2 char string cr-lf */
4792 if (flags & SCF_DO_SUBSTR) {
4793 /* Cannot expect anything... */
4794 scan_commit(pRExC_state, data, minlenp, is_inf);
4796 data->pos_delta += 1;
4797 data->longest = &(data->longest_float);
4800 else if (REGNODE_SIMPLE(OP(scan))) {
4802 if (flags & SCF_DO_SUBSTR) {
4803 scan_commit(pRExC_state, data, minlenp, is_inf);
4807 if (flags & SCF_DO_STCLASS) {
4809 SV* my_invlist = NULL;
4812 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4813 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4815 /* Some of the logic below assumes that switching
4816 locale on will only add false positives. */
4821 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
4826 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4827 ssc_match_all_cp(data->start_class);
4832 SV* REG_ANY_invlist = _new_invlist(2);
4833 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
4835 if (flags & SCF_DO_STCLASS_OR) {
4836 ssc_union(data->start_class,
4838 TRUE /* TRUE => invert, hence all but \n
4842 else if (flags & SCF_DO_STCLASS_AND) {
4843 ssc_intersection(data->start_class,
4845 TRUE /* TRUE => invert */
4847 ssc_clear_locale(data->start_class);
4849 SvREFCNT_dec_NN(REG_ANY_invlist);
4854 if (flags & SCF_DO_STCLASS_AND)
4855 ssc_and(pRExC_state, data->start_class,
4856 (regnode_charclass *) scan);
4858 ssc_or(pRExC_state, data->start_class,
4859 (regnode_charclass *) scan);
4867 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
4868 if (flags & SCF_DO_STCLASS_AND) {
4869 bool was_there = cBOOL(
4870 ANYOF_POSIXL_TEST(data->start_class,
4872 ANYOF_POSIXL_ZERO(data->start_class);
4873 if (was_there) { /* Do an AND */
4874 ANYOF_POSIXL_SET(data->start_class, namedclass);
4876 /* No individual code points can now match */
4877 data->start_class->invlist
4878 = sv_2mortal(_new_invlist(0));
4881 int complement = namedclass + ((invert) ? -1 : 1);
4883 assert(flags & SCF_DO_STCLASS_OR);
4885 /* If the complement of this class was already there,
4886 * the result is that they match all code points,
4887 * (\d + \D == everything). Remove the classes from
4888 * future consideration. Locale is not relevant in
4890 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
4891 ssc_match_all_cp(data->start_class);
4892 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
4893 ANYOF_POSIXL_CLEAR(data->start_class, complement);
4895 else { /* The usual case; just add this class to the
4897 ANYOF_POSIXL_SET(data->start_class, namedclass);
4902 case NPOSIXA: /* For these, we always know the exact set of
4907 if (FLAGS(scan) == _CC_ASCII) {
4908 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
4911 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
4912 PL_XPosix_ptrs[_CC_ASCII],
4923 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
4925 /* NPOSIXD matches all upper Latin1 code points unless the
4926 * target string being matched is UTF-8, which is
4927 * unknowable until match time. Since we are going to
4928 * invert, we want to get rid of all of them so that the
4929 * inversion will match all */
4930 if (OP(scan) == NPOSIXD) {
4931 _invlist_subtract(my_invlist, PL_UpperLatin1,
4937 if (flags & SCF_DO_STCLASS_AND) {
4938 ssc_intersection(data->start_class, my_invlist, invert);
4939 ssc_clear_locale(data->start_class);
4942 assert(flags & SCF_DO_STCLASS_OR);
4943 ssc_union(data->start_class, my_invlist, invert);
4945 SvREFCNT_dec(my_invlist);
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. These
5051 * assertions are zero-length, so can match an EMPTY
5053 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5054 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
5058 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5060 /* Positive Lookahead/lookbehind
5061 In this case we can do fixed string optimisation,
5062 but we must be careful about it. Note in the case of
5063 lookbehind the positions will be offset by the minimum
5064 length of the pattern, something we won't know about
5065 until after the recurse.
5067 SSize_t deltanext, fake = 0;
5071 /* We use SAVEFREEPV so that when the full compile
5072 is finished perl will clean up the allocated
5073 minlens when it's all done. This way we don't
5074 have to worry about freeing them when we know
5075 they wont be used, which would be a pain.
5078 Newx( minnextp, 1, SSize_t );
5079 SAVEFREEPV(minnextp);
5082 StructCopy(data, &data_fake, scan_data_t);
5083 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5086 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5087 data_fake.last_found=newSVsv(data->last_found);
5091 data_fake.last_closep = &fake;
5092 data_fake.flags = 0;
5093 data_fake.pos_delta = delta;
5095 data_fake.flags |= SF_IS_INF;
5096 if ( flags & SCF_DO_STCLASS && !scan->flags
5097 && OP(scan) == IFMATCH ) { /* Lookahead */
5098 ssc_init(pRExC_state, &intrnl);
5099 data_fake.start_class = &intrnl;
5100 f |= SCF_DO_STCLASS_AND;
5102 if (flags & SCF_WHILEM_VISITED_POS)
5103 f |= SCF_WHILEM_VISITED_POS;
5104 next = regnext(scan);
5105 nscan = NEXTOPER(NEXTOPER(scan));
5107 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5108 &deltanext, last, &data_fake,
5109 stopparen, recursed_depth, NULL,
5113 FAIL("Variable length lookbehind not implemented");
5115 else if (*minnextp > (I32)U8_MAX) {
5116 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5119 scan->flags = (U8)*minnextp;
5124 if (f & SCF_DO_STCLASS_AND) {
5125 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5126 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
5129 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5131 if (data_fake.flags & SF_HAS_EVAL)
5132 data->flags |= SF_HAS_EVAL;
5133 data->whilem_c = data_fake.whilem_c;
5134 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5135 if (RExC_rx->minlen<*minnextp)
5136 RExC_rx->minlen=*minnextp;
5137 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5138 SvREFCNT_dec_NN(data_fake.last_found);
5140 if ( data_fake.minlen_fixed != minlenp )
5142 data->offset_fixed= data_fake.offset_fixed;
5143 data->minlen_fixed= data_fake.minlen_fixed;
5144 data->lookbehind_fixed+= scan->flags;
5146 if ( data_fake.minlen_float != minlenp )
5148 data->minlen_float= data_fake.minlen_float;
5149 data->offset_float_min=data_fake.offset_float_min;
5150 data->offset_float_max=data_fake.offset_float_max;
5151 data->lookbehind_float+= scan->flags;
5158 else if (OP(scan) == OPEN) {
5159 if (stopparen != (I32)ARG(scan))
5162 else if (OP(scan) == CLOSE) {
5163 if (stopparen == (I32)ARG(scan)) {
5166 if ((I32)ARG(scan) == is_par) {
5167 next = regnext(scan);
5169 if ( next && (OP(next) != WHILEM) && next < last)
5170 is_par = 0; /* Disable optimization */
5173 *(data->last_closep) = ARG(scan);
5175 else if (OP(scan) == EVAL) {
5177 data->flags |= SF_HAS_EVAL;
5179 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5180 if (flags & SCF_DO_SUBSTR) {
5181 scan_commit(pRExC_state, data, minlenp, is_inf);
5182 flags &= ~SCF_DO_SUBSTR;
5184 if (data && OP(scan)==ACCEPT) {
5185 data->flags |= SCF_SEEN_ACCEPT;
5190 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5192 if (flags & SCF_DO_SUBSTR) {
5193 scan_commit(pRExC_state, data, minlenp, is_inf);
5194 data->longest = &(data->longest_float);
5196 is_inf = is_inf_internal = 1;
5197 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5198 ssc_anything(data->start_class);
5199 flags &= ~SCF_DO_STCLASS;
5201 else if (OP(scan) == GPOS) {
5202 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5203 !(delta || is_inf || (data && data->pos_delta)))
5205 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5206 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5207 if (RExC_rx->gofs < (STRLEN)min)
5208 RExC_rx->gofs = min;
5210 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5214 #ifdef TRIE_STUDY_OPT
5215 #ifdef FULL_TRIE_STUDY
5216 else if (PL_regkind[OP(scan)] == TRIE) {
5217 /* NOTE - There is similar code to this block above for handling
5218 BRANCH nodes on the initial study. If you change stuff here
5220 regnode *trie_node= scan;
5221 regnode *tail= regnext(scan);
5222 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5223 SSize_t max1 = 0, min1 = SSize_t_MAX;
5226 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5227 /* Cannot merge strings after this. */
5228 scan_commit(pRExC_state, data, minlenp, is_inf);
5230 if (flags & SCF_DO_STCLASS)
5231 ssc_init_zero(pRExC_state, &accum);
5237 const regnode *nextbranch= NULL;
5240 for ( word=1 ; word <= trie->wordcount ; word++)
5242 SSize_t deltanext=0, minnext=0, f = 0, fake;
5243 regnode_ssc this_class;
5245 data_fake.flags = 0;
5247 data_fake.whilem_c = data->whilem_c;
5248 data_fake.last_closep = data->last_closep;
5251 data_fake.last_closep = &fake;
5252 data_fake.pos_delta = delta;
5253 if (flags & SCF_DO_STCLASS) {
5254 ssc_init(pRExC_state, &this_class);
5255 data_fake.start_class = &this_class;
5256 f = SCF_DO_STCLASS_AND;
5258 if (flags & SCF_WHILEM_VISITED_POS)
5259 f |= SCF_WHILEM_VISITED_POS;
5261 if (trie->jump[word]) {
5263 nextbranch = trie_node + trie->jump[0];
5264 scan= trie_node + trie->jump[word];
5265 /* We go from the jump point to the branch that follows
5266 it. Note this means we need the vestigal unused
5267 branches even though they arent otherwise used. */
5268 minnext = study_chunk(pRExC_state, &scan, minlenp,
5269 &deltanext, (regnode *)nextbranch, &data_fake,
5270 stopparen, recursed_depth, NULL, f,depth+1);
5272 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5273 nextbranch= regnext((regnode*)nextbranch);
5275 if (min1 > (SSize_t)(minnext + trie->minlen))
5276 min1 = minnext + trie->minlen;
5277 if (deltanext == SSize_t_MAX) {
5278 is_inf = is_inf_internal = 1;
5280 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5281 max1 = minnext + deltanext + trie->maxlen;
5283 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5285 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5286 if ( stopmin > min + min1)
5287 stopmin = min + min1;
5288 flags &= ~SCF_DO_SUBSTR;
5290 data->flags |= SCF_SEEN_ACCEPT;
5293 if (data_fake.flags & SF_HAS_EVAL)
5294 data->flags |= SF_HAS_EVAL;
5295 data->whilem_c = data_fake.whilem_c;
5297 if (flags & SCF_DO_STCLASS)
5298 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5301 if (flags & SCF_DO_SUBSTR) {
5302 data->pos_min += min1;
5303 data->pos_delta += max1 - min1;
5304 if (max1 != min1 || is_inf)
5305 data->longest = &(data->longest_float);
5308 delta += max1 - min1;
5309 if (flags & SCF_DO_STCLASS_OR) {
5310 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5312 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5313 flags &= ~SCF_DO_STCLASS;
5316 else if (flags & SCF_DO_STCLASS_AND) {
5318 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5319 flags &= ~SCF_DO_STCLASS;
5322 /* Switch to OR mode: cache the old value of
5323 * data->start_class */
5325 StructCopy(data->start_class, and_withp, regnode_ssc);
5326 flags &= ~SCF_DO_STCLASS_AND;
5327 StructCopy(&accum, data->start_class, regnode_ssc);
5328 flags |= SCF_DO_STCLASS_OR;
5335 else if (PL_regkind[OP(scan)] == TRIE) {
5336 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5339 min += trie->minlen;
5340 delta += (trie->maxlen - trie->minlen);
5341 flags &= ~SCF_DO_STCLASS; /* xxx */
5342 if (flags & SCF_DO_SUBSTR) {
5343 /* Cannot expect anything... */
5344 scan_commit(pRExC_state, data, minlenp, is_inf);
5345 data->pos_min += trie->minlen;
5346 data->pos_delta += (trie->maxlen - trie->minlen);
5347 if (trie->maxlen != trie->minlen)
5348 data->longest = &(data->longest_float);
5350 if (trie->jump) /* no more substrings -- for now /grr*/
5351 flags &= ~SCF_DO_SUBSTR;
5353 #endif /* old or new */
5354 #endif /* TRIE_STUDY_OPT */
5356 /* Else: zero-length, ignore. */
5357 scan = regnext(scan);
5359 /* If we are exiting a recursion we can unset its recursed bit
5360 * and allow ourselves to enter it again - no danger of an
5361 * infinite loop there.
5362 if (stopparen > -1 && recursed) {
5363 DEBUG_STUDYDATA("unset:", data,depth);
5364 PAREN_UNSET( recursed, stopparen);
5368 DEBUG_STUDYDATA("frame-end:",data,depth);
5369 DEBUG_PEEP("fend", scan, depth);
5370 /* restore previous context */
5373 stopparen = frame->stop;
5374 recursed_depth = frame->prev_recursed_depth;
5377 frame = frame->prev;
5378 goto fake_study_recurse;
5383 DEBUG_STUDYDATA("pre-fin:",data,depth);
5386 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5388 if (flags & SCF_DO_SUBSTR && is_inf)
5389 data->pos_delta = SSize_t_MAX - data->pos_min;
5390 if (is_par > (I32)U8_MAX)
5392 if (is_par && pars==1 && data) {
5393 data->flags |= SF_IN_PAR;
5394 data->flags &= ~SF_HAS_PAR;
5396 else if (pars && data) {
5397 data->flags |= SF_HAS_PAR;
5398 data->flags &= ~SF_IN_PAR;
5400 if (flags & SCF_DO_STCLASS_OR)
5401 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5402 if (flags & SCF_TRIE_RESTUDY)
5403 data->flags |= SCF_TRIE_RESTUDY;
5405 DEBUG_STUDYDATA("post-fin:",data,depth);
5408 SSize_t final_minlen= min < stopmin ? min : stopmin;
5410 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) && (RExC_maxlen < final_minlen + delta)) {
5411 RExC_maxlen = final_minlen + delta;
5413 return final_minlen;
5419 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5421 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5423 PERL_ARGS_ASSERT_ADD_DATA;
5425 Renewc(RExC_rxi->data,
5426 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5427 char, struct reg_data);
5429 Renew(RExC_rxi->data->what, count + n, U8);
5431 Newx(RExC_rxi->data->what, n, U8);
5432 RExC_rxi->data->count = count + n;
5433 Copy(s, RExC_rxi->data->what + count, n, U8);
5437 /*XXX: todo make this not included in a non debugging perl */
5438 #ifndef PERL_IN_XSUB_RE
5440 Perl_reginitcolors(pTHX)
5443 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5445 char *t = savepv(s);
5449 t = strchr(t, '\t');
5455 PL_colors[i] = t = (char *)"";
5460 PL_colors[i++] = (char *)"";
5467 #ifdef TRIE_STUDY_OPT
5468 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5471 (data.flags & SCF_TRIE_RESTUDY) \
5479 #define CHECK_RESTUDY_GOTO_butfirst
5483 * pregcomp - compile a regular expression into internal code
5485 * Decides which engine's compiler to call based on the hint currently in
5489 #ifndef PERL_IN_XSUB_RE
5491 /* return the currently in-scope regex engine (or the default if none) */
5493 regexp_engine const *
5494 Perl_current_re_engine(pTHX)
5498 if (IN_PERL_COMPILETIME) {
5499 HV * const table = GvHV(PL_hintgv);
5502 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5503 return &PL_core_reg_engine;
5504 ptr = hv_fetchs(table, "regcomp", FALSE);
5505 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5506 return &PL_core_reg_engine;
5507 return INT2PTR(regexp_engine*,SvIV(*ptr));
5511 if (!PL_curcop->cop_hints_hash)
5512 return &PL_core_reg_engine;
5513 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5514 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5515 return &PL_core_reg_engine;
5516 return INT2PTR(regexp_engine*,SvIV(ptr));
5522 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5525 regexp_engine const *eng = current_re_engine();
5526 GET_RE_DEBUG_FLAGS_DECL;
5528 PERL_ARGS_ASSERT_PREGCOMP;
5530 /* Dispatch a request to compile a regexp to correct regexp engine. */
5532 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5535 return CALLREGCOMP_ENG(eng, pattern, flags);
5539 /* public(ish) entry point for the perl core's own regex compiling code.
5540 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5541 * pattern rather than a list of OPs, and uses the internal engine rather
5542 * than the current one */
5545 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5547 SV *pat = pattern; /* defeat constness! */
5548 PERL_ARGS_ASSERT_RE_COMPILE;
5549 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5550 #ifdef PERL_IN_XSUB_RE
5553 &PL_core_reg_engine,
5555 NULL, NULL, rx_flags, 0);
5559 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5560 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5561 * point to the realloced string and length.
5563 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5567 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5568 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5570 U8 *const src = (U8*)*pat_p;
5573 STRLEN s = 0, d = 0;
5575 GET_RE_DEBUG_FLAGS_DECL;
5577 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5578 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5580 Newx(dst, *plen_p * 2 + 1, U8);
5582 while (s < *plen_p) {
5583 if (NATIVE_BYTE_IS_INVARIANT(src[s]))
5586 dst[d++] = UTF8_EIGHT_BIT_HI(src[s]);
5587 dst[d] = UTF8_EIGHT_BIT_LO(src[s]);
5589 if (n < num_code_blocks) {
5590 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5591 pRExC_state->code_blocks[n].start = d;
5592 assert(dst[d] == '(');
5595 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5596 pRExC_state->code_blocks[n].end = d;
5597 assert(dst[d] == ')');
5607 *pat_p = (char*) dst;
5609 RExC_orig_utf8 = RExC_utf8 = 1;
5614 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5615 * while recording any code block indices, and handling overloading,
5616 * nested qr// objects etc. If pat is null, it will allocate a new
5617 * string, or just return the first arg, if there's only one.
5619 * Returns the malloced/updated pat.
5620 * patternp and pat_count is the array of SVs to be concatted;
5621 * oplist is the optional list of ops that generated the SVs;
5622 * recompile_p is a pointer to a boolean that will be set if
5623 * the regex will need to be recompiled.
5624 * delim, if non-null is an SV that will be inserted between each element
5628 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5629 SV *pat, SV ** const patternp, int pat_count,
5630 OP *oplist, bool *recompile_p, SV *delim)
5634 bool use_delim = FALSE;
5635 bool alloced = FALSE;
5637 /* if we know we have at least two args, create an empty string,
5638 * then concatenate args to that. For no args, return an empty string */
5639 if (!pat && pat_count != 1) {
5640 pat = newSVpvn("", 0);
5645 for (svp = patternp; svp < patternp + pat_count; svp++) {
5648 STRLEN orig_patlen = 0;
5650 SV *msv = use_delim ? delim : *svp;
5651 if (!msv) msv = &PL_sv_undef;
5653 /* if we've got a delimiter, we go round the loop twice for each
5654 * svp slot (except the last), using the delimiter the second
5663 if (SvTYPE(msv) == SVt_PVAV) {
5664 /* we've encountered an interpolated array within
5665 * the pattern, e.g. /...@a..../. Expand the list of elements,
5666 * then recursively append elements.
5667 * The code in this block is based on S_pushav() */
5669 AV *const av = (AV*)msv;
5670 const SSize_t maxarg = AvFILL(av) + 1;
5674 assert(oplist->op_type == OP_PADAV
5675 || oplist->op_type == OP_RV2AV);
5676 oplist = oplist->op_sibling;;
5679 if (SvRMAGICAL(av)) {
5682 Newx(array, maxarg, SV*);
5684 for (i=0; i < maxarg; i++) {
5685 SV ** const svp = av_fetch(av, i, FALSE);
5686 array[i] = svp ? *svp : &PL_sv_undef;
5690 array = AvARRAY(av);
5692 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5693 array, maxarg, NULL, recompile_p,
5695 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5701 /* we make the assumption here that each op in the list of
5702 * op_siblings maps to one SV pushed onto the stack,
5703 * except for code blocks, with have both an OP_NULL and
5705 * This allows us to match up the list of SVs against the
5706 * list of OPs to find the next code block.
5708 * Note that PUSHMARK PADSV PADSV ..
5710 * PADRANGE PADSV PADSV ..
5711 * so the alignment still works. */
5714 if (oplist->op_type == OP_NULL
5715 && (oplist->op_flags & OPf_SPECIAL))
5717 assert(n < pRExC_state->num_code_blocks);
5718 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5719 pRExC_state->code_blocks[n].block = oplist;
5720 pRExC_state->code_blocks[n].src_regex = NULL;
5723 oplist = oplist->op_sibling; /* skip CONST */
5726 oplist = oplist->op_sibling;;
5729 /* apply magic and QR overloading to arg */
5732 if (SvROK(msv) && SvAMAGIC(msv)) {
5733 SV *sv = AMG_CALLunary(msv, regexp_amg);
5737 if (SvTYPE(sv) != SVt_REGEXP)
5738 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5743 /* try concatenation overload ... */
5744 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5745 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5748 /* overloading involved: all bets are off over literal
5749 * code. Pretend we haven't seen it */
5750 pRExC_state->num_code_blocks -= n;
5754 /* ... or failing that, try "" overload */
5755 while (SvAMAGIC(msv)
5756 && (sv = AMG_CALLunary(msv, string_amg))
5760 && SvRV(msv) == SvRV(sv))
5765 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5769 /* this is a partially unrolled
5770 * sv_catsv_nomg(pat, msv);
5771 * that allows us to adjust code block indices if
5774 char *dst = SvPV_force_nomg(pat, dlen);
5776 if (SvUTF8(msv) && !SvUTF8(pat)) {
5777 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5778 sv_setpvn(pat, dst, dlen);
5781 sv_catsv_nomg(pat, msv);
5788 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5791 /* extract any code blocks within any embedded qr//'s */
5792 if (rx && SvTYPE(rx) == SVt_REGEXP
5793 && RX_ENGINE((REGEXP*)rx)->op_comp)
5796 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5797 if (ri->num_code_blocks) {
5799 /* the presence of an embedded qr// with code means
5800 * we should always recompile: the text of the
5801 * qr// may not have changed, but it may be a
5802 * different closure than last time */
5804 Renew(pRExC_state->code_blocks,
5805 pRExC_state->num_code_blocks + ri->num_code_blocks,
5806 struct reg_code_block);
5807 pRExC_state->num_code_blocks += ri->num_code_blocks;
5809 for (i=0; i < ri->num_code_blocks; i++) {
5810 struct reg_code_block *src, *dst;
5811 STRLEN offset = orig_patlen
5812 + ReANY((REGEXP *)rx)->pre_prefix;
5813 assert(n < pRExC_state->num_code_blocks);
5814 src = &ri->code_blocks[i];
5815 dst = &pRExC_state->code_blocks[n];
5816 dst->start = src->start + offset;
5817 dst->end = src->end + offset;
5818 dst->block = src->block;
5819 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5828 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5837 /* see if there are any run-time code blocks in the pattern.
5838 * False positives are allowed */
5841 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5842 char *pat, STRLEN plen)
5847 for (s = 0; s < plen; s++) {
5848 if (n < pRExC_state->num_code_blocks
5849 && s == pRExC_state->code_blocks[n].start)
5851 s = pRExC_state->code_blocks[n].end;
5855 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5857 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5859 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5866 /* Handle run-time code blocks. We will already have compiled any direct
5867 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5868 * copy of it, but with any literal code blocks blanked out and
5869 * appropriate chars escaped; then feed it into
5871 * eval "qr'modified_pattern'"
5875 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5879 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5881 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5882 * and merge them with any code blocks of the original regexp.
5884 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5885 * instead, just save the qr and return FALSE; this tells our caller that
5886 * the original pattern needs upgrading to utf8.
5890 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5891 char *pat, STRLEN plen)
5895 GET_RE_DEBUG_FLAGS_DECL;
5897 if (pRExC_state->runtime_code_qr) {
5898 /* this is the second time we've been called; this should
5899 * only happen if the main pattern got upgraded to utf8
5900 * during compilation; re-use the qr we compiled first time
5901 * round (which should be utf8 too)
5903 qr = pRExC_state->runtime_code_qr;
5904 pRExC_state->runtime_code_qr = NULL;
5905 assert(RExC_utf8 && SvUTF8(qr));
5911 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5915 /* determine how many extra chars we need for ' and \ escaping */
5916 for (s = 0; s < plen; s++) {
5917 if (pat[s] == '\'' || pat[s] == '\\')
5921 Newx(newpat, newlen, char);
5923 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5925 for (s = 0; s < plen; s++) {
5926 if (n < pRExC_state->num_code_blocks
5927 && s == pRExC_state->code_blocks[n].start)
5929 /* blank out literal code block */
5930 assert(pat[s] == '(');
5931 while (s <= pRExC_state->code_blocks[n].end) {
5939 if (pat[s] == '\'' || pat[s] == '\\')
5944 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5948 PerlIO_printf(Perl_debug_log,
5949 "%sre-parsing pattern for runtime code:%s %s\n",
5950 PL_colors[4],PL_colors[5],newpat);
5953 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5959 PUSHSTACKi(PERLSI_REQUIRE);
5960 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5961 * parsing qr''; normally only q'' does this. It also alters
5963 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5964 SvREFCNT_dec_NN(sv);
5969 SV * const errsv = ERRSV;
5970 if (SvTRUE_NN(errsv))
5972 Safefree(pRExC_state->code_blocks);
5973 /* use croak_sv ? */
5974 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
5977 assert(SvROK(qr_ref));
5979 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5980 /* the leaving below frees the tmp qr_ref.
5981 * Give qr a life of its own */
5989 if (!RExC_utf8 && SvUTF8(qr)) {
5990 /* first time through; the pattern got upgraded; save the
5991 * qr for the next time through */
5992 assert(!pRExC_state->runtime_code_qr);
5993 pRExC_state->runtime_code_qr = qr;
5998 /* extract any code blocks within the returned qr// */
6001 /* merge the main (r1) and run-time (r2) code blocks into one */
6003 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6004 struct reg_code_block *new_block, *dst;
6005 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6008 if (!r2->num_code_blocks) /* we guessed wrong */
6010 SvREFCNT_dec_NN(qr);
6015 r1->num_code_blocks + r2->num_code_blocks,
6016 struct reg_code_block);
6019 while ( i1 < r1->num_code_blocks
6020 || i2 < r2->num_code_blocks)
6022 struct reg_code_block *src;
6025 if (i1 == r1->num_code_blocks) {
6026 src = &r2->code_blocks[i2++];
6029 else if (i2 == r2->num_code_blocks)
6030 src = &r1->code_blocks[i1++];
6031 else if ( r1->code_blocks[i1].start
6032 < r2->code_blocks[i2].start)
6034 src = &r1->code_blocks[i1++];
6035 assert(src->end < r2->code_blocks[i2].start);
6038 assert( r1->code_blocks[i1].start
6039 > r2->code_blocks[i2].start);
6040 src = &r2->code_blocks[i2++];
6042 assert(src->end < r1->code_blocks[i1].start);
6045 assert(pat[src->start] == '(');
6046 assert(pat[src->end] == ')');
6047 dst->start = src->start;
6048 dst->end = src->end;
6049 dst->block = src->block;
6050 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6054 r1->num_code_blocks += r2->num_code_blocks;
6055 Safefree(r1->code_blocks);
6056 r1->code_blocks = new_block;
6059 SvREFCNT_dec_NN(qr);
6065 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6066 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6067 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6068 STRLEN longest_length, bool eol, bool meol)
6070 /* This is the common code for setting up the floating and fixed length
6071 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6072 * as to whether succeeded or not */
6077 if (! (longest_length
6078 || (eol /* Can't have SEOL and MULTI */
6079 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6081 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6082 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6087 /* copy the information about the longest from the reg_scan_data
6088 over to the program. */
6089 if (SvUTF8(sv_longest)) {
6090 *rx_utf8 = sv_longest;
6093 *rx_substr = sv_longest;
6096 /* end_shift is how many chars that must be matched that
6097 follow this item. We calculate it ahead of time as once the
6098 lookbehind offset is added in we lose the ability to correctly
6100 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6101 *rx_end_shift = ml - offset
6102 - longest_length + (SvTAIL(sv_longest) != 0)
6105 t = (eol/* Can't have SEOL and MULTI */
6106 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6107 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6113 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6114 * regular expression into internal code.
6115 * The pattern may be passed either as:
6116 * a list of SVs (patternp plus pat_count)
6117 * a list of OPs (expr)
6118 * If both are passed, the SV list is used, but the OP list indicates
6119 * which SVs are actually pre-compiled code blocks
6121 * The SVs in the list have magic and qr overloading applied to them (and
6122 * the list may be modified in-place with replacement SVs in the latter
6125 * If the pattern hasn't changed from old_re, then old_re will be
6128 * eng is the current engine. If that engine has an op_comp method, then
6129 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6130 * do the initial concatenation of arguments and pass on to the external
6133 * If is_bare_re is not null, set it to a boolean indicating whether the
6134 * arg list reduced (after overloading) to a single bare regex which has
6135 * been returned (i.e. /$qr/).
6137 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6139 * pm_flags contains the PMf_* flags, typically based on those from the
6140 * pm_flags field of the related PMOP. Currently we're only interested in
6141 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6143 * We can't allocate space until we know how big the compiled form will be,
6144 * but we can't compile it (and thus know how big it is) until we've got a
6145 * place to put the code. So we cheat: we compile it twice, once with code
6146 * generation turned off and size counting turned on, and once "for real".
6147 * This also means that we don't allocate space until we are sure that the
6148 * thing really will compile successfully, and we never have to move the
6149 * code and thus invalidate pointers into it. (Note that it has to be in
6150 * one piece because free() must be able to free it all.) [NB: not true in perl]
6152 * Beware that the optimization-preparation code in here knows about some
6153 * of the structure of the compiled regexp. [I'll say.]
6157 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6158 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6159 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6164 regexp_internal *ri;
6172 SV *code_blocksv = NULL;
6173 SV** new_patternp = patternp;
6175 /* these are all flags - maybe they should be turned
6176 * into a single int with different bit masks */
6177 I32 sawlookahead = 0;
6182 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6184 bool runtime_code = 0;
6186 RExC_state_t RExC_state;
6187 RExC_state_t * const pRExC_state = &RExC_state;
6188 #ifdef TRIE_STUDY_OPT
6190 RExC_state_t copyRExC_state;
6192 GET_RE_DEBUG_FLAGS_DECL;
6194 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6196 DEBUG_r(if (!PL_colorset) reginitcolors());
6198 #ifndef PERL_IN_XSUB_RE
6199 /* Initialize these here instead of as-needed, as is quick and avoids
6200 * having to test them each time otherwise */
6201 if (! PL_AboveLatin1) {
6202 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6203 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6204 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6205 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6206 PL_HasMultiCharFold =
6207 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6211 pRExC_state->code_blocks = NULL;
6212 pRExC_state->num_code_blocks = 0;
6215 *is_bare_re = FALSE;
6217 if (expr && (expr->op_type == OP_LIST ||
6218 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6219 /* allocate code_blocks if needed */
6223 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
6224 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6225 ncode++; /* count of DO blocks */
6227 pRExC_state->num_code_blocks = ncode;
6228 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6233 /* compile-time pattern with just OP_CONSTs and DO blocks */
6238 /* find how many CONSTs there are */
6241 if (expr->op_type == OP_CONST)
6244 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6245 if (o->op_type == OP_CONST)
6249 /* fake up an SV array */
6251 assert(!new_patternp);
6252 Newx(new_patternp, n, SV*);
6253 SAVEFREEPV(new_patternp);
6257 if (expr->op_type == OP_CONST)
6258 new_patternp[n] = cSVOPx_sv(expr);
6260 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6261 if (o->op_type == OP_CONST)
6262 new_patternp[n++] = cSVOPo_sv;
6267 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6268 "Assembling pattern from %d elements%s\n", pat_count,
6269 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6271 /* set expr to the first arg op */
6273 if (pRExC_state->num_code_blocks
6274 && expr->op_type != OP_CONST)
6276 expr = cLISTOPx(expr)->op_first;
6277 assert( expr->op_type == OP_PUSHMARK
6278 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6279 || expr->op_type == OP_PADRANGE);
6280 expr = expr->op_sibling;
6283 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6284 expr, &recompile, NULL);
6286 /* handle bare (possibly after overloading) regex: foo =~ $re */
6291 if (SvTYPE(re) == SVt_REGEXP) {
6295 Safefree(pRExC_state->code_blocks);
6296 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6297 "Precompiled pattern%s\n",
6298 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6304 exp = SvPV_nomg(pat, plen);
6306 if (!eng->op_comp) {
6307 if ((SvUTF8(pat) && IN_BYTES)
6308 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6310 /* make a temporary copy; either to convert to bytes,
6311 * or to avoid repeating get-magic / overloaded stringify */
6312 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6313 (IN_BYTES ? 0 : SvUTF8(pat)));
6315 Safefree(pRExC_state->code_blocks);
6316 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6319 /* ignore the utf8ness if the pattern is 0 length */
6320 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6321 RExC_uni_semantics = 0;
6322 RExC_contains_locale = 0;
6323 RExC_contains_i = 0;
6324 pRExC_state->runtime_code_qr = NULL;
6327 SV *dsv= sv_newmortal();
6328 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6329 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6330 PL_colors[4],PL_colors[5],s);
6334 /* we jump here if we upgrade the pattern to utf8 and have to
6337 if ((pm_flags & PMf_USE_RE_EVAL)
6338 /* this second condition covers the non-regex literal case,
6339 * i.e. $foo =~ '(?{})'. */
6340 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6342 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6344 /* return old regex if pattern hasn't changed */
6345 /* XXX: note in the below we have to check the flags as well as the
6348 * Things get a touch tricky as we have to compare the utf8 flag
6349 * independently from the compile flags. */
6353 && !!RX_UTF8(old_re) == !!RExC_utf8
6354 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6355 && RX_PRECOMP(old_re)
6356 && RX_PRELEN(old_re) == plen
6357 && memEQ(RX_PRECOMP(old_re), exp, plen)
6358 && !runtime_code /* with runtime code, always recompile */ )
6360 Safefree(pRExC_state->code_blocks);
6364 rx_flags = orig_rx_flags;
6366 if (rx_flags & PMf_FOLD) {
6367 RExC_contains_i = 1;
6369 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6371 /* Set to use unicode semantics if the pattern is in utf8 and has the
6372 * 'depends' charset specified, as it means unicode when utf8 */
6373 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6377 RExC_flags = rx_flags;
6378 RExC_pm_flags = pm_flags;
6381 if (TAINTING_get && TAINT_get)
6382 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6384 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6385 /* whoops, we have a non-utf8 pattern, whilst run-time code
6386 * got compiled as utf8. Try again with a utf8 pattern */
6387 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6388 pRExC_state->num_code_blocks);
6389 goto redo_first_pass;
6392 assert(!pRExC_state->runtime_code_qr);
6398 RExC_in_lookbehind = 0;
6399 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6401 RExC_override_recoding = 0;
6402 RExC_in_multi_char_class = 0;
6404 /* First pass: determine size, legality. */
6407 RExC_end = exp + plen;
6412 RExC_emit = (regnode *) &RExC_emit_dummy;
6413 RExC_whilem_seen = 0;
6414 RExC_open_parens = NULL;
6415 RExC_close_parens = NULL;
6417 RExC_paren_names = NULL;
6419 RExC_paren_name_list = NULL;
6421 RExC_recurse = NULL;
6422 RExC_study_chunk_recursed = NULL;
6423 RExC_study_chunk_recursed_bytes= 0;
6424 RExC_recurse_count = 0;
6425 pRExC_state->code_index = 0;
6427 #if 0 /* REGC() is (currently) a NOP at the first pass.
6428 * Clever compilers notice this and complain. --jhi */
6429 REGC((U8)REG_MAGIC, (char*)RExC_emit);
6432 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6434 RExC_lastparse=NULL;
6436 /* reg may croak on us, not giving us a chance to free
6437 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6438 need it to survive as long as the regexp (qr/(?{})/).
6439 We must check that code_blocksv is not already set, because we may
6440 have jumped back to restart the sizing pass. */
6441 if (pRExC_state->code_blocks && !code_blocksv) {
6442 code_blocksv = newSV_type(SVt_PV);
6443 SAVEFREESV(code_blocksv);
6444 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6445 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6447 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6448 /* It's possible to write a regexp in ascii that represents Unicode
6449 codepoints outside of the byte range, such as via \x{100}. If we
6450 detect such a sequence we have to convert the entire pattern to utf8
6451 and then recompile, as our sizing calculation will have been based
6452 on 1 byte == 1 character, but we will need to use utf8 to encode
6453 at least some part of the pattern, and therefore must convert the whole
6456 if (flags & RESTART_UTF8) {
6457 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6458 pRExC_state->num_code_blocks);
6459 goto redo_first_pass;
6461 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6464 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6467 PerlIO_printf(Perl_debug_log,
6468 "Required size %"IVdf" nodes\n"
6469 "Starting second pass (creation)\n",
6472 RExC_lastparse=NULL;
6475 /* The first pass could have found things that force Unicode semantics */
6476 if ((RExC_utf8 || RExC_uni_semantics)
6477 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6479 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6482 /* Small enough for pointer-storage convention?
6483 If extralen==0, this means that we will not need long jumps. */
6484 if (RExC_size >= 0x10000L && RExC_extralen)
6485 RExC_size += RExC_extralen;
6488 if (RExC_whilem_seen > 15)
6489 RExC_whilem_seen = 15;
6491 /* Allocate space and zero-initialize. Note, the two step process
6492 of zeroing when in debug mode, thus anything assigned has to
6493 happen after that */
6494 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6496 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6497 char, regexp_internal);
6498 if ( r == NULL || ri == NULL )
6499 FAIL("Regexp out of space");
6501 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6502 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6505 /* bulk initialize base fields with 0. */
6506 Zero(ri, sizeof(regexp_internal), char);
6509 /* non-zero initialization begins here */
6512 r->extflags = rx_flags;
6513 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6515 if (pm_flags & PMf_IS_QR) {
6516 ri->code_blocks = pRExC_state->code_blocks;
6517 ri->num_code_blocks = pRExC_state->num_code_blocks;
6522 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6523 if (pRExC_state->code_blocks[n].src_regex)
6524 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6525 SAVEFREEPV(pRExC_state->code_blocks);
6529 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6530 bool has_charset = (get_regex_charset(r->extflags)
6531 != REGEX_DEPENDS_CHARSET);
6533 /* The caret is output if there are any defaults: if not all the STD
6534 * flags are set, or if no character set specifier is needed */
6536 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6538 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6539 == REG_RUN_ON_COMMENT_SEEN);
6540 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6541 >> RXf_PMf_STD_PMMOD_SHIFT);
6542 const char *fptr = STD_PAT_MODS; /*"msix"*/
6544 /* Allocate for the worst case, which is all the std flags are turned
6545 * on. If more precision is desired, we could do a population count of
6546 * the flags set. This could be done with a small lookup table, or by
6547 * shifting, masking and adding, or even, when available, assembly
6548 * language for a machine-language population count.
6549 * We never output a minus, as all those are defaults, so are
6550 * covered by the caret */
6551 const STRLEN wraplen = plen + has_p + has_runon
6552 + has_default /* If needs a caret */
6554 /* If needs a character set specifier */
6555 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6556 + (sizeof(STD_PAT_MODS) - 1)
6557 + (sizeof("(?:)") - 1);
6559 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6560 r->xpv_len_u.xpvlenu_pv = p;
6562 SvFLAGS(rx) |= SVf_UTF8;
6565 /* If a default, cover it using the caret */
6567 *p++= DEFAULT_PAT_MOD;
6571 const char* const name = get_regex_charset_name(r->extflags, &len);
6572 Copy(name, p, len, char);
6576 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6579 while((ch = *fptr++)) {
6587 Copy(RExC_precomp, p, plen, char);
6588 assert ((RX_WRAPPED(rx) - p) < 16);
6589 r->pre_prefix = p - RX_WRAPPED(rx);
6595 SvCUR_set(rx, p - RX_WRAPPED(rx));
6599 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6601 /* setup various meta data about recursion, this all requires
6602 * RExC_npar to be correctly set, and a bit later on we clear it */
6603 if (RExC_seen & REG_RECURSE_SEEN) {
6604 Newxz(RExC_open_parens, RExC_npar,regnode *);
6605 SAVEFREEPV(RExC_open_parens);
6606 Newxz(RExC_close_parens,RExC_npar,regnode *);
6607 SAVEFREEPV(RExC_close_parens);
6609 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6610 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6611 * So its 1 if there are no parens. */
6612 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6613 ((RExC_npar & 0x07) != 0);
6614 Newx(RExC_study_chunk_recursed,
6615 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6616 SAVEFREEPV(RExC_study_chunk_recursed);
6619 /* Useful during FAIL. */
6620 #ifdef RE_TRACK_PATTERN_OFFSETS
6621 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6622 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6623 "%s %"UVuf" bytes for offset annotations.\n",
6624 ri->u.offsets ? "Got" : "Couldn't get",
6625 (UV)((2*RExC_size+1) * sizeof(U32))));
6627 SetProgLen(ri,RExC_size);
6632 /* Second pass: emit code. */
6633 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6634 RExC_pm_flags = pm_flags;
6636 RExC_end = exp + plen;
6639 RExC_emit_start = ri->program;
6640 RExC_emit = ri->program;
6641 RExC_emit_bound = ri->program + RExC_size + 1;
6642 pRExC_state->code_index = 0;
6644 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6645 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6647 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6649 /* XXXX To minimize changes to RE engine we always allocate
6650 3-units-long substrs field. */
6651 Newx(r->substrs, 1, struct reg_substr_data);
6652 if (RExC_recurse_count) {
6653 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6654 SAVEFREEPV(RExC_recurse);
6658 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6659 Zero(r->substrs, 1, struct reg_substr_data);
6660 if (RExC_study_chunk_recursed)
6661 Zero(RExC_study_chunk_recursed,
6662 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6664 #ifdef TRIE_STUDY_OPT
6666 StructCopy(&zero_scan_data, &data, scan_data_t);
6667 copyRExC_state = RExC_state;
6670 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6672 RExC_state = copyRExC_state;
6673 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6674 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6676 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6677 StructCopy(&zero_scan_data, &data, scan_data_t);
6680 StructCopy(&zero_scan_data, &data, scan_data_t);
6683 /* Dig out information for optimizations. */
6684 r->extflags = RExC_flags; /* was pm_op */
6685 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6688 SvUTF8_on(rx); /* Unicode in it? */
6689 ri->regstclass = NULL;
6690 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6691 r->intflags |= PREGf_NAUGHTY;
6692 scan = ri->program + 1; /* First BRANCH. */
6694 /* testing for BRANCH here tells us whether there is "must appear"
6695 data in the pattern. If there is then we can use it for optimisations */
6696 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6699 STRLEN longest_float_length, longest_fixed_length;
6700 regnode_ssc ch_class; /* pointed to by data */
6702 SSize_t last_close = 0; /* pointed to by data */
6703 regnode *first= scan;
6704 regnode *first_next= regnext(first);
6706 * Skip introductions and multiplicators >= 1
6707 * so that we can extract the 'meat' of the pattern that must
6708 * match in the large if() sequence following.
6709 * NOTE that EXACT is NOT covered here, as it is normally
6710 * picked up by the optimiser separately.
6712 * This is unfortunate as the optimiser isnt handling lookahead
6713 * properly currently.
6716 while ((OP(first) == OPEN && (sawopen = 1)) ||
6717 /* An OR of *one* alternative - should not happen now. */
6718 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6719 /* for now we can't handle lookbehind IFMATCH*/
6720 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6721 (OP(first) == PLUS) ||
6722 (OP(first) == MINMOD) ||
6723 /* An {n,m} with n>0 */
6724 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6725 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6728 * the only op that could be a regnode is PLUS, all the rest
6729 * will be regnode_1 or regnode_2.
6731 * (yves doesn't think this is true)
6733 if (OP(first) == PLUS)
6736 if (OP(first) == MINMOD)
6738 first += regarglen[OP(first)];
6740 first = NEXTOPER(first);
6741 first_next= regnext(first);
6744 /* Starting-point info. */
6746 DEBUG_PEEP("first:",first,0);
6747 /* Ignore EXACT as we deal with it later. */
6748 if (PL_regkind[OP(first)] == EXACT) {
6749 if (OP(first) == EXACT)
6750 NOOP; /* Empty, get anchored substr later. */
6752 ri->regstclass = first;
6755 else if (PL_regkind[OP(first)] == TRIE &&
6756 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6759 /* this can happen only on restudy */
6760 if ( OP(first) == TRIE ) {
6761 struct regnode_1 *trieop = (struct regnode_1 *)
6762 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6763 StructCopy(first,trieop,struct regnode_1);
6764 trie_op=(regnode *)trieop;
6766 struct regnode_charclass *trieop = (struct regnode_charclass *)
6767 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6768 StructCopy(first,trieop,struct regnode_charclass);
6769 trie_op=(regnode *)trieop;
6772 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6773 ri->regstclass = trie_op;
6776 else if (REGNODE_SIMPLE(OP(first)))
6777 ri->regstclass = first;
6778 else if (PL_regkind[OP(first)] == BOUND ||
6779 PL_regkind[OP(first)] == NBOUND)
6780 ri->regstclass = first;
6781 else if (PL_regkind[OP(first)] == BOL) {
6782 r->intflags |= (OP(first) == MBOL
6784 : (OP(first) == SBOL
6787 first = NEXTOPER(first);
6790 else if (OP(first) == GPOS) {
6791 r->intflags |= PREGf_ANCH_GPOS;
6792 first = NEXTOPER(first);
6795 else if ((!sawopen || !RExC_sawback) &&
6796 (OP(first) == STAR &&
6797 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6798 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
6800 /* turn .* into ^.* with an implied $*=1 */
6802 (OP(NEXTOPER(first)) == REG_ANY)
6805 r->intflags |= (type | PREGf_IMPLICIT);
6806 first = NEXTOPER(first);
6809 if (sawplus && !sawminmod && !sawlookahead
6810 && (!sawopen || !RExC_sawback)
6811 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6812 /* x+ must match at the 1st pos of run of x's */
6813 r->intflags |= PREGf_SKIP;
6815 /* Scan is after the zeroth branch, first is atomic matcher. */
6816 #ifdef TRIE_STUDY_OPT
6819 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6820 (IV)(first - scan + 1))
6824 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6825 (IV)(first - scan + 1))
6831 * If there's something expensive in the r.e., find the
6832 * longest literal string that must appear and make it the
6833 * regmust. Resolve ties in favor of later strings, since
6834 * the regstart check works with the beginning of the r.e.
6835 * and avoiding duplication strengthens checking. Not a
6836 * strong reason, but sufficient in the absence of others.
6837 * [Now we resolve ties in favor of the earlier string if
6838 * it happens that c_offset_min has been invalidated, since the
6839 * earlier string may buy us something the later one won't.]
6842 data.longest_fixed = newSVpvs("");
6843 data.longest_float = newSVpvs("");
6844 data.last_found = newSVpvs("");
6845 data.longest = &(data.longest_fixed);
6846 ENTER_with_name("study_chunk");
6847 SAVEFREESV(data.longest_fixed);
6848 SAVEFREESV(data.longest_float);
6849 SAVEFREESV(data.last_found);
6851 if (!ri->regstclass) {
6852 ssc_init(pRExC_state, &ch_class);
6853 data.start_class = &ch_class;
6854 stclass_flag = SCF_DO_STCLASS_AND;
6855 } else /* XXXX Check for BOUND? */
6857 data.last_closep = &last_close;
6860 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
6861 scan + RExC_size, /* Up to end */
6863 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6864 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6868 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6871 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6872 && data.last_start_min == 0 && data.last_end > 0
6873 && !RExC_seen_zerolen
6874 && !(RExC_seen & REG_VERBARG_SEEN)
6875 && !(RExC_seen & REG_GPOS_SEEN)
6877 r->extflags |= RXf_CHECK_ALL;
6879 scan_commit(pRExC_state, &data,&minlen,0);
6881 longest_float_length = CHR_SVLEN(data.longest_float);
6883 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6884 && data.offset_fixed == data.offset_float_min
6885 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6886 && S_setup_longest (aTHX_ pRExC_state,
6890 &(r->float_end_shift),
6891 data.lookbehind_float,
6892 data.offset_float_min,
6894 longest_float_length,
6895 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6896 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6898 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6899 r->float_max_offset = data.offset_float_max;
6900 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
6901 r->float_max_offset -= data.lookbehind_float;
6902 SvREFCNT_inc_simple_void_NN(data.longest_float);
6905 r->float_substr = r->float_utf8 = NULL;
6906 longest_float_length = 0;
6909 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6911 if (S_setup_longest (aTHX_ pRExC_state,
6913 &(r->anchored_utf8),
6914 &(r->anchored_substr),
6915 &(r->anchored_end_shift),
6916 data.lookbehind_fixed,
6919 longest_fixed_length,
6920 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6921 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6923 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6924 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6927 r->anchored_substr = r->anchored_utf8 = NULL;
6928 longest_fixed_length = 0;
6930 LEAVE_with_name("study_chunk");
6933 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6934 ri->regstclass = NULL;
6936 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6938 && ! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
6939 && !ssc_is_anything(data.start_class))
6941 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
6943 ssc_finalize(pRExC_state, data.start_class);
6945 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
6946 StructCopy(data.start_class,
6947 (regnode_ssc*)RExC_rxi->data->data[n],
6949 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6950 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6951 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6952 regprop(r, sv, (regnode*)data.start_class, NULL);
6953 PerlIO_printf(Perl_debug_log,
6954 "synthetic stclass \"%s\".\n",
6955 SvPVX_const(sv));});
6956 data.start_class = NULL;
6959 /* A temporary algorithm prefers floated substr to fixed one to dig
6961 if (longest_fixed_length > longest_float_length) {
6962 r->substrs->check_ix = 0;
6963 r->check_end_shift = r->anchored_end_shift;
6964 r->check_substr = r->anchored_substr;
6965 r->check_utf8 = r->anchored_utf8;
6966 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6967 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
6968 r->intflags |= PREGf_NOSCAN;
6971 r->substrs->check_ix = 1;
6972 r->check_end_shift = r->float_end_shift;
6973 r->check_substr = r->float_substr;
6974 r->check_utf8 = r->float_utf8;
6975 r->check_offset_min = r->float_min_offset;
6976 r->check_offset_max = r->float_max_offset;
6978 if ((r->check_substr || r->check_utf8) ) {
6979 r->extflags |= RXf_USE_INTUIT;
6980 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6981 r->extflags |= RXf_INTUIT_TAIL;
6983 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
6985 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6986 if ( (STRLEN)minlen < longest_float_length )
6987 minlen= longest_float_length;
6988 if ( (STRLEN)minlen < longest_fixed_length )
6989 minlen= longest_fixed_length;
6993 /* Several toplevels. Best we can is to set minlen. */
6995 regnode_ssc ch_class;
6996 SSize_t last_close = 0;
6998 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7000 scan = ri->program + 1;
7001 ssc_init(pRExC_state, &ch_class);
7002 data.start_class = &ch_class;
7003 data.last_closep = &last_close;
7006 minlen = study_chunk(pRExC_state,
7007 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7008 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7009 ? SCF_TRIE_DOING_RESTUDY
7013 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7015 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7016 = r->float_substr = r->float_utf8 = NULL;
7018 if (! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
7019 && ! ssc_is_anything(data.start_class))
7021 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7023 ssc_finalize(pRExC_state, data.start_class);
7025 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7026 StructCopy(data.start_class,
7027 (regnode_ssc*)RExC_rxi->data->data[n],
7029 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7030 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7031 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7032 regprop(r, sv, (regnode*)data.start_class, NULL);
7033 PerlIO_printf(Perl_debug_log,
7034 "synthetic stclass \"%s\".\n",
7035 SvPVX_const(sv));});
7036 data.start_class = NULL;
7040 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7041 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7042 r->maxlen = REG_INFTY;
7045 r->maxlen = RExC_maxlen;
7048 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7049 the "real" pattern. */
7051 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%ld\n",
7052 (IV)minlen, (IV)r->minlen, RExC_maxlen);
7054 r->minlenret = minlen;
7055 if (r->minlen < minlen)
7058 if (RExC_seen & REG_GPOS_SEEN)
7059 r->intflags |= PREGf_GPOS_SEEN;
7060 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7061 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7063 if (pRExC_state->num_code_blocks)
7064 r->extflags |= RXf_EVAL_SEEN;
7065 if (RExC_seen & REG_CANY_SEEN)
7066 r->intflags |= PREGf_CANY_SEEN;
7067 if (RExC_seen & REG_VERBARG_SEEN)
7069 r->intflags |= PREGf_VERBARG_SEEN;
7070 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7072 if (RExC_seen & REG_CUTGROUP_SEEN)
7073 r->intflags |= PREGf_CUTGROUP_SEEN;
7074 if (pm_flags & PMf_USE_RE_EVAL)
7075 r->intflags |= PREGf_USE_RE_EVAL;
7076 if (RExC_paren_names)
7077 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7079 RXp_PAREN_NAMES(r) = NULL;
7081 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7082 * so it can be used in pp.c */
7083 if (r->intflags & PREGf_ANCH)
7084 r->extflags |= RXf_IS_ANCHORED;
7088 /* this is used to identify "special" patterns that might result
7089 * in Perl NOT calling the regex engine and instead doing the match "itself",
7090 * particularly special cases in split//. By having the regex compiler
7091 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7092 * we avoid weird issues with equivalent patterns resulting in different behavior,
7093 * AND we allow non Perl engines to get the same optimizations by the setting the
7094 * flags appropriately - Yves */
7095 regnode *first = ri->program + 1;
7097 regnode *next = NEXTOPER(first);
7100 if (PL_regkind[fop] == NOTHING && nop == END)
7101 r->extflags |= RXf_NULL;
7102 else if (PL_regkind[fop] == BOL && nop == END)
7103 r->extflags |= RXf_START_ONLY;
7104 else if (fop == PLUS
7105 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7106 && OP(regnext(first)) == END)
7107 r->extflags |= RXf_WHITE;
7108 else if ( r->extflags & RXf_SPLIT
7110 && STR_LEN(first) == 1
7111 && *(STRING(first)) == ' '
7112 && OP(regnext(first)) == END )
7113 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7117 if (RExC_contains_locale) {
7118 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7122 if (RExC_paren_names) {
7123 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7124 ri->data->data[ri->name_list_idx]
7125 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7128 ri->name_list_idx = 0;
7130 if (RExC_recurse_count) {
7131 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7132 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7133 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7136 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7137 /* assume we don't need to swap parens around before we match */
7141 PerlIO_printf(Perl_debug_log,"Final program:\n");
7144 #ifdef RE_TRACK_PATTERN_OFFSETS
7145 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7146 const STRLEN len = ri->u.offsets[0];
7148 GET_RE_DEBUG_FLAGS_DECL;
7149 PerlIO_printf(Perl_debug_log,
7150 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7151 for (i = 1; i <= len; i++) {
7152 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7153 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7154 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7156 PerlIO_printf(Perl_debug_log, "\n");
7161 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7162 * by setting the regexp SV to readonly-only instead. If the
7163 * pattern's been recompiled, the USEDness should remain. */
7164 if (old_re && SvREADONLY(old_re))
7172 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7175 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7177 PERL_UNUSED_ARG(value);
7179 if (flags & RXapif_FETCH) {
7180 return reg_named_buff_fetch(rx, key, flags);
7181 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7182 Perl_croak_no_modify();
7184 } else if (flags & RXapif_EXISTS) {
7185 return reg_named_buff_exists(rx, key, flags)
7188 } else if (flags & RXapif_REGNAMES) {
7189 return reg_named_buff_all(rx, flags);
7190 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7191 return reg_named_buff_scalar(rx, flags);
7193 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7199 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7202 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7203 PERL_UNUSED_ARG(lastkey);
7205 if (flags & RXapif_FIRSTKEY)
7206 return reg_named_buff_firstkey(rx, flags);
7207 else if (flags & RXapif_NEXTKEY)
7208 return reg_named_buff_nextkey(rx, flags);
7210 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7217 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7220 AV *retarray = NULL;
7222 struct regexp *const rx = ReANY(r);
7224 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7226 if (flags & RXapif_ALL)
7229 if (rx && RXp_PAREN_NAMES(rx)) {
7230 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7233 SV* sv_dat=HeVAL(he_str);
7234 I32 *nums=(I32*)SvPVX(sv_dat);
7235 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7236 if ((I32)(rx->nparens) >= nums[i]
7237 && rx->offs[nums[i]].start != -1
7238 && rx->offs[nums[i]].end != -1)
7241 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7246 ret = newSVsv(&PL_sv_undef);
7249 av_push(retarray, ret);
7252 return newRV_noinc(MUTABLE_SV(retarray));
7259 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7262 struct regexp *const rx = ReANY(r);
7264 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7266 if (rx && RXp_PAREN_NAMES(rx)) {
7267 if (flags & RXapif_ALL) {
7268 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7270 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7272 SvREFCNT_dec_NN(sv);
7284 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7286 struct regexp *const rx = ReANY(r);
7288 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7290 if ( rx && RXp_PAREN_NAMES(rx) ) {
7291 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7293 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7300 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7302 struct regexp *const rx = ReANY(r);
7303 GET_RE_DEBUG_FLAGS_DECL;
7305 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7307 if (rx && RXp_PAREN_NAMES(rx)) {
7308 HV *hv = RXp_PAREN_NAMES(rx);
7310 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7313 SV* sv_dat = HeVAL(temphe);
7314 I32 *nums = (I32*)SvPVX(sv_dat);
7315 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7316 if ((I32)(rx->lastparen) >= nums[i] &&
7317 rx->offs[nums[i]].start != -1 &&
7318 rx->offs[nums[i]].end != -1)
7324 if (parno || flags & RXapif_ALL) {
7325 return newSVhek(HeKEY_hek(temphe));
7333 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7338 struct regexp *const rx = ReANY(r);
7340 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7342 if (rx && RXp_PAREN_NAMES(rx)) {
7343 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7344 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7345 } else if (flags & RXapif_ONE) {
7346 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7347 av = MUTABLE_AV(SvRV(ret));
7348 length = av_tindex(av);
7349 SvREFCNT_dec_NN(ret);
7350 return newSViv(length + 1);
7352 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7357 return &PL_sv_undef;
7361 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7363 struct regexp *const rx = ReANY(r);
7366 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7368 if (rx && RXp_PAREN_NAMES(rx)) {
7369 HV *hv= RXp_PAREN_NAMES(rx);
7371 (void)hv_iterinit(hv);
7372 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7375 SV* sv_dat = HeVAL(temphe);
7376 I32 *nums = (I32*)SvPVX(sv_dat);
7377 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7378 if ((I32)(rx->lastparen) >= nums[i] &&
7379 rx->offs[nums[i]].start != -1 &&
7380 rx->offs[nums[i]].end != -1)
7386 if (parno || flags & RXapif_ALL) {
7387 av_push(av, newSVhek(HeKEY_hek(temphe)));
7392 return newRV_noinc(MUTABLE_SV(av));
7396 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7399 struct regexp *const rx = ReANY(r);
7405 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7407 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7408 || n == RX_BUFF_IDX_CARET_FULLMATCH
7409 || n == RX_BUFF_IDX_CARET_POSTMATCH
7412 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7414 /* on something like
7417 * the KEEPCOPY is set on the PMOP rather than the regex */
7418 if (PL_curpm && r == PM_GETRE(PL_curpm))
7419 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7428 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7429 /* no need to distinguish between them any more */
7430 n = RX_BUFF_IDX_FULLMATCH;
7432 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7433 && rx->offs[0].start != -1)
7435 /* $`, ${^PREMATCH} */
7436 i = rx->offs[0].start;
7440 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7441 && rx->offs[0].end != -1)
7443 /* $', ${^POSTMATCH} */
7444 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7445 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7448 if ( 0 <= n && n <= (I32)rx->nparens &&
7449 (s1 = rx->offs[n].start) != -1 &&
7450 (t1 = rx->offs[n].end) != -1)
7452 /* $&, ${^MATCH}, $1 ... */
7454 s = rx->subbeg + s1 - rx->suboffset;
7459 assert(s >= rx->subbeg);
7460 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7462 #ifdef NO_TAINT_SUPPORT
7463 sv_setpvn(sv, s, i);
7465 const int oldtainted = TAINT_get;
7467 sv_setpvn(sv, s, i);
7468 TAINT_set(oldtainted);
7470 if ( (rx->intflags & PREGf_CANY_SEEN)
7471 ? (RXp_MATCH_UTF8(rx)
7472 && (!i || is_utf8_string((U8*)s, i)))
7473 : (RXp_MATCH_UTF8(rx)) )
7480 if (RXp_MATCH_TAINTED(rx)) {
7481 if (SvTYPE(sv) >= SVt_PVMG) {
7482 MAGIC* const mg = SvMAGIC(sv);
7485 SvMAGIC_set(sv, mg->mg_moremagic);
7487 if ((mgt = SvMAGIC(sv))) {
7488 mg->mg_moremagic = mgt;
7489 SvMAGIC_set(sv, mg);
7500 sv_setsv(sv,&PL_sv_undef);
7506 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7507 SV const * const value)
7509 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7511 PERL_UNUSED_ARG(rx);
7512 PERL_UNUSED_ARG(paren);
7513 PERL_UNUSED_ARG(value);
7516 Perl_croak_no_modify();
7520 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7523 struct regexp *const rx = ReANY(r);
7527 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7529 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7530 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7531 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7534 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7536 /* on something like
7539 * the KEEPCOPY is set on the PMOP rather than the regex */
7540 if (PL_curpm && r == PM_GETRE(PL_curpm))
7541 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7547 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7549 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7550 case RX_BUFF_IDX_PREMATCH: /* $` */
7551 if (rx->offs[0].start != -1) {
7552 i = rx->offs[0].start;
7561 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7562 case RX_BUFF_IDX_POSTMATCH: /* $' */
7563 if (rx->offs[0].end != -1) {
7564 i = rx->sublen - rx->offs[0].end;
7566 s1 = rx->offs[0].end;
7573 default: /* $& / ${^MATCH}, $1, $2, ... */
7574 if (paren <= (I32)rx->nparens &&
7575 (s1 = rx->offs[paren].start) != -1 &&
7576 (t1 = rx->offs[paren].end) != -1)
7582 if (ckWARN(WARN_UNINITIALIZED))
7583 report_uninit((const SV *)sv);
7588 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7589 const char * const s = rx->subbeg - rx->suboffset + s1;
7594 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7601 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7603 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7604 PERL_UNUSED_ARG(rx);
7608 return newSVpvs("Regexp");
7611 /* Scans the name of a named buffer from the pattern.
7612 * If flags is REG_RSN_RETURN_NULL returns null.
7613 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7614 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7615 * to the parsed name as looked up in the RExC_paren_names hash.
7616 * If there is an error throws a vFAIL().. type exception.
7619 #define REG_RSN_RETURN_NULL 0
7620 #define REG_RSN_RETURN_NAME 1
7621 #define REG_RSN_RETURN_DATA 2
7624 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7626 char *name_start = RExC_parse;
7628 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7630 assert (RExC_parse <= RExC_end);
7631 if (RExC_parse == RExC_end) NOOP;
7632 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7633 /* skip IDFIRST by using do...while */
7636 RExC_parse += UTF8SKIP(RExC_parse);
7637 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7641 } while (isWORDCHAR(*RExC_parse));
7643 RExC_parse++; /* so the <- from the vFAIL is after the offending
7645 vFAIL("Group name must start with a non-digit word character");
7649 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7650 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7651 if ( flags == REG_RSN_RETURN_NAME)
7653 else if (flags==REG_RSN_RETURN_DATA) {
7656 if ( ! sv_name ) /* should not happen*/
7657 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7658 if (RExC_paren_names)
7659 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7661 sv_dat = HeVAL(he_str);
7663 vFAIL("Reference to nonexistent named group");
7667 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7668 (unsigned long) flags);
7670 assert(0); /* NOT REACHED */
7675 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7676 int rem=(int)(RExC_end - RExC_parse); \
7685 if (RExC_lastparse!=RExC_parse) \
7686 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7689 iscut ? "..." : "<" \
7692 PerlIO_printf(Perl_debug_log,"%16s",""); \
7695 num = RExC_size + 1; \
7697 num=REG_NODE_NUM(RExC_emit); \
7698 if (RExC_lastnum!=num) \
7699 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7701 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7702 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7703 (int)((depth*2)), "", \
7707 RExC_lastparse=RExC_parse; \
7712 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7713 DEBUG_PARSE_MSG((funcname)); \
7714 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7716 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7717 DEBUG_PARSE_MSG((funcname)); \
7718 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7721 /* This section of code defines the inversion list object and its methods. The
7722 * interfaces are highly subject to change, so as much as possible is static to
7723 * this file. An inversion list is here implemented as a malloc'd C UV array
7724 * as an SVt_INVLIST scalar.
7726 * An inversion list for Unicode is an array of code points, sorted by ordinal
7727 * number. The zeroth element is the first code point in the list. The 1th
7728 * element is the first element beyond that not in the list. In other words,
7729 * the first range is
7730 * invlist[0]..(invlist[1]-1)
7731 * The other ranges follow. Thus every element whose index is divisible by two
7732 * marks the beginning of a range that is in the list, and every element not
7733 * divisible by two marks the beginning of a range not in the list. A single
7734 * element inversion list that contains the single code point N generally
7735 * consists of two elements
7738 * (The exception is when N is the highest representable value on the
7739 * machine, in which case the list containing just it would be a single
7740 * element, itself. By extension, if the last range in the list extends to
7741 * infinity, then the first element of that range will be in the inversion list
7742 * at a position that is divisible by two, and is the final element in the
7744 * Taking the complement (inverting) an inversion list is quite simple, if the
7745 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7746 * This implementation reserves an element at the beginning of each inversion
7747 * list to always contain 0; there is an additional flag in the header which
7748 * indicates if the list begins at the 0, or is offset to begin at the next
7751 * More about inversion lists can be found in "Unicode Demystified"
7752 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7753 * More will be coming when functionality is added later.
7755 * The inversion list data structure is currently implemented as an SV pointing
7756 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7757 * array of UV whose memory management is automatically handled by the existing
7758 * facilities for SV's.
7760 * Some of the methods should always be private to the implementation, and some
7761 * should eventually be made public */
7763 /* The header definitions are in F<inline_invlist.c> */
7765 PERL_STATIC_INLINE UV*
7766 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7768 /* Returns a pointer to the first element in the inversion list's array.
7769 * This is called upon initialization of an inversion list. Where the
7770 * array begins depends on whether the list has the code point U+0000 in it
7771 * or not. The other parameter tells it whether the code that follows this
7772 * call is about to put a 0 in the inversion list or not. The first
7773 * element is either the element reserved for 0, if TRUE, or the element
7774 * after it, if FALSE */
7776 bool* offset = get_invlist_offset_addr(invlist);
7777 UV* zero_addr = (UV *) SvPVX(invlist);
7779 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7782 assert(! _invlist_len(invlist));
7786 /* 1^1 = 0; 1^0 = 1 */
7787 *offset = 1 ^ will_have_0;
7788 return zero_addr + *offset;
7791 PERL_STATIC_INLINE UV*
7792 S_invlist_array(pTHX_ SV* const invlist)
7794 /* Returns the pointer to the inversion list's array. Every time the
7795 * length changes, this needs to be called in case malloc or realloc moved
7798 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7800 /* Must not be empty. If these fail, you probably didn't check for <len>
7801 * being non-zero before trying to get the array */
7802 assert(_invlist_len(invlist));
7804 /* The very first element always contains zero, The array begins either
7805 * there, or if the inversion list is offset, at the element after it.
7806 * The offset header field determines which; it contains 0 or 1 to indicate
7807 * how much additionally to add */
7808 assert(0 == *(SvPVX(invlist)));
7809 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7812 PERL_STATIC_INLINE void
7813 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7815 /* Sets the current number of elements stored in the inversion list.
7816 * Updates SvCUR correspondingly */
7818 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7820 assert(SvTYPE(invlist) == SVt_INVLIST);
7825 : TO_INTERNAL_SIZE(len + offset));
7826 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7829 PERL_STATIC_INLINE IV*
7830 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7832 /* Return the address of the IV that is reserved to hold the cached index
7835 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7837 assert(SvTYPE(invlist) == SVt_INVLIST);
7839 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7842 PERL_STATIC_INLINE IV
7843 S_invlist_previous_index(pTHX_ SV* const invlist)
7845 /* Returns cached index of previous search */
7847 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7849 return *get_invlist_previous_index_addr(invlist);
7852 PERL_STATIC_INLINE void
7853 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7855 /* Caches <index> for later retrieval */
7857 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7859 assert(index == 0 || index < (int) _invlist_len(invlist));
7861 *get_invlist_previous_index_addr(invlist) = index;
7864 PERL_STATIC_INLINE UV
7865 S_invlist_max(pTHX_ SV* const invlist)
7867 /* Returns the maximum number of elements storable in the inversion list's
7868 * array, without having to realloc() */
7870 PERL_ARGS_ASSERT_INVLIST_MAX;
7872 assert(SvTYPE(invlist) == SVt_INVLIST);
7874 /* Assumes worst case, in which the 0 element is not counted in the
7875 * inversion list, so subtracts 1 for that */
7876 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7877 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7878 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7881 #ifndef PERL_IN_XSUB_RE
7883 Perl__new_invlist(pTHX_ IV initial_size)
7886 /* Return a pointer to a newly constructed inversion list, with enough
7887 * space to store 'initial_size' elements. If that number is negative, a
7888 * system default is used instead */
7892 if (initial_size < 0) {
7896 /* Allocate the initial space */
7897 new_list = newSV_type(SVt_INVLIST);
7899 /* First 1 is in case the zero element isn't in the list; second 1 is for
7901 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7902 invlist_set_len(new_list, 0, 0);
7904 /* Force iterinit() to be used to get iteration to work */
7905 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7907 *get_invlist_previous_index_addr(new_list) = 0;
7913 Perl__new_invlist_C_array(pTHX_ const UV* const list)
7915 /* Return a pointer to a newly constructed inversion list, initialized to
7916 * point to <list>, which has to be in the exact correct inversion list
7917 * form, including internal fields. Thus this is a dangerous routine that
7918 * should not be used in the wrong hands. The passed in 'list' contains
7919 * several header fields at the beginning that are not part of the
7920 * inversion list body proper */
7922 const STRLEN length = (STRLEN) list[0];
7923 const UV version_id = list[1];
7924 const bool offset = cBOOL(list[2]);
7925 #define HEADER_LENGTH 3
7926 /* If any of the above changes in any way, you must change HEADER_LENGTH
7927 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7928 * perl -E 'say int(rand 2**31-1)'
7930 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7931 data structure type, so that one being
7932 passed in can be validated to be an
7933 inversion list of the correct vintage.
7936 SV* invlist = newSV_type(SVt_INVLIST);
7938 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7940 if (version_id != INVLIST_VERSION_ID) {
7941 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7944 /* The generated array passed in includes header elements that aren't part
7945 * of the list proper, so start it just after them */
7946 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7948 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7949 shouldn't touch it */
7951 *(get_invlist_offset_addr(invlist)) = offset;
7953 /* The 'length' passed to us is the physical number of elements in the
7954 * inversion list. But if there is an offset the logical number is one
7956 invlist_set_len(invlist, length - offset, offset);
7958 invlist_set_previous_index(invlist, 0);
7960 /* Initialize the iteration pointer. */
7961 invlist_iterfinish(invlist);
7963 SvREADONLY_on(invlist);
7967 #endif /* ifndef PERL_IN_XSUB_RE */
7970 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7972 /* Grow the maximum size of an inversion list */
7974 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7976 assert(SvTYPE(invlist) == SVt_INVLIST);
7978 /* Add one to account for the zero element at the beginning which may not
7979 * be counted by the calling parameters */
7980 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7983 PERL_STATIC_INLINE void
7984 S_invlist_trim(pTHX_ SV* const invlist)
7986 PERL_ARGS_ASSERT_INVLIST_TRIM;
7988 assert(SvTYPE(invlist) == SVt_INVLIST);
7990 /* Change the length of the inversion list to how many entries it currently
7992 SvPV_shrink_to_cur((SV *) invlist);
7996 S__append_range_to_invlist(pTHX_ SV* const invlist,
7997 const UV start, const UV end)
7999 /* Subject to change or removal. Append the range from 'start' to 'end' at
8000 * the end of the inversion list. The range must be above any existing
8004 UV max = invlist_max(invlist);
8005 UV len = _invlist_len(invlist);
8008 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8010 if (len == 0) { /* Empty lists must be initialized */
8011 offset = start != 0;
8012 array = _invlist_array_init(invlist, ! offset);
8015 /* Here, the existing list is non-empty. The current max entry in the
8016 * list is generally the first value not in the set, except when the
8017 * set extends to the end of permissible values, in which case it is
8018 * the first entry in that final set, and so this call is an attempt to
8019 * append out-of-order */
8021 UV final_element = len - 1;
8022 array = invlist_array(invlist);
8023 if (array[final_element] > start
8024 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8026 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",
8027 array[final_element], start,
8028 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8031 /* Here, it is a legal append. If the new range begins with the first
8032 * value not in the set, it is extending the set, so the new first
8033 * value not in the set is one greater than the newly extended range.
8035 offset = *get_invlist_offset_addr(invlist);
8036 if (array[final_element] == start) {
8037 if (end != UV_MAX) {
8038 array[final_element] = end + 1;
8041 /* But if the end is the maximum representable on the machine,
8042 * just let the range that this would extend to have no end */
8043 invlist_set_len(invlist, len - 1, offset);
8049 /* Here the new range doesn't extend any existing set. Add it */
8051 len += 2; /* Includes an element each for the start and end of range */
8053 /* If wll overflow the existing space, extend, which may cause the array to
8056 invlist_extend(invlist, len);
8058 /* Have to set len here to avoid assert failure in invlist_array() */
8059 invlist_set_len(invlist, len, offset);
8061 array = invlist_array(invlist);
8064 invlist_set_len(invlist, len, offset);
8067 /* The next item on the list starts the range, the one after that is
8068 * one past the new range. */
8069 array[len - 2] = start;
8070 if (end != UV_MAX) {
8071 array[len - 1] = end + 1;
8074 /* But if the end is the maximum representable on the machine, just let
8075 * the range have no end */
8076 invlist_set_len(invlist, len - 1, offset);
8080 #ifndef PERL_IN_XSUB_RE
8083 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
8085 /* Searches the inversion list for the entry that contains the input code
8086 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8087 * return value is the index into the list's array of the range that
8092 IV high = _invlist_len(invlist);
8093 const IV highest_element = high - 1;
8096 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8098 /* If list is empty, return failure. */
8103 /* (We can't get the array unless we know the list is non-empty) */
8104 array = invlist_array(invlist);
8106 mid = invlist_previous_index(invlist);
8107 assert(mid >=0 && mid <= highest_element);
8109 /* <mid> contains the cache of the result of the previous call to this
8110 * function (0 the first time). See if this call is for the same result,
8111 * or if it is for mid-1. This is under the theory that calls to this
8112 * function will often be for related code points that are near each other.
8113 * And benchmarks show that caching gives better results. We also test
8114 * here if the code point is within the bounds of the list. These tests
8115 * replace others that would have had to be made anyway to make sure that
8116 * the array bounds were not exceeded, and these give us extra information
8117 * at the same time */
8118 if (cp >= array[mid]) {
8119 if (cp >= array[highest_element]) {
8120 return highest_element;
8123 /* Here, array[mid] <= cp < array[highest_element]. This means that
8124 * the final element is not the answer, so can exclude it; it also
8125 * means that <mid> is not the final element, so can refer to 'mid + 1'
8127 if (cp < array[mid + 1]) {
8133 else { /* cp < aray[mid] */
8134 if (cp < array[0]) { /* Fail if outside the array */
8138 if (cp >= array[mid - 1]) {
8143 /* Binary search. What we are looking for is <i> such that
8144 * array[i] <= cp < array[i+1]
8145 * The loop below converges on the i+1. Note that there may not be an
8146 * (i+1)th element in the array, and things work nonetheless */
8147 while (low < high) {
8148 mid = (low + high) / 2;
8149 assert(mid <= highest_element);
8150 if (array[mid] <= cp) { /* cp >= array[mid] */
8153 /* We could do this extra test to exit the loop early.
8154 if (cp < array[low]) {
8159 else { /* cp < array[mid] */
8166 invlist_set_previous_index(invlist, high);
8171 Perl__invlist_populate_swatch(pTHX_ SV* const invlist,
8172 const UV start, const UV end, U8* swatch)
8174 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8175 * but is used when the swash has an inversion list. This makes this much
8176 * faster, as it uses a binary search instead of a linear one. This is
8177 * intimately tied to that function, and perhaps should be in utf8.c,
8178 * except it is intimately tied to inversion lists as well. It assumes
8179 * that <swatch> is all 0's on input */
8182 const IV len = _invlist_len(invlist);
8186 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8188 if (len == 0) { /* Empty inversion list */
8192 array = invlist_array(invlist);
8194 /* Find which element it is */
8195 i = _invlist_search(invlist, start);
8197 /* We populate from <start> to <end> */
8198 while (current < end) {
8201 /* The inversion list gives the results for every possible code point
8202 * after the first one in the list. Only those ranges whose index is
8203 * even are ones that the inversion list matches. For the odd ones,
8204 * and if the initial code point is not in the list, we have to skip
8205 * forward to the next element */
8206 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8208 if (i >= len) { /* Finished if beyond the end of the array */
8212 if (current >= end) { /* Finished if beyond the end of what we
8214 if (LIKELY(end < UV_MAX)) {
8218 /* We get here when the upper bound is the maximum
8219 * representable on the machine, and we are looking for just
8220 * that code point. Have to special case it */
8222 goto join_end_of_list;
8225 assert(current >= start);
8227 /* The current range ends one below the next one, except don't go past
8230 upper = (i < len && array[i] < end) ? array[i] : end;
8232 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8233 * for each code point in it */
8234 for (; current < upper; current++) {
8235 const STRLEN offset = (STRLEN)(current - start);
8236 swatch[offset >> 3] |= 1 << (offset & 7);
8241 /* Quit if at the end of the list */
8244 /* But first, have to deal with the highest possible code point on
8245 * the platform. The previous code assumes that <end> is one
8246 * beyond where we want to populate, but that is impossible at the
8247 * platform's infinity, so have to handle it specially */
8248 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8250 const STRLEN offset = (STRLEN)(end - start);
8251 swatch[offset >> 3] |= 1 << (offset & 7);
8256 /* Advance to the next range, which will be for code points not in the
8265 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8266 const bool complement_b, SV** output)
8268 /* Take the union of two inversion lists and point <output> to it. *output
8269 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8270 * the reference count to that list will be decremented if not already a
8271 * temporary (mortal); otherwise *output will be made correspondingly
8272 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8273 * second list is returned. If <complement_b> is TRUE, the union is taken
8274 * of the complement (inversion) of <b> instead of b itself.
8276 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8277 * Richard Gillam, published by Addison-Wesley, and explained at some
8278 * length there. The preface says to incorporate its examples into your
8279 * code at your own risk.
8281 * The algorithm is like a merge sort.
8283 * XXX A potential performance improvement is to keep track as we go along
8284 * if only one of the inputs contributes to the result, meaning the other
8285 * is a subset of that one. In that case, we can skip the final copy and
8286 * return the larger of the input lists, but then outside code might need
8287 * to keep track of whether to free the input list or not */
8289 const UV* array_a; /* a's array */
8291 UV len_a; /* length of a's array */
8294 SV* u; /* the resulting union */
8298 UV i_a = 0; /* current index into a's array */
8302 /* running count, as explained in the algorithm source book; items are
8303 * stopped accumulating and are output when the count changes to/from 0.
8304 * The count is incremented when we start a range that's in the set, and
8305 * decremented when we start a range that's not in the set. So its range
8306 * is 0 to 2. Only when the count is zero is something not in the set.
8310 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8313 /* If either one is empty, the union is the other one */
8314 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8315 bool make_temp = FALSE; /* Should we mortalize the result? */
8319 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8325 *output = invlist_clone(b);
8327 _invlist_invert(*output);
8329 } /* else *output already = b; */
8332 sv_2mortal(*output);
8336 else if ((len_b = _invlist_len(b)) == 0) {
8337 bool make_temp = FALSE;
8339 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8344 /* The complement of an empty list is a list that has everything in it,
8345 * so the union with <a> includes everything too */
8348 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8352 *output = _new_invlist(1);
8353 _append_range_to_invlist(*output, 0, UV_MAX);
8355 else if (*output != a) {
8356 *output = invlist_clone(a);
8358 /* else *output already = a; */
8361 sv_2mortal(*output);
8366 /* Here both lists exist and are non-empty */
8367 array_a = invlist_array(a);
8368 array_b = invlist_array(b);
8370 /* If are to take the union of 'a' with the complement of b, set it
8371 * up so are looking at b's complement. */
8374 /* To complement, we invert: if the first element is 0, remove it. To
8375 * do this, we just pretend the array starts one later */
8376 if (array_b[0] == 0) {
8382 /* But if the first element is not zero, we pretend the list starts
8383 * at the 0 that is always stored immediately before the array. */
8389 /* Size the union for the worst case: that the sets are completely
8391 u = _new_invlist(len_a + len_b);
8393 /* Will contain U+0000 if either component does */
8394 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8395 || (len_b > 0 && array_b[0] == 0));
8397 /* Go through each list item by item, stopping when exhausted one of
8399 while (i_a < len_a && i_b < len_b) {
8400 UV cp; /* The element to potentially add to the union's array */
8401 bool cp_in_set; /* is it in the the input list's set or not */
8403 /* We need to take one or the other of the two inputs for the union.
8404 * Since we are merging two sorted lists, we take the smaller of the
8405 * next items. In case of a tie, we take the one that is in its set
8406 * first. If we took one not in the set first, it would decrement the
8407 * count, possibly to 0 which would cause it to be output as ending the
8408 * range, and the next time through we would take the same number, and
8409 * output it again as beginning the next range. By doing it the
8410 * opposite way, there is no possibility that the count will be
8411 * momentarily decremented to 0, and thus the two adjoining ranges will
8412 * be seamlessly merged. (In a tie and both are in the set or both not
8413 * in the set, it doesn't matter which we take first.) */
8414 if (array_a[i_a] < array_b[i_b]
8415 || (array_a[i_a] == array_b[i_b]
8416 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8418 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8422 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8423 cp = array_b[i_b++];
8426 /* Here, have chosen which of the two inputs to look at. Only output
8427 * if the running count changes to/from 0, which marks the
8428 * beginning/end of a range in that's in the set */
8431 array_u[i_u++] = cp;
8438 array_u[i_u++] = cp;
8443 /* Here, we are finished going through at least one of the lists, which
8444 * means there is something remaining in at most one. We check if the list
8445 * that hasn't been exhausted is positioned such that we are in the middle
8446 * of a range in its set or not. (i_a and i_b point to the element beyond
8447 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8448 * is potentially more to output.
8449 * There are four cases:
8450 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8451 * in the union is entirely from the non-exhausted set.
8452 * 2) Both were in their sets, count is 2. Nothing further should
8453 * be output, as everything that remains will be in the exhausted
8454 * list's set, hence in the union; decrementing to 1 but not 0 insures
8456 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8457 * Nothing further should be output because the union includes
8458 * everything from the exhausted set. Not decrementing ensures that.
8459 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8460 * decrementing to 0 insures that we look at the remainder of the
8461 * non-exhausted set */
8462 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8463 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8468 /* The final length is what we've output so far, plus what else is about to
8469 * be output. (If 'count' is non-zero, then the input list we exhausted
8470 * has everything remaining up to the machine's limit in its set, and hence
8471 * in the union, so there will be no further output. */
8474 /* At most one of the subexpressions will be non-zero */
8475 len_u += (len_a - i_a) + (len_b - i_b);
8478 /* Set result to final length, which can change the pointer to array_u, so
8480 if (len_u != _invlist_len(u)) {
8481 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8483 array_u = invlist_array(u);
8486 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8487 * the other) ended with everything above it not in its set. That means
8488 * that the remaining part of the union is precisely the same as the
8489 * non-exhausted list, so can just copy it unchanged. (If both list were
8490 * exhausted at the same time, then the operations below will be both 0.)
8493 IV copy_count; /* At most one will have a non-zero copy count */
8494 if ((copy_count = len_a - i_a) > 0) {
8495 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8497 else if ((copy_count = len_b - i_b) > 0) {
8498 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8502 /* We may be removing a reference to one of the inputs. If so, the output
8503 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8504 * count decremented) */
8505 if (a == *output || b == *output) {
8506 assert(! invlist_is_iterating(*output));
8507 if ((SvTEMP(*output))) {
8511 SvREFCNT_dec_NN(*output);
8521 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8522 const bool complement_b, SV** i)
8524 /* Take the intersection of two inversion lists and point <i> to it. *i
8525 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8526 * the reference count to that list will be decremented if not already a
8527 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8528 * The first list, <a>, may be NULL, in which case an empty list is
8529 * returned. If <complement_b> is TRUE, the result will be the
8530 * intersection of <a> and the complement (or inversion) of <b> instead of
8533 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8534 * Richard Gillam, published by Addison-Wesley, and explained at some
8535 * length there. The preface says to incorporate its examples into your
8536 * code at your own risk. In fact, it had bugs
8538 * The algorithm is like a merge sort, and is essentially the same as the
8542 const UV* array_a; /* a's array */
8544 UV len_a; /* length of a's array */
8547 SV* r; /* the resulting intersection */
8551 UV i_a = 0; /* current index into a's array */
8555 /* running count, as explained in the algorithm source book; items are
8556 * stopped accumulating and are output when the count changes to/from 2.
8557 * The count is incremented when we start a range that's in the set, and
8558 * decremented when we start a range that's not in the set. So its range
8559 * is 0 to 2. Only when the count is 2 is something in the intersection.
8563 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8566 /* Special case if either one is empty */
8567 len_a = (a == NULL) ? 0 : _invlist_len(a);
8568 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8569 bool make_temp = FALSE;
8571 if (len_a != 0 && complement_b) {
8573 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8574 * be empty. Here, also we are using 'b's complement, which hence
8575 * must be every possible code point. Thus the intersection is
8579 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8584 *i = invlist_clone(a);
8586 /* else *i is already 'a' */
8594 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8595 * intersection must be empty */
8597 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8602 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8606 *i = _new_invlist(0);
8614 /* Here both lists exist and are non-empty */
8615 array_a = invlist_array(a);
8616 array_b = invlist_array(b);
8618 /* If are to take the intersection of 'a' with the complement of b, set it
8619 * up so are looking at b's complement. */
8622 /* To complement, we invert: if the first element is 0, remove it. To
8623 * do this, we just pretend the array starts one later */
8624 if (array_b[0] == 0) {
8630 /* But if the first element is not zero, we pretend the list starts
8631 * at the 0 that is always stored immediately before the array. */
8637 /* Size the intersection for the worst case: that the intersection ends up
8638 * fragmenting everything to be completely disjoint */
8639 r= _new_invlist(len_a + len_b);
8641 /* Will contain U+0000 iff both components do */
8642 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8643 && len_b > 0 && array_b[0] == 0);
8645 /* Go through each list item by item, stopping when exhausted one of
8647 while (i_a < len_a && i_b < len_b) {
8648 UV cp; /* The element to potentially add to the intersection's
8650 bool cp_in_set; /* Is it in the input list's set or not */
8652 /* We need to take one or the other of the two inputs for the
8653 * intersection. Since we are merging two sorted lists, we take the
8654 * smaller of the next items. In case of a tie, we take the one that
8655 * is not in its set first (a difference from the union algorithm). If
8656 * we took one in the set first, it would increment the count, possibly
8657 * to 2 which would cause it to be output as starting a range in the
8658 * intersection, and the next time through we would take that same
8659 * number, and output it again as ending the set. By doing it the
8660 * opposite of this, there is no possibility that the count will be
8661 * momentarily incremented to 2. (In a tie and both are in the set or
8662 * both not in the set, it doesn't matter which we take first.) */
8663 if (array_a[i_a] < array_b[i_b]
8664 || (array_a[i_a] == array_b[i_b]
8665 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8667 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8671 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8675 /* Here, have chosen which of the two inputs to look at. Only output
8676 * if the running count changes to/from 2, which marks the
8677 * beginning/end of a range that's in the intersection */
8681 array_r[i_r++] = cp;
8686 array_r[i_r++] = cp;
8692 /* Here, we are finished going through at least one of the lists, which
8693 * means there is something remaining in at most one. We check if the list
8694 * that has been exhausted is positioned such that we are in the middle
8695 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8696 * the ones we care about.) There are four cases:
8697 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8698 * nothing left in the intersection.
8699 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8700 * above 2. What should be output is exactly that which is in the
8701 * non-exhausted set, as everything it has is also in the intersection
8702 * set, and everything it doesn't have can't be in the intersection
8703 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8704 * gets incremented to 2. Like the previous case, the intersection is
8705 * everything that remains in the non-exhausted set.
8706 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8707 * remains 1. And the intersection has nothing more. */
8708 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8709 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8714 /* The final length is what we've output so far plus what else is in the
8715 * intersection. At most one of the subexpressions below will be non-zero
8719 len_r += (len_a - i_a) + (len_b - i_b);
8722 /* Set result to final length, which can change the pointer to array_r, so
8724 if (len_r != _invlist_len(r)) {
8725 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8727 array_r = invlist_array(r);
8730 /* Finish outputting any remaining */
8731 if (count >= 2) { /* At most one will have a non-zero copy count */
8733 if ((copy_count = len_a - i_a) > 0) {
8734 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8736 else if ((copy_count = len_b - i_b) > 0) {
8737 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8741 /* We may be removing a reference to one of the inputs. If so, the output
8742 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8743 * count decremented) */
8744 if (a == *i || b == *i) {
8745 assert(! invlist_is_iterating(*i));
8750 SvREFCNT_dec_NN(*i);
8760 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8762 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8763 * set. A pointer to the inversion list is returned. This may actually be
8764 * a new list, in which case the passed in one has been destroyed. The
8765 * passed in inversion list can be NULL, in which case a new one is created
8766 * with just the one range in it */
8771 if (invlist == NULL) {
8772 invlist = _new_invlist(2);
8776 len = _invlist_len(invlist);
8779 /* If comes after the final entry actually in the list, can just append it
8782 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8783 && start >= invlist_array(invlist)[len - 1]))
8785 _append_range_to_invlist(invlist, start, end);
8789 /* Here, can't just append things, create and return a new inversion list
8790 * which is the union of this range and the existing inversion list */
8791 range_invlist = _new_invlist(2);
8792 _append_range_to_invlist(range_invlist, start, end);
8794 _invlist_union(invlist, range_invlist, &invlist);
8796 /* The temporary can be freed */
8797 SvREFCNT_dec_NN(range_invlist);
8803 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
8804 UV** other_elements_ptr)
8806 /* Create and return an inversion list whose contents are to be populated
8807 * by the caller. The caller gives the number of elements (in 'size') and
8808 * the very first element ('element0'). This function will set
8809 * '*other_elements_ptr' to an array of UVs, where the remaining elements
8812 * Obviously there is some trust involved that the caller will properly
8813 * fill in the other elements of the array.
8815 * (The first element needs to be passed in, as the underlying code does
8816 * things differently depending on whether it is zero or non-zero) */
8818 SV* invlist = _new_invlist(size);
8821 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
8823 _append_range_to_invlist(invlist, element0, element0);
8824 offset = *get_invlist_offset_addr(invlist);
8826 invlist_set_len(invlist, size, offset);
8827 *other_elements_ptr = invlist_array(invlist) + 1;
8833 PERL_STATIC_INLINE SV*
8834 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8835 return _add_range_to_invlist(invlist, cp, cp);
8838 #ifndef PERL_IN_XSUB_RE
8840 Perl__invlist_invert(pTHX_ SV* const invlist)
8842 /* Complement the input inversion list. This adds a 0 if the list didn't
8843 * have a zero; removes it otherwise. As described above, the data
8844 * structure is set up so that this is very efficient */
8846 PERL_ARGS_ASSERT__INVLIST_INVERT;
8848 assert(! invlist_is_iterating(invlist));
8850 /* The inverse of matching nothing is matching everything */
8851 if (_invlist_len(invlist) == 0) {
8852 _append_range_to_invlist(invlist, 0, UV_MAX);
8856 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8861 PERL_STATIC_INLINE SV*
8862 S_invlist_clone(pTHX_ SV* const invlist)
8865 /* Return a new inversion list that is a copy of the input one, which is
8866 * unchanged. The new list will not be mortal even if the old one was. */
8868 /* Need to allocate extra space to accommodate Perl's addition of a
8869 * trailing NUL to SvPV's, since it thinks they are always strings */
8870 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8871 STRLEN physical_length = SvCUR(invlist);
8872 bool offset = *(get_invlist_offset_addr(invlist));
8874 PERL_ARGS_ASSERT_INVLIST_CLONE;
8876 *(get_invlist_offset_addr(new_invlist)) = offset;
8877 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8878 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8883 PERL_STATIC_INLINE STRLEN*
8884 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8886 /* Return the address of the UV that contains the current iteration
8889 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8891 assert(SvTYPE(invlist) == SVt_INVLIST);
8893 return &(((XINVLIST*) SvANY(invlist))->iterator);
8896 PERL_STATIC_INLINE void
8897 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8899 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8901 *get_invlist_iter_addr(invlist) = 0;
8904 PERL_STATIC_INLINE void
8905 S_invlist_iterfinish(pTHX_ SV* invlist)
8907 /* Terminate iterator for invlist. This is to catch development errors.
8908 * Any iteration that is interrupted before completed should call this
8909 * function. Functions that add code points anywhere else but to the end
8910 * of an inversion list assert that they are not in the middle of an
8911 * iteration. If they were, the addition would make the iteration
8912 * problematical: if the iteration hadn't reached the place where things
8913 * were being added, it would be ok */
8915 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8917 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8921 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8923 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8924 * This call sets in <*start> and <*end>, the next range in <invlist>.
8925 * Returns <TRUE> if successful and the next call will return the next
8926 * range; <FALSE> if was already at the end of the list. If the latter,
8927 * <*start> and <*end> are unchanged, and the next call to this function
8928 * will start over at the beginning of the list */
8930 STRLEN* pos = get_invlist_iter_addr(invlist);
8931 UV len = _invlist_len(invlist);
8934 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8937 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8941 array = invlist_array(invlist);
8943 *start = array[(*pos)++];
8949 *end = array[(*pos)++] - 1;
8955 PERL_STATIC_INLINE bool
8956 S_invlist_is_iterating(pTHX_ SV* const invlist)
8958 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8960 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8963 PERL_STATIC_INLINE UV
8964 S_invlist_highest(pTHX_ SV* const invlist)
8966 /* Returns the highest code point that matches an inversion list. This API
8967 * has an ambiguity, as it returns 0 under either the highest is actually
8968 * 0, or if the list is empty. If this distinction matters to you, check
8969 * for emptiness before calling this function */
8971 UV len = _invlist_len(invlist);
8974 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8980 array = invlist_array(invlist);
8982 /* The last element in the array in the inversion list always starts a
8983 * range that goes to infinity. That range may be for code points that are
8984 * matched in the inversion list, or it may be for ones that aren't
8985 * matched. In the latter case, the highest code point in the set is one
8986 * less than the beginning of this range; otherwise it is the final element
8987 * of this range: infinity */
8988 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8990 : array[len - 1] - 1;
8993 #ifndef PERL_IN_XSUB_RE
8995 Perl__invlist_contents(pTHX_ SV* const invlist)
8997 /* Get the contents of an inversion list into a string SV so that they can
8998 * be printed out. It uses the format traditionally done for debug tracing
9002 SV* output = newSVpvs("\n");
9004 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9006 assert(! invlist_is_iterating(invlist));
9008 invlist_iterinit(invlist);
9009 while (invlist_iternext(invlist, &start, &end)) {
9010 if (end == UV_MAX) {
9011 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9013 else if (end != start) {
9014 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9018 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9026 #ifndef PERL_IN_XSUB_RE
9028 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9029 const char * const indent, SV* const invlist)
9031 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9032 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9033 * the string 'indent'. The output looks like this:
9034 [0] 0x000A .. 0x000D
9036 [4] 0x2028 .. 0x2029
9037 [6] 0x3104 .. INFINITY
9038 * This means that the first range of code points matched by the list are
9039 * 0xA through 0xD; the second range contains only the single code point
9040 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9041 * are used to define each range (except if the final range extends to
9042 * infinity, only a single element is needed). The array index of the
9043 * first element for the corresponding range is given in brackets. */
9048 PERL_ARGS_ASSERT__INVLIST_DUMP;
9050 if (invlist_is_iterating(invlist)) {
9051 Perl_dump_indent(aTHX_ level, file,
9052 "%sCan't dump inversion list because is in middle of iterating\n",
9057 invlist_iterinit(invlist);
9058 while (invlist_iternext(invlist, &start, &end)) {
9059 if (end == UV_MAX) {
9060 Perl_dump_indent(aTHX_ level, file,
9061 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9062 indent, (UV)count, start);
9064 else if (end != start) {
9065 Perl_dump_indent(aTHX_ level, file,
9066 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9067 indent, (UV)count, start, end);
9070 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9071 indent, (UV)count, start);
9078 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9080 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9082 /* Return a boolean as to if the two passed in inversion lists are
9083 * identical. The final argument, if TRUE, says to take the complement of
9084 * the second inversion list before doing the comparison */
9086 const UV* array_a = invlist_array(a);
9087 const UV* array_b = invlist_array(b);
9088 UV len_a = _invlist_len(a);
9089 UV len_b = _invlist_len(b);
9091 UV i = 0; /* current index into the arrays */
9092 bool retval = TRUE; /* Assume are identical until proven otherwise */
9094 PERL_ARGS_ASSERT__INVLISTEQ;
9096 /* If are to compare 'a' with the complement of b, set it
9097 * up so are looking at b's complement. */
9100 /* The complement of nothing is everything, so <a> would have to have
9101 * just one element, starting at zero (ending at infinity) */
9103 return (len_a == 1 && array_a[0] == 0);
9105 else if (array_b[0] == 0) {
9107 /* Otherwise, to complement, we invert. Here, the first element is
9108 * 0, just remove it. To do this, we just pretend the array starts
9116 /* But if the first element is not zero, we pretend the list starts
9117 * at the 0 that is always stored immediately before the array. */
9123 /* Make sure that the lengths are the same, as well as the final element
9124 * before looping through the remainder. (Thus we test the length, final,
9125 * and first elements right off the bat) */
9126 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9129 else for (i = 0; i < len_a - 1; i++) {
9130 if (array_a[i] != array_b[i]) {
9140 #undef HEADER_LENGTH
9141 #undef TO_INTERNAL_SIZE
9142 #undef FROM_INTERNAL_SIZE
9143 #undef INVLIST_VERSION_ID
9145 /* End of inversion list object */
9148 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9150 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9151 * constructs, and updates RExC_flags with them. On input, RExC_parse
9152 * should point to the first flag; it is updated on output to point to the
9153 * final ')' or ':'. There needs to be at least one flag, or this will
9156 /* for (?g), (?gc), and (?o) warnings; warning
9157 about (?c) will warn about (?g) -- japhy */
9159 #define WASTED_O 0x01
9160 #define WASTED_G 0x02
9161 #define WASTED_C 0x04
9162 #define WASTED_GC (WASTED_G|WASTED_C)
9163 I32 wastedflags = 0x00;
9164 U32 posflags = 0, negflags = 0;
9165 U32 *flagsp = &posflags;
9166 char has_charset_modifier = '\0';
9168 bool has_use_defaults = FALSE;
9169 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9171 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9173 /* '^' as an initial flag sets certain defaults */
9174 if (UCHARAT(RExC_parse) == '^') {
9176 has_use_defaults = TRUE;
9177 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9178 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9179 ? REGEX_UNICODE_CHARSET
9180 : REGEX_DEPENDS_CHARSET);
9183 cs = get_regex_charset(RExC_flags);
9184 if (cs == REGEX_DEPENDS_CHARSET
9185 && (RExC_utf8 || RExC_uni_semantics))
9187 cs = REGEX_UNICODE_CHARSET;
9190 while (*RExC_parse) {
9191 /* && strchr("iogcmsx", *RExC_parse) */
9192 /* (?g), (?gc) and (?o) are useless here
9193 and must be globally applied -- japhy */
9194 switch (*RExC_parse) {
9196 /* Code for the imsx flags */
9197 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
9199 case LOCALE_PAT_MOD:
9200 if (has_charset_modifier) {
9201 goto excess_modifier;
9203 else if (flagsp == &negflags) {
9206 cs = REGEX_LOCALE_CHARSET;
9207 has_charset_modifier = LOCALE_PAT_MOD;
9209 case UNICODE_PAT_MOD:
9210 if (has_charset_modifier) {
9211 goto excess_modifier;
9213 else if (flagsp == &negflags) {
9216 cs = REGEX_UNICODE_CHARSET;
9217 has_charset_modifier = UNICODE_PAT_MOD;
9219 case ASCII_RESTRICT_PAT_MOD:
9220 if (flagsp == &negflags) {
9223 if (has_charset_modifier) {
9224 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9225 goto excess_modifier;
9227 /* Doubled modifier implies more restricted */
9228 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9231 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9233 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9235 case DEPENDS_PAT_MOD:
9236 if (has_use_defaults) {
9237 goto fail_modifiers;
9239 else if (flagsp == &negflags) {
9242 else if (has_charset_modifier) {
9243 goto excess_modifier;
9246 /* The dual charset means unicode semantics if the
9247 * pattern (or target, not known until runtime) are
9248 * utf8, or something in the pattern indicates unicode
9250 cs = (RExC_utf8 || RExC_uni_semantics)
9251 ? REGEX_UNICODE_CHARSET
9252 : REGEX_DEPENDS_CHARSET;
9253 has_charset_modifier = DEPENDS_PAT_MOD;
9257 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9258 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9260 else if (has_charset_modifier == *(RExC_parse - 1)) {
9261 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9265 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9270 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9273 case ONCE_PAT_MOD: /* 'o' */
9274 case GLOBAL_PAT_MOD: /* 'g' */
9275 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9276 const I32 wflagbit = *RExC_parse == 'o'
9279 if (! (wastedflags & wflagbit) ) {
9280 wastedflags |= wflagbit;
9281 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9284 "Useless (%s%c) - %suse /%c modifier",
9285 flagsp == &negflags ? "?-" : "?",
9287 flagsp == &negflags ? "don't " : "",
9294 case CONTINUE_PAT_MOD: /* 'c' */
9295 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9296 if (! (wastedflags & WASTED_C) ) {
9297 wastedflags |= WASTED_GC;
9298 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9301 "Useless (%sc) - %suse /gc modifier",
9302 flagsp == &negflags ? "?-" : "?",
9303 flagsp == &negflags ? "don't " : ""
9308 case KEEPCOPY_PAT_MOD: /* 'p' */
9309 if (flagsp == &negflags) {
9311 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9313 *flagsp |= RXf_PMf_KEEPCOPY;
9317 /* A flag is a default iff it is following a minus, so
9318 * if there is a minus, it means will be trying to
9319 * re-specify a default which is an error */
9320 if (has_use_defaults || flagsp == &negflags) {
9321 goto fail_modifiers;
9324 wastedflags = 0; /* reset so (?g-c) warns twice */
9328 RExC_flags |= posflags;
9329 RExC_flags &= ~negflags;
9330 set_regex_charset(&RExC_flags, cs);
9331 if (RExC_flags & RXf_PMf_FOLD) {
9332 RExC_contains_i = 1;
9338 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9339 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9340 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9341 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9350 - reg - regular expression, i.e. main body or parenthesized thing
9352 * Caller must absorb opening parenthesis.
9354 * Combining parenthesis handling with the base level of regular expression
9355 * is a trifle forced, but the need to tie the tails of the branches to what
9356 * follows makes it hard to avoid.
9358 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9360 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9362 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9365 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9366 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9367 needs to be restarted.
9368 Otherwise would only return NULL if regbranch() returns NULL, which
9371 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9372 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9373 * 2 is like 1, but indicates that nextchar() has been called to advance
9374 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9375 * this flag alerts us to the need to check for that */
9378 regnode *ret; /* Will be the head of the group. */
9381 regnode *ender = NULL;
9384 U32 oregflags = RExC_flags;
9385 bool have_branch = 0;
9387 I32 freeze_paren = 0;
9388 I32 after_freeze = 0;
9390 char * parse_start = RExC_parse; /* MJD */
9391 char * const oregcomp_parse = RExC_parse;
9393 GET_RE_DEBUG_FLAGS_DECL;
9395 PERL_ARGS_ASSERT_REG;
9396 DEBUG_PARSE("reg ");
9398 *flagp = 0; /* Tentatively. */
9401 /* Make an OPEN node, if parenthesized. */
9404 /* Under /x, space and comments can be gobbled up between the '(' and
9405 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9406 * intervening space, as the sequence is a token, and a token should be
9408 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9410 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9411 char *start_verb = RExC_parse;
9412 STRLEN verb_len = 0;
9413 char *start_arg = NULL;
9414 unsigned char op = 0;
9416 int internal_argval = 0; /* internal_argval is only useful if
9419 if (has_intervening_patws && SIZE_ONLY) {
9420 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
9422 while ( *RExC_parse && *RExC_parse != ')' ) {
9423 if ( *RExC_parse == ':' ) {
9424 start_arg = RExC_parse + 1;
9430 verb_len = RExC_parse - start_verb;
9433 while ( *RExC_parse && *RExC_parse != ')' )
9435 if ( *RExC_parse != ')' )
9436 vFAIL("Unterminated verb pattern argument");
9437 if ( RExC_parse == start_arg )
9440 if ( *RExC_parse != ')' )
9441 vFAIL("Unterminated verb pattern");
9444 switch ( *start_verb ) {
9445 case 'A': /* (*ACCEPT) */
9446 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9448 internal_argval = RExC_nestroot;
9451 case 'C': /* (*COMMIT) */
9452 if ( memEQs(start_verb,verb_len,"COMMIT") )
9455 case 'F': /* (*FAIL) */
9456 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9461 case ':': /* (*:NAME) */
9462 case 'M': /* (*MARK:NAME) */
9463 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9468 case 'P': /* (*PRUNE) */
9469 if ( memEQs(start_verb,verb_len,"PRUNE") )
9472 case 'S': /* (*SKIP) */
9473 if ( memEQs(start_verb,verb_len,"SKIP") )
9476 case 'T': /* (*THEN) */
9477 /* [19:06] <TimToady> :: is then */
9478 if ( memEQs(start_verb,verb_len,"THEN") ) {
9480 RExC_seen |= REG_CUTGROUP_SEEN;
9485 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9487 "Unknown verb pattern '%"UTF8f"'",
9488 UTF8fARG(UTF, verb_len, start_verb));
9491 if ( start_arg && internal_argval ) {
9492 vFAIL3("Verb pattern '%.*s' may not have an argument",
9493 verb_len, start_verb);
9494 } else if ( argok < 0 && !start_arg ) {
9495 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9496 verb_len, start_verb);
9498 ret = reganode(pRExC_state, op, internal_argval);
9499 if ( ! internal_argval && ! SIZE_ONLY ) {
9501 SV *sv = newSVpvn( start_arg,
9502 RExC_parse - start_arg);
9503 ARG(ret) = add_data( pRExC_state,
9505 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9512 if (!internal_argval)
9513 RExC_seen |= REG_VERBARG_SEEN;
9514 } else if ( start_arg ) {
9515 vFAIL3("Verb pattern '%.*s' may not have an argument",
9516 verb_len, start_verb);
9518 ret = reg_node(pRExC_state, op);
9520 nextchar(pRExC_state);
9523 else if (*RExC_parse == '?') { /* (?...) */
9524 bool is_logical = 0;
9525 const char * const seqstart = RExC_parse;
9526 if (has_intervening_patws && SIZE_ONLY) {
9527 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
9531 paren = *RExC_parse++;
9532 ret = NULL; /* For look-ahead/behind. */
9535 case 'P': /* (?P...) variants for those used to PCRE/Python */
9536 paren = *RExC_parse++;
9537 if ( paren == '<') /* (?P<...>) named capture */
9539 else if (paren == '>') { /* (?P>name) named recursion */
9540 goto named_recursion;
9542 else if (paren == '=') { /* (?P=...) named backref */
9543 /* this pretty much dupes the code for \k<NAME> in
9544 * regatom(), if you change this make sure you change that
9546 char* name_start = RExC_parse;
9548 SV *sv_dat = reg_scan_name(pRExC_state,
9549 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9550 if (RExC_parse == name_start || *RExC_parse != ')')
9551 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9552 vFAIL2("Sequence %.3s... not terminated",parse_start);
9555 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9556 RExC_rxi->data->data[num]=(void*)sv_dat;
9557 SvREFCNT_inc_simple_void(sv_dat);
9560 ret = reganode(pRExC_state,
9563 : (ASCII_FOLD_RESTRICTED)
9565 : (AT_LEAST_UNI_SEMANTICS)
9573 Set_Node_Offset(ret, parse_start+1);
9574 Set_Node_Cur_Length(ret, parse_start);
9576 nextchar(pRExC_state);
9580 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9581 vFAIL3("Sequence (%.*s...) not recognized",
9582 RExC_parse-seqstart, seqstart);
9584 case '<': /* (?<...) */
9585 if (*RExC_parse == '!')
9587 else if (*RExC_parse != '=')
9593 case '\'': /* (?'...') */
9594 name_start= RExC_parse;
9595 svname = reg_scan_name(pRExC_state,
9596 SIZE_ONLY /* reverse test from the others */
9597 ? REG_RSN_RETURN_NAME
9598 : REG_RSN_RETURN_NULL);
9599 if (RExC_parse == name_start || *RExC_parse != paren)
9600 vFAIL2("Sequence (?%c... not terminated",
9601 paren=='>' ? '<' : paren);
9605 if (!svname) /* shouldn't happen */
9607 "panic: reg_scan_name returned NULL");
9608 if (!RExC_paren_names) {
9609 RExC_paren_names= newHV();
9610 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9612 RExC_paren_name_list= newAV();
9613 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
9616 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
9618 sv_dat = HeVAL(he_str);
9620 /* croak baby croak */
9622 "panic: paren_name hash element allocation failed");
9623 } else if ( SvPOK(sv_dat) ) {
9624 /* (?|...) can mean we have dupes so scan to check
9625 its already been stored. Maybe a flag indicating
9626 we are inside such a construct would be useful,
9627 but the arrays are likely to be quite small, so
9628 for now we punt -- dmq */
9629 IV count = SvIV(sv_dat);
9630 I32 *pv = (I32*)SvPVX(sv_dat);
9632 for ( i = 0 ; i < count ; i++ ) {
9633 if ( pv[i] == RExC_npar ) {
9639 pv = (I32*)SvGROW(sv_dat,
9640 SvCUR(sv_dat) + sizeof(I32)+1);
9641 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
9642 pv[count] = RExC_npar;
9643 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
9646 (void)SvUPGRADE(sv_dat,SVt_PVNV);
9647 sv_setpvn(sv_dat, (char *)&(RExC_npar),
9650 SvIV_set(sv_dat, 1);
9653 /* Yes this does cause a memory leak in debugging Perls
9655 if (!av_store(RExC_paren_name_list,
9656 RExC_npar, SvREFCNT_inc(svname)))
9657 SvREFCNT_dec_NN(svname);
9660 /*sv_dump(sv_dat);*/
9662 nextchar(pRExC_state);
9664 goto capturing_parens;
9666 RExC_seen |= REG_LOOKBEHIND_SEEN;
9667 RExC_in_lookbehind++;
9669 case '=': /* (?=...) */
9670 RExC_seen_zerolen++;
9672 case '!': /* (?!...) */
9673 RExC_seen_zerolen++;
9674 if (*RExC_parse == ')') {
9675 ret=reg_node(pRExC_state, OPFAIL);
9676 nextchar(pRExC_state);
9680 case '|': /* (?|...) */
9681 /* branch reset, behave like a (?:...) except that
9682 buffers in alternations share the same numbers */
9684 after_freeze = freeze_paren = RExC_npar;
9686 case ':': /* (?:...) */
9687 case '>': /* (?>...) */
9689 case '$': /* (?$...) */
9690 case '@': /* (?@...) */
9691 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
9693 case '#': /* (?#...) */
9694 /* XXX As soon as we disallow separating the '?' and '*' (by
9695 * spaces or (?#...) comment), it is believed that this case
9696 * will be unreachable and can be removed. See
9698 while (*RExC_parse && *RExC_parse != ')')
9700 if (*RExC_parse != ')')
9701 FAIL("Sequence (?#... not terminated");
9702 nextchar(pRExC_state);
9705 case '0' : /* (?0) */
9706 case 'R' : /* (?R) */
9707 if (*RExC_parse != ')')
9708 FAIL("Sequence (?R) not terminated");
9709 ret = reg_node(pRExC_state, GOSTART);
9710 RExC_seen |= REG_GOSTART_SEEN;
9711 *flagp |= POSTPONED;
9712 nextchar(pRExC_state);
9715 { /* named and numeric backreferences */
9717 case '&': /* (?&NAME) */
9718 parse_start = RExC_parse - 1;
9721 SV *sv_dat = reg_scan_name(pRExC_state,
9722 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9723 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9725 if (RExC_parse == RExC_end || *RExC_parse != ')')
9726 vFAIL("Sequence (?&... not terminated");
9727 goto gen_recurse_regop;
9728 assert(0); /* NOT REACHED */
9730 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9732 vFAIL("Illegal pattern");
9734 goto parse_recursion;
9736 case '-': /* (?-1) */
9737 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9738 RExC_parse--; /* rewind to let it be handled later */
9742 case '1': case '2': case '3': case '4': /* (?1) */
9743 case '5': case '6': case '7': case '8': case '9':
9746 num = atoi(RExC_parse);
9747 parse_start = RExC_parse - 1; /* MJD */
9748 if (*RExC_parse == '-')
9750 while (isDIGIT(*RExC_parse))
9752 if (*RExC_parse!=')')
9753 vFAIL("Expecting close bracket");
9756 if ( paren == '-' ) {
9758 Diagram of capture buffer numbering.
9759 Top line is the normal capture buffer numbers
9760 Bottom line is the negative indexing as from
9764 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9768 num = RExC_npar + num;
9771 vFAIL("Reference to nonexistent group");
9773 } else if ( paren == '+' ) {
9774 num = RExC_npar + num - 1;
9777 ret = reganode(pRExC_state, GOSUB, num);
9779 if (num > (I32)RExC_rx->nparens) {
9781 vFAIL("Reference to nonexistent group");
9783 ARG2L_SET( ret, RExC_recurse_count++);
9785 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9786 "Recurse #%"UVuf" to %"IVdf"\n",
9787 (UV)ARG(ret), (IV)ARG2L(ret)));
9791 RExC_seen |= REG_RECURSE_SEEN;
9792 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9793 Set_Node_Offset(ret, parse_start); /* MJD */
9795 *flagp |= POSTPONED;
9796 nextchar(pRExC_state);
9798 } /* named and numeric backreferences */
9799 assert(0); /* NOT REACHED */
9801 case '?': /* (??...) */
9803 if (*RExC_parse != '{') {
9805 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9807 "Sequence (%"UTF8f"...) not recognized",
9808 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9811 *flagp |= POSTPONED;
9812 paren = *RExC_parse++;
9814 case '{': /* (?{...}) */
9817 struct reg_code_block *cb;
9819 RExC_seen_zerolen++;
9821 if ( !pRExC_state->num_code_blocks
9822 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9823 || pRExC_state->code_blocks[pRExC_state->code_index].start
9824 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9827 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9828 FAIL("panic: Sequence (?{...}): no code block found\n");
9829 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9831 /* this is a pre-compiled code block (?{...}) */
9832 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9833 RExC_parse = RExC_start + cb->end;
9836 if (cb->src_regex) {
9837 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
9838 RExC_rxi->data->data[n] =
9839 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9840 RExC_rxi->data->data[n+1] = (void*)o;
9843 n = add_data(pRExC_state,
9844 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
9845 RExC_rxi->data->data[n] = (void*)o;
9848 pRExC_state->code_index++;
9849 nextchar(pRExC_state);
9853 ret = reg_node(pRExC_state, LOGICAL);
9854 eval = reganode(pRExC_state, EVAL, n);
9857 /* for later propagation into (??{}) return value */
9858 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9860 REGTAIL(pRExC_state, ret, eval);
9861 /* deal with the length of this later - MJD */
9864 ret = reganode(pRExC_state, EVAL, n);
9865 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9866 Set_Node_Offset(ret, parse_start);
9869 case '(': /* (?(?{...})...) and (?(?=...)...) */
9872 if (RExC_parse[0] == '?') { /* (?(?...)) */
9873 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9874 || RExC_parse[1] == '<'
9875 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9879 ret = reg_node(pRExC_state, LOGICAL);
9883 tail = reg(pRExC_state, 1, &flag, depth+1);
9884 if (flag & RESTART_UTF8) {
9885 *flagp = RESTART_UTF8;
9888 REGTAIL(pRExC_state, ret, tail);
9892 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9893 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9895 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9896 char *name_start= RExC_parse++;
9898 SV *sv_dat=reg_scan_name(pRExC_state,
9899 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9900 if (RExC_parse == name_start || *RExC_parse != ch)
9901 vFAIL2("Sequence (?(%c... not terminated",
9902 (ch == '>' ? '<' : ch));
9905 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9906 RExC_rxi->data->data[num]=(void*)sv_dat;
9907 SvREFCNT_inc_simple_void(sv_dat);
9909 ret = reganode(pRExC_state,NGROUPP,num);
9910 goto insert_if_check_paren;
9912 else if (RExC_parse[0] == 'D' &&
9913 RExC_parse[1] == 'E' &&
9914 RExC_parse[2] == 'F' &&
9915 RExC_parse[3] == 'I' &&
9916 RExC_parse[4] == 'N' &&
9917 RExC_parse[5] == 'E')
9919 ret = reganode(pRExC_state,DEFINEP,0);
9922 goto insert_if_check_paren;
9924 else if (RExC_parse[0] == 'R') {
9927 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9928 parno = atoi(RExC_parse++);
9929 while (isDIGIT(*RExC_parse))
9931 } else if (RExC_parse[0] == '&') {
9934 sv_dat = reg_scan_name(pRExC_state,
9936 ? REG_RSN_RETURN_NULL
9937 : REG_RSN_RETURN_DATA);
9938 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9940 ret = reganode(pRExC_state,INSUBP,parno);
9941 goto insert_if_check_paren;
9943 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9947 parno = atoi(RExC_parse++);
9949 while (isDIGIT(*RExC_parse))
9951 ret = reganode(pRExC_state, GROUPP, parno);
9953 insert_if_check_paren:
9954 if (*(tmp = nextchar(pRExC_state)) != ')') {
9955 /* nextchar also skips comments, so undo its work
9956 * and skip over the the next character.
9959 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9960 vFAIL("Switch condition not recognized");
9963 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9964 br = regbranch(pRExC_state, &flags, 1,depth+1);
9966 if (flags & RESTART_UTF8) {
9967 *flagp = RESTART_UTF8;
9970 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9973 REGTAIL(pRExC_state, br, reganode(pRExC_state,
9975 c = *nextchar(pRExC_state);
9980 vFAIL("(?(DEFINE)....) does not allow branches");
9982 /* Fake one for optimizer. */
9983 lastbr = reganode(pRExC_state, IFTHEN, 0);
9985 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9986 if (flags & RESTART_UTF8) {
9987 *flagp = RESTART_UTF8;
9990 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9993 REGTAIL(pRExC_state, ret, lastbr);
9996 c = *nextchar(pRExC_state);
10001 vFAIL("Switch (?(condition)... contains too many branches");
10002 ender = reg_node(pRExC_state, TAIL);
10003 REGTAIL(pRExC_state, br, ender);
10005 REGTAIL(pRExC_state, lastbr, ender);
10006 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10009 REGTAIL(pRExC_state, ret, ender);
10010 RExC_size++; /* XXX WHY do we need this?!!
10011 For large programs it seems to be required
10012 but I can't figure out why. -- dmq*/
10016 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10017 vFAIL("Unknown switch condition (?(...))");
10020 case '[': /* (?[ ... ]) */
10021 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10024 RExC_parse--; /* for vFAIL to print correctly */
10025 vFAIL("Sequence (? incomplete");
10027 default: /* e.g., (?i) */
10030 parse_lparen_question_flags(pRExC_state);
10031 if (UCHARAT(RExC_parse) != ':') {
10032 nextchar(pRExC_state);
10037 nextchar(pRExC_state);
10047 ret = reganode(pRExC_state, OPEN, parno);
10049 if (!RExC_nestroot)
10050 RExC_nestroot = parno;
10051 if (RExC_seen & REG_RECURSE_SEEN
10052 && !RExC_open_parens[parno-1])
10054 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10055 "Setting open paren #%"IVdf" to %d\n",
10056 (IV)parno, REG_NODE_NUM(ret)));
10057 RExC_open_parens[parno-1]= ret;
10060 Set_Node_Length(ret, 1); /* MJD */
10061 Set_Node_Offset(ret, RExC_parse); /* MJD */
10069 /* Pick up the branches, linking them together. */
10070 parse_start = RExC_parse; /* MJD */
10071 br = regbranch(pRExC_state, &flags, 1,depth+1);
10073 /* branch_len = (paren != 0); */
10076 if (flags & RESTART_UTF8) {
10077 *flagp = RESTART_UTF8;
10080 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10082 if (*RExC_parse == '|') {
10083 if (!SIZE_ONLY && RExC_extralen) {
10084 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10087 reginsert(pRExC_state, BRANCH, br, depth+1);
10088 Set_Node_Length(br, paren != 0);
10089 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10093 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10095 else if (paren == ':') {
10096 *flagp |= flags&SIMPLE;
10098 if (is_open) { /* Starts with OPEN. */
10099 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10101 else if (paren != '?') /* Not Conditional */
10103 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10105 while (*RExC_parse == '|') {
10106 if (!SIZE_ONLY && RExC_extralen) {
10107 ender = reganode(pRExC_state, LONGJMP,0);
10109 /* Append to the previous. */
10110 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10113 RExC_extralen += 2; /* Account for LONGJMP. */
10114 nextchar(pRExC_state);
10115 if (freeze_paren) {
10116 if (RExC_npar > after_freeze)
10117 after_freeze = RExC_npar;
10118 RExC_npar = freeze_paren;
10120 br = regbranch(pRExC_state, &flags, 0, depth+1);
10123 if (flags & RESTART_UTF8) {
10124 *flagp = RESTART_UTF8;
10127 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10129 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10131 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10134 if (have_branch || paren != ':') {
10135 /* Make a closing node, and hook it on the end. */
10138 ender = reg_node(pRExC_state, TAIL);
10141 ender = reganode(pRExC_state, CLOSE, parno);
10142 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10143 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10144 "Setting close paren #%"IVdf" to %d\n",
10145 (IV)parno, REG_NODE_NUM(ender)));
10146 RExC_close_parens[parno-1]= ender;
10147 if (RExC_nestroot == parno)
10150 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10151 Set_Node_Length(ender,1); /* MJD */
10157 *flagp &= ~HASWIDTH;
10160 ender = reg_node(pRExC_state, SUCCEED);
10163 ender = reg_node(pRExC_state, END);
10165 assert(!RExC_opend); /* there can only be one! */
10166 RExC_opend = ender;
10170 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10171 SV * const mysv_val1=sv_newmortal();
10172 SV * const mysv_val2=sv_newmortal();
10173 DEBUG_PARSE_MSG("lsbr");
10174 regprop(RExC_rx, mysv_val1, lastbr, NULL);
10175 regprop(RExC_rx, mysv_val2, ender, NULL);
10176 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10177 SvPV_nolen_const(mysv_val1),
10178 (IV)REG_NODE_NUM(lastbr),
10179 SvPV_nolen_const(mysv_val2),
10180 (IV)REG_NODE_NUM(ender),
10181 (IV)(ender - lastbr)
10184 REGTAIL(pRExC_state, lastbr, ender);
10186 if (have_branch && !SIZE_ONLY) {
10187 char is_nothing= 1;
10189 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10191 /* Hook the tails of the branches to the closing node. */
10192 for (br = ret; br; br = regnext(br)) {
10193 const U8 op = PL_regkind[OP(br)];
10194 if (op == BRANCH) {
10195 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10196 if ( OP(NEXTOPER(br)) != NOTHING
10197 || regnext(NEXTOPER(br)) != ender)
10200 else if (op == BRANCHJ) {
10201 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10202 /* for now we always disable this optimisation * /
10203 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10204 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10210 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10211 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10212 SV * const mysv_val1=sv_newmortal();
10213 SV * const mysv_val2=sv_newmortal();
10214 DEBUG_PARSE_MSG("NADA");
10215 regprop(RExC_rx, mysv_val1, ret, NULL);
10216 regprop(RExC_rx, mysv_val2, ender, NULL);
10217 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10218 SvPV_nolen_const(mysv_val1),
10219 (IV)REG_NODE_NUM(ret),
10220 SvPV_nolen_const(mysv_val2),
10221 (IV)REG_NODE_NUM(ender),
10226 if (OP(ender) == TAIL) {
10231 for ( opt= br + 1; opt < ender ; opt++ )
10232 OP(opt)= OPTIMIZED;
10233 NEXT_OFF(br)= ender - br;
10241 static const char parens[] = "=!<,>";
10243 if (paren && (p = strchr(parens, paren))) {
10244 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10245 int flag = (p - parens) > 1;
10248 node = SUSPEND, flag = 0;
10249 reginsert(pRExC_state, node,ret, depth+1);
10250 Set_Node_Cur_Length(ret, parse_start);
10251 Set_Node_Offset(ret, parse_start + 1);
10253 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10257 /* Check for proper termination. */
10259 /* restore original flags, but keep (?p) */
10260 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10261 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10262 RExC_parse = oregcomp_parse;
10263 vFAIL("Unmatched (");
10266 else if (!paren && RExC_parse < RExC_end) {
10267 if (*RExC_parse == ')') {
10269 vFAIL("Unmatched )");
10272 FAIL("Junk on end of regexp"); /* "Can't happen". */
10273 assert(0); /* NOTREACHED */
10276 if (RExC_in_lookbehind) {
10277 RExC_in_lookbehind--;
10279 if (after_freeze > RExC_npar)
10280 RExC_npar = after_freeze;
10285 - regbranch - one alternative of an | operator
10287 * Implements the concatenation operator.
10289 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10293 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10297 regnode *chain = NULL;
10299 I32 flags = 0, c = 0;
10300 GET_RE_DEBUG_FLAGS_DECL;
10302 PERL_ARGS_ASSERT_REGBRANCH;
10304 DEBUG_PARSE("brnc");
10309 if (!SIZE_ONLY && RExC_extralen)
10310 ret = reganode(pRExC_state, BRANCHJ,0);
10312 ret = reg_node(pRExC_state, BRANCH);
10313 Set_Node_Length(ret, 1);
10317 if (!first && SIZE_ONLY)
10318 RExC_extralen += 1; /* BRANCHJ */
10320 *flagp = WORST; /* Tentatively. */
10323 nextchar(pRExC_state);
10324 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10325 flags &= ~TRYAGAIN;
10326 latest = regpiece(pRExC_state, &flags,depth+1);
10327 if (latest == NULL) {
10328 if (flags & TRYAGAIN)
10330 if (flags & RESTART_UTF8) {
10331 *flagp = RESTART_UTF8;
10334 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10336 else if (ret == NULL)
10338 *flagp |= flags&(HASWIDTH|POSTPONED);
10339 if (chain == NULL) /* First piece. */
10340 *flagp |= flags&SPSTART;
10343 REGTAIL(pRExC_state, chain, latest);
10348 if (chain == NULL) { /* Loop ran zero times. */
10349 chain = reg_node(pRExC_state, NOTHING);
10354 *flagp |= flags&SIMPLE;
10361 - regpiece - something followed by possible [*+?]
10363 * Note that the branching code sequences used for ? and the general cases
10364 * of * and + are somewhat optimized: they use the same NOTHING node as
10365 * both the endmarker for their branch list and the body of the last branch.
10366 * It might seem that this node could be dispensed with entirely, but the
10367 * endmarker role is not redundant.
10369 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10371 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10375 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10382 const char * const origparse = RExC_parse;
10384 I32 max = REG_INFTY;
10385 #ifdef RE_TRACK_PATTERN_OFFSETS
10388 const char *maxpos = NULL;
10390 /* Save the original in case we change the emitted regop to a FAIL. */
10391 regnode * const orig_emit = RExC_emit;
10393 GET_RE_DEBUG_FLAGS_DECL;
10395 PERL_ARGS_ASSERT_REGPIECE;
10397 DEBUG_PARSE("piec");
10399 ret = regatom(pRExC_state, &flags,depth+1);
10401 if (flags & (TRYAGAIN|RESTART_UTF8))
10402 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10404 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10410 if (op == '{' && regcurly(RExC_parse, FALSE)) {
10412 #ifdef RE_TRACK_PATTERN_OFFSETS
10413 parse_start = RExC_parse; /* MJD */
10415 next = RExC_parse + 1;
10416 while (isDIGIT(*next) || *next == ',') {
10417 if (*next == ',') {
10425 if (*next == '}') { /* got one */
10429 min = atoi(RExC_parse);
10430 if (*maxpos == ',')
10433 maxpos = RExC_parse;
10434 max = atoi(maxpos);
10435 if (!max && *maxpos != '0')
10436 max = REG_INFTY; /* meaning "infinity" */
10437 else if (max >= REG_INFTY)
10438 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10440 nextchar(pRExC_state);
10441 if (max < min) { /* If can't match, warn and optimize to fail
10444 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10446 /* We can't back off the size because we have to reserve
10447 * enough space for all the things we are about to throw
10448 * away, but we can shrink it by the ammount we are about
10449 * to re-use here */
10450 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10453 RExC_emit = orig_emit;
10455 ret = reg_node(pRExC_state, OPFAIL);
10458 else if (min == max
10459 && RExC_parse < RExC_end
10460 && (*RExC_parse == '?' || *RExC_parse == '+'))
10463 ckWARN2reg(RExC_parse + 1,
10464 "Useless use of greediness modifier '%c'",
10467 /* Absorb the modifier, so later code doesn't see nor use
10469 nextchar(pRExC_state);
10473 if ((flags&SIMPLE)) {
10474 RExC_naughty += 2 + RExC_naughty / 2;
10475 reginsert(pRExC_state, CURLY, ret, depth+1);
10476 Set_Node_Offset(ret, parse_start+1); /* MJD */
10477 Set_Node_Cur_Length(ret, parse_start);
10480 regnode * const w = reg_node(pRExC_state, WHILEM);
10483 REGTAIL(pRExC_state, ret, w);
10484 if (!SIZE_ONLY && RExC_extralen) {
10485 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10486 reginsert(pRExC_state, NOTHING,ret, depth+1);
10487 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10489 reginsert(pRExC_state, CURLYX,ret, depth+1);
10491 Set_Node_Offset(ret, parse_start+1);
10492 Set_Node_Length(ret,
10493 op == '{' ? (RExC_parse - parse_start) : 1);
10495 if (!SIZE_ONLY && RExC_extralen)
10496 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10497 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10499 RExC_whilem_seen++, RExC_extralen += 3;
10500 RExC_naughty += 4 + RExC_naughty; /* compound interest */
10507 *flagp |= HASWIDTH;
10509 ARG1_SET(ret, (U16)min);
10510 ARG2_SET(ret, (U16)max);
10512 if (max == REG_INFTY)
10513 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10519 if (!ISMULT1(op)) {
10524 #if 0 /* Now runtime fix should be reliable. */
10526 /* if this is reinstated, don't forget to put this back into perldiag:
10528 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10530 (F) The part of the regexp subject to either the * or + quantifier
10531 could match an empty string. The {#} shows in the regular
10532 expression about where the problem was discovered.
10536 if (!(flags&HASWIDTH) && op != '?')
10537 vFAIL("Regexp *+ operand could be empty");
10540 #ifdef RE_TRACK_PATTERN_OFFSETS
10541 parse_start = RExC_parse;
10543 nextchar(pRExC_state);
10545 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10547 if (op == '*' && (flags&SIMPLE)) {
10548 reginsert(pRExC_state, STAR, ret, depth+1);
10551 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10553 else if (op == '*') {
10557 else if (op == '+' && (flags&SIMPLE)) {
10558 reginsert(pRExC_state, PLUS, ret, depth+1);
10561 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10563 else if (op == '+') {
10567 else if (op == '?') {
10572 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10573 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10574 ckWARN2reg(RExC_parse,
10575 "%"UTF8f" matches null string many times",
10576 UTF8fARG(UTF, (RExC_parse >= origparse
10577 ? RExC_parse - origparse
10580 (void)ReREFCNT_inc(RExC_rx_sv);
10583 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10584 nextchar(pRExC_state);
10585 reginsert(pRExC_state, MINMOD, ret, depth+1);
10586 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10589 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10591 nextchar(pRExC_state);
10592 ender = reg_node(pRExC_state, SUCCEED);
10593 REGTAIL(pRExC_state, ret, ender);
10594 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10596 ender = reg_node(pRExC_state, TAIL);
10597 REGTAIL(pRExC_state, ret, ender);
10600 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10602 vFAIL("Nested quantifiers");
10609 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
10610 UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
10611 const bool strict /* Apply stricter parsing rules? */
10615 /* This is expected to be called by a parser routine that has recognized '\N'
10616 and needs to handle the rest. RExC_parse is expected to point at the first
10617 char following the N at the time of the call. On successful return,
10618 RExC_parse has been updated to point to just after the sequence identified
10619 by this routine, and <*flagp> has been updated.
10621 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
10624 \N may begin either a named sequence, or if outside a character class, mean
10625 to match a non-newline. For non single-quoted regexes, the tokenizer has
10626 attempted to decide which, and in the case of a named sequence, converted it
10627 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
10628 where c1... are the characters in the sequence. For single-quoted regexes,
10629 the tokenizer passes the \N sequence through unchanged; this code will not
10630 attempt to determine this nor expand those, instead raising a syntax error.
10631 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
10632 or there is no '}', it signals that this \N occurrence means to match a
10635 Only the \N{U+...} form should occur in a character class, for the same
10636 reason that '.' inside a character class means to just match a period: it
10637 just doesn't make sense.
10639 The function raises an error (via vFAIL), and doesn't return for various
10640 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
10641 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
10642 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
10643 only possible if node_p is non-NULL.
10646 If <valuep> is non-null, it means the caller can accept an input sequence
10647 consisting of a just a single code point; <*valuep> is set to that value
10648 if the input is such.
10650 If <node_p> is non-null it signifies that the caller can accept any other
10651 legal sequence (i.e., one that isn't just a single code point). <*node_p>
10653 1) \N means not-a-NL: points to a newly created REG_ANY node;
10654 2) \N{}: points to a new NOTHING node;
10655 3) otherwise: points to a new EXACT node containing the resolved
10657 Note that FALSE is returned for single code point sequences if <valuep> is
10661 char * endbrace; /* '}' following the name */
10663 char *endchar; /* Points to '.' or '}' ending cur char in the input
10665 bool has_multiple_chars; /* true if the input stream contains a sequence of
10666 more than one character */
10668 GET_RE_DEBUG_FLAGS_DECL;
10670 PERL_ARGS_ASSERT_GROK_BSLASH_N;
10672 GET_RE_DEBUG_FLAGS;
10674 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
10676 /* The [^\n] meaning of \N ignores spaces and comments under the /x
10677 * modifier. The other meaning does not, so use a temporary until we find
10678 * out which we are being called with */
10679 p = (RExC_flags & RXf_PMf_EXTENDED)
10680 ? regwhite( pRExC_state, RExC_parse )
10683 /* Disambiguate between \N meaning a named character versus \N meaning
10684 * [^\n]. The former is assumed when it can't be the latter. */
10685 if (*p != '{' || regcurly(p, FALSE)) {
10688 /* no bare \N allowed in a charclass */
10689 if (in_char_class) {
10690 vFAIL("\\N in a character class must be a named character: \\N{...}");
10694 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
10696 nextchar(pRExC_state);
10697 *node_p = reg_node(pRExC_state, REG_ANY);
10698 *flagp |= HASWIDTH|SIMPLE;
10700 Set_Node_Length(*node_p, 1); /* MJD */
10704 /* Here, we have decided it should be a named character or sequence */
10706 /* The test above made sure that the next real character is a '{', but
10707 * under the /x modifier, it could be separated by space (or a comment and
10708 * \n) and this is not allowed (for consistency with \x{...} and the
10709 * tokenizer handling of \N{NAME}). */
10710 if (*RExC_parse != '{') {
10711 vFAIL("Missing braces on \\N{}");
10714 RExC_parse++; /* Skip past the '{' */
10716 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
10717 || ! (endbrace == RExC_parse /* nothing between the {} */
10718 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below
10720 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg)
10723 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
10724 vFAIL("\\N{NAME} must be resolved by the lexer");
10727 if (endbrace == RExC_parse) { /* empty: \N{} */
10730 *node_p = reg_node(pRExC_state,NOTHING);
10732 else if (in_char_class) {
10733 if (SIZE_ONLY && in_char_class) {
10735 RExC_parse++; /* Position after the "}" */
10736 vFAIL("Zero length \\N{}");
10739 ckWARNreg(RExC_parse,
10740 "Ignoring zero length \\N{} in character class");
10748 nextchar(pRExC_state);
10752 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
10753 RExC_parse += 2; /* Skip past the 'U+' */
10755 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10757 /* Code points are separated by dots. If none, there is only one code
10758 * point, and is terminated by the brace */
10759 has_multiple_chars = (endchar < endbrace);
10761 if (valuep && (! has_multiple_chars || in_char_class)) {
10762 /* We only pay attention to the first char of
10763 multichar strings being returned in char classes. I kinda wonder
10764 if this makes sense as it does change the behaviour
10765 from earlier versions, OTOH that behaviour was broken
10766 as well. XXX Solution is to recharacterize as
10767 [rest-of-class]|multi1|multi2... */
10769 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10770 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10771 | PERL_SCAN_DISALLOW_PREFIX
10772 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10774 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10776 /* The tokenizer should have guaranteed validity, but it's possible to
10777 * bypass it by using single quoting, so check */
10778 if (length_of_hex == 0
10779 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10781 RExC_parse += length_of_hex; /* Includes all the valid */
10782 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10783 ? UTF8SKIP(RExC_parse)
10785 /* Guard against malformed utf8 */
10786 if (RExC_parse >= endchar) {
10787 RExC_parse = endchar;
10789 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10792 if (in_char_class && has_multiple_chars) {
10794 RExC_parse = endbrace;
10795 vFAIL("\\N{} in character class restricted to one character");
10798 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10802 RExC_parse = endbrace + 1;
10804 else if (! node_p || ! has_multiple_chars) {
10806 /* Here, the input is legal, but not according to the caller's
10807 * options. We fail without advancing the parse, so that the
10808 * caller can try again */
10814 /* What is done here is to convert this to a sub-pattern of the form
10815 * (?:\x{char1}\x{char2}...)
10816 * and then call reg recursively. That way, it retains its atomicness,
10817 * while not having to worry about special handling that some code
10818 * points may have. toke.c has converted the original Unicode values
10819 * to native, so that we can just pass on the hex values unchanged. We
10820 * do have to set a flag to keep recoding from happening in the
10823 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10825 char *orig_end = RExC_end;
10828 while (RExC_parse < endbrace) {
10830 /* Convert to notation the rest of the code understands */
10831 sv_catpv(substitute_parse, "\\x{");
10832 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10833 sv_catpv(substitute_parse, "}");
10835 /* Point to the beginning of the next character in the sequence. */
10836 RExC_parse = endchar + 1;
10837 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10839 sv_catpv(substitute_parse, ")");
10841 RExC_parse = SvPV(substitute_parse, len);
10843 /* Don't allow empty number */
10845 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10847 RExC_end = RExC_parse + len;
10849 /* The values are Unicode, and therefore not subject to recoding */
10850 RExC_override_recoding = 1;
10852 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10853 if (flags & RESTART_UTF8) {
10854 *flagp = RESTART_UTF8;
10857 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10860 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10862 RExC_parse = endbrace;
10863 RExC_end = orig_end;
10864 RExC_override_recoding = 0;
10866 nextchar(pRExC_state);
10876 * It returns the code point in utf8 for the value in *encp.
10877 * value: a code value in the source encoding
10878 * encp: a pointer to an Encode object
10880 * If the result from Encode is not a single character,
10881 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10884 S_reg_recode(pTHX_ const char value, SV **encp)
10887 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10888 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10889 const STRLEN newlen = SvCUR(sv);
10890 UV uv = UNICODE_REPLACEMENT;
10892 PERL_ARGS_ASSERT_REG_RECODE;
10896 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10899 if (!newlen || numlen != newlen) {
10900 uv = UNICODE_REPLACEMENT;
10906 PERL_STATIC_INLINE U8
10907 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10911 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10917 op = get_regex_charset(RExC_flags);
10918 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10919 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10920 been, so there is no hole */
10923 return op + EXACTF;
10926 PERL_STATIC_INLINE void
10927 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
10928 regnode *node, I32* flagp, STRLEN len, UV code_point,
10931 /* This knows the details about sizing an EXACTish node, setting flags for
10932 * it (by setting <*flagp>, and potentially populating it with a single
10935 * If <len> (the length in bytes) is non-zero, this function assumes that
10936 * the node has already been populated, and just does the sizing. In this
10937 * case <code_point> should be the final code point that has already been
10938 * placed into the node. This value will be ignored except that under some
10939 * circumstances <*flagp> is set based on it.
10941 * If <len> is zero, the function assumes that the node is to contain only
10942 * the single character given by <code_point> and calculates what <len>
10943 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10944 * additionally will populate the node's STRING with <code_point> or its
10947 * In both cases <*flagp> is appropriately set
10949 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10950 * 255, must be folded (the former only when the rules indicate it can
10953 * When it does the populating, it looks at the flag 'downgradable'. If
10954 * true with a node that folds, it checks if the single code point
10955 * participates in a fold, and if not downgrades the node to an EXACT.
10956 * This helps the optimizer */
10958 bool len_passed_in = cBOOL(len != 0);
10959 U8 character[UTF8_MAXBYTES_CASE+1];
10961 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10963 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
10964 * sizing difference, and is extra work that is thrown away */
10965 if (downgradable && ! PASS2) {
10966 downgradable = FALSE;
10969 if (! len_passed_in) {
10971 if (UNI_IS_INVARIANT(code_point)) {
10972 if (LOC || ! FOLD) { /* /l defers folding until runtime */
10973 *character = (U8) code_point;
10975 else { /* Here is /i and not /l (toFOLD() is defined on just
10976 ASCII, which isn't the same thing as INVARIANT on
10977 EBCDIC, but it works there, as the extra invariants
10978 fold to themselves) */
10979 *character = toFOLD((U8) code_point);
10981 /* We can downgrade to an EXACT node if this character
10982 * isn't a folding one. Note that this assumes that
10983 * nothing above Latin1 folds to some other invariant than
10984 * one of these alphabetics; otherwise we would also have
10986 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
10987 * || ASCII_FOLD_RESTRICTED))
10989 if (downgradable && PL_fold[code_point] == code_point) {
10995 else if (FOLD && (! LOC
10996 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
10997 { /* Folding, and ok to do so now */
10998 UV folded = _to_uni_fold_flags(
11002 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11003 ? FOLD_FLAGS_NOMIX_ASCII
11006 && folded == code_point
11007 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11012 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11014 /* Not folding this cp, and can output it directly */
11015 *character = UTF8_TWO_BYTE_HI(code_point);
11016 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11020 uvchr_to_utf8( character, code_point);
11021 len = UTF8SKIP(character);
11023 } /* Else pattern isn't UTF8. */
11025 *character = (U8) code_point;
11027 } /* Else is folded non-UTF8 */
11028 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11030 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11031 * comments at join_exact()); */
11032 *character = (U8) code_point;
11035 /* Can turn into an EXACT node if we know the fold at compile time,
11036 * and it folds to itself and doesn't particpate in other folds */
11039 && PL_fold_latin1[code_point] == code_point
11040 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11041 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11045 } /* else is Sharp s. May need to fold it */
11046 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11048 *(character + 1) = 's';
11052 *character = LATIN_SMALL_LETTER_SHARP_S;
11058 RExC_size += STR_SZ(len);
11061 RExC_emit += STR_SZ(len);
11062 STR_LEN(node) = len;
11063 if (! len_passed_in) {
11064 Copy((char *) character, STRING(node), len, char);
11068 *flagp |= HASWIDTH;
11070 /* A single character node is SIMPLE, except for the special-cased SHARP S
11072 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11073 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11074 || ! FOLD || ! DEPENDS_SEMANTICS))
11079 /* The OP may not be well defined in PASS1 */
11080 if (PASS2 && OP(node) == EXACTFL) {
11081 RExC_contains_locale = 1;
11086 /* return atoi(p), unless it's too big to sensibly be a backref,
11087 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11090 S_backref_value(char *p)
11094 for (;isDIGIT(*q); q++); /* calculate length of num */
11095 if (q - p == 0 || q - p > 9)
11102 - regatom - the lowest level
11104 Try to identify anything special at the start of the pattern. If there
11105 is, then handle it as required. This may involve generating a single regop,
11106 such as for an assertion; or it may involve recursing, such as to
11107 handle a () structure.
11109 If the string doesn't start with something special then we gobble up
11110 as much literal text as we can.
11112 Once we have been able to handle whatever type of thing started the
11113 sequence, we return.
11115 Note: we have to be careful with escapes, as they can be both literal
11116 and special, and in the case of \10 and friends, context determines which.
11118 A summary of the code structure is:
11120 switch (first_byte) {
11121 cases for each special:
11122 handle this special;
11125 switch (2nd byte) {
11126 cases for each unambiguous special:
11127 handle this special;
11129 cases for each ambigous special/literal:
11131 if (special) handle here
11133 default: // unambiguously literal:
11136 default: // is a literal char
11139 create EXACTish node for literal;
11140 while (more input and node isn't full) {
11141 switch (input_byte) {
11142 cases for each special;
11143 make sure parse pointer is set so that the next call to
11144 regatom will see this special first
11145 goto loopdone; // EXACTish node terminated by prev. char
11147 append char to EXACTISH node;
11149 get next input byte;
11153 return the generated node;
11155 Specifically there are two separate switches for handling
11156 escape sequences, with the one for handling literal escapes requiring
11157 a dummy entry for all of the special escapes that are actually handled
11160 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11162 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11164 Otherwise does not return NULL.
11168 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11171 regnode *ret = NULL;
11173 char *parse_start = RExC_parse;
11177 GET_RE_DEBUG_FLAGS_DECL;
11179 *flagp = WORST; /* Tentatively. */
11181 DEBUG_PARSE("atom");
11183 PERL_ARGS_ASSERT_REGATOM;
11186 switch ((U8)*RExC_parse) {
11188 RExC_seen_zerolen++;
11189 nextchar(pRExC_state);
11190 if (RExC_flags & RXf_PMf_MULTILINE)
11191 ret = reg_node(pRExC_state, MBOL);
11192 else if (RExC_flags & RXf_PMf_SINGLELINE)
11193 ret = reg_node(pRExC_state, SBOL);
11195 ret = reg_node(pRExC_state, BOL);
11196 Set_Node_Length(ret, 1); /* MJD */
11199 nextchar(pRExC_state);
11201 RExC_seen_zerolen++;
11202 if (RExC_flags & RXf_PMf_MULTILINE)
11203 ret = reg_node(pRExC_state, MEOL);
11204 else if (RExC_flags & RXf_PMf_SINGLELINE)
11205 ret = reg_node(pRExC_state, SEOL);
11207 ret = reg_node(pRExC_state, EOL);
11208 Set_Node_Length(ret, 1); /* MJD */
11211 nextchar(pRExC_state);
11212 if (RExC_flags & RXf_PMf_SINGLELINE)
11213 ret = reg_node(pRExC_state, SANY);
11215 ret = reg_node(pRExC_state, REG_ANY);
11216 *flagp |= HASWIDTH|SIMPLE;
11218 Set_Node_Length(ret, 1); /* MJD */
11222 char * const oregcomp_parse = ++RExC_parse;
11223 ret = regclass(pRExC_state, flagp,depth+1,
11224 FALSE, /* means parse the whole char class */
11225 TRUE, /* allow multi-char folds */
11226 FALSE, /* don't silence non-portable warnings. */
11228 if (*RExC_parse != ']') {
11229 RExC_parse = oregcomp_parse;
11230 vFAIL("Unmatched [");
11233 if (*flagp & RESTART_UTF8)
11235 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11238 nextchar(pRExC_state);
11239 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11243 nextchar(pRExC_state);
11244 ret = reg(pRExC_state, 2, &flags,depth+1);
11246 if (flags & TRYAGAIN) {
11247 if (RExC_parse == RExC_end) {
11248 /* Make parent create an empty node if needed. */
11249 *flagp |= TRYAGAIN;
11254 if (flags & RESTART_UTF8) {
11255 *flagp = RESTART_UTF8;
11258 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11261 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11265 if (flags & TRYAGAIN) {
11266 *flagp |= TRYAGAIN;
11269 vFAIL("Internal urp");
11270 /* Supposed to be caught earlier. */
11273 if (!regcurly(RExC_parse, FALSE)) {
11282 vFAIL("Quantifier follows nothing");
11287 This switch handles escape sequences that resolve to some kind
11288 of special regop and not to literal text. Escape sequnces that
11289 resolve to literal text are handled below in the switch marked
11292 Every entry in this switch *must* have a corresponding entry
11293 in the literal escape switch. However, the opposite is not
11294 required, as the default for this switch is to jump to the
11295 literal text handling code.
11297 switch ((U8)*++RExC_parse) {
11299 /* Special Escapes */
11301 RExC_seen_zerolen++;
11302 ret = reg_node(pRExC_state, SBOL);
11304 goto finish_meta_pat;
11306 ret = reg_node(pRExC_state, GPOS);
11307 RExC_seen |= REG_GPOS_SEEN;
11309 goto finish_meta_pat;
11311 RExC_seen_zerolen++;
11312 ret = reg_node(pRExC_state, KEEPS);
11314 /* XXX:dmq : disabling in-place substitution seems to
11315 * be necessary here to avoid cases of memory corruption, as
11316 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11318 RExC_seen |= REG_LOOKBEHIND_SEEN;
11319 goto finish_meta_pat;
11321 ret = reg_node(pRExC_state, SEOL);
11323 RExC_seen_zerolen++; /* Do not optimize RE away */
11324 goto finish_meta_pat;
11326 ret = reg_node(pRExC_state, EOS);
11328 RExC_seen_zerolen++; /* Do not optimize RE away */
11329 goto finish_meta_pat;
11331 ret = reg_node(pRExC_state, CANY);
11332 RExC_seen |= REG_CANY_SEEN;
11333 *flagp |= HASWIDTH|SIMPLE;
11334 goto finish_meta_pat;
11336 ret = reg_node(pRExC_state, CLUMP);
11337 *flagp |= HASWIDTH;
11338 goto finish_meta_pat;
11344 arg = ANYOF_WORDCHAR;
11348 RExC_seen_zerolen++;
11349 RExC_seen |= REG_LOOKBEHIND_SEEN;
11350 op = BOUND + get_regex_charset(RExC_flags);
11351 if (op > BOUNDA) { /* /aa is same as /a */
11354 else if (op == BOUNDL) {
11355 RExC_contains_locale = 1;
11357 ret = reg_node(pRExC_state, op);
11358 FLAGS(ret) = get_regex_charset(RExC_flags);
11360 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11361 /* diag_listed_as: Use "%s" instead of "%s" */
11362 vFAIL("Use \"\\b\\{\" instead of \"\\b{\"");
11364 goto finish_meta_pat;
11366 RExC_seen_zerolen++;
11367 RExC_seen |= REG_LOOKBEHIND_SEEN;
11368 op = NBOUND + get_regex_charset(RExC_flags);
11369 if (op > NBOUNDA) { /* /aa is same as /a */
11372 else if (op == NBOUNDL) {
11373 RExC_contains_locale = 1;
11375 ret = reg_node(pRExC_state, op);
11376 FLAGS(ret) = get_regex_charset(RExC_flags);
11378 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11379 /* diag_listed_as: Use "%s" instead of "%s" */
11380 vFAIL("Use \"\\B\\{\" instead of \"\\B{\"");
11382 goto finish_meta_pat;
11392 ret = reg_node(pRExC_state, LNBREAK);
11393 *flagp |= HASWIDTH|SIMPLE;
11394 goto finish_meta_pat;
11402 goto join_posix_op_known;
11408 arg = ANYOF_VERTWS;
11410 goto join_posix_op_known;
11420 op = POSIXD + get_regex_charset(RExC_flags);
11421 if (op > POSIXA) { /* /aa is same as /a */
11424 else if (op == POSIXL) {
11425 RExC_contains_locale = 1;
11428 join_posix_op_known:
11431 op += NPOSIXD - POSIXD;
11434 ret = reg_node(pRExC_state, op);
11436 FLAGS(ret) = namedclass_to_classnum(arg);
11439 *flagp |= HASWIDTH|SIMPLE;
11443 nextchar(pRExC_state);
11444 Set_Node_Length(ret, 2); /* MJD */
11450 char* parse_start = RExC_parse - 2;
11455 ret = regclass(pRExC_state, flagp,depth+1,
11456 TRUE, /* means just parse this element */
11457 FALSE, /* don't allow multi-char folds */
11458 FALSE, /* don't silence non-portable warnings.
11459 It would be a bug if these returned
11462 /* regclass() can only return RESTART_UTF8 if multi-char folds
11465 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11470 Set_Node_Offset(ret, parse_start + 2);
11471 Set_Node_Cur_Length(ret, parse_start);
11472 nextchar(pRExC_state);
11476 /* Handle \N and \N{NAME} with multiple code points here and not
11477 * below because it can be multicharacter. join_exact() will join
11478 * them up later on. Also this makes sure that things like
11479 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11480 * The options to the grok function call causes it to fail if the
11481 * sequence is just a single code point. We then go treat it as
11482 * just another character in the current EXACT node, and hence it
11483 * gets uniform treatment with all the other characters. The
11484 * special treatment for quantifiers is not needed for such single
11485 * character sequences */
11487 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
11488 FALSE /* not strict */ )) {
11489 if (*flagp & RESTART_UTF8)
11495 case 'k': /* Handle \k<NAME> and \k'NAME' */
11498 char ch= RExC_parse[1];
11499 if (ch != '<' && ch != '\'' && ch != '{') {
11501 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11502 vFAIL2("Sequence %.2s... not terminated",parse_start);
11504 /* this pretty much dupes the code for (?P=...) in reg(), if
11505 you change this make sure you change that */
11506 char* name_start = (RExC_parse += 2);
11508 SV *sv_dat = reg_scan_name(pRExC_state,
11509 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11510 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11511 if (RExC_parse == name_start || *RExC_parse != ch)
11512 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11513 vFAIL2("Sequence %.3s... not terminated",parse_start);
11516 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11517 RExC_rxi->data->data[num]=(void*)sv_dat;
11518 SvREFCNT_inc_simple_void(sv_dat);
11522 ret = reganode(pRExC_state,
11525 : (ASCII_FOLD_RESTRICTED)
11527 : (AT_LEAST_UNI_SEMANTICS)
11533 *flagp |= HASWIDTH;
11535 /* override incorrect value set in reganode MJD */
11536 Set_Node_Offset(ret, parse_start+1);
11537 Set_Node_Cur_Length(ret, parse_start);
11538 nextchar(pRExC_state);
11544 case '1': case '2': case '3': case '4':
11545 case '5': case '6': case '7': case '8': case '9':
11550 if (*RExC_parse == 'g') {
11554 if (*RExC_parse == '{') {
11558 if (*RExC_parse == '-') {
11562 if (hasbrace && !isDIGIT(*RExC_parse)) {
11563 if (isrel) RExC_parse--;
11565 goto parse_named_seq;
11568 num = S_backref_value(RExC_parse);
11570 vFAIL("Reference to invalid group 0");
11571 else if (num == I32_MAX) {
11572 if (isDIGIT(*RExC_parse))
11573 vFAIL("Reference to nonexistent group");
11575 vFAIL("Unterminated \\g... pattern");
11579 num = RExC_npar - num;
11581 vFAIL("Reference to nonexistent or unclosed group");
11585 num = S_backref_value(RExC_parse);
11586 /* bare \NNN might be backref or octal - if it is larger than or equal
11587 * RExC_npar then it is assumed to be and octal escape.
11588 * Note RExC_npar is +1 from the actual number of parens*/
11589 if (num == I32_MAX || (num > 9 && num >= RExC_npar
11590 && *RExC_parse != '8' && *RExC_parse != '9'))
11592 /* Probably a character specified in octal, e.g. \35 */
11597 /* at this point RExC_parse definitely points to a backref
11600 #ifdef RE_TRACK_PATTERN_OFFSETS
11601 char * const parse_start = RExC_parse - 1; /* MJD */
11603 while (isDIGIT(*RExC_parse))
11606 if (*RExC_parse != '}')
11607 vFAIL("Unterminated \\g{...} pattern");
11611 if (num > (I32)RExC_rx->nparens)
11612 vFAIL("Reference to nonexistent group");
11615 ret = reganode(pRExC_state,
11618 : (ASCII_FOLD_RESTRICTED)
11620 : (AT_LEAST_UNI_SEMANTICS)
11626 *flagp |= HASWIDTH;
11628 /* override incorrect value set in reganode MJD */
11629 Set_Node_Offset(ret, parse_start+1);
11630 Set_Node_Cur_Length(ret, parse_start);
11632 nextchar(pRExC_state);
11637 if (RExC_parse >= RExC_end)
11638 FAIL("Trailing \\");
11641 /* Do not generate "unrecognized" warnings here, we fall
11642 back into the quick-grab loop below */
11649 if (RExC_flags & RXf_PMf_EXTENDED) {
11650 if ( reg_skipcomment( pRExC_state ) )
11657 parse_start = RExC_parse - 1;
11666 #define MAX_NODE_STRING_SIZE 127
11667 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
11669 U8 upper_parse = MAX_NODE_STRING_SIZE;
11670 U8 node_type = compute_EXACTish(pRExC_state);
11671 bool next_is_quantifier;
11672 char * oldp = NULL;
11674 /* We can convert EXACTF nodes to EXACTFU if they contain only
11675 * characters that match identically regardless of the target
11676 * string's UTF8ness. The reason to do this is that EXACTF is not
11677 * trie-able, EXACTFU is.
11679 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
11680 * contain only above-Latin1 characters (hence must be in UTF8),
11681 * which don't participate in folds with Latin1-range characters,
11682 * as the latter's folds aren't known until runtime. (We don't
11683 * need to figure this out until pass 2) */
11684 bool maybe_exactfu = PASS2
11685 && (node_type == EXACTF || node_type == EXACTFL);
11687 /* If a folding node contains only code points that don't
11688 * participate in folds, it can be changed into an EXACT node,
11689 * which allows the optimizer more things to look for */
11692 ret = reg_node(pRExC_state, node_type);
11694 /* In pass1, folded, we use a temporary buffer instead of the
11695 * actual node, as the node doesn't exist yet */
11696 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
11702 /* We do the EXACTFish to EXACT node only if folding. (And we
11703 * don't need to figure this out until pass 2) */
11704 maybe_exact = FOLD && PASS2;
11706 /* XXX The node can hold up to 255 bytes, yet this only goes to
11707 * 127. I (khw) do not know why. Keeping it somewhat less than
11708 * 255 allows us to not have to worry about overflow due to
11709 * converting to utf8 and fold expansion, but that value is
11710 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
11711 * split up by this limit into a single one using the real max of
11712 * 255. Even at 127, this breaks under rare circumstances. If
11713 * folding, we do not want to split a node at a character that is a
11714 * non-final in a multi-char fold, as an input string could just
11715 * happen to want to match across the node boundary. The join
11716 * would solve that problem if the join actually happens. But a
11717 * series of more than two nodes in a row each of 127 would cause
11718 * the first join to succeed to get to 254, but then there wouldn't
11719 * be room for the next one, which could at be one of those split
11720 * multi-char folds. I don't know of any fool-proof solution. One
11721 * could back off to end with only a code point that isn't such a
11722 * non-final, but it is possible for there not to be any in the
11724 for (p = RExC_parse - 1;
11725 len < upper_parse && p < RExC_end;
11730 if (RExC_flags & RXf_PMf_EXTENDED)
11731 p = regwhite( pRExC_state, p );
11742 /* Literal Escapes Switch
11744 This switch is meant to handle escape sequences that
11745 resolve to a literal character.
11747 Every escape sequence that represents something
11748 else, like an assertion or a char class, is handled
11749 in the switch marked 'Special Escapes' above in this
11750 routine, but also has an entry here as anything that
11751 isn't explicitly mentioned here will be treated as
11752 an unescaped equivalent literal.
11755 switch ((U8)*++p) {
11756 /* These are all the special escapes. */
11757 case 'A': /* Start assertion */
11758 case 'b': case 'B': /* Word-boundary assertion*/
11759 case 'C': /* Single char !DANGEROUS! */
11760 case 'd': case 'D': /* digit class */
11761 case 'g': case 'G': /* generic-backref, pos assertion */
11762 case 'h': case 'H': /* HORIZWS */
11763 case 'k': case 'K': /* named backref, keep marker */
11764 case 'p': case 'P': /* Unicode property */
11765 case 'R': /* LNBREAK */
11766 case 's': case 'S': /* space class */
11767 case 'v': case 'V': /* VERTWS */
11768 case 'w': case 'W': /* word class */
11769 case 'X': /* eXtended Unicode "combining
11770 character sequence" */
11771 case 'z': case 'Z': /* End of line/string assertion */
11775 /* Anything after here is an escape that resolves to a
11776 literal. (Except digits, which may or may not)
11782 case 'N': /* Handle a single-code point named character. */
11783 /* The options cause it to fail if a multiple code
11784 * point sequence. Handle those in the switch() above
11786 RExC_parse = p + 1;
11787 if (! grok_bslash_N(pRExC_state, NULL, &ender,
11788 flagp, depth, FALSE,
11789 FALSE /* not strict */ ))
11791 if (*flagp & RESTART_UTF8)
11792 FAIL("panic: grok_bslash_N set RESTART_UTF8");
11793 RExC_parse = p = oldp;
11797 if (ender > 0xff) {
11814 ender = ASCII_TO_NATIVE('\033');
11824 const char* error_msg;
11826 bool valid = grok_bslash_o(&p,
11829 TRUE, /* out warnings */
11830 FALSE, /* not strict */
11831 TRUE, /* Output warnings
11836 RExC_parse = p; /* going to die anyway; point
11837 to exact spot of failure */
11841 if (PL_encoding && ender < 0x100) {
11842 goto recode_encoding;
11844 if (ender > 0xff) {
11851 UV result = UV_MAX; /* initialize to erroneous
11853 const char* error_msg;
11855 bool valid = grok_bslash_x(&p,
11858 TRUE, /* out warnings */
11859 FALSE, /* not strict */
11860 TRUE, /* Output warnings
11865 RExC_parse = p; /* going to die anyway; point
11866 to exact spot of failure */
11871 if (PL_encoding && ender < 0x100) {
11872 goto recode_encoding;
11874 if (ender > 0xff) {
11881 ender = grok_bslash_c(*p++, SIZE_ONLY);
11883 case '8': case '9': /* must be a backreference */
11886 case '1': case '2': case '3':case '4':
11887 case '5': case '6': case '7':
11888 /* When we parse backslash escapes there is ambiguity
11889 * between backreferences and octal escapes. Any escape
11890 * from \1 - \9 is a backreference, any multi-digit
11891 * escape which does not start with 0 and which when
11892 * evaluated as decimal could refer to an already
11893 * parsed capture buffer is a backslash. Anything else
11896 * Note this implies that \118 could be interpreted as
11897 * 118 OR as "\11" . "8" depending on whether there
11898 * were 118 capture buffers defined already in the
11901 /* NOTE, RExC_npar is 1 more than the actual number of
11902 * parens we have seen so far, hence the < RExC_npar below. */
11904 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
11905 { /* Not to be treated as an octal constant, go
11912 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11914 ender = grok_oct(p, &numlen, &flags, NULL);
11915 if (ender > 0xff) {
11919 if (SIZE_ONLY /* like \08, \178 */
11922 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11924 reg_warn_non_literal_string(
11926 form_short_octal_warning(p, numlen));
11929 if (PL_encoding && ender < 0x100)
11930 goto recode_encoding;
11933 if (! RExC_override_recoding) {
11934 SV* enc = PL_encoding;
11935 ender = reg_recode((const char)(U8)ender, &enc);
11936 if (!enc && SIZE_ONLY)
11937 ckWARNreg(p, "Invalid escape in the specified encoding");
11943 FAIL("Trailing \\");
11946 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11947 /* Include any { following the alpha to emphasize
11948 * that it could be part of an escape at some point
11950 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11951 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11953 goto normal_default;
11954 } /* End of switch on '\' */
11956 default: /* A literal character */
11959 && RExC_flags & RXf_PMf_EXTENDED
11960 && ckWARN_d(WARN_DEPRECATED)
11961 && is_PATWS_non_low_safe(p, RExC_end, UTF))
11963 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11964 "Escape literal pattern white space under /x");
11968 if (UTF8_IS_START(*p) && UTF) {
11970 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11971 &numlen, UTF8_ALLOW_DEFAULT);
11977 } /* End of switch on the literal */
11979 /* Here, have looked at the literal character and <ender>
11980 * contains its ordinal, <p> points to the character after it
11983 if ( RExC_flags & RXf_PMf_EXTENDED)
11984 p = regwhite( pRExC_state, p );
11986 /* If the next thing is a quantifier, it applies to this
11987 * character only, which means that this character has to be in
11988 * its own node and can't just be appended to the string in an
11989 * existing node, so if there are already other characters in
11990 * the node, close the node with just them, and set up to do
11991 * this character again next time through, when it will be the
11992 * only thing in its new node */
11993 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11999 if (! FOLD /* The simple case, just append the literal */
12000 || (LOC /* Also don't fold for tricky chars under /l */
12001 && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)))
12005 /* Normally, we don't need the representation of the
12006 * character in the sizing pass--just its size, but if
12007 * folding, we have to actually put the character out
12008 * even in the sizing pass, because the size could
12009 * change as we juggle things at the end of this loop
12010 * to avoid splitting a too-full node in the middle of
12011 * a potential multi-char fold [perl #123539] */
12012 const STRLEN unilen = (SIZE_ONLY && ! FOLD)
12014 : (uvchr_to_utf8((U8*)s, ender) - (U8*)s);
12020 /* The loop increments <len> each time, as all but this
12021 * path (and one other) through it add a single byte to
12022 * the EXACTish node. But this one has changed len to
12023 * be the correct final value, so subtract one to
12024 * cancel out the increment that follows */
12028 /* See comment above for [perl #123539] */
12029 *(s++) = (char) ender;
12032 REGC((char)ender, s++);
12035 /* Can get here if folding only if is one of the /l
12036 * characters whose fold depends on the locale. The
12037 * occurrence of any of these indicate that we can't
12038 * simplify things */
12040 maybe_exact = FALSE;
12041 maybe_exactfu = FALSE;
12046 /* See comments for join_exact() as to why we fold this
12047 * non-UTF at compile time */
12048 || (node_type == EXACTFU
12049 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12051 /* Here, are folding and are not UTF-8 encoded; therefore
12052 * the character must be in the range 0-255, and is not /l
12053 * (Not /l because we already handled these under /l in
12054 * is_PROBLEMATIC_LOCALE_FOLD_cp */
12055 if (IS_IN_SOME_FOLD_L1(ender)) {
12056 maybe_exact = FALSE;
12058 /* See if the character's fold differs between /d and
12059 * /u. This includes the multi-char fold SHARP S to
12062 && (PL_fold[ender] != PL_fold_latin1[ender]
12063 || ender == LATIN_SMALL_LETTER_SHARP_S
12065 && isARG2_lower_or_UPPER_ARG1('s', ender)
12066 && isARG2_lower_or_UPPER_ARG1('s',
12069 maybe_exactfu = FALSE;
12073 /* Even when folding, we store just the input character, as
12074 * we have an array that finds its fold quickly */
12075 *(s++) = (char) ender;
12077 else { /* FOLD and UTF */
12078 /* Unlike the non-fold case, we do actually have to
12079 * calculate the results here in pass 1. This is for two
12080 * reasons, the folded length may be longer than the
12081 * unfolded, and we have to calculate how many EXACTish
12082 * nodes it will take; and we may run out of room in a node
12083 * in the middle of a potential multi-char fold, and have
12084 * to back off accordingly. (Hence we can't use REGC for
12085 * the simple case just below.) */
12088 if (isASCII(ender)) {
12089 folded = toFOLD(ender);
12090 *(s)++ = (U8) folded;
12095 folded = _to_uni_fold_flags(
12099 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12100 ? FOLD_FLAGS_NOMIX_ASCII
12104 /* The loop increments <len> each time, as all but this
12105 * path (and one other) through it add a single byte to
12106 * the EXACTish node. But this one has changed len to
12107 * be the correct final value, so subtract one to
12108 * cancel out the increment that follows */
12109 len += foldlen - 1;
12111 /* If this node only contains non-folding code points so
12112 * far, see if this new one is also non-folding */
12114 if (folded != ender) {
12115 maybe_exact = FALSE;
12118 /* Here the fold is the original; we have to check
12119 * further to see if anything folds to it */
12120 if (_invlist_contains_cp(PL_utf8_foldable,
12123 maybe_exact = FALSE;
12130 if (next_is_quantifier) {
12132 /* Here, the next input is a quantifier, and to get here,
12133 * the current character is the only one in the node.
12134 * Also, here <len> doesn't include the final byte for this
12140 } /* End of loop through literal characters */
12142 /* Here we have either exhausted the input or ran out of room in
12143 * the node. (If we encountered a character that can't be in the
12144 * node, transfer is made directly to <loopdone>, and so we
12145 * wouldn't have fallen off the end of the loop.) In the latter
12146 * case, we artificially have to split the node into two, because
12147 * we just don't have enough space to hold everything. This
12148 * creates a problem if the final character participates in a
12149 * multi-character fold in the non-final position, as a match that
12150 * should have occurred won't, due to the way nodes are matched,
12151 * and our artificial boundary. So back off until we find a non-
12152 * problematic character -- one that isn't at the beginning or
12153 * middle of such a fold. (Either it doesn't participate in any
12154 * folds, or appears only in the final position of all the folds it
12155 * does participate in.) A better solution with far fewer false
12156 * positives, and that would fill the nodes more completely, would
12157 * be to actually have available all the multi-character folds to
12158 * test against, and to back-off only far enough to be sure that
12159 * this node isn't ending with a partial one. <upper_parse> is set
12160 * further below (if we need to reparse the node) to include just
12161 * up through that final non-problematic character that this code
12162 * identifies, so when it is set to less than the full node, we can
12163 * skip the rest of this */
12164 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12166 const STRLEN full_len = len;
12168 assert(len >= MAX_NODE_STRING_SIZE);
12170 /* Here, <s> points to the final byte of the final character.
12171 * Look backwards through the string until find a non-
12172 * problematic character */
12176 /* This has no multi-char folds to non-UTF characters */
12177 if (ASCII_FOLD_RESTRICTED) {
12181 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12185 if (! PL_NonL1NonFinalFold) {
12186 PL_NonL1NonFinalFold = _new_invlist_C_array(
12187 NonL1_Perl_Non_Final_Folds_invlist);
12190 /* Point to the first byte of the final character */
12191 s = (char *) utf8_hop((U8 *) s, -1);
12193 while (s >= s0) { /* Search backwards until find
12194 non-problematic char */
12195 if (UTF8_IS_INVARIANT(*s)) {
12197 /* There are no ascii characters that participate
12198 * in multi-char folds under /aa. In EBCDIC, the
12199 * non-ascii invariants are all control characters,
12200 * so don't ever participate in any folds. */
12201 if (ASCII_FOLD_RESTRICTED
12202 || ! IS_NON_FINAL_FOLD(*s))
12207 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12208 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12214 else if (! _invlist_contains_cp(
12215 PL_NonL1NonFinalFold,
12216 valid_utf8_to_uvchr((U8 *) s, NULL)))
12221 /* Here, the current character is problematic in that
12222 * it does occur in the non-final position of some
12223 * fold, so try the character before it, but have to
12224 * special case the very first byte in the string, so
12225 * we don't read outside the string */
12226 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12227 } /* End of loop backwards through the string */
12229 /* If there were only problematic characters in the string,
12230 * <s> will point to before s0, in which case the length
12231 * should be 0, otherwise include the length of the
12232 * non-problematic character just found */
12233 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12236 /* Here, have found the final character, if any, that is
12237 * non-problematic as far as ending the node without splitting
12238 * it across a potential multi-char fold. <len> contains the
12239 * number of bytes in the node up-to and including that
12240 * character, or is 0 if there is no such character, meaning
12241 * the whole node contains only problematic characters. In
12242 * this case, give up and just take the node as-is. We can't
12247 /* If the node ends in an 's' we make sure it stays EXACTF,
12248 * as if it turns into an EXACTFU, it could later get
12249 * joined with another 's' that would then wrongly match
12251 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
12253 maybe_exactfu = FALSE;
12257 /* Here, the node does contain some characters that aren't
12258 * problematic. If one such is the final character in the
12259 * node, we are done */
12260 if (len == full_len) {
12263 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12265 /* If the final character is problematic, but the
12266 * penultimate is not, back-off that last character to
12267 * later start a new node with it */
12272 /* Here, the final non-problematic character is earlier
12273 * in the input than the penultimate character. What we do
12274 * is reparse from the beginning, going up only as far as
12275 * this final ok one, thus guaranteeing that the node ends
12276 * in an acceptable character. The reason we reparse is
12277 * that we know how far in the character is, but we don't
12278 * know how to correlate its position with the input parse.
12279 * An alternate implementation would be to build that
12280 * correlation as we go along during the original parse,
12281 * but that would entail extra work for every node, whereas
12282 * this code gets executed only when the string is too
12283 * large for the node, and the final two characters are
12284 * problematic, an infrequent occurrence. Yet another
12285 * possible strategy would be to save the tail of the
12286 * string, and the next time regatom is called, initialize
12287 * with that. The problem with this is that unless you
12288 * back off one more character, you won't be guaranteed
12289 * regatom will get called again, unless regbranch,
12290 * regpiece ... are also changed. If you do back off that
12291 * extra character, so that there is input guaranteed to
12292 * force calling regatom, you can't handle the case where
12293 * just the first character in the node is acceptable. I
12294 * (khw) decided to try this method which doesn't have that
12295 * pitfall; if performance issues are found, we can do a
12296 * combination of the current approach plus that one */
12302 } /* End of verifying node ends with an appropriate char */
12304 loopdone: /* Jumped to when encounters something that shouldn't be in
12307 /* I (khw) don't know if you can get here with zero length, but the
12308 * old code handled this situation by creating a zero-length EXACT
12309 * node. Might as well be NOTHING instead */
12315 /* If 'maybe_exact' is still set here, means there are no
12316 * code points in the node that participate in folds;
12317 * similarly for 'maybe_exactfu' and code points that match
12318 * differently depending on UTF8ness of the target string
12319 * (for /u), or depending on locale for /l */
12323 else if (maybe_exactfu) {
12327 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12328 FALSE /* Don't look to see if could
12329 be turned into an EXACT
12330 node, as we have already
12335 RExC_parse = p - 1;
12336 Set_Node_Cur_Length(ret, parse_start);
12337 nextchar(pRExC_state);
12339 /* len is STRLEN which is unsigned, need to copy to signed */
12342 vFAIL("Internal disaster");
12345 } /* End of label 'defchar:' */
12347 } /* End of giant switch on input character */
12353 S_regwhite( RExC_state_t *pRExC_state, char *p )
12355 const char *e = RExC_end;
12357 PERL_ARGS_ASSERT_REGWHITE;
12362 else if (*p == '#') {
12365 if (*p++ == '\n') {
12371 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12380 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12382 /* Returns the next non-pattern-white space, non-comment character (the
12383 * latter only if 'recognize_comment is true) in the string p, which is
12384 * ended by RExC_end. If there is no line break ending a comment,
12385 * RExC_seen has added the REG_RUN_ON_COMMENT_SEEN flag; */
12386 const char *e = RExC_end;
12388 PERL_ARGS_ASSERT_REGPATWS;
12392 if ((len = is_PATWS_safe(p, e, UTF))) {
12395 else if (recognize_comment && *p == '#') {
12399 if (is_LNBREAK_safe(p, e, UTF)) {
12405 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12414 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12416 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12417 * sets up the bitmap and any flags, removing those code points from the
12418 * inversion list, setting it to NULL should it become completely empty */
12420 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12421 assert(PL_regkind[OP(node)] == ANYOF);
12423 ANYOF_BITMAP_ZERO(node);
12424 if (*invlist_ptr) {
12426 /* This gets set if we actually need to modify things */
12427 bool change_invlist = FALSE;
12431 /* Start looking through *invlist_ptr */
12432 invlist_iterinit(*invlist_ptr);
12433 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12437 if (end == UV_MAX && start <= 256) {
12438 ANYOF_FLAGS(node) |= ANYOF_ABOVE_LATIN1_ALL;
12440 else if (end >= 256) {
12441 ANYOF_FLAGS(node) |= ANYOF_UTF8;
12444 /* Quit if are above what we should change */
12449 change_invlist = TRUE;
12451 /* Set all the bits in the range, up to the max that we are doing */
12452 high = (end < 255) ? end : 255;
12453 for (i = start; i <= (int) high; i++) {
12454 if (! ANYOF_BITMAP_TEST(node, i)) {
12455 ANYOF_BITMAP_SET(node, i);
12459 invlist_iterfinish(*invlist_ptr);
12461 /* Done with loop; remove any code points that are in the bitmap from
12462 * *invlist_ptr; similarly for code points above latin1 if we have a
12463 * flag to match all of them anyways */
12464 if (change_invlist) {
12465 _invlist_subtract(*invlist_ptr, PL_Latin1, invlist_ptr);
12467 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
12468 _invlist_intersection(*invlist_ptr, PL_Latin1, invlist_ptr);
12471 /* If have completely emptied it, remove it completely */
12472 if (_invlist_len(*invlist_ptr) == 0) {
12473 SvREFCNT_dec_NN(*invlist_ptr);
12474 *invlist_ptr = NULL;
12479 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12480 Character classes ([:foo:]) can also be negated ([:^foo:]).
12481 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12482 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12483 but trigger failures because they are currently unimplemented. */
12485 #define POSIXCC_DONE(c) ((c) == ':')
12486 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12487 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12489 PERL_STATIC_INLINE I32
12490 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12493 I32 namedclass = OOB_NAMEDCLASS;
12495 PERL_ARGS_ASSERT_REGPPOSIXCC;
12497 if (value == '[' && RExC_parse + 1 < RExC_end &&
12498 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12499 POSIXCC(UCHARAT(RExC_parse)))
12501 const char c = UCHARAT(RExC_parse);
12502 char* const s = RExC_parse++;
12504 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12506 if (RExC_parse == RExC_end) {
12509 /* Try to give a better location for the error (than the end of
12510 * the string) by looking for the matching ']' */
12512 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12515 vFAIL2("Unmatched '%c' in POSIX class", c);
12517 /* Grandfather lone [:, [=, [. */
12521 const char* const t = RExC_parse++; /* skip over the c */
12524 if (UCHARAT(RExC_parse) == ']') {
12525 const char *posixcc = s + 1;
12526 RExC_parse++; /* skip over the ending ] */
12529 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12530 const I32 skip = t - posixcc;
12532 /* Initially switch on the length of the name. */
12535 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12536 this is the Perl \w
12538 namedclass = ANYOF_WORDCHAR;
12541 /* Names all of length 5. */
12542 /* alnum alpha ascii blank cntrl digit graph lower
12543 print punct space upper */
12544 /* Offset 4 gives the best switch position. */
12545 switch (posixcc[4]) {
12547 if (memEQ(posixcc, "alph", 4)) /* alpha */
12548 namedclass = ANYOF_ALPHA;
12551 if (memEQ(posixcc, "spac", 4)) /* space */
12552 namedclass = ANYOF_PSXSPC;
12555 if (memEQ(posixcc, "grap", 4)) /* graph */
12556 namedclass = ANYOF_GRAPH;
12559 if (memEQ(posixcc, "asci", 4)) /* ascii */
12560 namedclass = ANYOF_ASCII;
12563 if (memEQ(posixcc, "blan", 4)) /* blank */
12564 namedclass = ANYOF_BLANK;
12567 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12568 namedclass = ANYOF_CNTRL;
12571 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12572 namedclass = ANYOF_ALPHANUMERIC;
12575 if (memEQ(posixcc, "lowe", 4)) /* lower */
12576 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12577 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12578 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12581 if (memEQ(posixcc, "digi", 4)) /* digit */
12582 namedclass = ANYOF_DIGIT;
12583 else if (memEQ(posixcc, "prin", 4)) /* print */
12584 namedclass = ANYOF_PRINT;
12585 else if (memEQ(posixcc, "punc", 4)) /* punct */
12586 namedclass = ANYOF_PUNCT;
12591 if (memEQ(posixcc, "xdigit", 6))
12592 namedclass = ANYOF_XDIGIT;
12596 if (namedclass == OOB_NAMEDCLASS)
12598 "POSIX class [:%"UTF8f":] unknown",
12599 UTF8fARG(UTF, t - s - 1, s + 1));
12601 /* The #defines are structured so each complement is +1 to
12602 * the normal one */
12606 assert (posixcc[skip] == ':');
12607 assert (posixcc[skip+1] == ']');
12608 } else if (!SIZE_ONLY) {
12609 /* [[=foo=]] and [[.foo.]] are still future. */
12611 /* adjust RExC_parse so the warning shows after
12612 the class closes */
12613 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
12615 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
12618 /* Maternal grandfather:
12619 * "[:" ending in ":" but not in ":]" */
12621 vFAIL("Unmatched '[' in POSIX class");
12624 /* Grandfather lone [:, [=, [. */
12634 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
12636 /* This applies some heuristics at the current parse position (which should
12637 * be at a '[') to see if what follows might be intended to be a [:posix:]
12638 * class. It returns true if it really is a posix class, of course, but it
12639 * also can return true if it thinks that what was intended was a posix
12640 * class that didn't quite make it.
12642 * It will return true for
12644 * [:alphanumerics] (as long as the ] isn't followed immediately by a
12645 * ')' indicating the end of the (?[
12646 * [:any garbage including %^&$ punctuation:]
12648 * This is designed to be called only from S_handle_regex_sets; it could be
12649 * easily adapted to be called from the spot at the beginning of regclass()
12650 * that checks to see in a normal bracketed class if the surrounding []
12651 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
12652 * change long-standing behavior, so I (khw) didn't do that */
12653 char* p = RExC_parse + 1;
12654 char first_char = *p;
12656 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
12658 assert(*(p - 1) == '[');
12660 if (! POSIXCC(first_char)) {
12665 while (p < RExC_end && isWORDCHAR(*p)) p++;
12667 if (p >= RExC_end) {
12671 if (p - RExC_parse > 2 /* Got at least 1 word character */
12672 && (*p == first_char
12673 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
12678 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
12681 && p - RExC_parse > 2 /* [:] evaluates to colon;
12682 [::] is a bad posix class. */
12683 && first_char == *(p - 1));
12687 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
12688 I32 *flagp, U32 depth,
12689 char * const oregcomp_parse)
12691 /* Handle the (?[...]) construct to do set operations */
12694 UV start, end; /* End points of code point ranges */
12696 char *save_end, *save_parse;
12701 const bool save_fold = FOLD;
12703 GET_RE_DEBUG_FLAGS_DECL;
12705 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
12708 vFAIL("(?[...]) not valid in locale");
12710 RExC_uni_semantics = 1;
12712 /* This will return only an ANYOF regnode, or (unlikely) something smaller
12713 * (such as EXACT). Thus we can skip most everything if just sizing. We
12714 * call regclass to handle '[]' so as to not have to reinvent its parsing
12715 * rules here (throwing away the size it computes each time). And, we exit
12716 * upon an unescaped ']' that isn't one ending a regclass. To do both
12717 * these things, we need to realize that something preceded by a backslash
12718 * is escaped, so we have to keep track of backslashes */
12720 UV depth = 0; /* how many nested (?[...]) constructs */
12722 Perl_ck_warner_d(aTHX_
12723 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
12724 "The regex_sets feature is experimental" REPORT_LOCATION,
12725 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
12727 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
12728 RExC_precomp + (RExC_parse - RExC_precomp)));
12730 while (RExC_parse < RExC_end) {
12731 SV* current = NULL;
12732 RExC_parse = regpatws(pRExC_state, RExC_parse,
12733 TRUE); /* means recognize comments */
12734 switch (*RExC_parse) {
12736 if (RExC_parse[1] == '[') depth++, RExC_parse++;
12741 /* Skip the next byte (which could cause us to end up in
12742 * the middle of a UTF-8 character, but since none of those
12743 * are confusable with anything we currently handle in this
12744 * switch (invariants all), it's safe. We'll just hit the
12745 * default: case next time and keep on incrementing until
12746 * we find one of the invariants we do handle. */
12751 /* If this looks like it is a [:posix:] class, leave the
12752 * parse pointer at the '[' to fool regclass() into
12753 * thinking it is part of a '[[:posix:]]'. That function
12754 * will use strict checking to force a syntax error if it
12755 * doesn't work out to a legitimate class */
12756 bool is_posix_class
12757 = could_it_be_a_POSIX_class(pRExC_state);
12758 if (! is_posix_class) {
12762 /* regclass() can only return RESTART_UTF8 if multi-char
12763 folds are allowed. */
12764 if (!regclass(pRExC_state, flagp,depth+1,
12765 is_posix_class, /* parse the whole char
12766 class only if not a
12768 FALSE, /* don't allow multi-char folds */
12769 TRUE, /* silence non-portable warnings. */
12771 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12774 /* function call leaves parse pointing to the ']', except
12775 * if we faked it */
12776 if (is_posix_class) {
12780 SvREFCNT_dec(current); /* In case it returned something */
12785 if (depth--) break;
12787 if (RExC_parse < RExC_end
12788 && *RExC_parse == ')')
12790 node = reganode(pRExC_state, ANYOF, 0);
12791 RExC_size += ANYOF_SKIP;
12792 nextchar(pRExC_state);
12793 Set_Node_Length(node,
12794 RExC_parse - oregcomp_parse + 1); /* MJD */
12803 FAIL("Syntax error in (?[...])");
12806 /* Pass 2 only after this. Everything in this construct is a
12807 * metacharacter. Operands begin with either a '\' (for an escape
12808 * sequence), or a '[' for a bracketed character class. Any other
12809 * character should be an operator, or parenthesis for grouping. Both
12810 * types of operands are handled by calling regclass() to parse them. It
12811 * is called with a parameter to indicate to return the computed inversion
12812 * list. The parsing here is implemented via a stack. Each entry on the
12813 * stack is a single character representing one of the operators, or the
12814 * '('; or else a pointer to an operand inversion list. */
12816 #define IS_OPERAND(a) (! SvIOK(a))
12818 /* The stack starts empty. It is a syntax error if the first thing parsed
12819 * is a binary operator; everything else is pushed on the stack. When an
12820 * operand is parsed, the top of the stack is examined. If it is a binary
12821 * operator, the item before it should be an operand, and both are replaced
12822 * by the result of doing that operation on the new operand and the one on
12823 * the stack. Thus a sequence of binary operands is reduced to a single
12824 * one before the next one is parsed.
12826 * A unary operator may immediately follow a binary in the input, for
12829 * When an operand is parsed and the top of the stack is a unary operator,
12830 * the operation is performed, and then the stack is rechecked to see if
12831 * this new operand is part of a binary operation; if so, it is handled as
12834 * A '(' is simply pushed on the stack; it is valid only if the stack is
12835 * empty, or the top element of the stack is an operator or another '('
12836 * (for which the parenthesized expression will become an operand). By the
12837 * time the corresponding ')' is parsed everything in between should have
12838 * been parsed and evaluated to a single operand (or else is a syntax
12839 * error), and is handled as a regular operand */
12841 sv_2mortal((SV *)(stack = newAV()));
12843 while (RExC_parse < RExC_end) {
12844 I32 top_index = av_tindex(stack);
12846 SV* current = NULL;
12848 /* Skip white space */
12849 RExC_parse = regpatws(pRExC_state, RExC_parse,
12850 TRUE); /* means recognize comments */
12851 if (RExC_parse >= RExC_end) {
12852 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
12854 if ((curchar = UCHARAT(RExC_parse)) == ']') {
12861 if (av_tindex(stack) >= 0 /* This makes sure that we can
12862 safely subtract 1 from
12863 RExC_parse in the next clause.
12864 If we have something on the
12865 stack, we have parsed something
12867 && UCHARAT(RExC_parse - 1) == '('
12868 && RExC_parse < RExC_end)
12870 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
12871 * This happens when we have some thing like
12873 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
12875 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
12877 * Here we would be handling the interpolated
12878 * '$thai_or_lao'. We handle this by a recursive call to
12879 * ourselves which returns the inversion list the
12880 * interpolated expression evaluates to. We use the flags
12881 * from the interpolated pattern. */
12882 U32 save_flags = RExC_flags;
12883 const char * const save_parse = ++RExC_parse;
12885 parse_lparen_question_flags(pRExC_state);
12887 if (RExC_parse == save_parse /* Makes sure there was at
12888 least one flag (or this
12889 embedding wasn't compiled)
12891 || RExC_parse >= RExC_end - 4
12892 || UCHARAT(RExC_parse) != ':'
12893 || UCHARAT(++RExC_parse) != '('
12894 || UCHARAT(++RExC_parse) != '?'
12895 || UCHARAT(++RExC_parse) != '[')
12898 /* In combination with the above, this moves the
12899 * pointer to the point just after the first erroneous
12900 * character (or if there are no flags, to where they
12901 * should have been) */
12902 if (RExC_parse >= RExC_end - 4) {
12903 RExC_parse = RExC_end;
12905 else if (RExC_parse != save_parse) {
12906 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12908 vFAIL("Expecting '(?flags:(?[...'");
12911 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
12912 depth+1, oregcomp_parse);
12914 /* Here, 'current' contains the embedded expression's
12915 * inversion list, and RExC_parse points to the trailing
12916 * ']'; the next character should be the ')' which will be
12917 * paired with the '(' that has been put on the stack, so
12918 * the whole embedded expression reduces to '(operand)' */
12921 RExC_flags = save_flags;
12922 goto handle_operand;
12927 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12928 vFAIL("Unexpected character");
12931 /* regclass() can only return RESTART_UTF8 if multi-char
12932 folds are allowed. */
12933 if (!regclass(pRExC_state, flagp,depth+1,
12934 TRUE, /* means parse just the next thing */
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 /* regclass() will return with parsing just the \ sequence,
12941 * leaving the parse pointer at the next thing to parse */
12943 goto handle_operand;
12945 case '[': /* Is a bracketed character class */
12947 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
12949 if (! is_posix_class) {
12953 /* regclass() can only return RESTART_UTF8 if multi-char
12954 folds are allowed. */
12955 if(!regclass(pRExC_state, flagp,depth+1,
12956 is_posix_class, /* parse the whole char class
12957 only if not a posix class */
12958 FALSE, /* don't allow multi-char folds */
12959 FALSE, /* don't silence non-portable warnings. */
12961 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12963 /* function call leaves parse pointing to the ']', except if we
12965 if (is_posix_class) {
12969 goto handle_operand;
12978 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
12979 || ! IS_OPERAND(*top_ptr))
12982 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
12984 av_push(stack, newSVuv(curchar));
12988 av_push(stack, newSVuv(curchar));
12992 if (top_index >= 0) {
12993 top_ptr = av_fetch(stack, top_index, FALSE);
12995 if (IS_OPERAND(*top_ptr)) {
12997 vFAIL("Unexpected '(' with no preceding operator");
13000 av_push(stack, newSVuv(curchar));
13007 || ! (current = av_pop(stack))
13008 || ! IS_OPERAND(current)
13009 || ! (lparen = av_pop(stack))
13010 || IS_OPERAND(lparen)
13011 || SvUV(lparen) != '(')
13013 SvREFCNT_dec(current);
13015 vFAIL("Unexpected ')'");
13018 SvREFCNT_dec_NN(lparen);
13025 /* Here, we have an operand to process, in 'current' */
13027 if (top_index < 0) { /* Just push if stack is empty */
13028 av_push(stack, current);
13031 SV* top = av_pop(stack);
13033 char current_operator;
13035 if (IS_OPERAND(top)) {
13036 SvREFCNT_dec_NN(top);
13037 SvREFCNT_dec_NN(current);
13038 vFAIL("Operand with no preceding operator");
13040 current_operator = (char) SvUV(top);
13041 switch (current_operator) {
13042 case '(': /* Push the '(' back on followed by the new
13044 av_push(stack, top);
13045 av_push(stack, current);
13046 SvREFCNT_inc(top); /* Counters the '_dec' done
13047 just after the 'break', so
13048 it doesn't get wrongly freed
13053 _invlist_invert(current);
13055 /* Unlike binary operators, the top of the stack,
13056 * now that this unary one has been popped off, may
13057 * legally be an operator, and we now have operand
13060 SvREFCNT_dec_NN(top);
13061 goto handle_operand;
13064 prev = av_pop(stack);
13065 _invlist_intersection(prev,
13068 av_push(stack, current);
13073 prev = av_pop(stack);
13074 _invlist_union(prev, current, ¤t);
13075 av_push(stack, current);
13079 prev = av_pop(stack);;
13080 _invlist_subtract(prev, current, ¤t);
13081 av_push(stack, current);
13084 case '^': /* The union minus the intersection */
13090 prev = av_pop(stack);
13091 _invlist_union(prev, current, &u);
13092 _invlist_intersection(prev, current, &i);
13093 /* _invlist_subtract will overwrite current
13094 without freeing what it already contains */
13096 _invlist_subtract(u, i, ¤t);
13097 av_push(stack, current);
13098 SvREFCNT_dec_NN(i);
13099 SvREFCNT_dec_NN(u);
13100 SvREFCNT_dec_NN(element);
13105 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13107 SvREFCNT_dec_NN(top);
13108 SvREFCNT_dec(prev);
13112 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13115 if (av_tindex(stack) < 0 /* Was empty */
13116 || ((final = av_pop(stack)) == NULL)
13117 || ! IS_OPERAND(final)
13118 || av_tindex(stack) >= 0) /* More left on stack */
13120 vFAIL("Incomplete expression within '(?[ ])'");
13123 /* Here, 'final' is the resultant inversion list from evaluating the
13124 * expression. Return it if so requested */
13125 if (return_invlist) {
13126 *return_invlist = final;
13130 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13131 * expecting a string of ranges and individual code points */
13132 invlist_iterinit(final);
13133 result_string = newSVpvs("");
13134 while (invlist_iternext(final, &start, &end)) {
13135 if (start == end) {
13136 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13139 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13144 save_parse = RExC_parse;
13145 RExC_parse = SvPV(result_string, len);
13146 save_end = RExC_end;
13147 RExC_end = RExC_parse + len;
13149 /* We turn off folding around the call, as the class we have constructed
13150 * already has all folding taken into consideration, and we don't want
13151 * regclass() to add to that */
13152 RExC_flags &= ~RXf_PMf_FOLD;
13153 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13155 node = regclass(pRExC_state, flagp,depth+1,
13156 FALSE, /* means parse the whole char class */
13157 FALSE, /* don't allow multi-char folds */
13158 TRUE, /* silence non-portable warnings. The above may very
13159 well have generated non-portable code points, but
13160 they're valid on this machine */
13163 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13166 RExC_flags |= RXf_PMf_FOLD;
13168 RExC_parse = save_parse + 1;
13169 RExC_end = save_end;
13170 SvREFCNT_dec_NN(final);
13171 SvREFCNT_dec_NN(result_string);
13173 nextchar(pRExC_state);
13174 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13179 /* The names of properties whose definitions are not known at compile time are
13180 * stored in this SV, after a constant heading. So if the length has been
13181 * changed since initialization, then there is a run-time definition. */
13182 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13183 (SvCUR(listsv) != initial_listsv_len)
13186 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13187 const bool stop_at_1, /* Just parse the next thing, don't
13188 look for a full character class */
13189 bool allow_multi_folds,
13190 const bool silence_non_portable, /* Don't output warnings
13193 SV** ret_invlist) /* Return an inversion list, not a node */
13195 /* parse a bracketed class specification. Most of these will produce an
13196 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13197 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13198 * under /i with multi-character folds: it will be rewritten following the
13199 * paradigm of this example, where the <multi-fold>s are characters which
13200 * fold to multiple character sequences:
13201 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13202 * gets effectively rewritten as:
13203 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13204 * reg() gets called (recursively) on the rewritten version, and this
13205 * function will return what it constructs. (Actually the <multi-fold>s
13206 * aren't physically removed from the [abcdefghi], it's just that they are
13207 * ignored in the recursion by means of a flag:
13208 * <RExC_in_multi_char_class>.)
13210 * ANYOF nodes contain a bit map for the first 256 characters, with the
13211 * corresponding bit set if that character is in the list. For characters
13212 * above 255, a range list or swash is used. There are extra bits for \w,
13213 * etc. in locale ANYOFs, as what these match is not determinable at
13216 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13217 * to be restarted. This can only happen if ret_invlist is non-NULL.
13221 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13223 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13226 IV namedclass = OOB_NAMEDCLASS;
13227 char *rangebegin = NULL;
13228 bool need_class = 0;
13230 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13231 than just initialized. */
13232 SV* properties = NULL; /* Code points that match \p{} \P{} */
13233 SV* posixes = NULL; /* Code points that match classes like [:word:],
13234 extended beyond the Latin1 range. These have to
13235 be kept separate from other code points for much
13236 of this function because their handling is
13237 different under /i, and for most classes under
13239 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13240 separate for a while from the non-complemented
13241 versions because of complications with /d
13243 UV element_count = 0; /* Number of distinct elements in the class.
13244 Optimizations may be possible if this is tiny */
13245 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13246 character; used under /i */
13248 char * stop_ptr = RExC_end; /* where to stop parsing */
13249 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13251 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13253 /* Unicode properties are stored in a swash; this holds the current one
13254 * being parsed. If this swash is the only above-latin1 component of the
13255 * character class, an optimization is to pass it directly on to the
13256 * execution engine. Otherwise, it is set to NULL to indicate that there
13257 * are other things in the class that have to be dealt with at execution
13259 SV* swash = NULL; /* Code points that match \p{} \P{} */
13261 /* Set if a component of this character class is user-defined; just passed
13262 * on to the engine */
13263 bool has_user_defined_property = FALSE;
13265 /* inversion list of code points this node matches only when the target
13266 * string is in UTF-8. (Because is under /d) */
13267 SV* depends_list = NULL;
13269 /* Inversion list of code points this node matches regardless of things
13270 * like locale, folding, utf8ness of the target string */
13271 SV* cp_list = NULL;
13273 /* Like cp_list, but code points on this list need to be checked for things
13274 * that fold to/from them under /i */
13275 SV* cp_foldable_list = NULL;
13277 /* Like cp_list, but code points on this list are valid only when the
13278 * runtime locale is UTF-8 */
13279 SV* only_utf8_locale_list = NULL;
13282 /* In a range, counts how many 0-2 of the ends of it came from literals,
13283 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13284 UV literal_endpoint = 0;
13286 bool invert = FALSE; /* Is this class to be complemented */
13288 bool warn_super = ALWAYS_WARN_SUPER;
13290 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13291 case we need to change the emitted regop to an EXACT. */
13292 const char * orig_parse = RExC_parse;
13293 const SSize_t orig_size = RExC_size;
13294 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13295 GET_RE_DEBUG_FLAGS_DECL;
13297 PERL_ARGS_ASSERT_REGCLASS;
13299 PERL_UNUSED_ARG(depth);
13302 DEBUG_PARSE("clas");
13304 /* Assume we are going to generate an ANYOF node. */
13305 ret = reganode(pRExC_state, ANYOF, 0);
13308 RExC_size += ANYOF_SKIP;
13309 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13312 ANYOF_FLAGS(ret) = 0;
13314 RExC_emit += ANYOF_SKIP;
13315 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13316 initial_listsv_len = SvCUR(listsv);
13317 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13321 RExC_parse = regpatws(pRExC_state, RExC_parse,
13322 FALSE /* means don't recognize comments */);
13325 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13328 allow_multi_folds = FALSE;
13331 RExC_parse = regpatws(pRExC_state, RExC_parse,
13332 FALSE /* means don't recognize comments */);
13336 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13337 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13338 const char *s = RExC_parse;
13339 const char c = *s++;
13341 while (isWORDCHAR(*s))
13343 if (*s && c == *s && s[1] == ']') {
13344 SAVEFREESV(RExC_rx_sv);
13346 "POSIX syntax [%c %c] belongs inside character classes",
13348 (void)ReREFCNT_inc(RExC_rx_sv);
13352 /* If the caller wants us to just parse a single element, accomplish this
13353 * by faking the loop ending condition */
13354 if (stop_at_1 && RExC_end > RExC_parse) {
13355 stop_ptr = RExC_parse + 1;
13358 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13359 if (UCHARAT(RExC_parse) == ']')
13360 goto charclassloop;
13364 if (RExC_parse >= stop_ptr) {
13369 RExC_parse = regpatws(pRExC_state, RExC_parse,
13370 FALSE /* means don't recognize comments */);
13373 if (UCHARAT(RExC_parse) == ']') {
13379 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13380 save_value = value;
13381 save_prevvalue = prevvalue;
13384 rangebegin = RExC_parse;
13388 value = utf8n_to_uvchr((U8*)RExC_parse,
13389 RExC_end - RExC_parse,
13390 &numlen, UTF8_ALLOW_DEFAULT);
13391 RExC_parse += numlen;
13394 value = UCHARAT(RExC_parse++);
13397 && RExC_parse < RExC_end
13398 && POSIXCC(UCHARAT(RExC_parse)))
13400 namedclass = regpposixcc(pRExC_state, value, strict);
13402 else if (value == '\\') {
13404 value = utf8n_to_uvchr((U8*)RExC_parse,
13405 RExC_end - RExC_parse,
13406 &numlen, UTF8_ALLOW_DEFAULT);
13407 RExC_parse += numlen;
13410 value = UCHARAT(RExC_parse++);
13412 /* Some compilers cannot handle switching on 64-bit integer
13413 * values, therefore value cannot be an UV. Yes, this will
13414 * be a problem later if we want switch on Unicode.
13415 * A similar issue a little bit later when switching on
13416 * namedclass. --jhi */
13418 /* If the \ is escaping white space when white space is being
13419 * skipped, it means that that white space is wanted literally, and
13420 * is already in 'value'. Otherwise, need to translate the escape
13421 * into what it signifies. */
13422 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13424 case 'w': namedclass = ANYOF_WORDCHAR; break;
13425 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13426 case 's': namedclass = ANYOF_SPACE; break;
13427 case 'S': namedclass = ANYOF_NSPACE; break;
13428 case 'd': namedclass = ANYOF_DIGIT; break;
13429 case 'D': namedclass = ANYOF_NDIGIT; break;
13430 case 'v': namedclass = ANYOF_VERTWS; break;
13431 case 'V': namedclass = ANYOF_NVERTWS; break;
13432 case 'h': namedclass = ANYOF_HORIZWS; break;
13433 case 'H': namedclass = ANYOF_NHORIZWS; break;
13434 case 'N': /* Handle \N{NAME} in class */
13436 /* We only pay attention to the first char of
13437 multichar strings being returned. I kinda wonder
13438 if this makes sense as it does change the behaviour
13439 from earlier versions, OTOH that behaviour was broken
13441 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
13442 TRUE, /* => charclass */
13445 if (*flagp & RESTART_UTF8)
13446 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13456 /* We will handle any undefined properties ourselves */
13457 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
13458 /* And we actually would prefer to get
13459 * the straight inversion list of the
13460 * swash, since we will be accessing it
13461 * anyway, to save a little time */
13462 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
13464 if (RExC_parse >= RExC_end)
13465 vFAIL2("Empty \\%c{}", (U8)value);
13466 if (*RExC_parse == '{') {
13467 const U8 c = (U8)value;
13468 e = strchr(RExC_parse++, '}');
13470 vFAIL2("Missing right brace on \\%c{}", c);
13471 while (isSPACE(UCHARAT(RExC_parse)))
13473 if (e == RExC_parse)
13474 vFAIL2("Empty \\%c{}", c);
13475 n = e - RExC_parse;
13476 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
13488 if (UCHARAT(RExC_parse) == '^') {
13491 /* toggle. (The rhs xor gets the single bit that
13492 * differs between P and p; the other xor inverts just
13494 value ^= 'P' ^ 'p';
13496 while (isSPACE(UCHARAT(RExC_parse))) {
13501 /* Try to get the definition of the property into
13502 * <invlist>. If /i is in effect, the effective property
13503 * will have its name be <__NAME_i>. The design is
13504 * discussed in commit
13505 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
13506 formatted = Perl_form(aTHX_
13508 (FOLD) ? "__" : "",
13513 name = savepvn(formatted, strlen(formatted));
13515 /* Look up the property name, and get its swash and
13516 * inversion list, if the property is found */
13518 SvREFCNT_dec_NN(swash);
13520 swash = _core_swash_init("utf8", name, &PL_sv_undef,
13523 NULL, /* No inversion list */
13526 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
13528 SvREFCNT_dec_NN(swash);
13532 /* Here didn't find it. It could be a user-defined
13533 * property that will be available at run-time. If we
13534 * accept only compile-time properties, is an error;
13535 * otherwise add it to the list for run-time look up */
13537 RExC_parse = e + 1;
13539 "Property '%"UTF8f"' is unknown",
13540 UTF8fARG(UTF, n, name));
13542 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
13543 (value == 'p' ? '+' : '!'),
13544 UTF8fARG(UTF, n, name));
13545 has_user_defined_property = TRUE;
13547 /* We don't know yet, so have to assume that the
13548 * property could match something in the Latin1 range,
13549 * hence something that isn't utf8. Note that this
13550 * would cause things in <depends_list> to match
13551 * inappropriately, except that any \p{}, including
13552 * this one forces Unicode semantics, which means there
13553 * is no <depends_list> */
13554 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
13558 /* Here, did get the swash and its inversion list. If
13559 * the swash is from a user-defined property, then this
13560 * whole character class should be regarded as such */
13561 if (swash_init_flags
13562 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
13564 has_user_defined_property = TRUE;
13567 /* We warn on matching an above-Unicode code point
13568 * if the match would return true, except don't
13569 * warn for \p{All}, which has exactly one element
13571 (_invlist_contains_cp(invlist, 0x110000)
13572 && (! (_invlist_len(invlist) == 1
13573 && *invlist_array(invlist) == 0)))
13579 /* Invert if asking for the complement */
13580 if (value == 'P') {
13581 _invlist_union_complement_2nd(properties,
13585 /* The swash can't be used as-is, because we've
13586 * inverted things; delay removing it to here after
13587 * have copied its invlist above */
13588 SvREFCNT_dec_NN(swash);
13592 _invlist_union(properties, invlist, &properties);
13597 RExC_parse = e + 1;
13598 namedclass = ANYOF_UNIPROP; /* no official name, but it's
13601 /* \p means they want Unicode semantics */
13602 RExC_uni_semantics = 1;
13605 case 'n': value = '\n'; break;
13606 case 'r': value = '\r'; break;
13607 case 't': value = '\t'; break;
13608 case 'f': value = '\f'; break;
13609 case 'b': value = '\b'; break;
13610 case 'e': value = ASCII_TO_NATIVE('\033');break;
13611 case 'a': value = '\a'; break;
13613 RExC_parse--; /* function expects to be pointed at the 'o' */
13615 const char* error_msg;
13616 bool valid = grok_bslash_o(&RExC_parse,
13619 SIZE_ONLY, /* warnings in pass
13622 silence_non_portable,
13628 if (PL_encoding && value < 0x100) {
13629 goto recode_encoding;
13633 RExC_parse--; /* function expects to be pointed at the 'x' */
13635 const char* error_msg;
13636 bool valid = grok_bslash_x(&RExC_parse,
13639 TRUE, /* Output warnings */
13641 silence_non_portable,
13647 if (PL_encoding && value < 0x100)
13648 goto recode_encoding;
13651 value = grok_bslash_c(*RExC_parse++, SIZE_ONLY);
13653 case '0': case '1': case '2': case '3': case '4':
13654 case '5': case '6': case '7':
13656 /* Take 1-3 octal digits */
13657 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13658 numlen = (strict) ? 4 : 3;
13659 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
13660 RExC_parse += numlen;
13663 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13664 vFAIL("Need exactly 3 octal digits");
13666 else if (! SIZE_ONLY /* like \08, \178 */
13668 && RExC_parse < RExC_end
13669 && isDIGIT(*RExC_parse)
13670 && ckWARN(WARN_REGEXP))
13672 SAVEFREESV(RExC_rx_sv);
13673 reg_warn_non_literal_string(
13675 form_short_octal_warning(RExC_parse, numlen));
13676 (void)ReREFCNT_inc(RExC_rx_sv);
13679 if (PL_encoding && value < 0x100)
13680 goto recode_encoding;
13684 if (! RExC_override_recoding) {
13685 SV* enc = PL_encoding;
13686 value = reg_recode((const char)(U8)value, &enc);
13689 vFAIL("Invalid escape in the specified encoding");
13691 else if (SIZE_ONLY) {
13692 ckWARNreg(RExC_parse,
13693 "Invalid escape in the specified encoding");
13699 /* Allow \_ to not give an error */
13700 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
13702 vFAIL2("Unrecognized escape \\%c in character class",
13706 SAVEFREESV(RExC_rx_sv);
13707 ckWARN2reg(RExC_parse,
13708 "Unrecognized escape \\%c in character class passed through",
13710 (void)ReREFCNT_inc(RExC_rx_sv);
13714 } /* End of switch on char following backslash */
13715 } /* end of handling backslash escape sequences */
13718 literal_endpoint++;
13721 /* Here, we have the current token in 'value' */
13723 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
13726 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
13727 * literal, as is the character that began the false range, i.e.
13728 * the 'a' in the examples */
13731 const int w = (RExC_parse >= rangebegin)
13732 ? RExC_parse - rangebegin
13736 "False [] range \"%"UTF8f"\"",
13737 UTF8fARG(UTF, w, rangebegin));
13740 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
13741 ckWARN2reg(RExC_parse,
13742 "False [] range \"%"UTF8f"\"",
13743 UTF8fARG(UTF, w, rangebegin));
13744 (void)ReREFCNT_inc(RExC_rx_sv);
13745 cp_list = add_cp_to_invlist(cp_list, '-');
13746 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
13751 range = 0; /* this was not a true range */
13752 element_count += 2; /* So counts for three values */
13755 classnum = namedclass_to_classnum(namedclass);
13757 if (LOC && namedclass < ANYOF_POSIXL_MAX
13758 #ifndef HAS_ISASCII
13759 && classnum != _CC_ASCII
13762 /* What the Posix classes (like \w, [:space:]) match in locale
13763 * isn't knowable under locale until actual match time. Room
13764 * must be reserved (one time per outer bracketed class) to
13765 * store such classes. The space will contain a bit for each
13766 * named class that is to be matched against. This isn't
13767 * needed for \p{} and pseudo-classes, as they are not affected
13768 * by locale, and hence are dealt with separately */
13769 if (! need_class) {
13772 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13775 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13777 ANYOF_FLAGS(ret) |= ANYOF_POSIXL;
13778 ANYOF_POSIXL_ZERO(ret);
13781 /* See if it already matches the complement of this POSIX
13783 if ((ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13784 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
13788 posixl_matches_all = TRUE;
13789 break; /* No need to continue. Since it matches both
13790 e.g., \w and \W, it matches everything, and the
13791 bracketed class can be optimized into qr/./s */
13794 /* Add this class to those that should be checked at runtime */
13795 ANYOF_POSIXL_SET(ret, namedclass);
13797 /* The above-Latin1 characters are not subject to locale rules.
13798 * Just add them, in the second pass, to the
13799 * unconditionally-matched list */
13801 SV* scratch_list = NULL;
13803 /* Get the list of the above-Latin1 code points this
13805 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
13806 PL_XPosix_ptrs[classnum],
13808 /* Odd numbers are complements, like
13809 * NDIGIT, NASCII, ... */
13810 namedclass % 2 != 0,
13812 /* Checking if 'cp_list' is NULL first saves an extra
13813 * clone. Its reference count will be decremented at the
13814 * next union, etc, or if this is the only instance, at the
13815 * end of the routine */
13817 cp_list = scratch_list;
13820 _invlist_union(cp_list, scratch_list, &cp_list);
13821 SvREFCNT_dec_NN(scratch_list);
13823 continue; /* Go get next character */
13826 else if (! SIZE_ONLY) {
13828 /* Here, not in pass1 (in that pass we skip calculating the
13829 * contents of this class), and is /l, or is a POSIX class for
13830 * which /l doesn't matter (or is a Unicode property, which is
13831 * skipped here). */
13832 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
13833 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
13835 /* Here, should be \h, \H, \v, or \V. None of /d, /i
13836 * nor /l make a difference in what these match,
13837 * therefore we just add what they match to cp_list. */
13838 if (classnum != _CC_VERTSPACE) {
13839 assert( namedclass == ANYOF_HORIZWS
13840 || namedclass == ANYOF_NHORIZWS);
13842 /* It turns out that \h is just a synonym for
13844 classnum = _CC_BLANK;
13847 _invlist_union_maybe_complement_2nd(
13849 PL_XPosix_ptrs[classnum],
13850 namedclass % 2 != 0, /* Complement if odd
13851 (NHORIZWS, NVERTWS)
13856 else { /* Garden variety class. If is NASCII, NDIGIT, ...
13857 complement and use nposixes */
13858 SV** posixes_ptr = namedclass % 2 == 0
13861 SV** source_ptr = &PL_XPosix_ptrs[classnum];
13862 _invlist_union_maybe_complement_2nd(
13865 namedclass % 2 != 0,
13868 continue; /* Go get next character */
13870 } /* end of namedclass \blah */
13872 /* Here, we have a single value. If 'range' is set, it is the ending
13873 * of a range--check its validity. Later, we will handle each
13874 * individual code point in the range. If 'range' isn't set, this
13875 * could be the beginning of a range, so check for that by looking
13876 * ahead to see if the next real character to be processed is the range
13877 * indicator--the minus sign */
13880 RExC_parse = regpatws(pRExC_state, RExC_parse,
13881 FALSE /* means don't recognize comments */);
13885 if (prevvalue > value) /* b-a */ {
13886 const int w = RExC_parse - rangebegin;
13888 "Invalid [] range \"%"UTF8f"\"",
13889 UTF8fARG(UTF, w, rangebegin));
13890 range = 0; /* not a valid range */
13894 prevvalue = value; /* save the beginning of the potential range */
13895 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13896 && *RExC_parse == '-')
13898 char* next_char_ptr = RExC_parse + 1;
13899 if (skip_white) { /* Get the next real char after the '-' */
13900 next_char_ptr = regpatws(pRExC_state,
13902 FALSE); /* means don't recognize
13906 /* If the '-' is at the end of the class (just before the ']',
13907 * it is a literal minus; otherwise it is a range */
13908 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13909 RExC_parse = next_char_ptr;
13911 /* a bad range like \w-, [:word:]- ? */
13912 if (namedclass > OOB_NAMEDCLASS) {
13913 if (strict || ckWARN(WARN_REGEXP)) {
13915 RExC_parse >= rangebegin ?
13916 RExC_parse - rangebegin : 0;
13918 vFAIL4("False [] range \"%*.*s\"",
13923 "False [] range \"%*.*s\"",
13928 cp_list = add_cp_to_invlist(cp_list, '-');
13932 range = 1; /* yeah, it's a range! */
13933 continue; /* but do it the next time */
13938 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13941 /* non-Latin1 code point implies unicode semantics. Must be set in
13942 * pass1 so is there for the whole of pass 2 */
13944 RExC_uni_semantics = 1;
13947 /* Ready to process either the single value, or the completed range.
13948 * For single-valued non-inverted ranges, we consider the possibility
13949 * of multi-char folds. (We made a conscious decision to not do this
13950 * for the other cases because it can often lead to non-intuitive
13951 * results. For example, you have the peculiar case that:
13952 * "s s" =~ /^[^\xDF]+$/i => Y
13953 * "ss" =~ /^[^\xDF]+$/i => N
13955 * See [perl #89750] */
13956 if (FOLD && allow_multi_folds && value == prevvalue) {
13957 if (value == LATIN_SMALL_LETTER_SHARP_S
13958 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13961 /* Here <value> is indeed a multi-char fold. Get what it is */
13963 U8 foldbuf[UTF8_MAXBYTES_CASE];
13966 UV folded = _to_uni_fold_flags(
13970 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
13971 ? FOLD_FLAGS_NOMIX_ASCII
13975 /* Here, <folded> should be the first character of the
13976 * multi-char fold of <value>, with <foldbuf> containing the
13977 * whole thing. But, if this fold is not allowed (because of
13978 * the flags), <fold> will be the same as <value>, and should
13979 * be processed like any other character, so skip the special
13981 if (folded != value) {
13983 /* Skip if we are recursed, currently parsing the class
13984 * again. Otherwise add this character to the list of
13985 * multi-char folds. */
13986 if (! RExC_in_multi_char_class) {
13987 AV** this_array_ptr;
13989 STRLEN cp_count = utf8_length(foldbuf,
13990 foldbuf + foldlen);
13991 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13993 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13996 if (! multi_char_matches) {
13997 multi_char_matches = newAV();
14000 /* <multi_char_matches> is actually an array of arrays.
14001 * There will be one or two top-level elements: [2],
14002 * and/or [3]. The [2] element is an array, each
14003 * element thereof is a character which folds to TWO
14004 * characters; [3] is for folds to THREE characters.
14005 * (Unicode guarantees a maximum of 3 characters in any
14006 * fold.) When we rewrite the character class below,
14007 * we will do so such that the longest folds are
14008 * written first, so that it prefers the longest
14009 * matching strings first. This is done even if it
14010 * turns out that any quantifier is non-greedy, out of
14011 * programmer laziness. Tom Christiansen has agreed
14012 * that this is ok. This makes the test for the
14013 * ligature 'ffi' come before the test for 'ff' */
14014 if (av_exists(multi_char_matches, cp_count)) {
14015 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14017 this_array = *this_array_ptr;
14020 this_array = newAV();
14021 av_store(multi_char_matches, cp_count,
14024 av_push(this_array, multi_fold);
14027 /* This element should not be processed further in this
14030 value = save_value;
14031 prevvalue = save_prevvalue;
14037 /* Deal with this element of the class */
14040 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14043 SV* this_range = _new_invlist(1);
14044 _append_range_to_invlist(this_range, prevvalue, value);
14046 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
14047 * If this range was specified using something like 'i-j', we want
14048 * to include only the 'i' and the 'j', and not anything in
14049 * between, so exclude non-ASCII, non-alphabetics from it.
14050 * However, if the range was specified with something like
14051 * [\x89-\x91] or [\x89-j], all code points within it should be
14052 * included. literal_endpoint==2 means both ends of the range used
14053 * a literal character, not \x{foo} */
14054 if (literal_endpoint == 2
14055 && ((prevvalue >= 'a' && value <= 'z')
14056 || (prevvalue >= 'A' && value <= 'Z')))
14058 _invlist_intersection(this_range, PL_ASCII,
14061 /* Since this above only contains ascii, the intersection of it
14062 * with anything will still yield only ascii */
14063 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ALPHA],
14066 _invlist_union(cp_foldable_list, this_range, &cp_foldable_list);
14067 literal_endpoint = 0;
14071 range = 0; /* this range (if it was one) is done now */
14072 } /* End of loop through all the text within the brackets */
14074 /* If anything in the class expands to more than one character, we have to
14075 * deal with them by building up a substitute parse string, and recursively
14076 * calling reg() on it, instead of proceeding */
14077 if (multi_char_matches) {
14078 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14081 char *save_end = RExC_end;
14082 char *save_parse = RExC_parse;
14083 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14088 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14089 because too confusing */
14091 sv_catpv(substitute_parse, "(?:");
14095 /* Look at the longest folds first */
14096 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14098 if (av_exists(multi_char_matches, cp_count)) {
14099 AV** this_array_ptr;
14102 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14104 while ((this_sequence = av_pop(*this_array_ptr)) !=
14107 if (! first_time) {
14108 sv_catpv(substitute_parse, "|");
14110 first_time = FALSE;
14112 sv_catpv(substitute_parse, SvPVX(this_sequence));
14117 /* If the character class contains anything else besides these
14118 * multi-character folds, have to include it in recursive parsing */
14119 if (element_count) {
14120 sv_catpv(substitute_parse, "|[");
14121 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14122 sv_catpv(substitute_parse, "]");
14125 sv_catpv(substitute_parse, ")");
14128 /* This is a way to get the parse to skip forward a whole named
14129 * sequence instead of matching the 2nd character when it fails the
14131 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14135 RExC_parse = SvPV(substitute_parse, len);
14136 RExC_end = RExC_parse + len;
14137 RExC_in_multi_char_class = 1;
14138 RExC_emit = (regnode *)orig_emit;
14140 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14142 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14144 RExC_parse = save_parse;
14145 RExC_end = save_end;
14146 RExC_in_multi_char_class = 0;
14147 SvREFCNT_dec_NN(multi_char_matches);
14151 /* Here, we've gone through the entire class and dealt with multi-char
14152 * folds. We are now in a position that we can do some checks to see if we
14153 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14154 * Currently we only do two checks:
14155 * 1) is in the unlikely event that the user has specified both, eg. \w and
14156 * \W under /l, then the class matches everything. (This optimization
14157 * is done only to make the optimizer code run later work.)
14158 * 2) if the character class contains only a single element (including a
14159 * single range), we see if there is an equivalent node for it.
14160 * Other checks are possible */
14161 if (! ret_invlist /* Can't optimize if returning the constructed
14163 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14168 if (UNLIKELY(posixl_matches_all)) {
14171 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14172 \w or [:digit:] or \p{foo}
14175 /* All named classes are mapped into POSIXish nodes, with its FLAG
14176 * argument giving which class it is */
14177 switch ((I32)namedclass) {
14178 case ANYOF_UNIPROP:
14181 /* These don't depend on the charset modifiers. They always
14182 * match under /u rules */
14183 case ANYOF_NHORIZWS:
14184 case ANYOF_HORIZWS:
14185 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14188 case ANYOF_NVERTWS:
14193 /* The actual POSIXish node for all the rest depends on the
14194 * charset modifier. The ones in the first set depend only on
14195 * ASCII or, if available on this platform, locale */
14199 op = (LOC) ? POSIXL : POSIXA;
14210 /* under /a could be alpha */
14212 if (ASCII_RESTRICTED) {
14213 namedclass = ANYOF_ALPHA + (namedclass % 2);
14221 /* The rest have more possibilities depending on the charset.
14222 * We take advantage of the enum ordering of the charset
14223 * modifiers to get the exact node type, */
14225 op = POSIXD + get_regex_charset(RExC_flags);
14226 if (op > POSIXA) { /* /aa is same as /a */
14231 /* The odd numbered ones are the complements of the
14232 * next-lower even number one */
14233 if (namedclass % 2 == 1) {
14237 arg = namedclass_to_classnum(namedclass);
14241 else if (value == prevvalue) {
14243 /* Here, the class consists of just a single code point */
14246 if (! LOC && value == '\n') {
14247 op = REG_ANY; /* Optimize [^\n] */
14248 *flagp |= HASWIDTH|SIMPLE;
14252 else if (value < 256 || UTF) {
14254 /* Optimize a single value into an EXACTish node, but not if it
14255 * would require converting the pattern to UTF-8. */
14256 op = compute_EXACTish(pRExC_state);
14258 } /* Otherwise is a range */
14259 else if (! LOC) { /* locale could vary these */
14260 if (prevvalue == '0') {
14261 if (value == '9') {
14268 /* Here, we have changed <op> away from its initial value iff we found
14269 * an optimization */
14272 /* Throw away this ANYOF regnode, and emit the calculated one,
14273 * which should correspond to the beginning, not current, state of
14275 const char * cur_parse = RExC_parse;
14276 RExC_parse = (char *)orig_parse;
14280 /* To get locale nodes to not use the full ANYOF size would
14281 * require moving the code above that writes the portions
14282 * of it that aren't in other nodes to after this point.
14283 * e.g. ANYOF_POSIXL_SET */
14284 RExC_size = orig_size;
14288 RExC_emit = (regnode *)orig_emit;
14289 if (PL_regkind[op] == POSIXD) {
14290 if (op == POSIXL) {
14291 RExC_contains_locale = 1;
14294 op += NPOSIXD - POSIXD;
14299 ret = reg_node(pRExC_state, op);
14301 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14305 *flagp |= HASWIDTH|SIMPLE;
14307 else if (PL_regkind[op] == EXACT) {
14308 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14309 TRUE /* downgradable to EXACT */
14313 RExC_parse = (char *) cur_parse;
14315 SvREFCNT_dec(posixes);
14316 SvREFCNT_dec(nposixes);
14317 SvREFCNT_dec(cp_list);
14318 SvREFCNT_dec(cp_foldable_list);
14325 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14327 /* If folding, we calculate all characters that could fold to or from the
14328 * ones already on the list */
14329 if (cp_foldable_list) {
14331 UV start, end; /* End points of code point ranges */
14333 SV* fold_intersection = NULL;
14336 /* Our calculated list will be for Unicode rules. For locale
14337 * matching, we have to keep a separate list that is consulted at
14338 * runtime only when the locale indicates Unicode rules. For
14339 * non-locale, we just use to the general list */
14341 use_list = &only_utf8_locale_list;
14344 use_list = &cp_list;
14347 /* Only the characters in this class that participate in folds need
14348 * be checked. Get the intersection of this class and all the
14349 * possible characters that are foldable. This can quickly narrow
14350 * down a large class */
14351 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14352 &fold_intersection);
14354 /* The folds for all the Latin1 characters are hard-coded into this
14355 * program, but we have to go out to disk to get the others. */
14356 if (invlist_highest(cp_foldable_list) >= 256) {
14358 /* This is a hash that for a particular fold gives all
14359 * characters that are involved in it */
14360 if (! PL_utf8_foldclosures) {
14362 /* If the folds haven't been read in, call a fold function
14364 if (! PL_utf8_tofold) {
14365 U8 dummy[UTF8_MAXBYTES_CASE+1];
14367 /* This string is just a short named one above \xff */
14368 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
14369 assert(PL_utf8_tofold); /* Verify that worked */
14371 PL_utf8_foldclosures
14372 = _swash_inversion_hash(PL_utf8_tofold);
14376 /* Now look at the foldable characters in this class individually */
14377 invlist_iterinit(fold_intersection);
14378 while (invlist_iternext(fold_intersection, &start, &end)) {
14381 /* Look at every character in the range */
14382 for (j = start; j <= end; j++) {
14383 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14389 /* We have the latin1 folding rules hard-coded here so
14390 * that an innocent-looking character class, like
14391 * /[ks]/i won't have to go out to disk to find the
14392 * possible matches. XXX It would be better to
14393 * generate these via regen, in case a new version of
14394 * the Unicode standard adds new mappings, though that
14395 * is not really likely, and may be caught by the
14396 * default: case of the switch below. */
14398 if (IS_IN_SOME_FOLD_L1(j)) {
14400 /* ASCII is always matched; non-ASCII is matched
14401 * only under Unicode rules (which could happen
14402 * under /l if the locale is a UTF-8 one */
14403 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
14404 *use_list = add_cp_to_invlist(*use_list,
14405 PL_fold_latin1[j]);
14409 add_cp_to_invlist(depends_list,
14410 PL_fold_latin1[j]);
14414 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
14415 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
14417 /* Certain Latin1 characters have matches outside
14418 * Latin1. To get here, <j> is one of those
14419 * characters. None of these matches is valid for
14420 * ASCII characters under /aa, which is why the 'if'
14421 * just above excludes those. These matches only
14422 * happen when the target string is utf8. The code
14423 * below adds the single fold closures for <j> to the
14424 * inversion list. */
14430 add_cp_to_invlist(*use_list, KELVIN_SIGN);
14434 *use_list = add_cp_to_invlist(*use_list,
14435 LATIN_SMALL_LETTER_LONG_S);
14438 *use_list = add_cp_to_invlist(*use_list,
14439 GREEK_CAPITAL_LETTER_MU);
14440 *use_list = add_cp_to_invlist(*use_list,
14441 GREEK_SMALL_LETTER_MU);
14443 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14444 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14446 add_cp_to_invlist(*use_list, ANGSTROM_SIGN);
14448 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14449 *use_list = add_cp_to_invlist(*use_list,
14450 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14452 case LATIN_SMALL_LETTER_SHARP_S:
14453 *use_list = add_cp_to_invlist(*use_list,
14454 LATIN_CAPITAL_LETTER_SHARP_S);
14456 case 'F': case 'f':
14457 case 'I': case 'i':
14458 case 'L': case 'l':
14459 case 'T': case 't':
14460 case 'A': case 'a':
14461 case 'H': case 'h':
14462 case 'J': case 'j':
14463 case 'N': case 'n':
14464 case 'W': case 'w':
14465 case 'Y': case 'y':
14466 /* These all are targets of multi-character
14467 * folds from code points that require UTF8
14468 * to express, so they can't match unless
14469 * the target string is in UTF-8, so no
14470 * action here is necessary, as regexec.c
14471 * properly handles the general case for
14472 * UTF-8 matching and multi-char folds */
14475 /* Use deprecated warning to increase the
14476 * chances of this being output */
14477 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
14484 /* Here is an above Latin1 character. We don't have the
14485 * rules hard-coded for it. First, get its fold. This is
14486 * the simple fold, as the multi-character folds have been
14487 * handled earlier and separated out */
14488 _to_uni_fold_flags(j, foldbuf, &foldlen,
14489 (ASCII_FOLD_RESTRICTED)
14490 ? FOLD_FLAGS_NOMIX_ASCII
14493 /* Single character fold of above Latin1. Add everything in
14494 * its fold closure to the list that this node should match.
14495 * The fold closures data structure is a hash with the keys
14496 * being the UTF-8 of every character that is folded to, like
14497 * 'k', and the values each an array of all code points that
14498 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
14499 * Multi-character folds are not included */
14500 if ((listp = hv_fetch(PL_utf8_foldclosures,
14501 (char *) foldbuf, foldlen, FALSE)))
14503 AV* list = (AV*) *listp;
14505 for (k = 0; k <= av_tindex(list); k++) {
14506 SV** c_p = av_fetch(list, k, FALSE);
14509 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
14513 /* /aa doesn't allow folds between ASCII and non- */
14514 if ((ASCII_FOLD_RESTRICTED
14515 && (isASCII(c) != isASCII(j))))
14520 /* Folds under /l which cross the 255/256 boundary
14521 * are added to a separate list. (These are valid
14522 * only when the locale is UTF-8.) */
14523 if (c < 256 && LOC) {
14524 *use_list = add_cp_to_invlist(*use_list, c);
14528 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
14530 cp_list = add_cp_to_invlist(cp_list, c);
14533 /* Similarly folds involving non-ascii Latin1
14534 * characters under /d are added to their list */
14535 depends_list = add_cp_to_invlist(depends_list,
14542 SvREFCNT_dec_NN(fold_intersection);
14545 /* Now that we have finished adding all the folds, there is no reason
14546 * to keep the foldable list separate */
14547 _invlist_union(cp_list, cp_foldable_list, &cp_list);
14548 SvREFCNT_dec_NN(cp_foldable_list);
14551 /* And combine the result (if any) with any inversion list from posix
14552 * classes. The lists are kept separate up to now because we don't want to
14553 * fold the classes (folding of those is automatically handled by the swash
14554 * fetching code) */
14555 if (posixes || nposixes) {
14556 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
14557 /* Under /a and /aa, nothing above ASCII matches these */
14558 _invlist_intersection(posixes,
14559 PL_XPosix_ptrs[_CC_ASCII],
14563 if (DEPENDS_SEMANTICS) {
14564 /* Under /d, everything in the upper half of the Latin1 range
14565 * matches these complements */
14566 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_NON_ASCII_ALL;
14568 else if (AT_LEAST_ASCII_RESTRICTED) {
14569 /* Under /a and /aa, everything above ASCII matches these
14571 _invlist_union_complement_2nd(nposixes,
14572 PL_XPosix_ptrs[_CC_ASCII],
14576 _invlist_union(posixes, nposixes, &posixes);
14577 SvREFCNT_dec_NN(nposixes);
14580 posixes = nposixes;
14583 if (! DEPENDS_SEMANTICS) {
14585 _invlist_union(cp_list, posixes, &cp_list);
14586 SvREFCNT_dec_NN(posixes);
14593 /* Under /d, we put into a separate list the Latin1 things that
14594 * match only when the target string is utf8 */
14595 SV* nonascii_but_latin1_properties = NULL;
14596 _invlist_intersection(posixes, PL_UpperLatin1,
14597 &nonascii_but_latin1_properties);
14598 _invlist_subtract(posixes, nonascii_but_latin1_properties,
14601 _invlist_union(cp_list, posixes, &cp_list);
14602 SvREFCNT_dec_NN(posixes);
14608 if (depends_list) {
14609 _invlist_union(depends_list, nonascii_but_latin1_properties,
14611 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
14614 depends_list = nonascii_but_latin1_properties;
14619 /* And combine the result (if any) with any inversion list from properties.
14620 * The lists are kept separate up to now so that we can distinguish the two
14621 * in regards to matching above-Unicode. A run-time warning is generated
14622 * if a Unicode property is matched against a non-Unicode code point. But,
14623 * we allow user-defined properties to match anything, without any warning,
14624 * and we also suppress the warning if there is a portion of the character
14625 * class that isn't a Unicode property, and which matches above Unicode, \W
14626 * or [\x{110000}] for example.
14627 * (Note that in this case, unlike the Posix one above, there is no
14628 * <depends_list>, because having a Unicode property forces Unicode
14633 /* If it matters to the final outcome, see if a non-property
14634 * component of the class matches above Unicode. If so, the
14635 * warning gets suppressed. This is true even if just a single
14636 * such code point is specified, as though not strictly correct if
14637 * another such code point is matched against, the fact that they
14638 * are using above-Unicode code points indicates they should know
14639 * the issues involved */
14641 warn_super = ! (invert
14642 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
14645 _invlist_union(properties, cp_list, &cp_list);
14646 SvREFCNT_dec_NN(properties);
14649 cp_list = properties;
14653 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
14657 /* Here, we have calculated what code points should be in the character
14660 * Now we can see about various optimizations. Fold calculation (which we
14661 * did above) needs to take place before inversion. Otherwise /[^k]/i
14662 * would invert to include K, which under /i would match k, which it
14663 * shouldn't. Therefore we can't invert folded locale now, as it won't be
14664 * folded until runtime */
14666 /* If we didn't do folding, it's because some information isn't available
14667 * until runtime; set the run-time fold flag for these. (We don't have to
14668 * worry about properties folding, as that is taken care of by the swash
14669 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
14670 * locales, or the class matches at least one 0-255 range code point */
14672 if (only_utf8_locale_list) {
14673 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14675 else if (cp_list) { /* Look to see if there a 0-255 code point is in
14678 invlist_iterinit(cp_list);
14679 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
14680 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14682 invlist_iterfinish(cp_list);
14686 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
14687 * at compile time. Besides not inverting folded locale now, we can't
14688 * invert if there are things such as \w, which aren't known until runtime
14692 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14694 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14696 _invlist_invert(cp_list);
14698 /* Any swash can't be used as-is, because we've inverted things */
14700 SvREFCNT_dec_NN(swash);
14704 /* Clear the invert flag since have just done it here */
14709 *ret_invlist = cp_list;
14710 SvREFCNT_dec(swash);
14712 /* Discard the generated node */
14714 RExC_size = orig_size;
14717 RExC_emit = orig_emit;
14722 /* Some character classes are equivalent to other nodes. Such nodes take
14723 * up less room and generally fewer operations to execute than ANYOF nodes.
14724 * Above, we checked for and optimized into some such equivalents for
14725 * certain common classes that are easy to test. Getting to this point in
14726 * the code means that the class didn't get optimized there. Since this
14727 * code is only executed in Pass 2, it is too late to save space--it has
14728 * been allocated in Pass 1, and currently isn't given back. But turning
14729 * things into an EXACTish node can allow the optimizer to join it to any
14730 * adjacent such nodes. And if the class is equivalent to things like /./,
14731 * expensive run-time swashes can be avoided. Now that we have more
14732 * complete information, we can find things necessarily missed by the
14733 * earlier code. I (khw) am not sure how much to look for here. It would
14734 * be easy, but perhaps too slow, to check any candidates against all the
14735 * node types they could possibly match using _invlistEQ(). */
14740 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14741 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14743 /* We don't optimize if we are supposed to make sure all non-Unicode
14744 * code points raise a warning, as only ANYOF nodes have this check.
14746 && ! ((ANYOF_FLAGS(ret) | ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
14749 U8 op = END; /* The optimzation node-type */
14750 const char * cur_parse= RExC_parse;
14752 invlist_iterinit(cp_list);
14753 if (! invlist_iternext(cp_list, &start, &end)) {
14755 /* Here, the list is empty. This happens, for example, when a
14756 * Unicode property is the only thing in the character class, and
14757 * it doesn't match anything. (perluniprops.pod notes such
14760 *flagp |= HASWIDTH|SIMPLE;
14762 else if (start == end) { /* The range is a single code point */
14763 if (! invlist_iternext(cp_list, &start, &end)
14765 /* Don't do this optimization if it would require changing
14766 * the pattern to UTF-8 */
14767 && (start < 256 || UTF))
14769 /* Here, the list contains a single code point. Can optimize
14770 * into an EXACTish node */
14779 /* A locale node under folding with one code point can be
14780 * an EXACTFL, as its fold won't be calculated until
14786 /* Here, we are generally folding, but there is only one
14787 * code point to match. If we have to, we use an EXACT
14788 * node, but it would be better for joining with adjacent
14789 * nodes in the optimization pass if we used the same
14790 * EXACTFish node that any such are likely to be. We can
14791 * do this iff the code point doesn't participate in any
14792 * folds. For example, an EXACTF of a colon is the same as
14793 * an EXACT one, since nothing folds to or from a colon. */
14795 if (IS_IN_SOME_FOLD_L1(value)) {
14800 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
14805 /* If we haven't found the node type, above, it means we
14806 * can use the prevailing one */
14808 op = compute_EXACTish(pRExC_state);
14813 else if (start == 0) {
14814 if (end == UV_MAX) {
14816 *flagp |= HASWIDTH|SIMPLE;
14819 else if (end == '\n' - 1
14820 && invlist_iternext(cp_list, &start, &end)
14821 && start == '\n' + 1 && end == UV_MAX)
14824 *flagp |= HASWIDTH|SIMPLE;
14828 invlist_iterfinish(cp_list);
14831 RExC_parse = (char *)orig_parse;
14832 RExC_emit = (regnode *)orig_emit;
14834 ret = reg_node(pRExC_state, op);
14836 RExC_parse = (char *)cur_parse;
14838 if (PL_regkind[op] == EXACT) {
14839 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14840 TRUE /* downgradable to EXACT */
14844 SvREFCNT_dec_NN(cp_list);
14849 /* Here, <cp_list> contains all the code points we can determine at
14850 * compile time that match under all conditions. Go through it, and
14851 * for things that belong in the bitmap, put them there, and delete from
14852 * <cp_list>. While we are at it, see if everything above 255 is in the
14853 * list, and if so, set a flag to speed up execution */
14855 populate_ANYOF_from_invlist(ret, &cp_list);
14858 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
14861 /* Here, the bitmap has been populated with all the Latin1 code points that
14862 * always match. Can now add to the overall list those that match only
14863 * when the target string is UTF-8 (<depends_list>). */
14864 if (depends_list) {
14866 _invlist_union(cp_list, depends_list, &cp_list);
14867 SvREFCNT_dec_NN(depends_list);
14870 cp_list = depends_list;
14872 ANYOF_FLAGS(ret) |= ANYOF_UTF8;
14875 /* If there is a swash and more than one element, we can't use the swash in
14876 * the optimization below. */
14877 if (swash && element_count > 1) {
14878 SvREFCNT_dec_NN(swash);
14882 set_ANYOF_arg(pRExC_state, ret, cp_list,
14883 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14885 only_utf8_locale_list,
14886 swash, has_user_defined_property);
14888 *flagp |= HASWIDTH|SIMPLE;
14890 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
14891 RExC_contains_locale = 1;
14897 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14900 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
14901 regnode* const node,
14903 SV* const runtime_defns,
14904 SV* const only_utf8_locale_list,
14906 const bool has_user_defined_property)
14908 /* Sets the arg field of an ANYOF-type node 'node', using information about
14909 * the node passed-in. If there is nothing outside the node's bitmap, the
14910 * arg is set to ANYOF_NONBITMAP_EMPTY. Otherwise, it sets the argument to
14911 * the count returned by add_data(), having allocated and stored an array,
14912 * av, that that count references, as follows:
14913 * av[0] stores the character class description in its textual form.
14914 * This is used later (regexec.c:Perl_regclass_swash()) to
14915 * initialize the appropriate swash, and is also useful for dumping
14916 * the regnode. This is set to &PL_sv_undef if the textual
14917 * description is not needed at run-time (as happens if the other
14918 * elements completely define the class)
14919 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
14920 * computed from av[0]. But if no further computation need be done,
14921 * the swash is stored here now (and av[0] is &PL_sv_undef).
14922 * av[2] stores the inversion list of code points that match only if the
14923 * current locale is UTF-8
14924 * av[3] stores the cp_list inversion list for use in addition or instead
14925 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
14926 * (Otherwise everything needed is already in av[0] and av[1])
14927 * av[4] is set if any component of the class is from a user-defined
14928 * property; used only if av[3] exists */
14932 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
14934 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
14935 assert(! (ANYOF_FLAGS(node)
14936 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8)));
14937 ARG_SET(node, ANYOF_NONBITMAP_EMPTY);
14940 AV * const av = newAV();
14943 assert(ANYOF_FLAGS(node)
14944 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8|ANYOF_LOC_FOLD));
14946 av_store(av, 0, (runtime_defns)
14947 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
14949 av_store(av, 1, swash);
14950 SvREFCNT_dec_NN(cp_list);
14953 av_store(av, 1, &PL_sv_undef);
14955 av_store(av, 3, cp_list);
14956 av_store(av, 4, newSVuv(has_user_defined_property));
14960 if (only_utf8_locale_list) {
14961 av_store(av, 2, only_utf8_locale_list);
14964 av_store(av, 2, &PL_sv_undef);
14967 rv = newRV_noinc(MUTABLE_SV(av));
14968 n = add_data(pRExC_state, STR_WITH_LEN("s"));
14969 RExC_rxi->data->data[n] = (void*)rv;
14975 /* reg_skipcomment()
14977 Absorbs an /x style # comments from the input stream.
14978 Returns true if there is more text remaining in the stream.
14979 Will set the REG_RUN_ON_COMMENT_SEEN flag if the comment
14980 terminates the pattern without including a newline.
14982 Note its the callers responsibility to ensure that we are
14983 actually in /x mode
14988 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14992 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14994 while (RExC_parse < RExC_end)
14995 if (*RExC_parse++ == '\n') {
15000 /* we ran off the end of the pattern without ending
15001 the comment, so we have to add an \n when wrapping */
15002 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
15010 Advances the parse position, and optionally absorbs
15011 "whitespace" from the inputstream.
15013 Without /x "whitespace" means (?#...) style comments only,
15014 with /x this means (?#...) and # comments and whitespace proper.
15016 Returns the RExC_parse point from BEFORE the scan occurs.
15018 This is the /x friendly way of saying RExC_parse++.
15022 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
15024 char* const retval = RExC_parse++;
15026 PERL_ARGS_ASSERT_NEXTCHAR;
15029 if (RExC_end - RExC_parse >= 3
15030 && *RExC_parse == '('
15031 && RExC_parse[1] == '?'
15032 && RExC_parse[2] == '#')
15034 while (*RExC_parse != ')') {
15035 if (RExC_parse == RExC_end)
15036 FAIL("Sequence (?#... not terminated");
15042 if (RExC_flags & RXf_PMf_EXTENDED) {
15043 if (isSPACE(*RExC_parse)) {
15047 else if (*RExC_parse == '#') {
15048 if ( reg_skipcomment( pRExC_state ) )
15057 - reg_node - emit a node
15059 STATIC regnode * /* Location. */
15060 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
15064 regnode * const ret = RExC_emit;
15065 GET_RE_DEBUG_FLAGS_DECL;
15067 PERL_ARGS_ASSERT_REG_NODE;
15070 SIZE_ALIGN(RExC_size);
15074 if (RExC_emit >= RExC_emit_bound)
15075 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15076 op, RExC_emit, RExC_emit_bound);
15078 NODE_ALIGN_FILL(ret);
15080 FILL_ADVANCE_NODE(ptr, op);
15081 #ifdef RE_TRACK_PATTERN_OFFSETS
15082 if (RExC_offsets) { /* MJD */
15084 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
15085 "reg_node", __LINE__,
15087 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
15088 ? "Overwriting end of array!\n" : "OK",
15089 (UV)(RExC_emit - RExC_emit_start),
15090 (UV)(RExC_parse - RExC_start),
15091 (UV)RExC_offsets[0]));
15092 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
15100 - reganode - emit a node with an argument
15102 STATIC regnode * /* Location. */
15103 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
15107 regnode * const ret = RExC_emit;
15108 GET_RE_DEBUG_FLAGS_DECL;
15110 PERL_ARGS_ASSERT_REGANODE;
15113 SIZE_ALIGN(RExC_size);
15118 assert(2==regarglen[op]+1);
15120 Anything larger than this has to allocate the extra amount.
15121 If we changed this to be:
15123 RExC_size += (1 + regarglen[op]);
15125 then it wouldn't matter. Its not clear what side effect
15126 might come from that so its not done so far.
15131 if (RExC_emit >= RExC_emit_bound)
15132 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15133 op, RExC_emit, RExC_emit_bound);
15135 NODE_ALIGN_FILL(ret);
15137 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15138 #ifdef RE_TRACK_PATTERN_OFFSETS
15139 if (RExC_offsets) { /* MJD */
15141 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15145 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
15146 "Overwriting end of array!\n" : "OK",
15147 (UV)(RExC_emit - RExC_emit_start),
15148 (UV)(RExC_parse - RExC_start),
15149 (UV)RExC_offsets[0]));
15150 Set_Cur_Node_Offset;
15158 - reguni - emit (if appropriate) a Unicode character
15160 PERL_STATIC_INLINE STRLEN
15161 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
15165 PERL_ARGS_ASSERT_REGUNI;
15167 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
15171 - reginsert - insert an operator in front of already-emitted operand
15173 * Means relocating the operand.
15176 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15182 const int offset = regarglen[(U8)op];
15183 const int size = NODE_STEP_REGNODE + offset;
15184 GET_RE_DEBUG_FLAGS_DECL;
15186 PERL_ARGS_ASSERT_REGINSERT;
15187 PERL_UNUSED_ARG(depth);
15188 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15189 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15198 if (RExC_open_parens) {
15200 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15201 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15202 if ( RExC_open_parens[paren] >= opnd ) {
15203 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15204 RExC_open_parens[paren] += size;
15206 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15208 if ( RExC_close_parens[paren] >= opnd ) {
15209 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15210 RExC_close_parens[paren] += size;
15212 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15217 while (src > opnd) {
15218 StructCopy(--src, --dst, regnode);
15219 #ifdef RE_TRACK_PATTERN_OFFSETS
15220 if (RExC_offsets) { /* MJD 20010112 */
15222 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15226 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15227 ? "Overwriting end of array!\n" : "OK",
15228 (UV)(src - RExC_emit_start),
15229 (UV)(dst - RExC_emit_start),
15230 (UV)RExC_offsets[0]));
15231 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15232 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15238 place = opnd; /* Op node, where operand used to be. */
15239 #ifdef RE_TRACK_PATTERN_OFFSETS
15240 if (RExC_offsets) { /* MJD */
15242 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15246 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15247 ? "Overwriting end of array!\n" : "OK",
15248 (UV)(place - RExC_emit_start),
15249 (UV)(RExC_parse - RExC_start),
15250 (UV)RExC_offsets[0]));
15251 Set_Node_Offset(place, RExC_parse);
15252 Set_Node_Length(place, 1);
15255 src = NEXTOPER(place);
15256 FILL_ADVANCE_NODE(place, op);
15257 Zero(src, offset, regnode);
15261 - regtail - set the next-pointer at the end of a node chain of p to val.
15262 - SEE ALSO: regtail_study
15264 /* TODO: All three parms should be const */
15266 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15267 const regnode *val,U32 depth)
15271 GET_RE_DEBUG_FLAGS_DECL;
15273 PERL_ARGS_ASSERT_REGTAIL;
15275 PERL_UNUSED_ARG(depth);
15281 /* Find last node. */
15284 regnode * const temp = regnext(scan);
15286 SV * const mysv=sv_newmortal();
15287 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15288 regprop(RExC_rx, mysv, scan, NULL);
15289 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15290 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
15291 (temp == NULL ? "->" : ""),
15292 (temp == NULL ? PL_reg_name[OP(val)] : "")
15300 if (reg_off_by_arg[OP(scan)]) {
15301 ARG_SET(scan, val - scan);
15304 NEXT_OFF(scan) = val - scan;
15310 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15311 - Look for optimizable sequences at the same time.
15312 - currently only looks for EXACT chains.
15314 This is experimental code. The idea is to use this routine to perform
15315 in place optimizations on branches and groups as they are constructed,
15316 with the long term intention of removing optimization from study_chunk so
15317 that it is purely analytical.
15319 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15320 to control which is which.
15323 /* TODO: All four parms should be const */
15326 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15327 const regnode *val,U32 depth)
15332 #ifdef EXPERIMENTAL_INPLACESCAN
15335 GET_RE_DEBUG_FLAGS_DECL;
15337 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15343 /* Find last node. */
15347 regnode * const temp = regnext(scan);
15348 #ifdef EXPERIMENTAL_INPLACESCAN
15349 if (PL_regkind[OP(scan)] == EXACT) {
15350 bool unfolded_multi_char; /* Unexamined in this routine */
15351 if (join_exact(pRExC_state, scan, &min,
15352 &unfolded_multi_char, 1, val, depth+1))
15357 switch (OP(scan)) {
15360 case EXACTFA_NO_TRIE:
15365 if( exact == PSEUDO )
15367 else if ( exact != OP(scan) )
15376 SV * const mysv=sv_newmortal();
15377 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
15378 regprop(RExC_rx, mysv, scan, NULL);
15379 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
15380 SvPV_nolen_const(mysv),
15381 REG_NODE_NUM(scan),
15382 PL_reg_name[exact]);
15389 SV * const mysv_val=sv_newmortal();
15390 DEBUG_PARSE_MSG("");
15391 regprop(RExC_rx, mysv_val, val, NULL);
15392 PerlIO_printf(Perl_debug_log,
15393 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
15394 SvPV_nolen_const(mysv_val),
15395 (IV)REG_NODE_NUM(val),
15399 if (reg_off_by_arg[OP(scan)]) {
15400 ARG_SET(scan, val - scan);
15403 NEXT_OFF(scan) = val - scan;
15411 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
15416 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
15421 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15423 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
15424 if (flags & (1<<bit)) {
15425 if (!set++ && lead)
15426 PerlIO_printf(Perl_debug_log, "%s",lead);
15427 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
15432 PerlIO_printf(Perl_debug_log, "\n");
15434 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15439 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
15445 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15447 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
15448 if (flags & (1<<bit)) {
15449 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
15452 if (!set++ && lead)
15453 PerlIO_printf(Perl_debug_log, "%s",lead);
15454 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
15457 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
15458 if (!set++ && lead) {
15459 PerlIO_printf(Perl_debug_log, "%s",lead);
15462 case REGEX_UNICODE_CHARSET:
15463 PerlIO_printf(Perl_debug_log, "UNICODE");
15465 case REGEX_LOCALE_CHARSET:
15466 PerlIO_printf(Perl_debug_log, "LOCALE");
15468 case REGEX_ASCII_RESTRICTED_CHARSET:
15469 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
15471 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
15472 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
15475 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
15481 PerlIO_printf(Perl_debug_log, "\n");
15483 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15489 Perl_regdump(pTHX_ const regexp *r)
15493 SV * const sv = sv_newmortal();
15494 SV *dsv= sv_newmortal();
15495 RXi_GET_DECL(r,ri);
15496 GET_RE_DEBUG_FLAGS_DECL;
15498 PERL_ARGS_ASSERT_REGDUMP;
15500 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
15502 /* Header fields of interest. */
15503 if (r->anchored_substr) {
15504 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
15505 RE_SV_DUMPLEN(r->anchored_substr), 30);
15506 PerlIO_printf(Perl_debug_log,
15507 "anchored %s%s at %"IVdf" ",
15508 s, RE_SV_TAIL(r->anchored_substr),
15509 (IV)r->anchored_offset);
15510 } else if (r->anchored_utf8) {
15511 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
15512 RE_SV_DUMPLEN(r->anchored_utf8), 30);
15513 PerlIO_printf(Perl_debug_log,
15514 "anchored utf8 %s%s at %"IVdf" ",
15515 s, RE_SV_TAIL(r->anchored_utf8),
15516 (IV)r->anchored_offset);
15518 if (r->float_substr) {
15519 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
15520 RE_SV_DUMPLEN(r->float_substr), 30);
15521 PerlIO_printf(Perl_debug_log,
15522 "floating %s%s at %"IVdf"..%"UVuf" ",
15523 s, RE_SV_TAIL(r->float_substr),
15524 (IV)r->float_min_offset, (UV)r->float_max_offset);
15525 } else if (r->float_utf8) {
15526 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
15527 RE_SV_DUMPLEN(r->float_utf8), 30);
15528 PerlIO_printf(Perl_debug_log,
15529 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
15530 s, RE_SV_TAIL(r->float_utf8),
15531 (IV)r->float_min_offset, (UV)r->float_max_offset);
15533 if (r->check_substr || r->check_utf8)
15534 PerlIO_printf(Perl_debug_log,
15536 (r->check_substr == r->float_substr
15537 && r->check_utf8 == r->float_utf8
15538 ? "(checking floating" : "(checking anchored"));
15539 if (r->intflags & PREGf_NOSCAN)
15540 PerlIO_printf(Perl_debug_log, " noscan");
15541 if (r->extflags & RXf_CHECK_ALL)
15542 PerlIO_printf(Perl_debug_log, " isall");
15543 if (r->check_substr || r->check_utf8)
15544 PerlIO_printf(Perl_debug_log, ") ");
15546 if (ri->regstclass) {
15547 regprop(r, sv, ri->regstclass, NULL);
15548 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
15550 if (r->intflags & PREGf_ANCH) {
15551 PerlIO_printf(Perl_debug_log, "anchored");
15552 if (r->intflags & PREGf_ANCH_BOL)
15553 PerlIO_printf(Perl_debug_log, "(BOL)");
15554 if (r->intflags & PREGf_ANCH_MBOL)
15555 PerlIO_printf(Perl_debug_log, "(MBOL)");
15556 if (r->intflags & PREGf_ANCH_SBOL)
15557 PerlIO_printf(Perl_debug_log, "(SBOL)");
15558 if (r->intflags & PREGf_ANCH_GPOS)
15559 PerlIO_printf(Perl_debug_log, "(GPOS)");
15560 PerlIO_putc(Perl_debug_log, ' ');
15562 if (r->intflags & PREGf_GPOS_SEEN)
15563 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
15564 if (r->intflags & PREGf_SKIP)
15565 PerlIO_printf(Perl_debug_log, "plus ");
15566 if (r->intflags & PREGf_IMPLICIT)
15567 PerlIO_printf(Perl_debug_log, "implicit ");
15568 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
15569 if (r->extflags & RXf_EVAL_SEEN)
15570 PerlIO_printf(Perl_debug_log, "with eval ");
15571 PerlIO_printf(Perl_debug_log, "\n");
15573 regdump_extflags("r->extflags: ",r->extflags);
15574 regdump_intflags("r->intflags: ",r->intflags);
15577 PERL_ARGS_ASSERT_REGDUMP;
15578 PERL_UNUSED_CONTEXT;
15579 PERL_UNUSED_ARG(r);
15580 #endif /* DEBUGGING */
15584 - regprop - printable representation of opcode, with run time support
15588 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo)
15594 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
15595 static const char * const anyofs[] = {
15596 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
15597 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
15598 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
15599 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
15600 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
15601 || _CC_VERTSPACE != 16
15602 #error Need to adjust order of anyofs[]
15639 RXi_GET_DECL(prog,progi);
15640 GET_RE_DEBUG_FLAGS_DECL;
15642 PERL_ARGS_ASSERT_REGPROP;
15646 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
15647 /* It would be nice to FAIL() here, but this may be called from
15648 regexec.c, and it would be hard to supply pRExC_state. */
15649 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
15650 (int)OP(o), (int)REGNODE_MAX);
15651 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
15653 k = PL_regkind[OP(o)];
15656 sv_catpvs(sv, " ");
15657 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
15658 * is a crude hack but it may be the best for now since
15659 * we have no flag "this EXACTish node was UTF-8"
15661 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
15662 PERL_PV_ESCAPE_UNI_DETECT |
15663 PERL_PV_ESCAPE_NONASCII |
15664 PERL_PV_PRETTY_ELLIPSES |
15665 PERL_PV_PRETTY_LTGT |
15666 PERL_PV_PRETTY_NOCLEAR
15668 } else if (k == TRIE) {
15669 /* print the details of the trie in dumpuntil instead, as
15670 * progi->data isn't available here */
15671 const char op = OP(o);
15672 const U32 n = ARG(o);
15673 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
15674 (reg_ac_data *)progi->data->data[n] :
15676 const reg_trie_data * const trie
15677 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
15679 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
15680 DEBUG_TRIE_COMPILE_r(
15681 Perl_sv_catpvf(aTHX_ sv,
15682 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
15683 (UV)trie->startstate,
15684 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
15685 (UV)trie->wordcount,
15688 (UV)TRIE_CHARCOUNT(trie),
15689 (UV)trie->uniquecharcount
15692 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
15693 sv_catpvs(sv, "[");
15694 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
15696 : TRIE_BITMAP(trie));
15697 sv_catpvs(sv, "]");
15700 } else if (k == CURLY) {
15701 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
15702 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
15703 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
15705 else if (k == WHILEM && o->flags) /* Ordinal/of */
15706 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
15707 else if (k == REF || k == OPEN || k == CLOSE
15708 || k == GROUPP || OP(o)==ACCEPT)
15710 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
15711 if ( RXp_PAREN_NAMES(prog) ) {
15712 if ( k != REF || (OP(o) < NREF)) {
15713 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
15714 SV **name= av_fetch(list, ARG(o), 0 );
15716 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15719 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
15720 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
15721 I32 *nums=(I32*)SvPVX(sv_dat);
15722 SV **name= av_fetch(list, nums[0], 0 );
15725 for ( n=0; n<SvIVX(sv_dat); n++ ) {
15726 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
15727 (n ? "," : ""), (IV)nums[n]);
15729 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15733 if ( k == REF && reginfo) {
15734 U32 n = ARG(o); /* which paren pair */
15735 I32 ln = prog->offs[n].start;
15736 if (prog->lastparen < n || ln == -1)
15737 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
15738 else if (ln == prog->offs[n].end)
15739 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
15741 const char *s = reginfo->strbeg + ln;
15742 Perl_sv_catpvf(aTHX_ sv, ": ");
15743 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
15744 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
15747 } else if (k == GOSUB)
15748 /* Paren and offset */
15749 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
15750 else if (k == VERB) {
15752 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
15753 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
15754 } else if (k == LOGICAL)
15755 /* 2: embedded, otherwise 1 */
15756 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
15757 else if (k == ANYOF) {
15758 const U8 flags = ANYOF_FLAGS(o);
15762 if (flags & ANYOF_LOCALE_FLAGS)
15763 sv_catpvs(sv, "{loc}");
15764 if (flags & ANYOF_LOC_FOLD)
15765 sv_catpvs(sv, "{i}");
15766 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
15767 if (flags & ANYOF_INVERT)
15768 sv_catpvs(sv, "^");
15770 /* output what the standard cp 0-255 bitmap matches */
15771 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
15773 /* output any special charclass tests (used entirely under use
15775 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
15777 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
15778 if (ANYOF_POSIXL_TEST(o,i)) {
15779 sv_catpv(sv, anyofs[i]);
15785 if ((flags & (ANYOF_ABOVE_LATIN1_ALL
15787 |ANYOF_NONBITMAP_NON_UTF8
15791 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
15792 if (flags & ANYOF_INVERT)
15793 /*make sure the invert info is in each */
15794 sv_catpvs(sv, "^");
15797 if (flags & ANYOF_NON_UTF8_NON_ASCII_ALL) {
15798 sv_catpvs(sv, "{non-utf8-latin1-all}");
15801 /* output information about the unicode matching */
15802 if (flags & ANYOF_ABOVE_LATIN1_ALL)
15803 sv_catpvs(sv, "{unicode_all}");
15804 else if (ARG(o) != ANYOF_NONBITMAP_EMPTY) {
15805 SV *lv; /* Set if there is something outside the bit map. */
15806 bool byte_output = FALSE; /* If something in the bitmap has
15808 SV *only_utf8_locale;
15810 /* Get the stuff that wasn't in the bitmap */
15811 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
15812 &lv, &only_utf8_locale);
15813 if (lv && lv != &PL_sv_undef) {
15814 char *s = savesvpv(lv);
15815 char * const origs = s;
15817 while (*s && *s != '\n')
15821 const char * const t = ++s;
15823 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
15824 sv_catpvs(sv, "{outside bitmap}");
15827 sv_catpvs(sv, "{utf8}");
15831 sv_catpvs(sv, " ");
15837 /* Truncate very long output */
15838 if (s - origs > 256) {
15839 Perl_sv_catpvf(aTHX_ sv,
15841 (int) (s - origs - 1),
15847 else if (*s == '\t') {
15861 SvREFCNT_dec_NN(lv);
15864 if ((flags & ANYOF_LOC_FOLD)
15865 && only_utf8_locale
15866 && only_utf8_locale != &PL_sv_undef)
15869 int max_entries = 256;
15871 sv_catpvs(sv, "{utf8 locale}");
15872 invlist_iterinit(only_utf8_locale);
15873 while (invlist_iternext(only_utf8_locale,
15875 put_range(sv, start, end);
15877 if (max_entries < 0) {
15878 sv_catpvs(sv, "...");
15882 invlist_iterfinish(only_utf8_locale);
15887 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
15889 else if (k == POSIXD || k == NPOSIXD) {
15890 U8 index = FLAGS(o) * 2;
15891 if (index < C_ARRAY_LENGTH(anyofs)) {
15892 if (*anyofs[index] != '[') {
15895 sv_catpv(sv, anyofs[index]);
15896 if (*anyofs[index] != '[') {
15901 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
15904 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
15905 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
15907 PERL_UNUSED_CONTEXT;
15908 PERL_UNUSED_ARG(sv);
15909 PERL_UNUSED_ARG(o);
15910 PERL_UNUSED_ARG(prog);
15911 PERL_UNUSED_ARG(reginfo);
15912 #endif /* DEBUGGING */
15918 Perl_re_intuit_string(pTHX_ REGEXP * const r)
15919 { /* Assume that RE_INTUIT is set */
15921 struct regexp *const prog = ReANY(r);
15922 GET_RE_DEBUG_FLAGS_DECL;
15924 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
15925 PERL_UNUSED_CONTEXT;
15929 const char * const s = SvPV_nolen_const(prog->check_substr
15930 ? prog->check_substr : prog->check_utf8);
15932 if (!PL_colorset) reginitcolors();
15933 PerlIO_printf(Perl_debug_log,
15934 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
15936 prog->check_substr ? "" : "utf8 ",
15937 PL_colors[5],PL_colors[0],
15940 (strlen(s) > 60 ? "..." : ""));
15943 return prog->check_substr ? prog->check_substr : prog->check_utf8;
15949 handles refcounting and freeing the perl core regexp structure. When
15950 it is necessary to actually free the structure the first thing it
15951 does is call the 'free' method of the regexp_engine associated to
15952 the regexp, allowing the handling of the void *pprivate; member
15953 first. (This routine is not overridable by extensions, which is why
15954 the extensions free is called first.)
15956 See regdupe and regdupe_internal if you change anything here.
15958 #ifndef PERL_IN_XSUB_RE
15960 Perl_pregfree(pTHX_ REGEXP *r)
15966 Perl_pregfree2(pTHX_ REGEXP *rx)
15969 struct regexp *const r = ReANY(rx);
15970 GET_RE_DEBUG_FLAGS_DECL;
15972 PERL_ARGS_ASSERT_PREGFREE2;
15974 if (r->mother_re) {
15975 ReREFCNT_dec(r->mother_re);
15977 CALLREGFREE_PVT(rx); /* free the private data */
15978 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15979 Safefree(r->xpv_len_u.xpvlenu_pv);
15982 SvREFCNT_dec(r->anchored_substr);
15983 SvREFCNT_dec(r->anchored_utf8);
15984 SvREFCNT_dec(r->float_substr);
15985 SvREFCNT_dec(r->float_utf8);
15986 Safefree(r->substrs);
15988 RX_MATCH_COPY_FREE(rx);
15989 #ifdef PERL_ANY_COW
15990 SvREFCNT_dec(r->saved_copy);
15993 SvREFCNT_dec(r->qr_anoncv);
15994 rx->sv_u.svu_rx = 0;
15999 This is a hacky workaround to the structural issue of match results
16000 being stored in the regexp structure which is in turn stored in
16001 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
16002 could be PL_curpm in multiple contexts, and could require multiple
16003 result sets being associated with the pattern simultaneously, such
16004 as when doing a recursive match with (??{$qr})
16006 The solution is to make a lightweight copy of the regexp structure
16007 when a qr// is returned from the code executed by (??{$qr}) this
16008 lightweight copy doesn't actually own any of its data except for
16009 the starp/end and the actual regexp structure itself.
16015 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
16017 struct regexp *ret;
16018 struct regexp *const r = ReANY(rx);
16019 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
16021 PERL_ARGS_ASSERT_REG_TEMP_COPY;
16024 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
16026 SvOK_off((SV *)ret_x);
16028 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
16029 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
16030 made both spots point to the same regexp body.) */
16031 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
16032 assert(!SvPVX(ret_x));
16033 ret_x->sv_u.svu_rx = temp->sv_any;
16034 temp->sv_any = NULL;
16035 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
16036 SvREFCNT_dec_NN(temp);
16037 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
16038 ing below will not set it. */
16039 SvCUR_set(ret_x, SvCUR(rx));
16042 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
16043 sv_force_normal(sv) is called. */
16045 ret = ReANY(ret_x);
16047 SvFLAGS(ret_x) |= SvUTF8(rx);
16048 /* We share the same string buffer as the original regexp, on which we
16049 hold a reference count, incremented when mother_re is set below.
16050 The string pointer is copied here, being part of the regexp struct.
16052 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
16053 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
16055 const I32 npar = r->nparens+1;
16056 Newx(ret->offs, npar, regexp_paren_pair);
16057 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16060 Newx(ret->substrs, 1, struct reg_substr_data);
16061 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16063 SvREFCNT_inc_void(ret->anchored_substr);
16064 SvREFCNT_inc_void(ret->anchored_utf8);
16065 SvREFCNT_inc_void(ret->float_substr);
16066 SvREFCNT_inc_void(ret->float_utf8);
16068 /* check_substr and check_utf8, if non-NULL, point to either their
16069 anchored or float namesakes, and don't hold a second reference. */
16071 RX_MATCH_COPIED_off(ret_x);
16072 #ifdef PERL_ANY_COW
16073 ret->saved_copy = NULL;
16075 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
16076 SvREFCNT_inc_void(ret->qr_anoncv);
16082 /* regfree_internal()
16084 Free the private data in a regexp. This is overloadable by
16085 extensions. Perl takes care of the regexp structure in pregfree(),
16086 this covers the *pprivate pointer which technically perl doesn't
16087 know about, however of course we have to handle the
16088 regexp_internal structure when no extension is in use.
16090 Note this is called before freeing anything in the regexp
16095 Perl_regfree_internal(pTHX_ REGEXP * const rx)
16098 struct regexp *const r = ReANY(rx);
16099 RXi_GET_DECL(r,ri);
16100 GET_RE_DEBUG_FLAGS_DECL;
16102 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
16108 SV *dsv= sv_newmortal();
16109 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
16110 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
16111 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
16112 PL_colors[4],PL_colors[5],s);
16115 #ifdef RE_TRACK_PATTERN_OFFSETS
16117 Safefree(ri->u.offsets); /* 20010421 MJD */
16119 if (ri->code_blocks) {
16121 for (n = 0; n < ri->num_code_blocks; n++)
16122 SvREFCNT_dec(ri->code_blocks[n].src_regex);
16123 Safefree(ri->code_blocks);
16127 int n = ri->data->count;
16130 /* If you add a ->what type here, update the comment in regcomp.h */
16131 switch (ri->data->what[n]) {
16137 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
16140 Safefree(ri->data->data[n]);
16146 { /* Aho Corasick add-on structure for a trie node.
16147 Used in stclass optimization only */
16149 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16151 refcount = --aho->refcount;
16154 PerlMemShared_free(aho->states);
16155 PerlMemShared_free(aho->fail);
16156 /* do this last!!!! */
16157 PerlMemShared_free(ri->data->data[n]);
16158 PerlMemShared_free(ri->regstclass);
16164 /* trie structure. */
16166 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16168 refcount = --trie->refcount;
16171 PerlMemShared_free(trie->charmap);
16172 PerlMemShared_free(trie->states);
16173 PerlMemShared_free(trie->trans);
16175 PerlMemShared_free(trie->bitmap);
16177 PerlMemShared_free(trie->jump);
16178 PerlMemShared_free(trie->wordinfo);
16179 /* do this last!!!! */
16180 PerlMemShared_free(ri->data->data[n]);
16185 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16186 ri->data->what[n]);
16189 Safefree(ri->data->what);
16190 Safefree(ri->data);
16196 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16197 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16198 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16201 re_dup - duplicate a regexp.
16203 This routine is expected to clone a given regexp structure. It is only
16204 compiled under USE_ITHREADS.
16206 After all of the core data stored in struct regexp is duplicated
16207 the regexp_engine.dupe method is used to copy any private data
16208 stored in the *pprivate pointer. This allows extensions to handle
16209 any duplication it needs to do.
16211 See pregfree() and regfree_internal() if you change anything here.
16213 #if defined(USE_ITHREADS)
16214 #ifndef PERL_IN_XSUB_RE
16216 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16220 const struct regexp *r = ReANY(sstr);
16221 struct regexp *ret = ReANY(dstr);
16223 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16225 npar = r->nparens+1;
16226 Newx(ret->offs, npar, regexp_paren_pair);
16227 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16229 if (ret->substrs) {
16230 /* Do it this way to avoid reading from *r after the StructCopy().
16231 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16232 cache, it doesn't matter. */
16233 const bool anchored = r->check_substr
16234 ? r->check_substr == r->anchored_substr
16235 : r->check_utf8 == r->anchored_utf8;
16236 Newx(ret->substrs, 1, struct reg_substr_data);
16237 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16239 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16240 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16241 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16242 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16244 /* check_substr and check_utf8, if non-NULL, point to either their
16245 anchored or float namesakes, and don't hold a second reference. */
16247 if (ret->check_substr) {
16249 assert(r->check_utf8 == r->anchored_utf8);
16250 ret->check_substr = ret->anchored_substr;
16251 ret->check_utf8 = ret->anchored_utf8;
16253 assert(r->check_substr == r->float_substr);
16254 assert(r->check_utf8 == r->float_utf8);
16255 ret->check_substr = ret->float_substr;
16256 ret->check_utf8 = ret->float_utf8;
16258 } else if (ret->check_utf8) {
16260 ret->check_utf8 = ret->anchored_utf8;
16262 ret->check_utf8 = ret->float_utf8;
16267 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16268 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16271 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16273 if (RX_MATCH_COPIED(dstr))
16274 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16276 ret->subbeg = NULL;
16277 #ifdef PERL_ANY_COW
16278 ret->saved_copy = NULL;
16281 /* Whether mother_re be set or no, we need to copy the string. We
16282 cannot refrain from copying it when the storage points directly to
16283 our mother regexp, because that's
16284 1: a buffer in a different thread
16285 2: something we no longer hold a reference on
16286 so we need to copy it locally. */
16287 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16288 ret->mother_re = NULL;
16290 #endif /* PERL_IN_XSUB_RE */
16295 This is the internal complement to regdupe() which is used to copy
16296 the structure pointed to by the *pprivate pointer in the regexp.
16297 This is the core version of the extension overridable cloning hook.
16298 The regexp structure being duplicated will be copied by perl prior
16299 to this and will be provided as the regexp *r argument, however
16300 with the /old/ structures pprivate pointer value. Thus this routine
16301 may override any copying normally done by perl.
16303 It returns a pointer to the new regexp_internal structure.
16307 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
16310 struct regexp *const r = ReANY(rx);
16311 regexp_internal *reti;
16313 RXi_GET_DECL(r,ri);
16315 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
16319 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
16320 char, regexp_internal);
16321 Copy(ri->program, reti->program, len+1, regnode);
16323 reti->num_code_blocks = ri->num_code_blocks;
16324 if (ri->code_blocks) {
16326 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
16327 struct reg_code_block);
16328 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
16329 struct reg_code_block);
16330 for (n = 0; n < ri->num_code_blocks; n++)
16331 reti->code_blocks[n].src_regex = (REGEXP*)
16332 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
16335 reti->code_blocks = NULL;
16337 reti->regstclass = NULL;
16340 struct reg_data *d;
16341 const int count = ri->data->count;
16344 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
16345 char, struct reg_data);
16346 Newx(d->what, count, U8);
16349 for (i = 0; i < count; i++) {
16350 d->what[i] = ri->data->what[i];
16351 switch (d->what[i]) {
16352 /* see also regcomp.h and regfree_internal() */
16353 case 'a': /* actually an AV, but the dup function is identical. */
16357 case 'u': /* actually an HV, but the dup function is identical. */
16358 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
16361 /* This is cheating. */
16362 Newx(d->data[i], 1, regnode_ssc);
16363 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
16364 reti->regstclass = (regnode*)d->data[i];
16367 /* Trie stclasses are readonly and can thus be shared
16368 * without duplication. We free the stclass in pregfree
16369 * when the corresponding reg_ac_data struct is freed.
16371 reti->regstclass= ri->regstclass;
16375 ((reg_trie_data*)ri->data->data[i])->refcount++;
16380 d->data[i] = ri->data->data[i];
16383 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
16384 ri->data->what[i]);
16393 reti->name_list_idx = ri->name_list_idx;
16395 #ifdef RE_TRACK_PATTERN_OFFSETS
16396 if (ri->u.offsets) {
16397 Newx(reti->u.offsets, 2*len+1, U32);
16398 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
16401 SetProgLen(reti,len);
16404 return (void*)reti;
16407 #endif /* USE_ITHREADS */
16409 #ifndef PERL_IN_XSUB_RE
16412 - regnext - dig the "next" pointer out of a node
16415 Perl_regnext(pTHX_ regnode *p)
16423 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
16424 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16425 (int)OP(p), (int)REGNODE_MAX);
16428 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
16437 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
16440 STRLEN l1 = strlen(pat1);
16441 STRLEN l2 = strlen(pat2);
16444 const char *message;
16446 PERL_ARGS_ASSERT_RE_CROAK2;
16452 Copy(pat1, buf, l1 , char);
16453 Copy(pat2, buf + l1, l2 , char);
16454 buf[l1 + l2] = '\n';
16455 buf[l1 + l2 + 1] = '\0';
16456 va_start(args, pat2);
16457 msv = vmess(buf, &args);
16459 message = SvPV_const(msv,l1);
16462 Copy(message, buf, l1 , char);
16463 /* l1-1 to avoid \n */
16464 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
16467 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
16469 #ifndef PERL_IN_XSUB_RE
16471 Perl_save_re_context(pTHX)
16475 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
16477 const REGEXP * const rx = PM_GETRE(PL_curpm);
16480 for (i = 1; i <= RX_NPARENS(rx); i++) {
16481 char digits[TYPE_CHARS(long)];
16482 const STRLEN len = my_snprintf(digits, sizeof(digits),
16484 GV *const *const gvp
16485 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
16488 GV * const gv = *gvp;
16489 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
16501 S_put_byte(pTHX_ SV *sv, int c)
16503 PERL_ARGS_ASSERT_PUT_BYTE;
16507 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
16508 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
16509 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
16510 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
16511 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
16514 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
16519 const char string = c;
16520 if (c == '-' || c == ']' || c == '\\' || c == '^')
16521 sv_catpvs(sv, "\\");
16522 sv_catpvn(sv, &string, 1);
16527 S_put_range(pTHX_ SV *sv, UV start, UV end)
16530 /* Appends to 'sv' a displayable version of the range of code points from
16531 * 'start' to 'end' */
16533 assert(start <= end);
16535 PERL_ARGS_ASSERT_PUT_RANGE;
16537 if (end - start < 3) { /* Individual chars in short ranges */
16538 for (; start <= end; start++)
16539 put_byte(sv, start);
16541 else if ( end > 255
16542 || ! isALPHANUMERIC(start)
16543 || ! isALPHANUMERIC(end)
16544 || isDIGIT(start) != isDIGIT(end)
16545 || isUPPER(start) != isUPPER(end)
16546 || isLOWER(start) != isLOWER(end)
16548 /* This final test should get optimized out except on EBCDIC
16549 * platforms, where it causes ranges that cross discontinuities
16550 * like i/j to be shown as hex instead of the misleading,
16551 * e.g. H-K (since that range includes more than H, I, J, K).
16553 || (end - start) != NATIVE_TO_ASCII(end) - NATIVE_TO_ASCII(start))
16555 Perl_sv_catpvf(aTHX_ sv, "\\x{%02" UVXf "}-\\x{%02" UVXf "}",
16557 (end < 256) ? end : 255);
16559 else { /* Here, the ends of the range are both digits, or both uppercase,
16560 or both lowercase; and there's no discontinuity in the range
16561 (which could happen on EBCDIC platforms) */
16562 put_byte(sv, start);
16563 sv_catpvs(sv, "-");
16569 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
16571 /* Appends to 'sv' a displayable version of the innards of the bracketed
16572 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
16573 * output anything */
16576 bool has_output_anything = FALSE;
16578 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
16580 for (i = 0; i < 256; i++) {
16581 if (BITMAP_TEST((U8 *) bitmap,i)) {
16583 /* The character at index i should be output. Find the next
16584 * character that should NOT be output */
16586 for (j = i + 1; j < 256; j++) {
16587 if (! BITMAP_TEST((U8 *) bitmap, j)) {
16592 /* Everything between them is a single range that should be output
16594 put_range(sv, i, j - 1);
16595 has_output_anything = TRUE;
16600 return has_output_anything;
16603 #define CLEAR_OPTSTART \
16604 if (optstart) STMT_START { \
16605 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
16606 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
16610 #define DUMPUNTIL(b,e) \
16612 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
16614 STATIC const regnode *
16615 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
16616 const regnode *last, const regnode *plast,
16617 SV* sv, I32 indent, U32 depth)
16620 U8 op = PSEUDO; /* Arbitrary non-END op. */
16621 const regnode *next;
16622 const regnode *optstart= NULL;
16624 RXi_GET_DECL(r,ri);
16625 GET_RE_DEBUG_FLAGS_DECL;
16627 PERL_ARGS_ASSERT_DUMPUNTIL;
16629 #ifdef DEBUG_DUMPUNTIL
16630 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
16631 last ? last-start : 0,plast ? plast-start : 0);
16634 if (plast && plast < last)
16637 while (PL_regkind[op] != END && (!last || node < last)) {
16638 /* While that wasn't END last time... */
16641 if (op == CLOSE || op == WHILEM)
16643 next = regnext((regnode *)node);
16646 if (OP(node) == OPTIMIZED) {
16647 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
16654 regprop(r, sv, node, NULL);
16655 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
16656 (int)(2*indent + 1), "", SvPVX_const(sv));
16658 if (OP(node) != OPTIMIZED) {
16659 if (next == NULL) /* Next ptr. */
16660 PerlIO_printf(Perl_debug_log, " (0)");
16661 else if (PL_regkind[(U8)op] == BRANCH
16662 && PL_regkind[OP(next)] != BRANCH )
16663 PerlIO_printf(Perl_debug_log, " (FAIL)");
16665 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
16666 (void)PerlIO_putc(Perl_debug_log, '\n');
16670 if (PL_regkind[(U8)op] == BRANCHJ) {
16673 const regnode *nnode = (OP(next) == LONGJMP
16674 ? regnext((regnode *)next)
16676 if (last && nnode > last)
16678 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
16681 else if (PL_regkind[(U8)op] == BRANCH) {
16683 DUMPUNTIL(NEXTOPER(node), next);
16685 else if ( PL_regkind[(U8)op] == TRIE ) {
16686 const regnode *this_trie = node;
16687 const char op = OP(node);
16688 const U32 n = ARG(node);
16689 const reg_ac_data * const ac = op>=AHOCORASICK ?
16690 (reg_ac_data *)ri->data->data[n] :
16692 const reg_trie_data * const trie =
16693 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
16695 AV *const trie_words
16696 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
16698 const regnode *nextbranch= NULL;
16701 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
16702 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
16704 PerlIO_printf(Perl_debug_log, "%*s%s ",
16705 (int)(2*(indent+3)), "",
16707 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
16708 SvCUR(*elem_ptr), 60,
16709 PL_colors[0], PL_colors[1],
16711 ? PERL_PV_ESCAPE_UNI
16713 | PERL_PV_PRETTY_ELLIPSES
16714 | PERL_PV_PRETTY_LTGT
16719 U16 dist= trie->jump[word_idx+1];
16720 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
16721 (UV)((dist ? this_trie + dist : next) - start));
16724 nextbranch= this_trie + trie->jump[0];
16725 DUMPUNTIL(this_trie + dist, nextbranch);
16727 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
16728 nextbranch= regnext((regnode *)nextbranch);
16730 PerlIO_printf(Perl_debug_log, "\n");
16733 if (last && next > last)
16738 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
16739 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
16740 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
16742 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
16744 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
16746 else if ( op == PLUS || op == STAR) {
16747 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
16749 else if (PL_regkind[(U8)op] == ANYOF) {
16750 /* arglen 1 + class block */
16751 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_POSIXL)
16752 ? ANYOF_POSIXL_SKIP
16754 node = NEXTOPER(node);
16756 else if (PL_regkind[(U8)op] == EXACT) {
16757 /* Literal string, where present. */
16758 node += NODE_SZ_STR(node) - 1;
16759 node = NEXTOPER(node);
16762 node = NEXTOPER(node);
16763 node += regarglen[(U8)op];
16765 if (op == CURLYX || op == OPEN)
16769 #ifdef DEBUG_DUMPUNTIL
16770 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
16775 #endif /* DEBUGGING */
16779 * c-indentation-style: bsd
16780 * c-basic-offset: 4
16781 * indent-tabs-mode: nil
16784 * ex: set ts=8 sts=4 sw=4 et: