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 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1449 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1452 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1455 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1456 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1457 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1458 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1459 ? (TRIE_LIST_CUR( idx ) - 1) \
1465 dump_trie(trie,widecharmap,revcharmap)
1466 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1467 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1469 These routines dump out a trie in a somewhat readable format.
1470 The _interim_ variants are used for debugging the interim
1471 tables that are used to generate the final compressed
1472 representation which is what dump_trie expects.
1474 Part of the reason for their existence is to provide a form
1475 of documentation as to how the different representations function.
1480 Dumps the final compressed table form of the trie to Perl_debug_log.
1481 Used for debugging make_trie().
1485 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1486 AV *revcharmap, U32 depth)
1489 SV *sv=sv_newmortal();
1490 int colwidth= widecharmap ? 6 : 4;
1492 GET_RE_DEBUG_FLAGS_DECL;
1494 PERL_ARGS_ASSERT_DUMP_TRIE;
1496 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1497 (int)depth * 2 + 2,"",
1498 "Match","Base","Ofs" );
1500 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1501 SV ** const tmp = av_fetch( revcharmap, state, 0);
1503 PerlIO_printf( Perl_debug_log, "%*s",
1505 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1506 PL_colors[0], PL_colors[1],
1507 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1508 PERL_PV_ESCAPE_FIRSTCHAR
1513 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1514 (int)depth * 2 + 2,"");
1516 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1517 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1518 PerlIO_printf( Perl_debug_log, "\n");
1520 for( state = 1 ; state < trie->statecount ; state++ ) {
1521 const U32 base = trie->states[ state ].trans.base;
1523 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1524 (int)depth * 2 + 2,"", (UV)state);
1526 if ( trie->states[ state ].wordnum ) {
1527 PerlIO_printf( Perl_debug_log, " W%4X",
1528 trie->states[ state ].wordnum );
1530 PerlIO_printf( Perl_debug_log, "%6s", "" );
1533 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1538 while( ( base + ofs < trie->uniquecharcount ) ||
1539 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1540 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1544 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1546 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1547 if ( ( base + ofs >= trie->uniquecharcount )
1548 && ( base + ofs - trie->uniquecharcount
1550 && trie->trans[ base + ofs
1551 - trie->uniquecharcount ].check == state )
1553 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1555 (UV)trie->trans[ base + ofs
1556 - trie->uniquecharcount ].next );
1558 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1562 PerlIO_printf( Perl_debug_log, "]");
1565 PerlIO_printf( Perl_debug_log, "\n" );
1567 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1569 for (word=1; word <= trie->wordcount; word++) {
1570 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1571 (int)word, (int)(trie->wordinfo[word].prev),
1572 (int)(trie->wordinfo[word].len));
1574 PerlIO_printf(Perl_debug_log, "\n" );
1577 Dumps a fully constructed but uncompressed trie in list form.
1578 List tries normally only are used for construction when the number of
1579 possible chars (trie->uniquecharcount) is very high.
1580 Used for debugging make_trie().
1583 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1584 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1588 SV *sv=sv_newmortal();
1589 int colwidth= widecharmap ? 6 : 4;
1590 GET_RE_DEBUG_FLAGS_DECL;
1592 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1594 /* print out the table precompression. */
1595 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1596 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1597 "------:-----+-----------------\n" );
1599 for( state=1 ; state < next_alloc ; state ++ ) {
1602 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1603 (int)depth * 2 + 2,"", (UV)state );
1604 if ( ! trie->states[ state ].wordnum ) {
1605 PerlIO_printf( Perl_debug_log, "%5s| ","");
1607 PerlIO_printf( Perl_debug_log, "W%4x| ",
1608 trie->states[ state ].wordnum
1611 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1612 SV ** const tmp = av_fetch( revcharmap,
1613 TRIE_LIST_ITEM(state,charid).forid, 0);
1615 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1617 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1619 PL_colors[0], PL_colors[1],
1620 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1621 | PERL_PV_ESCAPE_FIRSTCHAR
1623 TRIE_LIST_ITEM(state,charid).forid,
1624 (UV)TRIE_LIST_ITEM(state,charid).newstate
1627 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1628 (int)((depth * 2) + 14), "");
1631 PerlIO_printf( Perl_debug_log, "\n");
1636 Dumps a fully constructed but uncompressed trie in table form.
1637 This is the normal DFA style state transition table, with a few
1638 twists to facilitate compression later.
1639 Used for debugging make_trie().
1642 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1643 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1648 SV *sv=sv_newmortal();
1649 int colwidth= widecharmap ? 6 : 4;
1650 GET_RE_DEBUG_FLAGS_DECL;
1652 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1655 print out the table precompression so that we can do a visual check
1656 that they are identical.
1659 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1661 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1662 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1664 PerlIO_printf( Perl_debug_log, "%*s",
1666 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1667 PL_colors[0], PL_colors[1],
1668 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1669 PERL_PV_ESCAPE_FIRSTCHAR
1675 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1677 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1678 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1681 PerlIO_printf( Perl_debug_log, "\n" );
1683 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1685 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1686 (int)depth * 2 + 2,"",
1687 (UV)TRIE_NODENUM( state ) );
1689 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1690 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1692 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1694 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1696 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1697 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1698 (UV)trie->trans[ state ].check );
1700 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1701 (UV)trie->trans[ state ].check,
1702 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1710 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1711 startbranch: the first branch in the whole branch sequence
1712 first : start branch of sequence of branch-exact nodes.
1713 May be the same as startbranch
1714 last : Thing following the last branch.
1715 May be the same as tail.
1716 tail : item following the branch sequence
1717 count : words in the sequence
1718 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1719 depth : indent depth
1721 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1723 A trie is an N'ary tree where the branches are determined by digital
1724 decomposition of the key. IE, at the root node you look up the 1st character and
1725 follow that branch repeat until you find the end of the branches. Nodes can be
1726 marked as "accepting" meaning they represent a complete word. Eg:
1730 would convert into the following structure. Numbers represent states, letters
1731 following numbers represent valid transitions on the letter from that state, if
1732 the number is in square brackets it represents an accepting state, otherwise it
1733 will be in parenthesis.
1735 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1739 (1) +-i->(6)-+-s->[7]
1741 +-s->(3)-+-h->(4)-+-e->[5]
1743 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1745 This shows that when matching against the string 'hers' we will begin at state 1
1746 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1747 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1748 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1749 single traverse. We store a mapping from accepting to state to which word was
1750 matched, and then when we have multiple possibilities we try to complete the
1751 rest of the regex in the order in which they occured in the alternation.
1753 The only prior NFA like behaviour that would be changed by the TRIE support is
1754 the silent ignoring of duplicate alternations which are of the form:
1756 / (DUPE|DUPE) X? (?{ ... }) Y /x
1758 Thus EVAL blocks following a trie may be called a different number of times with
1759 and without the optimisation. With the optimisations dupes will be silently
1760 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1761 the following demonstrates:
1763 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1765 which prints out 'word' three times, but
1767 'words'=~/(word|word|word)(?{ print $1 })S/
1769 which doesnt print it out at all. This is due to other optimisations kicking in.
1771 Example of what happens on a structural level:
1773 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1775 1: CURLYM[1] {1,32767}(18)
1786 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1787 and should turn into:
1789 1: CURLYM[1] {1,32767}(18)
1791 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1799 Cases where tail != last would be like /(?foo|bar)baz/:
1809 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1810 and would end up looking like:
1813 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1820 d = uvchr_to_utf8_flags(d, uv, 0);
1822 is the recommended Unicode-aware way of saying
1827 #define TRIE_STORE_REVCHAR(val) \
1830 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1831 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1832 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1833 SvCUR_set(zlopp, kapow - flrbbbbb); \
1836 av_push(revcharmap, zlopp); \
1838 char ooooff = (char)val; \
1839 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1843 /* This gets the next character from the input, folding it if not already
1845 #define TRIE_READ_CHAR STMT_START { \
1848 /* if it is UTF then it is either already folded, or does not need \
1850 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1852 else if (folder == PL_fold_latin1) { \
1853 /* This folder implies Unicode rules, which in the range expressible \
1854 * by not UTF is the lower case, with the two exceptions, one of \
1855 * which should have been taken care of before calling this */ \
1856 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1857 uvc = toLOWER_L1(*uc); \
1858 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1861 /* raw data, will be folded later if needed */ \
1869 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1870 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1871 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1872 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1874 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1875 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1876 TRIE_LIST_CUR( state )++; \
1879 #define TRIE_LIST_NEW(state) STMT_START { \
1880 Newxz( trie->states[ state ].trans.list, \
1881 4, reg_trie_trans_le ); \
1882 TRIE_LIST_CUR( state ) = 1; \
1883 TRIE_LIST_LEN( state ) = 4; \
1886 #define TRIE_HANDLE_WORD(state) STMT_START { \
1887 U16 dupe= trie->states[ state ].wordnum; \
1888 regnode * const noper_next = regnext( noper ); \
1891 /* store the word for dumping */ \
1893 if (OP(noper) != NOTHING) \
1894 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1896 tmp = newSVpvn_utf8( "", 0, UTF ); \
1897 av_push( trie_words, tmp ); \
1901 trie->wordinfo[curword].prev = 0; \
1902 trie->wordinfo[curword].len = wordlen; \
1903 trie->wordinfo[curword].accept = state; \
1905 if ( noper_next < tail ) { \
1907 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1909 trie->jump[curword] = (U16)(noper_next - convert); \
1911 jumper = noper_next; \
1913 nextbranch= regnext(cur); \
1917 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1918 /* chain, so that when the bits of chain are later */\
1919 /* linked together, the dups appear in the chain */\
1920 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1921 trie->wordinfo[dupe].prev = curword; \
1923 /* we haven't inserted this word yet. */ \
1924 trie->states[ state ].wordnum = curword; \
1929 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1930 ( ( base + charid >= ucharcount \
1931 && base + charid < ubound \
1932 && state == trie->trans[ base - ucharcount + charid ].check \
1933 && trie->trans[ base - ucharcount + charid ].next ) \
1934 ? trie->trans[ base - ucharcount + charid ].next \
1935 : ( state==1 ? special : 0 ) \
1939 #define MADE_JUMP_TRIE 2
1940 #define MADE_EXACT_TRIE 4
1943 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1944 regnode *first, regnode *last, regnode *tail,
1945 U32 word_count, U32 flags, U32 depth)
1948 /* first pass, loop through and scan words */
1949 reg_trie_data *trie;
1950 HV *widecharmap = NULL;
1951 AV *revcharmap = newAV();
1957 regnode *jumper = NULL;
1958 regnode *nextbranch = NULL;
1959 regnode *convert = NULL;
1960 U32 *prev_states; /* temp array mapping each state to previous one */
1961 /* we just use folder as a flag in utf8 */
1962 const U8 * folder = NULL;
1965 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1966 AV *trie_words = NULL;
1967 /* along with revcharmap, this only used during construction but both are
1968 * useful during debugging so we store them in the struct when debugging.
1971 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1972 STRLEN trie_charcount=0;
1974 SV *re_trie_maxbuff;
1975 GET_RE_DEBUG_FLAGS_DECL;
1977 PERL_ARGS_ASSERT_MAKE_TRIE;
1979 PERL_UNUSED_ARG(depth);
1986 case EXACTFU: folder = PL_fold_latin1; break;
1987 case EXACTF: folder = PL_fold; break;
1988 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1991 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1993 trie->startstate = 1;
1994 trie->wordcount = word_count;
1995 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1996 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1998 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1999 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2000 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2003 trie_words = newAV();
2006 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2007 if (!SvIOK(re_trie_maxbuff)) {
2008 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2010 DEBUG_TRIE_COMPILE_r({
2011 PerlIO_printf( Perl_debug_log,
2012 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2013 (int)depth * 2 + 2, "",
2014 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2015 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2018 /* Find the node we are going to overwrite */
2019 if ( first == startbranch && OP( last ) != BRANCH ) {
2020 /* whole branch chain */
2023 /* branch sub-chain */
2024 convert = NEXTOPER( first );
2027 /* -- First loop and Setup --
2029 We first traverse the branches and scan each word to determine if it
2030 contains widechars, and how many unique chars there are, this is
2031 important as we have to build a table with at least as many columns as we
2034 We use an array of integers to represent the character codes 0..255
2035 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2036 the native representation of the character value as the key and IV's for
2039 *TODO* If we keep track of how many times each character is used we can
2040 remap the columns so that the table compression later on is more
2041 efficient in terms of memory by ensuring the most common value is in the
2042 middle and the least common are on the outside. IMO this would be better
2043 than a most to least common mapping as theres a decent chance the most
2044 common letter will share a node with the least common, meaning the node
2045 will not be compressible. With a middle is most common approach the worst
2046 case is when we have the least common nodes twice.
2050 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2051 regnode *noper = NEXTOPER( cur );
2052 const U8 *uc = (U8*)STRING( noper );
2053 const U8 *e = uc + STR_LEN( noper );
2055 U32 wordlen = 0; /* required init */
2056 STRLEN minbytes = 0;
2057 STRLEN maxbytes = 0;
2058 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2061 if (OP(noper) == NOTHING) {
2062 regnode *noper_next= regnext(noper);
2063 if (noper_next != tail && OP(noper_next) == flags) {
2065 uc= (U8*)STRING(noper);
2066 e= uc + STR_LEN(noper);
2067 trie->minlen= STR_LEN(noper);
2074 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2075 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2076 regardless of encoding */
2077 if (OP( noper ) == EXACTFU_SS) {
2078 /* false positives are ok, so just set this */
2079 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2082 for ( ; uc < e ; uc += len ) {
2083 TRIE_CHARCOUNT(trie)++;
2086 /* Acummulate to the current values, the range in the number of
2087 * bytes that this character could match. The max is presumed to
2088 * be the same as the folded input (which TRIE_READ_CHAR returns),
2089 * except that when this is not in UTF-8, it could be matched
2090 * against a string which is UTF-8, and the variant characters
2091 * could be 2 bytes instead of the 1 here. Likewise, for the
2092 * minimum number of bytes when not folded. When folding, the min
2093 * is assumed to be 1 byte could fold to match the single character
2094 * here, or in the case of a multi-char fold, 1 byte can fold to
2095 * the whole sequence. 'foldlen' is used to denote whether we are
2096 * in such a sequence, skipping the min setting if so. XXX TODO
2097 * Use the exact list of what folds to each character, from
2098 * PL_utf8_foldclosures */
2100 maxbytes += UTF8SKIP(uc);
2102 /* A non-UTF-8 string could be 1 byte to match our 2 */
2103 minbytes += (UTF8_IS_DOWNGRADEABLE_START(*uc))
2109 foldlen -= UTF8SKIP(uc);
2112 foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e);
2118 maxbytes += (UNI_IS_INVARIANT(*uc))
2129 foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e);
2136 U8 folded= folder[ (U8) uvc ];
2137 if ( !trie->charmap[ folded ] ) {
2138 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2139 TRIE_STORE_REVCHAR( folded );
2142 if ( !trie->charmap[ uvc ] ) {
2143 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2144 TRIE_STORE_REVCHAR( uvc );
2147 /* store the codepoint in the bitmap, and its folded
2149 TRIE_BITMAP_SET(trie, uvc);
2151 /* store the folded codepoint */
2152 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2155 /* store first byte of utf8 representation of
2156 variant codepoints */
2157 if (! UVCHR_IS_INVARIANT(uvc)) {
2158 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2161 set_bit = 0; /* We've done our bit :-) */
2166 widecharmap = newHV();
2168 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2171 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2173 if ( !SvTRUE( *svpp ) ) {
2174 sv_setiv( *svpp, ++trie->uniquecharcount );
2175 TRIE_STORE_REVCHAR(uvc);
2179 if( cur == first ) {
2180 trie->minlen = minbytes;
2181 trie->maxlen = maxbytes;
2182 } else if (minbytes < trie->minlen) {
2183 trie->minlen = minbytes;
2184 } else if (maxbytes > trie->maxlen) {
2185 trie->maxlen = maxbytes;
2187 } /* end first pass */
2188 DEBUG_TRIE_COMPILE_r(
2189 PerlIO_printf( Perl_debug_log,
2190 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2191 (int)depth * 2 + 2,"",
2192 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2193 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2194 (int)trie->minlen, (int)trie->maxlen )
2198 We now know what we are dealing with in terms of unique chars and
2199 string sizes so we can calculate how much memory a naive
2200 representation using a flat table will take. If it's over a reasonable
2201 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2202 conservative but potentially much slower representation using an array
2205 At the end we convert both representations into the same compressed
2206 form that will be used in regexec.c for matching with. The latter
2207 is a form that cannot be used to construct with but has memory
2208 properties similar to the list form and access properties similar
2209 to the table form making it both suitable for fast searches and
2210 small enough that its feasable to store for the duration of a program.
2212 See the comment in the code where the compressed table is produced
2213 inplace from the flat tabe representation for an explanation of how
2214 the compression works.
2219 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2222 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2223 > SvIV(re_trie_maxbuff) )
2226 Second Pass -- Array Of Lists Representation
2228 Each state will be represented by a list of charid:state records
2229 (reg_trie_trans_le) the first such element holds the CUR and LEN
2230 points of the allocated array. (See defines above).
2232 We build the initial structure using the lists, and then convert
2233 it into the compressed table form which allows faster lookups
2234 (but cant be modified once converted).
2237 STRLEN transcount = 1;
2239 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2240 "%*sCompiling trie using list compiler\n",
2241 (int)depth * 2 + 2, ""));
2243 trie->states = (reg_trie_state *)
2244 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2245 sizeof(reg_trie_state) );
2249 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2251 regnode *noper = NEXTOPER( cur );
2252 U8 *uc = (U8*)STRING( noper );
2253 const U8 *e = uc + STR_LEN( noper );
2254 U32 state = 1; /* required init */
2255 U16 charid = 0; /* sanity init */
2256 U32 wordlen = 0; /* required init */
2258 if (OP(noper) == NOTHING) {
2259 regnode *noper_next= regnext(noper);
2260 if (noper_next != tail && OP(noper_next) == flags) {
2262 uc= (U8*)STRING(noper);
2263 e= uc + STR_LEN(noper);
2267 if (OP(noper) != NOTHING) {
2268 for ( ; uc < e ; uc += len ) {
2273 charid = trie->charmap[ uvc ];
2275 SV** const svpp = hv_fetch( widecharmap,
2282 charid=(U16)SvIV( *svpp );
2285 /* charid is now 0 if we dont know the char read, or
2286 * nonzero if we do */
2293 if ( !trie->states[ state ].trans.list ) {
2294 TRIE_LIST_NEW( state );
2297 check <= TRIE_LIST_USED( state );
2300 if ( TRIE_LIST_ITEM( state, check ).forid
2303 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2308 newstate = next_alloc++;
2309 prev_states[newstate] = state;
2310 TRIE_LIST_PUSH( state, charid, newstate );
2315 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2319 TRIE_HANDLE_WORD(state);
2321 } /* end second pass */
2323 /* next alloc is the NEXT state to be allocated */
2324 trie->statecount = next_alloc;
2325 trie->states = (reg_trie_state *)
2326 PerlMemShared_realloc( trie->states,
2328 * sizeof(reg_trie_state) );
2330 /* and now dump it out before we compress it */
2331 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2332 revcharmap, next_alloc,
2336 trie->trans = (reg_trie_trans *)
2337 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2344 for( state=1 ; state < next_alloc ; state ++ ) {
2348 DEBUG_TRIE_COMPILE_MORE_r(
2349 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2353 if (trie->states[state].trans.list) {
2354 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2358 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2359 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2360 if ( forid < minid ) {
2362 } else if ( forid > maxid ) {
2366 if ( transcount < tp + maxid - minid + 1) {
2368 trie->trans = (reg_trie_trans *)
2369 PerlMemShared_realloc( trie->trans,
2371 * sizeof(reg_trie_trans) );
2372 Zero( trie->trans + (transcount / 2),
2376 base = trie->uniquecharcount + tp - minid;
2377 if ( maxid == minid ) {
2379 for ( ; zp < tp ; zp++ ) {
2380 if ( ! trie->trans[ zp ].next ) {
2381 base = trie->uniquecharcount + zp - minid;
2382 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2384 trie->trans[ zp ].check = state;
2390 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2392 trie->trans[ tp ].check = state;
2397 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2398 const U32 tid = base
2399 - trie->uniquecharcount
2400 + TRIE_LIST_ITEM( state, idx ).forid;
2401 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2403 trie->trans[ tid ].check = state;
2405 tp += ( maxid - minid + 1 );
2407 Safefree(trie->states[ state ].trans.list);
2410 DEBUG_TRIE_COMPILE_MORE_r(
2411 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2414 trie->states[ state ].trans.base=base;
2416 trie->lasttrans = tp + 1;
2420 Second Pass -- Flat Table Representation.
2422 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2423 each. We know that we will need Charcount+1 trans at most to store
2424 the data (one row per char at worst case) So we preallocate both
2425 structures assuming worst case.
2427 We then construct the trie using only the .next slots of the entry
2430 We use the .check field of the first entry of the node temporarily
2431 to make compression both faster and easier by keeping track of how
2432 many non zero fields are in the node.
2434 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2437 There are two terms at use here: state as a TRIE_NODEIDX() which is
2438 a number representing the first entry of the node, and state as a
2439 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2440 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2441 if there are 2 entrys per node. eg:
2449 The table is internally in the right hand, idx form. However as we
2450 also have to deal with the states array which is indexed by nodenum
2451 we have to use TRIE_NODENUM() to convert.
2454 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2455 "%*sCompiling trie using table compiler\n",
2456 (int)depth * 2 + 2, ""));
2458 trie->trans = (reg_trie_trans *)
2459 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2460 * trie->uniquecharcount + 1,
2461 sizeof(reg_trie_trans) );
2462 trie->states = (reg_trie_state *)
2463 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2464 sizeof(reg_trie_state) );
2465 next_alloc = trie->uniquecharcount + 1;
2468 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2470 regnode *noper = NEXTOPER( cur );
2471 const U8 *uc = (U8*)STRING( noper );
2472 const U8 *e = uc + STR_LEN( noper );
2474 U32 state = 1; /* required init */
2476 U16 charid = 0; /* sanity init */
2477 U32 accept_state = 0; /* sanity init */
2479 U32 wordlen = 0; /* required init */
2481 if (OP(noper) == NOTHING) {
2482 regnode *noper_next= regnext(noper);
2483 if (noper_next != tail && OP(noper_next) == flags) {
2485 uc= (U8*)STRING(noper);
2486 e= uc + STR_LEN(noper);
2490 if ( OP(noper) != NOTHING ) {
2491 for ( ; uc < e ; uc += len ) {
2496 charid = trie->charmap[ uvc ];
2498 SV* const * const svpp = hv_fetch( widecharmap,
2502 charid = svpp ? (U16)SvIV(*svpp) : 0;
2506 if ( !trie->trans[ state + charid ].next ) {
2507 trie->trans[ state + charid ].next = next_alloc;
2508 trie->trans[ state ].check++;
2509 prev_states[TRIE_NODENUM(next_alloc)]
2510 = TRIE_NODENUM(state);
2511 next_alloc += trie->uniquecharcount;
2513 state = trie->trans[ state + charid ].next;
2515 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2517 /* charid is now 0 if we dont know the char read, or
2518 * nonzero if we do */
2521 accept_state = TRIE_NODENUM( state );
2522 TRIE_HANDLE_WORD(accept_state);
2524 } /* end second pass */
2526 /* and now dump it out before we compress it */
2527 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2529 next_alloc, depth+1));
2533 * Inplace compress the table.*
2535 For sparse data sets the table constructed by the trie algorithm will
2536 be mostly 0/FAIL transitions or to put it another way mostly empty.
2537 (Note that leaf nodes will not contain any transitions.)
2539 This algorithm compresses the tables by eliminating most such
2540 transitions, at the cost of a modest bit of extra work during lookup:
2542 - Each states[] entry contains a .base field which indicates the
2543 index in the state[] array wheres its transition data is stored.
2545 - If .base is 0 there are no valid transitions from that node.
2547 - If .base is nonzero then charid is added to it to find an entry in
2550 -If trans[states[state].base+charid].check!=state then the
2551 transition is taken to be a 0/Fail transition. Thus if there are fail
2552 transitions at the front of the node then the .base offset will point
2553 somewhere inside the previous nodes data (or maybe even into a node
2554 even earlier), but the .check field determines if the transition is
2558 The following process inplace converts the table to the compressed
2559 table: We first do not compress the root node 1,and mark all its
2560 .check pointers as 1 and set its .base pointer as 1 as well. This
2561 allows us to do a DFA construction from the compressed table later,
2562 and ensures that any .base pointers we calculate later are greater
2565 - We set 'pos' to indicate the first entry of the second node.
2567 - We then iterate over the columns of the node, finding the first and
2568 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2569 and set the .check pointers accordingly, and advance pos
2570 appropriately and repreat for the next node. Note that when we copy
2571 the next pointers we have to convert them from the original
2572 NODEIDX form to NODENUM form as the former is not valid post
2575 - If a node has no transitions used we mark its base as 0 and do not
2576 advance the pos pointer.
2578 - If a node only has one transition we use a second pointer into the
2579 structure to fill in allocated fail transitions from other states.
2580 This pointer is independent of the main pointer and scans forward
2581 looking for null transitions that are allocated to a state. When it
2582 finds one it writes the single transition into the "hole". If the
2583 pointer doesnt find one the single transition is appended as normal.
2585 - Once compressed we can Renew/realloc the structures to release the
2588 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2589 specifically Fig 3.47 and the associated pseudocode.
2593 const U32 laststate = TRIE_NODENUM( next_alloc );
2596 trie->statecount = laststate;
2598 for ( state = 1 ; state < laststate ; state++ ) {
2600 const U32 stateidx = TRIE_NODEIDX( state );
2601 const U32 o_used = trie->trans[ stateidx ].check;
2602 U32 used = trie->trans[ stateidx ].check;
2603 trie->trans[ stateidx ].check = 0;
2606 used && charid < trie->uniquecharcount;
2609 if ( flag || trie->trans[ stateidx + charid ].next ) {
2610 if ( trie->trans[ stateidx + charid ].next ) {
2612 for ( ; zp < pos ; zp++ ) {
2613 if ( ! trie->trans[ zp ].next ) {
2617 trie->states[ state ].trans.base
2619 + trie->uniquecharcount
2621 trie->trans[ zp ].next
2622 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2624 trie->trans[ zp ].check = state;
2625 if ( ++zp > pos ) pos = zp;
2632 trie->states[ state ].trans.base
2633 = pos + trie->uniquecharcount - charid ;
2635 trie->trans[ pos ].next
2636 = SAFE_TRIE_NODENUM(
2637 trie->trans[ stateidx + charid ].next );
2638 trie->trans[ pos ].check = state;
2643 trie->lasttrans = pos + 1;
2644 trie->states = (reg_trie_state *)
2645 PerlMemShared_realloc( trie->states, laststate
2646 * sizeof(reg_trie_state) );
2647 DEBUG_TRIE_COMPILE_MORE_r(
2648 PerlIO_printf( Perl_debug_log,
2649 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2650 (int)depth * 2 + 2,"",
2651 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2655 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2658 } /* end table compress */
2660 DEBUG_TRIE_COMPILE_MORE_r(
2661 PerlIO_printf(Perl_debug_log,
2662 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2663 (int)depth * 2 + 2, "",
2664 (UV)trie->statecount,
2665 (UV)trie->lasttrans)
2667 /* resize the trans array to remove unused space */
2668 trie->trans = (reg_trie_trans *)
2669 PerlMemShared_realloc( trie->trans, trie->lasttrans
2670 * sizeof(reg_trie_trans) );
2672 { /* Modify the program and insert the new TRIE node */
2673 U8 nodetype =(U8)(flags & 0xFF);
2677 regnode *optimize = NULL;
2678 #ifdef RE_TRACK_PATTERN_OFFSETS
2681 U32 mjd_nodelen = 0;
2682 #endif /* RE_TRACK_PATTERN_OFFSETS */
2683 #endif /* DEBUGGING */
2685 This means we convert either the first branch or the first Exact,
2686 depending on whether the thing following (in 'last') is a branch
2687 or not and whther first is the startbranch (ie is it a sub part of
2688 the alternation or is it the whole thing.)
2689 Assuming its a sub part we convert the EXACT otherwise we convert
2690 the whole branch sequence, including the first.
2692 /* Find the node we are going to overwrite */
2693 if ( first != startbranch || OP( last ) == BRANCH ) {
2694 /* branch sub-chain */
2695 NEXT_OFF( first ) = (U16)(last - first);
2696 #ifdef RE_TRACK_PATTERN_OFFSETS
2698 mjd_offset= Node_Offset((convert));
2699 mjd_nodelen= Node_Length((convert));
2702 /* whole branch chain */
2704 #ifdef RE_TRACK_PATTERN_OFFSETS
2707 const regnode *nop = NEXTOPER( convert );
2708 mjd_offset= Node_Offset((nop));
2709 mjd_nodelen= Node_Length((nop));
2713 PerlIO_printf(Perl_debug_log,
2714 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2715 (int)depth * 2 + 2, "",
2716 (UV)mjd_offset, (UV)mjd_nodelen)
2719 /* But first we check to see if there is a common prefix we can
2720 split out as an EXACT and put in front of the TRIE node. */
2721 trie->startstate= 1;
2722 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2724 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2728 const U32 base = trie->states[ state ].trans.base;
2730 if ( trie->states[state].wordnum )
2733 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2734 if ( ( base + ofs >= trie->uniquecharcount ) &&
2735 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2736 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2738 if ( ++count > 1 ) {
2739 SV **tmp = av_fetch( revcharmap, ofs, 0);
2740 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2741 if ( state == 1 ) break;
2743 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2745 PerlIO_printf(Perl_debug_log,
2746 "%*sNew Start State=%"UVuf" Class: [",
2747 (int)depth * 2 + 2, "",
2750 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2751 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2753 TRIE_BITMAP_SET(trie,*ch);
2755 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2757 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2761 TRIE_BITMAP_SET(trie,*ch);
2763 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2764 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2770 SV **tmp = av_fetch( revcharmap, idx, 0);
2772 char *ch = SvPV( *tmp, len );
2774 SV *sv=sv_newmortal();
2775 PerlIO_printf( Perl_debug_log,
2776 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2777 (int)depth * 2 + 2, "",
2779 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2780 PL_colors[0], PL_colors[1],
2781 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2782 PERL_PV_ESCAPE_FIRSTCHAR
2787 OP( convert ) = nodetype;
2788 str=STRING(convert);
2791 STR_LEN(convert) += len;
2797 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2802 trie->prefixlen = (state-1);
2804 regnode *n = convert+NODE_SZ_STR(convert);
2805 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2806 trie->startstate = state;
2807 trie->minlen -= (state - 1);
2808 trie->maxlen -= (state - 1);
2810 /* At least the UNICOS C compiler choked on this
2811 * being argument to DEBUG_r(), so let's just have
2814 #ifdef PERL_EXT_RE_BUILD
2820 regnode *fix = convert;
2821 U32 word = trie->wordcount;
2823 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2824 while( ++fix < n ) {
2825 Set_Node_Offset_Length(fix, 0, 0);
2828 SV ** const tmp = av_fetch( trie_words, word, 0 );
2830 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2831 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2833 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2841 NEXT_OFF(convert) = (U16)(tail - convert);
2842 DEBUG_r(optimize= n);
2848 if ( trie->maxlen ) {
2849 NEXT_OFF( convert ) = (U16)(tail - convert);
2850 ARG_SET( convert, data_slot );
2851 /* Store the offset to the first unabsorbed branch in
2852 jump[0], which is otherwise unused by the jump logic.
2853 We use this when dumping a trie and during optimisation. */
2855 trie->jump[0] = (U16)(nextbranch - convert);
2857 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2858 * and there is a bitmap
2859 * and the first "jump target" node we found leaves enough room
2860 * then convert the TRIE node into a TRIEC node, with the bitmap
2861 * embedded inline in the opcode - this is hypothetically faster.
2863 if ( !trie->states[trie->startstate].wordnum
2865 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2867 OP( convert ) = TRIEC;
2868 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2869 PerlMemShared_free(trie->bitmap);
2872 OP( convert ) = TRIE;
2874 /* store the type in the flags */
2875 convert->flags = nodetype;
2879 + regarglen[ OP( convert ) ];
2881 /* XXX We really should free up the resource in trie now,
2882 as we won't use them - (which resources?) dmq */
2884 /* needed for dumping*/
2885 DEBUG_r(if (optimize) {
2886 regnode *opt = convert;
2888 while ( ++opt < optimize) {
2889 Set_Node_Offset_Length(opt,0,0);
2892 Try to clean up some of the debris left after the
2895 while( optimize < jumper ) {
2896 mjd_nodelen += Node_Length((optimize));
2897 OP( optimize ) = OPTIMIZED;
2898 Set_Node_Offset_Length(optimize,0,0);
2901 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2903 } /* end node insert */
2905 /* Finish populating the prev field of the wordinfo array. Walk back
2906 * from each accept state until we find another accept state, and if
2907 * so, point the first word's .prev field at the second word. If the
2908 * second already has a .prev field set, stop now. This will be the
2909 * case either if we've already processed that word's accept state,
2910 * or that state had multiple words, and the overspill words were
2911 * already linked up earlier.
2918 for (word=1; word <= trie->wordcount; word++) {
2920 if (trie->wordinfo[word].prev)
2922 state = trie->wordinfo[word].accept;
2924 state = prev_states[state];
2927 prev = trie->states[state].wordnum;
2931 trie->wordinfo[word].prev = prev;
2933 Safefree(prev_states);
2937 /* and now dump out the compressed format */
2938 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2940 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2942 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2943 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2945 SvREFCNT_dec_NN(revcharmap);
2949 : trie->startstate>1
2955 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2957 /* The Trie is constructed and compressed now so we can build a fail array if
2960 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2962 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2966 We find the fail state for each state in the trie, this state is the longest
2967 proper suffix of the current state's 'word' that is also a proper prefix of
2968 another word in our trie. State 1 represents the word '' and is thus the
2969 default fail state. This allows the DFA not to have to restart after its
2970 tried and failed a word at a given point, it simply continues as though it
2971 had been matching the other word in the first place.
2973 'abcdgu'=~/abcdefg|cdgu/
2974 When we get to 'd' we are still matching the first word, we would encounter
2975 'g' which would fail, which would bring us to the state representing 'd' in
2976 the second word where we would try 'g' and succeed, proceeding to match
2979 /* add a fail transition */
2980 const U32 trie_offset = ARG(source);
2981 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2983 const U32 ucharcount = trie->uniquecharcount;
2984 const U32 numstates = trie->statecount;
2985 const U32 ubound = trie->lasttrans + ucharcount;
2989 U32 base = trie->states[ 1 ].trans.base;
2992 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
2993 GET_RE_DEBUG_FLAGS_DECL;
2995 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2997 PERL_UNUSED_ARG(depth);
3001 ARG_SET( stclass, data_slot );
3002 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3003 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3004 aho->trie=trie_offset;
3005 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3006 Copy( trie->states, aho->states, numstates, reg_trie_state );
3007 Newxz( q, numstates, U32);
3008 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3011 /* initialize fail[0..1] to be 1 so that we always have
3012 a valid final fail state */
3013 fail[ 0 ] = fail[ 1 ] = 1;
3015 for ( charid = 0; charid < ucharcount ; charid++ ) {
3016 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3018 q[ q_write ] = newstate;
3019 /* set to point at the root */
3020 fail[ q[ q_write++ ] ]=1;
3023 while ( q_read < q_write) {
3024 const U32 cur = q[ q_read++ % numstates ];
3025 base = trie->states[ cur ].trans.base;
3027 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3028 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3030 U32 fail_state = cur;
3033 fail_state = fail[ fail_state ];
3034 fail_base = aho->states[ fail_state ].trans.base;
3035 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3037 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3038 fail[ ch_state ] = fail_state;
3039 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3041 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3043 q[ q_write++ % numstates] = ch_state;
3047 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3048 when we fail in state 1, this allows us to use the
3049 charclass scan to find a valid start char. This is based on the principle
3050 that theres a good chance the string being searched contains lots of stuff
3051 that cant be a start char.
3053 fail[ 0 ] = fail[ 1 ] = 0;
3054 DEBUG_TRIE_COMPILE_r({
3055 PerlIO_printf(Perl_debug_log,
3056 "%*sStclass Failtable (%"UVuf" states): 0",
3057 (int)(depth * 2), "", (UV)numstates
3059 for( q_read=1; q_read<numstates; q_read++ ) {
3060 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3062 PerlIO_printf(Perl_debug_log, "\n");
3065 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3069 #define DEBUG_PEEP(str,scan,depth) \
3070 DEBUG_OPTIMISE_r({if (scan){ \
3071 SV * const mysv=sv_newmortal(); \
3072 regnode *Next = regnext(scan); \
3073 regprop(RExC_rx, mysv, scan, NULL); \
3074 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3075 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3076 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3080 /* The below joins as many adjacent EXACTish nodes as possible into a single
3081 * one. The regop may be changed if the node(s) contain certain sequences that
3082 * require special handling. The joining is only done if:
3083 * 1) there is room in the current conglomerated node to entirely contain the
3085 * 2) they are the exact same node type
3087 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3088 * these get optimized out
3090 * If a node is to match under /i (folded), the number of characters it matches
3091 * can be different than its character length if it contains a multi-character
3092 * fold. *min_subtract is set to the total delta number of characters of the
3095 * And *unfolded_multi_char is set to indicate whether or not the node contains
3096 * an unfolded multi-char fold. This happens when whether the fold is valid or
3097 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3098 * SMALL LETTER SHARP S, as only if the target string being matched against
3099 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3100 * folding rules depend on the locale in force at runtime. (Multi-char folds
3101 * whose components are all above the Latin1 range are not run-time locale
3102 * dependent, and have already been folded by the time this function is
3105 * This is as good a place as any to discuss the design of handling these
3106 * multi-character fold sequences. It's been wrong in Perl for a very long
3107 * time. There are three code points in Unicode whose multi-character folds
3108 * were long ago discovered to mess things up. The previous designs for
3109 * dealing with these involved assigning a special node for them. This
3110 * approach doesn't always work, as evidenced by this example:
3111 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3112 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3113 * would match just the \xDF, it won't be able to handle the case where a
3114 * successful match would have to cross the node's boundary. The new approach
3115 * that hopefully generally solves the problem generates an EXACTFU_SS node
3116 * that is "sss" in this case.
3118 * It turns out that there are problems with all multi-character folds, and not
3119 * just these three. Now the code is general, for all such cases. The
3120 * approach taken is:
3121 * 1) This routine examines each EXACTFish node that could contain multi-
3122 * character folded sequences. Since a single character can fold into
3123 * such a sequence, the minimum match length for this node is less than
3124 * the number of characters in the node. This routine returns in
3125 * *min_subtract how many characters to subtract from the the actual
3126 * length of the string to get a real minimum match length; it is 0 if
3127 * there are no multi-char foldeds. This delta is used by the caller to
3128 * adjust the min length of the match, and the delta between min and max,
3129 * so that the optimizer doesn't reject these possibilities based on size
3131 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3132 * is used for an EXACTFU node that contains at least one "ss" sequence in
3133 * it. For non-UTF-8 patterns and strings, this is the only case where
3134 * there is a possible fold length change. That means that a regular
3135 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3136 * with length changes, and so can be processed faster. regexec.c takes
3137 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3138 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3139 * known until runtime). This saves effort in regex matching. However,
3140 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3141 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3142 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3143 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3144 * possibilities for the non-UTF8 patterns are quite simple, except for
3145 * the sharp s. All the ones that don't involve a UTF-8 target string are
3146 * members of a fold-pair, and arrays are set up for all of them so that
3147 * the other member of the pair can be found quickly. Code elsewhere in
3148 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3149 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3150 * described in the next item.
3151 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3152 * validity of the fold won't be known until runtime, and so must remain
3153 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3154 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3155 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3156 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3157 * The reason this is a problem is that the optimizer part of regexec.c
3158 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3159 * that a character in the pattern corresponds to at most a single
3160 * character in the target string. (And I do mean character, and not byte
3161 * here, unlike other parts of the documentation that have never been
3162 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3163 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3164 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3165 * nodes, violate the assumption, and they are the only instances where it
3166 * is violated. I'm reluctant to try to change the assumption, as the
3167 * code involved is impenetrable to me (khw), so instead the code here
3168 * punts. This routine examines EXACTFL nodes, and (when the pattern
3169 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3170 * boolean indicating whether or not the node contains such a fold. When
3171 * it is true, the caller sets a flag that later causes the optimizer in
3172 * this file to not set values for the floating and fixed string lengths,
3173 * and thus avoids the optimizer code in regexec.c that makes the invalid
3174 * assumption. Thus, there is no optimization based on string lengths for
3175 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3176 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3177 * assumption is wrong only in these cases is that all other non-UTF-8
3178 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3179 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3180 * EXACTF nodes because we don't know at compile time if it actually
3181 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3182 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3183 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3184 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3185 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3186 * string would require the pattern to be forced into UTF-8, the overhead
3187 * of which we want to avoid. Similarly the unfolded multi-char folds in
3188 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3191 * Similarly, the code that generates tries doesn't currently handle
3192 * not-already-folded multi-char folds, and it looks like a pain to change
3193 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3194 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3195 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3196 * using /iaa matching will be doing so almost entirely with ASCII
3197 * strings, so this should rarely be encountered in practice */
3199 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3200 if (PL_regkind[OP(scan)] == EXACT) \
3201 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3204 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3205 UV *min_subtract, bool *unfolded_multi_char,
3206 U32 flags,regnode *val, U32 depth)
3208 /* Merge several consecutive EXACTish nodes into one. */
3209 regnode *n = regnext(scan);
3211 regnode *next = scan + NODE_SZ_STR(scan);
3215 regnode *stop = scan;
3216 GET_RE_DEBUG_FLAGS_DECL;
3218 PERL_UNUSED_ARG(depth);
3221 PERL_ARGS_ASSERT_JOIN_EXACT;
3222 #ifndef EXPERIMENTAL_INPLACESCAN
3223 PERL_UNUSED_ARG(flags);
3224 PERL_UNUSED_ARG(val);
3226 DEBUG_PEEP("join",scan,depth);
3228 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3229 * EXACT ones that are mergeable to the current one. */
3231 && (PL_regkind[OP(n)] == NOTHING
3232 || (stringok && OP(n) == OP(scan)))
3234 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3237 if (OP(n) == TAIL || n > next)
3239 if (PL_regkind[OP(n)] == NOTHING) {
3240 DEBUG_PEEP("skip:",n,depth);
3241 NEXT_OFF(scan) += NEXT_OFF(n);
3242 next = n + NODE_STEP_REGNODE;
3249 else if (stringok) {
3250 const unsigned int oldl = STR_LEN(scan);
3251 regnode * const nnext = regnext(n);
3253 /* XXX I (khw) kind of doubt that this works on platforms (should
3254 * Perl ever run on one) where U8_MAX is above 255 because of lots
3255 * of other assumptions */
3256 /* Don't join if the sum can't fit into a single node */
3257 if (oldl + STR_LEN(n) > U8_MAX)
3260 DEBUG_PEEP("merg",n,depth);
3263 NEXT_OFF(scan) += NEXT_OFF(n);
3264 STR_LEN(scan) += STR_LEN(n);
3265 next = n + NODE_SZ_STR(n);
3266 /* Now we can overwrite *n : */
3267 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3275 #ifdef EXPERIMENTAL_INPLACESCAN
3276 if (flags && !NEXT_OFF(n)) {
3277 DEBUG_PEEP("atch", val, depth);
3278 if (reg_off_by_arg[OP(n)]) {
3279 ARG_SET(n, val - n);
3282 NEXT_OFF(n) = val - n;
3290 *unfolded_multi_char = FALSE;
3292 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3293 * can now analyze for sequences of problematic code points. (Prior to
3294 * this final joining, sequences could have been split over boundaries, and
3295 * hence missed). The sequences only happen in folding, hence for any
3296 * non-EXACT EXACTish node */
3297 if (OP(scan) != EXACT) {
3298 U8* s0 = (U8*) STRING(scan);
3300 U8* s_end = s0 + STR_LEN(scan);
3302 int total_count_delta = 0; /* Total delta number of characters that
3303 multi-char folds expand to */
3305 /* One pass is made over the node's string looking for all the
3306 * possibilities. To avoid some tests in the loop, there are two main
3307 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3312 if (OP(scan) == EXACTFL) {
3315 /* An EXACTFL node would already have been changed to another
3316 * node type unless there is at least one character in it that
3317 * is problematic; likely a character whose fold definition
3318 * won't be known until runtime, and so has yet to be folded.
3319 * For all but the UTF-8 locale, folds are 1-1 in length, but
3320 * to handle the UTF-8 case, we need to create a temporary
3321 * folded copy using UTF-8 locale rules in order to analyze it.
3322 * This is because our macros that look to see if a sequence is
3323 * a multi-char fold assume everything is folded (otherwise the
3324 * tests in those macros would be too complicated and slow).
3325 * Note that here, the non-problematic folds will have already
3326 * been done, so we can just copy such characters. We actually
3327 * don't completely fold the EXACTFL string. We skip the
3328 * unfolded multi-char folds, as that would just create work
3329 * below to figure out the size they already are */
3331 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3334 STRLEN s_len = UTF8SKIP(s);
3335 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3336 Copy(s, d, s_len, U8);
3339 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3340 *unfolded_multi_char = TRUE;
3341 Copy(s, d, s_len, U8);
3344 else if (isASCII(*s)) {
3345 *(d++) = toFOLD(*s);
3349 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3355 /* Point the remainder of the routine to look at our temporary
3359 } /* End of creating folded copy of EXACTFL string */
3361 /* Examine the string for a multi-character fold sequence. UTF-8
3362 * patterns have all characters pre-folded by the time this code is
3364 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3365 length sequence we are looking for is 2 */
3367 int count = 0; /* How many characters in a multi-char fold */
3368 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3369 if (! len) { /* Not a multi-char fold: get next char */
3374 /* Nodes with 'ss' require special handling, except for
3375 * EXACTFA-ish for which there is no multi-char fold to this */
3376 if (len == 2 && *s == 's' && *(s+1) == 's'
3377 && OP(scan) != EXACTFA
3378 && OP(scan) != EXACTFA_NO_TRIE)
3381 if (OP(scan) != EXACTFL) {
3382 OP(scan) = EXACTFU_SS;
3386 else { /* Here is a generic multi-char fold. */
3387 U8* multi_end = s + len;
3389 /* Count how many characters in it. In the case of /aa, no
3390 * folds which contain ASCII code points are allowed, so
3391 * check for those, and skip if found. */
3392 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3393 count = utf8_length(s, multi_end);
3397 while (s < multi_end) {
3400 goto next_iteration;
3410 /* The delta is how long the sequence is minus 1 (1 is how long
3411 * the character that folds to the sequence is) */
3412 total_count_delta += count - 1;
3416 /* We created a temporary folded copy of the string in EXACTFL
3417 * nodes. Therefore we need to be sure it doesn't go below zero,
3418 * as the real string could be shorter */
3419 if (OP(scan) == EXACTFL) {
3420 int total_chars = utf8_length((U8*) STRING(scan),
3421 (U8*) STRING(scan) + STR_LEN(scan));
3422 if (total_count_delta > total_chars) {
3423 total_count_delta = total_chars;
3427 *min_subtract += total_count_delta;
3430 else if (OP(scan) == EXACTFA) {
3432 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3433 * fold to the ASCII range (and there are no existing ones in the
3434 * upper latin1 range). But, as outlined in the comments preceding
3435 * this function, we need to flag any occurrences of the sharp s.
3436 * This character forbids trie formation (because of added
3439 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3440 OP(scan) = EXACTFA_NO_TRIE;
3441 *unfolded_multi_char = TRUE;
3450 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3451 * folds that are all Latin1. As explained in the comments
3452 * preceding this function, we look also for the sharp s in EXACTF
3453 * and EXACTFL nodes; it can be in the final position. Otherwise
3454 * we can stop looking 1 byte earlier because have to find at least
3455 * two characters for a multi-fold */
3456 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3461 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3462 if (! len) { /* Not a multi-char fold. */
3463 if (*s == LATIN_SMALL_LETTER_SHARP_S
3464 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3466 *unfolded_multi_char = TRUE;
3473 && isARG2_lower_or_UPPER_ARG1('s', *s)
3474 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3477 /* EXACTF nodes need to know that the minimum length
3478 * changed so that a sharp s in the string can match this
3479 * ss in the pattern, but they remain EXACTF nodes, as they
3480 * won't match this unless the target string is is UTF-8,
3481 * which we don't know until runtime. EXACTFL nodes can't
3482 * transform into EXACTFU nodes */
3483 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3484 OP(scan) = EXACTFU_SS;
3488 *min_subtract += len - 1;
3495 /* Allow dumping but overwriting the collection of skipped
3496 * ops and/or strings with fake optimized ops */
3497 n = scan + NODE_SZ_STR(scan);
3505 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3509 /* REx optimizer. Converts nodes into quicker variants "in place".
3510 Finds fixed substrings. */
3512 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3513 to the position after last scanned or to NULL. */
3515 #define INIT_AND_WITHP \
3516 assert(!and_withp); \
3517 Newx(and_withp,1, regnode_ssc); \
3518 SAVEFREEPV(and_withp)
3520 /* this is a chain of data about sub patterns we are processing that
3521 need to be handled separately/specially in study_chunk. Its so
3522 we can simulate recursion without losing state. */
3524 typedef struct scan_frame {
3525 regnode *last; /* last node to process in this frame */
3526 regnode *next; /* next node to process when last is reached */
3527 struct scan_frame *prev; /*previous frame*/
3528 U32 prev_recursed_depth;
3529 I32 stop; /* what stopparen do we use */
3534 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3535 SSize_t *minlenp, SSize_t *deltap,
3540 regnode_ssc *and_withp,
3541 U32 flags, U32 depth)
3542 /* scanp: Start here (read-write). */
3543 /* deltap: Write maxlen-minlen here. */
3544 /* last: Stop before this one. */
3545 /* data: string data about the pattern */
3546 /* stopparen: treat close N as END */
3547 /* recursed: which subroutines have we recursed into */
3548 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3551 /* There must be at least this number of characters to match */
3554 regnode *scan = *scanp, *next;
3556 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3557 int is_inf_internal = 0; /* The studied chunk is infinite */
3558 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3559 scan_data_t data_fake;
3560 SV *re_trie_maxbuff = NULL;
3561 regnode *first_non_open = scan;
3562 SSize_t stopmin = SSize_t_MAX;
3563 scan_frame *frame = NULL;
3564 GET_RE_DEBUG_FLAGS_DECL;
3566 PERL_ARGS_ASSERT_STUDY_CHUNK;
3569 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3572 while (first_non_open && OP(first_non_open) == OPEN)
3573 first_non_open=regnext(first_non_open);
3578 while ( scan && OP(scan) != END && scan < last ){
3579 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3580 node length to get a real minimum (because
3581 the folded version may be shorter) */
3582 bool unfolded_multi_char = FALSE;
3583 /* Peephole optimizer: */
3584 DEBUG_OPTIMISE_MORE_r(
3586 PerlIO_printf(Perl_debug_log,
3587 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3588 ((int) depth*2), "", (long)stopparen,
3589 (unsigned long)depth, (unsigned long)recursed_depth);
3590 if (recursed_depth) {
3593 for ( j = 0 ; j < recursed_depth ; j++ ) {
3594 PerlIO_printf(Perl_debug_log,"[");
3595 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3596 PerlIO_printf(Perl_debug_log,"%d",
3597 PAREN_TEST(RExC_study_chunk_recursed +
3598 (j * RExC_study_chunk_recursed_bytes), i)
3601 PerlIO_printf(Perl_debug_log,"]");
3604 PerlIO_printf(Perl_debug_log,"\n");
3607 DEBUG_STUDYDATA("Peep:", data, depth);
3608 DEBUG_PEEP("Peep", scan, depth);
3611 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3612 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3613 * by a different invocation of reg() -- Yves
3615 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3617 /* Follow the next-chain of the current node and optimize
3618 away all the NOTHINGs from it. */
3619 if (OP(scan) != CURLYX) {
3620 const int max = (reg_off_by_arg[OP(scan)]
3622 /* I32 may be smaller than U16 on CRAYs! */
3623 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3624 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3628 /* Skip NOTHING and LONGJMP. */
3629 while ((n = regnext(n))
3630 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3631 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3632 && off + noff < max)
3634 if (reg_off_by_arg[OP(scan)])
3637 NEXT_OFF(scan) = off;
3642 /* The principal pseudo-switch. Cannot be a switch, since we
3643 look into several different things. */
3644 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3645 || OP(scan) == IFTHEN) {
3646 next = regnext(scan);
3648 /* demq: the op(next)==code check is to see if we have
3649 * "branch-branch" AFAICT */
3651 if (OP(next) == code || code == IFTHEN) {
3652 /* NOTE - There is similar code to this block below for
3653 * handling TRIE nodes on a re-study. If you change stuff here
3654 * check there too. */
3655 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3657 regnode * const startbranch=scan;
3659 if (flags & SCF_DO_SUBSTR) {
3660 /* Cannot merge strings after this. */
3661 scan_commit(pRExC_state, data, minlenp, is_inf);
3664 if (flags & SCF_DO_STCLASS)
3665 ssc_init_zero(pRExC_state, &accum);
3667 while (OP(scan) == code) {
3668 SSize_t deltanext, minnext, fake;
3670 regnode_ssc this_class;
3673 data_fake.flags = 0;
3675 data_fake.whilem_c = data->whilem_c;
3676 data_fake.last_closep = data->last_closep;
3679 data_fake.last_closep = &fake;
3681 data_fake.pos_delta = delta;
3682 next = regnext(scan);
3683 scan = NEXTOPER(scan);
3685 scan = NEXTOPER(scan);
3686 if (flags & SCF_DO_STCLASS) {
3687 ssc_init(pRExC_state, &this_class);
3688 data_fake.start_class = &this_class;
3689 f = SCF_DO_STCLASS_AND;
3691 if (flags & SCF_WHILEM_VISITED_POS)
3692 f |= SCF_WHILEM_VISITED_POS;
3694 /* we suppose the run is continuous, last=next...*/
3695 minnext = study_chunk(pRExC_state, &scan, minlenp,
3696 &deltanext, next, &data_fake, stopparen,
3697 recursed_depth, NULL, f,depth+1);
3700 if (deltanext == SSize_t_MAX) {
3701 is_inf = is_inf_internal = 1;
3703 } else if (max1 < minnext + deltanext)
3704 max1 = minnext + deltanext;
3706 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3708 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3709 if ( stopmin > minnext)
3710 stopmin = min + min1;
3711 flags &= ~SCF_DO_SUBSTR;
3713 data->flags |= SCF_SEEN_ACCEPT;
3716 if (data_fake.flags & SF_HAS_EVAL)
3717 data->flags |= SF_HAS_EVAL;
3718 data->whilem_c = data_fake.whilem_c;
3720 if (flags & SCF_DO_STCLASS)
3721 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3723 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3725 if (flags & SCF_DO_SUBSTR) {
3726 data->pos_min += min1;
3727 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3728 data->pos_delta = SSize_t_MAX;
3730 data->pos_delta += max1 - min1;
3731 if (max1 != min1 || is_inf)
3732 data->longest = &(data->longest_float);
3735 if (delta == SSize_t_MAX
3736 || SSize_t_MAX - delta - (max1 - min1) < 0)
3737 delta = SSize_t_MAX;
3739 delta += max1 - min1;
3740 if (flags & SCF_DO_STCLASS_OR) {
3741 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3743 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3744 flags &= ~SCF_DO_STCLASS;
3747 else if (flags & SCF_DO_STCLASS_AND) {
3749 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3750 flags &= ~SCF_DO_STCLASS;
3753 /* Switch to OR mode: cache the old value of
3754 * data->start_class */
3756 StructCopy(data->start_class, and_withp, regnode_ssc);
3757 flags &= ~SCF_DO_STCLASS_AND;
3758 StructCopy(&accum, data->start_class, regnode_ssc);
3759 flags |= SCF_DO_STCLASS_OR;
3763 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3764 OP( startbranch ) == BRANCH )
3768 Assuming this was/is a branch we are dealing with: 'scan'
3769 now points at the item that follows the branch sequence,
3770 whatever it is. We now start at the beginning of the
3771 sequence and look for subsequences of
3777 which would be constructed from a pattern like
3780 If we can find such a subsequence we need to turn the first
3781 element into a trie and then add the subsequent branch exact
3782 strings to the trie.
3786 1. patterns where the whole set of branches can be
3789 2. patterns where only a subset can be converted.
3791 In case 1 we can replace the whole set with a single regop
3792 for the trie. In case 2 we need to keep the start and end
3795 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3796 becomes BRANCH TRIE; BRANCH X;
3798 There is an additional case, that being where there is a
3799 common prefix, which gets split out into an EXACT like node
3800 preceding the TRIE node.
3802 If x(1..n)==tail then we can do a simple trie, if not we make
3803 a "jump" trie, such that when we match the appropriate word
3804 we "jump" to the appropriate tail node. Essentially we turn
3805 a nested if into a case structure of sorts.
3810 if (!re_trie_maxbuff) {
3811 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3812 if (!SvIOK(re_trie_maxbuff))
3813 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3815 if ( SvIV(re_trie_maxbuff)>=0 ) {
3817 regnode *first = (regnode *)NULL;
3818 regnode *last = (regnode *)NULL;
3819 regnode *tail = scan;
3824 SV * const mysv = sv_newmortal(); /* for dumping */
3826 /* var tail is used because there may be a TAIL
3827 regop in the way. Ie, the exacts will point to the
3828 thing following the TAIL, but the last branch will
3829 point at the TAIL. So we advance tail. If we
3830 have nested (?:) we may have to move through several
3834 while ( OP( tail ) == TAIL ) {
3835 /* this is the TAIL generated by (?:) */
3836 tail = regnext( tail );
3840 DEBUG_TRIE_COMPILE_r({
3841 regprop(RExC_rx, mysv, tail, NULL);
3842 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3843 (int)depth * 2 + 2, "",
3844 "Looking for TRIE'able sequences. Tail node is: ",
3845 SvPV_nolen_const( mysv )
3851 Step through the branches
3852 cur represents each branch,
3853 noper is the first thing to be matched as part
3855 noper_next is the regnext() of that node.
3857 We normally handle a case like this
3858 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3859 support building with NOJUMPTRIE, which restricts
3860 the trie logic to structures like /FOO|BAR/.
3862 If noper is a trieable nodetype then the branch is
3863 a possible optimization target. If we are building
3864 under NOJUMPTRIE then we require that noper_next is
3865 the same as scan (our current position in the regex
3868 Once we have two or more consecutive such branches
3869 we can create a trie of the EXACT's contents and
3870 stitch it in place into the program.
3872 If the sequence represents all of the branches in
3873 the alternation we replace the entire thing with a
3876 Otherwise when it is a subsequence we need to
3877 stitch it in place and replace only the relevant
3878 branches. This means the first branch has to remain
3879 as it is used by the alternation logic, and its
3880 next pointer, and needs to be repointed at the item
3881 on the branch chain following the last branch we
3882 have optimized away.
3884 This could be either a BRANCH, in which case the
3885 subsequence is internal, or it could be the item
3886 following the branch sequence in which case the
3887 subsequence is at the end (which does not
3888 necessarily mean the first node is the start of the
3891 TRIE_TYPE(X) is a define which maps the optype to a
3895 ----------------+-----------
3899 EXACTFU_SS | EXACTFU
3904 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3905 ( EXACT == (X) ) ? EXACT : \
3906 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3907 ( EXACTFA == (X) ) ? EXACTFA : \
3910 /* dont use tail as the end marker for this traverse */
3911 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3912 regnode * const noper = NEXTOPER( cur );
3913 U8 noper_type = OP( noper );
3914 U8 noper_trietype = TRIE_TYPE( noper_type );
3915 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3916 regnode * const noper_next = regnext( noper );
3917 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3918 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3921 DEBUG_TRIE_COMPILE_r({
3922 regprop(RExC_rx, mysv, cur, NULL);
3923 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3924 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3926 regprop(RExC_rx, mysv, noper, NULL);
3927 PerlIO_printf( Perl_debug_log, " -> %s",
3928 SvPV_nolen_const(mysv));
3931 regprop(RExC_rx, mysv, noper_next, NULL);
3932 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3933 SvPV_nolen_const(mysv));
3935 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3936 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3937 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3941 /* Is noper a trieable nodetype that can be merged
3942 * with the current trie (if there is one)? */
3946 ( noper_trietype == NOTHING)
3947 || ( trietype == NOTHING )
3948 || ( trietype == noper_trietype )
3951 && noper_next == tail
3955 /* Handle mergable triable node Either we are
3956 * the first node in a new trieable sequence,
3957 * in which case we do some bookkeeping,
3958 * otherwise we update the end pointer. */
3961 if ( noper_trietype == NOTHING ) {
3962 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3963 regnode * const noper_next = regnext( noper );
3964 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3965 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3968 if ( noper_next_trietype ) {
3969 trietype = noper_next_trietype;
3970 } else if (noper_next_type) {
3971 /* a NOTHING regop is 1 regop wide.
3972 * We need at least two for a trie
3973 * so we can't merge this in */
3977 trietype = noper_trietype;
3980 if ( trietype == NOTHING )
3981 trietype = noper_trietype;
3986 } /* end handle mergable triable node */
3988 /* handle unmergable node -
3989 * noper may either be a triable node which can
3990 * not be tried together with the current trie,
3991 * or a non triable node */
3993 /* If last is set and trietype is not
3994 * NOTHING then we have found at least two
3995 * triable branch sequences in a row of a
3996 * similar trietype so we can turn them
3997 * into a trie. If/when we allow NOTHING to
3998 * start a trie sequence this condition
3999 * will be required, and it isn't expensive
4000 * so we leave it in for now. */
4001 if ( trietype && trietype != NOTHING )
4002 make_trie( pRExC_state,
4003 startbranch, first, cur, tail,
4004 count, trietype, depth+1 );
4005 last = NULL; /* note: we clear/update
4006 first, trietype etc below,
4007 so we dont do it here */
4011 && noper_next == tail
4014 /* noper is triable, so we can start a new
4018 trietype = noper_trietype;
4020 /* if we already saw a first but the
4021 * current node is not triable then we have
4022 * to reset the first information. */
4027 } /* end handle unmergable node */
4028 } /* loop over branches */
4029 DEBUG_TRIE_COMPILE_r({
4030 regprop(RExC_rx, mysv, cur, NULL);
4031 PerlIO_printf( Perl_debug_log,
4032 "%*s- %s (%d) <SCAN FINISHED>\n",
4034 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4037 if ( last && trietype ) {
4038 if ( trietype != NOTHING ) {
4039 /* the last branch of the sequence was part of
4040 * a trie, so we have to construct it here
4041 * outside of the loop */
4042 made= make_trie( pRExC_state, startbranch,
4043 first, scan, tail, count,
4044 trietype, depth+1 );
4045 #ifdef TRIE_STUDY_OPT
4046 if ( ((made == MADE_EXACT_TRIE &&
4047 startbranch == first)
4048 || ( first_non_open == first )) &&
4050 flags |= SCF_TRIE_RESTUDY;
4051 if ( startbranch == first
4054 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4059 /* at this point we know whatever we have is a
4060 * NOTHING sequence/branch AND if 'startbranch'
4061 * is 'first' then we can turn the whole thing
4064 if ( startbranch == first ) {
4066 /* the entire thing is a NOTHING sequence,
4067 * something like this: (?:|) So we can
4068 * turn it into a plain NOTHING op. */
4069 DEBUG_TRIE_COMPILE_r({
4070 regprop(RExC_rx, mysv, cur, NULL);
4071 PerlIO_printf( Perl_debug_log,
4072 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4073 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4076 OP(startbranch)= NOTHING;
4077 NEXT_OFF(startbranch)= tail - startbranch;
4078 for ( opt= startbranch + 1; opt < tail ; opt++ )
4082 } /* end if ( last) */
4083 } /* TRIE_MAXBUF is non zero */
4088 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4089 scan = NEXTOPER(NEXTOPER(scan));
4090 } else /* single branch is optimized. */
4091 scan = NEXTOPER(scan);
4093 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4094 scan_frame *newframe = NULL;
4098 U32 my_recursed_depth= recursed_depth;
4100 if (OP(scan) != SUSPEND) {
4101 /* set the pointer */
4102 if (OP(scan) == GOSUB) {
4104 RExC_recurse[ARG2L(scan)] = scan;
4105 start = RExC_open_parens[paren-1];
4106 end = RExC_close_parens[paren-1];
4109 start = RExC_rxi->program + 1;
4114 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4116 if (!recursed_depth) {
4117 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4119 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4120 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4121 RExC_study_chunk_recursed_bytes, U8);
4123 /* we havent recursed into this paren yet, so recurse into it */
4124 DEBUG_STUDYDATA("set:", data,depth);
4125 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4126 my_recursed_depth= recursed_depth + 1;
4127 Newx(newframe,1,scan_frame);
4129 DEBUG_STUDYDATA("inf:", data,depth);
4130 /* some form of infinite recursion, assume infinite length
4132 if (flags & SCF_DO_SUBSTR) {
4133 scan_commit(pRExC_state, data, minlenp, is_inf);
4134 data->longest = &(data->longest_float);
4136 is_inf = is_inf_internal = 1;
4137 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4138 ssc_anything(data->start_class);
4139 flags &= ~SCF_DO_STCLASS;
4142 Newx(newframe,1,scan_frame);
4145 end = regnext(scan);
4150 SAVEFREEPV(newframe);
4151 newframe->next = regnext(scan);
4152 newframe->last = last;
4153 newframe->stop = stopparen;
4154 newframe->prev = frame;
4155 newframe->prev_recursed_depth = recursed_depth;
4157 DEBUG_STUDYDATA("frame-new:",data,depth);
4158 DEBUG_PEEP("fnew", scan, depth);
4165 recursed_depth= my_recursed_depth;
4170 else if (OP(scan) == EXACT) {
4171 SSize_t l = STR_LEN(scan);
4174 const U8 * const s = (U8*)STRING(scan);
4175 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4176 l = utf8_length(s, s + l);
4178 uc = *((U8*)STRING(scan));
4181 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4182 /* The code below prefers earlier match for fixed
4183 offset, later match for variable offset. */
4184 if (data->last_end == -1) { /* Update the start info. */
4185 data->last_start_min = data->pos_min;
4186 data->last_start_max = is_inf
4187 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4189 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4191 SvUTF8_on(data->last_found);
4193 SV * const sv = data->last_found;
4194 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4195 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4196 if (mg && mg->mg_len >= 0)
4197 mg->mg_len += utf8_length((U8*)STRING(scan),
4198 (U8*)STRING(scan)+STR_LEN(scan));
4200 data->last_end = data->pos_min + l;
4201 data->pos_min += l; /* As in the first entry. */
4202 data->flags &= ~SF_BEFORE_EOL;
4205 /* ANDing the code point leaves at most it, and not in locale, and
4206 * can't match null string */
4207 if (flags & SCF_DO_STCLASS_AND) {
4208 ssc_cp_and(data->start_class, uc);
4209 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4210 ssc_clear_locale(data->start_class);
4212 else if (flags & SCF_DO_STCLASS_OR) {
4213 ssc_add_cp(data->start_class, uc);
4214 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4216 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4217 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4219 flags &= ~SCF_DO_STCLASS;
4221 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
4222 SSize_t l = STR_LEN(scan);
4223 UV uc = *((U8*)STRING(scan));
4224 SV* EXACTF_invlist = _new_invlist(4); /* Start out big enough for 2
4225 separate code points */
4227 /* Search for fixed substrings supports EXACT only. */
4228 if (flags & SCF_DO_SUBSTR) {
4230 scan_commit(pRExC_state, data, minlenp, is_inf);
4233 const U8 * const s = (U8 *)STRING(scan);
4234 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4235 l = utf8_length(s, s + l);
4237 if (unfolded_multi_char) {
4238 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4240 min += l - min_subtract;
4242 delta += min_subtract;
4243 if (flags & SCF_DO_SUBSTR) {
4244 data->pos_min += l - min_subtract;
4245 if (data->pos_min < 0) {
4248 data->pos_delta += min_subtract;
4250 data->longest = &(data->longest_float);
4253 if (OP(scan) == EXACTFL) {
4255 /* We don't know what the folds are; it could be anything. XXX
4256 * Actually, we only support UTF-8 encoding for code points
4257 * above Latin1, so we could know what those folds are. */
4258 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4262 else { /* Non-locale EXACTFish */
4263 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4264 if (flags & SCF_DO_STCLASS_AND) {
4265 ssc_clear_locale(data->start_class);
4267 if (uc < 256) { /* We know what the Latin1 folds are ... */
4268 if (IS_IN_SOME_FOLD_L1(uc)) { /* For instance, we
4269 know if anything folds
4271 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist,
4272 PL_fold_latin1[uc]);
4273 if (OP(scan) != EXACTFA) { /* The folds below aren't
4275 if (isARG2_lower_or_UPPER_ARG1('s', uc)) {
4277 = add_cp_to_invlist(EXACTF_invlist,
4278 LATIN_SMALL_LETTER_SHARP_S);
4280 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
4282 = add_cp_to_invlist(EXACTF_invlist, 's');
4284 = add_cp_to_invlist(EXACTF_invlist, 'S');
4288 /* We also know if there are above-Latin1 code points
4289 * that fold to this (none legal for ASCII and /iaa) */
4290 if ((! isASCII(uc) || OP(scan) != EXACTFA)
4291 && HAS_NONLATIN1_FOLD_CLOSURE(uc))
4293 /* XXX We could know exactly what does fold to this
4294 * if the reverse folds are loaded, as currently in
4296 _invlist_union(EXACTF_invlist,
4302 else { /* Non-locale, above Latin1. XXX We don't currently
4303 know what participates in folds with this, so have
4304 to assume anything could */
4306 /* XXX We could know exactly what does fold to this if the
4307 * reverse folds are loaded, as currently in S_regclass().
4308 * But we do know that under /iaa nothing in the ASCII
4309 * range can participate */
4310 if (OP(scan) == EXACTFA) {
4311 _invlist_union_complement_2nd(EXACTF_invlist,
4312 PL_XPosix_ptrs[_CC_ASCII],
4316 EXACTF_invlist = _add_range_to_invlist(EXACTF_invlist,
4321 if (flags & SCF_DO_STCLASS_AND) {
4322 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4323 ANYOF_POSIXL_ZERO(data->start_class);
4324 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4326 else if (flags & SCF_DO_STCLASS_OR) {
4327 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4328 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4330 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4331 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4333 flags &= ~SCF_DO_STCLASS;
4334 SvREFCNT_dec(EXACTF_invlist);
4336 else if (REGNODE_VARIES(OP(scan))) {
4337 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4338 I32 fl = 0, f = flags;
4339 regnode * const oscan = scan;
4340 regnode_ssc this_class;
4341 regnode_ssc *oclass = NULL;
4342 I32 next_is_eval = 0;
4344 switch (PL_regkind[OP(scan)]) {
4345 case WHILEM: /* End of (?:...)* . */
4346 scan = NEXTOPER(scan);
4349 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4350 next = NEXTOPER(scan);
4351 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
4353 maxcount = REG_INFTY;
4354 next = regnext(scan);
4355 scan = NEXTOPER(scan);
4359 if (flags & SCF_DO_SUBSTR)
4364 if (flags & SCF_DO_STCLASS) {
4366 maxcount = REG_INFTY;
4367 next = regnext(scan);
4368 scan = NEXTOPER(scan);
4371 if (flags & SCF_DO_SUBSTR) {
4372 scan_commit(pRExC_state, data, minlenp, is_inf);
4373 /* Cannot extend fixed substrings */
4374 data->longest = &(data->longest_float);
4376 is_inf = is_inf_internal = 1;
4377 scan = regnext(scan);
4378 goto optimize_curly_tail;
4380 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4381 && (scan->flags == stopparen))
4386 mincount = ARG1(scan);
4387 maxcount = ARG2(scan);
4389 next = regnext(scan);
4390 if (OP(scan) == CURLYX) {
4391 I32 lp = (data ? *(data->last_closep) : 0);
4392 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4394 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4395 next_is_eval = (OP(scan) == EVAL);
4397 if (flags & SCF_DO_SUBSTR) {
4399 scan_commit(pRExC_state, data, minlenp, is_inf);
4400 /* Cannot extend fixed substrings */
4401 pos_before = data->pos_min;
4405 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4407 data->flags |= SF_IS_INF;
4409 if (flags & SCF_DO_STCLASS) {
4410 ssc_init(pRExC_state, &this_class);
4411 oclass = data->start_class;
4412 data->start_class = &this_class;
4413 f |= SCF_DO_STCLASS_AND;
4414 f &= ~SCF_DO_STCLASS_OR;
4416 /* Exclude from super-linear cache processing any {n,m}
4417 regops for which the combination of input pos and regex
4418 pos is not enough information to determine if a match
4421 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4422 regex pos at the \s*, the prospects for a match depend not
4423 only on the input position but also on how many (bar\s*)
4424 repeats into the {4,8} we are. */
4425 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4426 f &= ~SCF_WHILEM_VISITED_POS;
4428 /* This will finish on WHILEM, setting scan, or on NULL: */
4429 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4430 last, data, stopparen, recursed_depth, NULL,
4432 ? (f & ~SCF_DO_SUBSTR)
4436 if (flags & SCF_DO_STCLASS)
4437 data->start_class = oclass;
4438 if (mincount == 0 || minnext == 0) {
4439 if (flags & SCF_DO_STCLASS_OR) {
4440 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4442 else if (flags & SCF_DO_STCLASS_AND) {
4443 /* Switch to OR mode: cache the old value of
4444 * data->start_class */
4446 StructCopy(data->start_class, and_withp, regnode_ssc);
4447 flags &= ~SCF_DO_STCLASS_AND;
4448 StructCopy(&this_class, data->start_class, regnode_ssc);
4449 flags |= SCF_DO_STCLASS_OR;
4450 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
4452 } else { /* Non-zero len */
4453 if (flags & SCF_DO_STCLASS_OR) {
4454 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4455 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4457 else if (flags & SCF_DO_STCLASS_AND)
4458 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4459 flags &= ~SCF_DO_STCLASS;
4461 if (!scan) /* It was not CURLYX, but CURLY. */
4463 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4464 /* ? quantifier ok, except for (?{ ... }) */
4465 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4466 && (minnext == 0) && (deltanext == 0)
4467 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4468 && maxcount <= REG_INFTY/3) /* Complement check for big
4471 /* Fatal warnings may leak the regexp without this: */
4472 SAVEFREESV(RExC_rx_sv);
4473 ckWARNreg(RExC_parse,
4474 "Quantifier unexpected on zero-length expression");
4475 (void)ReREFCNT_inc(RExC_rx_sv);
4478 min += minnext * mincount;
4479 is_inf_internal |= deltanext == SSize_t_MAX
4480 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4481 is_inf |= is_inf_internal;
4483 delta = SSize_t_MAX;
4485 delta += (minnext + deltanext) * maxcount
4486 - minnext * mincount;
4488 /* Try powerful optimization CURLYX => CURLYN. */
4489 if ( OP(oscan) == CURLYX && data
4490 && data->flags & SF_IN_PAR
4491 && !(data->flags & SF_HAS_EVAL)
4492 && !deltanext && minnext == 1 ) {
4493 /* Try to optimize to CURLYN. */
4494 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4495 regnode * const nxt1 = nxt;
4502 if (!REGNODE_SIMPLE(OP(nxt))
4503 && !(PL_regkind[OP(nxt)] == EXACT
4504 && STR_LEN(nxt) == 1))
4510 if (OP(nxt) != CLOSE)
4512 if (RExC_open_parens) {
4513 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4514 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4516 /* Now we know that nxt2 is the only contents: */
4517 oscan->flags = (U8)ARG(nxt);
4519 OP(nxt1) = NOTHING; /* was OPEN. */
4522 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4523 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4524 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4525 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4526 OP(nxt + 1) = OPTIMIZED; /* was count. */
4527 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4532 /* Try optimization CURLYX => CURLYM. */
4533 if ( OP(oscan) == CURLYX && data
4534 && !(data->flags & SF_HAS_PAR)
4535 && !(data->flags & SF_HAS_EVAL)
4536 && !deltanext /* atom is fixed width */
4537 && minnext != 0 /* CURLYM can't handle zero width */
4539 /* Nor characters whose fold at run-time may be
4540 * multi-character */
4541 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4543 /* XXXX How to optimize if data == 0? */
4544 /* Optimize to a simpler form. */
4545 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4549 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4550 && (OP(nxt2) != WHILEM))
4552 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4553 /* Need to optimize away parenths. */
4554 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4555 /* Set the parenth number. */
4556 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4558 oscan->flags = (U8)ARG(nxt);
4559 if (RExC_open_parens) {
4560 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4561 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4563 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4564 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4567 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4568 OP(nxt + 1) = OPTIMIZED; /* was count. */
4569 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4570 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4573 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4574 regnode *nnxt = regnext(nxt1);
4576 if (reg_off_by_arg[OP(nxt1)])
4577 ARG_SET(nxt1, nxt2 - nxt1);
4578 else if (nxt2 - nxt1 < U16_MAX)
4579 NEXT_OFF(nxt1) = nxt2 - nxt1;
4581 OP(nxt) = NOTHING; /* Cannot beautify */
4586 /* Optimize again: */
4587 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4588 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4593 else if ((OP(oscan) == CURLYX)
4594 && (flags & SCF_WHILEM_VISITED_POS)
4595 /* See the comment on a similar expression above.
4596 However, this time it's not a subexpression
4597 we care about, but the expression itself. */
4598 && (maxcount == REG_INFTY)
4599 && data && ++data->whilem_c < 16) {
4600 /* This stays as CURLYX, we can put the count/of pair. */
4601 /* Find WHILEM (as in regexec.c) */
4602 regnode *nxt = oscan + NEXT_OFF(oscan);
4604 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4606 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4607 | (RExC_whilem_seen << 4)); /* On WHILEM */
4609 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4611 if (flags & SCF_DO_SUBSTR) {
4612 SV *last_str = NULL;
4613 STRLEN last_chrs = 0;
4614 int counted = mincount != 0;
4616 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4618 SSize_t b = pos_before >= data->last_start_min
4619 ? pos_before : data->last_start_min;
4621 const char * const s = SvPV_const(data->last_found, l);
4622 SSize_t old = b - data->last_start_min;
4625 old = utf8_hop((U8*)s, old) - (U8*)s;
4627 /* Get the added string: */
4628 last_str = newSVpvn_utf8(s + old, l, UTF);
4629 last_chrs = UTF ? utf8_length((U8*)(s + old),
4630 (U8*)(s + old + l)) : l;
4631 if (deltanext == 0 && pos_before == b) {
4632 /* What was added is a constant string */
4635 SvGROW(last_str, (mincount * l) + 1);
4636 repeatcpy(SvPVX(last_str) + l,
4637 SvPVX_const(last_str), l,
4639 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4640 /* Add additional parts. */
4641 SvCUR_set(data->last_found,
4642 SvCUR(data->last_found) - l);
4643 sv_catsv(data->last_found, last_str);
4645 SV * sv = data->last_found;
4647 SvUTF8(sv) && SvMAGICAL(sv) ?
4648 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4649 if (mg && mg->mg_len >= 0)
4650 mg->mg_len += last_chrs * (mincount-1);
4652 last_chrs *= mincount;
4653 data->last_end += l * (mincount - 1);
4656 /* start offset must point into the last copy */
4657 data->last_start_min += minnext * (mincount - 1);
4658 data->last_start_max += is_inf ? SSize_t_MAX
4659 : (maxcount - 1) * (minnext + data->pos_delta);
4662 /* It is counted once already... */
4663 data->pos_min += minnext * (mincount - counted);
4665 PerlIO_printf(Perl_debug_log, "counted=%"UVdf" deltanext=%"UVdf
4666 " SSize_t_MAX=%"UVdf" minnext=%"UVdf
4667 " maxcount=%"UVdf" mincount=%"UVdf"\n",
4668 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4670 if (deltanext != SSize_t_MAX)
4671 PerlIO_printf(Perl_debug_log, "LHS=%"UVdf" RHS=%"UVdf"\n",
4672 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4673 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4675 if (deltanext == SSize_t_MAX
4676 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4677 data->pos_delta = SSize_t_MAX;
4679 data->pos_delta += - counted * deltanext +
4680 (minnext + deltanext) * maxcount - minnext * mincount;
4681 if (mincount != maxcount) {
4682 /* Cannot extend fixed substrings found inside
4684 scan_commit(pRExC_state, data, minlenp, is_inf);
4685 if (mincount && last_str) {
4686 SV * const sv = data->last_found;
4687 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4688 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4692 sv_setsv(sv, last_str);
4693 data->last_end = data->pos_min;
4694 data->last_start_min = data->pos_min - last_chrs;
4695 data->last_start_max = is_inf
4697 : data->pos_min + data->pos_delta - last_chrs;
4699 data->longest = &(data->longest_float);
4701 SvREFCNT_dec(last_str);
4703 if (data && (fl & SF_HAS_EVAL))
4704 data->flags |= SF_HAS_EVAL;
4705 optimize_curly_tail:
4706 if (OP(oscan) != CURLYX) {
4707 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4709 NEXT_OFF(oscan) += NEXT_OFF(next);
4715 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4720 if (flags & SCF_DO_SUBSTR) {
4721 /* Cannot expect anything... */
4722 scan_commit(pRExC_state, data, minlenp, is_inf);
4723 data->longest = &(data->longest_float);
4725 is_inf = is_inf_internal = 1;
4726 if (flags & SCF_DO_STCLASS_OR) {
4727 if (OP(scan) == CLUMP) {
4728 /* Actually is any start char, but very few code points
4729 * aren't start characters */
4730 ssc_match_all_cp(data->start_class);
4733 ssc_anything(data->start_class);
4736 flags &= ~SCF_DO_STCLASS;
4740 else if (OP(scan) == LNBREAK) {
4741 if (flags & SCF_DO_STCLASS) {
4742 if (flags & SCF_DO_STCLASS_AND) {
4743 ssc_intersection(data->start_class,
4744 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
4745 ssc_clear_locale(data->start_class);
4746 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4748 else if (flags & SCF_DO_STCLASS_OR) {
4749 ssc_union(data->start_class,
4750 PL_XPosix_ptrs[_CC_VERTSPACE],
4752 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4754 /* See commit msg for
4755 * 749e076fceedeb708a624933726e7989f2302f6a */
4756 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4758 flags &= ~SCF_DO_STCLASS;
4761 delta++; /* Because of the 2 char string cr-lf */
4762 if (flags & SCF_DO_SUBSTR) {
4763 /* Cannot expect anything... */
4764 scan_commit(pRExC_state, data, minlenp, is_inf);
4766 data->pos_delta += 1;
4767 data->longest = &(data->longest_float);
4770 else if (REGNODE_SIMPLE(OP(scan))) {
4772 if (flags & SCF_DO_SUBSTR) {
4773 scan_commit(pRExC_state, data, minlenp, is_inf);
4777 if (flags & SCF_DO_STCLASS) {
4779 SV* my_invlist = sv_2mortal(_new_invlist(0));
4782 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4783 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4785 /* Some of the logic below assumes that switching
4786 locale on will only add false positives. */
4791 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
4796 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4797 ssc_match_all_cp(data->start_class);
4802 SV* REG_ANY_invlist = _new_invlist(2);
4803 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
4805 if (flags & SCF_DO_STCLASS_OR) {
4806 ssc_union(data->start_class,
4808 TRUE /* TRUE => invert, hence all but \n
4812 else if (flags & SCF_DO_STCLASS_AND) {
4813 ssc_intersection(data->start_class,
4815 TRUE /* TRUE => invert */
4817 ssc_clear_locale(data->start_class);
4819 SvREFCNT_dec_NN(REG_ANY_invlist);
4824 if (flags & SCF_DO_STCLASS_AND)
4825 ssc_and(pRExC_state, data->start_class,
4826 (regnode_charclass *) scan);
4828 ssc_or(pRExC_state, data->start_class,
4829 (regnode_charclass *) scan);
4837 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
4838 if (flags & SCF_DO_STCLASS_AND) {
4839 bool was_there = cBOOL(
4840 ANYOF_POSIXL_TEST(data->start_class,
4842 ANYOF_POSIXL_ZERO(data->start_class);
4843 if (was_there) { /* Do an AND */
4844 ANYOF_POSIXL_SET(data->start_class, namedclass);
4846 /* No individual code points can now match */
4847 data->start_class->invlist
4848 = sv_2mortal(_new_invlist(0));
4851 int complement = namedclass + ((invert) ? -1 : 1);
4853 assert(flags & SCF_DO_STCLASS_OR);
4855 /* If the complement of this class was already there,
4856 * the result is that they match all code points,
4857 * (\d + \D == everything). Remove the classes from
4858 * future consideration. Locale is not relevant in
4860 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
4861 ssc_match_all_cp(data->start_class);
4862 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
4863 ANYOF_POSIXL_CLEAR(data->start_class, complement);
4865 else { /* The usual case; just add this class to the
4867 ANYOF_POSIXL_SET(data->start_class, namedclass);
4872 case NPOSIXA: /* For these, we always know the exact set of
4877 if (FLAGS(scan) == _CC_ASCII) {
4878 my_invlist = PL_XPosix_ptrs[_CC_ASCII];
4881 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
4882 PL_XPosix_ptrs[_CC_ASCII],
4893 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
4895 /* NPOSIXD matches all upper Latin1 code points unless the
4896 * target string being matched is UTF-8, which is
4897 * unknowable until match time. Since we are going to
4898 * invert, we want to get rid of all of them so that the
4899 * inversion will match all */
4900 if (OP(scan) == NPOSIXD) {
4901 _invlist_subtract(my_invlist, PL_UpperLatin1,
4907 if (flags & SCF_DO_STCLASS_AND) {
4908 ssc_intersection(data->start_class, my_invlist, invert);
4909 ssc_clear_locale(data->start_class);
4912 assert(flags & SCF_DO_STCLASS_OR);
4913 ssc_union(data->start_class, my_invlist, invert);
4916 if (flags & SCF_DO_STCLASS_OR)
4917 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4918 flags &= ~SCF_DO_STCLASS;
4921 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4922 data->flags |= (OP(scan) == MEOL
4925 scan_commit(pRExC_state, data, minlenp, is_inf);
4928 else if ( PL_regkind[OP(scan)] == BRANCHJ
4929 /* Lookbehind, or need to calculate parens/evals/stclass: */
4930 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4931 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4932 if ( OP(scan) == UNLESSM &&
4934 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4935 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4938 regnode *upto= regnext(scan);
4940 SV * const mysv_val=sv_newmortal();
4941 DEBUG_STUDYDATA("OPFAIL",data,depth);
4943 /*DEBUG_PARSE_MSG("opfail");*/
4944 regprop(RExC_rx, mysv_val, upto, NULL);
4945 PerlIO_printf(Perl_debug_log,
4946 "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4947 SvPV_nolen_const(mysv_val),
4948 (IV)REG_NODE_NUM(upto),
4953 NEXT_OFF(scan) = upto - scan;
4954 for (opt= scan + 1; opt < upto ; opt++)
4955 OP(opt) = OPTIMIZED;
4959 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4960 || OP(scan) == UNLESSM )
4962 /* Negative Lookahead/lookbehind
4963 In this case we can't do fixed string optimisation.
4966 SSize_t deltanext, minnext, fake = 0;
4971 data_fake.flags = 0;
4973 data_fake.whilem_c = data->whilem_c;
4974 data_fake.last_closep = data->last_closep;
4977 data_fake.last_closep = &fake;
4978 data_fake.pos_delta = delta;
4979 if ( flags & SCF_DO_STCLASS && !scan->flags
4980 && OP(scan) == IFMATCH ) { /* Lookahead */
4981 ssc_init(pRExC_state, &intrnl);
4982 data_fake.start_class = &intrnl;
4983 f |= SCF_DO_STCLASS_AND;
4985 if (flags & SCF_WHILEM_VISITED_POS)
4986 f |= SCF_WHILEM_VISITED_POS;
4987 next = regnext(scan);
4988 nscan = NEXTOPER(NEXTOPER(scan));
4989 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4990 last, &data_fake, stopparen,
4991 recursed_depth, NULL, f, depth+1);
4994 FAIL("Variable length lookbehind not implemented");
4996 else if (minnext > (I32)U8_MAX) {
4997 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5000 scan->flags = (U8)minnext;
5003 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5005 if (data_fake.flags & SF_HAS_EVAL)
5006 data->flags |= SF_HAS_EVAL;
5007 data->whilem_c = data_fake.whilem_c;
5009 if (f & SCF_DO_STCLASS_AND) {
5010 if (flags & SCF_DO_STCLASS_OR) {
5011 /* OR before, AND after: ideally we would recurse with
5012 * data_fake to get the AND applied by study of the
5013 * remainder of the pattern, and then derecurse;
5014 * *** HACK *** for now just treat as "no information".
5015 * See [perl #56690].
5017 ssc_init(pRExC_state, data->start_class);
5019 /* AND before and after: combine and continue */
5020 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5024 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5026 /* Positive Lookahead/lookbehind
5027 In this case we can do fixed string optimisation,
5028 but we must be careful about it. Note in the case of
5029 lookbehind the positions will be offset by the minimum
5030 length of the pattern, something we won't know about
5031 until after the recurse.
5033 SSize_t deltanext, fake = 0;
5037 /* We use SAVEFREEPV so that when the full compile
5038 is finished perl will clean up the allocated
5039 minlens when it's all done. This way we don't
5040 have to worry about freeing them when we know
5041 they wont be used, which would be a pain.
5044 Newx( minnextp, 1, SSize_t );
5045 SAVEFREEPV(minnextp);
5048 StructCopy(data, &data_fake, scan_data_t);
5049 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5052 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5053 data_fake.last_found=newSVsv(data->last_found);
5057 data_fake.last_closep = &fake;
5058 data_fake.flags = 0;
5059 data_fake.pos_delta = delta;
5061 data_fake.flags |= SF_IS_INF;
5062 if ( flags & SCF_DO_STCLASS && !scan->flags
5063 && OP(scan) == IFMATCH ) { /* Lookahead */
5064 ssc_init(pRExC_state, &intrnl);
5065 data_fake.start_class = &intrnl;
5066 f |= SCF_DO_STCLASS_AND;
5068 if (flags & SCF_WHILEM_VISITED_POS)
5069 f |= SCF_WHILEM_VISITED_POS;
5070 next = regnext(scan);
5071 nscan = NEXTOPER(NEXTOPER(scan));
5073 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5074 &deltanext, last, &data_fake,
5075 stopparen, recursed_depth, NULL,
5079 FAIL("Variable length lookbehind not implemented");
5081 else if (*minnextp > (I32)U8_MAX) {
5082 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5085 scan->flags = (U8)*minnextp;
5090 if (f & SCF_DO_STCLASS_AND) {
5091 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5094 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5096 if (data_fake.flags & SF_HAS_EVAL)
5097 data->flags |= SF_HAS_EVAL;
5098 data->whilem_c = data_fake.whilem_c;
5099 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5100 if (RExC_rx->minlen<*minnextp)
5101 RExC_rx->minlen=*minnextp;
5102 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5103 SvREFCNT_dec_NN(data_fake.last_found);
5105 if ( data_fake.minlen_fixed != minlenp )
5107 data->offset_fixed= data_fake.offset_fixed;
5108 data->minlen_fixed= data_fake.minlen_fixed;
5109 data->lookbehind_fixed+= scan->flags;
5111 if ( data_fake.minlen_float != minlenp )
5113 data->minlen_float= data_fake.minlen_float;
5114 data->offset_float_min=data_fake.offset_float_min;
5115 data->offset_float_max=data_fake.offset_float_max;
5116 data->lookbehind_float+= scan->flags;
5123 else if (OP(scan) == OPEN) {
5124 if (stopparen != (I32)ARG(scan))
5127 else if (OP(scan) == CLOSE) {
5128 if (stopparen == (I32)ARG(scan)) {
5131 if ((I32)ARG(scan) == is_par) {
5132 next = regnext(scan);
5134 if ( next && (OP(next) != WHILEM) && next < last)
5135 is_par = 0; /* Disable optimization */
5138 *(data->last_closep) = ARG(scan);
5140 else if (OP(scan) == EVAL) {
5142 data->flags |= SF_HAS_EVAL;
5144 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5145 if (flags & SCF_DO_SUBSTR) {
5146 scan_commit(pRExC_state, data, minlenp, is_inf);
5147 flags &= ~SCF_DO_SUBSTR;
5149 if (data && OP(scan)==ACCEPT) {
5150 data->flags |= SCF_SEEN_ACCEPT;
5155 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5157 if (flags & SCF_DO_SUBSTR) {
5158 scan_commit(pRExC_state, data, minlenp, is_inf);
5159 data->longest = &(data->longest_float);
5161 is_inf = is_inf_internal = 1;
5162 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5163 ssc_anything(data->start_class);
5164 flags &= ~SCF_DO_STCLASS;
5166 else if (OP(scan) == GPOS) {
5167 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5168 !(delta || is_inf || (data && data->pos_delta)))
5170 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5171 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5172 if (RExC_rx->gofs < (STRLEN)min)
5173 RExC_rx->gofs = min;
5175 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5179 #ifdef TRIE_STUDY_OPT
5180 #ifdef FULL_TRIE_STUDY
5181 else if (PL_regkind[OP(scan)] == TRIE) {
5182 /* NOTE - There is similar code to this block above for handling
5183 BRANCH nodes on the initial study. If you change stuff here
5185 regnode *trie_node= scan;
5186 regnode *tail= regnext(scan);
5187 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5188 SSize_t max1 = 0, min1 = SSize_t_MAX;
5191 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5192 /* Cannot merge strings after this. */
5193 scan_commit(pRExC_state, data, minlenp, is_inf);
5195 if (flags & SCF_DO_STCLASS)
5196 ssc_init_zero(pRExC_state, &accum);
5202 const regnode *nextbranch= NULL;
5205 for ( word=1 ; word <= trie->wordcount ; word++)
5207 SSize_t deltanext=0, minnext=0, f = 0, fake;
5208 regnode_ssc this_class;
5210 data_fake.flags = 0;
5212 data_fake.whilem_c = data->whilem_c;
5213 data_fake.last_closep = data->last_closep;
5216 data_fake.last_closep = &fake;
5217 data_fake.pos_delta = delta;
5218 if (flags & SCF_DO_STCLASS) {
5219 ssc_init(pRExC_state, &this_class);
5220 data_fake.start_class = &this_class;
5221 f = SCF_DO_STCLASS_AND;
5223 if (flags & SCF_WHILEM_VISITED_POS)
5224 f |= SCF_WHILEM_VISITED_POS;
5226 if (trie->jump[word]) {
5228 nextbranch = trie_node + trie->jump[0];
5229 scan= trie_node + trie->jump[word];
5230 /* We go from the jump point to the branch that follows
5231 it. Note this means we need the vestigal unused
5232 branches even though they arent otherwise used. */
5233 minnext = study_chunk(pRExC_state, &scan, minlenp,
5234 &deltanext, (regnode *)nextbranch, &data_fake,
5235 stopparen, recursed_depth, NULL, f,depth+1);
5237 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5238 nextbranch= regnext((regnode*)nextbranch);
5240 if (min1 > (SSize_t)(minnext + trie->minlen))
5241 min1 = minnext + trie->minlen;
5242 if (deltanext == SSize_t_MAX) {
5243 is_inf = is_inf_internal = 1;
5245 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5246 max1 = minnext + deltanext + trie->maxlen;
5248 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5250 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5251 if ( stopmin > min + min1)
5252 stopmin = min + min1;
5253 flags &= ~SCF_DO_SUBSTR;
5255 data->flags |= SCF_SEEN_ACCEPT;
5258 if (data_fake.flags & SF_HAS_EVAL)
5259 data->flags |= SF_HAS_EVAL;
5260 data->whilem_c = data_fake.whilem_c;
5262 if (flags & SCF_DO_STCLASS)
5263 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5266 if (flags & SCF_DO_SUBSTR) {
5267 data->pos_min += min1;
5268 data->pos_delta += max1 - min1;
5269 if (max1 != min1 || is_inf)
5270 data->longest = &(data->longest_float);
5273 delta += max1 - min1;
5274 if (flags & SCF_DO_STCLASS_OR) {
5275 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5277 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5278 flags &= ~SCF_DO_STCLASS;
5281 else if (flags & SCF_DO_STCLASS_AND) {
5283 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5284 flags &= ~SCF_DO_STCLASS;
5287 /* Switch to OR mode: cache the old value of
5288 * data->start_class */
5290 StructCopy(data->start_class, and_withp, regnode_ssc);
5291 flags &= ~SCF_DO_STCLASS_AND;
5292 StructCopy(&accum, data->start_class, regnode_ssc);
5293 flags |= SCF_DO_STCLASS_OR;
5300 else if (PL_regkind[OP(scan)] == TRIE) {
5301 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5304 min += trie->minlen;
5305 delta += (trie->maxlen - trie->minlen);
5306 flags &= ~SCF_DO_STCLASS; /* xxx */
5307 if (flags & SCF_DO_SUBSTR) {
5308 /* Cannot expect anything... */
5309 scan_commit(pRExC_state, data, minlenp, is_inf);
5310 data->pos_min += trie->minlen;
5311 data->pos_delta += (trie->maxlen - trie->minlen);
5312 if (trie->maxlen != trie->minlen)
5313 data->longest = &(data->longest_float);
5315 if (trie->jump) /* no more substrings -- for now /grr*/
5316 flags &= ~SCF_DO_SUBSTR;
5318 #endif /* old or new */
5319 #endif /* TRIE_STUDY_OPT */
5321 /* Else: zero-length, ignore. */
5322 scan = regnext(scan);
5324 /* If we are exiting a recursion we can unset its recursed bit
5325 * and allow ourselves to enter it again - no danger of an
5326 * infinite loop there.
5327 if (stopparen > -1 && recursed) {
5328 DEBUG_STUDYDATA("unset:", data,depth);
5329 PAREN_UNSET( recursed, stopparen);
5333 DEBUG_STUDYDATA("frame-end:",data,depth);
5334 DEBUG_PEEP("fend", scan, depth);
5335 /* restore previous context */
5338 stopparen = frame->stop;
5339 recursed_depth = frame->prev_recursed_depth;
5342 frame = frame->prev;
5343 goto fake_study_recurse;
5348 DEBUG_STUDYDATA("pre-fin:",data,depth);
5351 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5353 if (flags & SCF_DO_SUBSTR && is_inf)
5354 data->pos_delta = SSize_t_MAX - data->pos_min;
5355 if (is_par > (I32)U8_MAX)
5357 if (is_par && pars==1 && data) {
5358 data->flags |= SF_IN_PAR;
5359 data->flags &= ~SF_HAS_PAR;
5361 else if (pars && data) {
5362 data->flags |= SF_HAS_PAR;
5363 data->flags &= ~SF_IN_PAR;
5365 if (flags & SCF_DO_STCLASS_OR)
5366 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5367 if (flags & SCF_TRIE_RESTUDY)
5368 data->flags |= SCF_TRIE_RESTUDY;
5370 DEBUG_STUDYDATA("post-fin:",data,depth);
5373 SSize_t final_minlen= min < stopmin ? min : stopmin;
5375 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) && (RExC_maxlen < final_minlen + delta)) {
5376 RExC_maxlen = final_minlen + delta;
5378 return final_minlen;
5384 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5386 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5388 PERL_ARGS_ASSERT_ADD_DATA;
5390 Renewc(RExC_rxi->data,
5391 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5392 char, struct reg_data);
5394 Renew(RExC_rxi->data->what, count + n, U8);
5396 Newx(RExC_rxi->data->what, n, U8);
5397 RExC_rxi->data->count = count + n;
5398 Copy(s, RExC_rxi->data->what + count, n, U8);
5402 /*XXX: todo make this not included in a non debugging perl */
5403 #ifndef PERL_IN_XSUB_RE
5405 Perl_reginitcolors(pTHX)
5408 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5410 char *t = savepv(s);
5414 t = strchr(t, '\t');
5420 PL_colors[i] = t = (char *)"";
5425 PL_colors[i++] = (char *)"";
5432 #ifdef TRIE_STUDY_OPT
5433 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5436 (data.flags & SCF_TRIE_RESTUDY) \
5444 #define CHECK_RESTUDY_GOTO_butfirst
5448 * pregcomp - compile a regular expression into internal code
5450 * Decides which engine's compiler to call based on the hint currently in
5454 #ifndef PERL_IN_XSUB_RE
5456 /* return the currently in-scope regex engine (or the default if none) */
5458 regexp_engine const *
5459 Perl_current_re_engine(pTHX)
5463 if (IN_PERL_COMPILETIME) {
5464 HV * const table = GvHV(PL_hintgv);
5467 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5468 return &PL_core_reg_engine;
5469 ptr = hv_fetchs(table, "regcomp", FALSE);
5470 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5471 return &PL_core_reg_engine;
5472 return INT2PTR(regexp_engine*,SvIV(*ptr));
5476 if (!PL_curcop->cop_hints_hash)
5477 return &PL_core_reg_engine;
5478 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5479 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5480 return &PL_core_reg_engine;
5481 return INT2PTR(regexp_engine*,SvIV(ptr));
5487 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5490 regexp_engine const *eng = current_re_engine();
5491 GET_RE_DEBUG_FLAGS_DECL;
5493 PERL_ARGS_ASSERT_PREGCOMP;
5495 /* Dispatch a request to compile a regexp to correct regexp engine. */
5497 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5500 return CALLREGCOMP_ENG(eng, pattern, flags);
5504 /* public(ish) entry point for the perl core's own regex compiling code.
5505 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5506 * pattern rather than a list of OPs, and uses the internal engine rather
5507 * than the current one */
5510 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5512 SV *pat = pattern; /* defeat constness! */
5513 PERL_ARGS_ASSERT_RE_COMPILE;
5514 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5515 #ifdef PERL_IN_XSUB_RE
5518 &PL_core_reg_engine,
5520 NULL, NULL, rx_flags, 0);
5524 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5525 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5526 * point to the realloced string and length.
5528 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5532 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5533 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5535 U8 *const src = (U8*)*pat_p;
5538 STRLEN s = 0, d = 0;
5540 GET_RE_DEBUG_FLAGS_DECL;
5542 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5543 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5545 Newx(dst, *plen_p * 2 + 1, U8);
5547 while (s < *plen_p) {
5548 if (NATIVE_BYTE_IS_INVARIANT(src[s]))
5551 dst[d++] = UTF8_EIGHT_BIT_HI(src[s]);
5552 dst[d] = UTF8_EIGHT_BIT_LO(src[s]);
5554 if (n < num_code_blocks) {
5555 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5556 pRExC_state->code_blocks[n].start = d;
5557 assert(dst[d] == '(');
5560 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5561 pRExC_state->code_blocks[n].end = d;
5562 assert(dst[d] == ')');
5572 *pat_p = (char*) dst;
5574 RExC_orig_utf8 = RExC_utf8 = 1;
5579 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5580 * while recording any code block indices, and handling overloading,
5581 * nested qr// objects etc. If pat is null, it will allocate a new
5582 * string, or just return the first arg, if there's only one.
5584 * Returns the malloced/updated pat.
5585 * patternp and pat_count is the array of SVs to be concatted;
5586 * oplist is the optional list of ops that generated the SVs;
5587 * recompile_p is a pointer to a boolean that will be set if
5588 * the regex will need to be recompiled.
5589 * delim, if non-null is an SV that will be inserted between each element
5593 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5594 SV *pat, SV ** const patternp, int pat_count,
5595 OP *oplist, bool *recompile_p, SV *delim)
5599 bool use_delim = FALSE;
5600 bool alloced = FALSE;
5602 /* if we know we have at least two args, create an empty string,
5603 * then concatenate args to that. For no args, return an empty string */
5604 if (!pat && pat_count != 1) {
5605 pat = newSVpvn("", 0);
5610 for (svp = patternp; svp < patternp + pat_count; svp++) {
5613 STRLEN orig_patlen = 0;
5615 SV *msv = use_delim ? delim : *svp;
5616 if (!msv) msv = &PL_sv_undef;
5618 /* if we've got a delimiter, we go round the loop twice for each
5619 * svp slot (except the last), using the delimiter the second
5628 if (SvTYPE(msv) == SVt_PVAV) {
5629 /* we've encountered an interpolated array within
5630 * the pattern, e.g. /...@a..../. Expand the list of elements,
5631 * then recursively append elements.
5632 * The code in this block is based on S_pushav() */
5634 AV *const av = (AV*)msv;
5635 const SSize_t maxarg = AvFILL(av) + 1;
5639 assert(oplist->op_type == OP_PADAV
5640 || oplist->op_type == OP_RV2AV);
5641 oplist = oplist->op_sibling;;
5644 if (SvRMAGICAL(av)) {
5647 Newx(array, maxarg, SV*);
5649 for (i=0; i < maxarg; i++) {
5650 SV ** const svp = av_fetch(av, i, FALSE);
5651 array[i] = svp ? *svp : &PL_sv_undef;
5655 array = AvARRAY(av);
5657 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5658 array, maxarg, NULL, recompile_p,
5660 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5666 /* we make the assumption here that each op in the list of
5667 * op_siblings maps to one SV pushed onto the stack,
5668 * except for code blocks, with have both an OP_NULL and
5670 * This allows us to match up the list of SVs against the
5671 * list of OPs to find the next code block.
5673 * Note that PUSHMARK PADSV PADSV ..
5675 * PADRANGE PADSV PADSV ..
5676 * so the alignment still works. */
5679 if (oplist->op_type == OP_NULL
5680 && (oplist->op_flags & OPf_SPECIAL))
5682 assert(n < pRExC_state->num_code_blocks);
5683 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5684 pRExC_state->code_blocks[n].block = oplist;
5685 pRExC_state->code_blocks[n].src_regex = NULL;
5688 oplist = oplist->op_sibling; /* skip CONST */
5691 oplist = oplist->op_sibling;;
5694 /* apply magic and QR overloading to arg */
5697 if (SvROK(msv) && SvAMAGIC(msv)) {
5698 SV *sv = AMG_CALLunary(msv, regexp_amg);
5702 if (SvTYPE(sv) != SVt_REGEXP)
5703 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5708 /* try concatenation overload ... */
5709 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5710 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5713 /* overloading involved: all bets are off over literal
5714 * code. Pretend we haven't seen it */
5715 pRExC_state->num_code_blocks -= n;
5719 /* ... or failing that, try "" overload */
5720 while (SvAMAGIC(msv)
5721 && (sv = AMG_CALLunary(msv, string_amg))
5725 && SvRV(msv) == SvRV(sv))
5730 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5734 /* this is a partially unrolled
5735 * sv_catsv_nomg(pat, msv);
5736 * that allows us to adjust code block indices if
5739 char *dst = SvPV_force_nomg(pat, dlen);
5741 if (SvUTF8(msv) && !SvUTF8(pat)) {
5742 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5743 sv_setpvn(pat, dst, dlen);
5746 sv_catsv_nomg(pat, msv);
5753 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5756 /* extract any code blocks within any embedded qr//'s */
5757 if (rx && SvTYPE(rx) == SVt_REGEXP
5758 && RX_ENGINE((REGEXP*)rx)->op_comp)
5761 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5762 if (ri->num_code_blocks) {
5764 /* the presence of an embedded qr// with code means
5765 * we should always recompile: the text of the
5766 * qr// may not have changed, but it may be a
5767 * different closure than last time */
5769 Renew(pRExC_state->code_blocks,
5770 pRExC_state->num_code_blocks + ri->num_code_blocks,
5771 struct reg_code_block);
5772 pRExC_state->num_code_blocks += ri->num_code_blocks;
5774 for (i=0; i < ri->num_code_blocks; i++) {
5775 struct reg_code_block *src, *dst;
5776 STRLEN offset = orig_patlen
5777 + ReANY((REGEXP *)rx)->pre_prefix;
5778 assert(n < pRExC_state->num_code_blocks);
5779 src = &ri->code_blocks[i];
5780 dst = &pRExC_state->code_blocks[n];
5781 dst->start = src->start + offset;
5782 dst->end = src->end + offset;
5783 dst->block = src->block;
5784 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5793 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5802 /* see if there are any run-time code blocks in the pattern.
5803 * False positives are allowed */
5806 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5807 char *pat, STRLEN plen)
5812 for (s = 0; s < plen; s++) {
5813 if (n < pRExC_state->num_code_blocks
5814 && s == pRExC_state->code_blocks[n].start)
5816 s = pRExC_state->code_blocks[n].end;
5820 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5822 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5824 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5831 /* Handle run-time code blocks. We will already have compiled any direct
5832 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5833 * copy of it, but with any literal code blocks blanked out and
5834 * appropriate chars escaped; then feed it into
5836 * eval "qr'modified_pattern'"
5840 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5844 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5846 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5847 * and merge them with any code blocks of the original regexp.
5849 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5850 * instead, just save the qr and return FALSE; this tells our caller that
5851 * the original pattern needs upgrading to utf8.
5855 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5856 char *pat, STRLEN plen)
5860 GET_RE_DEBUG_FLAGS_DECL;
5862 if (pRExC_state->runtime_code_qr) {
5863 /* this is the second time we've been called; this should
5864 * only happen if the main pattern got upgraded to utf8
5865 * during compilation; re-use the qr we compiled first time
5866 * round (which should be utf8 too)
5868 qr = pRExC_state->runtime_code_qr;
5869 pRExC_state->runtime_code_qr = NULL;
5870 assert(RExC_utf8 && SvUTF8(qr));
5876 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5880 /* determine how many extra chars we need for ' and \ escaping */
5881 for (s = 0; s < plen; s++) {
5882 if (pat[s] == '\'' || pat[s] == '\\')
5886 Newx(newpat, newlen, char);
5888 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5890 for (s = 0; s < plen; s++) {
5891 if (n < pRExC_state->num_code_blocks
5892 && s == pRExC_state->code_blocks[n].start)
5894 /* blank out literal code block */
5895 assert(pat[s] == '(');
5896 while (s <= pRExC_state->code_blocks[n].end) {
5904 if (pat[s] == '\'' || pat[s] == '\\')
5909 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5913 PerlIO_printf(Perl_debug_log,
5914 "%sre-parsing pattern for runtime code:%s %s\n",
5915 PL_colors[4],PL_colors[5],newpat);
5918 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5924 PUSHSTACKi(PERLSI_REQUIRE);
5925 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5926 * parsing qr''; normally only q'' does this. It also alters
5928 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5929 SvREFCNT_dec_NN(sv);
5934 SV * const errsv = ERRSV;
5935 if (SvTRUE_NN(errsv))
5937 Safefree(pRExC_state->code_blocks);
5938 /* use croak_sv ? */
5939 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
5942 assert(SvROK(qr_ref));
5944 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5945 /* the leaving below frees the tmp qr_ref.
5946 * Give qr a life of its own */
5954 if (!RExC_utf8 && SvUTF8(qr)) {
5955 /* first time through; the pattern got upgraded; save the
5956 * qr for the next time through */
5957 assert(!pRExC_state->runtime_code_qr);
5958 pRExC_state->runtime_code_qr = qr;
5963 /* extract any code blocks within the returned qr// */
5966 /* merge the main (r1) and run-time (r2) code blocks into one */
5968 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5969 struct reg_code_block *new_block, *dst;
5970 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5973 if (!r2->num_code_blocks) /* we guessed wrong */
5975 SvREFCNT_dec_NN(qr);
5980 r1->num_code_blocks + r2->num_code_blocks,
5981 struct reg_code_block);
5984 while ( i1 < r1->num_code_blocks
5985 || i2 < r2->num_code_blocks)
5987 struct reg_code_block *src;
5990 if (i1 == r1->num_code_blocks) {
5991 src = &r2->code_blocks[i2++];
5994 else if (i2 == r2->num_code_blocks)
5995 src = &r1->code_blocks[i1++];
5996 else if ( r1->code_blocks[i1].start
5997 < r2->code_blocks[i2].start)
5999 src = &r1->code_blocks[i1++];
6000 assert(src->end < r2->code_blocks[i2].start);
6003 assert( r1->code_blocks[i1].start
6004 > r2->code_blocks[i2].start);
6005 src = &r2->code_blocks[i2++];
6007 assert(src->end < r1->code_blocks[i1].start);
6010 assert(pat[src->start] == '(');
6011 assert(pat[src->end] == ')');
6012 dst->start = src->start;
6013 dst->end = src->end;
6014 dst->block = src->block;
6015 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6019 r1->num_code_blocks += r2->num_code_blocks;
6020 Safefree(r1->code_blocks);
6021 r1->code_blocks = new_block;
6024 SvREFCNT_dec_NN(qr);
6030 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6031 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6032 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6033 STRLEN longest_length, bool eol, bool meol)
6035 /* This is the common code for setting up the floating and fixed length
6036 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6037 * as to whether succeeded or not */
6042 if (! (longest_length
6043 || (eol /* Can't have SEOL and MULTI */
6044 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6046 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6047 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6052 /* copy the information about the longest from the reg_scan_data
6053 over to the program. */
6054 if (SvUTF8(sv_longest)) {
6055 *rx_utf8 = sv_longest;
6058 *rx_substr = sv_longest;
6061 /* end_shift is how many chars that must be matched that
6062 follow this item. We calculate it ahead of time as once the
6063 lookbehind offset is added in we lose the ability to correctly
6065 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6066 *rx_end_shift = ml - offset
6067 - longest_length + (SvTAIL(sv_longest) != 0)
6070 t = (eol/* Can't have SEOL and MULTI */
6071 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6072 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6078 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6079 * regular expression into internal code.
6080 * The pattern may be passed either as:
6081 * a list of SVs (patternp plus pat_count)
6082 * a list of OPs (expr)
6083 * If both are passed, the SV list is used, but the OP list indicates
6084 * which SVs are actually pre-compiled code blocks
6086 * The SVs in the list have magic and qr overloading applied to them (and
6087 * the list may be modified in-place with replacement SVs in the latter
6090 * If the pattern hasn't changed from old_re, then old_re will be
6093 * eng is the current engine. If that engine has an op_comp method, then
6094 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6095 * do the initial concatenation of arguments and pass on to the external
6098 * If is_bare_re is not null, set it to a boolean indicating whether the
6099 * arg list reduced (after overloading) to a single bare regex which has
6100 * been returned (i.e. /$qr/).
6102 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6104 * pm_flags contains the PMf_* flags, typically based on those from the
6105 * pm_flags field of the related PMOP. Currently we're only interested in
6106 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6108 * We can't allocate space until we know how big the compiled form will be,
6109 * but we can't compile it (and thus know how big it is) until we've got a
6110 * place to put the code. So we cheat: we compile it twice, once with code
6111 * generation turned off and size counting turned on, and once "for real".
6112 * This also means that we don't allocate space until we are sure that the
6113 * thing really will compile successfully, and we never have to move the
6114 * code and thus invalidate pointers into it. (Note that it has to be in
6115 * one piece because free() must be able to free it all.) [NB: not true in perl]
6117 * Beware that the optimization-preparation code in here knows about some
6118 * of the structure of the compiled regexp. [I'll say.]
6122 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6123 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6124 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6129 regexp_internal *ri;
6137 SV *code_blocksv = NULL;
6138 SV** new_patternp = patternp;
6140 /* these are all flags - maybe they should be turned
6141 * into a single int with different bit masks */
6142 I32 sawlookahead = 0;
6147 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6149 bool runtime_code = 0;
6151 RExC_state_t RExC_state;
6152 RExC_state_t * const pRExC_state = &RExC_state;
6153 #ifdef TRIE_STUDY_OPT
6155 RExC_state_t copyRExC_state;
6157 GET_RE_DEBUG_FLAGS_DECL;
6159 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6161 DEBUG_r(if (!PL_colorset) reginitcolors());
6163 #ifndef PERL_IN_XSUB_RE
6164 /* Initialize these here instead of as-needed, as is quick and avoids
6165 * having to test them each time otherwise */
6166 if (! PL_AboveLatin1) {
6167 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6168 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6169 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6170 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6171 PL_HasMultiCharFold =
6172 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6176 pRExC_state->code_blocks = NULL;
6177 pRExC_state->num_code_blocks = 0;
6180 *is_bare_re = FALSE;
6182 if (expr && (expr->op_type == OP_LIST ||
6183 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6184 /* allocate code_blocks if needed */
6188 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
6189 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6190 ncode++; /* count of DO blocks */
6192 pRExC_state->num_code_blocks = ncode;
6193 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6198 /* compile-time pattern with just OP_CONSTs and DO blocks */
6203 /* find how many CONSTs there are */
6206 if (expr->op_type == OP_CONST)
6209 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6210 if (o->op_type == OP_CONST)
6214 /* fake up an SV array */
6216 assert(!new_patternp);
6217 Newx(new_patternp, n, SV*);
6218 SAVEFREEPV(new_patternp);
6222 if (expr->op_type == OP_CONST)
6223 new_patternp[n] = cSVOPx_sv(expr);
6225 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
6226 if (o->op_type == OP_CONST)
6227 new_patternp[n++] = cSVOPo_sv;
6232 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6233 "Assembling pattern from %d elements%s\n", pat_count,
6234 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6236 /* set expr to the first arg op */
6238 if (pRExC_state->num_code_blocks
6239 && expr->op_type != OP_CONST)
6241 expr = cLISTOPx(expr)->op_first;
6242 assert( expr->op_type == OP_PUSHMARK
6243 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6244 || expr->op_type == OP_PADRANGE);
6245 expr = expr->op_sibling;
6248 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6249 expr, &recompile, NULL);
6251 /* handle bare (possibly after overloading) regex: foo =~ $re */
6256 if (SvTYPE(re) == SVt_REGEXP) {
6260 Safefree(pRExC_state->code_blocks);
6261 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6262 "Precompiled pattern%s\n",
6263 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6269 exp = SvPV_nomg(pat, plen);
6271 if (!eng->op_comp) {
6272 if ((SvUTF8(pat) && IN_BYTES)
6273 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6275 /* make a temporary copy; either to convert to bytes,
6276 * or to avoid repeating get-magic / overloaded stringify */
6277 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6278 (IN_BYTES ? 0 : SvUTF8(pat)));
6280 Safefree(pRExC_state->code_blocks);
6281 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6284 /* ignore the utf8ness if the pattern is 0 length */
6285 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6286 RExC_uni_semantics = 0;
6287 RExC_contains_locale = 0;
6288 RExC_contains_i = 0;
6289 pRExC_state->runtime_code_qr = NULL;
6292 SV *dsv= sv_newmortal();
6293 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6294 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6295 PL_colors[4],PL_colors[5],s);
6299 /* we jump here if we upgrade the pattern to utf8 and have to
6302 if ((pm_flags & PMf_USE_RE_EVAL)
6303 /* this second condition covers the non-regex literal case,
6304 * i.e. $foo =~ '(?{})'. */
6305 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6307 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6309 /* return old regex if pattern hasn't changed */
6310 /* XXX: note in the below we have to check the flags as well as the
6313 * Things get a touch tricky as we have to compare the utf8 flag
6314 * independently from the compile flags. */
6318 && !!RX_UTF8(old_re) == !!RExC_utf8
6319 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6320 && RX_PRECOMP(old_re)
6321 && RX_PRELEN(old_re) == plen
6322 && memEQ(RX_PRECOMP(old_re), exp, plen)
6323 && !runtime_code /* with runtime code, always recompile */ )
6325 Safefree(pRExC_state->code_blocks);
6329 rx_flags = orig_rx_flags;
6331 if (rx_flags & PMf_FOLD) {
6332 RExC_contains_i = 1;
6334 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6336 /* Set to use unicode semantics if the pattern is in utf8 and has the
6337 * 'depends' charset specified, as it means unicode when utf8 */
6338 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6342 RExC_flags = rx_flags;
6343 RExC_pm_flags = pm_flags;
6346 if (TAINTING_get && TAINT_get)
6347 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6349 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6350 /* whoops, we have a non-utf8 pattern, whilst run-time code
6351 * got compiled as utf8. Try again with a utf8 pattern */
6352 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6353 pRExC_state->num_code_blocks);
6354 goto redo_first_pass;
6357 assert(!pRExC_state->runtime_code_qr);
6363 RExC_in_lookbehind = 0;
6364 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6366 RExC_override_recoding = 0;
6367 RExC_in_multi_char_class = 0;
6369 /* First pass: determine size, legality. */
6372 RExC_end = exp + plen;
6377 RExC_emit = (regnode *) &RExC_emit_dummy;
6378 RExC_whilem_seen = 0;
6379 RExC_open_parens = NULL;
6380 RExC_close_parens = NULL;
6382 RExC_paren_names = NULL;
6384 RExC_paren_name_list = NULL;
6386 RExC_recurse = NULL;
6387 RExC_study_chunk_recursed = NULL;
6388 RExC_study_chunk_recursed_bytes= 0;
6389 RExC_recurse_count = 0;
6390 pRExC_state->code_index = 0;
6392 #if 0 /* REGC() is (currently) a NOP at the first pass.
6393 * Clever compilers notice this and complain. --jhi */
6394 REGC((U8)REG_MAGIC, (char*)RExC_emit);
6397 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6399 RExC_lastparse=NULL;
6401 /* reg may croak on us, not giving us a chance to free
6402 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6403 need it to survive as long as the regexp (qr/(?{})/).
6404 We must check that code_blocksv is not already set, because we may
6405 have jumped back to restart the sizing pass. */
6406 if (pRExC_state->code_blocks && !code_blocksv) {
6407 code_blocksv = newSV_type(SVt_PV);
6408 SAVEFREESV(code_blocksv);
6409 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6410 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6412 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6413 /* It's possible to write a regexp in ascii that represents Unicode
6414 codepoints outside of the byte range, such as via \x{100}. If we
6415 detect such a sequence we have to convert the entire pattern to utf8
6416 and then recompile, as our sizing calculation will have been based
6417 on 1 byte == 1 character, but we will need to use utf8 to encode
6418 at least some part of the pattern, and therefore must convert the whole
6421 if (flags & RESTART_UTF8) {
6422 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6423 pRExC_state->num_code_blocks);
6424 goto redo_first_pass;
6426 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6429 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6432 PerlIO_printf(Perl_debug_log,
6433 "Required size %"IVdf" nodes\n"
6434 "Starting second pass (creation)\n",
6437 RExC_lastparse=NULL;
6440 /* The first pass could have found things that force Unicode semantics */
6441 if ((RExC_utf8 || RExC_uni_semantics)
6442 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6444 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6447 /* Small enough for pointer-storage convention?
6448 If extralen==0, this means that we will not need long jumps. */
6449 if (RExC_size >= 0x10000L && RExC_extralen)
6450 RExC_size += RExC_extralen;
6453 if (RExC_whilem_seen > 15)
6454 RExC_whilem_seen = 15;
6456 /* Allocate space and zero-initialize. Note, the two step process
6457 of zeroing when in debug mode, thus anything assigned has to
6458 happen after that */
6459 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6461 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6462 char, regexp_internal);
6463 if ( r == NULL || ri == NULL )
6464 FAIL("Regexp out of space");
6466 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6467 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6470 /* bulk initialize base fields with 0. */
6471 Zero(ri, sizeof(regexp_internal), char);
6474 /* non-zero initialization begins here */
6477 r->extflags = rx_flags;
6478 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6480 if (pm_flags & PMf_IS_QR) {
6481 ri->code_blocks = pRExC_state->code_blocks;
6482 ri->num_code_blocks = pRExC_state->num_code_blocks;
6487 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6488 if (pRExC_state->code_blocks[n].src_regex)
6489 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6490 SAVEFREEPV(pRExC_state->code_blocks);
6494 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6495 bool has_charset = (get_regex_charset(r->extflags)
6496 != REGEX_DEPENDS_CHARSET);
6498 /* The caret is output if there are any defaults: if not all the STD
6499 * flags are set, or if no character set specifier is needed */
6501 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6503 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6504 == REG_RUN_ON_COMMENT_SEEN);
6505 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6506 >> RXf_PMf_STD_PMMOD_SHIFT);
6507 const char *fptr = STD_PAT_MODS; /*"msix"*/
6509 /* Allocate for the worst case, which is all the std flags are turned
6510 * on. If more precision is desired, we could do a population count of
6511 * the flags set. This could be done with a small lookup table, or by
6512 * shifting, masking and adding, or even, when available, assembly
6513 * language for a machine-language population count.
6514 * We never output a minus, as all those are defaults, so are
6515 * covered by the caret */
6516 const STRLEN wraplen = plen + has_p + has_runon
6517 + has_default /* If needs a caret */
6519 /* If needs a character set specifier */
6520 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6521 + (sizeof(STD_PAT_MODS) - 1)
6522 + (sizeof("(?:)") - 1);
6524 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6525 r->xpv_len_u.xpvlenu_pv = p;
6527 SvFLAGS(rx) |= SVf_UTF8;
6530 /* If a default, cover it using the caret */
6532 *p++= DEFAULT_PAT_MOD;
6536 const char* const name = get_regex_charset_name(r->extflags, &len);
6537 Copy(name, p, len, char);
6541 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6544 while((ch = *fptr++)) {
6552 Copy(RExC_precomp, p, plen, char);
6553 assert ((RX_WRAPPED(rx) - p) < 16);
6554 r->pre_prefix = p - RX_WRAPPED(rx);
6560 SvCUR_set(rx, p - RX_WRAPPED(rx));
6564 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6566 /* setup various meta data about recursion, this all requires
6567 * RExC_npar to be correctly set, and a bit later on we clear it */
6568 if (RExC_seen & REG_RECURSE_SEEN) {
6569 Newxz(RExC_open_parens, RExC_npar,regnode *);
6570 SAVEFREEPV(RExC_open_parens);
6571 Newxz(RExC_close_parens,RExC_npar,regnode *);
6572 SAVEFREEPV(RExC_close_parens);
6574 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6575 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6576 * So its 1 if there are no parens. */
6577 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6578 ((RExC_npar & 0x07) != 0);
6579 Newx(RExC_study_chunk_recursed,
6580 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6581 SAVEFREEPV(RExC_study_chunk_recursed);
6584 /* Useful during FAIL. */
6585 #ifdef RE_TRACK_PATTERN_OFFSETS
6586 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6587 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6588 "%s %"UVuf" bytes for offset annotations.\n",
6589 ri->u.offsets ? "Got" : "Couldn't get",
6590 (UV)((2*RExC_size+1) * sizeof(U32))));
6592 SetProgLen(ri,RExC_size);
6597 /* Second pass: emit code. */
6598 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6599 RExC_pm_flags = pm_flags;
6601 RExC_end = exp + plen;
6604 RExC_emit_start = ri->program;
6605 RExC_emit = ri->program;
6606 RExC_emit_bound = ri->program + RExC_size + 1;
6607 pRExC_state->code_index = 0;
6609 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6610 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6612 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6614 /* XXXX To minimize changes to RE engine we always allocate
6615 3-units-long substrs field. */
6616 Newx(r->substrs, 1, struct reg_substr_data);
6617 if (RExC_recurse_count) {
6618 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6619 SAVEFREEPV(RExC_recurse);
6623 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6624 Zero(r->substrs, 1, struct reg_substr_data);
6625 if (RExC_study_chunk_recursed)
6626 Zero(RExC_study_chunk_recursed,
6627 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6629 #ifdef TRIE_STUDY_OPT
6631 StructCopy(&zero_scan_data, &data, scan_data_t);
6632 copyRExC_state = RExC_state;
6635 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6637 RExC_state = copyRExC_state;
6638 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6639 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6641 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6642 StructCopy(&zero_scan_data, &data, scan_data_t);
6645 StructCopy(&zero_scan_data, &data, scan_data_t);
6648 /* Dig out information for optimizations. */
6649 r->extflags = RExC_flags; /* was pm_op */
6650 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6653 SvUTF8_on(rx); /* Unicode in it? */
6654 ri->regstclass = NULL;
6655 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6656 r->intflags |= PREGf_NAUGHTY;
6657 scan = ri->program + 1; /* First BRANCH. */
6659 /* testing for BRANCH here tells us whether there is "must appear"
6660 data in the pattern. If there is then we can use it for optimisations */
6661 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6664 STRLEN longest_float_length, longest_fixed_length;
6665 regnode_ssc ch_class; /* pointed to by data */
6667 SSize_t last_close = 0; /* pointed to by data */
6668 regnode *first= scan;
6669 regnode *first_next= regnext(first);
6671 * Skip introductions and multiplicators >= 1
6672 * so that we can extract the 'meat' of the pattern that must
6673 * match in the large if() sequence following.
6674 * NOTE that EXACT is NOT covered here, as it is normally
6675 * picked up by the optimiser separately.
6677 * This is unfortunate as the optimiser isnt handling lookahead
6678 * properly currently.
6681 while ((OP(first) == OPEN && (sawopen = 1)) ||
6682 /* An OR of *one* alternative - should not happen now. */
6683 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6684 /* for now we can't handle lookbehind IFMATCH*/
6685 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6686 (OP(first) == PLUS) ||
6687 (OP(first) == MINMOD) ||
6688 /* An {n,m} with n>0 */
6689 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6690 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6693 * the only op that could be a regnode is PLUS, all the rest
6694 * will be regnode_1 or regnode_2.
6696 * (yves doesn't think this is true)
6698 if (OP(first) == PLUS)
6701 if (OP(first) == MINMOD)
6703 first += regarglen[OP(first)];
6705 first = NEXTOPER(first);
6706 first_next= regnext(first);
6709 /* Starting-point info. */
6711 DEBUG_PEEP("first:",first,0);
6712 /* Ignore EXACT as we deal with it later. */
6713 if (PL_regkind[OP(first)] == EXACT) {
6714 if (OP(first) == EXACT)
6715 NOOP; /* Empty, get anchored substr later. */
6717 ri->regstclass = first;
6720 else if (PL_regkind[OP(first)] == TRIE &&
6721 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6724 /* this can happen only on restudy */
6725 if ( OP(first) == TRIE ) {
6726 struct regnode_1 *trieop = (struct regnode_1 *)
6727 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6728 StructCopy(first,trieop,struct regnode_1);
6729 trie_op=(regnode *)trieop;
6731 struct regnode_charclass *trieop = (struct regnode_charclass *)
6732 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6733 StructCopy(first,trieop,struct regnode_charclass);
6734 trie_op=(regnode *)trieop;
6737 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6738 ri->regstclass = trie_op;
6741 else if (REGNODE_SIMPLE(OP(first)))
6742 ri->regstclass = first;
6743 else if (PL_regkind[OP(first)] == BOUND ||
6744 PL_regkind[OP(first)] == NBOUND)
6745 ri->regstclass = first;
6746 else if (PL_regkind[OP(first)] == BOL) {
6747 r->intflags |= (OP(first) == MBOL
6749 : (OP(first) == SBOL
6752 first = NEXTOPER(first);
6755 else if (OP(first) == GPOS) {
6756 r->intflags |= PREGf_ANCH_GPOS;
6757 first = NEXTOPER(first);
6760 else if ((!sawopen || !RExC_sawback) &&
6761 (OP(first) == STAR &&
6762 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6763 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
6765 /* turn .* into ^.* with an implied $*=1 */
6767 (OP(NEXTOPER(first)) == REG_ANY)
6770 r->intflags |= (type | PREGf_IMPLICIT);
6771 first = NEXTOPER(first);
6774 if (sawplus && !sawminmod && !sawlookahead
6775 && (!sawopen || !RExC_sawback)
6776 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6777 /* x+ must match at the 1st pos of run of x's */
6778 r->intflags |= PREGf_SKIP;
6780 /* Scan is after the zeroth branch, first is atomic matcher. */
6781 #ifdef TRIE_STUDY_OPT
6784 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6785 (IV)(first - scan + 1))
6789 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6790 (IV)(first - scan + 1))
6796 * If there's something expensive in the r.e., find the
6797 * longest literal string that must appear and make it the
6798 * regmust. Resolve ties in favor of later strings, since
6799 * the regstart check works with the beginning of the r.e.
6800 * and avoiding duplication strengthens checking. Not a
6801 * strong reason, but sufficient in the absence of others.
6802 * [Now we resolve ties in favor of the earlier string if
6803 * it happens that c_offset_min has been invalidated, since the
6804 * earlier string may buy us something the later one won't.]
6807 data.longest_fixed = newSVpvs("");
6808 data.longest_float = newSVpvs("");
6809 data.last_found = newSVpvs("");
6810 data.longest = &(data.longest_fixed);
6811 ENTER_with_name("study_chunk");
6812 SAVEFREESV(data.longest_fixed);
6813 SAVEFREESV(data.longest_float);
6814 SAVEFREESV(data.last_found);
6816 if (!ri->regstclass) {
6817 ssc_init(pRExC_state, &ch_class);
6818 data.start_class = &ch_class;
6819 stclass_flag = SCF_DO_STCLASS_AND;
6820 } else /* XXXX Check for BOUND? */
6822 data.last_closep = &last_close;
6825 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
6826 scan + RExC_size, /* Up to end */
6828 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6829 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6833 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6836 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6837 && data.last_start_min == 0 && data.last_end > 0
6838 && !RExC_seen_zerolen
6839 && !(RExC_seen & REG_VERBARG_SEEN)
6840 && !(RExC_seen & REG_GPOS_SEEN)
6842 r->extflags |= RXf_CHECK_ALL;
6844 scan_commit(pRExC_state, &data,&minlen,0);
6846 longest_float_length = CHR_SVLEN(data.longest_float);
6848 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6849 && data.offset_fixed == data.offset_float_min
6850 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6851 && S_setup_longest (aTHX_ pRExC_state,
6855 &(r->float_end_shift),
6856 data.lookbehind_float,
6857 data.offset_float_min,
6859 longest_float_length,
6860 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6861 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6863 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6864 r->float_max_offset = data.offset_float_max;
6865 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
6866 r->float_max_offset -= data.lookbehind_float;
6867 SvREFCNT_inc_simple_void_NN(data.longest_float);
6870 r->float_substr = r->float_utf8 = NULL;
6871 longest_float_length = 0;
6874 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6876 if (S_setup_longest (aTHX_ pRExC_state,
6878 &(r->anchored_utf8),
6879 &(r->anchored_substr),
6880 &(r->anchored_end_shift),
6881 data.lookbehind_fixed,
6884 longest_fixed_length,
6885 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6886 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6888 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6889 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6892 r->anchored_substr = r->anchored_utf8 = NULL;
6893 longest_fixed_length = 0;
6895 LEAVE_with_name("study_chunk");
6898 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6899 ri->regstclass = NULL;
6901 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6903 && ! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
6904 && !ssc_is_anything(data.start_class))
6906 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
6908 ssc_finalize(pRExC_state, data.start_class);
6910 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
6911 StructCopy(data.start_class,
6912 (regnode_ssc*)RExC_rxi->data->data[n],
6914 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6915 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6916 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6917 regprop(r, sv, (regnode*)data.start_class, NULL);
6918 PerlIO_printf(Perl_debug_log,
6919 "synthetic stclass \"%s\".\n",
6920 SvPVX_const(sv));});
6921 data.start_class = NULL;
6924 /* A temporary algorithm prefers floated substr to fixed one to dig
6926 if (longest_fixed_length > longest_float_length) {
6927 r->substrs->check_ix = 0;
6928 r->check_end_shift = r->anchored_end_shift;
6929 r->check_substr = r->anchored_substr;
6930 r->check_utf8 = r->anchored_utf8;
6931 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6932 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
6933 r->intflags |= PREGf_NOSCAN;
6936 r->substrs->check_ix = 1;
6937 r->check_end_shift = r->float_end_shift;
6938 r->check_substr = r->float_substr;
6939 r->check_utf8 = r->float_utf8;
6940 r->check_offset_min = r->float_min_offset;
6941 r->check_offset_max = r->float_max_offset;
6943 if ((r->check_substr || r->check_utf8) ) {
6944 r->extflags |= RXf_USE_INTUIT;
6945 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6946 r->extflags |= RXf_INTUIT_TAIL;
6948 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
6950 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6951 if ( (STRLEN)minlen < longest_float_length )
6952 minlen= longest_float_length;
6953 if ( (STRLEN)minlen < longest_fixed_length )
6954 minlen= longest_fixed_length;
6958 /* Several toplevels. Best we can is to set minlen. */
6960 regnode_ssc ch_class;
6961 SSize_t last_close = 0;
6963 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6965 scan = ri->program + 1;
6966 ssc_init(pRExC_state, &ch_class);
6967 data.start_class = &ch_class;
6968 data.last_closep = &last_close;
6971 minlen = study_chunk(pRExC_state,
6972 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
6973 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
6974 ? SCF_TRIE_DOING_RESTUDY
6978 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6980 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6981 = r->float_substr = r->float_utf8 = NULL;
6983 if (! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
6984 && ! ssc_is_anything(data.start_class))
6986 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
6988 ssc_finalize(pRExC_state, data.start_class);
6990 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
6991 StructCopy(data.start_class,
6992 (regnode_ssc*)RExC_rxi->data->data[n],
6994 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6995 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6996 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6997 regprop(r, sv, (regnode*)data.start_class, NULL);
6998 PerlIO_printf(Perl_debug_log,
6999 "synthetic stclass \"%s\".\n",
7000 SvPVX_const(sv));});
7001 data.start_class = NULL;
7005 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7006 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7007 r->maxlen = REG_INFTY;
7010 r->maxlen = RExC_maxlen;
7013 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7014 the "real" pattern. */
7016 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%ld\n",
7017 (IV)minlen, (IV)r->minlen, RExC_maxlen);
7019 r->minlenret = minlen;
7020 if (r->minlen < minlen)
7023 if (RExC_seen & REG_GPOS_SEEN)
7024 r->intflags |= PREGf_GPOS_SEEN;
7025 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7026 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7028 if (pRExC_state->num_code_blocks)
7029 r->extflags |= RXf_EVAL_SEEN;
7030 if (RExC_seen & REG_CANY_SEEN)
7031 r->intflags |= PREGf_CANY_SEEN;
7032 if (RExC_seen & REG_VERBARG_SEEN)
7034 r->intflags |= PREGf_VERBARG_SEEN;
7035 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7037 if (RExC_seen & REG_CUTGROUP_SEEN)
7038 r->intflags |= PREGf_CUTGROUP_SEEN;
7039 if (pm_flags & PMf_USE_RE_EVAL)
7040 r->intflags |= PREGf_USE_RE_EVAL;
7041 if (RExC_paren_names)
7042 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7044 RXp_PAREN_NAMES(r) = NULL;
7046 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7047 * so it can be used in pp.c */
7048 if (r->intflags & PREGf_ANCH)
7049 r->extflags |= RXf_IS_ANCHORED;
7053 /* this is used to identify "special" patterns that might result
7054 * in Perl NOT calling the regex engine and instead doing the match "itself",
7055 * particularly special cases in split//. By having the regex compiler
7056 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7057 * we avoid weird issues with equivalent patterns resulting in different behavior,
7058 * AND we allow non Perl engines to get the same optimizations by the setting the
7059 * flags appropriately - Yves */
7060 regnode *first = ri->program + 1;
7062 regnode *next = NEXTOPER(first);
7065 if (PL_regkind[fop] == NOTHING && nop == END)
7066 r->extflags |= RXf_NULL;
7067 else if (PL_regkind[fop] == BOL && nop == END)
7068 r->extflags |= RXf_START_ONLY;
7069 else if (fop == PLUS
7070 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7071 && OP(regnext(first)) == END)
7072 r->extflags |= RXf_WHITE;
7073 else if ( r->extflags & RXf_SPLIT
7075 && STR_LEN(first) == 1
7076 && *(STRING(first)) == ' '
7077 && OP(regnext(first)) == END )
7078 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7082 if (RExC_contains_locale) {
7083 RXp_EXTFLAGS(r) |= RXf_TAINTED_SEEN;
7087 if (RExC_paren_names) {
7088 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7089 ri->data->data[ri->name_list_idx]
7090 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7093 ri->name_list_idx = 0;
7095 if (RExC_recurse_count) {
7096 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7097 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7098 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7101 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7102 /* assume we don't need to swap parens around before we match */
7106 PerlIO_printf(Perl_debug_log,"Final program:\n");
7109 #ifdef RE_TRACK_PATTERN_OFFSETS
7110 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7111 const STRLEN len = ri->u.offsets[0];
7113 GET_RE_DEBUG_FLAGS_DECL;
7114 PerlIO_printf(Perl_debug_log,
7115 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7116 for (i = 1; i <= len; i++) {
7117 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7118 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7119 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7121 PerlIO_printf(Perl_debug_log, "\n");
7126 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7127 * by setting the regexp SV to readonly-only instead. If the
7128 * pattern's been recompiled, the USEDness should remain. */
7129 if (old_re && SvREADONLY(old_re))
7137 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7140 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7142 PERL_UNUSED_ARG(value);
7144 if (flags & RXapif_FETCH) {
7145 return reg_named_buff_fetch(rx, key, flags);
7146 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7147 Perl_croak_no_modify();
7149 } else if (flags & RXapif_EXISTS) {
7150 return reg_named_buff_exists(rx, key, flags)
7153 } else if (flags & RXapif_REGNAMES) {
7154 return reg_named_buff_all(rx, flags);
7155 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7156 return reg_named_buff_scalar(rx, flags);
7158 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7164 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7167 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7168 PERL_UNUSED_ARG(lastkey);
7170 if (flags & RXapif_FIRSTKEY)
7171 return reg_named_buff_firstkey(rx, flags);
7172 else if (flags & RXapif_NEXTKEY)
7173 return reg_named_buff_nextkey(rx, flags);
7175 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7182 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7185 AV *retarray = NULL;
7187 struct regexp *const rx = ReANY(r);
7189 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7191 if (flags & RXapif_ALL)
7194 if (rx && RXp_PAREN_NAMES(rx)) {
7195 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7198 SV* sv_dat=HeVAL(he_str);
7199 I32 *nums=(I32*)SvPVX(sv_dat);
7200 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7201 if ((I32)(rx->nparens) >= nums[i]
7202 && rx->offs[nums[i]].start != -1
7203 && rx->offs[nums[i]].end != -1)
7206 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7211 ret = newSVsv(&PL_sv_undef);
7214 av_push(retarray, ret);
7217 return newRV_noinc(MUTABLE_SV(retarray));
7224 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7227 struct regexp *const rx = ReANY(r);
7229 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7231 if (rx && RXp_PAREN_NAMES(rx)) {
7232 if (flags & RXapif_ALL) {
7233 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7235 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7237 SvREFCNT_dec_NN(sv);
7249 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7251 struct regexp *const rx = ReANY(r);
7253 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7255 if ( rx && RXp_PAREN_NAMES(rx) ) {
7256 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7258 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7265 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7267 struct regexp *const rx = ReANY(r);
7268 GET_RE_DEBUG_FLAGS_DECL;
7270 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7272 if (rx && RXp_PAREN_NAMES(rx)) {
7273 HV *hv = RXp_PAREN_NAMES(rx);
7275 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7278 SV* sv_dat = HeVAL(temphe);
7279 I32 *nums = (I32*)SvPVX(sv_dat);
7280 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7281 if ((I32)(rx->lastparen) >= nums[i] &&
7282 rx->offs[nums[i]].start != -1 &&
7283 rx->offs[nums[i]].end != -1)
7289 if (parno || flags & RXapif_ALL) {
7290 return newSVhek(HeKEY_hek(temphe));
7298 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7303 struct regexp *const rx = ReANY(r);
7305 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7307 if (rx && RXp_PAREN_NAMES(rx)) {
7308 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7309 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7310 } else if (flags & RXapif_ONE) {
7311 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7312 av = MUTABLE_AV(SvRV(ret));
7313 length = av_tindex(av);
7314 SvREFCNT_dec_NN(ret);
7315 return newSViv(length + 1);
7317 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7322 return &PL_sv_undef;
7326 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7328 struct regexp *const rx = ReANY(r);
7331 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7333 if (rx && RXp_PAREN_NAMES(rx)) {
7334 HV *hv= RXp_PAREN_NAMES(rx);
7336 (void)hv_iterinit(hv);
7337 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7340 SV* sv_dat = HeVAL(temphe);
7341 I32 *nums = (I32*)SvPVX(sv_dat);
7342 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7343 if ((I32)(rx->lastparen) >= nums[i] &&
7344 rx->offs[nums[i]].start != -1 &&
7345 rx->offs[nums[i]].end != -1)
7351 if (parno || flags & RXapif_ALL) {
7352 av_push(av, newSVhek(HeKEY_hek(temphe)));
7357 return newRV_noinc(MUTABLE_SV(av));
7361 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7364 struct regexp *const rx = ReANY(r);
7370 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7372 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7373 || n == RX_BUFF_IDX_CARET_FULLMATCH
7374 || n == RX_BUFF_IDX_CARET_POSTMATCH
7377 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7379 /* on something like
7382 * the KEEPCOPY is set on the PMOP rather than the regex */
7383 if (PL_curpm && r == PM_GETRE(PL_curpm))
7384 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7393 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7394 /* no need to distinguish between them any more */
7395 n = RX_BUFF_IDX_FULLMATCH;
7397 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7398 && rx->offs[0].start != -1)
7400 /* $`, ${^PREMATCH} */
7401 i = rx->offs[0].start;
7405 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7406 && rx->offs[0].end != -1)
7408 /* $', ${^POSTMATCH} */
7409 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7410 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7413 if ( 0 <= n && n <= (I32)rx->nparens &&
7414 (s1 = rx->offs[n].start) != -1 &&
7415 (t1 = rx->offs[n].end) != -1)
7417 /* $&, ${^MATCH}, $1 ... */
7419 s = rx->subbeg + s1 - rx->suboffset;
7424 assert(s >= rx->subbeg);
7425 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7427 #ifdef NO_TAINT_SUPPORT
7428 sv_setpvn(sv, s, i);
7430 const int oldtainted = TAINT_get;
7432 sv_setpvn(sv, s, i);
7433 TAINT_set(oldtainted);
7435 if ( (rx->intflags & PREGf_CANY_SEEN)
7436 ? (RXp_MATCH_UTF8(rx)
7437 && (!i || is_utf8_string((U8*)s, i)))
7438 : (RXp_MATCH_UTF8(rx)) )
7445 if (RXp_MATCH_TAINTED(rx)) {
7446 if (SvTYPE(sv) >= SVt_PVMG) {
7447 MAGIC* const mg = SvMAGIC(sv);
7450 SvMAGIC_set(sv, mg->mg_moremagic);
7452 if ((mgt = SvMAGIC(sv))) {
7453 mg->mg_moremagic = mgt;
7454 SvMAGIC_set(sv, mg);
7465 sv_setsv(sv,&PL_sv_undef);
7471 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7472 SV const * const value)
7474 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7476 PERL_UNUSED_ARG(rx);
7477 PERL_UNUSED_ARG(paren);
7478 PERL_UNUSED_ARG(value);
7481 Perl_croak_no_modify();
7485 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7488 struct regexp *const rx = ReANY(r);
7492 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7494 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7495 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7496 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7499 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7501 /* on something like
7504 * the KEEPCOPY is set on the PMOP rather than the regex */
7505 if (PL_curpm && r == PM_GETRE(PL_curpm))
7506 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7512 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7514 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7515 case RX_BUFF_IDX_PREMATCH: /* $` */
7516 if (rx->offs[0].start != -1) {
7517 i = rx->offs[0].start;
7526 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7527 case RX_BUFF_IDX_POSTMATCH: /* $' */
7528 if (rx->offs[0].end != -1) {
7529 i = rx->sublen - rx->offs[0].end;
7531 s1 = rx->offs[0].end;
7538 default: /* $& / ${^MATCH}, $1, $2, ... */
7539 if (paren <= (I32)rx->nparens &&
7540 (s1 = rx->offs[paren].start) != -1 &&
7541 (t1 = rx->offs[paren].end) != -1)
7547 if (ckWARN(WARN_UNINITIALIZED))
7548 report_uninit((const SV *)sv);
7553 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7554 const char * const s = rx->subbeg - rx->suboffset + s1;
7559 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7566 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7568 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7569 PERL_UNUSED_ARG(rx);
7573 return newSVpvs("Regexp");
7576 /* Scans the name of a named buffer from the pattern.
7577 * If flags is REG_RSN_RETURN_NULL returns null.
7578 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7579 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7580 * to the parsed name as looked up in the RExC_paren_names hash.
7581 * If there is an error throws a vFAIL().. type exception.
7584 #define REG_RSN_RETURN_NULL 0
7585 #define REG_RSN_RETURN_NAME 1
7586 #define REG_RSN_RETURN_DATA 2
7589 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7591 char *name_start = RExC_parse;
7593 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7595 assert (RExC_parse <= RExC_end);
7596 if (RExC_parse == RExC_end) NOOP;
7597 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7598 /* skip IDFIRST by using do...while */
7601 RExC_parse += UTF8SKIP(RExC_parse);
7602 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7606 } while (isWORDCHAR(*RExC_parse));
7608 RExC_parse++; /* so the <- from the vFAIL is after the offending
7610 vFAIL("Group name must start with a non-digit word character");
7614 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7615 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7616 if ( flags == REG_RSN_RETURN_NAME)
7618 else if (flags==REG_RSN_RETURN_DATA) {
7621 if ( ! sv_name ) /* should not happen*/
7622 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7623 if (RExC_paren_names)
7624 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7626 sv_dat = HeVAL(he_str);
7628 vFAIL("Reference to nonexistent named group");
7632 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7633 (unsigned long) flags);
7635 assert(0); /* NOT REACHED */
7640 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7641 int rem=(int)(RExC_end - RExC_parse); \
7650 if (RExC_lastparse!=RExC_parse) \
7651 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7654 iscut ? "..." : "<" \
7657 PerlIO_printf(Perl_debug_log,"%16s",""); \
7660 num = RExC_size + 1; \
7662 num=REG_NODE_NUM(RExC_emit); \
7663 if (RExC_lastnum!=num) \
7664 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7666 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7667 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7668 (int)((depth*2)), "", \
7672 RExC_lastparse=RExC_parse; \
7677 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7678 DEBUG_PARSE_MSG((funcname)); \
7679 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7681 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7682 DEBUG_PARSE_MSG((funcname)); \
7683 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7686 /* This section of code defines the inversion list object and its methods. The
7687 * interfaces are highly subject to change, so as much as possible is static to
7688 * this file. An inversion list is here implemented as a malloc'd C UV array
7689 * as an SVt_INVLIST scalar.
7691 * An inversion list for Unicode is an array of code points, sorted by ordinal
7692 * number. The zeroth element is the first code point in the list. The 1th
7693 * element is the first element beyond that not in the list. In other words,
7694 * the first range is
7695 * invlist[0]..(invlist[1]-1)
7696 * The other ranges follow. Thus every element whose index is divisible by two
7697 * marks the beginning of a range that is in the list, and every element not
7698 * divisible by two marks the beginning of a range not in the list. A single
7699 * element inversion list that contains the single code point N generally
7700 * consists of two elements
7703 * (The exception is when N is the highest representable value on the
7704 * machine, in which case the list containing just it would be a single
7705 * element, itself. By extension, if the last range in the list extends to
7706 * infinity, then the first element of that range will be in the inversion list
7707 * at a position that is divisible by two, and is the final element in the
7709 * Taking the complement (inverting) an inversion list is quite simple, if the
7710 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7711 * This implementation reserves an element at the beginning of each inversion
7712 * list to always contain 0; there is an additional flag in the header which
7713 * indicates if the list begins at the 0, or is offset to begin at the next
7716 * More about inversion lists can be found in "Unicode Demystified"
7717 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7718 * More will be coming when functionality is added later.
7720 * The inversion list data structure is currently implemented as an SV pointing
7721 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7722 * array of UV whose memory management is automatically handled by the existing
7723 * facilities for SV's.
7725 * Some of the methods should always be private to the implementation, and some
7726 * should eventually be made public */
7728 /* The header definitions are in F<inline_invlist.c> */
7730 PERL_STATIC_INLINE UV*
7731 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7733 /* Returns a pointer to the first element in the inversion list's array.
7734 * This is called upon initialization of an inversion list. Where the
7735 * array begins depends on whether the list has the code point U+0000 in it
7736 * or not. The other parameter tells it whether the code that follows this
7737 * call is about to put a 0 in the inversion list or not. The first
7738 * element is either the element reserved for 0, if TRUE, or the element
7739 * after it, if FALSE */
7741 bool* offset = get_invlist_offset_addr(invlist);
7742 UV* zero_addr = (UV *) SvPVX(invlist);
7744 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7747 assert(! _invlist_len(invlist));
7751 /* 1^1 = 0; 1^0 = 1 */
7752 *offset = 1 ^ will_have_0;
7753 return zero_addr + *offset;
7756 PERL_STATIC_INLINE UV*
7757 S_invlist_array(pTHX_ SV* const invlist)
7759 /* Returns the pointer to the inversion list's array. Every time the
7760 * length changes, this needs to be called in case malloc or realloc moved
7763 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7765 /* Must not be empty. If these fail, you probably didn't check for <len>
7766 * being non-zero before trying to get the array */
7767 assert(_invlist_len(invlist));
7769 /* The very first element always contains zero, The array begins either
7770 * there, or if the inversion list is offset, at the element after it.
7771 * The offset header field determines which; it contains 0 or 1 to indicate
7772 * how much additionally to add */
7773 assert(0 == *(SvPVX(invlist)));
7774 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7777 PERL_STATIC_INLINE void
7778 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7780 /* Sets the current number of elements stored in the inversion list.
7781 * Updates SvCUR correspondingly */
7783 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7785 assert(SvTYPE(invlist) == SVt_INVLIST);
7790 : TO_INTERNAL_SIZE(len + offset));
7791 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7794 PERL_STATIC_INLINE IV*
7795 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7797 /* Return the address of the IV that is reserved to hold the cached index
7800 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7802 assert(SvTYPE(invlist) == SVt_INVLIST);
7804 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7807 PERL_STATIC_INLINE IV
7808 S_invlist_previous_index(pTHX_ SV* const invlist)
7810 /* Returns cached index of previous search */
7812 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7814 return *get_invlist_previous_index_addr(invlist);
7817 PERL_STATIC_INLINE void
7818 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7820 /* Caches <index> for later retrieval */
7822 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7824 assert(index == 0 || index < (int) _invlist_len(invlist));
7826 *get_invlist_previous_index_addr(invlist) = index;
7829 PERL_STATIC_INLINE UV
7830 S_invlist_max(pTHX_ SV* const invlist)
7832 /* Returns the maximum number of elements storable in the inversion list's
7833 * array, without having to realloc() */
7835 PERL_ARGS_ASSERT_INVLIST_MAX;
7837 assert(SvTYPE(invlist) == SVt_INVLIST);
7839 /* Assumes worst case, in which the 0 element is not counted in the
7840 * inversion list, so subtracts 1 for that */
7841 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7842 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7843 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7846 #ifndef PERL_IN_XSUB_RE
7848 Perl__new_invlist(pTHX_ IV initial_size)
7851 /* Return a pointer to a newly constructed inversion list, with enough
7852 * space to store 'initial_size' elements. If that number is negative, a
7853 * system default is used instead */
7857 if (initial_size < 0) {
7861 /* Allocate the initial space */
7862 new_list = newSV_type(SVt_INVLIST);
7864 /* First 1 is in case the zero element isn't in the list; second 1 is for
7866 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7867 invlist_set_len(new_list, 0, 0);
7869 /* Force iterinit() to be used to get iteration to work */
7870 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7872 *get_invlist_previous_index_addr(new_list) = 0;
7878 Perl__new_invlist_C_array(pTHX_ const UV* const list)
7880 /* Return a pointer to a newly constructed inversion list, initialized to
7881 * point to <list>, which has to be in the exact correct inversion list
7882 * form, including internal fields. Thus this is a dangerous routine that
7883 * should not be used in the wrong hands. The passed in 'list' contains
7884 * several header fields at the beginning that are not part of the
7885 * inversion list body proper */
7887 const STRLEN length = (STRLEN) list[0];
7888 const UV version_id = list[1];
7889 const bool offset = cBOOL(list[2]);
7890 #define HEADER_LENGTH 3
7891 /* If any of the above changes in any way, you must change HEADER_LENGTH
7892 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7893 * perl -E 'say int(rand 2**31-1)'
7895 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7896 data structure type, so that one being
7897 passed in can be validated to be an
7898 inversion list of the correct vintage.
7901 SV* invlist = newSV_type(SVt_INVLIST);
7903 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7905 if (version_id != INVLIST_VERSION_ID) {
7906 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7909 /* The generated array passed in includes header elements that aren't part
7910 * of the list proper, so start it just after them */
7911 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7913 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7914 shouldn't touch it */
7916 *(get_invlist_offset_addr(invlist)) = offset;
7918 /* The 'length' passed to us is the physical number of elements in the
7919 * inversion list. But if there is an offset the logical number is one
7921 invlist_set_len(invlist, length - offset, offset);
7923 invlist_set_previous_index(invlist, 0);
7925 /* Initialize the iteration pointer. */
7926 invlist_iterfinish(invlist);
7928 SvREADONLY_on(invlist);
7932 #endif /* ifndef PERL_IN_XSUB_RE */
7935 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7937 /* Grow the maximum size of an inversion list */
7939 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7941 assert(SvTYPE(invlist) == SVt_INVLIST);
7943 /* Add one to account for the zero element at the beginning which may not
7944 * be counted by the calling parameters */
7945 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7948 PERL_STATIC_INLINE void
7949 S_invlist_trim(pTHX_ SV* const invlist)
7951 PERL_ARGS_ASSERT_INVLIST_TRIM;
7953 assert(SvTYPE(invlist) == SVt_INVLIST);
7955 /* Change the length of the inversion list to how many entries it currently
7957 SvPV_shrink_to_cur((SV *) invlist);
7961 S__append_range_to_invlist(pTHX_ SV* const invlist,
7962 const UV start, const UV end)
7964 /* Subject to change or removal. Append the range from 'start' to 'end' at
7965 * the end of the inversion list. The range must be above any existing
7969 UV max = invlist_max(invlist);
7970 UV len = _invlist_len(invlist);
7973 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7975 if (len == 0) { /* Empty lists must be initialized */
7976 offset = start != 0;
7977 array = _invlist_array_init(invlist, ! offset);
7980 /* Here, the existing list is non-empty. The current max entry in the
7981 * list is generally the first value not in the set, except when the
7982 * set extends to the end of permissible values, in which case it is
7983 * the first entry in that final set, and so this call is an attempt to
7984 * append out-of-order */
7986 UV final_element = len - 1;
7987 array = invlist_array(invlist);
7988 if (array[final_element] > start
7989 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7991 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",
7992 array[final_element], start,
7993 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7996 /* Here, it is a legal append. If the new range begins with the first
7997 * value not in the set, it is extending the set, so the new first
7998 * value not in the set is one greater than the newly extended range.
8000 offset = *get_invlist_offset_addr(invlist);
8001 if (array[final_element] == start) {
8002 if (end != UV_MAX) {
8003 array[final_element] = end + 1;
8006 /* But if the end is the maximum representable on the machine,
8007 * just let the range that this would extend to have no end */
8008 invlist_set_len(invlist, len - 1, offset);
8014 /* Here the new range doesn't extend any existing set. Add it */
8016 len += 2; /* Includes an element each for the start and end of range */
8018 /* If wll overflow the existing space, extend, which may cause the array to
8021 invlist_extend(invlist, len);
8023 /* Have to set len here to avoid assert failure in invlist_array() */
8024 invlist_set_len(invlist, len, offset);
8026 array = invlist_array(invlist);
8029 invlist_set_len(invlist, len, offset);
8032 /* The next item on the list starts the range, the one after that is
8033 * one past the new range. */
8034 array[len - 2] = start;
8035 if (end != UV_MAX) {
8036 array[len - 1] = end + 1;
8039 /* But if the end is the maximum representable on the machine, just let
8040 * the range have no end */
8041 invlist_set_len(invlist, len - 1, offset);
8045 #ifndef PERL_IN_XSUB_RE
8048 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
8050 /* Searches the inversion list for the entry that contains the input code
8051 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8052 * return value is the index into the list's array of the range that
8057 IV high = _invlist_len(invlist);
8058 const IV highest_element = high - 1;
8061 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8063 /* If list is empty, return failure. */
8068 /* (We can't get the array unless we know the list is non-empty) */
8069 array = invlist_array(invlist);
8071 mid = invlist_previous_index(invlist);
8072 assert(mid >=0 && mid <= highest_element);
8074 /* <mid> contains the cache of the result of the previous call to this
8075 * function (0 the first time). See if this call is for the same result,
8076 * or if it is for mid-1. This is under the theory that calls to this
8077 * function will often be for related code points that are near each other.
8078 * And benchmarks show that caching gives better results. We also test
8079 * here if the code point is within the bounds of the list. These tests
8080 * replace others that would have had to be made anyway to make sure that
8081 * the array bounds were not exceeded, and these give us extra information
8082 * at the same time */
8083 if (cp >= array[mid]) {
8084 if (cp >= array[highest_element]) {
8085 return highest_element;
8088 /* Here, array[mid] <= cp < array[highest_element]. This means that
8089 * the final element is not the answer, so can exclude it; it also
8090 * means that <mid> is not the final element, so can refer to 'mid + 1'
8092 if (cp < array[mid + 1]) {
8098 else { /* cp < aray[mid] */
8099 if (cp < array[0]) { /* Fail if outside the array */
8103 if (cp >= array[mid - 1]) {
8108 /* Binary search. What we are looking for is <i> such that
8109 * array[i] <= cp < array[i+1]
8110 * The loop below converges on the i+1. Note that there may not be an
8111 * (i+1)th element in the array, and things work nonetheless */
8112 while (low < high) {
8113 mid = (low + high) / 2;
8114 assert(mid <= highest_element);
8115 if (array[mid] <= cp) { /* cp >= array[mid] */
8118 /* We could do this extra test to exit the loop early.
8119 if (cp < array[low]) {
8124 else { /* cp < array[mid] */
8131 invlist_set_previous_index(invlist, high);
8136 Perl__invlist_populate_swatch(pTHX_ SV* const invlist,
8137 const UV start, const UV end, U8* swatch)
8139 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8140 * but is used when the swash has an inversion list. This makes this much
8141 * faster, as it uses a binary search instead of a linear one. This is
8142 * intimately tied to that function, and perhaps should be in utf8.c,
8143 * except it is intimately tied to inversion lists as well. It assumes
8144 * that <swatch> is all 0's on input */
8147 const IV len = _invlist_len(invlist);
8151 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8153 if (len == 0) { /* Empty inversion list */
8157 array = invlist_array(invlist);
8159 /* Find which element it is */
8160 i = _invlist_search(invlist, start);
8162 /* We populate from <start> to <end> */
8163 while (current < end) {
8166 /* The inversion list gives the results for every possible code point
8167 * after the first one in the list. Only those ranges whose index is
8168 * even are ones that the inversion list matches. For the odd ones,
8169 * and if the initial code point is not in the list, we have to skip
8170 * forward to the next element */
8171 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8173 if (i >= len) { /* Finished if beyond the end of the array */
8177 if (current >= end) { /* Finished if beyond the end of what we
8179 if (LIKELY(end < UV_MAX)) {
8183 /* We get here when the upper bound is the maximum
8184 * representable on the machine, and we are looking for just
8185 * that code point. Have to special case it */
8187 goto join_end_of_list;
8190 assert(current >= start);
8192 /* The current range ends one below the next one, except don't go past
8195 upper = (i < len && array[i] < end) ? array[i] : end;
8197 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8198 * for each code point in it */
8199 for (; current < upper; current++) {
8200 const STRLEN offset = (STRLEN)(current - start);
8201 swatch[offset >> 3] |= 1 << (offset & 7);
8206 /* Quit if at the end of the list */
8209 /* But first, have to deal with the highest possible code point on
8210 * the platform. The previous code assumes that <end> is one
8211 * beyond where we want to populate, but that is impossible at the
8212 * platform's infinity, so have to handle it specially */
8213 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8215 const STRLEN offset = (STRLEN)(end - start);
8216 swatch[offset >> 3] |= 1 << (offset & 7);
8221 /* Advance to the next range, which will be for code points not in the
8230 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8231 const bool complement_b, SV** output)
8233 /* Take the union of two inversion lists and point <output> to it. *output
8234 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8235 * the reference count to that list will be decremented if not already a
8236 * temporary (mortal); otherwise *output will be made correspondingly
8237 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8238 * second list is returned. If <complement_b> is TRUE, the union is taken
8239 * of the complement (inversion) of <b> instead of b itself.
8241 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8242 * Richard Gillam, published by Addison-Wesley, and explained at some
8243 * length there. The preface says to incorporate its examples into your
8244 * code at your own risk.
8246 * The algorithm is like a merge sort.
8248 * XXX A potential performance improvement is to keep track as we go along
8249 * if only one of the inputs contributes to the result, meaning the other
8250 * is a subset of that one. In that case, we can skip the final copy and
8251 * return the larger of the input lists, but then outside code might need
8252 * to keep track of whether to free the input list or not */
8254 const UV* array_a; /* a's array */
8256 UV len_a; /* length of a's array */
8259 SV* u; /* the resulting union */
8263 UV i_a = 0; /* current index into a's array */
8267 /* running count, as explained in the algorithm source book; items are
8268 * stopped accumulating and are output when the count changes to/from 0.
8269 * The count is incremented when we start a range that's in the set, and
8270 * decremented when we start a range that's not in the set. So its range
8271 * is 0 to 2. Only when the count is zero is something not in the set.
8275 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8278 /* If either one is empty, the union is the other one */
8279 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8280 bool make_temp = FALSE; /* Should we mortalize the result? */
8284 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8290 *output = invlist_clone(b);
8292 _invlist_invert(*output);
8294 } /* else *output already = b; */
8297 sv_2mortal(*output);
8301 else if ((len_b = _invlist_len(b)) == 0) {
8302 bool make_temp = FALSE;
8304 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8309 /* The complement of an empty list is a list that has everything in it,
8310 * so the union with <a> includes everything too */
8313 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8317 *output = _new_invlist(1);
8318 _append_range_to_invlist(*output, 0, UV_MAX);
8320 else if (*output != a) {
8321 *output = invlist_clone(a);
8323 /* else *output already = a; */
8326 sv_2mortal(*output);
8331 /* Here both lists exist and are non-empty */
8332 array_a = invlist_array(a);
8333 array_b = invlist_array(b);
8335 /* If are to take the union of 'a' with the complement of b, set it
8336 * up so are looking at b's complement. */
8339 /* To complement, we invert: if the first element is 0, remove it. To
8340 * do this, we just pretend the array starts one later */
8341 if (array_b[0] == 0) {
8347 /* But if the first element is not zero, we pretend the list starts
8348 * at the 0 that is always stored immediately before the array. */
8354 /* Size the union for the worst case: that the sets are completely
8356 u = _new_invlist(len_a + len_b);
8358 /* Will contain U+0000 if either component does */
8359 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8360 || (len_b > 0 && array_b[0] == 0));
8362 /* Go through each list item by item, stopping when exhausted one of
8364 while (i_a < len_a && i_b < len_b) {
8365 UV cp; /* The element to potentially add to the union's array */
8366 bool cp_in_set; /* is it in the the input list's set or not */
8368 /* We need to take one or the other of the two inputs for the union.
8369 * Since we are merging two sorted lists, we take the smaller of the
8370 * next items. In case of a tie, we take the one that is in its set
8371 * first. If we took one not in the set first, it would decrement the
8372 * count, possibly to 0 which would cause it to be output as ending the
8373 * range, and the next time through we would take the same number, and
8374 * output it again as beginning the next range. By doing it the
8375 * opposite way, there is no possibility that the count will be
8376 * momentarily decremented to 0, and thus the two adjoining ranges will
8377 * be seamlessly merged. (In a tie and both are in the set or both not
8378 * in the set, it doesn't matter which we take first.) */
8379 if (array_a[i_a] < array_b[i_b]
8380 || (array_a[i_a] == array_b[i_b]
8381 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8383 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8387 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8388 cp = array_b[i_b++];
8391 /* Here, have chosen which of the two inputs to look at. Only output
8392 * if the running count changes to/from 0, which marks the
8393 * beginning/end of a range in that's in the set */
8396 array_u[i_u++] = cp;
8403 array_u[i_u++] = cp;
8408 /* Here, we are finished going through at least one of the lists, which
8409 * means there is something remaining in at most one. We check if the list
8410 * that hasn't been exhausted is positioned such that we are in the middle
8411 * of a range in its set or not. (i_a and i_b point to the element beyond
8412 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8413 * is potentially more to output.
8414 * There are four cases:
8415 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8416 * in the union is entirely from the non-exhausted set.
8417 * 2) Both were in their sets, count is 2. Nothing further should
8418 * be output, as everything that remains will be in the exhausted
8419 * list's set, hence in the union; decrementing to 1 but not 0 insures
8421 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8422 * Nothing further should be output because the union includes
8423 * everything from the exhausted set. Not decrementing ensures that.
8424 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8425 * decrementing to 0 insures that we look at the remainder of the
8426 * non-exhausted set */
8427 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8428 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8433 /* The final length is what we've output so far, plus what else is about to
8434 * be output. (If 'count' is non-zero, then the input list we exhausted
8435 * has everything remaining up to the machine's limit in its set, and hence
8436 * in the union, so there will be no further output. */
8439 /* At most one of the subexpressions will be non-zero */
8440 len_u += (len_a - i_a) + (len_b - i_b);
8443 /* Set result to final length, which can change the pointer to array_u, so
8445 if (len_u != _invlist_len(u)) {
8446 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8448 array_u = invlist_array(u);
8451 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8452 * the other) ended with everything above it not in its set. That means
8453 * that the remaining part of the union is precisely the same as the
8454 * non-exhausted list, so can just copy it unchanged. (If both list were
8455 * exhausted at the same time, then the operations below will be both 0.)
8458 IV copy_count; /* At most one will have a non-zero copy count */
8459 if ((copy_count = len_a - i_a) > 0) {
8460 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8462 else if ((copy_count = len_b - i_b) > 0) {
8463 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8467 /* We may be removing a reference to one of the inputs. If so, the output
8468 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8469 * count decremented) */
8470 if (a == *output || b == *output) {
8471 assert(! invlist_is_iterating(*output));
8472 if ((SvTEMP(*output))) {
8476 SvREFCNT_dec_NN(*output);
8486 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8487 const bool complement_b, SV** i)
8489 /* Take the intersection of two inversion lists and point <i> to it. *i
8490 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8491 * the reference count to that list will be decremented if not already a
8492 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8493 * The first list, <a>, may be NULL, in which case an empty list is
8494 * returned. If <complement_b> is TRUE, the result will be the
8495 * intersection of <a> and the complement (or inversion) of <b> instead of
8498 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8499 * Richard Gillam, published by Addison-Wesley, and explained at some
8500 * length there. The preface says to incorporate its examples into your
8501 * code at your own risk. In fact, it had bugs
8503 * The algorithm is like a merge sort, and is essentially the same as the
8507 const UV* array_a; /* a's array */
8509 UV len_a; /* length of a's array */
8512 SV* r; /* the resulting intersection */
8516 UV i_a = 0; /* current index into a's array */
8520 /* running count, as explained in the algorithm source book; items are
8521 * stopped accumulating and are output when the count changes to/from 2.
8522 * The count is incremented when we start a range that's in the set, and
8523 * decremented when we start a range that's not in the set. So its range
8524 * is 0 to 2. Only when the count is 2 is something in the intersection.
8528 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8531 /* Special case if either one is empty */
8532 len_a = (a == NULL) ? 0 : _invlist_len(a);
8533 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8534 bool make_temp = FALSE;
8536 if (len_a != 0 && complement_b) {
8538 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8539 * be empty. Here, also we are using 'b's complement, which hence
8540 * must be every possible code point. Thus the intersection is
8544 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8549 *i = invlist_clone(a);
8551 /* else *i is already 'a' */
8559 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8560 * intersection must be empty */
8562 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8567 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8571 *i = _new_invlist(0);
8579 /* Here both lists exist and are non-empty */
8580 array_a = invlist_array(a);
8581 array_b = invlist_array(b);
8583 /* If are to take the intersection of 'a' with the complement of b, set it
8584 * up so are looking at b's complement. */
8587 /* To complement, we invert: if the first element is 0, remove it. To
8588 * do this, we just pretend the array starts one later */
8589 if (array_b[0] == 0) {
8595 /* But if the first element is not zero, we pretend the list starts
8596 * at the 0 that is always stored immediately before the array. */
8602 /* Size the intersection for the worst case: that the intersection ends up
8603 * fragmenting everything to be completely disjoint */
8604 r= _new_invlist(len_a + len_b);
8606 /* Will contain U+0000 iff both components do */
8607 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8608 && len_b > 0 && array_b[0] == 0);
8610 /* Go through each list item by item, stopping when exhausted one of
8612 while (i_a < len_a && i_b < len_b) {
8613 UV cp; /* The element to potentially add to the intersection's
8615 bool cp_in_set; /* Is it in the input list's set or not */
8617 /* We need to take one or the other of the two inputs for the
8618 * intersection. Since we are merging two sorted lists, we take the
8619 * smaller of the next items. In case of a tie, we take the one that
8620 * is not in its set first (a difference from the union algorithm). If
8621 * we took one in the set first, it would increment the count, possibly
8622 * to 2 which would cause it to be output as starting a range in the
8623 * intersection, and the next time through we would take that same
8624 * number, and output it again as ending the set. By doing it the
8625 * opposite of this, there is no possibility that the count will be
8626 * momentarily incremented to 2. (In a tie and both are in the set or
8627 * both not in the set, it doesn't matter which we take first.) */
8628 if (array_a[i_a] < array_b[i_b]
8629 || (array_a[i_a] == array_b[i_b]
8630 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8632 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8636 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8640 /* Here, have chosen which of the two inputs to look at. Only output
8641 * if the running count changes to/from 2, which marks the
8642 * beginning/end of a range that's in the intersection */
8646 array_r[i_r++] = cp;
8651 array_r[i_r++] = cp;
8657 /* Here, we are finished going through at least one of the lists, which
8658 * means there is something remaining in at most one. We check if the list
8659 * that has been exhausted is positioned such that we are in the middle
8660 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8661 * the ones we care about.) There are four cases:
8662 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8663 * nothing left in the intersection.
8664 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8665 * above 2. What should be output is exactly that which is in the
8666 * non-exhausted set, as everything it has is also in the intersection
8667 * set, and everything it doesn't have can't be in the intersection
8668 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8669 * gets incremented to 2. Like the previous case, the intersection is
8670 * everything that remains in the non-exhausted set.
8671 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8672 * remains 1. And the intersection has nothing more. */
8673 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8674 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8679 /* The final length is what we've output so far plus what else is in the
8680 * intersection. At most one of the subexpressions below will be non-zero
8684 len_r += (len_a - i_a) + (len_b - i_b);
8687 /* Set result to final length, which can change the pointer to array_r, so
8689 if (len_r != _invlist_len(r)) {
8690 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8692 array_r = invlist_array(r);
8695 /* Finish outputting any remaining */
8696 if (count >= 2) { /* At most one will have a non-zero copy count */
8698 if ((copy_count = len_a - i_a) > 0) {
8699 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8701 else if ((copy_count = len_b - i_b) > 0) {
8702 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8706 /* We may be removing a reference to one of the inputs. If so, the output
8707 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8708 * count decremented) */
8709 if (a == *i || b == *i) {
8710 assert(! invlist_is_iterating(*i));
8715 SvREFCNT_dec_NN(*i);
8725 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8727 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8728 * set. A pointer to the inversion list is returned. This may actually be
8729 * a new list, in which case the passed in one has been destroyed. The
8730 * passed in inversion list can be NULL, in which case a new one is created
8731 * with just the one range in it */
8736 if (invlist == NULL) {
8737 invlist = _new_invlist(2);
8741 len = _invlist_len(invlist);
8744 /* If comes after the final entry actually in the list, can just append it
8747 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8748 && start >= invlist_array(invlist)[len - 1]))
8750 _append_range_to_invlist(invlist, start, end);
8754 /* Here, can't just append things, create and return a new inversion list
8755 * which is the union of this range and the existing inversion list */
8756 range_invlist = _new_invlist(2);
8757 _append_range_to_invlist(range_invlist, start, end);
8759 _invlist_union(invlist, range_invlist, &invlist);
8761 /* The temporary can be freed */
8762 SvREFCNT_dec_NN(range_invlist);
8768 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
8769 UV** other_elements_ptr)
8771 /* Create and return an inversion list whose contents are to be populated
8772 * by the caller. The caller gives the number of elements (in 'size') and
8773 * the very first element ('element0'). This function will set
8774 * '*other_elements_ptr' to an array of UVs, where the remaining elements
8777 * Obviously there is some trust involved that the caller will properly
8778 * fill in the other elements of the array.
8780 * (The first element needs to be passed in, as the underlying code does
8781 * things differently depending on whether it is zero or non-zero) */
8783 SV* invlist = _new_invlist(size);
8786 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
8788 _append_range_to_invlist(invlist, element0, element0);
8789 offset = *get_invlist_offset_addr(invlist);
8791 invlist_set_len(invlist, size, offset);
8792 *other_elements_ptr = invlist_array(invlist) + 1;
8798 PERL_STATIC_INLINE SV*
8799 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8800 return _add_range_to_invlist(invlist, cp, cp);
8803 #ifndef PERL_IN_XSUB_RE
8805 Perl__invlist_invert(pTHX_ SV* const invlist)
8807 /* Complement the input inversion list. This adds a 0 if the list didn't
8808 * have a zero; removes it otherwise. As described above, the data
8809 * structure is set up so that this is very efficient */
8811 PERL_ARGS_ASSERT__INVLIST_INVERT;
8813 assert(! invlist_is_iterating(invlist));
8815 /* The inverse of matching nothing is matching everything */
8816 if (_invlist_len(invlist) == 0) {
8817 _append_range_to_invlist(invlist, 0, UV_MAX);
8821 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8826 PERL_STATIC_INLINE SV*
8827 S_invlist_clone(pTHX_ SV* const invlist)
8830 /* Return a new inversion list that is a copy of the input one, which is
8831 * unchanged. The new list will not be mortal even if the old one was. */
8833 /* Need to allocate extra space to accommodate Perl's addition of a
8834 * trailing NUL to SvPV's, since it thinks they are always strings */
8835 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8836 STRLEN physical_length = SvCUR(invlist);
8837 bool offset = *(get_invlist_offset_addr(invlist));
8839 PERL_ARGS_ASSERT_INVLIST_CLONE;
8841 *(get_invlist_offset_addr(new_invlist)) = offset;
8842 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8843 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8848 PERL_STATIC_INLINE STRLEN*
8849 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8851 /* Return the address of the UV that contains the current iteration
8854 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8856 assert(SvTYPE(invlist) == SVt_INVLIST);
8858 return &(((XINVLIST*) SvANY(invlist))->iterator);
8861 PERL_STATIC_INLINE void
8862 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8864 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8866 *get_invlist_iter_addr(invlist) = 0;
8869 PERL_STATIC_INLINE void
8870 S_invlist_iterfinish(pTHX_ SV* invlist)
8872 /* Terminate iterator for invlist. This is to catch development errors.
8873 * Any iteration that is interrupted before completed should call this
8874 * function. Functions that add code points anywhere else but to the end
8875 * of an inversion list assert that they are not in the middle of an
8876 * iteration. If they were, the addition would make the iteration
8877 * problematical: if the iteration hadn't reached the place where things
8878 * were being added, it would be ok */
8880 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8882 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8886 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8888 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8889 * This call sets in <*start> and <*end>, the next range in <invlist>.
8890 * Returns <TRUE> if successful and the next call will return the next
8891 * range; <FALSE> if was already at the end of the list. If the latter,
8892 * <*start> and <*end> are unchanged, and the next call to this function
8893 * will start over at the beginning of the list */
8895 STRLEN* pos = get_invlist_iter_addr(invlist);
8896 UV len = _invlist_len(invlist);
8899 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8902 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8906 array = invlist_array(invlist);
8908 *start = array[(*pos)++];
8914 *end = array[(*pos)++] - 1;
8920 PERL_STATIC_INLINE bool
8921 S_invlist_is_iterating(pTHX_ SV* const invlist)
8923 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8925 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8928 PERL_STATIC_INLINE UV
8929 S_invlist_highest(pTHX_ SV* const invlist)
8931 /* Returns the highest code point that matches an inversion list. This API
8932 * has an ambiguity, as it returns 0 under either the highest is actually
8933 * 0, or if the list is empty. If this distinction matters to you, check
8934 * for emptiness before calling this function */
8936 UV len = _invlist_len(invlist);
8939 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8945 array = invlist_array(invlist);
8947 /* The last element in the array in the inversion list always starts a
8948 * range that goes to infinity. That range may be for code points that are
8949 * matched in the inversion list, or it may be for ones that aren't
8950 * matched. In the latter case, the highest code point in the set is one
8951 * less than the beginning of this range; otherwise it is the final element
8952 * of this range: infinity */
8953 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8955 : array[len - 1] - 1;
8958 #ifndef PERL_IN_XSUB_RE
8960 Perl__invlist_contents(pTHX_ SV* const invlist)
8962 /* Get the contents of an inversion list into a string SV so that they can
8963 * be printed out. It uses the format traditionally done for debug tracing
8967 SV* output = newSVpvs("\n");
8969 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8971 assert(! invlist_is_iterating(invlist));
8973 invlist_iterinit(invlist);
8974 while (invlist_iternext(invlist, &start, &end)) {
8975 if (end == UV_MAX) {
8976 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8978 else if (end != start) {
8979 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8983 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8991 #ifndef PERL_IN_XSUB_RE
8993 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
8994 const char * const indent, SV* const invlist)
8996 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
8997 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
8998 * the string 'indent'. The output looks like this:
8999 [0] 0x000A .. 0x000D
9001 [4] 0x2028 .. 0x2029
9002 [6] 0x3104 .. INFINITY
9003 * This means that the first range of code points matched by the list are
9004 * 0xA through 0xD; the second range contains only the single code point
9005 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9006 * are used to define each range (except if the final range extends to
9007 * infinity, only a single element is needed). The array index of the
9008 * first element for the corresponding range is given in brackets. */
9013 PERL_ARGS_ASSERT__INVLIST_DUMP;
9015 if (invlist_is_iterating(invlist)) {
9016 Perl_dump_indent(aTHX_ level, file,
9017 "%sCan't dump inversion list because is in middle of iterating\n",
9022 invlist_iterinit(invlist);
9023 while (invlist_iternext(invlist, &start, &end)) {
9024 if (end == UV_MAX) {
9025 Perl_dump_indent(aTHX_ level, file,
9026 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9027 indent, (UV)count, start);
9029 else if (end != start) {
9030 Perl_dump_indent(aTHX_ level, file,
9031 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9032 indent, (UV)count, start, end);
9035 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9036 indent, (UV)count, start);
9043 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9045 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9047 /* Return a boolean as to if the two passed in inversion lists are
9048 * identical. The final argument, if TRUE, says to take the complement of
9049 * the second inversion list before doing the comparison */
9051 const UV* array_a = invlist_array(a);
9052 const UV* array_b = invlist_array(b);
9053 UV len_a = _invlist_len(a);
9054 UV len_b = _invlist_len(b);
9056 UV i = 0; /* current index into the arrays */
9057 bool retval = TRUE; /* Assume are identical until proven otherwise */
9059 PERL_ARGS_ASSERT__INVLISTEQ;
9061 /* If are to compare 'a' with the complement of b, set it
9062 * up so are looking at b's complement. */
9065 /* The complement of nothing is everything, so <a> would have to have
9066 * just one element, starting at zero (ending at infinity) */
9068 return (len_a == 1 && array_a[0] == 0);
9070 else if (array_b[0] == 0) {
9072 /* Otherwise, to complement, we invert. Here, the first element is
9073 * 0, just remove it. To do this, we just pretend the array starts
9081 /* But if the first element is not zero, we pretend the list starts
9082 * at the 0 that is always stored immediately before the array. */
9088 /* Make sure that the lengths are the same, as well as the final element
9089 * before looping through the remainder. (Thus we test the length, final,
9090 * and first elements right off the bat) */
9091 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9094 else for (i = 0; i < len_a - 1; i++) {
9095 if (array_a[i] != array_b[i]) {
9105 #undef HEADER_LENGTH
9106 #undef TO_INTERNAL_SIZE
9107 #undef FROM_INTERNAL_SIZE
9108 #undef INVLIST_VERSION_ID
9110 /* End of inversion list object */
9113 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9115 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9116 * constructs, and updates RExC_flags with them. On input, RExC_parse
9117 * should point to the first flag; it is updated on output to point to the
9118 * final ')' or ':'. There needs to be at least one flag, or this will
9121 /* for (?g), (?gc), and (?o) warnings; warning
9122 about (?c) will warn about (?g) -- japhy */
9124 #define WASTED_O 0x01
9125 #define WASTED_G 0x02
9126 #define WASTED_C 0x04
9127 #define WASTED_GC (WASTED_G|WASTED_C)
9128 I32 wastedflags = 0x00;
9129 U32 posflags = 0, negflags = 0;
9130 U32 *flagsp = &posflags;
9131 char has_charset_modifier = '\0';
9133 bool has_use_defaults = FALSE;
9134 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9136 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9138 /* '^' as an initial flag sets certain defaults */
9139 if (UCHARAT(RExC_parse) == '^') {
9141 has_use_defaults = TRUE;
9142 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9143 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9144 ? REGEX_UNICODE_CHARSET
9145 : REGEX_DEPENDS_CHARSET);
9148 cs = get_regex_charset(RExC_flags);
9149 if (cs == REGEX_DEPENDS_CHARSET
9150 && (RExC_utf8 || RExC_uni_semantics))
9152 cs = REGEX_UNICODE_CHARSET;
9155 while (*RExC_parse) {
9156 /* && strchr("iogcmsx", *RExC_parse) */
9157 /* (?g), (?gc) and (?o) are useless here
9158 and must be globally applied -- japhy */
9159 switch (*RExC_parse) {
9161 /* Code for the imsx flags */
9162 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
9164 case LOCALE_PAT_MOD:
9165 if (has_charset_modifier) {
9166 goto excess_modifier;
9168 else if (flagsp == &negflags) {
9171 cs = REGEX_LOCALE_CHARSET;
9172 has_charset_modifier = LOCALE_PAT_MOD;
9174 case UNICODE_PAT_MOD:
9175 if (has_charset_modifier) {
9176 goto excess_modifier;
9178 else if (flagsp == &negflags) {
9181 cs = REGEX_UNICODE_CHARSET;
9182 has_charset_modifier = UNICODE_PAT_MOD;
9184 case ASCII_RESTRICT_PAT_MOD:
9185 if (flagsp == &negflags) {
9188 if (has_charset_modifier) {
9189 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9190 goto excess_modifier;
9192 /* Doubled modifier implies more restricted */
9193 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9196 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9198 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9200 case DEPENDS_PAT_MOD:
9201 if (has_use_defaults) {
9202 goto fail_modifiers;
9204 else if (flagsp == &negflags) {
9207 else if (has_charset_modifier) {
9208 goto excess_modifier;
9211 /* The dual charset means unicode semantics if the
9212 * pattern (or target, not known until runtime) are
9213 * utf8, or something in the pattern indicates unicode
9215 cs = (RExC_utf8 || RExC_uni_semantics)
9216 ? REGEX_UNICODE_CHARSET
9217 : REGEX_DEPENDS_CHARSET;
9218 has_charset_modifier = DEPENDS_PAT_MOD;
9222 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9223 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9225 else if (has_charset_modifier == *(RExC_parse - 1)) {
9226 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9230 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9235 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9238 case ONCE_PAT_MOD: /* 'o' */
9239 case GLOBAL_PAT_MOD: /* 'g' */
9240 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9241 const I32 wflagbit = *RExC_parse == 'o'
9244 if (! (wastedflags & wflagbit) ) {
9245 wastedflags |= wflagbit;
9246 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9249 "Useless (%s%c) - %suse /%c modifier",
9250 flagsp == &negflags ? "?-" : "?",
9252 flagsp == &negflags ? "don't " : "",
9259 case CONTINUE_PAT_MOD: /* 'c' */
9260 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9261 if (! (wastedflags & WASTED_C) ) {
9262 wastedflags |= WASTED_GC;
9263 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9266 "Useless (%sc) - %suse /gc modifier",
9267 flagsp == &negflags ? "?-" : "?",
9268 flagsp == &negflags ? "don't " : ""
9273 case KEEPCOPY_PAT_MOD: /* 'p' */
9274 if (flagsp == &negflags) {
9276 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9278 *flagsp |= RXf_PMf_KEEPCOPY;
9282 /* A flag is a default iff it is following a minus, so
9283 * if there is a minus, it means will be trying to
9284 * re-specify a default which is an error */
9285 if (has_use_defaults || flagsp == &negflags) {
9286 goto fail_modifiers;
9289 wastedflags = 0; /* reset so (?g-c) warns twice */
9293 RExC_flags |= posflags;
9294 RExC_flags &= ~negflags;
9295 set_regex_charset(&RExC_flags, cs);
9296 if (RExC_flags & RXf_PMf_FOLD) {
9297 RExC_contains_i = 1;
9303 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9304 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9305 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9306 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9315 - reg - regular expression, i.e. main body or parenthesized thing
9317 * Caller must absorb opening parenthesis.
9319 * Combining parenthesis handling with the base level of regular expression
9320 * is a trifle forced, but the need to tie the tails of the branches to what
9321 * follows makes it hard to avoid.
9323 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9325 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9327 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9330 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9331 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9332 needs to be restarted.
9333 Otherwise would only return NULL if regbranch() returns NULL, which
9336 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9337 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9338 * 2 is like 1, but indicates that nextchar() has been called to advance
9339 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9340 * this flag alerts us to the need to check for that */
9343 regnode *ret; /* Will be the head of the group. */
9346 regnode *ender = NULL;
9349 U32 oregflags = RExC_flags;
9350 bool have_branch = 0;
9352 I32 freeze_paren = 0;
9353 I32 after_freeze = 0;
9355 char * parse_start = RExC_parse; /* MJD */
9356 char * const oregcomp_parse = RExC_parse;
9358 GET_RE_DEBUG_FLAGS_DECL;
9360 PERL_ARGS_ASSERT_REG;
9361 DEBUG_PARSE("reg ");
9363 *flagp = 0; /* Tentatively. */
9366 /* Make an OPEN node, if parenthesized. */
9369 /* Under /x, space and comments can be gobbled up between the '(' and
9370 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9371 * intervening space, as the sequence is a token, and a token should be
9373 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9375 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9376 char *start_verb = RExC_parse;
9377 STRLEN verb_len = 0;
9378 char *start_arg = NULL;
9379 unsigned char op = 0;
9381 int internal_argval = 0; /* internal_argval is only useful if
9384 if (has_intervening_patws && SIZE_ONLY) {
9385 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
9387 while ( *RExC_parse && *RExC_parse != ')' ) {
9388 if ( *RExC_parse == ':' ) {
9389 start_arg = RExC_parse + 1;
9395 verb_len = RExC_parse - start_verb;
9398 while ( *RExC_parse && *RExC_parse != ')' )
9400 if ( *RExC_parse != ')' )
9401 vFAIL("Unterminated verb pattern argument");
9402 if ( RExC_parse == start_arg )
9405 if ( *RExC_parse != ')' )
9406 vFAIL("Unterminated verb pattern");
9409 switch ( *start_verb ) {
9410 case 'A': /* (*ACCEPT) */
9411 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9413 internal_argval = RExC_nestroot;
9416 case 'C': /* (*COMMIT) */
9417 if ( memEQs(start_verb,verb_len,"COMMIT") )
9420 case 'F': /* (*FAIL) */
9421 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9426 case ':': /* (*:NAME) */
9427 case 'M': /* (*MARK:NAME) */
9428 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9433 case 'P': /* (*PRUNE) */
9434 if ( memEQs(start_verb,verb_len,"PRUNE") )
9437 case 'S': /* (*SKIP) */
9438 if ( memEQs(start_verb,verb_len,"SKIP") )
9441 case 'T': /* (*THEN) */
9442 /* [19:06] <TimToady> :: is then */
9443 if ( memEQs(start_verb,verb_len,"THEN") ) {
9445 RExC_seen |= REG_CUTGROUP_SEEN;
9450 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9452 "Unknown verb pattern '%"UTF8f"'",
9453 UTF8fARG(UTF, verb_len, start_verb));
9456 if ( start_arg && internal_argval ) {
9457 vFAIL3("Verb pattern '%.*s' may not have an argument",
9458 verb_len, start_verb);
9459 } else if ( argok < 0 && !start_arg ) {
9460 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9461 verb_len, start_verb);
9463 ret = reganode(pRExC_state, op, internal_argval);
9464 if ( ! internal_argval && ! SIZE_ONLY ) {
9466 SV *sv = newSVpvn( start_arg,
9467 RExC_parse - start_arg);
9468 ARG(ret) = add_data( pRExC_state,
9470 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9477 if (!internal_argval)
9478 RExC_seen |= REG_VERBARG_SEEN;
9479 } else if ( start_arg ) {
9480 vFAIL3("Verb pattern '%.*s' may not have an argument",
9481 verb_len, start_verb);
9483 ret = reg_node(pRExC_state, op);
9485 nextchar(pRExC_state);
9488 else if (*RExC_parse == '?') { /* (?...) */
9489 bool is_logical = 0;
9490 const char * const seqstart = RExC_parse;
9491 if (has_intervening_patws && SIZE_ONLY) {
9492 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
9496 paren = *RExC_parse++;
9497 ret = NULL; /* For look-ahead/behind. */
9500 case 'P': /* (?P...) variants for those used to PCRE/Python */
9501 paren = *RExC_parse++;
9502 if ( paren == '<') /* (?P<...>) named capture */
9504 else if (paren == '>') { /* (?P>name) named recursion */
9505 goto named_recursion;
9507 else if (paren == '=') { /* (?P=...) named backref */
9508 /* this pretty much dupes the code for \k<NAME> in
9509 * regatom(), if you change this make sure you change that
9511 char* name_start = RExC_parse;
9513 SV *sv_dat = reg_scan_name(pRExC_state,
9514 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9515 if (RExC_parse == name_start || *RExC_parse != ')')
9516 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9517 vFAIL2("Sequence %.3s... not terminated",parse_start);
9520 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9521 RExC_rxi->data->data[num]=(void*)sv_dat;
9522 SvREFCNT_inc_simple_void(sv_dat);
9525 ret = reganode(pRExC_state,
9528 : (ASCII_FOLD_RESTRICTED)
9530 : (AT_LEAST_UNI_SEMANTICS)
9538 Set_Node_Offset(ret, parse_start+1);
9539 Set_Node_Cur_Length(ret, parse_start);
9541 nextchar(pRExC_state);
9545 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9546 vFAIL3("Sequence (%.*s...) not recognized",
9547 RExC_parse-seqstart, seqstart);
9549 case '<': /* (?<...) */
9550 if (*RExC_parse == '!')
9552 else if (*RExC_parse != '=')
9558 case '\'': /* (?'...') */
9559 name_start= RExC_parse;
9560 svname = reg_scan_name(pRExC_state,
9561 SIZE_ONLY /* reverse test from the others */
9562 ? REG_RSN_RETURN_NAME
9563 : REG_RSN_RETURN_NULL);
9564 if (RExC_parse == name_start || *RExC_parse != paren)
9565 vFAIL2("Sequence (?%c... not terminated",
9566 paren=='>' ? '<' : paren);
9570 if (!svname) /* shouldn't happen */
9572 "panic: reg_scan_name returned NULL");
9573 if (!RExC_paren_names) {
9574 RExC_paren_names= newHV();
9575 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9577 RExC_paren_name_list= newAV();
9578 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
9581 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
9583 sv_dat = HeVAL(he_str);
9585 /* croak baby croak */
9587 "panic: paren_name hash element allocation failed");
9588 } else if ( SvPOK(sv_dat) ) {
9589 /* (?|...) can mean we have dupes so scan to check
9590 its already been stored. Maybe a flag indicating
9591 we are inside such a construct would be useful,
9592 but the arrays are likely to be quite small, so
9593 for now we punt -- dmq */
9594 IV count = SvIV(sv_dat);
9595 I32 *pv = (I32*)SvPVX(sv_dat);
9597 for ( i = 0 ; i < count ; i++ ) {
9598 if ( pv[i] == RExC_npar ) {
9604 pv = (I32*)SvGROW(sv_dat,
9605 SvCUR(sv_dat) + sizeof(I32)+1);
9606 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
9607 pv[count] = RExC_npar;
9608 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
9611 (void)SvUPGRADE(sv_dat,SVt_PVNV);
9612 sv_setpvn(sv_dat, (char *)&(RExC_npar),
9615 SvIV_set(sv_dat, 1);
9618 /* Yes this does cause a memory leak in debugging Perls
9620 if (!av_store(RExC_paren_name_list,
9621 RExC_npar, SvREFCNT_inc(svname)))
9622 SvREFCNT_dec_NN(svname);
9625 /*sv_dump(sv_dat);*/
9627 nextchar(pRExC_state);
9629 goto capturing_parens;
9631 RExC_seen |= REG_LOOKBEHIND_SEEN;
9632 RExC_in_lookbehind++;
9634 case '=': /* (?=...) */
9635 RExC_seen_zerolen++;
9637 case '!': /* (?!...) */
9638 RExC_seen_zerolen++;
9639 if (*RExC_parse == ')') {
9640 ret=reg_node(pRExC_state, OPFAIL);
9641 nextchar(pRExC_state);
9645 case '|': /* (?|...) */
9646 /* branch reset, behave like a (?:...) except that
9647 buffers in alternations share the same numbers */
9649 after_freeze = freeze_paren = RExC_npar;
9651 case ':': /* (?:...) */
9652 case '>': /* (?>...) */
9654 case '$': /* (?$...) */
9655 case '@': /* (?@...) */
9656 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
9658 case '#': /* (?#...) */
9659 /* XXX As soon as we disallow separating the '?' and '*' (by
9660 * spaces or (?#...) comment), it is believed that this case
9661 * will be unreachable and can be removed. See
9663 while (*RExC_parse && *RExC_parse != ')')
9665 if (*RExC_parse != ')')
9666 FAIL("Sequence (?#... not terminated");
9667 nextchar(pRExC_state);
9670 case '0' : /* (?0) */
9671 case 'R' : /* (?R) */
9672 if (*RExC_parse != ')')
9673 FAIL("Sequence (?R) not terminated");
9674 ret = reg_node(pRExC_state, GOSTART);
9675 RExC_seen |= REG_GOSTART_SEEN;
9676 *flagp |= POSTPONED;
9677 nextchar(pRExC_state);
9680 { /* named and numeric backreferences */
9682 case '&': /* (?&NAME) */
9683 parse_start = RExC_parse - 1;
9686 SV *sv_dat = reg_scan_name(pRExC_state,
9687 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9688 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9690 if (RExC_parse == RExC_end || *RExC_parse != ')')
9691 vFAIL("Sequence (?&... not terminated");
9692 goto gen_recurse_regop;
9693 assert(0); /* NOT REACHED */
9695 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9697 vFAIL("Illegal pattern");
9699 goto parse_recursion;
9701 case '-': /* (?-1) */
9702 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9703 RExC_parse--; /* rewind to let it be handled later */
9707 case '1': case '2': case '3': case '4': /* (?1) */
9708 case '5': case '6': case '7': case '8': case '9':
9711 num = atoi(RExC_parse);
9712 parse_start = RExC_parse - 1; /* MJD */
9713 if (*RExC_parse == '-')
9715 while (isDIGIT(*RExC_parse))
9717 if (*RExC_parse!=')')
9718 vFAIL("Expecting close bracket");
9721 if ( paren == '-' ) {
9723 Diagram of capture buffer numbering.
9724 Top line is the normal capture buffer numbers
9725 Bottom line is the negative indexing as from
9729 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9733 num = RExC_npar + num;
9736 vFAIL("Reference to nonexistent group");
9738 } else if ( paren == '+' ) {
9739 num = RExC_npar + num - 1;
9742 ret = reganode(pRExC_state, GOSUB, num);
9744 if (num > (I32)RExC_rx->nparens) {
9746 vFAIL("Reference to nonexistent group");
9748 ARG2L_SET( ret, RExC_recurse_count++);
9750 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9751 "Recurse #%"UVuf" to %"IVdf"\n",
9752 (UV)ARG(ret), (IV)ARG2L(ret)));
9756 RExC_seen |= REG_RECURSE_SEEN;
9757 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9758 Set_Node_Offset(ret, parse_start); /* MJD */
9760 *flagp |= POSTPONED;
9761 nextchar(pRExC_state);
9763 } /* named and numeric backreferences */
9764 assert(0); /* NOT REACHED */
9766 case '?': /* (??...) */
9768 if (*RExC_parse != '{') {
9770 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9772 "Sequence (%"UTF8f"...) not recognized",
9773 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9776 *flagp |= POSTPONED;
9777 paren = *RExC_parse++;
9779 case '{': /* (?{...}) */
9782 struct reg_code_block *cb;
9784 RExC_seen_zerolen++;
9786 if ( !pRExC_state->num_code_blocks
9787 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9788 || pRExC_state->code_blocks[pRExC_state->code_index].start
9789 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9792 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9793 FAIL("panic: Sequence (?{...}): no code block found\n");
9794 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9796 /* this is a pre-compiled code block (?{...}) */
9797 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9798 RExC_parse = RExC_start + cb->end;
9801 if (cb->src_regex) {
9802 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
9803 RExC_rxi->data->data[n] =
9804 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9805 RExC_rxi->data->data[n+1] = (void*)o;
9808 n = add_data(pRExC_state,
9809 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
9810 RExC_rxi->data->data[n] = (void*)o;
9813 pRExC_state->code_index++;
9814 nextchar(pRExC_state);
9818 ret = reg_node(pRExC_state, LOGICAL);
9819 eval = reganode(pRExC_state, EVAL, n);
9822 /* for later propagation into (??{}) return value */
9823 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9825 REGTAIL(pRExC_state, ret, eval);
9826 /* deal with the length of this later - MJD */
9829 ret = reganode(pRExC_state, EVAL, n);
9830 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9831 Set_Node_Offset(ret, parse_start);
9834 case '(': /* (?(?{...})...) and (?(?=...)...) */
9837 if (RExC_parse[0] == '?') { /* (?(?...)) */
9838 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9839 || RExC_parse[1] == '<'
9840 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9844 ret = reg_node(pRExC_state, LOGICAL);
9848 tail = reg(pRExC_state, 1, &flag, depth+1);
9849 if (flag & RESTART_UTF8) {
9850 *flagp = RESTART_UTF8;
9853 REGTAIL(pRExC_state, ret, tail);
9857 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9858 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9860 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9861 char *name_start= RExC_parse++;
9863 SV *sv_dat=reg_scan_name(pRExC_state,
9864 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9865 if (RExC_parse == name_start || *RExC_parse != ch)
9866 vFAIL2("Sequence (?(%c... not terminated",
9867 (ch == '>' ? '<' : ch));
9870 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9871 RExC_rxi->data->data[num]=(void*)sv_dat;
9872 SvREFCNT_inc_simple_void(sv_dat);
9874 ret = reganode(pRExC_state,NGROUPP,num);
9875 goto insert_if_check_paren;
9877 else if (RExC_parse[0] == 'D' &&
9878 RExC_parse[1] == 'E' &&
9879 RExC_parse[2] == 'F' &&
9880 RExC_parse[3] == 'I' &&
9881 RExC_parse[4] == 'N' &&
9882 RExC_parse[5] == 'E')
9884 ret = reganode(pRExC_state,DEFINEP,0);
9887 goto insert_if_check_paren;
9889 else if (RExC_parse[0] == 'R') {
9892 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9893 parno = atoi(RExC_parse++);
9894 while (isDIGIT(*RExC_parse))
9896 } else if (RExC_parse[0] == '&') {
9899 sv_dat = reg_scan_name(pRExC_state,
9901 ? REG_RSN_RETURN_NULL
9902 : REG_RSN_RETURN_DATA);
9903 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9905 ret = reganode(pRExC_state,INSUBP,parno);
9906 goto insert_if_check_paren;
9908 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9912 parno = atoi(RExC_parse++);
9914 while (isDIGIT(*RExC_parse))
9916 ret = reganode(pRExC_state, GROUPP, parno);
9918 insert_if_check_paren:
9919 if (*(tmp = nextchar(pRExC_state)) != ')') {
9920 /* nextchar also skips comments, so undo its work
9921 * and skip over the the next character.
9924 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9925 vFAIL("Switch condition not recognized");
9928 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9929 br = regbranch(pRExC_state, &flags, 1,depth+1);
9931 if (flags & RESTART_UTF8) {
9932 *flagp = RESTART_UTF8;
9935 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9938 REGTAIL(pRExC_state, br, reganode(pRExC_state,
9940 c = *nextchar(pRExC_state);
9945 vFAIL("(?(DEFINE)....) does not allow branches");
9947 /* Fake one for optimizer. */
9948 lastbr = reganode(pRExC_state, IFTHEN, 0);
9950 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9951 if (flags & RESTART_UTF8) {
9952 *flagp = RESTART_UTF8;
9955 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9958 REGTAIL(pRExC_state, ret, lastbr);
9961 c = *nextchar(pRExC_state);
9966 vFAIL("Switch (?(condition)... contains too many branches");
9967 ender = reg_node(pRExC_state, TAIL);
9968 REGTAIL(pRExC_state, br, ender);
9970 REGTAIL(pRExC_state, lastbr, ender);
9971 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9974 REGTAIL(pRExC_state, ret, ender);
9975 RExC_size++; /* XXX WHY do we need this?!!
9976 For large programs it seems to be required
9977 but I can't figure out why. -- dmq*/
9981 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9982 vFAIL("Unknown switch condition (?(...))");
9985 case '[': /* (?[ ... ]) */
9986 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9989 RExC_parse--; /* for vFAIL to print correctly */
9990 vFAIL("Sequence (? incomplete");
9992 default: /* e.g., (?i) */
9995 parse_lparen_question_flags(pRExC_state);
9996 if (UCHARAT(RExC_parse) != ':') {
9997 nextchar(pRExC_state);
10002 nextchar(pRExC_state);
10012 ret = reganode(pRExC_state, OPEN, parno);
10014 if (!RExC_nestroot)
10015 RExC_nestroot = parno;
10016 if (RExC_seen & REG_RECURSE_SEEN
10017 && !RExC_open_parens[parno-1])
10019 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10020 "Setting open paren #%"IVdf" to %d\n",
10021 (IV)parno, REG_NODE_NUM(ret)));
10022 RExC_open_parens[parno-1]= ret;
10025 Set_Node_Length(ret, 1); /* MJD */
10026 Set_Node_Offset(ret, RExC_parse); /* MJD */
10034 /* Pick up the branches, linking them together. */
10035 parse_start = RExC_parse; /* MJD */
10036 br = regbranch(pRExC_state, &flags, 1,depth+1);
10038 /* branch_len = (paren != 0); */
10041 if (flags & RESTART_UTF8) {
10042 *flagp = RESTART_UTF8;
10045 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10047 if (*RExC_parse == '|') {
10048 if (!SIZE_ONLY && RExC_extralen) {
10049 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10052 reginsert(pRExC_state, BRANCH, br, depth+1);
10053 Set_Node_Length(br, paren != 0);
10054 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10058 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10060 else if (paren == ':') {
10061 *flagp |= flags&SIMPLE;
10063 if (is_open) { /* Starts with OPEN. */
10064 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10066 else if (paren != '?') /* Not Conditional */
10068 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10070 while (*RExC_parse == '|') {
10071 if (!SIZE_ONLY && RExC_extralen) {
10072 ender = reganode(pRExC_state, LONGJMP,0);
10074 /* Append to the previous. */
10075 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10078 RExC_extralen += 2; /* Account for LONGJMP. */
10079 nextchar(pRExC_state);
10080 if (freeze_paren) {
10081 if (RExC_npar > after_freeze)
10082 after_freeze = RExC_npar;
10083 RExC_npar = freeze_paren;
10085 br = regbranch(pRExC_state, &flags, 0, depth+1);
10088 if (flags & RESTART_UTF8) {
10089 *flagp = RESTART_UTF8;
10092 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10094 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10096 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10099 if (have_branch || paren != ':') {
10100 /* Make a closing node, and hook it on the end. */
10103 ender = reg_node(pRExC_state, TAIL);
10106 ender = reganode(pRExC_state, CLOSE, parno);
10107 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10108 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10109 "Setting close paren #%"IVdf" to %d\n",
10110 (IV)parno, REG_NODE_NUM(ender)));
10111 RExC_close_parens[parno-1]= ender;
10112 if (RExC_nestroot == parno)
10115 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10116 Set_Node_Length(ender,1); /* MJD */
10122 *flagp &= ~HASWIDTH;
10125 ender = reg_node(pRExC_state, SUCCEED);
10128 ender = reg_node(pRExC_state, END);
10130 assert(!RExC_opend); /* there can only be one! */
10131 RExC_opend = ender;
10135 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10136 SV * const mysv_val1=sv_newmortal();
10137 SV * const mysv_val2=sv_newmortal();
10138 DEBUG_PARSE_MSG("lsbr");
10139 regprop(RExC_rx, mysv_val1, lastbr, NULL);
10140 regprop(RExC_rx, mysv_val2, ender, NULL);
10141 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10142 SvPV_nolen_const(mysv_val1),
10143 (IV)REG_NODE_NUM(lastbr),
10144 SvPV_nolen_const(mysv_val2),
10145 (IV)REG_NODE_NUM(ender),
10146 (IV)(ender - lastbr)
10149 REGTAIL(pRExC_state, lastbr, ender);
10151 if (have_branch && !SIZE_ONLY) {
10152 char is_nothing= 1;
10154 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10156 /* Hook the tails of the branches to the closing node. */
10157 for (br = ret; br; br = regnext(br)) {
10158 const U8 op = PL_regkind[OP(br)];
10159 if (op == BRANCH) {
10160 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10161 if ( OP(NEXTOPER(br)) != NOTHING
10162 || regnext(NEXTOPER(br)) != ender)
10165 else if (op == BRANCHJ) {
10166 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10167 /* for now we always disable this optimisation * /
10168 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10169 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10175 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10176 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10177 SV * const mysv_val1=sv_newmortal();
10178 SV * const mysv_val2=sv_newmortal();
10179 DEBUG_PARSE_MSG("NADA");
10180 regprop(RExC_rx, mysv_val1, ret, NULL);
10181 regprop(RExC_rx, mysv_val2, ender, NULL);
10182 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10183 SvPV_nolen_const(mysv_val1),
10184 (IV)REG_NODE_NUM(ret),
10185 SvPV_nolen_const(mysv_val2),
10186 (IV)REG_NODE_NUM(ender),
10191 if (OP(ender) == TAIL) {
10196 for ( opt= br + 1; opt < ender ; opt++ )
10197 OP(opt)= OPTIMIZED;
10198 NEXT_OFF(br)= ender - br;
10206 static const char parens[] = "=!<,>";
10208 if (paren && (p = strchr(parens, paren))) {
10209 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10210 int flag = (p - parens) > 1;
10213 node = SUSPEND, flag = 0;
10214 reginsert(pRExC_state, node,ret, depth+1);
10215 Set_Node_Cur_Length(ret, parse_start);
10216 Set_Node_Offset(ret, parse_start + 1);
10218 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10222 /* Check for proper termination. */
10224 /* restore original flags, but keep (?p) */
10225 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10226 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10227 RExC_parse = oregcomp_parse;
10228 vFAIL("Unmatched (");
10231 else if (!paren && RExC_parse < RExC_end) {
10232 if (*RExC_parse == ')') {
10234 vFAIL("Unmatched )");
10237 FAIL("Junk on end of regexp"); /* "Can't happen". */
10238 assert(0); /* NOTREACHED */
10241 if (RExC_in_lookbehind) {
10242 RExC_in_lookbehind--;
10244 if (after_freeze > RExC_npar)
10245 RExC_npar = after_freeze;
10250 - regbranch - one alternative of an | operator
10252 * Implements the concatenation operator.
10254 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10258 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10262 regnode *chain = NULL;
10264 I32 flags = 0, c = 0;
10265 GET_RE_DEBUG_FLAGS_DECL;
10267 PERL_ARGS_ASSERT_REGBRANCH;
10269 DEBUG_PARSE("brnc");
10274 if (!SIZE_ONLY && RExC_extralen)
10275 ret = reganode(pRExC_state, BRANCHJ,0);
10277 ret = reg_node(pRExC_state, BRANCH);
10278 Set_Node_Length(ret, 1);
10282 if (!first && SIZE_ONLY)
10283 RExC_extralen += 1; /* BRANCHJ */
10285 *flagp = WORST; /* Tentatively. */
10288 nextchar(pRExC_state);
10289 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10290 flags &= ~TRYAGAIN;
10291 latest = regpiece(pRExC_state, &flags,depth+1);
10292 if (latest == NULL) {
10293 if (flags & TRYAGAIN)
10295 if (flags & RESTART_UTF8) {
10296 *flagp = RESTART_UTF8;
10299 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10301 else if (ret == NULL)
10303 *flagp |= flags&(HASWIDTH|POSTPONED);
10304 if (chain == NULL) /* First piece. */
10305 *flagp |= flags&SPSTART;
10308 REGTAIL(pRExC_state, chain, latest);
10313 if (chain == NULL) { /* Loop ran zero times. */
10314 chain = reg_node(pRExC_state, NOTHING);
10319 *flagp |= flags&SIMPLE;
10326 - regpiece - something followed by possible [*+?]
10328 * Note that the branching code sequences used for ? and the general cases
10329 * of * and + are somewhat optimized: they use the same NOTHING node as
10330 * both the endmarker for their branch list and the body of the last branch.
10331 * It might seem that this node could be dispensed with entirely, but the
10332 * endmarker role is not redundant.
10334 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10336 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10340 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10347 const char * const origparse = RExC_parse;
10349 I32 max = REG_INFTY;
10350 #ifdef RE_TRACK_PATTERN_OFFSETS
10353 const char *maxpos = NULL;
10355 /* Save the original in case we change the emitted regop to a FAIL. */
10356 regnode * const orig_emit = RExC_emit;
10358 GET_RE_DEBUG_FLAGS_DECL;
10360 PERL_ARGS_ASSERT_REGPIECE;
10362 DEBUG_PARSE("piec");
10364 ret = regatom(pRExC_state, &flags,depth+1);
10366 if (flags & (TRYAGAIN|RESTART_UTF8))
10367 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10369 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10375 if (op == '{' && regcurly(RExC_parse, FALSE)) {
10377 #ifdef RE_TRACK_PATTERN_OFFSETS
10378 parse_start = RExC_parse; /* MJD */
10380 next = RExC_parse + 1;
10381 while (isDIGIT(*next) || *next == ',') {
10382 if (*next == ',') {
10390 if (*next == '}') { /* got one */
10394 min = atoi(RExC_parse);
10395 if (*maxpos == ',')
10398 maxpos = RExC_parse;
10399 max = atoi(maxpos);
10400 if (!max && *maxpos != '0')
10401 max = REG_INFTY; /* meaning "infinity" */
10402 else if (max >= REG_INFTY)
10403 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10405 nextchar(pRExC_state);
10406 if (max < min) { /* If can't match, warn and optimize to fail
10409 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10411 /* We can't back off the size because we have to reserve
10412 * enough space for all the things we are about to throw
10413 * away, but we can shrink it by the ammount we are about
10414 * to re-use here */
10415 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10418 RExC_emit = orig_emit;
10420 ret = reg_node(pRExC_state, OPFAIL);
10423 else if (min == max
10424 && RExC_parse < RExC_end
10425 && (*RExC_parse == '?' || *RExC_parse == '+'))
10428 ckWARN2reg(RExC_parse + 1,
10429 "Useless use of greediness modifier '%c'",
10432 /* Absorb the modifier, so later code doesn't see nor use
10434 nextchar(pRExC_state);
10438 if ((flags&SIMPLE)) {
10439 RExC_naughty += 2 + RExC_naughty / 2;
10440 reginsert(pRExC_state, CURLY, ret, depth+1);
10441 Set_Node_Offset(ret, parse_start+1); /* MJD */
10442 Set_Node_Cur_Length(ret, parse_start);
10445 regnode * const w = reg_node(pRExC_state, WHILEM);
10448 REGTAIL(pRExC_state, ret, w);
10449 if (!SIZE_ONLY && RExC_extralen) {
10450 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10451 reginsert(pRExC_state, NOTHING,ret, depth+1);
10452 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10454 reginsert(pRExC_state, CURLYX,ret, depth+1);
10456 Set_Node_Offset(ret, parse_start+1);
10457 Set_Node_Length(ret,
10458 op == '{' ? (RExC_parse - parse_start) : 1);
10460 if (!SIZE_ONLY && RExC_extralen)
10461 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10462 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10464 RExC_whilem_seen++, RExC_extralen += 3;
10465 RExC_naughty += 4 + RExC_naughty; /* compound interest */
10472 *flagp |= HASWIDTH;
10474 ARG1_SET(ret, (U16)min);
10475 ARG2_SET(ret, (U16)max);
10477 if (max == REG_INFTY)
10478 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10484 if (!ISMULT1(op)) {
10489 #if 0 /* Now runtime fix should be reliable. */
10491 /* if this is reinstated, don't forget to put this back into perldiag:
10493 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10495 (F) The part of the regexp subject to either the * or + quantifier
10496 could match an empty string. The {#} shows in the regular
10497 expression about where the problem was discovered.
10501 if (!(flags&HASWIDTH) && op != '?')
10502 vFAIL("Regexp *+ operand could be empty");
10505 #ifdef RE_TRACK_PATTERN_OFFSETS
10506 parse_start = RExC_parse;
10508 nextchar(pRExC_state);
10510 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10512 if (op == '*' && (flags&SIMPLE)) {
10513 reginsert(pRExC_state, STAR, ret, depth+1);
10516 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10518 else if (op == '*') {
10522 else if (op == '+' && (flags&SIMPLE)) {
10523 reginsert(pRExC_state, PLUS, ret, depth+1);
10526 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10528 else if (op == '+') {
10532 else if (op == '?') {
10537 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10538 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10539 ckWARN2reg(RExC_parse,
10540 "%"UTF8f" matches null string many times",
10541 UTF8fARG(UTF, (RExC_parse >= origparse
10542 ? RExC_parse - origparse
10545 (void)ReREFCNT_inc(RExC_rx_sv);
10548 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10549 nextchar(pRExC_state);
10550 reginsert(pRExC_state, MINMOD, ret, depth+1);
10551 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10554 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10556 nextchar(pRExC_state);
10557 ender = reg_node(pRExC_state, SUCCEED);
10558 REGTAIL(pRExC_state, ret, ender);
10559 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10561 ender = reg_node(pRExC_state, TAIL);
10562 REGTAIL(pRExC_state, ret, ender);
10565 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10567 vFAIL("Nested quantifiers");
10574 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
10575 UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
10576 const bool strict /* Apply stricter parsing rules? */
10580 /* This is expected to be called by a parser routine that has recognized '\N'
10581 and needs to handle the rest. RExC_parse is expected to point at the first
10582 char following the N at the time of the call. On successful return,
10583 RExC_parse has been updated to point to just after the sequence identified
10584 by this routine, and <*flagp> has been updated.
10586 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
10589 \N may begin either a named sequence, or if outside a character class, mean
10590 to match a non-newline. For non single-quoted regexes, the tokenizer has
10591 attempted to decide which, and in the case of a named sequence, converted it
10592 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
10593 where c1... are the characters in the sequence. For single-quoted regexes,
10594 the tokenizer passes the \N sequence through unchanged; this code will not
10595 attempt to determine this nor expand those, instead raising a syntax error.
10596 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
10597 or there is no '}', it signals that this \N occurrence means to match a
10600 Only the \N{U+...} form should occur in a character class, for the same
10601 reason that '.' inside a character class means to just match a period: it
10602 just doesn't make sense.
10604 The function raises an error (via vFAIL), and doesn't return for various
10605 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
10606 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
10607 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
10608 only possible if node_p is non-NULL.
10611 If <valuep> is non-null, it means the caller can accept an input sequence
10612 consisting of a just a single code point; <*valuep> is set to that value
10613 if the input is such.
10615 If <node_p> is non-null it signifies that the caller can accept any other
10616 legal sequence (i.e., one that isn't just a single code point). <*node_p>
10618 1) \N means not-a-NL: points to a newly created REG_ANY node;
10619 2) \N{}: points to a new NOTHING node;
10620 3) otherwise: points to a new EXACT node containing the resolved
10622 Note that FALSE is returned for single code point sequences if <valuep> is
10626 char * endbrace; /* '}' following the name */
10628 char *endchar; /* Points to '.' or '}' ending cur char in the input
10630 bool has_multiple_chars; /* true if the input stream contains a sequence of
10631 more than one character */
10633 GET_RE_DEBUG_FLAGS_DECL;
10635 PERL_ARGS_ASSERT_GROK_BSLASH_N;
10637 GET_RE_DEBUG_FLAGS;
10639 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
10641 /* The [^\n] meaning of \N ignores spaces and comments under the /x
10642 * modifier. The other meaning does not, so use a temporary until we find
10643 * out which we are being called with */
10644 p = (RExC_flags & RXf_PMf_EXTENDED)
10645 ? regwhite( pRExC_state, RExC_parse )
10648 /* Disambiguate between \N meaning a named character versus \N meaning
10649 * [^\n]. The former is assumed when it can't be the latter. */
10650 if (*p != '{' || regcurly(p, FALSE)) {
10653 /* no bare \N allowed in a charclass */
10654 if (in_char_class) {
10655 vFAIL("\\N in a character class must be a named character: \\N{...}");
10659 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
10661 nextchar(pRExC_state);
10662 *node_p = reg_node(pRExC_state, REG_ANY);
10663 *flagp |= HASWIDTH|SIMPLE;
10665 Set_Node_Length(*node_p, 1); /* MJD */
10669 /* Here, we have decided it should be a named character or sequence */
10671 /* The test above made sure that the next real character is a '{', but
10672 * under the /x modifier, it could be separated by space (or a comment and
10673 * \n) and this is not allowed (for consistency with \x{...} and the
10674 * tokenizer handling of \N{NAME}). */
10675 if (*RExC_parse != '{') {
10676 vFAIL("Missing braces on \\N{}");
10679 RExC_parse++; /* Skip past the '{' */
10681 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
10682 || ! (endbrace == RExC_parse /* nothing between the {} */
10683 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below
10685 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg)
10688 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
10689 vFAIL("\\N{NAME} must be resolved by the lexer");
10692 if (endbrace == RExC_parse) { /* empty: \N{} */
10695 *node_p = reg_node(pRExC_state,NOTHING);
10697 else if (in_char_class) {
10698 if (SIZE_ONLY && in_char_class) {
10700 RExC_parse++; /* Position after the "}" */
10701 vFAIL("Zero length \\N{}");
10704 ckWARNreg(RExC_parse,
10705 "Ignoring zero length \\N{} in character class");
10713 nextchar(pRExC_state);
10717 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
10718 RExC_parse += 2; /* Skip past the 'U+' */
10720 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10722 /* Code points are separated by dots. If none, there is only one code
10723 * point, and is terminated by the brace */
10724 has_multiple_chars = (endchar < endbrace);
10726 if (valuep && (! has_multiple_chars || in_char_class)) {
10727 /* We only pay attention to the first char of
10728 multichar strings being returned in char classes. I kinda wonder
10729 if this makes sense as it does change the behaviour
10730 from earlier versions, OTOH that behaviour was broken
10731 as well. XXX Solution is to recharacterize as
10732 [rest-of-class]|multi1|multi2... */
10734 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10735 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10736 | PERL_SCAN_DISALLOW_PREFIX
10737 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10739 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10741 /* The tokenizer should have guaranteed validity, but it's possible to
10742 * bypass it by using single quoting, so check */
10743 if (length_of_hex == 0
10744 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10746 RExC_parse += length_of_hex; /* Includes all the valid */
10747 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10748 ? UTF8SKIP(RExC_parse)
10750 /* Guard against malformed utf8 */
10751 if (RExC_parse >= endchar) {
10752 RExC_parse = endchar;
10754 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10757 if (in_char_class && has_multiple_chars) {
10759 RExC_parse = endbrace;
10760 vFAIL("\\N{} in character class restricted to one character");
10763 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10767 RExC_parse = endbrace + 1;
10769 else if (! node_p || ! has_multiple_chars) {
10771 /* Here, the input is legal, but not according to the caller's
10772 * options. We fail without advancing the parse, so that the
10773 * caller can try again */
10779 /* What is done here is to convert this to a sub-pattern of the form
10780 * (?:\x{char1}\x{char2}...)
10781 * and then call reg recursively. That way, it retains its atomicness,
10782 * while not having to worry about special handling that some code
10783 * points may have. toke.c has converted the original Unicode values
10784 * to native, so that we can just pass on the hex values unchanged. We
10785 * do have to set a flag to keep recoding from happening in the
10788 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10790 char *orig_end = RExC_end;
10793 while (RExC_parse < endbrace) {
10795 /* Convert to notation the rest of the code understands */
10796 sv_catpv(substitute_parse, "\\x{");
10797 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10798 sv_catpv(substitute_parse, "}");
10800 /* Point to the beginning of the next character in the sequence. */
10801 RExC_parse = endchar + 1;
10802 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10804 sv_catpv(substitute_parse, ")");
10806 RExC_parse = SvPV(substitute_parse, len);
10808 /* Don't allow empty number */
10810 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10812 RExC_end = RExC_parse + len;
10814 /* The values are Unicode, and therefore not subject to recoding */
10815 RExC_override_recoding = 1;
10817 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10818 if (flags & RESTART_UTF8) {
10819 *flagp = RESTART_UTF8;
10822 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10825 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10827 RExC_parse = endbrace;
10828 RExC_end = orig_end;
10829 RExC_override_recoding = 0;
10831 nextchar(pRExC_state);
10841 * It returns the code point in utf8 for the value in *encp.
10842 * value: a code value in the source encoding
10843 * encp: a pointer to an Encode object
10845 * If the result from Encode is not a single character,
10846 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10849 S_reg_recode(pTHX_ const char value, SV **encp)
10852 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10853 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10854 const STRLEN newlen = SvCUR(sv);
10855 UV uv = UNICODE_REPLACEMENT;
10857 PERL_ARGS_ASSERT_REG_RECODE;
10861 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10864 if (!newlen || numlen != newlen) {
10865 uv = UNICODE_REPLACEMENT;
10871 PERL_STATIC_INLINE U8
10872 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10876 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10882 op = get_regex_charset(RExC_flags);
10883 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10884 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10885 been, so there is no hole */
10888 return op + EXACTF;
10891 PERL_STATIC_INLINE void
10892 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
10893 regnode *node, I32* flagp, STRLEN len, UV code_point,
10896 /* This knows the details about sizing an EXACTish node, setting flags for
10897 * it (by setting <*flagp>, and potentially populating it with a single
10900 * If <len> (the length in bytes) is non-zero, this function assumes that
10901 * the node has already been populated, and just does the sizing. In this
10902 * case <code_point> should be the final code point that has already been
10903 * placed into the node. This value will be ignored except that under some
10904 * circumstances <*flagp> is set based on it.
10906 * If <len> is zero, the function assumes that the node is to contain only
10907 * the single character given by <code_point> and calculates what <len>
10908 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10909 * additionally will populate the node's STRING with <code_point> or its
10912 * In both cases <*flagp> is appropriately set
10914 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10915 * 255, must be folded (the former only when the rules indicate it can
10918 * When it does the populating, it looks at the flag 'downgradable'. If
10919 * true with a node that folds, it checks if the single code point
10920 * participates in a fold, and if not downgrades the node to an EXACT.
10921 * This helps the optimizer */
10923 bool len_passed_in = cBOOL(len != 0);
10924 U8 character[UTF8_MAXBYTES_CASE+1];
10926 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10928 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
10929 * sizing difference, and is extra work that is thrown away */
10930 if (downgradable && ! PASS2) {
10931 downgradable = FALSE;
10934 if (! len_passed_in) {
10936 if (UNI_IS_INVARIANT(code_point)) {
10937 if (LOC || ! FOLD) { /* /l defers folding until runtime */
10938 *character = (U8) code_point;
10940 else { /* Here is /i and not /l (toFOLD() is defined on just
10941 ASCII, which isn't the same thing as INVARIANT on
10942 EBCDIC, but it works there, as the extra invariants
10943 fold to themselves) */
10944 *character = toFOLD((U8) code_point);
10946 && *character == code_point
10947 && ! HAS_NONLATIN1_FOLD_CLOSURE(code_point))
10954 else if (FOLD && (! LOC
10955 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
10956 { /* Folding, and ok to do so now */
10957 UV folded = _to_uni_fold_flags(
10961 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
10962 ? FOLD_FLAGS_NOMIX_ASCII
10965 && folded == code_point
10966 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
10971 else if (code_point <= MAX_UTF8_TWO_BYTE) {
10973 /* Not folding this cp, and can output it directly */
10974 *character = UTF8_TWO_BYTE_HI(code_point);
10975 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
10979 uvchr_to_utf8( character, code_point);
10980 len = UTF8SKIP(character);
10982 } /* Else pattern isn't UTF8. */
10984 *character = (U8) code_point;
10986 } /* Else is folded non-UTF8 */
10987 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
10989 /* We don't fold any non-UTF8 except possibly the Sharp s (see
10990 * comments at join_exact()); */
10991 *character = (U8) code_point;
10994 /* Can turn into an EXACT node if we know the fold at compile time,
10995 * and it folds to itself and doesn't particpate in other folds */
10998 && PL_fold_latin1[code_point] == code_point
10999 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11000 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11004 } /* else is Sharp s. May need to fold it */
11005 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11007 *(character + 1) = 's';
11011 *character = LATIN_SMALL_LETTER_SHARP_S;
11017 RExC_size += STR_SZ(len);
11020 RExC_emit += STR_SZ(len);
11021 STR_LEN(node) = len;
11022 if (! len_passed_in) {
11023 Copy((char *) character, STRING(node), len, char);
11027 *flagp |= HASWIDTH;
11029 /* A single character node is SIMPLE, except for the special-cased SHARP S
11031 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11032 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11033 || ! FOLD || ! DEPENDS_SEMANTICS))
11038 /* The OP may not be well defined in PASS1 */
11039 if (PASS2 && OP(node) == EXACTFL) {
11040 RExC_contains_locale = 1;
11045 /* return atoi(p), unless it's too big to sensibly be a backref,
11046 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11049 S_backref_value(char *p)
11053 for (;isDIGIT(*q); q++); /* calculate length of num */
11054 if (q - p == 0 || q - p > 9)
11061 - regatom - the lowest level
11063 Try to identify anything special at the start of the pattern. If there
11064 is, then handle it as required. This may involve generating a single regop,
11065 such as for an assertion; or it may involve recursing, such as to
11066 handle a () structure.
11068 If the string doesn't start with something special then we gobble up
11069 as much literal text as we can.
11071 Once we have been able to handle whatever type of thing started the
11072 sequence, we return.
11074 Note: we have to be careful with escapes, as they can be both literal
11075 and special, and in the case of \10 and friends, context determines which.
11077 A summary of the code structure is:
11079 switch (first_byte) {
11080 cases for each special:
11081 handle this special;
11084 switch (2nd byte) {
11085 cases for each unambiguous special:
11086 handle this special;
11088 cases for each ambigous special/literal:
11090 if (special) handle here
11092 default: // unambiguously literal:
11095 default: // is a literal char
11098 create EXACTish node for literal;
11099 while (more input and node isn't full) {
11100 switch (input_byte) {
11101 cases for each special;
11102 make sure parse pointer is set so that the next call to
11103 regatom will see this special first
11104 goto loopdone; // EXACTish node terminated by prev. char
11106 append char to EXACTISH node;
11108 get next input byte;
11112 return the generated node;
11114 Specifically there are two separate switches for handling
11115 escape sequences, with the one for handling literal escapes requiring
11116 a dummy entry for all of the special escapes that are actually handled
11119 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11121 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11123 Otherwise does not return NULL.
11127 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11130 regnode *ret = NULL;
11132 char *parse_start = RExC_parse;
11136 GET_RE_DEBUG_FLAGS_DECL;
11138 *flagp = WORST; /* Tentatively. */
11140 DEBUG_PARSE("atom");
11142 PERL_ARGS_ASSERT_REGATOM;
11145 switch ((U8)*RExC_parse) {
11147 RExC_seen_zerolen++;
11148 nextchar(pRExC_state);
11149 if (RExC_flags & RXf_PMf_MULTILINE)
11150 ret = reg_node(pRExC_state, MBOL);
11151 else if (RExC_flags & RXf_PMf_SINGLELINE)
11152 ret = reg_node(pRExC_state, SBOL);
11154 ret = reg_node(pRExC_state, BOL);
11155 Set_Node_Length(ret, 1); /* MJD */
11158 nextchar(pRExC_state);
11160 RExC_seen_zerolen++;
11161 if (RExC_flags & RXf_PMf_MULTILINE)
11162 ret = reg_node(pRExC_state, MEOL);
11163 else if (RExC_flags & RXf_PMf_SINGLELINE)
11164 ret = reg_node(pRExC_state, SEOL);
11166 ret = reg_node(pRExC_state, EOL);
11167 Set_Node_Length(ret, 1); /* MJD */
11170 nextchar(pRExC_state);
11171 if (RExC_flags & RXf_PMf_SINGLELINE)
11172 ret = reg_node(pRExC_state, SANY);
11174 ret = reg_node(pRExC_state, REG_ANY);
11175 *flagp |= HASWIDTH|SIMPLE;
11177 Set_Node_Length(ret, 1); /* MJD */
11181 char * const oregcomp_parse = ++RExC_parse;
11182 ret = regclass(pRExC_state, flagp,depth+1,
11183 FALSE, /* means parse the whole char class */
11184 TRUE, /* allow multi-char folds */
11185 FALSE, /* don't silence non-portable warnings. */
11187 if (*RExC_parse != ']') {
11188 RExC_parse = oregcomp_parse;
11189 vFAIL("Unmatched [");
11192 if (*flagp & RESTART_UTF8)
11194 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11197 nextchar(pRExC_state);
11198 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11202 nextchar(pRExC_state);
11203 ret = reg(pRExC_state, 2, &flags,depth+1);
11205 if (flags & TRYAGAIN) {
11206 if (RExC_parse == RExC_end) {
11207 /* Make parent create an empty node if needed. */
11208 *flagp |= TRYAGAIN;
11213 if (flags & RESTART_UTF8) {
11214 *flagp = RESTART_UTF8;
11217 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11220 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11224 if (flags & TRYAGAIN) {
11225 *flagp |= TRYAGAIN;
11228 vFAIL("Internal urp");
11229 /* Supposed to be caught earlier. */
11232 if (!regcurly(RExC_parse, FALSE)) {
11241 vFAIL("Quantifier follows nothing");
11246 This switch handles escape sequences that resolve to some kind
11247 of special regop and not to literal text. Escape sequnces that
11248 resolve to literal text are handled below in the switch marked
11251 Every entry in this switch *must* have a corresponding entry
11252 in the literal escape switch. However, the opposite is not
11253 required, as the default for this switch is to jump to the
11254 literal text handling code.
11256 switch ((U8)*++RExC_parse) {
11258 /* Special Escapes */
11260 RExC_seen_zerolen++;
11261 ret = reg_node(pRExC_state, SBOL);
11263 goto finish_meta_pat;
11265 ret = reg_node(pRExC_state, GPOS);
11266 RExC_seen |= REG_GPOS_SEEN;
11268 goto finish_meta_pat;
11270 RExC_seen_zerolen++;
11271 ret = reg_node(pRExC_state, KEEPS);
11273 /* XXX:dmq : disabling in-place substitution seems to
11274 * be necessary here to avoid cases of memory corruption, as
11275 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11277 RExC_seen |= REG_LOOKBEHIND_SEEN;
11278 goto finish_meta_pat;
11280 ret = reg_node(pRExC_state, SEOL);
11282 RExC_seen_zerolen++; /* Do not optimize RE away */
11283 goto finish_meta_pat;
11285 ret = reg_node(pRExC_state, EOS);
11287 RExC_seen_zerolen++; /* Do not optimize RE away */
11288 goto finish_meta_pat;
11290 ret = reg_node(pRExC_state, CANY);
11291 RExC_seen |= REG_CANY_SEEN;
11292 *flagp |= HASWIDTH|SIMPLE;
11293 goto finish_meta_pat;
11295 ret = reg_node(pRExC_state, CLUMP);
11296 *flagp |= HASWIDTH;
11297 goto finish_meta_pat;
11303 arg = ANYOF_WORDCHAR;
11307 RExC_seen_zerolen++;
11308 RExC_seen |= REG_LOOKBEHIND_SEEN;
11309 op = BOUND + get_regex_charset(RExC_flags);
11310 if (op > BOUNDA) { /* /aa is same as /a */
11313 else if (op == BOUNDL) {
11314 RExC_contains_locale = 1;
11316 ret = reg_node(pRExC_state, op);
11317 FLAGS(ret) = get_regex_charset(RExC_flags);
11319 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11320 /* diag_listed_as: Use "%s" instead of "%s" */
11321 vFAIL("Use \"\\b\\{\" instead of \"\\b{\"");
11323 goto finish_meta_pat;
11325 RExC_seen_zerolen++;
11326 RExC_seen |= REG_LOOKBEHIND_SEEN;
11327 op = NBOUND + get_regex_charset(RExC_flags);
11328 if (op > NBOUNDA) { /* /aa is same as /a */
11331 else if (op == NBOUNDL) {
11332 RExC_contains_locale = 1;
11334 ret = reg_node(pRExC_state, op);
11335 FLAGS(ret) = get_regex_charset(RExC_flags);
11337 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11338 /* diag_listed_as: Use "%s" instead of "%s" */
11339 vFAIL("Use \"\\B\\{\" instead of \"\\B{\"");
11341 goto finish_meta_pat;
11351 ret = reg_node(pRExC_state, LNBREAK);
11352 *flagp |= HASWIDTH|SIMPLE;
11353 goto finish_meta_pat;
11361 goto join_posix_op_known;
11367 arg = ANYOF_VERTWS;
11369 goto join_posix_op_known;
11379 op = POSIXD + get_regex_charset(RExC_flags);
11380 if (op > POSIXA) { /* /aa is same as /a */
11383 else if (op == POSIXL) {
11384 RExC_contains_locale = 1;
11387 join_posix_op_known:
11390 op += NPOSIXD - POSIXD;
11393 ret = reg_node(pRExC_state, op);
11395 FLAGS(ret) = namedclass_to_classnum(arg);
11398 *flagp |= HASWIDTH|SIMPLE;
11402 nextchar(pRExC_state);
11403 Set_Node_Length(ret, 2); /* MJD */
11409 char* parse_start = RExC_parse - 2;
11414 ret = regclass(pRExC_state, flagp,depth+1,
11415 TRUE, /* means just parse this element */
11416 FALSE, /* don't allow multi-char folds */
11417 FALSE, /* don't silence non-portable warnings.
11418 It would be a bug if these returned
11421 /* regclass() can only return RESTART_UTF8 if multi-char folds
11424 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11429 Set_Node_Offset(ret, parse_start + 2);
11430 Set_Node_Cur_Length(ret, parse_start);
11431 nextchar(pRExC_state);
11435 /* Handle \N and \N{NAME} with multiple code points here and not
11436 * below because it can be multicharacter. join_exact() will join
11437 * them up later on. Also this makes sure that things like
11438 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11439 * The options to the grok function call causes it to fail if the
11440 * sequence is just a single code point. We then go treat it as
11441 * just another character in the current EXACT node, and hence it
11442 * gets uniform treatment with all the other characters. The
11443 * special treatment for quantifiers is not needed for such single
11444 * character sequences */
11446 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
11447 FALSE /* not strict */ )) {
11448 if (*flagp & RESTART_UTF8)
11454 case 'k': /* Handle \k<NAME> and \k'NAME' */
11457 char ch= RExC_parse[1];
11458 if (ch != '<' && ch != '\'' && ch != '{') {
11460 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11461 vFAIL2("Sequence %.2s... not terminated",parse_start);
11463 /* this pretty much dupes the code for (?P=...) in reg(), if
11464 you change this make sure you change that */
11465 char* name_start = (RExC_parse += 2);
11467 SV *sv_dat = reg_scan_name(pRExC_state,
11468 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11469 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11470 if (RExC_parse == name_start || *RExC_parse != ch)
11471 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11472 vFAIL2("Sequence %.3s... not terminated",parse_start);
11475 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11476 RExC_rxi->data->data[num]=(void*)sv_dat;
11477 SvREFCNT_inc_simple_void(sv_dat);
11481 ret = reganode(pRExC_state,
11484 : (ASCII_FOLD_RESTRICTED)
11486 : (AT_LEAST_UNI_SEMANTICS)
11492 *flagp |= HASWIDTH;
11494 /* override incorrect value set in reganode MJD */
11495 Set_Node_Offset(ret, parse_start+1);
11496 Set_Node_Cur_Length(ret, parse_start);
11497 nextchar(pRExC_state);
11503 case '1': case '2': case '3': case '4':
11504 case '5': case '6': case '7': case '8': case '9':
11509 if (*RExC_parse == 'g') {
11513 if (*RExC_parse == '{') {
11517 if (*RExC_parse == '-') {
11521 if (hasbrace && !isDIGIT(*RExC_parse)) {
11522 if (isrel) RExC_parse--;
11524 goto parse_named_seq;
11527 num = S_backref_value(RExC_parse);
11529 vFAIL("Reference to invalid group 0");
11530 else if (num == I32_MAX) {
11531 if (isDIGIT(*RExC_parse))
11532 vFAIL("Reference to nonexistent group");
11534 vFAIL("Unterminated \\g... pattern");
11538 num = RExC_npar - num;
11540 vFAIL("Reference to nonexistent or unclosed group");
11544 num = S_backref_value(RExC_parse);
11545 /* bare \NNN might be backref or octal - if it is larger than or equal
11546 * RExC_npar then it is assumed to be and octal escape.
11547 * Note RExC_npar is +1 from the actual number of parens*/
11548 if (num == I32_MAX || (num > 9 && num >= RExC_npar
11549 && *RExC_parse != '8' && *RExC_parse != '9'))
11551 /* Probably a character specified in octal, e.g. \35 */
11556 /* at this point RExC_parse definitely points to a backref
11559 #ifdef RE_TRACK_PATTERN_OFFSETS
11560 char * const parse_start = RExC_parse - 1; /* MJD */
11562 while (isDIGIT(*RExC_parse))
11565 if (*RExC_parse != '}')
11566 vFAIL("Unterminated \\g{...} pattern");
11570 if (num > (I32)RExC_rx->nparens)
11571 vFAIL("Reference to nonexistent group");
11574 ret = reganode(pRExC_state,
11577 : (ASCII_FOLD_RESTRICTED)
11579 : (AT_LEAST_UNI_SEMANTICS)
11585 *flagp |= HASWIDTH;
11587 /* override incorrect value set in reganode MJD */
11588 Set_Node_Offset(ret, parse_start+1);
11589 Set_Node_Cur_Length(ret, parse_start);
11591 nextchar(pRExC_state);
11596 if (RExC_parse >= RExC_end)
11597 FAIL("Trailing \\");
11600 /* Do not generate "unrecognized" warnings here, we fall
11601 back into the quick-grab loop below */
11608 if (RExC_flags & RXf_PMf_EXTENDED) {
11609 if ( reg_skipcomment( pRExC_state ) )
11616 parse_start = RExC_parse - 1;
11625 #define MAX_NODE_STRING_SIZE 127
11626 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
11628 U8 upper_parse = MAX_NODE_STRING_SIZE;
11629 U8 node_type = compute_EXACTish(pRExC_state);
11630 bool next_is_quantifier;
11631 char * oldp = NULL;
11633 /* We can convert EXACTF nodes to EXACTFU if they contain only
11634 * characters that match identically regardless of the target
11635 * string's UTF8ness. The reason to do this is that EXACTF is not
11636 * trie-able, EXACTFU is.
11638 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
11639 * contain only above-Latin1 characters (hence must be in UTF8),
11640 * which don't participate in folds with Latin1-range characters,
11641 * as the latter's folds aren't known until runtime. (We don't
11642 * need to figure this out until pass 2) */
11643 bool maybe_exactfu = PASS2
11644 && (node_type == EXACTF || node_type == EXACTFL);
11646 /* If a folding node contains only code points that don't
11647 * participate in folds, it can be changed into an EXACT node,
11648 * which allows the optimizer more things to look for */
11651 ret = reg_node(pRExC_state, node_type);
11653 /* In pass1, folded, we use a temporary buffer instead of the
11654 * actual node, as the node doesn't exist yet */
11655 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
11661 /* We do the EXACTFish to EXACT node only if folding. (And we
11662 * don't need to figure this out until pass 2) */
11663 maybe_exact = FOLD && PASS2;
11665 /* XXX The node can hold up to 255 bytes, yet this only goes to
11666 * 127. I (khw) do not know why. Keeping it somewhat less than
11667 * 255 allows us to not have to worry about overflow due to
11668 * converting to utf8 and fold expansion, but that value is
11669 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
11670 * split up by this limit into a single one using the real max of
11671 * 255. Even at 127, this breaks under rare circumstances. If
11672 * folding, we do not want to split a node at a character that is a
11673 * non-final in a multi-char fold, as an input string could just
11674 * happen to want to match across the node boundary. The join
11675 * would solve that problem if the join actually happens. But a
11676 * series of more than two nodes in a row each of 127 would cause
11677 * the first join to succeed to get to 254, but then there wouldn't
11678 * be room for the next one, which could at be one of those split
11679 * multi-char folds. I don't know of any fool-proof solution. One
11680 * could back off to end with only a code point that isn't such a
11681 * non-final, but it is possible for there not to be any in the
11683 for (p = RExC_parse - 1;
11684 len < upper_parse && p < RExC_end;
11689 if (RExC_flags & RXf_PMf_EXTENDED)
11690 p = regwhite( pRExC_state, p );
11701 /* Literal Escapes Switch
11703 This switch is meant to handle escape sequences that
11704 resolve to a literal character.
11706 Every escape sequence that represents something
11707 else, like an assertion or a char class, is handled
11708 in the switch marked 'Special Escapes' above in this
11709 routine, but also has an entry here as anything that
11710 isn't explicitly mentioned here will be treated as
11711 an unescaped equivalent literal.
11714 switch ((U8)*++p) {
11715 /* These are all the special escapes. */
11716 case 'A': /* Start assertion */
11717 case 'b': case 'B': /* Word-boundary assertion*/
11718 case 'C': /* Single char !DANGEROUS! */
11719 case 'd': case 'D': /* digit class */
11720 case 'g': case 'G': /* generic-backref, pos assertion */
11721 case 'h': case 'H': /* HORIZWS */
11722 case 'k': case 'K': /* named backref, keep marker */
11723 case 'p': case 'P': /* Unicode property */
11724 case 'R': /* LNBREAK */
11725 case 's': case 'S': /* space class */
11726 case 'v': case 'V': /* VERTWS */
11727 case 'w': case 'W': /* word class */
11728 case 'X': /* eXtended Unicode "combining
11729 character sequence" */
11730 case 'z': case 'Z': /* End of line/string assertion */
11734 /* Anything after here is an escape that resolves to a
11735 literal. (Except digits, which may or may not)
11741 case 'N': /* Handle a single-code point named character. */
11742 /* The options cause it to fail if a multiple code
11743 * point sequence. Handle those in the switch() above
11745 RExC_parse = p + 1;
11746 if (! grok_bslash_N(pRExC_state, NULL, &ender,
11747 flagp, depth, FALSE,
11748 FALSE /* not strict */ ))
11750 if (*flagp & RESTART_UTF8)
11751 FAIL("panic: grok_bslash_N set RESTART_UTF8");
11752 RExC_parse = p = oldp;
11756 if (ender > 0xff) {
11773 ender = ASCII_TO_NATIVE('\033');
11783 const char* error_msg;
11785 bool valid = grok_bslash_o(&p,
11788 TRUE, /* out warnings */
11789 FALSE, /* not strict */
11790 TRUE, /* Output warnings
11795 RExC_parse = p; /* going to die anyway; point
11796 to exact spot of failure */
11800 if (PL_encoding && ender < 0x100) {
11801 goto recode_encoding;
11803 if (ender > 0xff) {
11810 UV result = UV_MAX; /* initialize to erroneous
11812 const char* error_msg;
11814 bool valid = grok_bslash_x(&p,
11817 TRUE, /* out warnings */
11818 FALSE, /* not strict */
11819 TRUE, /* Output warnings
11824 RExC_parse = p; /* going to die anyway; point
11825 to exact spot of failure */
11830 if (PL_encoding && ender < 0x100) {
11831 goto recode_encoding;
11833 if (ender > 0xff) {
11840 ender = grok_bslash_c(*p++, SIZE_ONLY);
11842 case '8': case '9': /* must be a backreference */
11845 case '1': case '2': case '3':case '4':
11846 case '5': case '6': case '7':
11847 /* When we parse backslash escapes there is ambiguity
11848 * between backreferences and octal escapes. Any escape
11849 * from \1 - \9 is a backreference, any multi-digit
11850 * escape which does not start with 0 and which when
11851 * evaluated as decimal could refer to an already
11852 * parsed capture buffer is a backslash. Anything else
11855 * Note this implies that \118 could be interpreted as
11856 * 118 OR as "\11" . "8" depending on whether there
11857 * were 118 capture buffers defined already in the
11860 /* NOTE, RExC_npar is 1 more than the actual number of
11861 * parens we have seen so far, hence the < RExC_npar below. */
11863 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
11864 { /* Not to be treated as an octal constant, go
11871 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11873 ender = grok_oct(p, &numlen, &flags, NULL);
11874 if (ender > 0xff) {
11878 if (SIZE_ONLY /* like \08, \178 */
11881 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11883 reg_warn_non_literal_string(
11885 form_short_octal_warning(p, numlen));
11888 if (PL_encoding && ender < 0x100)
11889 goto recode_encoding;
11892 if (! RExC_override_recoding) {
11893 SV* enc = PL_encoding;
11894 ender = reg_recode((const char)(U8)ender, &enc);
11895 if (!enc && SIZE_ONLY)
11896 ckWARNreg(p, "Invalid escape in the specified encoding");
11902 FAIL("Trailing \\");
11905 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11906 /* Include any { following the alpha to emphasize
11907 * that it could be part of an escape at some point
11909 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11910 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11912 goto normal_default;
11913 } /* End of switch on '\' */
11915 default: /* A literal character */
11918 && RExC_flags & RXf_PMf_EXTENDED
11919 && ckWARN_d(WARN_DEPRECATED)
11920 && is_PATWS_non_low(p, UTF))
11922 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11923 "Escape literal pattern white space under /x");
11927 if (UTF8_IS_START(*p) && UTF) {
11929 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11930 &numlen, UTF8_ALLOW_DEFAULT);
11936 } /* End of switch on the literal */
11938 /* Here, have looked at the literal character and <ender>
11939 * contains its ordinal, <p> points to the character after it
11942 if ( RExC_flags & RXf_PMf_EXTENDED)
11943 p = regwhite( pRExC_state, p );
11945 /* If the next thing is a quantifier, it applies to this
11946 * character only, which means that this character has to be in
11947 * its own node and can't just be appended to the string in an
11948 * existing node, so if there are already other characters in
11949 * the node, close the node with just them, and set up to do
11950 * this character again next time through, when it will be the
11951 * only thing in its new node */
11952 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11958 if (! FOLD /* The simple case, just append the literal */
11959 || (LOC /* Also don't fold for tricky chars under /l */
11960 && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)))
11963 const STRLEN unilen = reguni(pRExC_state, ender, s);
11969 /* The loop increments <len> each time, as all but this
11970 * path (and one other) through it add a single byte to
11971 * the EXACTish node. But this one has changed len to
11972 * be the correct final value, so subtract one to
11973 * cancel out the increment that follows */
11977 REGC((char)ender, s++);
11980 /* Can get here if folding only if is one of the /l
11981 * characters whose fold depends on the locale. The
11982 * occurrence of any of these indicate that we can't
11983 * simplify things */
11985 maybe_exact = FALSE;
11986 maybe_exactfu = FALSE;
11991 /* See comments for join_exact() as to why we fold this
11992 * non-UTF at compile time */
11993 || (node_type == EXACTFU
11994 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11996 /* Here, are folding and are not UTF-8 encoded; therefore
11997 * the character must be in the range 0-255, and is not /l
11998 * (Not /l because we already handled these under /l in
11999 * is_PROBLEMATIC_LOCALE_FOLD_cp */
12000 if (IS_IN_SOME_FOLD_L1(ender)) {
12001 maybe_exact = FALSE;
12003 /* See if the character's fold differs between /d and
12004 * /u. This includes the multi-char fold SHARP S to
12007 && (PL_fold[ender] != PL_fold_latin1[ender]
12008 || ender == LATIN_SMALL_LETTER_SHARP_S
12010 && isARG2_lower_or_UPPER_ARG1('s', ender)
12011 && isARG2_lower_or_UPPER_ARG1('s',
12014 maybe_exactfu = FALSE;
12018 /* Even when folding, we store just the input character, as
12019 * we have an array that finds its fold quickly */
12020 *(s++) = (char) ender;
12022 else { /* FOLD and UTF */
12023 /* Unlike the non-fold case, we do actually have to
12024 * calculate the results here in pass 1. This is for two
12025 * reasons, the folded length may be longer than the
12026 * unfolded, and we have to calculate how many EXACTish
12027 * nodes it will take; and we may run out of room in a node
12028 * in the middle of a potential multi-char fold, and have
12029 * to back off accordingly. (Hence we can't use REGC for
12030 * the simple case just below.) */
12033 if (isASCII(ender)) {
12034 folded = toFOLD(ender);
12035 *(s)++ = (U8) folded;
12040 folded = _to_uni_fold_flags(
12044 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12045 ? FOLD_FLAGS_NOMIX_ASCII
12049 /* The loop increments <len> each time, as all but this
12050 * path (and one other) through it add a single byte to
12051 * the EXACTish node. But this one has changed len to
12052 * be the correct final value, so subtract one to
12053 * cancel out the increment that follows */
12054 len += foldlen - 1;
12056 /* If this node only contains non-folding code points so
12057 * far, see if this new one is also non-folding */
12059 if (folded != ender) {
12060 maybe_exact = FALSE;
12063 /* Here the fold is the original; we have to check
12064 * further to see if anything folds to it */
12065 if (_invlist_contains_cp(PL_utf8_foldable,
12068 maybe_exact = FALSE;
12075 if (next_is_quantifier) {
12077 /* Here, the next input is a quantifier, and to get here,
12078 * the current character is the only one in the node.
12079 * Also, here <len> doesn't include the final byte for this
12085 } /* End of loop through literal characters */
12087 /* Here we have either exhausted the input or ran out of room in
12088 * the node. (If we encountered a character that can't be in the
12089 * node, transfer is made directly to <loopdone>, and so we
12090 * wouldn't have fallen off the end of the loop.) In the latter
12091 * case, we artificially have to split the node into two, because
12092 * we just don't have enough space to hold everything. This
12093 * creates a problem if the final character participates in a
12094 * multi-character fold in the non-final position, as a match that
12095 * should have occurred won't, due to the way nodes are matched,
12096 * and our artificial boundary. So back off until we find a non-
12097 * problematic character -- one that isn't at the beginning or
12098 * middle of such a fold. (Either it doesn't participate in any
12099 * folds, or appears only in the final position of all the folds it
12100 * does participate in.) A better solution with far fewer false
12101 * positives, and that would fill the nodes more completely, would
12102 * be to actually have available all the multi-character folds to
12103 * test against, and to back-off only far enough to be sure that
12104 * this node isn't ending with a partial one. <upper_parse> is set
12105 * further below (if we need to reparse the node) to include just
12106 * up through that final non-problematic character that this code
12107 * identifies, so when it is set to less than the full node, we can
12108 * skip the rest of this */
12109 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12111 const STRLEN full_len = len;
12113 assert(len >= MAX_NODE_STRING_SIZE);
12115 /* Here, <s> points to the final byte of the final character.
12116 * Look backwards through the string until find a non-
12117 * problematic character */
12121 /* This has no multi-char folds to non-UTF characters */
12122 if (ASCII_FOLD_RESTRICTED) {
12126 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12130 if (! PL_NonL1NonFinalFold) {
12131 PL_NonL1NonFinalFold = _new_invlist_C_array(
12132 NonL1_Perl_Non_Final_Folds_invlist);
12135 /* Point to the first byte of the final character */
12136 s = (char *) utf8_hop((U8 *) s, -1);
12138 while (s >= s0) { /* Search backwards until find
12139 non-problematic char */
12140 if (UTF8_IS_INVARIANT(*s)) {
12142 /* There are no ascii characters that participate
12143 * in multi-char folds under /aa. In EBCDIC, the
12144 * non-ascii invariants are all control characters,
12145 * so don't ever participate in any folds. */
12146 if (ASCII_FOLD_RESTRICTED
12147 || ! IS_NON_FINAL_FOLD(*s))
12152 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12153 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12159 else if (! _invlist_contains_cp(
12160 PL_NonL1NonFinalFold,
12161 valid_utf8_to_uvchr((U8 *) s, NULL)))
12166 /* Here, the current character is problematic in that
12167 * it does occur in the non-final position of some
12168 * fold, so try the character before it, but have to
12169 * special case the very first byte in the string, so
12170 * we don't read outside the string */
12171 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12172 } /* End of loop backwards through the string */
12174 /* If there were only problematic characters in the string,
12175 * <s> will point to before s0, in which case the length
12176 * should be 0, otherwise include the length of the
12177 * non-problematic character just found */
12178 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12181 /* Here, have found the final character, if any, that is
12182 * non-problematic as far as ending the node without splitting
12183 * it across a potential multi-char fold. <len> contains the
12184 * number of bytes in the node up-to and including that
12185 * character, or is 0 if there is no such character, meaning
12186 * the whole node contains only problematic characters. In
12187 * this case, give up and just take the node as-is. We can't
12192 /* If the node ends in an 's' we make sure it stays EXACTF,
12193 * as if it turns into an EXACTFU, it could later get
12194 * joined with another 's' that would then wrongly match
12196 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
12198 maybe_exactfu = FALSE;
12202 /* Here, the node does contain some characters that aren't
12203 * problematic. If one such is the final character in the
12204 * node, we are done */
12205 if (len == full_len) {
12208 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12210 /* If the final character is problematic, but the
12211 * penultimate is not, back-off that last character to
12212 * later start a new node with it */
12217 /* Here, the final non-problematic character is earlier
12218 * in the input than the penultimate character. What we do
12219 * is reparse from the beginning, going up only as far as
12220 * this final ok one, thus guaranteeing that the node ends
12221 * in an acceptable character. The reason we reparse is
12222 * that we know how far in the character is, but we don't
12223 * know how to correlate its position with the input parse.
12224 * An alternate implementation would be to build that
12225 * correlation as we go along during the original parse,
12226 * but that would entail extra work for every node, whereas
12227 * this code gets executed only when the string is too
12228 * large for the node, and the final two characters are
12229 * problematic, an infrequent occurrence. Yet another
12230 * possible strategy would be to save the tail of the
12231 * string, and the next time regatom is called, initialize
12232 * with that. The problem with this is that unless you
12233 * back off one more character, you won't be guaranteed
12234 * regatom will get called again, unless regbranch,
12235 * regpiece ... are also changed. If you do back off that
12236 * extra character, so that there is input guaranteed to
12237 * force calling regatom, you can't handle the case where
12238 * just the first character in the node is acceptable. I
12239 * (khw) decided to try this method which doesn't have that
12240 * pitfall; if performance issues are found, we can do a
12241 * combination of the current approach plus that one */
12247 } /* End of verifying node ends with an appropriate char */
12249 loopdone: /* Jumped to when encounters something that shouldn't be in
12252 /* I (khw) don't know if you can get here with zero length, but the
12253 * old code handled this situation by creating a zero-length EXACT
12254 * node. Might as well be NOTHING instead */
12260 /* If 'maybe_exact' is still set here, means there are no
12261 * code points in the node that participate in folds;
12262 * similarly for 'maybe_exactfu' and code points that match
12263 * differently depending on UTF8ness of the target string
12264 * (for /u), or depending on locale for /l */
12268 else if (maybe_exactfu) {
12272 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12273 FALSE /* Don't look to see if could
12274 be turned into an EXACT
12275 node, as we have already
12280 RExC_parse = p - 1;
12281 Set_Node_Cur_Length(ret, parse_start);
12282 nextchar(pRExC_state);
12284 /* len is STRLEN which is unsigned, need to copy to signed */
12287 vFAIL("Internal disaster");
12290 } /* End of label 'defchar:' */
12292 } /* End of giant switch on input character */
12298 S_regwhite( RExC_state_t *pRExC_state, char *p )
12300 const char *e = RExC_end;
12302 PERL_ARGS_ASSERT_REGWHITE;
12307 else if (*p == '#') {
12310 if (*p++ == '\n') {
12316 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12325 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12327 /* Returns the next non-pattern-white space, non-comment character (the
12328 * latter only if 'recognize_comment is true) in the string p, which is
12329 * ended by RExC_end. If there is no line break ending a comment,
12330 * RExC_seen has added the REG_RUN_ON_COMMENT_SEEN flag; */
12331 const char *e = RExC_end;
12333 PERL_ARGS_ASSERT_REGPATWS;
12337 if ((len = is_PATWS_safe(p, e, UTF))) {
12340 else if (recognize_comment && *p == '#') {
12344 if (is_LNBREAK_safe(p, e, UTF)) {
12350 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
12359 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12361 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12362 * sets up the bitmap and any flags, removing those code points from the
12363 * inversion list, setting it to NULL should it become completely empty */
12365 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12366 assert(PL_regkind[OP(node)] == ANYOF);
12368 ANYOF_BITMAP_ZERO(node);
12369 if (*invlist_ptr) {
12371 /* This gets set if we actually need to modify things */
12372 bool change_invlist = FALSE;
12376 /* Start looking through *invlist_ptr */
12377 invlist_iterinit(*invlist_ptr);
12378 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12382 if (end == UV_MAX && start <= 256) {
12383 ANYOF_FLAGS(node) |= ANYOF_ABOVE_LATIN1_ALL;
12385 else if (end >= 256) {
12386 ANYOF_FLAGS(node) |= ANYOF_UTF8;
12389 /* Quit if are above what we should change */
12394 change_invlist = TRUE;
12396 /* Set all the bits in the range, up to the max that we are doing */
12397 high = (end < 255) ? end : 255;
12398 for (i = start; i <= (int) high; i++) {
12399 if (! ANYOF_BITMAP_TEST(node, i)) {
12400 ANYOF_BITMAP_SET(node, i);
12404 invlist_iterfinish(*invlist_ptr);
12406 /* Done with loop; remove any code points that are in the bitmap from
12407 * *invlist_ptr; similarly for code points above latin1 if we have a
12408 * flag to match all of them anyways */
12409 if (change_invlist) {
12410 _invlist_subtract(*invlist_ptr, PL_Latin1, invlist_ptr);
12412 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
12413 _invlist_intersection(*invlist_ptr, PL_Latin1, invlist_ptr);
12416 /* If have completely emptied it, remove it completely */
12417 if (_invlist_len(*invlist_ptr) == 0) {
12418 SvREFCNT_dec_NN(*invlist_ptr);
12419 *invlist_ptr = NULL;
12424 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12425 Character classes ([:foo:]) can also be negated ([:^foo:]).
12426 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12427 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12428 but trigger failures because they are currently unimplemented. */
12430 #define POSIXCC_DONE(c) ((c) == ':')
12431 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12432 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12434 PERL_STATIC_INLINE I32
12435 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12438 I32 namedclass = OOB_NAMEDCLASS;
12440 PERL_ARGS_ASSERT_REGPPOSIXCC;
12442 if (value == '[' && RExC_parse + 1 < RExC_end &&
12443 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12444 POSIXCC(UCHARAT(RExC_parse)))
12446 const char c = UCHARAT(RExC_parse);
12447 char* const s = RExC_parse++;
12449 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12451 if (RExC_parse == RExC_end) {
12454 /* Try to give a better location for the error (than the end of
12455 * the string) by looking for the matching ']' */
12457 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12460 vFAIL2("Unmatched '%c' in POSIX class", c);
12462 /* Grandfather lone [:, [=, [. */
12466 const char* const t = RExC_parse++; /* skip over the c */
12469 if (UCHARAT(RExC_parse) == ']') {
12470 const char *posixcc = s + 1;
12471 RExC_parse++; /* skip over the ending ] */
12474 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12475 const I32 skip = t - posixcc;
12477 /* Initially switch on the length of the name. */
12480 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12481 this is the Perl \w
12483 namedclass = ANYOF_WORDCHAR;
12486 /* Names all of length 5. */
12487 /* alnum alpha ascii blank cntrl digit graph lower
12488 print punct space upper */
12489 /* Offset 4 gives the best switch position. */
12490 switch (posixcc[4]) {
12492 if (memEQ(posixcc, "alph", 4)) /* alpha */
12493 namedclass = ANYOF_ALPHA;
12496 if (memEQ(posixcc, "spac", 4)) /* space */
12497 namedclass = ANYOF_PSXSPC;
12500 if (memEQ(posixcc, "grap", 4)) /* graph */
12501 namedclass = ANYOF_GRAPH;
12504 if (memEQ(posixcc, "asci", 4)) /* ascii */
12505 namedclass = ANYOF_ASCII;
12508 if (memEQ(posixcc, "blan", 4)) /* blank */
12509 namedclass = ANYOF_BLANK;
12512 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12513 namedclass = ANYOF_CNTRL;
12516 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12517 namedclass = ANYOF_ALPHANUMERIC;
12520 if (memEQ(posixcc, "lowe", 4)) /* lower */
12521 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12522 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12523 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12526 if (memEQ(posixcc, "digi", 4)) /* digit */
12527 namedclass = ANYOF_DIGIT;
12528 else if (memEQ(posixcc, "prin", 4)) /* print */
12529 namedclass = ANYOF_PRINT;
12530 else if (memEQ(posixcc, "punc", 4)) /* punct */
12531 namedclass = ANYOF_PUNCT;
12536 if (memEQ(posixcc, "xdigit", 6))
12537 namedclass = ANYOF_XDIGIT;
12541 if (namedclass == OOB_NAMEDCLASS)
12543 "POSIX class [:%"UTF8f":] unknown",
12544 UTF8fARG(UTF, t - s - 1, s + 1));
12546 /* The #defines are structured so each complement is +1 to
12547 * the normal one */
12551 assert (posixcc[skip] == ':');
12552 assert (posixcc[skip+1] == ']');
12553 } else if (!SIZE_ONLY) {
12554 /* [[=foo=]] and [[.foo.]] are still future. */
12556 /* adjust RExC_parse so the warning shows after
12557 the class closes */
12558 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
12560 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
12563 /* Maternal grandfather:
12564 * "[:" ending in ":" but not in ":]" */
12566 vFAIL("Unmatched '[' in POSIX class");
12569 /* Grandfather lone [:, [=, [. */
12579 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
12581 /* This applies some heuristics at the current parse position (which should
12582 * be at a '[') to see if what follows might be intended to be a [:posix:]
12583 * class. It returns true if it really is a posix class, of course, but it
12584 * also can return true if it thinks that what was intended was a posix
12585 * class that didn't quite make it.
12587 * It will return true for
12589 * [:alphanumerics] (as long as the ] isn't followed immediately by a
12590 * ')' indicating the end of the (?[
12591 * [:any garbage including %^&$ punctuation:]
12593 * This is designed to be called only from S_handle_regex_sets; it could be
12594 * easily adapted to be called from the spot at the beginning of regclass()
12595 * that checks to see in a normal bracketed class if the surrounding []
12596 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
12597 * change long-standing behavior, so I (khw) didn't do that */
12598 char* p = RExC_parse + 1;
12599 char first_char = *p;
12601 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
12603 assert(*(p - 1) == '[');
12605 if (! POSIXCC(first_char)) {
12610 while (p < RExC_end && isWORDCHAR(*p)) p++;
12612 if (p >= RExC_end) {
12616 if (p - RExC_parse > 2 /* Got at least 1 word character */
12617 && (*p == first_char
12618 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
12623 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
12626 && p - RExC_parse > 2 /* [:] evaluates to colon;
12627 [::] is a bad posix class. */
12628 && first_char == *(p - 1));
12632 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
12633 I32 *flagp, U32 depth,
12634 char * const oregcomp_parse)
12636 /* Handle the (?[...]) construct to do set operations */
12639 UV start, end; /* End points of code point ranges */
12641 char *save_end, *save_parse;
12646 const bool save_fold = FOLD;
12648 GET_RE_DEBUG_FLAGS_DECL;
12650 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
12653 vFAIL("(?[...]) not valid in locale");
12655 RExC_uni_semantics = 1;
12657 /* This will return only an ANYOF regnode, or (unlikely) something smaller
12658 * (such as EXACT). Thus we can skip most everything if just sizing. We
12659 * call regclass to handle '[]' so as to not have to reinvent its parsing
12660 * rules here (throwing away the size it computes each time). And, we exit
12661 * upon an unescaped ']' that isn't one ending a regclass. To do both
12662 * these things, we need to realize that something preceded by a backslash
12663 * is escaped, so we have to keep track of backslashes */
12665 UV depth = 0; /* how many nested (?[...]) constructs */
12667 Perl_ck_warner_d(aTHX_
12668 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
12669 "The regex_sets feature is experimental" REPORT_LOCATION,
12670 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
12672 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
12673 RExC_precomp + (RExC_parse - RExC_precomp)));
12675 while (RExC_parse < RExC_end) {
12676 SV* current = NULL;
12677 RExC_parse = regpatws(pRExC_state, RExC_parse,
12678 TRUE); /* means recognize comments */
12679 switch (*RExC_parse) {
12681 if (RExC_parse[1] == '[') depth++, RExC_parse++;
12686 /* Skip the next byte (which could cause us to end up in
12687 * the middle of a UTF-8 character, but since none of those
12688 * are confusable with anything we currently handle in this
12689 * switch (invariants all), it's safe. We'll just hit the
12690 * default: case next time and keep on incrementing until
12691 * we find one of the invariants we do handle. */
12696 /* If this looks like it is a [:posix:] class, leave the
12697 * parse pointer at the '[' to fool regclass() into
12698 * thinking it is part of a '[[:posix:]]'. That function
12699 * will use strict checking to force a syntax error if it
12700 * doesn't work out to a legitimate class */
12701 bool is_posix_class
12702 = could_it_be_a_POSIX_class(pRExC_state);
12703 if (! is_posix_class) {
12707 /* regclass() can only return RESTART_UTF8 if multi-char
12708 folds are allowed. */
12709 if (!regclass(pRExC_state, flagp,depth+1,
12710 is_posix_class, /* parse the whole char
12711 class only if not a
12713 FALSE, /* don't allow multi-char folds */
12714 TRUE, /* silence non-portable warnings. */
12716 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12719 /* function call leaves parse pointing to the ']', except
12720 * if we faked it */
12721 if (is_posix_class) {
12725 SvREFCNT_dec(current); /* In case it returned something */
12730 if (depth--) break;
12732 if (RExC_parse < RExC_end
12733 && *RExC_parse == ')')
12735 node = reganode(pRExC_state, ANYOF, 0);
12736 RExC_size += ANYOF_SKIP;
12737 nextchar(pRExC_state);
12738 Set_Node_Length(node,
12739 RExC_parse - oregcomp_parse + 1); /* MJD */
12748 FAIL("Syntax error in (?[...])");
12751 /* Pass 2 only after this. Everything in this construct is a
12752 * metacharacter. Operands begin with either a '\' (for an escape
12753 * sequence), or a '[' for a bracketed character class. Any other
12754 * character should be an operator, or parenthesis for grouping. Both
12755 * types of operands are handled by calling regclass() to parse them. It
12756 * is called with a parameter to indicate to return the computed inversion
12757 * list. The parsing here is implemented via a stack. Each entry on the
12758 * stack is a single character representing one of the operators, or the
12759 * '('; or else a pointer to an operand inversion list. */
12761 #define IS_OPERAND(a) (! SvIOK(a))
12763 /* The stack starts empty. It is a syntax error if the first thing parsed
12764 * is a binary operator; everything else is pushed on the stack. When an
12765 * operand is parsed, the top of the stack is examined. If it is a binary
12766 * operator, the item before it should be an operand, and both are replaced
12767 * by the result of doing that operation on the new operand and the one on
12768 * the stack. Thus a sequence of binary operands is reduced to a single
12769 * one before the next one is parsed.
12771 * A unary operator may immediately follow a binary in the input, for
12774 * When an operand is parsed and the top of the stack is a unary operator,
12775 * the operation is performed, and then the stack is rechecked to see if
12776 * this new operand is part of a binary operation; if so, it is handled as
12779 * A '(' is simply pushed on the stack; it is valid only if the stack is
12780 * empty, or the top element of the stack is an operator or another '('
12781 * (for which the parenthesized expression will become an operand). By the
12782 * time the corresponding ')' is parsed everything in between should have
12783 * been parsed and evaluated to a single operand (or else is a syntax
12784 * error), and is handled as a regular operand */
12786 sv_2mortal((SV *)(stack = newAV()));
12788 while (RExC_parse < RExC_end) {
12789 I32 top_index = av_tindex(stack);
12791 SV* current = NULL;
12793 /* Skip white space */
12794 RExC_parse = regpatws(pRExC_state, RExC_parse,
12795 TRUE); /* means recognize comments */
12796 if (RExC_parse >= RExC_end) {
12797 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
12799 if ((curchar = UCHARAT(RExC_parse)) == ']') {
12806 if (av_tindex(stack) >= 0 /* This makes sure that we can
12807 safely subtract 1 from
12808 RExC_parse in the next clause.
12809 If we have something on the
12810 stack, we have parsed something
12812 && UCHARAT(RExC_parse - 1) == '('
12813 && RExC_parse < RExC_end)
12815 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
12816 * This happens when we have some thing like
12818 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
12820 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
12822 * Here we would be handling the interpolated
12823 * '$thai_or_lao'. We handle this by a recursive call to
12824 * ourselves which returns the inversion list the
12825 * interpolated expression evaluates to. We use the flags
12826 * from the interpolated pattern. */
12827 U32 save_flags = RExC_flags;
12828 const char * const save_parse = ++RExC_parse;
12830 parse_lparen_question_flags(pRExC_state);
12832 if (RExC_parse == save_parse /* Makes sure there was at
12833 least one flag (or this
12834 embedding wasn't compiled)
12836 || RExC_parse >= RExC_end - 4
12837 || UCHARAT(RExC_parse) != ':'
12838 || UCHARAT(++RExC_parse) != '('
12839 || UCHARAT(++RExC_parse) != '?'
12840 || UCHARAT(++RExC_parse) != '[')
12843 /* In combination with the above, this moves the
12844 * pointer to the point just after the first erroneous
12845 * character (or if there are no flags, to where they
12846 * should have been) */
12847 if (RExC_parse >= RExC_end - 4) {
12848 RExC_parse = RExC_end;
12850 else if (RExC_parse != save_parse) {
12851 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12853 vFAIL("Expecting '(?flags:(?[...'");
12856 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
12857 depth+1, oregcomp_parse);
12859 /* Here, 'current' contains the embedded expression's
12860 * inversion list, and RExC_parse points to the trailing
12861 * ']'; the next character should be the ')' which will be
12862 * paired with the '(' that has been put on the stack, so
12863 * the whole embedded expression reduces to '(operand)' */
12866 RExC_flags = save_flags;
12867 goto handle_operand;
12872 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12873 vFAIL("Unexpected character");
12876 /* regclass() can only return RESTART_UTF8 if multi-char
12877 folds are allowed. */
12878 if (!regclass(pRExC_state, flagp,depth+1,
12879 TRUE, /* means parse just the next thing */
12880 FALSE, /* don't allow multi-char folds */
12881 FALSE, /* don't silence non-portable warnings. */
12883 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12885 /* regclass() will return with parsing just the \ sequence,
12886 * leaving the parse pointer at the next thing to parse */
12888 goto handle_operand;
12890 case '[': /* Is a bracketed character class */
12892 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
12894 if (! is_posix_class) {
12898 /* regclass() can only return RESTART_UTF8 if multi-char
12899 folds are allowed. */
12900 if(!regclass(pRExC_state, flagp,depth+1,
12901 is_posix_class, /* parse the whole char class
12902 only if not a posix class */
12903 FALSE, /* don't allow multi-char folds */
12904 FALSE, /* don't silence non-portable warnings. */
12906 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12908 /* function call leaves parse pointing to the ']', except if we
12910 if (is_posix_class) {
12914 goto handle_operand;
12923 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
12924 || ! IS_OPERAND(*top_ptr))
12927 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
12929 av_push(stack, newSVuv(curchar));
12933 av_push(stack, newSVuv(curchar));
12937 if (top_index >= 0) {
12938 top_ptr = av_fetch(stack, top_index, FALSE);
12940 if (IS_OPERAND(*top_ptr)) {
12942 vFAIL("Unexpected '(' with no preceding operator");
12945 av_push(stack, newSVuv(curchar));
12952 || ! (current = av_pop(stack))
12953 || ! IS_OPERAND(current)
12954 || ! (lparen = av_pop(stack))
12955 || IS_OPERAND(lparen)
12956 || SvUV(lparen) != '(')
12958 SvREFCNT_dec(current);
12960 vFAIL("Unexpected ')'");
12963 SvREFCNT_dec_NN(lparen);
12970 /* Here, we have an operand to process, in 'current' */
12972 if (top_index < 0) { /* Just push if stack is empty */
12973 av_push(stack, current);
12976 SV* top = av_pop(stack);
12978 char current_operator;
12980 if (IS_OPERAND(top)) {
12981 SvREFCNT_dec_NN(top);
12982 SvREFCNT_dec_NN(current);
12983 vFAIL("Operand with no preceding operator");
12985 current_operator = (char) SvUV(top);
12986 switch (current_operator) {
12987 case '(': /* Push the '(' back on followed by the new
12989 av_push(stack, top);
12990 av_push(stack, current);
12991 SvREFCNT_inc(top); /* Counters the '_dec' done
12992 just after the 'break', so
12993 it doesn't get wrongly freed
12998 _invlist_invert(current);
13000 /* Unlike binary operators, the top of the stack,
13001 * now that this unary one has been popped off, may
13002 * legally be an operator, and we now have operand
13005 SvREFCNT_dec_NN(top);
13006 goto handle_operand;
13009 prev = av_pop(stack);
13010 _invlist_intersection(prev,
13013 av_push(stack, current);
13018 prev = av_pop(stack);
13019 _invlist_union(prev, current, ¤t);
13020 av_push(stack, current);
13024 prev = av_pop(stack);;
13025 _invlist_subtract(prev, current, ¤t);
13026 av_push(stack, current);
13029 case '^': /* The union minus the intersection */
13035 prev = av_pop(stack);
13036 _invlist_union(prev, current, &u);
13037 _invlist_intersection(prev, current, &i);
13038 /* _invlist_subtract will overwrite current
13039 without freeing what it already contains */
13041 _invlist_subtract(u, i, ¤t);
13042 av_push(stack, current);
13043 SvREFCNT_dec_NN(i);
13044 SvREFCNT_dec_NN(u);
13045 SvREFCNT_dec_NN(element);
13050 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13052 SvREFCNT_dec_NN(top);
13053 SvREFCNT_dec(prev);
13057 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13060 if (av_tindex(stack) < 0 /* Was empty */
13061 || ((final = av_pop(stack)) == NULL)
13062 || ! IS_OPERAND(final)
13063 || av_tindex(stack) >= 0) /* More left on stack */
13065 vFAIL("Incomplete expression within '(?[ ])'");
13068 /* Here, 'final' is the resultant inversion list from evaluating the
13069 * expression. Return it if so requested */
13070 if (return_invlist) {
13071 *return_invlist = final;
13075 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13076 * expecting a string of ranges and individual code points */
13077 invlist_iterinit(final);
13078 result_string = newSVpvs("");
13079 while (invlist_iternext(final, &start, &end)) {
13080 if (start == end) {
13081 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13084 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13089 save_parse = RExC_parse;
13090 RExC_parse = SvPV(result_string, len);
13091 save_end = RExC_end;
13092 RExC_end = RExC_parse + len;
13094 /* We turn off folding around the call, as the class we have constructed
13095 * already has all folding taken into consideration, and we don't want
13096 * regclass() to add to that */
13097 RExC_flags &= ~RXf_PMf_FOLD;
13098 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13100 node = regclass(pRExC_state, flagp,depth+1,
13101 FALSE, /* means parse the whole char class */
13102 FALSE, /* don't allow multi-char folds */
13103 TRUE, /* silence non-portable warnings. The above may very
13104 well have generated non-portable code points, but
13105 they're valid on this machine */
13108 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13111 RExC_flags |= RXf_PMf_FOLD;
13113 RExC_parse = save_parse + 1;
13114 RExC_end = save_end;
13115 SvREFCNT_dec_NN(final);
13116 SvREFCNT_dec_NN(result_string);
13118 nextchar(pRExC_state);
13119 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13124 /* The names of properties whose definitions are not known at compile time are
13125 * stored in this SV, after a constant heading. So if the length has been
13126 * changed since initialization, then there is a run-time definition. */
13127 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13128 (SvCUR(listsv) != initial_listsv_len)
13131 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13132 const bool stop_at_1, /* Just parse the next thing, don't
13133 look for a full character class */
13134 bool allow_multi_folds,
13135 const bool silence_non_portable, /* Don't output warnings
13138 SV** ret_invlist) /* Return an inversion list, not a node */
13140 /* parse a bracketed class specification. Most of these will produce an
13141 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13142 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13143 * under /i with multi-character folds: it will be rewritten following the
13144 * paradigm of this example, where the <multi-fold>s are characters which
13145 * fold to multiple character sequences:
13146 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13147 * gets effectively rewritten as:
13148 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13149 * reg() gets called (recursively) on the rewritten version, and this
13150 * function will return what it constructs. (Actually the <multi-fold>s
13151 * aren't physically removed from the [abcdefghi], it's just that they are
13152 * ignored in the recursion by means of a flag:
13153 * <RExC_in_multi_char_class>.)
13155 * ANYOF nodes contain a bit map for the first 256 characters, with the
13156 * corresponding bit set if that character is in the list. For characters
13157 * above 255, a range list or swash is used. There are extra bits for \w,
13158 * etc. in locale ANYOFs, as what these match is not determinable at
13161 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13162 * to be restarted. This can only happen if ret_invlist is non-NULL.
13166 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13168 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13171 IV namedclass = OOB_NAMEDCLASS;
13172 char *rangebegin = NULL;
13173 bool need_class = 0;
13175 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13176 than just initialized. */
13177 SV* properties = NULL; /* Code points that match \p{} \P{} */
13178 SV* posixes = NULL; /* Code points that match classes like [:word:],
13179 extended beyond the Latin1 range. These have to
13180 be kept separate from other code points for much
13181 of this function because their handling is
13182 different under /i, and for most classes under
13184 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13185 separate for a while from the non-complemented
13186 versions because of complications with /d
13188 UV element_count = 0; /* Number of distinct elements in the class.
13189 Optimizations may be possible if this is tiny */
13190 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13191 character; used under /i */
13193 char * stop_ptr = RExC_end; /* where to stop parsing */
13194 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13196 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13198 /* Unicode properties are stored in a swash; this holds the current one
13199 * being parsed. If this swash is the only above-latin1 component of the
13200 * character class, an optimization is to pass it directly on to the
13201 * execution engine. Otherwise, it is set to NULL to indicate that there
13202 * are other things in the class that have to be dealt with at execution
13204 SV* swash = NULL; /* Code points that match \p{} \P{} */
13206 /* Set if a component of this character class is user-defined; just passed
13207 * on to the engine */
13208 bool has_user_defined_property = FALSE;
13210 /* inversion list of code points this node matches only when the target
13211 * string is in UTF-8. (Because is under /d) */
13212 SV* depends_list = NULL;
13214 /* Inversion list of code points this node matches regardless of things
13215 * like locale, folding, utf8ness of the target string */
13216 SV* cp_list = NULL;
13218 /* Like cp_list, but code points on this list need to be checked for things
13219 * that fold to/from them under /i */
13220 SV* cp_foldable_list = NULL;
13222 /* Like cp_list, but code points on this list are valid only when the
13223 * runtime locale is UTF-8 */
13224 SV* only_utf8_locale_list = NULL;
13227 /* In a range, counts how many 0-2 of the ends of it came from literals,
13228 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13229 UV literal_endpoint = 0;
13231 bool invert = FALSE; /* Is this class to be complemented */
13233 bool warn_super = ALWAYS_WARN_SUPER;
13235 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13236 case we need to change the emitted regop to an EXACT. */
13237 const char * orig_parse = RExC_parse;
13238 const SSize_t orig_size = RExC_size;
13239 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13240 GET_RE_DEBUG_FLAGS_DECL;
13242 PERL_ARGS_ASSERT_REGCLASS;
13244 PERL_UNUSED_ARG(depth);
13247 DEBUG_PARSE("clas");
13249 /* Assume we are going to generate an ANYOF node. */
13250 ret = reganode(pRExC_state, ANYOF, 0);
13253 RExC_size += ANYOF_SKIP;
13254 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13257 ANYOF_FLAGS(ret) = 0;
13259 RExC_emit += ANYOF_SKIP;
13260 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13261 initial_listsv_len = SvCUR(listsv);
13262 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13266 RExC_parse = regpatws(pRExC_state, RExC_parse,
13267 FALSE /* means don't recognize comments */);
13270 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13273 allow_multi_folds = FALSE;
13276 RExC_parse = regpatws(pRExC_state, RExC_parse,
13277 FALSE /* means don't recognize comments */);
13281 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13282 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13283 const char *s = RExC_parse;
13284 const char c = *s++;
13286 while (isWORDCHAR(*s))
13288 if (*s && c == *s && s[1] == ']') {
13289 SAVEFREESV(RExC_rx_sv);
13291 "POSIX syntax [%c %c] belongs inside character classes",
13293 (void)ReREFCNT_inc(RExC_rx_sv);
13297 /* If the caller wants us to just parse a single element, accomplish this
13298 * by faking the loop ending condition */
13299 if (stop_at_1 && RExC_end > RExC_parse) {
13300 stop_ptr = RExC_parse + 1;
13303 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13304 if (UCHARAT(RExC_parse) == ']')
13305 goto charclassloop;
13309 if (RExC_parse >= stop_ptr) {
13314 RExC_parse = regpatws(pRExC_state, RExC_parse,
13315 FALSE /* means don't recognize comments */);
13318 if (UCHARAT(RExC_parse) == ']') {
13324 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13325 save_value = value;
13326 save_prevvalue = prevvalue;
13329 rangebegin = RExC_parse;
13333 value = utf8n_to_uvchr((U8*)RExC_parse,
13334 RExC_end - RExC_parse,
13335 &numlen, UTF8_ALLOW_DEFAULT);
13336 RExC_parse += numlen;
13339 value = UCHARAT(RExC_parse++);
13342 && RExC_parse < RExC_end
13343 && POSIXCC(UCHARAT(RExC_parse)))
13345 namedclass = regpposixcc(pRExC_state, value, strict);
13347 else if (value == '\\') {
13349 value = utf8n_to_uvchr((U8*)RExC_parse,
13350 RExC_end - RExC_parse,
13351 &numlen, UTF8_ALLOW_DEFAULT);
13352 RExC_parse += numlen;
13355 value = UCHARAT(RExC_parse++);
13357 /* Some compilers cannot handle switching on 64-bit integer
13358 * values, therefore value cannot be an UV. Yes, this will
13359 * be a problem later if we want switch on Unicode.
13360 * A similar issue a little bit later when switching on
13361 * namedclass. --jhi */
13363 /* If the \ is escaping white space when white space is being
13364 * skipped, it means that that white space is wanted literally, and
13365 * is already in 'value'. Otherwise, need to translate the escape
13366 * into what it signifies. */
13367 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13369 case 'w': namedclass = ANYOF_WORDCHAR; break;
13370 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13371 case 's': namedclass = ANYOF_SPACE; break;
13372 case 'S': namedclass = ANYOF_NSPACE; break;
13373 case 'd': namedclass = ANYOF_DIGIT; break;
13374 case 'D': namedclass = ANYOF_NDIGIT; break;
13375 case 'v': namedclass = ANYOF_VERTWS; break;
13376 case 'V': namedclass = ANYOF_NVERTWS; break;
13377 case 'h': namedclass = ANYOF_HORIZWS; break;
13378 case 'H': namedclass = ANYOF_NHORIZWS; break;
13379 case 'N': /* Handle \N{NAME} in class */
13381 /* We only pay attention to the first char of
13382 multichar strings being returned. I kinda wonder
13383 if this makes sense as it does change the behaviour
13384 from earlier versions, OTOH that behaviour was broken
13386 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
13387 TRUE, /* => charclass */
13390 if (*flagp & RESTART_UTF8)
13391 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13401 /* We will handle any undefined properties ourselves */
13402 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
13403 /* And we actually would prefer to get
13404 * the straight inversion list of the
13405 * swash, since we will be accessing it
13406 * anyway, to save a little time */
13407 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
13409 if (RExC_parse >= RExC_end)
13410 vFAIL2("Empty \\%c{}", (U8)value);
13411 if (*RExC_parse == '{') {
13412 const U8 c = (U8)value;
13413 e = strchr(RExC_parse++, '}');
13415 vFAIL2("Missing right brace on \\%c{}", c);
13416 while (isSPACE(UCHARAT(RExC_parse)))
13418 if (e == RExC_parse)
13419 vFAIL2("Empty \\%c{}", c);
13420 n = e - RExC_parse;
13421 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
13433 if (UCHARAT(RExC_parse) == '^') {
13436 /* toggle. (The rhs xor gets the single bit that
13437 * differs between P and p; the other xor inverts just
13439 value ^= 'P' ^ 'p';
13441 while (isSPACE(UCHARAT(RExC_parse))) {
13446 /* Try to get the definition of the property into
13447 * <invlist>. If /i is in effect, the effective property
13448 * will have its name be <__NAME_i>. The design is
13449 * discussed in commit
13450 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
13451 formatted = Perl_form(aTHX_
13453 (FOLD) ? "__" : "",
13458 name = savepvn(formatted, strlen(formatted));
13460 /* Look up the property name, and get its swash and
13461 * inversion list, if the property is found */
13463 SvREFCNT_dec_NN(swash);
13465 swash = _core_swash_init("utf8", name, &PL_sv_undef,
13468 NULL, /* No inversion list */
13471 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
13473 SvREFCNT_dec_NN(swash);
13477 /* Here didn't find it. It could be a user-defined
13478 * property that will be available at run-time. If we
13479 * accept only compile-time properties, is an error;
13480 * otherwise add it to the list for run-time look up */
13482 RExC_parse = e + 1;
13484 "Property '%"UTF8f"' is unknown",
13485 UTF8fARG(UTF, n, name));
13487 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
13488 (value == 'p' ? '+' : '!'),
13489 UTF8fARG(UTF, n, name));
13490 has_user_defined_property = TRUE;
13492 /* We don't know yet, so have to assume that the
13493 * property could match something in the Latin1 range,
13494 * hence something that isn't utf8. Note that this
13495 * would cause things in <depends_list> to match
13496 * inappropriately, except that any \p{}, including
13497 * this one forces Unicode semantics, which means there
13498 * is no <depends_list> */
13499 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
13503 /* Here, did get the swash and its inversion list. If
13504 * the swash is from a user-defined property, then this
13505 * whole character class should be regarded as such */
13506 if (swash_init_flags
13507 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
13509 has_user_defined_property = TRUE;
13512 /* We warn on matching an above-Unicode code point
13513 * if the match would return true, except don't
13514 * warn for \p{All}, which has exactly one element
13516 (_invlist_contains_cp(invlist, 0x110000)
13517 && (! (_invlist_len(invlist) == 1
13518 && *invlist_array(invlist) == 0)))
13524 /* Invert if asking for the complement */
13525 if (value == 'P') {
13526 _invlist_union_complement_2nd(properties,
13530 /* The swash can't be used as-is, because we've
13531 * inverted things; delay removing it to here after
13532 * have copied its invlist above */
13533 SvREFCNT_dec_NN(swash);
13537 _invlist_union(properties, invlist, &properties);
13542 RExC_parse = e + 1;
13543 namedclass = ANYOF_UNIPROP; /* no official name, but it's
13546 /* \p means they want Unicode semantics */
13547 RExC_uni_semantics = 1;
13550 case 'n': value = '\n'; break;
13551 case 'r': value = '\r'; break;
13552 case 't': value = '\t'; break;
13553 case 'f': value = '\f'; break;
13554 case 'b': value = '\b'; break;
13555 case 'e': value = ASCII_TO_NATIVE('\033');break;
13556 case 'a': value = '\a'; break;
13558 RExC_parse--; /* function expects to be pointed at the 'o' */
13560 const char* error_msg;
13561 bool valid = grok_bslash_o(&RExC_parse,
13564 SIZE_ONLY, /* warnings in pass
13567 silence_non_portable,
13573 if (PL_encoding && value < 0x100) {
13574 goto recode_encoding;
13578 RExC_parse--; /* function expects to be pointed at the 'x' */
13580 const char* error_msg;
13581 bool valid = grok_bslash_x(&RExC_parse,
13584 TRUE, /* Output warnings */
13586 silence_non_portable,
13592 if (PL_encoding && value < 0x100)
13593 goto recode_encoding;
13596 value = grok_bslash_c(*RExC_parse++, SIZE_ONLY);
13598 case '0': case '1': case '2': case '3': case '4':
13599 case '5': case '6': case '7':
13601 /* Take 1-3 octal digits */
13602 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13603 numlen = (strict) ? 4 : 3;
13604 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
13605 RExC_parse += numlen;
13608 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13609 vFAIL("Need exactly 3 octal digits");
13611 else if (! SIZE_ONLY /* like \08, \178 */
13613 && RExC_parse < RExC_end
13614 && isDIGIT(*RExC_parse)
13615 && ckWARN(WARN_REGEXP))
13617 SAVEFREESV(RExC_rx_sv);
13618 reg_warn_non_literal_string(
13620 form_short_octal_warning(RExC_parse, numlen));
13621 (void)ReREFCNT_inc(RExC_rx_sv);
13624 if (PL_encoding && value < 0x100)
13625 goto recode_encoding;
13629 if (! RExC_override_recoding) {
13630 SV* enc = PL_encoding;
13631 value = reg_recode((const char)(U8)value, &enc);
13634 vFAIL("Invalid escape in the specified encoding");
13636 else if (SIZE_ONLY) {
13637 ckWARNreg(RExC_parse,
13638 "Invalid escape in the specified encoding");
13644 /* Allow \_ to not give an error */
13645 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
13647 vFAIL2("Unrecognized escape \\%c in character class",
13651 SAVEFREESV(RExC_rx_sv);
13652 ckWARN2reg(RExC_parse,
13653 "Unrecognized escape \\%c in character class passed through",
13655 (void)ReREFCNT_inc(RExC_rx_sv);
13659 } /* End of switch on char following backslash */
13660 } /* end of handling backslash escape sequences */
13663 literal_endpoint++;
13666 /* Here, we have the current token in 'value' */
13668 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
13671 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
13672 * literal, as is the character that began the false range, i.e.
13673 * the 'a' in the examples */
13676 const int w = (RExC_parse >= rangebegin)
13677 ? RExC_parse - rangebegin
13681 "False [] range \"%"UTF8f"\"",
13682 UTF8fARG(UTF, w, rangebegin));
13685 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
13686 ckWARN2reg(RExC_parse,
13687 "False [] range \"%"UTF8f"\"",
13688 UTF8fARG(UTF, w, rangebegin));
13689 (void)ReREFCNT_inc(RExC_rx_sv);
13690 cp_list = add_cp_to_invlist(cp_list, '-');
13691 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
13696 range = 0; /* this was not a true range */
13697 element_count += 2; /* So counts for three values */
13700 classnum = namedclass_to_classnum(namedclass);
13702 if (LOC && namedclass < ANYOF_POSIXL_MAX
13703 #ifndef HAS_ISASCII
13704 && classnum != _CC_ASCII
13707 /* What the Posix classes (like \w, [:space:]) match in locale
13708 * isn't knowable under locale until actual match time. Room
13709 * must be reserved (one time per outer bracketed class) to
13710 * store such classes. The space will contain a bit for each
13711 * named class that is to be matched against. This isn't
13712 * needed for \p{} and pseudo-classes, as they are not affected
13713 * by locale, and hence are dealt with separately */
13714 if (! need_class) {
13717 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13720 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13722 ANYOF_FLAGS(ret) |= ANYOF_POSIXL;
13723 ANYOF_POSIXL_ZERO(ret);
13726 /* See if it already matches the complement of this POSIX
13728 if ((ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13729 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
13733 posixl_matches_all = TRUE;
13734 break; /* No need to continue. Since it matches both
13735 e.g., \w and \W, it matches everything, and the
13736 bracketed class can be optimized into qr/./s */
13739 /* Add this class to those that should be checked at runtime */
13740 ANYOF_POSIXL_SET(ret, namedclass);
13742 /* The above-Latin1 characters are not subject to locale rules.
13743 * Just add them, in the second pass, to the
13744 * unconditionally-matched list */
13746 SV* scratch_list = NULL;
13748 /* Get the list of the above-Latin1 code points this
13750 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
13751 PL_XPosix_ptrs[classnum],
13753 /* Odd numbers are complements, like
13754 * NDIGIT, NASCII, ... */
13755 namedclass % 2 != 0,
13757 /* Checking if 'cp_list' is NULL first saves an extra
13758 * clone. Its reference count will be decremented at the
13759 * next union, etc, or if this is the only instance, at the
13760 * end of the routine */
13762 cp_list = scratch_list;
13765 _invlist_union(cp_list, scratch_list, &cp_list);
13766 SvREFCNT_dec_NN(scratch_list);
13768 continue; /* Go get next character */
13771 else if (! SIZE_ONLY) {
13773 /* Here, not in pass1 (in that pass we skip calculating the
13774 * contents of this class), and is /l, or is a POSIX class for
13775 * which /l doesn't matter (or is a Unicode property, which is
13776 * skipped here). */
13777 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
13778 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
13780 /* Here, should be \h, \H, \v, or \V. None of /d, /i
13781 * nor /l make a difference in what these match,
13782 * therefore we just add what they match to cp_list. */
13783 if (classnum != _CC_VERTSPACE) {
13784 assert( namedclass == ANYOF_HORIZWS
13785 || namedclass == ANYOF_NHORIZWS);
13787 /* It turns out that \h is just a synonym for
13789 classnum = _CC_BLANK;
13792 _invlist_union_maybe_complement_2nd(
13794 PL_XPosix_ptrs[classnum],
13795 namedclass % 2 != 0, /* Complement if odd
13796 (NHORIZWS, NVERTWS)
13801 else { /* Garden variety class. If is NASCII, NDIGIT, ...
13802 complement and use nposixes */
13803 SV** posixes_ptr = namedclass % 2 == 0
13806 SV** source_ptr = &PL_XPosix_ptrs[classnum];
13807 _invlist_union_maybe_complement_2nd(
13810 namedclass % 2 != 0,
13813 continue; /* Go get next character */
13815 } /* end of namedclass \blah */
13817 /* Here, we have a single value. If 'range' is set, it is the ending
13818 * of a range--check its validity. Later, we will handle each
13819 * individual code point in the range. If 'range' isn't set, this
13820 * could be the beginning of a range, so check for that by looking
13821 * ahead to see if the next real character to be processed is the range
13822 * indicator--the minus sign */
13825 RExC_parse = regpatws(pRExC_state, RExC_parse,
13826 FALSE /* means don't recognize comments */);
13830 if (prevvalue > value) /* b-a */ {
13831 const int w = RExC_parse - rangebegin;
13833 "Invalid [] range \"%"UTF8f"\"",
13834 UTF8fARG(UTF, w, rangebegin));
13835 range = 0; /* not a valid range */
13839 prevvalue = value; /* save the beginning of the potential range */
13840 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13841 && *RExC_parse == '-')
13843 char* next_char_ptr = RExC_parse + 1;
13844 if (skip_white) { /* Get the next real char after the '-' */
13845 next_char_ptr = regpatws(pRExC_state,
13847 FALSE); /* means don't recognize
13851 /* If the '-' is at the end of the class (just before the ']',
13852 * it is a literal minus; otherwise it is a range */
13853 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13854 RExC_parse = next_char_ptr;
13856 /* a bad range like \w-, [:word:]- ? */
13857 if (namedclass > OOB_NAMEDCLASS) {
13858 if (strict || ckWARN(WARN_REGEXP)) {
13860 RExC_parse >= rangebegin ?
13861 RExC_parse - rangebegin : 0;
13863 vFAIL4("False [] range \"%*.*s\"",
13868 "False [] range \"%*.*s\"",
13873 cp_list = add_cp_to_invlist(cp_list, '-');
13877 range = 1; /* yeah, it's a range! */
13878 continue; /* but do it the next time */
13883 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13886 /* non-Latin1 code point implies unicode semantics. Must be set in
13887 * pass1 so is there for the whole of pass 2 */
13889 RExC_uni_semantics = 1;
13892 /* Ready to process either the single value, or the completed range.
13893 * For single-valued non-inverted ranges, we consider the possibility
13894 * of multi-char folds. (We made a conscious decision to not do this
13895 * for the other cases because it can often lead to non-intuitive
13896 * results. For example, you have the peculiar case that:
13897 * "s s" =~ /^[^\xDF]+$/i => Y
13898 * "ss" =~ /^[^\xDF]+$/i => N
13900 * See [perl #89750] */
13901 if (FOLD && allow_multi_folds && value == prevvalue) {
13902 if (value == LATIN_SMALL_LETTER_SHARP_S
13903 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13906 /* Here <value> is indeed a multi-char fold. Get what it is */
13908 U8 foldbuf[UTF8_MAXBYTES_CASE];
13911 UV folded = _to_uni_fold_flags(
13915 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
13916 ? FOLD_FLAGS_NOMIX_ASCII
13920 /* Here, <folded> should be the first character of the
13921 * multi-char fold of <value>, with <foldbuf> containing the
13922 * whole thing. But, if this fold is not allowed (because of
13923 * the flags), <fold> will be the same as <value>, and should
13924 * be processed like any other character, so skip the special
13926 if (folded != value) {
13928 /* Skip if we are recursed, currently parsing the class
13929 * again. Otherwise add this character to the list of
13930 * multi-char folds. */
13931 if (! RExC_in_multi_char_class) {
13932 AV** this_array_ptr;
13934 STRLEN cp_count = utf8_length(foldbuf,
13935 foldbuf + foldlen);
13936 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13938 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13941 if (! multi_char_matches) {
13942 multi_char_matches = newAV();
13945 /* <multi_char_matches> is actually an array of arrays.
13946 * There will be one or two top-level elements: [2],
13947 * and/or [3]. The [2] element is an array, each
13948 * element thereof is a character which folds to TWO
13949 * characters; [3] is for folds to THREE characters.
13950 * (Unicode guarantees a maximum of 3 characters in any
13951 * fold.) When we rewrite the character class below,
13952 * we will do so such that the longest folds are
13953 * written first, so that it prefers the longest
13954 * matching strings first. This is done even if it
13955 * turns out that any quantifier is non-greedy, out of
13956 * programmer laziness. Tom Christiansen has agreed
13957 * that this is ok. This makes the test for the
13958 * ligature 'ffi' come before the test for 'ff' */
13959 if (av_exists(multi_char_matches, cp_count)) {
13960 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13962 this_array = *this_array_ptr;
13965 this_array = newAV();
13966 av_store(multi_char_matches, cp_count,
13969 av_push(this_array, multi_fold);
13972 /* This element should not be processed further in this
13975 value = save_value;
13976 prevvalue = save_prevvalue;
13982 /* Deal with this element of the class */
13985 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
13988 SV* this_range = _new_invlist(1);
13989 _append_range_to_invlist(this_range, prevvalue, value);
13991 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13992 * If this range was specified using something like 'i-j', we want
13993 * to include only the 'i' and the 'j', and not anything in
13994 * between, so exclude non-ASCII, non-alphabetics from it.
13995 * However, if the range was specified with something like
13996 * [\x89-\x91] or [\x89-j], all code points within it should be
13997 * included. literal_endpoint==2 means both ends of the range used
13998 * a literal character, not \x{foo} */
13999 if (literal_endpoint == 2
14000 && ((prevvalue >= 'a' && value <= 'z')
14001 || (prevvalue >= 'A' && value <= 'Z')))
14003 _invlist_intersection(this_range, PL_ASCII,
14006 /* Since this above only contains ascii, the intersection of it
14007 * with anything will still yield only ascii */
14008 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ALPHA],
14011 _invlist_union(cp_foldable_list, this_range, &cp_foldable_list);
14012 literal_endpoint = 0;
14016 range = 0; /* this range (if it was one) is done now */
14017 } /* End of loop through all the text within the brackets */
14019 /* If anything in the class expands to more than one character, we have to
14020 * deal with them by building up a substitute parse string, and recursively
14021 * calling reg() on it, instead of proceeding */
14022 if (multi_char_matches) {
14023 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14026 char *save_end = RExC_end;
14027 char *save_parse = RExC_parse;
14028 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14033 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14034 because too confusing */
14036 sv_catpv(substitute_parse, "(?:");
14040 /* Look at the longest folds first */
14041 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14043 if (av_exists(multi_char_matches, cp_count)) {
14044 AV** this_array_ptr;
14047 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14049 while ((this_sequence = av_pop(*this_array_ptr)) !=
14052 if (! first_time) {
14053 sv_catpv(substitute_parse, "|");
14055 first_time = FALSE;
14057 sv_catpv(substitute_parse, SvPVX(this_sequence));
14062 /* If the character class contains anything else besides these
14063 * multi-character folds, have to include it in recursive parsing */
14064 if (element_count) {
14065 sv_catpv(substitute_parse, "|[");
14066 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14067 sv_catpv(substitute_parse, "]");
14070 sv_catpv(substitute_parse, ")");
14073 /* This is a way to get the parse to skip forward a whole named
14074 * sequence instead of matching the 2nd character when it fails the
14076 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14080 RExC_parse = SvPV(substitute_parse, len);
14081 RExC_end = RExC_parse + len;
14082 RExC_in_multi_char_class = 1;
14083 RExC_emit = (regnode *)orig_emit;
14085 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14087 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14089 RExC_parse = save_parse;
14090 RExC_end = save_end;
14091 RExC_in_multi_char_class = 0;
14092 SvREFCNT_dec_NN(multi_char_matches);
14096 /* Here, we've gone through the entire class and dealt with multi-char
14097 * folds. We are now in a position that we can do some checks to see if we
14098 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14099 * Currently we only do two checks:
14100 * 1) is in the unlikely event that the user has specified both, eg. \w and
14101 * \W under /l, then the class matches everything. (This optimization
14102 * is done only to make the optimizer code run later work.)
14103 * 2) if the character class contains only a single element (including a
14104 * single range), we see if there is an equivalent node for it.
14105 * Other checks are possible */
14106 if (! ret_invlist /* Can't optimize if returning the constructed
14108 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14113 if (UNLIKELY(posixl_matches_all)) {
14116 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14117 \w or [:digit:] or \p{foo}
14120 /* All named classes are mapped into POSIXish nodes, with its FLAG
14121 * argument giving which class it is */
14122 switch ((I32)namedclass) {
14123 case ANYOF_UNIPROP:
14126 /* These don't depend on the charset modifiers. They always
14127 * match under /u rules */
14128 case ANYOF_NHORIZWS:
14129 case ANYOF_HORIZWS:
14130 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14133 case ANYOF_NVERTWS:
14138 /* The actual POSIXish node for all the rest depends on the
14139 * charset modifier. The ones in the first set depend only on
14140 * ASCII or, if available on this platform, locale */
14144 op = (LOC) ? POSIXL : POSIXA;
14155 /* under /a could be alpha */
14157 if (ASCII_RESTRICTED) {
14158 namedclass = ANYOF_ALPHA + (namedclass % 2);
14166 /* The rest have more possibilities depending on the charset.
14167 * We take advantage of the enum ordering of the charset
14168 * modifiers to get the exact node type, */
14170 op = POSIXD + get_regex_charset(RExC_flags);
14171 if (op > POSIXA) { /* /aa is same as /a */
14176 /* The odd numbered ones are the complements of the
14177 * next-lower even number one */
14178 if (namedclass % 2 == 1) {
14182 arg = namedclass_to_classnum(namedclass);
14186 else if (value == prevvalue) {
14188 /* Here, the class consists of just a single code point */
14191 if (! LOC && value == '\n') {
14192 op = REG_ANY; /* Optimize [^\n] */
14193 *flagp |= HASWIDTH|SIMPLE;
14197 else if (value < 256 || UTF) {
14199 /* Optimize a single value into an EXACTish node, but not if it
14200 * would require converting the pattern to UTF-8. */
14201 op = compute_EXACTish(pRExC_state);
14203 } /* Otherwise is a range */
14204 else if (! LOC) { /* locale could vary these */
14205 if (prevvalue == '0') {
14206 if (value == '9') {
14213 /* Here, we have changed <op> away from its initial value iff we found
14214 * an optimization */
14217 /* Throw away this ANYOF regnode, and emit the calculated one,
14218 * which should correspond to the beginning, not current, state of
14220 const char * cur_parse = RExC_parse;
14221 RExC_parse = (char *)orig_parse;
14225 /* To get locale nodes to not use the full ANYOF size would
14226 * require moving the code above that writes the portions
14227 * of it that aren't in other nodes to after this point.
14228 * e.g. ANYOF_POSIXL_SET */
14229 RExC_size = orig_size;
14233 RExC_emit = (regnode *)orig_emit;
14234 if (PL_regkind[op] == POSIXD) {
14235 if (op == POSIXL) {
14236 RExC_contains_locale = 1;
14239 op += NPOSIXD - POSIXD;
14244 ret = reg_node(pRExC_state, op);
14246 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14250 *flagp |= HASWIDTH|SIMPLE;
14252 else if (PL_regkind[op] == EXACT) {
14253 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14254 TRUE /* downgradable to EXACT */
14258 RExC_parse = (char *) cur_parse;
14260 SvREFCNT_dec(posixes);
14261 SvREFCNT_dec(nposixes);
14262 SvREFCNT_dec(cp_list);
14263 SvREFCNT_dec(cp_foldable_list);
14270 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14272 /* If folding, we calculate all characters that could fold to or from the
14273 * ones already on the list */
14274 if (cp_foldable_list) {
14276 UV start, end; /* End points of code point ranges */
14278 SV* fold_intersection = NULL;
14281 /* Our calculated list will be for Unicode rules. For locale
14282 * matching, we have to keep a separate list that is consulted at
14283 * runtime only when the locale indicates Unicode rules. For
14284 * non-locale, we just use to the general list */
14286 use_list = &only_utf8_locale_list;
14289 use_list = &cp_list;
14292 /* Only the characters in this class that participate in folds need
14293 * be checked. Get the intersection of this class and all the
14294 * possible characters that are foldable. This can quickly narrow
14295 * down a large class */
14296 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14297 &fold_intersection);
14299 /* The folds for all the Latin1 characters are hard-coded into this
14300 * program, but we have to go out to disk to get the others. */
14301 if (invlist_highest(cp_foldable_list) >= 256) {
14303 /* This is a hash that for a particular fold gives all
14304 * characters that are involved in it */
14305 if (! PL_utf8_foldclosures) {
14307 /* If the folds haven't been read in, call a fold function
14309 if (! PL_utf8_tofold) {
14310 U8 dummy[UTF8_MAXBYTES_CASE+1];
14312 /* This string is just a short named one above \xff */
14313 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
14314 assert(PL_utf8_tofold); /* Verify that worked */
14316 PL_utf8_foldclosures
14317 = _swash_inversion_hash(PL_utf8_tofold);
14321 /* Now look at the foldable characters in this class individually */
14322 invlist_iterinit(fold_intersection);
14323 while (invlist_iternext(fold_intersection, &start, &end)) {
14326 /* Look at every character in the range */
14327 for (j = start; j <= end; j++) {
14328 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14334 /* We have the latin1 folding rules hard-coded here so
14335 * that an innocent-looking character class, like
14336 * /[ks]/i won't have to go out to disk to find the
14337 * possible matches. XXX It would be better to
14338 * generate these via regen, in case a new version of
14339 * the Unicode standard adds new mappings, though that
14340 * is not really likely, and may be caught by the
14341 * default: case of the switch below. */
14343 if (IS_IN_SOME_FOLD_L1(j)) {
14345 /* ASCII is always matched; non-ASCII is matched
14346 * only under Unicode rules (which could happen
14347 * under /l if the locale is a UTF-8 one */
14348 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
14349 *use_list = add_cp_to_invlist(*use_list,
14350 PL_fold_latin1[j]);
14354 add_cp_to_invlist(depends_list,
14355 PL_fold_latin1[j]);
14359 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
14360 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
14362 /* Certain Latin1 characters have matches outside
14363 * Latin1. To get here, <j> is one of those
14364 * characters. None of these matches is valid for
14365 * ASCII characters under /aa, which is why the 'if'
14366 * just above excludes those. These matches only
14367 * happen when the target string is utf8. The code
14368 * below adds the single fold closures for <j> to the
14369 * inversion list. */
14375 add_cp_to_invlist(*use_list, KELVIN_SIGN);
14379 *use_list = add_cp_to_invlist(*use_list,
14380 LATIN_SMALL_LETTER_LONG_S);
14383 *use_list = add_cp_to_invlist(*use_list,
14384 GREEK_CAPITAL_LETTER_MU);
14385 *use_list = add_cp_to_invlist(*use_list,
14386 GREEK_SMALL_LETTER_MU);
14388 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14389 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14391 add_cp_to_invlist(*use_list, ANGSTROM_SIGN);
14393 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14394 *use_list = add_cp_to_invlist(*use_list,
14395 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14397 case LATIN_SMALL_LETTER_SHARP_S:
14398 *use_list = add_cp_to_invlist(*use_list,
14399 LATIN_CAPITAL_LETTER_SHARP_S);
14401 case 'F': case 'f':
14402 case 'I': case 'i':
14403 case 'L': case 'l':
14404 case 'T': case 't':
14405 case 'A': case 'a':
14406 case 'H': case 'h':
14407 case 'J': case 'j':
14408 case 'N': case 'n':
14409 case 'W': case 'w':
14410 case 'Y': case 'y':
14411 /* These all are targets of multi-character
14412 * folds from code points that require UTF8
14413 * to express, so they can't match unless
14414 * the target string is in UTF-8, so no
14415 * action here is necessary, as regexec.c
14416 * properly handles the general case for
14417 * UTF-8 matching and multi-char folds */
14420 /* Use deprecated warning to increase the
14421 * chances of this being output */
14422 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
14429 /* Here is an above Latin1 character. We don't have the
14430 * rules hard-coded for it. First, get its fold. This is
14431 * the simple fold, as the multi-character folds have been
14432 * handled earlier and separated out */
14433 _to_uni_fold_flags(j, foldbuf, &foldlen,
14434 (ASCII_FOLD_RESTRICTED)
14435 ? FOLD_FLAGS_NOMIX_ASCII
14438 /* Single character fold of above Latin1. Add everything in
14439 * its fold closure to the list that this node should match.
14440 * The fold closures data structure is a hash with the keys
14441 * being the UTF-8 of every character that is folded to, like
14442 * 'k', and the values each an array of all code points that
14443 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
14444 * Multi-character folds are not included */
14445 if ((listp = hv_fetch(PL_utf8_foldclosures,
14446 (char *) foldbuf, foldlen, FALSE)))
14448 AV* list = (AV*) *listp;
14450 for (k = 0; k <= av_tindex(list); k++) {
14451 SV** c_p = av_fetch(list, k, FALSE);
14454 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
14458 /* /aa doesn't allow folds between ASCII and non- */
14459 if ((ASCII_FOLD_RESTRICTED
14460 && (isASCII(c) != isASCII(j))))
14465 /* Folds under /l which cross the 255/256 boundary
14466 * are added to a separate list. (These are valid
14467 * only when the locale is UTF-8.) */
14468 if (c < 256 && LOC) {
14469 *use_list = add_cp_to_invlist(*use_list, c);
14473 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
14475 cp_list = add_cp_to_invlist(cp_list, c);
14478 /* Similarly folds involving non-ascii Latin1
14479 * characters under /d are added to their list */
14480 depends_list = add_cp_to_invlist(depends_list,
14487 SvREFCNT_dec_NN(fold_intersection);
14490 /* Now that we have finished adding all the folds, there is no reason
14491 * to keep the foldable list separate */
14492 _invlist_union(cp_list, cp_foldable_list, &cp_list);
14493 SvREFCNT_dec_NN(cp_foldable_list);
14496 /* And combine the result (if any) with any inversion list from posix
14497 * classes. The lists are kept separate up to now because we don't want to
14498 * fold the classes (folding of those is automatically handled by the swash
14499 * fetching code) */
14500 if (posixes || nposixes) {
14501 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
14502 /* Under /a and /aa, nothing above ASCII matches these */
14503 _invlist_intersection(posixes,
14504 PL_XPosix_ptrs[_CC_ASCII],
14508 if (DEPENDS_SEMANTICS) {
14509 /* Under /d, everything in the upper half of the Latin1 range
14510 * matches these complements */
14511 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_NON_ASCII_ALL;
14513 else if (AT_LEAST_ASCII_RESTRICTED) {
14514 /* Under /a and /aa, everything above ASCII matches these
14516 _invlist_union_complement_2nd(nposixes,
14517 PL_XPosix_ptrs[_CC_ASCII],
14521 _invlist_union(posixes, nposixes, &posixes);
14522 SvREFCNT_dec_NN(nposixes);
14525 posixes = nposixes;
14528 if (! DEPENDS_SEMANTICS) {
14530 _invlist_union(cp_list, posixes, &cp_list);
14531 SvREFCNT_dec_NN(posixes);
14538 /* Under /d, we put into a separate list the Latin1 things that
14539 * match only when the target string is utf8 */
14540 SV* nonascii_but_latin1_properties = NULL;
14541 _invlist_intersection(posixes, PL_UpperLatin1,
14542 &nonascii_but_latin1_properties);
14543 _invlist_subtract(posixes, nonascii_but_latin1_properties,
14546 _invlist_union(cp_list, posixes, &cp_list);
14547 SvREFCNT_dec_NN(posixes);
14553 if (depends_list) {
14554 _invlist_union(depends_list, nonascii_but_latin1_properties,
14556 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
14559 depends_list = nonascii_but_latin1_properties;
14564 /* And combine the result (if any) with any inversion list from properties.
14565 * The lists are kept separate up to now so that we can distinguish the two
14566 * in regards to matching above-Unicode. A run-time warning is generated
14567 * if a Unicode property is matched against a non-Unicode code point. But,
14568 * we allow user-defined properties to match anything, without any warning,
14569 * and we also suppress the warning if there is a portion of the character
14570 * class that isn't a Unicode property, and which matches above Unicode, \W
14571 * or [\x{110000}] for example.
14572 * (Note that in this case, unlike the Posix one above, there is no
14573 * <depends_list>, because having a Unicode property forces Unicode
14578 /* If it matters to the final outcome, see if a non-property
14579 * component of the class matches above Unicode. If so, the
14580 * warning gets suppressed. This is true even if just a single
14581 * such code point is specified, as though not strictly correct if
14582 * another such code point is matched against, the fact that they
14583 * are using above-Unicode code points indicates they should know
14584 * the issues involved */
14586 warn_super = ! (invert
14587 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
14590 _invlist_union(properties, cp_list, &cp_list);
14591 SvREFCNT_dec_NN(properties);
14594 cp_list = properties;
14598 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
14602 /* Here, we have calculated what code points should be in the character
14605 * Now we can see about various optimizations. Fold calculation (which we
14606 * did above) needs to take place before inversion. Otherwise /[^k]/i
14607 * would invert to include K, which under /i would match k, which it
14608 * shouldn't. Therefore we can't invert folded locale now, as it won't be
14609 * folded until runtime */
14611 /* If we didn't do folding, it's because some information isn't available
14612 * until runtime; set the run-time fold flag for these. (We don't have to
14613 * worry about properties folding, as that is taken care of by the swash
14614 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
14615 * locales, or the class matches at least one 0-255 range code point */
14617 if (only_utf8_locale_list) {
14618 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14620 else if (cp_list) { /* Look to see if there a 0-255 code point is in
14623 invlist_iterinit(cp_list);
14624 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
14625 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14627 invlist_iterfinish(cp_list);
14631 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
14632 * at compile time. Besides not inverting folded locale now, we can't
14633 * invert if there are things such as \w, which aren't known until runtime
14636 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14638 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14640 _invlist_invert(cp_list);
14642 /* Any swash can't be used as-is, because we've inverted things */
14644 SvREFCNT_dec_NN(swash);
14648 /* Clear the invert flag since have just done it here */
14653 *ret_invlist = cp_list;
14654 SvREFCNT_dec(swash);
14656 /* Discard the generated node */
14658 RExC_size = orig_size;
14661 RExC_emit = orig_emit;
14666 /* Some character classes are equivalent to other nodes. Such nodes take
14667 * up less room and generally fewer operations to execute than ANYOF nodes.
14668 * Above, we checked for and optimized into some such equivalents for
14669 * certain common classes that are easy to test. Getting to this point in
14670 * the code means that the class didn't get optimized there. Since this
14671 * code is only executed in Pass 2, it is too late to save space--it has
14672 * been allocated in Pass 1, and currently isn't given back. But turning
14673 * things into an EXACTish node can allow the optimizer to join it to any
14674 * adjacent such nodes. And if the class is equivalent to things like /./,
14675 * expensive run-time swashes can be avoided. Now that we have more
14676 * complete information, we can find things necessarily missed by the
14677 * earlier code. I (khw) am not sure how much to look for here. It would
14678 * be easy, but perhaps too slow, to check any candidates against all the
14679 * node types they could possibly match using _invlistEQ(). */
14684 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14685 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14687 /* We don't optimize if we are supposed to make sure all non-Unicode
14688 * code points raise a warning, as only ANYOF nodes have this check.
14690 && ! ((ANYOF_FLAGS(ret) | ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
14693 U8 op = END; /* The optimzation node-type */
14694 const char * cur_parse= RExC_parse;
14696 invlist_iterinit(cp_list);
14697 if (! invlist_iternext(cp_list, &start, &end)) {
14699 /* Here, the list is empty. This happens, for example, when a
14700 * Unicode property is the only thing in the character class, and
14701 * it doesn't match anything. (perluniprops.pod notes such
14704 *flagp |= HASWIDTH|SIMPLE;
14706 else if (start == end) { /* The range is a single code point */
14707 if (! invlist_iternext(cp_list, &start, &end)
14709 /* Don't do this optimization if it would require changing
14710 * the pattern to UTF-8 */
14711 && (start < 256 || UTF))
14713 /* Here, the list contains a single code point. Can optimize
14714 * into an EXACTish node */
14723 /* A locale node under folding with one code point can be
14724 * an EXACTFL, as its fold won't be calculated until
14730 /* Here, we are generally folding, but there is only one
14731 * code point to match. If we have to, we use an EXACT
14732 * node, but it would be better for joining with adjacent
14733 * nodes in the optimization pass if we used the same
14734 * EXACTFish node that any such are likely to be. We can
14735 * do this iff the code point doesn't participate in any
14736 * folds. For example, an EXACTF of a colon is the same as
14737 * an EXACT one, since nothing folds to or from a colon. */
14739 if (IS_IN_SOME_FOLD_L1(value)) {
14744 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
14749 /* If we haven't found the node type, above, it means we
14750 * can use the prevailing one */
14752 op = compute_EXACTish(pRExC_state);
14757 else if (start == 0) {
14758 if (end == UV_MAX) {
14760 *flagp |= HASWIDTH|SIMPLE;
14763 else if (end == '\n' - 1
14764 && invlist_iternext(cp_list, &start, &end)
14765 && start == '\n' + 1 && end == UV_MAX)
14768 *flagp |= HASWIDTH|SIMPLE;
14772 invlist_iterfinish(cp_list);
14775 RExC_parse = (char *)orig_parse;
14776 RExC_emit = (regnode *)orig_emit;
14778 ret = reg_node(pRExC_state, op);
14780 RExC_parse = (char *)cur_parse;
14782 if (PL_regkind[op] == EXACT) {
14783 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14784 TRUE /* downgradable to EXACT */
14788 SvREFCNT_dec_NN(cp_list);
14793 /* Here, <cp_list> contains all the code points we can determine at
14794 * compile time that match under all conditions. Go through it, and
14795 * for things that belong in the bitmap, put them there, and delete from
14796 * <cp_list>. While we are at it, see if everything above 255 is in the
14797 * list, and if so, set a flag to speed up execution */
14799 populate_ANYOF_from_invlist(ret, &cp_list);
14802 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
14805 /* Here, the bitmap has been populated with all the Latin1 code points that
14806 * always match. Can now add to the overall list those that match only
14807 * when the target string is UTF-8 (<depends_list>). */
14808 if (depends_list) {
14810 _invlist_union(cp_list, depends_list, &cp_list);
14811 SvREFCNT_dec_NN(depends_list);
14814 cp_list = depends_list;
14816 ANYOF_FLAGS(ret) |= ANYOF_UTF8;
14819 /* If there is a swash and more than one element, we can't use the swash in
14820 * the optimization below. */
14821 if (swash && element_count > 1) {
14822 SvREFCNT_dec_NN(swash);
14826 set_ANYOF_arg(pRExC_state, ret, cp_list,
14827 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14829 only_utf8_locale_list,
14830 swash, has_user_defined_property);
14832 *flagp |= HASWIDTH|SIMPLE;
14834 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
14835 RExC_contains_locale = 1;
14841 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14844 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
14845 regnode* const node,
14847 SV* const runtime_defns,
14848 SV* const only_utf8_locale_list,
14850 const bool has_user_defined_property)
14852 /* Sets the arg field of an ANYOF-type node 'node', using information about
14853 * the node passed-in. If there is nothing outside the node's bitmap, the
14854 * arg is set to ANYOF_NONBITMAP_EMPTY. Otherwise, it sets the argument to
14855 * the count returned by add_data(), having allocated and stored an array,
14856 * av, that that count references, as follows:
14857 * av[0] stores the character class description in its textual form.
14858 * This is used later (regexec.c:Perl_regclass_swash()) to
14859 * initialize the appropriate swash, and is also useful for dumping
14860 * the regnode. This is set to &PL_sv_undef if the textual
14861 * description is not needed at run-time (as happens if the other
14862 * elements completely define the class)
14863 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
14864 * computed from av[0]. But if no further computation need be done,
14865 * the swash is stored here now (and av[0] is &PL_sv_undef).
14866 * av[2] stores the inversion list of code points that match only if the
14867 * current locale is UTF-8
14868 * av[3] stores the cp_list inversion list for use in addition or instead
14869 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
14870 * (Otherwise everything needed is already in av[0] and av[1])
14871 * av[4] is set if any component of the class is from a user-defined
14872 * property; used only if av[3] exists */
14876 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
14878 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
14879 assert(! (ANYOF_FLAGS(node)
14880 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8)));
14881 ARG_SET(node, ANYOF_NONBITMAP_EMPTY);
14884 AV * const av = newAV();
14887 assert(ANYOF_FLAGS(node)
14888 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8|ANYOF_LOC_FOLD));
14890 av_store(av, 0, (runtime_defns)
14891 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
14893 av_store(av, 1, swash);
14894 SvREFCNT_dec_NN(cp_list);
14897 av_store(av, 1, &PL_sv_undef);
14899 av_store(av, 3, cp_list);
14900 av_store(av, 4, newSVuv(has_user_defined_property));
14904 if (only_utf8_locale_list) {
14905 av_store(av, 2, only_utf8_locale_list);
14908 av_store(av, 2, &PL_sv_undef);
14911 rv = newRV_noinc(MUTABLE_SV(av));
14912 n = add_data(pRExC_state, STR_WITH_LEN("s"));
14913 RExC_rxi->data->data[n] = (void*)rv;
14919 /* reg_skipcomment()
14921 Absorbs an /x style # comments from the input stream.
14922 Returns true if there is more text remaining in the stream.
14923 Will set the REG_RUN_ON_COMMENT_SEEN flag if the comment
14924 terminates the pattern without including a newline.
14926 Note its the callers responsibility to ensure that we are
14927 actually in /x mode
14932 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14936 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14938 while (RExC_parse < RExC_end)
14939 if (*RExC_parse++ == '\n') {
14944 /* we ran off the end of the pattern without ending
14945 the comment, so we have to add an \n when wrapping */
14946 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
14954 Advances the parse position, and optionally absorbs
14955 "whitespace" from the inputstream.
14957 Without /x "whitespace" means (?#...) style comments only,
14958 with /x this means (?#...) and # comments and whitespace proper.
14960 Returns the RExC_parse point from BEFORE the scan occurs.
14962 This is the /x friendly way of saying RExC_parse++.
14966 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14968 char* const retval = RExC_parse++;
14970 PERL_ARGS_ASSERT_NEXTCHAR;
14973 if (RExC_end - RExC_parse >= 3
14974 && *RExC_parse == '('
14975 && RExC_parse[1] == '?'
14976 && RExC_parse[2] == '#')
14978 while (*RExC_parse != ')') {
14979 if (RExC_parse == RExC_end)
14980 FAIL("Sequence (?#... not terminated");
14986 if (RExC_flags & RXf_PMf_EXTENDED) {
14987 if (isSPACE(*RExC_parse)) {
14991 else if (*RExC_parse == '#') {
14992 if ( reg_skipcomment( pRExC_state ) )
15001 - reg_node - emit a node
15003 STATIC regnode * /* Location. */
15004 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
15008 regnode * const ret = RExC_emit;
15009 GET_RE_DEBUG_FLAGS_DECL;
15011 PERL_ARGS_ASSERT_REG_NODE;
15014 SIZE_ALIGN(RExC_size);
15018 if (RExC_emit >= RExC_emit_bound)
15019 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15020 op, RExC_emit, RExC_emit_bound);
15022 NODE_ALIGN_FILL(ret);
15024 FILL_ADVANCE_NODE(ptr, op);
15025 #ifdef RE_TRACK_PATTERN_OFFSETS
15026 if (RExC_offsets) { /* MJD */
15028 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
15029 "reg_node", __LINE__,
15031 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
15032 ? "Overwriting end of array!\n" : "OK",
15033 (UV)(RExC_emit - RExC_emit_start),
15034 (UV)(RExC_parse - RExC_start),
15035 (UV)RExC_offsets[0]));
15036 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
15044 - reganode - emit a node with an argument
15046 STATIC regnode * /* Location. */
15047 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
15051 regnode * const ret = RExC_emit;
15052 GET_RE_DEBUG_FLAGS_DECL;
15054 PERL_ARGS_ASSERT_REGANODE;
15057 SIZE_ALIGN(RExC_size);
15062 assert(2==regarglen[op]+1);
15064 Anything larger than this has to allocate the extra amount.
15065 If we changed this to be:
15067 RExC_size += (1 + regarglen[op]);
15069 then it wouldn't matter. Its not clear what side effect
15070 might come from that so its not done so far.
15075 if (RExC_emit >= RExC_emit_bound)
15076 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15077 op, RExC_emit, RExC_emit_bound);
15079 NODE_ALIGN_FILL(ret);
15081 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15082 #ifdef RE_TRACK_PATTERN_OFFSETS
15083 if (RExC_offsets) { /* MJD */
15085 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15089 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
15090 "Overwriting end of array!\n" : "OK",
15091 (UV)(RExC_emit - RExC_emit_start),
15092 (UV)(RExC_parse - RExC_start),
15093 (UV)RExC_offsets[0]));
15094 Set_Cur_Node_Offset;
15102 - reguni - emit (if appropriate) a Unicode character
15104 PERL_STATIC_INLINE STRLEN
15105 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
15109 PERL_ARGS_ASSERT_REGUNI;
15111 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
15115 - reginsert - insert an operator in front of already-emitted operand
15117 * Means relocating the operand.
15120 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15126 const int offset = regarglen[(U8)op];
15127 const int size = NODE_STEP_REGNODE + offset;
15128 GET_RE_DEBUG_FLAGS_DECL;
15130 PERL_ARGS_ASSERT_REGINSERT;
15131 PERL_UNUSED_ARG(depth);
15132 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15133 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15142 if (RExC_open_parens) {
15144 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15145 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15146 if ( RExC_open_parens[paren] >= opnd ) {
15147 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15148 RExC_open_parens[paren] += size;
15150 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15152 if ( RExC_close_parens[paren] >= opnd ) {
15153 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15154 RExC_close_parens[paren] += size;
15156 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15161 while (src > opnd) {
15162 StructCopy(--src, --dst, regnode);
15163 #ifdef RE_TRACK_PATTERN_OFFSETS
15164 if (RExC_offsets) { /* MJD 20010112 */
15166 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15170 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15171 ? "Overwriting end of array!\n" : "OK",
15172 (UV)(src - RExC_emit_start),
15173 (UV)(dst - RExC_emit_start),
15174 (UV)RExC_offsets[0]));
15175 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15176 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15182 place = opnd; /* Op node, where operand used to be. */
15183 #ifdef RE_TRACK_PATTERN_OFFSETS
15184 if (RExC_offsets) { /* MJD */
15186 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15190 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15191 ? "Overwriting end of array!\n" : "OK",
15192 (UV)(place - RExC_emit_start),
15193 (UV)(RExC_parse - RExC_start),
15194 (UV)RExC_offsets[0]));
15195 Set_Node_Offset(place, RExC_parse);
15196 Set_Node_Length(place, 1);
15199 src = NEXTOPER(place);
15200 FILL_ADVANCE_NODE(place, op);
15201 Zero(src, offset, regnode);
15205 - regtail - set the next-pointer at the end of a node chain of p to val.
15206 - SEE ALSO: regtail_study
15208 /* TODO: All three parms should be const */
15210 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15211 const regnode *val,U32 depth)
15215 GET_RE_DEBUG_FLAGS_DECL;
15217 PERL_ARGS_ASSERT_REGTAIL;
15219 PERL_UNUSED_ARG(depth);
15225 /* Find last node. */
15228 regnode * const temp = regnext(scan);
15230 SV * const mysv=sv_newmortal();
15231 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15232 regprop(RExC_rx, mysv, scan, NULL);
15233 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15234 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
15235 (temp == NULL ? "->" : ""),
15236 (temp == NULL ? PL_reg_name[OP(val)] : "")
15244 if (reg_off_by_arg[OP(scan)]) {
15245 ARG_SET(scan, val - scan);
15248 NEXT_OFF(scan) = val - scan;
15254 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15255 - Look for optimizable sequences at the same time.
15256 - currently only looks for EXACT chains.
15258 This is experimental code. The idea is to use this routine to perform
15259 in place optimizations on branches and groups as they are constructed,
15260 with the long term intention of removing optimization from study_chunk so
15261 that it is purely analytical.
15263 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15264 to control which is which.
15267 /* TODO: All four parms should be const */
15270 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15271 const regnode *val,U32 depth)
15276 #ifdef EXPERIMENTAL_INPLACESCAN
15279 GET_RE_DEBUG_FLAGS_DECL;
15281 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15287 /* Find last node. */
15291 regnode * const temp = regnext(scan);
15292 #ifdef EXPERIMENTAL_INPLACESCAN
15293 if (PL_regkind[OP(scan)] == EXACT) {
15294 bool unfolded_multi_char; /* Unexamined in this routine */
15295 if (join_exact(pRExC_state, scan, &min,
15296 &unfolded_multi_char, 1, val, depth+1))
15301 switch (OP(scan)) {
15304 case EXACTFA_NO_TRIE:
15309 if( exact == PSEUDO )
15311 else if ( exact != OP(scan) )
15320 SV * const mysv=sv_newmortal();
15321 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
15322 regprop(RExC_rx, mysv, scan, NULL);
15323 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
15324 SvPV_nolen_const(mysv),
15325 REG_NODE_NUM(scan),
15326 PL_reg_name[exact]);
15333 SV * const mysv_val=sv_newmortal();
15334 DEBUG_PARSE_MSG("");
15335 regprop(RExC_rx, mysv_val, val, NULL);
15336 PerlIO_printf(Perl_debug_log,
15337 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
15338 SvPV_nolen_const(mysv_val),
15339 (IV)REG_NODE_NUM(val),
15343 if (reg_off_by_arg[OP(scan)]) {
15344 ARG_SET(scan, val - scan);
15347 NEXT_OFF(scan) = val - scan;
15355 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
15360 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
15365 for (bit=0; bit<32; bit++) {
15366 if (flags & (1<<bit)) {
15367 if (!set++ && lead)
15368 PerlIO_printf(Perl_debug_log, "%s",lead);
15369 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
15374 PerlIO_printf(Perl_debug_log, "\n");
15376 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15381 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
15387 for (bit=0; bit<32; bit++) {
15388 if (flags & (1<<bit)) {
15389 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
15392 if (!set++ && lead)
15393 PerlIO_printf(Perl_debug_log, "%s",lead);
15394 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
15397 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
15398 if (!set++ && lead) {
15399 PerlIO_printf(Perl_debug_log, "%s",lead);
15402 case REGEX_UNICODE_CHARSET:
15403 PerlIO_printf(Perl_debug_log, "UNICODE");
15405 case REGEX_LOCALE_CHARSET:
15406 PerlIO_printf(Perl_debug_log, "LOCALE");
15408 case REGEX_ASCII_RESTRICTED_CHARSET:
15409 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
15411 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
15412 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
15415 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
15421 PerlIO_printf(Perl_debug_log, "\n");
15423 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15429 Perl_regdump(pTHX_ const regexp *r)
15433 SV * const sv = sv_newmortal();
15434 SV *dsv= sv_newmortal();
15435 RXi_GET_DECL(r,ri);
15436 GET_RE_DEBUG_FLAGS_DECL;
15438 PERL_ARGS_ASSERT_REGDUMP;
15440 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
15442 /* Header fields of interest. */
15443 if (r->anchored_substr) {
15444 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
15445 RE_SV_DUMPLEN(r->anchored_substr), 30);
15446 PerlIO_printf(Perl_debug_log,
15447 "anchored %s%s at %"IVdf" ",
15448 s, RE_SV_TAIL(r->anchored_substr),
15449 (IV)r->anchored_offset);
15450 } else if (r->anchored_utf8) {
15451 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
15452 RE_SV_DUMPLEN(r->anchored_utf8), 30);
15453 PerlIO_printf(Perl_debug_log,
15454 "anchored utf8 %s%s at %"IVdf" ",
15455 s, RE_SV_TAIL(r->anchored_utf8),
15456 (IV)r->anchored_offset);
15458 if (r->float_substr) {
15459 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
15460 RE_SV_DUMPLEN(r->float_substr), 30);
15461 PerlIO_printf(Perl_debug_log,
15462 "floating %s%s at %"IVdf"..%"UVuf" ",
15463 s, RE_SV_TAIL(r->float_substr),
15464 (IV)r->float_min_offset, (UV)r->float_max_offset);
15465 } else if (r->float_utf8) {
15466 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
15467 RE_SV_DUMPLEN(r->float_utf8), 30);
15468 PerlIO_printf(Perl_debug_log,
15469 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
15470 s, RE_SV_TAIL(r->float_utf8),
15471 (IV)r->float_min_offset, (UV)r->float_max_offset);
15473 if (r->check_substr || r->check_utf8)
15474 PerlIO_printf(Perl_debug_log,
15476 (r->check_substr == r->float_substr
15477 && r->check_utf8 == r->float_utf8
15478 ? "(checking floating" : "(checking anchored"));
15479 if (r->intflags & PREGf_NOSCAN)
15480 PerlIO_printf(Perl_debug_log, " noscan");
15481 if (r->extflags & RXf_CHECK_ALL)
15482 PerlIO_printf(Perl_debug_log, " isall");
15483 if (r->check_substr || r->check_utf8)
15484 PerlIO_printf(Perl_debug_log, ") ");
15486 if (ri->regstclass) {
15487 regprop(r, sv, ri->regstclass, NULL);
15488 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
15490 if (r->intflags & PREGf_ANCH) {
15491 PerlIO_printf(Perl_debug_log, "anchored");
15492 if (r->intflags & PREGf_ANCH_BOL)
15493 PerlIO_printf(Perl_debug_log, "(BOL)");
15494 if (r->intflags & PREGf_ANCH_MBOL)
15495 PerlIO_printf(Perl_debug_log, "(MBOL)");
15496 if (r->intflags & PREGf_ANCH_SBOL)
15497 PerlIO_printf(Perl_debug_log, "(SBOL)");
15498 if (r->intflags & PREGf_ANCH_GPOS)
15499 PerlIO_printf(Perl_debug_log, "(GPOS)");
15500 PerlIO_putc(Perl_debug_log, ' ');
15502 if (r->intflags & PREGf_GPOS_SEEN)
15503 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
15504 if (r->intflags & PREGf_SKIP)
15505 PerlIO_printf(Perl_debug_log, "plus ");
15506 if (r->intflags & PREGf_IMPLICIT)
15507 PerlIO_printf(Perl_debug_log, "implicit ");
15508 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
15509 if (r->extflags & RXf_EVAL_SEEN)
15510 PerlIO_printf(Perl_debug_log, "with eval ");
15511 PerlIO_printf(Perl_debug_log, "\n");
15513 regdump_extflags("r->extflags: ",r->extflags);
15514 regdump_intflags("r->intflags: ",r->intflags);
15517 PERL_ARGS_ASSERT_REGDUMP;
15518 PERL_UNUSED_CONTEXT;
15519 PERL_UNUSED_ARG(r);
15520 #endif /* DEBUGGING */
15524 - regprop - printable representation of opcode, with run time support
15528 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo)
15534 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
15535 static const char * const anyofs[] = {
15536 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
15537 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
15538 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
15539 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
15540 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
15541 || _CC_VERTSPACE != 16
15542 #error Need to adjust order of anyofs[]
15579 RXi_GET_DECL(prog,progi);
15580 GET_RE_DEBUG_FLAGS_DECL;
15582 PERL_ARGS_ASSERT_REGPROP;
15586 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
15587 /* It would be nice to FAIL() here, but this may be called from
15588 regexec.c, and it would be hard to supply pRExC_state. */
15589 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
15590 (int)OP(o), (int)REGNODE_MAX);
15591 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
15593 k = PL_regkind[OP(o)];
15596 sv_catpvs(sv, " ");
15597 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
15598 * is a crude hack but it may be the best for now since
15599 * we have no flag "this EXACTish node was UTF-8"
15601 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
15602 PERL_PV_ESCAPE_UNI_DETECT |
15603 PERL_PV_ESCAPE_NONASCII |
15604 PERL_PV_PRETTY_ELLIPSES |
15605 PERL_PV_PRETTY_LTGT |
15606 PERL_PV_PRETTY_NOCLEAR
15608 } else if (k == TRIE) {
15609 /* print the details of the trie in dumpuntil instead, as
15610 * progi->data isn't available here */
15611 const char op = OP(o);
15612 const U32 n = ARG(o);
15613 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
15614 (reg_ac_data *)progi->data->data[n] :
15616 const reg_trie_data * const trie
15617 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
15619 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
15620 DEBUG_TRIE_COMPILE_r(
15621 Perl_sv_catpvf(aTHX_ sv,
15622 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
15623 (UV)trie->startstate,
15624 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
15625 (UV)trie->wordcount,
15628 (UV)TRIE_CHARCOUNT(trie),
15629 (UV)trie->uniquecharcount
15632 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
15633 sv_catpvs(sv, "[");
15634 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
15636 : TRIE_BITMAP(trie));
15637 sv_catpvs(sv, "]");
15640 } else if (k == CURLY) {
15641 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
15642 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
15643 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
15645 else if (k == WHILEM && o->flags) /* Ordinal/of */
15646 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
15647 else if (k == REF || k == OPEN || k == CLOSE
15648 || k == GROUPP || OP(o)==ACCEPT)
15650 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
15651 if ( RXp_PAREN_NAMES(prog) ) {
15652 if ( k != REF || (OP(o) < NREF)) {
15653 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
15654 SV **name= av_fetch(list, ARG(o), 0 );
15656 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15659 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
15660 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
15661 I32 *nums=(I32*)SvPVX(sv_dat);
15662 SV **name= av_fetch(list, nums[0], 0 );
15665 for ( n=0; n<SvIVX(sv_dat); n++ ) {
15666 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
15667 (n ? "," : ""), (IV)nums[n]);
15669 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15673 if ( k == REF && reginfo) {
15674 U32 n = ARG(o); /* which paren pair */
15675 I32 ln = prog->offs[n].start;
15676 if (prog->lastparen < n || ln == -1)
15677 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
15678 else if (ln == prog->offs[n].end)
15679 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
15681 const char *s = reginfo->strbeg + ln;
15682 Perl_sv_catpvf(aTHX_ sv, ": ");
15683 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
15684 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
15687 } else if (k == GOSUB)
15688 /* Paren and offset */
15689 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
15690 else if (k == VERB) {
15692 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
15693 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
15694 } else if (k == LOGICAL)
15695 /* 2: embedded, otherwise 1 */
15696 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
15697 else if (k == ANYOF) {
15698 const U8 flags = ANYOF_FLAGS(o);
15702 if (flags & ANYOF_LOCALE_FLAGS)
15703 sv_catpvs(sv, "{loc}");
15704 if (flags & ANYOF_LOC_FOLD)
15705 sv_catpvs(sv, "{i}");
15706 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
15707 if (flags & ANYOF_INVERT)
15708 sv_catpvs(sv, "^");
15710 /* output what the standard cp 0-255 bitmap matches */
15711 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
15713 /* output any special charclass tests (used entirely under use
15715 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
15717 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
15718 if (ANYOF_POSIXL_TEST(o,i)) {
15719 sv_catpv(sv, anyofs[i]);
15725 if ((flags & (ANYOF_ABOVE_LATIN1_ALL
15727 |ANYOF_NONBITMAP_NON_UTF8
15731 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
15732 if (flags & ANYOF_INVERT)
15733 /*make sure the invert info is in each */
15734 sv_catpvs(sv, "^");
15737 if (flags & ANYOF_NON_UTF8_NON_ASCII_ALL) {
15738 sv_catpvs(sv, "{non-utf8-latin1-all}");
15741 /* output information about the unicode matching */
15742 if (flags & ANYOF_ABOVE_LATIN1_ALL)
15743 sv_catpvs(sv, "{unicode_all}");
15744 else if (ARG(o) != ANYOF_NONBITMAP_EMPTY) {
15745 SV *lv; /* Set if there is something outside the bit map. */
15746 bool byte_output = FALSE; /* If something in the bitmap has
15748 SV *only_utf8_locale;
15750 /* Get the stuff that wasn't in the bitmap */
15751 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
15752 &lv, &only_utf8_locale);
15753 if (lv && lv != &PL_sv_undef) {
15754 char *s = savesvpv(lv);
15755 char * const origs = s;
15757 while (*s && *s != '\n')
15761 const char * const t = ++s;
15763 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
15764 sv_catpvs(sv, "{outside bitmap}");
15767 sv_catpvs(sv, "{utf8}");
15771 sv_catpvs(sv, " ");
15777 /* Truncate very long output */
15778 if (s - origs > 256) {
15779 Perl_sv_catpvf(aTHX_ sv,
15781 (int) (s - origs - 1),
15787 else if (*s == '\t') {
15801 SvREFCNT_dec_NN(lv);
15804 if ((flags & ANYOF_LOC_FOLD)
15805 && only_utf8_locale
15806 && only_utf8_locale != &PL_sv_undef)
15809 int max_entries = 256;
15811 sv_catpvs(sv, "{utf8 locale}");
15812 invlist_iterinit(only_utf8_locale);
15813 while (invlist_iternext(only_utf8_locale,
15815 put_range(sv, start, end);
15817 if (max_entries < 0) {
15818 sv_catpvs(sv, "...");
15822 invlist_iterfinish(only_utf8_locale);
15827 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
15829 else if (k == POSIXD || k == NPOSIXD) {
15830 U8 index = FLAGS(o) * 2;
15831 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
15832 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
15835 if (*anyofs[index] != '[') {
15838 sv_catpv(sv, anyofs[index]);
15839 if (*anyofs[index] != '[') {
15844 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
15845 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
15847 PERL_UNUSED_CONTEXT;
15848 PERL_UNUSED_ARG(sv);
15849 PERL_UNUSED_ARG(o);
15850 PERL_UNUSED_ARG(prog);
15851 PERL_UNUSED_ARG(reginfo);
15852 #endif /* DEBUGGING */
15858 Perl_re_intuit_string(pTHX_ REGEXP * const r)
15859 { /* Assume that RE_INTUIT is set */
15861 struct regexp *const prog = ReANY(r);
15862 GET_RE_DEBUG_FLAGS_DECL;
15864 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
15865 PERL_UNUSED_CONTEXT;
15869 const char * const s = SvPV_nolen_const(prog->check_substr
15870 ? prog->check_substr : prog->check_utf8);
15872 if (!PL_colorset) reginitcolors();
15873 PerlIO_printf(Perl_debug_log,
15874 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
15876 prog->check_substr ? "" : "utf8 ",
15877 PL_colors[5],PL_colors[0],
15880 (strlen(s) > 60 ? "..." : ""));
15883 return prog->check_substr ? prog->check_substr : prog->check_utf8;
15889 handles refcounting and freeing the perl core regexp structure. When
15890 it is necessary to actually free the structure the first thing it
15891 does is call the 'free' method of the regexp_engine associated to
15892 the regexp, allowing the handling of the void *pprivate; member
15893 first. (This routine is not overridable by extensions, which is why
15894 the extensions free is called first.)
15896 See regdupe and regdupe_internal if you change anything here.
15898 #ifndef PERL_IN_XSUB_RE
15900 Perl_pregfree(pTHX_ REGEXP *r)
15906 Perl_pregfree2(pTHX_ REGEXP *rx)
15909 struct regexp *const r = ReANY(rx);
15910 GET_RE_DEBUG_FLAGS_DECL;
15912 PERL_ARGS_ASSERT_PREGFREE2;
15914 if (r->mother_re) {
15915 ReREFCNT_dec(r->mother_re);
15917 CALLREGFREE_PVT(rx); /* free the private data */
15918 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15919 Safefree(r->xpv_len_u.xpvlenu_pv);
15922 SvREFCNT_dec(r->anchored_substr);
15923 SvREFCNT_dec(r->anchored_utf8);
15924 SvREFCNT_dec(r->float_substr);
15925 SvREFCNT_dec(r->float_utf8);
15926 Safefree(r->substrs);
15928 RX_MATCH_COPY_FREE(rx);
15929 #ifdef PERL_ANY_COW
15930 SvREFCNT_dec(r->saved_copy);
15933 SvREFCNT_dec(r->qr_anoncv);
15934 rx->sv_u.svu_rx = 0;
15939 This is a hacky workaround to the structural issue of match results
15940 being stored in the regexp structure which is in turn stored in
15941 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15942 could be PL_curpm in multiple contexts, and could require multiple
15943 result sets being associated with the pattern simultaneously, such
15944 as when doing a recursive match with (??{$qr})
15946 The solution is to make a lightweight copy of the regexp structure
15947 when a qr// is returned from the code executed by (??{$qr}) this
15948 lightweight copy doesn't actually own any of its data except for
15949 the starp/end and the actual regexp structure itself.
15955 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15957 struct regexp *ret;
15958 struct regexp *const r = ReANY(rx);
15959 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15961 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15964 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15966 SvOK_off((SV *)ret_x);
15968 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15969 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15970 made both spots point to the same regexp body.) */
15971 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15972 assert(!SvPVX(ret_x));
15973 ret_x->sv_u.svu_rx = temp->sv_any;
15974 temp->sv_any = NULL;
15975 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15976 SvREFCNT_dec_NN(temp);
15977 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15978 ing below will not set it. */
15979 SvCUR_set(ret_x, SvCUR(rx));
15982 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15983 sv_force_normal(sv) is called. */
15985 ret = ReANY(ret_x);
15987 SvFLAGS(ret_x) |= SvUTF8(rx);
15988 /* We share the same string buffer as the original regexp, on which we
15989 hold a reference count, incremented when mother_re is set below.
15990 The string pointer is copied here, being part of the regexp struct.
15992 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15993 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15995 const I32 npar = r->nparens+1;
15996 Newx(ret->offs, npar, regexp_paren_pair);
15997 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16000 Newx(ret->substrs, 1, struct reg_substr_data);
16001 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16003 SvREFCNT_inc_void(ret->anchored_substr);
16004 SvREFCNT_inc_void(ret->anchored_utf8);
16005 SvREFCNT_inc_void(ret->float_substr);
16006 SvREFCNT_inc_void(ret->float_utf8);
16008 /* check_substr and check_utf8, if non-NULL, point to either their
16009 anchored or float namesakes, and don't hold a second reference. */
16011 RX_MATCH_COPIED_off(ret_x);
16012 #ifdef PERL_ANY_COW
16013 ret->saved_copy = NULL;
16015 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
16016 SvREFCNT_inc_void(ret->qr_anoncv);
16022 /* regfree_internal()
16024 Free the private data in a regexp. This is overloadable by
16025 extensions. Perl takes care of the regexp structure in pregfree(),
16026 this covers the *pprivate pointer which technically perl doesn't
16027 know about, however of course we have to handle the
16028 regexp_internal structure when no extension is in use.
16030 Note this is called before freeing anything in the regexp
16035 Perl_regfree_internal(pTHX_ REGEXP * const rx)
16038 struct regexp *const r = ReANY(rx);
16039 RXi_GET_DECL(r,ri);
16040 GET_RE_DEBUG_FLAGS_DECL;
16042 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
16048 SV *dsv= sv_newmortal();
16049 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
16050 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
16051 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
16052 PL_colors[4],PL_colors[5],s);
16055 #ifdef RE_TRACK_PATTERN_OFFSETS
16057 Safefree(ri->u.offsets); /* 20010421 MJD */
16059 if (ri->code_blocks) {
16061 for (n = 0; n < ri->num_code_blocks; n++)
16062 SvREFCNT_dec(ri->code_blocks[n].src_regex);
16063 Safefree(ri->code_blocks);
16067 int n = ri->data->count;
16070 /* If you add a ->what type here, update the comment in regcomp.h */
16071 switch (ri->data->what[n]) {
16077 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
16080 Safefree(ri->data->data[n]);
16086 { /* Aho Corasick add-on structure for a trie node.
16087 Used in stclass optimization only */
16089 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16091 refcount = --aho->refcount;
16094 PerlMemShared_free(aho->states);
16095 PerlMemShared_free(aho->fail);
16096 /* do this last!!!! */
16097 PerlMemShared_free(ri->data->data[n]);
16098 PerlMemShared_free(ri->regstclass);
16104 /* trie structure. */
16106 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16108 refcount = --trie->refcount;
16111 PerlMemShared_free(trie->charmap);
16112 PerlMemShared_free(trie->states);
16113 PerlMemShared_free(trie->trans);
16115 PerlMemShared_free(trie->bitmap);
16117 PerlMemShared_free(trie->jump);
16118 PerlMemShared_free(trie->wordinfo);
16119 /* do this last!!!! */
16120 PerlMemShared_free(ri->data->data[n]);
16125 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16126 ri->data->what[n]);
16129 Safefree(ri->data->what);
16130 Safefree(ri->data);
16136 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16137 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16138 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16141 re_dup - duplicate a regexp.
16143 This routine is expected to clone a given regexp structure. It is only
16144 compiled under USE_ITHREADS.
16146 After all of the core data stored in struct regexp is duplicated
16147 the regexp_engine.dupe method is used to copy any private data
16148 stored in the *pprivate pointer. This allows extensions to handle
16149 any duplication it needs to do.
16151 See pregfree() and regfree_internal() if you change anything here.
16153 #if defined(USE_ITHREADS)
16154 #ifndef PERL_IN_XSUB_RE
16156 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16160 const struct regexp *r = ReANY(sstr);
16161 struct regexp *ret = ReANY(dstr);
16163 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16165 npar = r->nparens+1;
16166 Newx(ret->offs, npar, regexp_paren_pair);
16167 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16169 if (ret->substrs) {
16170 /* Do it this way to avoid reading from *r after the StructCopy().
16171 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16172 cache, it doesn't matter. */
16173 const bool anchored = r->check_substr
16174 ? r->check_substr == r->anchored_substr
16175 : r->check_utf8 == r->anchored_utf8;
16176 Newx(ret->substrs, 1, struct reg_substr_data);
16177 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16179 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16180 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16181 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16182 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16184 /* check_substr and check_utf8, if non-NULL, point to either their
16185 anchored or float namesakes, and don't hold a second reference. */
16187 if (ret->check_substr) {
16189 assert(r->check_utf8 == r->anchored_utf8);
16190 ret->check_substr = ret->anchored_substr;
16191 ret->check_utf8 = ret->anchored_utf8;
16193 assert(r->check_substr == r->float_substr);
16194 assert(r->check_utf8 == r->float_utf8);
16195 ret->check_substr = ret->float_substr;
16196 ret->check_utf8 = ret->float_utf8;
16198 } else if (ret->check_utf8) {
16200 ret->check_utf8 = ret->anchored_utf8;
16202 ret->check_utf8 = ret->float_utf8;
16207 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16208 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16211 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16213 if (RX_MATCH_COPIED(dstr))
16214 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16216 ret->subbeg = NULL;
16217 #ifdef PERL_ANY_COW
16218 ret->saved_copy = NULL;
16221 /* Whether mother_re be set or no, we need to copy the string. We
16222 cannot refrain from copying it when the storage points directly to
16223 our mother regexp, because that's
16224 1: a buffer in a different thread
16225 2: something we no longer hold a reference on
16226 so we need to copy it locally. */
16227 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16228 ret->mother_re = NULL;
16230 #endif /* PERL_IN_XSUB_RE */
16235 This is the internal complement to regdupe() which is used to copy
16236 the structure pointed to by the *pprivate pointer in the regexp.
16237 This is the core version of the extension overridable cloning hook.
16238 The regexp structure being duplicated will be copied by perl prior
16239 to this and will be provided as the regexp *r argument, however
16240 with the /old/ structures pprivate pointer value. Thus this routine
16241 may override any copying normally done by perl.
16243 It returns a pointer to the new regexp_internal structure.
16247 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
16250 struct regexp *const r = ReANY(rx);
16251 regexp_internal *reti;
16253 RXi_GET_DECL(r,ri);
16255 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
16259 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
16260 char, regexp_internal);
16261 Copy(ri->program, reti->program, len+1, regnode);
16263 reti->num_code_blocks = ri->num_code_blocks;
16264 if (ri->code_blocks) {
16266 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
16267 struct reg_code_block);
16268 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
16269 struct reg_code_block);
16270 for (n = 0; n < ri->num_code_blocks; n++)
16271 reti->code_blocks[n].src_regex = (REGEXP*)
16272 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
16275 reti->code_blocks = NULL;
16277 reti->regstclass = NULL;
16280 struct reg_data *d;
16281 const int count = ri->data->count;
16284 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
16285 char, struct reg_data);
16286 Newx(d->what, count, U8);
16289 for (i = 0; i < count; i++) {
16290 d->what[i] = ri->data->what[i];
16291 switch (d->what[i]) {
16292 /* see also regcomp.h and regfree_internal() */
16293 case 'a': /* actually an AV, but the dup function is identical. */
16297 case 'u': /* actually an HV, but the dup function is identical. */
16298 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
16301 /* This is cheating. */
16302 Newx(d->data[i], 1, regnode_ssc);
16303 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
16304 reti->regstclass = (regnode*)d->data[i];
16307 /* Trie stclasses are readonly and can thus be shared
16308 * without duplication. We free the stclass in pregfree
16309 * when the corresponding reg_ac_data struct is freed.
16311 reti->regstclass= ri->regstclass;
16315 ((reg_trie_data*)ri->data->data[i])->refcount++;
16320 d->data[i] = ri->data->data[i];
16323 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
16324 ri->data->what[i]);
16333 reti->name_list_idx = ri->name_list_idx;
16335 #ifdef RE_TRACK_PATTERN_OFFSETS
16336 if (ri->u.offsets) {
16337 Newx(reti->u.offsets, 2*len+1, U32);
16338 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
16341 SetProgLen(reti,len);
16344 return (void*)reti;
16347 #endif /* USE_ITHREADS */
16349 #ifndef PERL_IN_XSUB_RE
16352 - regnext - dig the "next" pointer out of a node
16355 Perl_regnext(pTHX_ regnode *p)
16363 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
16364 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16365 (int)OP(p), (int)REGNODE_MAX);
16368 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
16377 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
16380 STRLEN l1 = strlen(pat1);
16381 STRLEN l2 = strlen(pat2);
16384 const char *message;
16386 PERL_ARGS_ASSERT_RE_CROAK2;
16392 Copy(pat1, buf, l1 , char);
16393 Copy(pat2, buf + l1, l2 , char);
16394 buf[l1 + l2] = '\n';
16395 buf[l1 + l2 + 1] = '\0';
16396 va_start(args, pat2);
16397 msv = vmess(buf, &args);
16399 message = SvPV_const(msv,l1);
16402 Copy(message, buf, l1 , char);
16403 /* l1-1 to avoid \n */
16404 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
16407 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
16409 #ifndef PERL_IN_XSUB_RE
16411 Perl_save_re_context(pTHX)
16415 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
16417 const REGEXP * const rx = PM_GETRE(PL_curpm);
16420 for (i = 1; i <= RX_NPARENS(rx); i++) {
16421 char digits[TYPE_CHARS(long)];
16422 const STRLEN len = my_snprintf(digits, sizeof(digits),
16424 GV *const *const gvp
16425 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
16428 GV * const gv = *gvp;
16429 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
16441 S_put_byte(pTHX_ SV *sv, int c)
16443 PERL_ARGS_ASSERT_PUT_BYTE;
16447 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
16448 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
16449 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
16450 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
16451 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
16454 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
16459 const char string = c;
16460 if (c == '-' || c == ']' || c == '\\' || c == '^')
16461 sv_catpvs(sv, "\\");
16462 sv_catpvn(sv, &string, 1);
16467 S_put_range(pTHX_ SV *sv, UV start, UV end)
16470 /* Appends to 'sv' a displayable version of the range of code points from
16471 * 'start' to 'end' */
16473 assert(start <= end);
16475 PERL_ARGS_ASSERT_PUT_RANGE;
16477 if (end - start < 3) { /* Individual chars in short ranges */
16478 for (; start <= end; start++)
16479 put_byte(sv, start);
16481 else if ( end > 255
16482 || ! isALPHANUMERIC(start)
16483 || ! isALPHANUMERIC(end)
16484 || isDIGIT(start) != isDIGIT(end)
16485 || isUPPER(start) != isUPPER(end)
16486 || isLOWER(start) != isLOWER(end)
16488 /* This final test should get optimized out except on EBCDIC
16489 * platforms, where it causes ranges that cross discontinuities
16490 * like i/j to be shown as hex instead of the misleading,
16491 * e.g. H-K (since that range includes more than H, I, J, K).
16493 || (end - start) != NATIVE_TO_ASCII(end) - NATIVE_TO_ASCII(start))
16495 Perl_sv_catpvf(aTHX_ sv, "\\x{%02" UVXf "}-\\x{%02" UVXf "}",
16497 (end < 256) ? end : 255);
16499 else { /* Here, the ends of the range are both digits, or both uppercase,
16500 or both lowercase; and there's no discontinuity in the range
16501 (which could happen on EBCDIC platforms) */
16502 put_byte(sv, start);
16503 sv_catpvs(sv, "-");
16509 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
16511 /* Appends to 'sv' a displayable version of the innards of the bracketed
16512 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
16513 * output anything */
16516 bool has_output_anything = FALSE;
16518 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
16520 for (i = 0; i < 256; i++) {
16521 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
16523 /* The character at index i should be output. Find the next
16524 * character that should NOT be output */
16526 for (j = i + 1; j <= 256; j++) {
16527 if (! BITMAP_TEST((U8 *) bitmap, j)) {
16532 /* Everything between them is a single range that should be output
16534 put_range(sv, i, j - 1);
16535 has_output_anything = TRUE;
16540 return has_output_anything;
16543 #define CLEAR_OPTSTART \
16544 if (optstart) STMT_START { \
16545 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
16546 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
16550 #define DUMPUNTIL(b,e) \
16552 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
16554 STATIC const regnode *
16555 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
16556 const regnode *last, const regnode *plast,
16557 SV* sv, I32 indent, U32 depth)
16560 U8 op = PSEUDO; /* Arbitrary non-END op. */
16561 const regnode *next;
16562 const regnode *optstart= NULL;
16564 RXi_GET_DECL(r,ri);
16565 GET_RE_DEBUG_FLAGS_DECL;
16567 PERL_ARGS_ASSERT_DUMPUNTIL;
16569 #ifdef DEBUG_DUMPUNTIL
16570 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
16571 last ? last-start : 0,plast ? plast-start : 0);
16574 if (plast && plast < last)
16577 while (PL_regkind[op] != END && (!last || node < last)) {
16578 /* While that wasn't END last time... */
16581 if (op == CLOSE || op == WHILEM)
16583 next = regnext((regnode *)node);
16586 if (OP(node) == OPTIMIZED) {
16587 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
16594 regprop(r, sv, node, NULL);
16595 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
16596 (int)(2*indent + 1), "", SvPVX_const(sv));
16598 if (OP(node) != OPTIMIZED) {
16599 if (next == NULL) /* Next ptr. */
16600 PerlIO_printf(Perl_debug_log, " (0)");
16601 else if (PL_regkind[(U8)op] == BRANCH
16602 && PL_regkind[OP(next)] != BRANCH )
16603 PerlIO_printf(Perl_debug_log, " (FAIL)");
16605 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
16606 (void)PerlIO_putc(Perl_debug_log, '\n');
16610 if (PL_regkind[(U8)op] == BRANCHJ) {
16613 const regnode *nnode = (OP(next) == LONGJMP
16614 ? regnext((regnode *)next)
16616 if (last && nnode > last)
16618 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
16621 else if (PL_regkind[(U8)op] == BRANCH) {
16623 DUMPUNTIL(NEXTOPER(node), next);
16625 else if ( PL_regkind[(U8)op] == TRIE ) {
16626 const regnode *this_trie = node;
16627 const char op = OP(node);
16628 const U32 n = ARG(node);
16629 const reg_ac_data * const ac = op>=AHOCORASICK ?
16630 (reg_ac_data *)ri->data->data[n] :
16632 const reg_trie_data * const trie =
16633 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
16635 AV *const trie_words
16636 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
16638 const regnode *nextbranch= NULL;
16641 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
16642 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
16644 PerlIO_printf(Perl_debug_log, "%*s%s ",
16645 (int)(2*(indent+3)), "",
16647 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
16648 SvCUR(*elem_ptr), 60,
16649 PL_colors[0], PL_colors[1],
16651 ? PERL_PV_ESCAPE_UNI
16653 | PERL_PV_PRETTY_ELLIPSES
16654 | PERL_PV_PRETTY_LTGT
16659 U16 dist= trie->jump[word_idx+1];
16660 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
16661 (UV)((dist ? this_trie + dist : next) - start));
16664 nextbranch= this_trie + trie->jump[0];
16665 DUMPUNTIL(this_trie + dist, nextbranch);
16667 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
16668 nextbranch= regnext((regnode *)nextbranch);
16670 PerlIO_printf(Perl_debug_log, "\n");
16673 if (last && next > last)
16678 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
16679 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
16680 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
16682 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
16684 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
16686 else if ( op == PLUS || op == STAR) {
16687 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
16689 else if (PL_regkind[(U8)op] == ANYOF) {
16690 /* arglen 1 + class block */
16691 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_POSIXL)
16692 ? ANYOF_POSIXL_SKIP
16694 node = NEXTOPER(node);
16696 else if (PL_regkind[(U8)op] == EXACT) {
16697 /* Literal string, where present. */
16698 node += NODE_SZ_STR(node) - 1;
16699 node = NEXTOPER(node);
16702 node = NEXTOPER(node);
16703 node += regarglen[(U8)op];
16705 if (op == CURLYX || op == OPEN)
16709 #ifdef DEBUG_DUMPUNTIL
16710 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
16715 #endif /* DEBUGGING */
16719 * c-indentation-style: bsd
16720 * c-basic-offset: 4
16721 * indent-tabs-mode: nil
16724 * ex: set ts=8 sts=4 sw=4 et: