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 HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
97 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
98 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
99 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
102 #define STATIC static
106 struct RExC_state_t {
107 U32 flags; /* RXf_* are we folding, multilining? */
108 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
109 char *precomp; /* uncompiled string. */
110 REGEXP *rx_sv; /* The SV that is the regexp. */
111 regexp *rx; /* perl core regexp structure */
112 regexp_internal *rxi; /* internal data for regexp object
114 char *start; /* Start of input for compile */
115 char *end; /* End of input for compile */
116 char *parse; /* Input-scan pointer. */
117 SSize_t whilem_seen; /* number of WHILEM in this expr */
118 regnode *emit_start; /* Start of emitted-code area */
119 regnode *emit_bound; /* First regnode outside of the
121 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
122 implies compiling, so don't emit */
123 regnode_ssc emit_dummy; /* placeholder for emit to point to;
124 large enough for the largest
125 non-EXACTish node, so can use it as
127 I32 naughty; /* How bad is this pattern? */
128 I32 sawback; /* Did we see \1, ...? */
130 SSize_t size; /* Code size. */
131 I32 npar; /* Capture buffer count, (OPEN) plus
132 one. ("par" 0 is the whole
134 I32 nestroot; /* root parens we are in - used by
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
148 HV *paren_names; /* Paren names */
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
152 U8 *study_chunk_recursed; /* bitmap of which parens we have moved
154 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
158 I32 override_recoding;
159 I32 in_multi_char_class;
160 struct reg_code_block *code_blocks; /* positions of literal (?{})
162 int num_code_blocks; /* size of code_blocks[] */
163 int code_index; /* next code_blocks[] slot */
164 SSize_t maxlen; /* mininum possible number of chars in string to match */
165 #ifdef ADD_TO_REGEXEC
166 char *starttry; /* -Dr: where regtry was called. */
167 #define RExC_starttry (pRExC_state->starttry)
169 SV *runtime_code_qr; /* qr with the runtime code blocks */
171 const char *lastparse;
173 AV *paren_name_list; /* idx -> name */
174 #define RExC_lastparse (pRExC_state->lastparse)
175 #define RExC_lastnum (pRExC_state->lastnum)
176 #define RExC_paren_name_list (pRExC_state->paren_name_list)
180 #define RExC_flags (pRExC_state->flags)
181 #define RExC_pm_flags (pRExC_state->pm_flags)
182 #define RExC_precomp (pRExC_state->precomp)
183 #define RExC_rx_sv (pRExC_state->rx_sv)
184 #define RExC_rx (pRExC_state->rx)
185 #define RExC_rxi (pRExC_state->rxi)
186 #define RExC_start (pRExC_state->start)
187 #define RExC_end (pRExC_state->end)
188 #define RExC_parse (pRExC_state->parse)
189 #define RExC_whilem_seen (pRExC_state->whilem_seen)
190 #ifdef RE_TRACK_PATTERN_OFFSETS
191 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
194 #define RExC_emit (pRExC_state->emit)
195 #define RExC_emit_dummy (pRExC_state->emit_dummy)
196 #define RExC_emit_start (pRExC_state->emit_start)
197 #define RExC_emit_bound (pRExC_state->emit_bound)
198 #define RExC_naughty (pRExC_state->naughty)
199 #define RExC_sawback (pRExC_state->sawback)
200 #define RExC_seen (pRExC_state->seen)
201 #define RExC_size (pRExC_state->size)
202 #define RExC_maxlen (pRExC_state->maxlen)
203 #define RExC_npar (pRExC_state->npar)
204 #define RExC_nestroot (pRExC_state->nestroot)
205 #define RExC_extralen (pRExC_state->extralen)
206 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
207 #define RExC_utf8 (pRExC_state->utf8)
208 #define RExC_uni_semantics (pRExC_state->uni_semantics)
209 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
210 #define RExC_open_parens (pRExC_state->open_parens)
211 #define RExC_close_parens (pRExC_state->close_parens)
212 #define RExC_opend (pRExC_state->opend)
213 #define RExC_paren_names (pRExC_state->paren_names)
214 #define RExC_recurse (pRExC_state->recurse)
215 #define RExC_recurse_count (pRExC_state->recurse_count)
216 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
217 #define RExC_study_chunk_recursed_bytes \
218 (pRExC_state->study_chunk_recursed_bytes)
219 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
220 #define RExC_contains_locale (pRExC_state->contains_locale)
221 #define RExC_contains_i (pRExC_state->contains_i)
222 #define RExC_override_recoding (pRExC_state->override_recoding)
223 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
226 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
227 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
228 ((*s) == '{' && regcurly(s)))
231 * Flags to be passed up and down.
233 #define WORST 0 /* Worst case. */
234 #define HASWIDTH 0x01 /* Known to match non-null strings. */
236 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
237 * character. (There needs to be a case: in the switch statement in regexec.c
238 * for any node marked SIMPLE.) Note that this is not the same thing as
241 #define SPSTART 0x04 /* Starts with * or + */
242 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
243 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
244 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
246 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
248 /* whether trie related optimizations are enabled */
249 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
250 #define TRIE_STUDY_OPT
251 #define FULL_TRIE_STUDY
257 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
258 #define PBITVAL(paren) (1 << ((paren) & 7))
259 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
260 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
261 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
263 #define REQUIRE_UTF8 STMT_START { \
265 *flagp = RESTART_UTF8; \
270 /* This converts the named class defined in regcomp.h to its equivalent class
271 * number defined in handy.h. */
272 #define namedclass_to_classnum(class) ((int) ((class) / 2))
273 #define classnum_to_namedclass(classnum) ((classnum) * 2)
275 #define _invlist_union_complement_2nd(a, b, output) \
276 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
277 #define _invlist_intersection_complement_2nd(a, b, output) \
278 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
280 /* About scan_data_t.
282 During optimisation we recurse through the regexp program performing
283 various inplace (keyhole style) optimisations. In addition study_chunk
284 and scan_commit populate this data structure with information about
285 what strings MUST appear in the pattern. We look for the longest
286 string that must appear at a fixed location, and we look for the
287 longest string that may appear at a floating location. So for instance
292 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
293 strings (because they follow a .* construct). study_chunk will identify
294 both FOO and BAR as being the longest fixed and floating strings respectively.
296 The strings can be composites, for instance
300 will result in a composite fixed substring 'foo'.
302 For each string some basic information is maintained:
304 - offset or min_offset
305 This is the position the string must appear at, or not before.
306 It also implicitly (when combined with minlenp) tells us how many
307 characters must match before the string we are searching for.
308 Likewise when combined with minlenp and the length of the string it
309 tells us how many characters must appear after the string we have
313 Only used for floating strings. This is the rightmost point that
314 the string can appear at. If set to SSize_t_MAX it indicates that the
315 string can occur infinitely far to the right.
318 A pointer to the minimum number of characters of the pattern that the
319 string was found inside. This is important as in the case of positive
320 lookahead or positive lookbehind we can have multiple patterns
325 The minimum length of the pattern overall is 3, the minimum length
326 of the lookahead part is 3, but the minimum length of the part that
327 will actually match is 1. So 'FOO's minimum length is 3, but the
328 minimum length for the F is 1. This is important as the minimum length
329 is used to determine offsets in front of and behind the string being
330 looked for. Since strings can be composites this is the length of the
331 pattern at the time it was committed with a scan_commit. Note that
332 the length is calculated by study_chunk, so that the minimum lengths
333 are not known until the full pattern has been compiled, thus the
334 pointer to the value.
338 In the case of lookbehind the string being searched for can be
339 offset past the start point of the final matching string.
340 If this value was just blithely removed from the min_offset it would
341 invalidate some of the calculations for how many chars must match
342 before or after (as they are derived from min_offset and minlen and
343 the length of the string being searched for).
344 When the final pattern is compiled and the data is moved from the
345 scan_data_t structure into the regexp structure the information
346 about lookbehind is factored in, with the information that would
347 have been lost precalculated in the end_shift field for the
350 The fields pos_min and pos_delta are used to store the minimum offset
351 and the delta to the maximum offset at the current point in the pattern.
355 typedef struct scan_data_t {
356 /*I32 len_min; unused */
357 /*I32 len_delta; unused */
361 SSize_t last_end; /* min value, <0 unless valid. */
362 SSize_t last_start_min;
363 SSize_t last_start_max;
364 SV **longest; /* Either &l_fixed, or &l_float. */
365 SV *longest_fixed; /* longest fixed string found in pattern */
366 SSize_t offset_fixed; /* offset where it starts */
367 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
368 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
369 SV *longest_float; /* longest floating string found in pattern */
370 SSize_t offset_float_min; /* earliest point in string it can appear */
371 SSize_t offset_float_max; /* latest point in string it can appear */
372 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
373 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
376 SSize_t *last_closep;
377 regnode_ssc *start_class;
380 /* The below is perhaps overboard, but this allows us to save a test at the
381 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
382 * and 'a' differ by a single bit; the same with the upper and lower case of
383 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
384 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
385 * then inverts it to form a mask, with just a single 0, in the bit position
386 * where the upper- and lowercase differ. XXX There are about 40 other
387 * instances in the Perl core where this micro-optimization could be used.
388 * Should decide if maintenance cost is worse, before changing those
390 * Returns a boolean as to whether or not 'v' is either a lowercase or
391 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
392 * compile-time constant, the generated code is better than some optimizing
393 * compilers figure out, amounting to a mask and test. The results are
394 * meaningless if 'c' is not one of [A-Za-z] */
395 #define isARG2_lower_or_UPPER_ARG1(c, v) \
396 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
399 * Forward declarations for pregcomp()'s friends.
402 static const scan_data_t zero_scan_data =
403 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
405 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
406 #define SF_BEFORE_SEOL 0x0001
407 #define SF_BEFORE_MEOL 0x0002
408 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
409 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
411 #define SF_FIX_SHIFT_EOL (+2)
412 #define SF_FL_SHIFT_EOL (+4)
414 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
415 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
417 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
418 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
419 #define SF_IS_INF 0x0040
420 #define SF_HAS_PAR 0x0080
421 #define SF_IN_PAR 0x0100
422 #define SF_HAS_EVAL 0x0200
423 #define SCF_DO_SUBSTR 0x0400
424 #define SCF_DO_STCLASS_AND 0x0800
425 #define SCF_DO_STCLASS_OR 0x1000
426 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
427 #define SCF_WHILEM_VISITED_POS 0x2000
429 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
430 #define SCF_SEEN_ACCEPT 0x8000
431 #define SCF_TRIE_DOING_RESTUDY 0x10000
433 #define UTF cBOOL(RExC_utf8)
435 /* The enums for all these are ordered so things work out correctly */
436 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
437 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
438 == REGEX_DEPENDS_CHARSET)
439 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
440 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
441 >= REGEX_UNICODE_CHARSET)
442 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
443 == REGEX_ASCII_RESTRICTED_CHARSET)
444 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
445 >= REGEX_ASCII_RESTRICTED_CHARSET)
446 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
447 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
449 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
451 /* For programs that want to be strictly Unicode compatible by dying if any
452 * attempt is made to match a non-Unicode code point against a Unicode
454 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
456 #define OOB_NAMEDCLASS -1
458 /* There is no code point that is out-of-bounds, so this is problematic. But
459 * its only current use is to initialize a variable that is always set before
461 #define OOB_UNICODE 0xDEADBEEF
463 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
464 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
467 /* length of regex to show in messages that don't mark a position within */
468 #define RegexLengthToShowInErrorMessages 127
471 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
472 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
473 * op/pragma/warn/regcomp.
475 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
476 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
478 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
479 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
481 #define REPORT_LOCATION_ARGS(offset) \
482 UTF8fARG(UTF, offset, RExC_precomp), \
483 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
486 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
487 * arg. Show regex, up to a maximum length. If it's too long, chop and add
490 #define _FAIL(code) STMT_START { \
491 const char *ellipses = ""; \
492 IV len = RExC_end - RExC_precomp; \
495 SAVEFREESV(RExC_rx_sv); \
496 if (len > RegexLengthToShowInErrorMessages) { \
497 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
498 len = RegexLengthToShowInErrorMessages - 10; \
504 #define FAIL(msg) _FAIL( \
505 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
506 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
508 #define FAIL2(msg,arg) _FAIL( \
509 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
510 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
513 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
515 #define Simple_vFAIL(m) STMT_START { \
516 const IV offset = RExC_parse - RExC_precomp; \
517 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
518 m, REPORT_LOCATION_ARGS(offset)); \
522 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
524 #define vFAIL(m) STMT_START { \
526 SAVEFREESV(RExC_rx_sv); \
531 * Like Simple_vFAIL(), but accepts two arguments.
533 #define Simple_vFAIL2(m,a1) STMT_START { \
534 const IV offset = RExC_parse - RExC_precomp; \
535 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
536 REPORT_LOCATION_ARGS(offset)); \
540 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
542 #define vFAIL2(m,a1) STMT_START { \
544 SAVEFREESV(RExC_rx_sv); \
545 Simple_vFAIL2(m, a1); \
550 * Like Simple_vFAIL(), but accepts three arguments.
552 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
553 const IV offset = RExC_parse - RExC_precomp; \
554 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
555 REPORT_LOCATION_ARGS(offset)); \
559 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
561 #define vFAIL3(m,a1,a2) STMT_START { \
563 SAVEFREESV(RExC_rx_sv); \
564 Simple_vFAIL3(m, a1, a2); \
568 * Like Simple_vFAIL(), but accepts four arguments.
570 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
571 const IV offset = RExC_parse - RExC_precomp; \
572 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
573 REPORT_LOCATION_ARGS(offset)); \
576 #define vFAIL4(m,a1,a2,a3) STMT_START { \
578 SAVEFREESV(RExC_rx_sv); \
579 Simple_vFAIL4(m, a1, a2, a3); \
582 /* A specialized version of vFAIL2 that works with UTF8f */
583 #define vFAIL2utf8f(m, a1) STMT_START { \
584 const IV offset = RExC_parse - RExC_precomp; \
586 SAVEFREESV(RExC_rx_sv); \
587 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
588 REPORT_LOCATION_ARGS(offset)); \
592 /* m is not necessarily a "literal string", in this macro */
593 #define reg_warn_non_literal_string(loc, m) STMT_START { \
594 const IV offset = loc - RExC_precomp; \
595 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
596 m, REPORT_LOCATION_ARGS(offset)); \
599 #define ckWARNreg(loc,m) STMT_START { \
600 const IV offset = loc - RExC_precomp; \
601 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
602 REPORT_LOCATION_ARGS(offset)); \
605 #define vWARN_dep(loc, m) STMT_START { \
606 const IV offset = loc - RExC_precomp; \
607 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
608 REPORT_LOCATION_ARGS(offset)); \
611 #define ckWARNdep(loc,m) STMT_START { \
612 const IV offset = loc - RExC_precomp; \
613 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
615 REPORT_LOCATION_ARGS(offset)); \
618 #define ckWARNregdep(loc,m) STMT_START { \
619 const IV offset = loc - RExC_precomp; \
620 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
622 REPORT_LOCATION_ARGS(offset)); \
625 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
626 const IV offset = loc - RExC_precomp; \
627 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
629 a1, REPORT_LOCATION_ARGS(offset)); \
632 #define ckWARN2reg(loc, m, a1) STMT_START { \
633 const IV offset = loc - RExC_precomp; \
634 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, REPORT_LOCATION_ARGS(offset)); \
638 #define vWARN3(loc, m, a1, a2) STMT_START { \
639 const IV offset = loc - RExC_precomp; \
640 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
641 a1, a2, REPORT_LOCATION_ARGS(offset)); \
644 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
645 const IV offset = loc - RExC_precomp; \
646 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
647 a1, a2, REPORT_LOCATION_ARGS(offset)); \
650 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
651 const IV offset = loc - RExC_precomp; \
652 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
653 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
656 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
657 const IV offset = loc - RExC_precomp; \
658 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
659 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
662 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
663 const IV offset = loc - RExC_precomp; \
664 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
665 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
669 /* Allow for side effects in s */
670 #define REGC(c,s) STMT_START { \
671 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
674 /* Macros for recording node offsets. 20001227 mjd@plover.com
675 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
676 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
677 * Element 0 holds the number n.
678 * Position is 1 indexed.
680 #ifndef RE_TRACK_PATTERN_OFFSETS
681 #define Set_Node_Offset_To_R(node,byte)
682 #define Set_Node_Offset(node,byte)
683 #define Set_Cur_Node_Offset
684 #define Set_Node_Length_To_R(node,len)
685 #define Set_Node_Length(node,len)
686 #define Set_Node_Cur_Length(node,start)
687 #define Node_Offset(n)
688 #define Node_Length(n)
689 #define Set_Node_Offset_Length(node,offset,len)
690 #define ProgLen(ri) ri->u.proglen
691 #define SetProgLen(ri,x) ri->u.proglen = x
693 #define ProgLen(ri) ri->u.offsets[0]
694 #define SetProgLen(ri,x) ri->u.offsets[0] = x
695 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
697 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
698 __LINE__, (int)(node), (int)(byte))); \
700 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
703 RExC_offsets[2*(node)-1] = (byte); \
708 #define Set_Node_Offset(node,byte) \
709 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
710 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
712 #define Set_Node_Length_To_R(node,len) STMT_START { \
714 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
715 __LINE__, (int)(node), (int)(len))); \
717 Perl_croak(aTHX_ "value of node is %d in Length macro", \
720 RExC_offsets[2*(node)] = (len); \
725 #define Set_Node_Length(node,len) \
726 Set_Node_Length_To_R((node)-RExC_emit_start, len)
727 #define Set_Node_Cur_Length(node, start) \
728 Set_Node_Length(node, RExC_parse - start)
730 /* Get offsets and lengths */
731 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
732 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
734 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
735 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
736 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
740 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
741 #define EXPERIMENTAL_INPLACESCAN
742 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
744 #define DEBUG_RExC_seen() \
745 DEBUG_OPTIMISE_MORE_r({ \
746 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
748 if (RExC_seen & REG_ZERO_LEN_SEEN) \
749 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
751 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
752 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
754 if (RExC_seen & REG_GPOS_SEEN) \
755 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
757 if (RExC_seen & REG_CANY_SEEN) \
758 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
760 if (RExC_seen & REG_RECURSE_SEEN) \
761 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
763 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
764 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
766 if (RExC_seen & REG_VERBARG_SEEN) \
767 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
769 if (RExC_seen & REG_CUTGROUP_SEEN) \
770 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
772 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
773 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
775 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
776 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
778 if (RExC_seen & REG_GOSTART_SEEN) \
779 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
781 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
782 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
784 PerlIO_printf(Perl_debug_log,"\n"); \
787 #define DEBUG_STUDYDATA(str,data,depth) \
788 DEBUG_OPTIMISE_MORE_r(if(data){ \
789 PerlIO_printf(Perl_debug_log, \
790 "%*s" str "Pos:%"IVdf"/%"IVdf \
791 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
792 (int)(depth)*2, "", \
793 (IV)((data)->pos_min), \
794 (IV)((data)->pos_delta), \
795 (UV)((data)->flags), \
796 (IV)((data)->whilem_c), \
797 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
798 is_inf ? "INF " : "" \
800 if ((data)->last_found) \
801 PerlIO_printf(Perl_debug_log, \
802 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
803 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
804 SvPVX_const((data)->last_found), \
805 (IV)((data)->last_end), \
806 (IV)((data)->last_start_min), \
807 (IV)((data)->last_start_max), \
808 ((data)->longest && \
809 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
810 SvPVX_const((data)->longest_fixed), \
811 (IV)((data)->offset_fixed), \
812 ((data)->longest && \
813 (data)->longest==&((data)->longest_float)) ? "*" : "", \
814 SvPVX_const((data)->longest_float), \
815 (IV)((data)->offset_float_min), \
816 (IV)((data)->offset_float_max) \
818 PerlIO_printf(Perl_debug_log,"\n"); \
821 /* Mark that we cannot extend a found fixed substring at this point.
822 Update the longest found anchored substring and the longest found
823 floating substrings if needed. */
826 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
827 SSize_t *minlenp, int is_inf)
829 const STRLEN l = CHR_SVLEN(data->last_found);
830 const STRLEN old_l = CHR_SVLEN(*data->longest);
831 GET_RE_DEBUG_FLAGS_DECL;
833 PERL_ARGS_ASSERT_SCAN_COMMIT;
835 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
836 SvSetMagicSV(*data->longest, data->last_found);
837 if (*data->longest == data->longest_fixed) {
838 data->offset_fixed = l ? data->last_start_min : data->pos_min;
839 if (data->flags & SF_BEFORE_EOL)
841 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
843 data->flags &= ~SF_FIX_BEFORE_EOL;
844 data->minlen_fixed=minlenp;
845 data->lookbehind_fixed=0;
847 else { /* *data->longest == data->longest_float */
848 data->offset_float_min = l ? data->last_start_min : data->pos_min;
849 data->offset_float_max = (l
850 ? data->last_start_max
851 : (data->pos_delta == SSize_t_MAX
853 : data->pos_min + data->pos_delta));
855 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
856 data->offset_float_max = SSize_t_MAX;
857 if (data->flags & SF_BEFORE_EOL)
859 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
861 data->flags &= ~SF_FL_BEFORE_EOL;
862 data->minlen_float=minlenp;
863 data->lookbehind_float=0;
866 SvCUR_set(data->last_found, 0);
868 SV * const sv = data->last_found;
869 if (SvUTF8(sv) && SvMAGICAL(sv)) {
870 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
876 data->flags &= ~SF_BEFORE_EOL;
877 DEBUG_STUDYDATA("commit: ",data,0);
880 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
881 * list that describes which code points it matches */
884 S_ssc_anything(pTHX_ regnode_ssc *ssc)
886 /* Set the SSC 'ssc' to match an empty string or any code point */
888 PERL_ARGS_ASSERT_SSC_ANYTHING;
890 assert(is_ANYOF_SYNTHETIC(ssc));
892 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
893 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
894 ANYOF_FLAGS(ssc) |= ANYOF_EMPTY_STRING; /* Plus match empty string */
898 S_ssc_is_anything(const regnode_ssc *ssc)
900 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
901 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
902 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
903 * in any way, so there's no point in using it */
908 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
910 assert(is_ANYOF_SYNTHETIC(ssc));
912 if (! (ANYOF_FLAGS(ssc) & ANYOF_EMPTY_STRING)) {
916 /* See if the list consists solely of the range 0 - Infinity */
917 invlist_iterinit(ssc->invlist);
918 ret = invlist_iternext(ssc->invlist, &start, &end)
922 invlist_iterfinish(ssc->invlist);
928 /* If e.g., both \w and \W are set, matches everything */
929 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
931 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
932 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
942 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
944 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
945 * string, any code point, or any posix class under locale */
947 PERL_ARGS_ASSERT_SSC_INIT;
949 Zero(ssc, 1, regnode_ssc);
950 set_ANYOF_SYNTHETIC(ssc);
951 ARG_SET(ssc, ANYOF_NONBITMAP_EMPTY);
954 /* If any portion of the regex is to operate under locale rules,
955 * initialization includes it. The reason this isn't done for all regexes
956 * is that the optimizer was written under the assumption that locale was
957 * all-or-nothing. Given the complexity and lack of documentation in the
958 * optimizer, and that there are inadequate test cases for locale, many
959 * parts of it may not work properly, it is safest to avoid locale unless
961 if (RExC_contains_locale) {
962 ANYOF_POSIXL_SETALL(ssc);
965 ANYOF_POSIXL_ZERO(ssc);
970 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
971 const regnode_ssc *ssc)
973 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
974 * to the list of code points matched, and locale posix classes; hence does
975 * not check its flags) */
980 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
982 assert(is_ANYOF_SYNTHETIC(ssc));
984 invlist_iterinit(ssc->invlist);
985 ret = invlist_iternext(ssc->invlist, &start, &end)
989 invlist_iterfinish(ssc->invlist);
995 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1003 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1004 const regnode_charclass* const node)
1006 /* Returns a mortal inversion list defining which code points are matched
1007 * by 'node', which is of type ANYOF. Handles complementing the result if
1008 * appropriate. If some code points aren't knowable at this time, the
1009 * returned list must, and will, contain every code point that is a
1012 SV* invlist = sv_2mortal(_new_invlist(0));
1013 SV* only_utf8_locale_invlist = NULL;
1015 const U32 n = ARG(node);
1016 bool new_node_has_latin1 = FALSE;
1018 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1020 /* Look at the data structure created by S_set_ANYOF_arg() */
1021 if (n != ANYOF_NONBITMAP_EMPTY) {
1022 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1023 AV * const av = MUTABLE_AV(SvRV(rv));
1024 SV **const ary = AvARRAY(av);
1025 assert(RExC_rxi->data->what[n] == 's');
1027 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1028 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1030 else if (ary[0] && ary[0] != &PL_sv_undef) {
1032 /* Here, no compile-time swash, and there are things that won't be
1033 * known until runtime -- we have to assume it could be anything */
1034 return _add_range_to_invlist(invlist, 0, UV_MAX);
1036 else if (ary[3] && ary[3] != &PL_sv_undef) {
1038 /* Here no compile-time swash, and no run-time only data. Use the
1039 * node's inversion list */
1040 invlist = sv_2mortal(invlist_clone(ary[3]));
1043 /* Get the code points valid only under UTF-8 locales */
1044 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1045 && ary[2] && ary[2] != &PL_sv_undef)
1047 only_utf8_locale_invlist = ary[2];
1051 /* An ANYOF node contains a bitmap for the first 256 code points, and an
1052 * inversion list for the others, but if there are code points that should
1053 * match only conditionally on the target string being UTF-8, those are
1054 * placed in the inversion list, and not the bitmap. Since there are
1055 * circumstances under which they could match, they are included in the
1056 * SSC. But if the ANYOF node is to be inverted, we have to exclude them
1057 * here, so that when we invert below, the end result actually does include
1058 * them. (Think about "\xe0" =~ /[^\xc0]/di;). We have to do this here
1059 * before we add the unconditionally matched code points */
1060 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1061 _invlist_intersection_complement_2nd(invlist,
1066 /* Add in the points from the bit map */
1067 for (i = 0; i < 256; i++) {
1068 if (ANYOF_BITMAP_TEST(node, i)) {
1069 invlist = add_cp_to_invlist(invlist, i);
1070 new_node_has_latin1 = TRUE;
1074 /* If this can match all upper Latin1 code points, have to add them
1076 if (ANYOF_FLAGS(node) & ANYOF_NON_UTF8_NON_ASCII_ALL) {
1077 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1080 /* Similarly for these */
1081 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
1082 invlist = _add_range_to_invlist(invlist, 256, UV_MAX);
1085 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1086 _invlist_invert(invlist);
1088 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1090 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1091 * locale. We can skip this if there are no 0-255 at all. */
1092 _invlist_union(invlist, PL_Latin1, &invlist);
1095 /* Similarly add the UTF-8 locale possible matches. These have to be
1096 * deferred until after the non-UTF-8 locale ones are taken care of just
1097 * above, or it leads to wrong results under ANYOF_INVERT */
1098 if (only_utf8_locale_invlist) {
1099 _invlist_union_maybe_complement_2nd(invlist,
1100 only_utf8_locale_invlist,
1101 ANYOF_FLAGS(node) & ANYOF_INVERT,
1108 /* These two functions currently do the exact same thing */
1109 #define ssc_init_zero ssc_init
1111 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1112 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1114 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1115 * should not be inverted. 'and_with->flags & ANYOF_POSIXL' should be 0 if
1116 * 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1119 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1120 const regnode_charclass *and_with)
1122 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1123 * another SSC or a regular ANYOF class. Can create false positives. */
1128 PERL_ARGS_ASSERT_SSC_AND;
1130 assert(is_ANYOF_SYNTHETIC(ssc));
1132 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1133 * the code point inversion list and just the relevant flags */
1134 if (is_ANYOF_SYNTHETIC(and_with)) {
1135 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1136 anded_flags = ANYOF_FLAGS(and_with);
1138 /* XXX This is a kludge around what appears to be deficiencies in the
1139 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1140 * there are paths through the optimizer where it doesn't get weeded
1141 * out when it should. And if we don't make some extra provision for
1142 * it like the code just below, it doesn't get added when it should.
1143 * This solution is to add it only when AND'ing, which is here, and
1144 * only when what is being AND'ed is the pristine, original node
1145 * matching anything. Thus it is like adding it to ssc_anything() but
1146 * only when the result is to be AND'ed. Probably the same solution
1147 * could be adopted for the same problem we have with /l matching,
1148 * which is solved differently in S_ssc_init(), and that would lead to
1149 * fewer false positives than that solution has. But if this solution
1150 * creates bugs, the consequences are only that a warning isn't raised
1151 * that should be; while the consequences for having /l bugs is
1152 * incorrect matches */
1153 if (ssc_is_anything((regnode_ssc *)and_with)) {
1154 anded_flags |= ANYOF_WARN_SUPER;
1158 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1159 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1162 ANYOF_FLAGS(ssc) &= anded_flags;
1164 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1165 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1166 * 'and_with' may be inverted. When not inverted, we have the situation of
1168 * (C1 | P1) & (C2 | P2)
1169 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1170 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1171 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1172 * <= ((C1 & C2) | P1 | P2)
1173 * Alternatively, the last few steps could be:
1174 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1175 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1176 * <= (C1 | C2 | (P1 & P2))
1177 * We favor the second approach if either P1 or P2 is non-empty. This is
1178 * because these components are a barrier to doing optimizations, as what
1179 * they match cannot be known until the moment of matching as they are
1180 * dependent on the current locale, 'AND"ing them likely will reduce or
1182 * But we can do better if we know that C1,P1 are in their initial state (a
1183 * frequent occurrence), each matching everything:
1184 * (<everything>) & (C2 | P2) = C2 | P2
1185 * Similarly, if C2,P2 are in their initial state (again a frequent
1186 * occurrence), the result is a no-op
1187 * (C1 | P1) & (<everything>) = C1 | P1
1190 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1191 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1192 * <= (C1 & ~C2) | (P1 & ~P2)
1195 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1196 && ! is_ANYOF_SYNTHETIC(and_with))
1200 ssc_intersection(ssc,
1202 FALSE /* Has already been inverted */
1205 /* If either P1 or P2 is empty, the intersection will be also; can skip
1207 if (! (ANYOF_FLAGS(and_with) & ANYOF_POSIXL)) {
1208 ANYOF_POSIXL_ZERO(ssc);
1210 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1212 /* Note that the Posix class component P from 'and_with' actually
1214 * P = Pa | Pb | ... | Pn
1215 * where each component is one posix class, such as in [\w\s].
1217 * ~P = ~(Pa | Pb | ... | Pn)
1218 * = ~Pa & ~Pb & ... & ~Pn
1219 * <= ~Pa | ~Pb | ... | ~Pn
1220 * The last is something we can easily calculate, but unfortunately
1221 * is likely to have many false positives. We could do better
1222 * in some (but certainly not all) instances if two classes in
1223 * P have known relationships. For example
1224 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1226 * :lower: & :print: = :lower:
1227 * And similarly for classes that must be disjoint. For example,
1228 * since \s and \w can have no elements in common based on rules in
1229 * the POSIX standard,
1230 * \w & ^\S = nothing
1231 * Unfortunately, some vendor locales do not meet the Posix
1232 * standard, in particular almost everything by Microsoft.
1233 * The loop below just changes e.g., \w into \W and vice versa */
1235 regnode_charclass_posixl temp;
1236 int add = 1; /* To calculate the index of the complement */
1238 ANYOF_POSIXL_ZERO(&temp);
1239 for (i = 0; i < ANYOF_MAX; i++) {
1241 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1242 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1244 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1245 ANYOF_POSIXL_SET(&temp, i + add);
1247 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1249 ANYOF_POSIXL_AND(&temp, ssc);
1251 } /* else ssc already has no posixes */
1252 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1253 in its initial state */
1254 else if (! is_ANYOF_SYNTHETIC(and_with)
1255 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1257 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1258 * copy it over 'ssc' */
1259 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1260 if (is_ANYOF_SYNTHETIC(and_with)) {
1261 StructCopy(and_with, ssc, regnode_ssc);
1264 ssc->invlist = anded_cp_list;
1265 ANYOF_POSIXL_ZERO(ssc);
1266 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1267 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1271 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1272 || (ANYOF_FLAGS(and_with) & ANYOF_POSIXL))
1274 /* One or the other of P1, P2 is non-empty. */
1275 if (ANYOF_FLAGS(and_with) & ANYOF_POSIXL) {
1276 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1278 ssc_union(ssc, anded_cp_list, FALSE);
1280 else { /* P1 = P2 = empty */
1281 ssc_intersection(ssc, anded_cp_list, FALSE);
1287 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1288 const regnode_charclass *or_with)
1290 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1291 * another SSC or a regular ANYOF class. Can create false positives if
1292 * 'or_with' is to be inverted. */
1297 PERL_ARGS_ASSERT_SSC_OR;
1299 assert(is_ANYOF_SYNTHETIC(ssc));
1301 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1302 * the code point inversion list and just the relevant flags */
1303 if (is_ANYOF_SYNTHETIC(or_with)) {
1304 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1305 ored_flags = ANYOF_FLAGS(or_with);
1308 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1309 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1312 ANYOF_FLAGS(ssc) |= ored_flags;
1314 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1315 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1316 * 'or_with' may be inverted. When not inverted, we have the simple
1317 * situation of computing:
1318 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1319 * If P1|P2 yields a situation with both a class and its complement are
1320 * set, like having both \w and \W, this matches all code points, and we
1321 * can delete these from the P component of the ssc going forward. XXX We
1322 * might be able to delete all the P components, but I (khw) am not certain
1323 * about this, and it is better to be safe.
1326 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1327 * <= (C1 | P1) | ~C2
1328 * <= (C1 | ~C2) | P1
1329 * (which results in actually simpler code than the non-inverted case)
1332 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1333 && ! is_ANYOF_SYNTHETIC(or_with))
1335 /* We ignore P2, leaving P1 going forward */
1336 } /* else Not inverted */
1337 else if (ANYOF_FLAGS(or_with) & ANYOF_POSIXL) {
1338 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1339 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1341 for (i = 0; i < ANYOF_MAX; i += 2) {
1342 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1344 ssc_match_all_cp(ssc);
1345 ANYOF_POSIXL_CLEAR(ssc, i);
1346 ANYOF_POSIXL_CLEAR(ssc, i+1);
1354 FALSE /* Already has been inverted */
1358 PERL_STATIC_INLINE void
1359 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1361 PERL_ARGS_ASSERT_SSC_UNION;
1363 assert(is_ANYOF_SYNTHETIC(ssc));
1365 _invlist_union_maybe_complement_2nd(ssc->invlist,
1371 PERL_STATIC_INLINE void
1372 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1374 const bool invert2nd)
1376 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1378 assert(is_ANYOF_SYNTHETIC(ssc));
1380 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1386 PERL_STATIC_INLINE void
1387 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1389 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1391 assert(is_ANYOF_SYNTHETIC(ssc));
1393 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1396 PERL_STATIC_INLINE void
1397 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1399 /* AND just the single code point 'cp' into the SSC 'ssc' */
1401 SV* cp_list = _new_invlist(2);
1403 PERL_ARGS_ASSERT_SSC_CP_AND;
1405 assert(is_ANYOF_SYNTHETIC(ssc));
1407 cp_list = add_cp_to_invlist(cp_list, cp);
1408 ssc_intersection(ssc, cp_list,
1409 FALSE /* Not inverted */
1411 SvREFCNT_dec_NN(cp_list);
1414 PERL_STATIC_INLINE void
1415 S_ssc_clear_locale(regnode_ssc *ssc)
1417 /* Set the SSC 'ssc' to not match any locale things */
1418 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1420 assert(is_ANYOF_SYNTHETIC(ssc));
1422 ANYOF_POSIXL_ZERO(ssc);
1423 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1427 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1429 /* The inversion list in the SSC is marked mortal; now we need a more
1430 * permanent copy, which is stored the same way that is done in a regular
1431 * ANYOF node, with the first 256 code points in a bit map */
1433 SV* invlist = invlist_clone(ssc->invlist);
1435 PERL_ARGS_ASSERT_SSC_FINALIZE;
1437 assert(is_ANYOF_SYNTHETIC(ssc));
1439 /* The code in this file assumes that all but these flags aren't relevant
1440 * to the SSC, except ANYOF_EMPTY_STRING, which should be cleared by the
1441 * time we reach here */
1442 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1444 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1446 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1447 NULL, NULL, NULL, FALSE);
1449 /* Make sure is clone-safe */
1450 ssc->invlist = NULL;
1452 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1453 ANYOF_FLAGS(ssc) |= ANYOF_POSIXL;
1456 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1459 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1460 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1461 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1462 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1463 ? (TRIE_LIST_CUR( idx ) - 1) \
1469 dump_trie(trie,widecharmap,revcharmap)
1470 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1471 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1473 These routines dump out a trie in a somewhat readable format.
1474 The _interim_ variants are used for debugging the interim
1475 tables that are used to generate the final compressed
1476 representation which is what dump_trie expects.
1478 Part of the reason for their existence is to provide a form
1479 of documentation as to how the different representations function.
1484 Dumps the final compressed table form of the trie to Perl_debug_log.
1485 Used for debugging make_trie().
1489 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1490 AV *revcharmap, U32 depth)
1493 SV *sv=sv_newmortal();
1494 int colwidth= widecharmap ? 6 : 4;
1496 GET_RE_DEBUG_FLAGS_DECL;
1498 PERL_ARGS_ASSERT_DUMP_TRIE;
1500 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1501 (int)depth * 2 + 2,"",
1502 "Match","Base","Ofs" );
1504 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1505 SV ** const tmp = av_fetch( revcharmap, state, 0);
1507 PerlIO_printf( Perl_debug_log, "%*s",
1509 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1510 PL_colors[0], PL_colors[1],
1511 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1512 PERL_PV_ESCAPE_FIRSTCHAR
1517 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1518 (int)depth * 2 + 2,"");
1520 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1521 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1522 PerlIO_printf( Perl_debug_log, "\n");
1524 for( state = 1 ; state < trie->statecount ; state++ ) {
1525 const U32 base = trie->states[ state ].trans.base;
1527 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1528 (int)depth * 2 + 2,"", (UV)state);
1530 if ( trie->states[ state ].wordnum ) {
1531 PerlIO_printf( Perl_debug_log, " W%4X",
1532 trie->states[ state ].wordnum );
1534 PerlIO_printf( Perl_debug_log, "%6s", "" );
1537 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1542 while( ( base + ofs < trie->uniquecharcount ) ||
1543 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1544 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1548 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1550 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1551 if ( ( base + ofs >= trie->uniquecharcount )
1552 && ( base + ofs - trie->uniquecharcount
1554 && trie->trans[ base + ofs
1555 - trie->uniquecharcount ].check == state )
1557 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1559 (UV)trie->trans[ base + ofs
1560 - trie->uniquecharcount ].next );
1562 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1566 PerlIO_printf( Perl_debug_log, "]");
1569 PerlIO_printf( Perl_debug_log, "\n" );
1571 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1573 for (word=1; word <= trie->wordcount; word++) {
1574 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1575 (int)word, (int)(trie->wordinfo[word].prev),
1576 (int)(trie->wordinfo[word].len));
1578 PerlIO_printf(Perl_debug_log, "\n" );
1581 Dumps a fully constructed but uncompressed trie in list form.
1582 List tries normally only are used for construction when the number of
1583 possible chars (trie->uniquecharcount) is very high.
1584 Used for debugging make_trie().
1587 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1588 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1592 SV *sv=sv_newmortal();
1593 int colwidth= widecharmap ? 6 : 4;
1594 GET_RE_DEBUG_FLAGS_DECL;
1596 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1598 /* print out the table precompression. */
1599 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1600 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1601 "------:-----+-----------------\n" );
1603 for( state=1 ; state < next_alloc ; state ++ ) {
1606 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1607 (int)depth * 2 + 2,"", (UV)state );
1608 if ( ! trie->states[ state ].wordnum ) {
1609 PerlIO_printf( Perl_debug_log, "%5s| ","");
1611 PerlIO_printf( Perl_debug_log, "W%4x| ",
1612 trie->states[ state ].wordnum
1615 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1616 SV ** const tmp = av_fetch( revcharmap,
1617 TRIE_LIST_ITEM(state,charid).forid, 0);
1619 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1621 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1623 PL_colors[0], PL_colors[1],
1624 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1625 | PERL_PV_ESCAPE_FIRSTCHAR
1627 TRIE_LIST_ITEM(state,charid).forid,
1628 (UV)TRIE_LIST_ITEM(state,charid).newstate
1631 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1632 (int)((depth * 2) + 14), "");
1635 PerlIO_printf( Perl_debug_log, "\n");
1640 Dumps a fully constructed but uncompressed trie in table form.
1641 This is the normal DFA style state transition table, with a few
1642 twists to facilitate compression later.
1643 Used for debugging make_trie().
1646 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1647 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1652 SV *sv=sv_newmortal();
1653 int colwidth= widecharmap ? 6 : 4;
1654 GET_RE_DEBUG_FLAGS_DECL;
1656 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1659 print out the table precompression so that we can do a visual check
1660 that they are identical.
1663 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1665 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1666 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1668 PerlIO_printf( Perl_debug_log, "%*s",
1670 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1671 PL_colors[0], PL_colors[1],
1672 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1673 PERL_PV_ESCAPE_FIRSTCHAR
1679 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1681 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1682 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1685 PerlIO_printf( Perl_debug_log, "\n" );
1687 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1689 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1690 (int)depth * 2 + 2,"",
1691 (UV)TRIE_NODENUM( state ) );
1693 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1694 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1696 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1698 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1700 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1701 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1702 (UV)trie->trans[ state ].check );
1704 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1705 (UV)trie->trans[ state ].check,
1706 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1714 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1715 startbranch: the first branch in the whole branch sequence
1716 first : start branch of sequence of branch-exact nodes.
1717 May be the same as startbranch
1718 last : Thing following the last branch.
1719 May be the same as tail.
1720 tail : item following the branch sequence
1721 count : words in the sequence
1722 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS)/
1723 depth : indent depth
1725 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1727 A trie is an N'ary tree where the branches are determined by digital
1728 decomposition of the key. IE, at the root node you look up the 1st character and
1729 follow that branch repeat until you find the end of the branches. Nodes can be
1730 marked as "accepting" meaning they represent a complete word. Eg:
1734 would convert into the following structure. Numbers represent states, letters
1735 following numbers represent valid transitions on the letter from that state, if
1736 the number is in square brackets it represents an accepting state, otherwise it
1737 will be in parenthesis.
1739 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1743 (1) +-i->(6)-+-s->[7]
1745 +-s->(3)-+-h->(4)-+-e->[5]
1747 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1749 This shows that when matching against the string 'hers' we will begin at state 1
1750 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1751 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1752 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1753 single traverse. We store a mapping from accepting to state to which word was
1754 matched, and then when we have multiple possibilities we try to complete the
1755 rest of the regex in the order in which they occured in the alternation.
1757 The only prior NFA like behaviour that would be changed by the TRIE support is
1758 the silent ignoring of duplicate alternations which are of the form:
1760 / (DUPE|DUPE) X? (?{ ... }) Y /x
1762 Thus EVAL blocks following a trie may be called a different number of times with
1763 and without the optimisation. With the optimisations dupes will be silently
1764 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1765 the following demonstrates:
1767 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1769 which prints out 'word' three times, but
1771 'words'=~/(word|word|word)(?{ print $1 })S/
1773 which doesnt print it out at all. This is due to other optimisations kicking in.
1775 Example of what happens on a structural level:
1777 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1779 1: CURLYM[1] {1,32767}(18)
1790 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1791 and should turn into:
1793 1: CURLYM[1] {1,32767}(18)
1795 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1803 Cases where tail != last would be like /(?foo|bar)baz/:
1813 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1814 and would end up looking like:
1817 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1824 d = uvchr_to_utf8_flags(d, uv, 0);
1826 is the recommended Unicode-aware way of saying
1831 #define TRIE_STORE_REVCHAR(val) \
1834 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1835 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1836 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
1837 SvCUR_set(zlopp, kapow - flrbbbbb); \
1840 av_push(revcharmap, zlopp); \
1842 char ooooff = (char)val; \
1843 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1847 /* This gets the next character from the input, folding it if not already
1849 #define TRIE_READ_CHAR STMT_START { \
1852 /* if it is UTF then it is either already folded, or does not need \
1854 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
1856 else if (folder == PL_fold_latin1) { \
1857 /* This folder implies Unicode rules, which in the range expressible \
1858 * by not UTF is the lower case, with the two exceptions, one of \
1859 * which should have been taken care of before calling this */ \
1860 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
1861 uvc = toLOWER_L1(*uc); \
1862 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
1865 /* raw data, will be folded later if needed */ \
1873 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1874 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1875 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1876 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1878 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1879 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1880 TRIE_LIST_CUR( state )++; \
1883 #define TRIE_LIST_NEW(state) STMT_START { \
1884 Newxz( trie->states[ state ].trans.list, \
1885 4, reg_trie_trans_le ); \
1886 TRIE_LIST_CUR( state ) = 1; \
1887 TRIE_LIST_LEN( state ) = 4; \
1890 #define TRIE_HANDLE_WORD(state) STMT_START { \
1891 U16 dupe= trie->states[ state ].wordnum; \
1892 regnode * const noper_next = regnext( noper ); \
1895 /* store the word for dumping */ \
1897 if (OP(noper) != NOTHING) \
1898 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1900 tmp = newSVpvn_utf8( "", 0, UTF ); \
1901 av_push( trie_words, tmp ); \
1905 trie->wordinfo[curword].prev = 0; \
1906 trie->wordinfo[curword].len = wordlen; \
1907 trie->wordinfo[curword].accept = state; \
1909 if ( noper_next < tail ) { \
1911 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
1913 trie->jump[curword] = (U16)(noper_next - convert); \
1915 jumper = noper_next; \
1917 nextbranch= regnext(cur); \
1921 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1922 /* chain, so that when the bits of chain are later */\
1923 /* linked together, the dups appear in the chain */\
1924 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1925 trie->wordinfo[dupe].prev = curword; \
1927 /* we haven't inserted this word yet. */ \
1928 trie->states[ state ].wordnum = curword; \
1933 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1934 ( ( base + charid >= ucharcount \
1935 && base + charid < ubound \
1936 && state == trie->trans[ base - ucharcount + charid ].check \
1937 && trie->trans[ base - ucharcount + charid ].next ) \
1938 ? trie->trans[ base - ucharcount + charid ].next \
1939 : ( state==1 ? special : 0 ) \
1943 #define MADE_JUMP_TRIE 2
1944 #define MADE_EXACT_TRIE 4
1947 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
1948 regnode *first, regnode *last, regnode *tail,
1949 U32 word_count, U32 flags, U32 depth)
1951 /* first pass, loop through and scan words */
1952 reg_trie_data *trie;
1953 HV *widecharmap = NULL;
1954 AV *revcharmap = newAV();
1960 regnode *jumper = NULL;
1961 regnode *nextbranch = NULL;
1962 regnode *convert = NULL;
1963 U32 *prev_states; /* temp array mapping each state to previous one */
1964 /* we just use folder as a flag in utf8 */
1965 const U8 * folder = NULL;
1968 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
1969 AV *trie_words = NULL;
1970 /* along with revcharmap, this only used during construction but both are
1971 * useful during debugging so we store them in the struct when debugging.
1974 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
1975 STRLEN trie_charcount=0;
1977 SV *re_trie_maxbuff;
1978 GET_RE_DEBUG_FLAGS_DECL;
1980 PERL_ARGS_ASSERT_MAKE_TRIE;
1982 PERL_UNUSED_ARG(depth);
1989 case EXACTFU: folder = PL_fold_latin1; break;
1990 case EXACTF: folder = PL_fold; break;
1991 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1994 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1996 trie->startstate = 1;
1997 trie->wordcount = word_count;
1998 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1999 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2001 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2002 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2003 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2006 trie_words = newAV();
2009 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2010 assert(re_trie_maxbuff);
2011 if (!SvIOK(re_trie_maxbuff)) {
2012 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2014 DEBUG_TRIE_COMPILE_r({
2015 PerlIO_printf( Perl_debug_log,
2016 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2017 (int)depth * 2 + 2, "",
2018 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2019 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2022 /* Find the node we are going to overwrite */
2023 if ( first == startbranch && OP( last ) != BRANCH ) {
2024 /* whole branch chain */
2027 /* branch sub-chain */
2028 convert = NEXTOPER( first );
2031 /* -- First loop and Setup --
2033 We first traverse the branches and scan each word to determine if it
2034 contains widechars, and how many unique chars there are, this is
2035 important as we have to build a table with at least as many columns as we
2038 We use an array of integers to represent the character codes 0..255
2039 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2040 the native representation of the character value as the key and IV's for
2043 *TODO* If we keep track of how many times each character is used we can
2044 remap the columns so that the table compression later on is more
2045 efficient in terms of memory by ensuring the most common value is in the
2046 middle and the least common are on the outside. IMO this would be better
2047 than a most to least common mapping as theres a decent chance the most
2048 common letter will share a node with the least common, meaning the node
2049 will not be compressible. With a middle is most common approach the worst
2050 case is when we have the least common nodes twice.
2054 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2055 regnode *noper = NEXTOPER( cur );
2056 const U8 *uc = (U8*)STRING( noper );
2057 const U8 *e = uc + STR_LEN( noper );
2059 U32 wordlen = 0; /* required init */
2060 STRLEN minchars = 0;
2061 STRLEN maxchars = 0;
2062 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2065 if (OP(noper) == NOTHING) {
2066 regnode *noper_next= regnext(noper);
2067 if (noper_next != tail && OP(noper_next) == flags) {
2069 uc= (U8*)STRING(noper);
2070 e= uc + STR_LEN(noper);
2071 trie->minlen= STR_LEN(noper);
2078 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2079 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2080 regardless of encoding */
2081 if (OP( noper ) == EXACTFU_SS) {
2082 /* false positives are ok, so just set this */
2083 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2086 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2088 TRIE_CHARCOUNT(trie)++;
2091 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2092 * is in effect. Under /i, this character can match itself, or
2093 * anything that folds to it. If not under /i, it can match just
2094 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2095 * all fold to k, and all are single characters. But some folds
2096 * expand to more than one character, so for example LATIN SMALL
2097 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2098 * the string beginning at 'uc' is 'ffi', it could be matched by
2099 * three characters, or just by the one ligature character. (It
2100 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2101 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2102 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2103 * match.) The trie needs to know the minimum and maximum number
2104 * of characters that could match so that it can use size alone to
2105 * quickly reject many match attempts. The max is simple: it is
2106 * the number of folded characters in this branch (since a fold is
2107 * never shorter than what folds to it. */
2111 /* And the min is equal to the max if not under /i (indicated by
2112 * 'folder' being NULL), or there are no multi-character folds. If
2113 * there is a multi-character fold, the min is incremented just
2114 * once, for the character that folds to the sequence. Each
2115 * character in the sequence needs to be added to the list below of
2116 * characters in the trie, but we count only the first towards the
2117 * min number of characters needed. This is done through the
2118 * variable 'foldlen', which is returned by the macros that look
2119 * for these sequences as the number of bytes the sequence
2120 * occupies. Each time through the loop, we decrement 'foldlen' by
2121 * how many bytes the current char occupies. Only when it reaches
2122 * 0 do we increment 'minchars' or look for another multi-character
2124 if (folder == NULL) {
2127 else if (foldlen > 0) {
2128 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2133 /* See if *uc is the beginning of a multi-character fold. If
2134 * so, we decrement the length remaining to look at, to account
2135 * for the current character this iteration. (We can use 'uc'
2136 * instead of the fold returned by TRIE_READ_CHAR because for
2137 * non-UTF, the latin1_safe macro is smart enough to account
2138 * for all the unfolded characters, and because for UTF, the
2139 * string will already have been folded earlier in the
2140 * compilation process */
2142 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2143 foldlen -= UTF8SKIP(uc);
2146 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2151 /* The current character (and any potential folds) should be added
2152 * to the possible matching characters for this position in this
2156 U8 folded= folder[ (U8) uvc ];
2157 if ( !trie->charmap[ folded ] ) {
2158 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2159 TRIE_STORE_REVCHAR( folded );
2162 if ( !trie->charmap[ uvc ] ) {
2163 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2164 TRIE_STORE_REVCHAR( uvc );
2167 /* store the codepoint in the bitmap, and its folded
2169 TRIE_BITMAP_SET(trie, uvc);
2171 /* store the folded codepoint */
2172 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2175 /* store first byte of utf8 representation of
2176 variant codepoints */
2177 if (! UVCHR_IS_INVARIANT(uvc)) {
2178 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2181 set_bit = 0; /* We've done our bit :-) */
2185 /* XXX We could come up with the list of code points that fold
2186 * to this using PL_utf8_foldclosures, except not for
2187 * multi-char folds, as there may be multiple combinations
2188 * there that could work, which needs to wait until runtime to
2189 * resolve (The comment about LIGATURE FFI above is such an
2194 widecharmap = newHV();
2196 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2199 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2201 if ( !SvTRUE( *svpp ) ) {
2202 sv_setiv( *svpp, ++trie->uniquecharcount );
2203 TRIE_STORE_REVCHAR(uvc);
2206 } /* end loop through characters in this branch of the trie */
2208 /* We take the min and max for this branch and combine to find the min
2209 * and max for all branches processed so far */
2210 if( cur == first ) {
2211 trie->minlen = minchars;
2212 trie->maxlen = maxchars;
2213 } else if (minchars < trie->minlen) {
2214 trie->minlen = minchars;
2215 } else if (maxchars > trie->maxlen) {
2216 trie->maxlen = maxchars;
2218 } /* end first pass */
2219 DEBUG_TRIE_COMPILE_r(
2220 PerlIO_printf( Perl_debug_log,
2221 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2222 (int)depth * 2 + 2,"",
2223 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2224 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2225 (int)trie->minlen, (int)trie->maxlen )
2229 We now know what we are dealing with in terms of unique chars and
2230 string sizes so we can calculate how much memory a naive
2231 representation using a flat table will take. If it's over a reasonable
2232 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2233 conservative but potentially much slower representation using an array
2236 At the end we convert both representations into the same compressed
2237 form that will be used in regexec.c for matching with. The latter
2238 is a form that cannot be used to construct with but has memory
2239 properties similar to the list form and access properties similar
2240 to the table form making it both suitable for fast searches and
2241 small enough that its feasable to store for the duration of a program.
2243 See the comment in the code where the compressed table is produced
2244 inplace from the flat tabe representation for an explanation of how
2245 the compression works.
2250 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2253 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2254 > SvIV(re_trie_maxbuff) )
2257 Second Pass -- Array Of Lists Representation
2259 Each state will be represented by a list of charid:state records
2260 (reg_trie_trans_le) the first such element holds the CUR and LEN
2261 points of the allocated array. (See defines above).
2263 We build the initial structure using the lists, and then convert
2264 it into the compressed table form which allows faster lookups
2265 (but cant be modified once converted).
2268 STRLEN transcount = 1;
2270 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2271 "%*sCompiling trie using list compiler\n",
2272 (int)depth * 2 + 2, ""));
2274 trie->states = (reg_trie_state *)
2275 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2276 sizeof(reg_trie_state) );
2280 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2282 regnode *noper = NEXTOPER( cur );
2283 U8 *uc = (U8*)STRING( noper );
2284 const U8 *e = uc + STR_LEN( noper );
2285 U32 state = 1; /* required init */
2286 U16 charid = 0; /* sanity init */
2287 U32 wordlen = 0; /* required init */
2289 if (OP(noper) == NOTHING) {
2290 regnode *noper_next= regnext(noper);
2291 if (noper_next != tail && OP(noper_next) == flags) {
2293 uc= (U8*)STRING(noper);
2294 e= uc + STR_LEN(noper);
2298 if (OP(noper) != NOTHING) {
2299 for ( ; uc < e ; uc += len ) {
2304 charid = trie->charmap[ uvc ];
2306 SV** const svpp = hv_fetch( widecharmap,
2313 charid=(U16)SvIV( *svpp );
2316 /* charid is now 0 if we dont know the char read, or
2317 * nonzero if we do */
2324 if ( !trie->states[ state ].trans.list ) {
2325 TRIE_LIST_NEW( state );
2328 check <= TRIE_LIST_USED( state );
2331 if ( TRIE_LIST_ITEM( state, check ).forid
2334 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2339 newstate = next_alloc++;
2340 prev_states[newstate] = state;
2341 TRIE_LIST_PUSH( state, charid, newstate );
2346 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2350 TRIE_HANDLE_WORD(state);
2352 } /* end second pass */
2354 /* next alloc is the NEXT state to be allocated */
2355 trie->statecount = next_alloc;
2356 trie->states = (reg_trie_state *)
2357 PerlMemShared_realloc( trie->states,
2359 * sizeof(reg_trie_state) );
2361 /* and now dump it out before we compress it */
2362 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2363 revcharmap, next_alloc,
2367 trie->trans = (reg_trie_trans *)
2368 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2375 for( state=1 ; state < next_alloc ; state ++ ) {
2379 DEBUG_TRIE_COMPILE_MORE_r(
2380 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2384 if (trie->states[state].trans.list) {
2385 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2389 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2390 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2391 if ( forid < minid ) {
2393 } else if ( forid > maxid ) {
2397 if ( transcount < tp + maxid - minid + 1) {
2399 trie->trans = (reg_trie_trans *)
2400 PerlMemShared_realloc( trie->trans,
2402 * sizeof(reg_trie_trans) );
2403 Zero( trie->trans + (transcount / 2),
2407 base = trie->uniquecharcount + tp - minid;
2408 if ( maxid == minid ) {
2410 for ( ; zp < tp ; zp++ ) {
2411 if ( ! trie->trans[ zp ].next ) {
2412 base = trie->uniquecharcount + zp - minid;
2413 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2415 trie->trans[ zp ].check = state;
2421 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2423 trie->trans[ tp ].check = state;
2428 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2429 const U32 tid = base
2430 - trie->uniquecharcount
2431 + TRIE_LIST_ITEM( state, idx ).forid;
2432 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2434 trie->trans[ tid ].check = state;
2436 tp += ( maxid - minid + 1 );
2438 Safefree(trie->states[ state ].trans.list);
2441 DEBUG_TRIE_COMPILE_MORE_r(
2442 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2445 trie->states[ state ].trans.base=base;
2447 trie->lasttrans = tp + 1;
2451 Second Pass -- Flat Table Representation.
2453 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2454 each. We know that we will need Charcount+1 trans at most to store
2455 the data (one row per char at worst case) So we preallocate both
2456 structures assuming worst case.
2458 We then construct the trie using only the .next slots of the entry
2461 We use the .check field of the first entry of the node temporarily
2462 to make compression both faster and easier by keeping track of how
2463 many non zero fields are in the node.
2465 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2468 There are two terms at use here: state as a TRIE_NODEIDX() which is
2469 a number representing the first entry of the node, and state as a
2470 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2471 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2472 if there are 2 entrys per node. eg:
2480 The table is internally in the right hand, idx form. However as we
2481 also have to deal with the states array which is indexed by nodenum
2482 we have to use TRIE_NODENUM() to convert.
2485 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2486 "%*sCompiling trie using table compiler\n",
2487 (int)depth * 2 + 2, ""));
2489 trie->trans = (reg_trie_trans *)
2490 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2491 * trie->uniquecharcount + 1,
2492 sizeof(reg_trie_trans) );
2493 trie->states = (reg_trie_state *)
2494 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2495 sizeof(reg_trie_state) );
2496 next_alloc = trie->uniquecharcount + 1;
2499 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2501 regnode *noper = NEXTOPER( cur );
2502 const U8 *uc = (U8*)STRING( noper );
2503 const U8 *e = uc + STR_LEN( noper );
2505 U32 state = 1; /* required init */
2507 U16 charid = 0; /* sanity init */
2508 U32 accept_state = 0; /* sanity init */
2510 U32 wordlen = 0; /* required init */
2512 if (OP(noper) == NOTHING) {
2513 regnode *noper_next= regnext(noper);
2514 if (noper_next != tail && OP(noper_next) == flags) {
2516 uc= (U8*)STRING(noper);
2517 e= uc + STR_LEN(noper);
2521 if ( OP(noper) != NOTHING ) {
2522 for ( ; uc < e ; uc += len ) {
2527 charid = trie->charmap[ uvc ];
2529 SV* const * const svpp = hv_fetch( widecharmap,
2533 charid = svpp ? (U16)SvIV(*svpp) : 0;
2537 if ( !trie->trans[ state + charid ].next ) {
2538 trie->trans[ state + charid ].next = next_alloc;
2539 trie->trans[ state ].check++;
2540 prev_states[TRIE_NODENUM(next_alloc)]
2541 = TRIE_NODENUM(state);
2542 next_alloc += trie->uniquecharcount;
2544 state = trie->trans[ state + charid ].next;
2546 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2548 /* charid is now 0 if we dont know the char read, or
2549 * nonzero if we do */
2552 accept_state = TRIE_NODENUM( state );
2553 TRIE_HANDLE_WORD(accept_state);
2555 } /* end second pass */
2557 /* and now dump it out before we compress it */
2558 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2560 next_alloc, depth+1));
2564 * Inplace compress the table.*
2566 For sparse data sets the table constructed by the trie algorithm will
2567 be mostly 0/FAIL transitions or to put it another way mostly empty.
2568 (Note that leaf nodes will not contain any transitions.)
2570 This algorithm compresses the tables by eliminating most such
2571 transitions, at the cost of a modest bit of extra work during lookup:
2573 - Each states[] entry contains a .base field which indicates the
2574 index in the state[] array wheres its transition data is stored.
2576 - If .base is 0 there are no valid transitions from that node.
2578 - If .base is nonzero then charid is added to it to find an entry in
2581 -If trans[states[state].base+charid].check!=state then the
2582 transition is taken to be a 0/Fail transition. Thus if there are fail
2583 transitions at the front of the node then the .base offset will point
2584 somewhere inside the previous nodes data (or maybe even into a node
2585 even earlier), but the .check field determines if the transition is
2589 The following process inplace converts the table to the compressed
2590 table: We first do not compress the root node 1,and mark all its
2591 .check pointers as 1 and set its .base pointer as 1 as well. This
2592 allows us to do a DFA construction from the compressed table later,
2593 and ensures that any .base pointers we calculate later are greater
2596 - We set 'pos' to indicate the first entry of the second node.
2598 - We then iterate over the columns of the node, finding the first and
2599 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2600 and set the .check pointers accordingly, and advance pos
2601 appropriately and repreat for the next node. Note that when we copy
2602 the next pointers we have to convert them from the original
2603 NODEIDX form to NODENUM form as the former is not valid post
2606 - If a node has no transitions used we mark its base as 0 and do not
2607 advance the pos pointer.
2609 - If a node only has one transition we use a second pointer into the
2610 structure to fill in allocated fail transitions from other states.
2611 This pointer is independent of the main pointer and scans forward
2612 looking for null transitions that are allocated to a state. When it
2613 finds one it writes the single transition into the "hole". If the
2614 pointer doesnt find one the single transition is appended as normal.
2616 - Once compressed we can Renew/realloc the structures to release the
2619 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2620 specifically Fig 3.47 and the associated pseudocode.
2624 const U32 laststate = TRIE_NODENUM( next_alloc );
2627 trie->statecount = laststate;
2629 for ( state = 1 ; state < laststate ; state++ ) {
2631 const U32 stateidx = TRIE_NODEIDX( state );
2632 const U32 o_used = trie->trans[ stateidx ].check;
2633 U32 used = trie->trans[ stateidx ].check;
2634 trie->trans[ stateidx ].check = 0;
2637 used && charid < trie->uniquecharcount;
2640 if ( flag || trie->trans[ stateidx + charid ].next ) {
2641 if ( trie->trans[ stateidx + charid ].next ) {
2643 for ( ; zp < pos ; zp++ ) {
2644 if ( ! trie->trans[ zp ].next ) {
2648 trie->states[ state ].trans.base
2650 + trie->uniquecharcount
2652 trie->trans[ zp ].next
2653 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2655 trie->trans[ zp ].check = state;
2656 if ( ++zp > pos ) pos = zp;
2663 trie->states[ state ].trans.base
2664 = pos + trie->uniquecharcount - charid ;
2666 trie->trans[ pos ].next
2667 = SAFE_TRIE_NODENUM(
2668 trie->trans[ stateidx + charid ].next );
2669 trie->trans[ pos ].check = state;
2674 trie->lasttrans = pos + 1;
2675 trie->states = (reg_trie_state *)
2676 PerlMemShared_realloc( trie->states, laststate
2677 * sizeof(reg_trie_state) );
2678 DEBUG_TRIE_COMPILE_MORE_r(
2679 PerlIO_printf( Perl_debug_log,
2680 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2681 (int)depth * 2 + 2,"",
2682 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2686 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2689 } /* end table compress */
2691 DEBUG_TRIE_COMPILE_MORE_r(
2692 PerlIO_printf(Perl_debug_log,
2693 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2694 (int)depth * 2 + 2, "",
2695 (UV)trie->statecount,
2696 (UV)trie->lasttrans)
2698 /* resize the trans array to remove unused space */
2699 trie->trans = (reg_trie_trans *)
2700 PerlMemShared_realloc( trie->trans, trie->lasttrans
2701 * sizeof(reg_trie_trans) );
2703 { /* Modify the program and insert the new TRIE node */
2704 U8 nodetype =(U8)(flags & 0xFF);
2708 regnode *optimize = NULL;
2709 #ifdef RE_TRACK_PATTERN_OFFSETS
2712 U32 mjd_nodelen = 0;
2713 #endif /* RE_TRACK_PATTERN_OFFSETS */
2714 #endif /* DEBUGGING */
2716 This means we convert either the first branch or the first Exact,
2717 depending on whether the thing following (in 'last') is a branch
2718 or not and whther first is the startbranch (ie is it a sub part of
2719 the alternation or is it the whole thing.)
2720 Assuming its a sub part we convert the EXACT otherwise we convert
2721 the whole branch sequence, including the first.
2723 /* Find the node we are going to overwrite */
2724 if ( first != startbranch || OP( last ) == BRANCH ) {
2725 /* branch sub-chain */
2726 NEXT_OFF( first ) = (U16)(last - first);
2727 #ifdef RE_TRACK_PATTERN_OFFSETS
2729 mjd_offset= Node_Offset((convert));
2730 mjd_nodelen= Node_Length((convert));
2733 /* whole branch chain */
2735 #ifdef RE_TRACK_PATTERN_OFFSETS
2738 const regnode *nop = NEXTOPER( convert );
2739 mjd_offset= Node_Offset((nop));
2740 mjd_nodelen= Node_Length((nop));
2744 PerlIO_printf(Perl_debug_log,
2745 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2746 (int)depth * 2 + 2, "",
2747 (UV)mjd_offset, (UV)mjd_nodelen)
2750 /* But first we check to see if there is a common prefix we can
2751 split out as an EXACT and put in front of the TRIE node. */
2752 trie->startstate= 1;
2753 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2755 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2759 const U32 base = trie->states[ state ].trans.base;
2761 if ( trie->states[state].wordnum )
2764 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2765 if ( ( base + ofs >= trie->uniquecharcount ) &&
2766 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2767 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2769 if ( ++count > 1 ) {
2770 SV **tmp = av_fetch( revcharmap, ofs, 0);
2771 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2772 if ( state == 1 ) break;
2774 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2776 PerlIO_printf(Perl_debug_log,
2777 "%*sNew Start State=%"UVuf" Class: [",
2778 (int)depth * 2 + 2, "",
2781 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2782 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2784 TRIE_BITMAP_SET(trie,*ch);
2786 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2788 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2792 TRIE_BITMAP_SET(trie,*ch);
2794 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2795 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2801 SV **tmp = av_fetch( revcharmap, idx, 0);
2803 char *ch = SvPV( *tmp, len );
2805 SV *sv=sv_newmortal();
2806 PerlIO_printf( Perl_debug_log,
2807 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2808 (int)depth * 2 + 2, "",
2810 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2811 PL_colors[0], PL_colors[1],
2812 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2813 PERL_PV_ESCAPE_FIRSTCHAR
2818 OP( convert ) = nodetype;
2819 str=STRING(convert);
2822 STR_LEN(convert) += len;
2828 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2833 trie->prefixlen = (state-1);
2835 regnode *n = convert+NODE_SZ_STR(convert);
2836 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2837 trie->startstate = state;
2838 trie->minlen -= (state - 1);
2839 trie->maxlen -= (state - 1);
2841 /* At least the UNICOS C compiler choked on this
2842 * being argument to DEBUG_r(), so let's just have
2845 #ifdef PERL_EXT_RE_BUILD
2851 regnode *fix = convert;
2852 U32 word = trie->wordcount;
2854 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2855 while( ++fix < n ) {
2856 Set_Node_Offset_Length(fix, 0, 0);
2859 SV ** const tmp = av_fetch( trie_words, word, 0 );
2861 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2862 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2864 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2872 NEXT_OFF(convert) = (U16)(tail - convert);
2873 DEBUG_r(optimize= n);
2879 if ( trie->maxlen ) {
2880 NEXT_OFF( convert ) = (U16)(tail - convert);
2881 ARG_SET( convert, data_slot );
2882 /* Store the offset to the first unabsorbed branch in
2883 jump[0], which is otherwise unused by the jump logic.
2884 We use this when dumping a trie and during optimisation. */
2886 trie->jump[0] = (U16)(nextbranch - convert);
2888 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2889 * and there is a bitmap
2890 * and the first "jump target" node we found leaves enough room
2891 * then convert the TRIE node into a TRIEC node, with the bitmap
2892 * embedded inline in the opcode - this is hypothetically faster.
2894 if ( !trie->states[trie->startstate].wordnum
2896 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2898 OP( convert ) = TRIEC;
2899 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2900 PerlMemShared_free(trie->bitmap);
2903 OP( convert ) = TRIE;
2905 /* store the type in the flags */
2906 convert->flags = nodetype;
2910 + regarglen[ OP( convert ) ];
2912 /* XXX We really should free up the resource in trie now,
2913 as we won't use them - (which resources?) dmq */
2915 /* needed for dumping*/
2916 DEBUG_r(if (optimize) {
2917 regnode *opt = convert;
2919 while ( ++opt < optimize) {
2920 Set_Node_Offset_Length(opt,0,0);
2923 Try to clean up some of the debris left after the
2926 while( optimize < jumper ) {
2927 mjd_nodelen += Node_Length((optimize));
2928 OP( optimize ) = OPTIMIZED;
2929 Set_Node_Offset_Length(optimize,0,0);
2932 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2934 } /* end node insert */
2936 /* Finish populating the prev field of the wordinfo array. Walk back
2937 * from each accept state until we find another accept state, and if
2938 * so, point the first word's .prev field at the second word. If the
2939 * second already has a .prev field set, stop now. This will be the
2940 * case either if we've already processed that word's accept state,
2941 * or that state had multiple words, and the overspill words were
2942 * already linked up earlier.
2949 for (word=1; word <= trie->wordcount; word++) {
2951 if (trie->wordinfo[word].prev)
2953 state = trie->wordinfo[word].accept;
2955 state = prev_states[state];
2958 prev = trie->states[state].wordnum;
2962 trie->wordinfo[word].prev = prev;
2964 Safefree(prev_states);
2968 /* and now dump out the compressed format */
2969 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2971 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2973 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2974 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2976 SvREFCNT_dec_NN(revcharmap);
2980 : trie->startstate>1
2986 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
2988 /* The Trie is constructed and compressed now so we can build a fail array if
2991 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
2993 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
2997 We find the fail state for each state in the trie, this state is the longest
2998 proper suffix of the current state's 'word' that is also a proper prefix of
2999 another word in our trie. State 1 represents the word '' and is thus the
3000 default fail state. This allows the DFA not to have to restart after its
3001 tried and failed a word at a given point, it simply continues as though it
3002 had been matching the other word in the first place.
3004 'abcdgu'=~/abcdefg|cdgu/
3005 When we get to 'd' we are still matching the first word, we would encounter
3006 'g' which would fail, which would bring us to the state representing 'd' in
3007 the second word where we would try 'g' and succeed, proceeding to match
3010 /* add a fail transition */
3011 const U32 trie_offset = ARG(source);
3012 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3014 const U32 ucharcount = trie->uniquecharcount;
3015 const U32 numstates = trie->statecount;
3016 const U32 ubound = trie->lasttrans + ucharcount;
3020 U32 base = trie->states[ 1 ].trans.base;
3023 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3025 GET_RE_DEBUG_FLAGS_DECL;
3027 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3028 PERL_UNUSED_CONTEXT;
3030 PERL_UNUSED_ARG(depth);
3033 if ( OP(source) == TRIE ) {
3034 struct regnode_1 *op = (struct regnode_1 *)
3035 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3036 StructCopy(source,op,struct regnode_1);
3037 stclass = (regnode *)op;
3039 struct regnode_charclass *op = (struct regnode_charclass *)
3040 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3041 StructCopy(source,op,struct regnode_charclass);
3042 stclass = (regnode *)op;
3044 OP(stclass)+=2; /* covert the TRIE type to its AHO-CORASICK equivalent */
3046 ARG_SET( stclass, data_slot );
3047 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3048 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3049 aho->trie=trie_offset;
3050 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3051 Copy( trie->states, aho->states, numstates, reg_trie_state );
3052 Newxz( q, numstates, U32);
3053 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3056 /* initialize fail[0..1] to be 1 so that we always have
3057 a valid final fail state */
3058 fail[ 0 ] = fail[ 1 ] = 1;
3060 for ( charid = 0; charid < ucharcount ; charid++ ) {
3061 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3063 q[ q_write ] = newstate;
3064 /* set to point at the root */
3065 fail[ q[ q_write++ ] ]=1;
3068 while ( q_read < q_write) {
3069 const U32 cur = q[ q_read++ % numstates ];
3070 base = trie->states[ cur ].trans.base;
3072 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3073 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3075 U32 fail_state = cur;
3078 fail_state = fail[ fail_state ];
3079 fail_base = aho->states[ fail_state ].trans.base;
3080 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3082 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3083 fail[ ch_state ] = fail_state;
3084 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3086 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3088 q[ q_write++ % numstates] = ch_state;
3092 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3093 when we fail in state 1, this allows us to use the
3094 charclass scan to find a valid start char. This is based on the principle
3095 that theres a good chance the string being searched contains lots of stuff
3096 that cant be a start char.
3098 fail[ 0 ] = fail[ 1 ] = 0;
3099 DEBUG_TRIE_COMPILE_r({
3100 PerlIO_printf(Perl_debug_log,
3101 "%*sStclass Failtable (%"UVuf" states): 0",
3102 (int)(depth * 2), "", (UV)numstates
3104 for( q_read=1; q_read<numstates; q_read++ ) {
3105 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3107 PerlIO_printf(Perl_debug_log, "\n");
3110 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3115 #define DEBUG_PEEP(str,scan,depth) \
3116 DEBUG_OPTIMISE_r({if (scan){ \
3117 SV * const mysv=sv_newmortal(); \
3118 regnode *Next = regnext(scan); \
3119 regprop(RExC_rx, mysv, scan, NULL); \
3120 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
3121 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
3122 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3126 /* The below joins as many adjacent EXACTish nodes as possible into a single
3127 * one. The regop may be changed if the node(s) contain certain sequences that
3128 * require special handling. The joining is only done if:
3129 * 1) there is room in the current conglomerated node to entirely contain the
3131 * 2) they are the exact same node type
3133 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3134 * these get optimized out
3136 * If a node is to match under /i (folded), the number of characters it matches
3137 * can be different than its character length if it contains a multi-character
3138 * fold. *min_subtract is set to the total delta number of characters of the
3141 * And *unfolded_multi_char is set to indicate whether or not the node contains
3142 * an unfolded multi-char fold. This happens when whether the fold is valid or
3143 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3144 * SMALL LETTER SHARP S, as only if the target string being matched against
3145 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3146 * folding rules depend on the locale in force at runtime. (Multi-char folds
3147 * whose components are all above the Latin1 range are not run-time locale
3148 * dependent, and have already been folded by the time this function is
3151 * This is as good a place as any to discuss the design of handling these
3152 * multi-character fold sequences. It's been wrong in Perl for a very long
3153 * time. There are three code points in Unicode whose multi-character folds
3154 * were long ago discovered to mess things up. The previous designs for
3155 * dealing with these involved assigning a special node for them. This
3156 * approach doesn't always work, as evidenced by this example:
3157 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3158 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3159 * would match just the \xDF, it won't be able to handle the case where a
3160 * successful match would have to cross the node's boundary. The new approach
3161 * that hopefully generally solves the problem generates an EXACTFU_SS node
3162 * that is "sss" in this case.
3164 * It turns out that there are problems with all multi-character folds, and not
3165 * just these three. Now the code is general, for all such cases. The
3166 * approach taken is:
3167 * 1) This routine examines each EXACTFish node that could contain multi-
3168 * character folded sequences. Since a single character can fold into
3169 * such a sequence, the minimum match length for this node is less than
3170 * the number of characters in the node. This routine returns in
3171 * *min_subtract how many characters to subtract from the the actual
3172 * length of the string to get a real minimum match length; it is 0 if
3173 * there are no multi-char foldeds. This delta is used by the caller to
3174 * adjust the min length of the match, and the delta between min and max,
3175 * so that the optimizer doesn't reject these possibilities based on size
3177 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3178 * is used for an EXACTFU node that contains at least one "ss" sequence in
3179 * it. For non-UTF-8 patterns and strings, this is the only case where
3180 * there is a possible fold length change. That means that a regular
3181 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3182 * with length changes, and so can be processed faster. regexec.c takes
3183 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3184 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3185 * known until runtime). This saves effort in regex matching. However,
3186 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3187 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3188 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3189 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3190 * possibilities for the non-UTF8 patterns are quite simple, except for
3191 * the sharp s. All the ones that don't involve a UTF-8 target string are
3192 * members of a fold-pair, and arrays are set up for all of them so that
3193 * the other member of the pair can be found quickly. Code elsewhere in
3194 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3195 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3196 * described in the next item.
3197 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3198 * validity of the fold won't be known until runtime, and so must remain
3199 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3200 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3201 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3202 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3203 * The reason this is a problem is that the optimizer part of regexec.c
3204 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3205 * that a character in the pattern corresponds to at most a single
3206 * character in the target string. (And I do mean character, and not byte
3207 * here, unlike other parts of the documentation that have never been
3208 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3209 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3210 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3211 * nodes, violate the assumption, and they are the only instances where it
3212 * is violated. I'm reluctant to try to change the assumption, as the
3213 * code involved is impenetrable to me (khw), so instead the code here
3214 * punts. This routine examines EXACTFL nodes, and (when the pattern
3215 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3216 * boolean indicating whether or not the node contains such a fold. When
3217 * it is true, the caller sets a flag that later causes the optimizer in
3218 * this file to not set values for the floating and fixed string lengths,
3219 * and thus avoids the optimizer code in regexec.c that makes the invalid
3220 * assumption. Thus, there is no optimization based on string lengths for
3221 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3222 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3223 * assumption is wrong only in these cases is that all other non-UTF-8
3224 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3225 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3226 * EXACTF nodes because we don't know at compile time if it actually
3227 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3228 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3229 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3230 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3231 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3232 * string would require the pattern to be forced into UTF-8, the overhead
3233 * of which we want to avoid. Similarly the unfolded multi-char folds in
3234 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3237 * Similarly, the code that generates tries doesn't currently handle
3238 * not-already-folded multi-char folds, and it looks like a pain to change
3239 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3240 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3241 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3242 * using /iaa matching will be doing so almost entirely with ASCII
3243 * strings, so this should rarely be encountered in practice */
3245 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3246 if (PL_regkind[OP(scan)] == EXACT) \
3247 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3250 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3251 UV *min_subtract, bool *unfolded_multi_char,
3252 U32 flags,regnode *val, U32 depth)
3254 /* Merge several consecutive EXACTish nodes into one. */
3255 regnode *n = regnext(scan);
3257 regnode *next = scan + NODE_SZ_STR(scan);
3261 regnode *stop = scan;
3262 GET_RE_DEBUG_FLAGS_DECL;
3264 PERL_UNUSED_ARG(depth);
3267 PERL_ARGS_ASSERT_JOIN_EXACT;
3268 #ifndef EXPERIMENTAL_INPLACESCAN
3269 PERL_UNUSED_ARG(flags);
3270 PERL_UNUSED_ARG(val);
3272 DEBUG_PEEP("join",scan,depth);
3274 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3275 * EXACT ones that are mergeable to the current one. */
3277 && (PL_regkind[OP(n)] == NOTHING
3278 || (stringok && OP(n) == OP(scan)))
3280 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3283 if (OP(n) == TAIL || n > next)
3285 if (PL_regkind[OP(n)] == NOTHING) {
3286 DEBUG_PEEP("skip:",n,depth);
3287 NEXT_OFF(scan) += NEXT_OFF(n);
3288 next = n + NODE_STEP_REGNODE;
3295 else if (stringok) {
3296 const unsigned int oldl = STR_LEN(scan);
3297 regnode * const nnext = regnext(n);
3299 /* XXX I (khw) kind of doubt that this works on platforms (should
3300 * Perl ever run on one) where U8_MAX is above 255 because of lots
3301 * of other assumptions */
3302 /* Don't join if the sum can't fit into a single node */
3303 if (oldl + STR_LEN(n) > U8_MAX)
3306 DEBUG_PEEP("merg",n,depth);
3309 NEXT_OFF(scan) += NEXT_OFF(n);
3310 STR_LEN(scan) += STR_LEN(n);
3311 next = n + NODE_SZ_STR(n);
3312 /* Now we can overwrite *n : */
3313 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3321 #ifdef EXPERIMENTAL_INPLACESCAN
3322 if (flags && !NEXT_OFF(n)) {
3323 DEBUG_PEEP("atch", val, depth);
3324 if (reg_off_by_arg[OP(n)]) {
3325 ARG_SET(n, val - n);
3328 NEXT_OFF(n) = val - n;
3336 *unfolded_multi_char = FALSE;
3338 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3339 * can now analyze for sequences of problematic code points. (Prior to
3340 * this final joining, sequences could have been split over boundaries, and
3341 * hence missed). The sequences only happen in folding, hence for any
3342 * non-EXACT EXACTish node */
3343 if (OP(scan) != EXACT) {
3344 U8* s0 = (U8*) STRING(scan);
3346 U8* s_end = s0 + STR_LEN(scan);
3348 int total_count_delta = 0; /* Total delta number of characters that
3349 multi-char folds expand to */
3351 /* One pass is made over the node's string looking for all the
3352 * possibilities. To avoid some tests in the loop, there are two main
3353 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3358 if (OP(scan) == EXACTFL) {
3361 /* An EXACTFL node would already have been changed to another
3362 * node type unless there is at least one character in it that
3363 * is problematic; likely a character whose fold definition
3364 * won't be known until runtime, and so has yet to be folded.
3365 * For all but the UTF-8 locale, folds are 1-1 in length, but
3366 * to handle the UTF-8 case, we need to create a temporary
3367 * folded copy using UTF-8 locale rules in order to analyze it.
3368 * This is because our macros that look to see if a sequence is
3369 * a multi-char fold assume everything is folded (otherwise the
3370 * tests in those macros would be too complicated and slow).
3371 * Note that here, the non-problematic folds will have already
3372 * been done, so we can just copy such characters. We actually
3373 * don't completely fold the EXACTFL string. We skip the
3374 * unfolded multi-char folds, as that would just create work
3375 * below to figure out the size they already are */
3377 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3380 STRLEN s_len = UTF8SKIP(s);
3381 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3382 Copy(s, d, s_len, U8);
3385 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3386 *unfolded_multi_char = TRUE;
3387 Copy(s, d, s_len, U8);
3390 else if (isASCII(*s)) {
3391 *(d++) = toFOLD(*s);
3395 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3401 /* Point the remainder of the routine to look at our temporary
3405 } /* End of creating folded copy of EXACTFL string */
3407 /* Examine the string for a multi-character fold sequence. UTF-8
3408 * patterns have all characters pre-folded by the time this code is
3410 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3411 length sequence we are looking for is 2 */
3413 int count = 0; /* How many characters in a multi-char fold */
3414 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3415 if (! len) { /* Not a multi-char fold: get next char */
3420 /* Nodes with 'ss' require special handling, except for
3421 * EXACTFA-ish for which there is no multi-char fold to this */
3422 if (len == 2 && *s == 's' && *(s+1) == 's'
3423 && OP(scan) != EXACTFA
3424 && OP(scan) != EXACTFA_NO_TRIE)
3427 if (OP(scan) != EXACTFL) {
3428 OP(scan) = EXACTFU_SS;
3432 else { /* Here is a generic multi-char fold. */
3433 U8* multi_end = s + len;
3435 /* Count how many characters are in it. In the case of
3436 * /aa, no folds which contain ASCII code points are
3437 * allowed, so check for those, and skip if found. */
3438 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3439 count = utf8_length(s, multi_end);
3443 while (s < multi_end) {
3446 goto next_iteration;
3456 /* The delta is how long the sequence is minus 1 (1 is how long
3457 * the character that folds to the sequence is) */
3458 total_count_delta += count - 1;
3462 /* We created a temporary folded copy of the string in EXACTFL
3463 * nodes. Therefore we need to be sure it doesn't go below zero,
3464 * as the real string could be shorter */
3465 if (OP(scan) == EXACTFL) {
3466 int total_chars = utf8_length((U8*) STRING(scan),
3467 (U8*) STRING(scan) + STR_LEN(scan));
3468 if (total_count_delta > total_chars) {
3469 total_count_delta = total_chars;
3473 *min_subtract += total_count_delta;
3476 else if (OP(scan) == EXACTFA) {
3478 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3479 * fold to the ASCII range (and there are no existing ones in the
3480 * upper latin1 range). But, as outlined in the comments preceding
3481 * this function, we need to flag any occurrences of the sharp s.
3482 * This character forbids trie formation (because of added
3485 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3486 OP(scan) = EXACTFA_NO_TRIE;
3487 *unfolded_multi_char = TRUE;
3496 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3497 * folds that are all Latin1. As explained in the comments
3498 * preceding this function, we look also for the sharp s in EXACTF
3499 * and EXACTFL nodes; it can be in the final position. Otherwise
3500 * we can stop looking 1 byte earlier because have to find at least
3501 * two characters for a multi-fold */
3502 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3507 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3508 if (! len) { /* Not a multi-char fold. */
3509 if (*s == LATIN_SMALL_LETTER_SHARP_S
3510 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3512 *unfolded_multi_char = TRUE;
3519 && isARG2_lower_or_UPPER_ARG1('s', *s)
3520 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
3523 /* EXACTF nodes need to know that the minimum length
3524 * changed so that a sharp s in the string can match this
3525 * ss in the pattern, but they remain EXACTF nodes, as they
3526 * won't match this unless the target string is is UTF-8,
3527 * which we don't know until runtime. EXACTFL nodes can't
3528 * transform into EXACTFU nodes */
3529 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3530 OP(scan) = EXACTFU_SS;
3534 *min_subtract += len - 1;
3541 /* Allow dumping but overwriting the collection of skipped
3542 * ops and/or strings with fake optimized ops */
3543 n = scan + NODE_SZ_STR(scan);
3551 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3555 /* REx optimizer. Converts nodes into quicker variants "in place".
3556 Finds fixed substrings. */
3558 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3559 to the position after last scanned or to NULL. */
3561 #define INIT_AND_WITHP \
3562 assert(!and_withp); \
3563 Newx(and_withp,1, regnode_ssc); \
3564 SAVEFREEPV(and_withp)
3566 /* this is a chain of data about sub patterns we are processing that
3567 need to be handled separately/specially in study_chunk. Its so
3568 we can simulate recursion without losing state. */
3570 typedef struct scan_frame {
3571 regnode *last; /* last node to process in this frame */
3572 regnode *next; /* next node to process when last is reached */
3573 struct scan_frame *prev; /*previous frame*/
3574 U32 prev_recursed_depth;
3575 I32 stop; /* what stopparen do we use */
3580 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3581 SSize_t *minlenp, SSize_t *deltap,
3586 regnode_ssc *and_withp,
3587 U32 flags, U32 depth)
3588 /* scanp: Start here (read-write). */
3589 /* deltap: Write maxlen-minlen here. */
3590 /* last: Stop before this one. */
3591 /* data: string data about the pattern */
3592 /* stopparen: treat close N as END */
3593 /* recursed: which subroutines have we recursed into */
3594 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3596 /* There must be at least this number of characters to match */
3599 regnode *scan = *scanp, *next;
3601 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3602 int is_inf_internal = 0; /* The studied chunk is infinite */
3603 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3604 scan_data_t data_fake;
3605 SV *re_trie_maxbuff = NULL;
3606 regnode *first_non_open = scan;
3607 SSize_t stopmin = SSize_t_MAX;
3608 scan_frame *frame = NULL;
3609 GET_RE_DEBUG_FLAGS_DECL;
3611 PERL_ARGS_ASSERT_STUDY_CHUNK;
3614 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3617 while (first_non_open && OP(first_non_open) == OPEN)
3618 first_non_open=regnext(first_non_open);
3623 while ( scan && OP(scan) != END && scan < last ){
3624 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3625 node length to get a real minimum (because
3626 the folded version may be shorter) */
3627 bool unfolded_multi_char = FALSE;
3628 /* Peephole optimizer: */
3629 DEBUG_OPTIMISE_MORE_r(
3631 PerlIO_printf(Perl_debug_log,
3632 "%*sstudy_chunk stopparen=%ld depth=%lu recursed_depth=%lu ",
3633 ((int) depth*2), "", (long)stopparen,
3634 (unsigned long)depth, (unsigned long)recursed_depth);
3635 if (recursed_depth) {
3638 for ( j = 0 ; j < recursed_depth ; j++ ) {
3639 PerlIO_printf(Perl_debug_log,"[");
3640 for ( i = 0 ; i < (U32)RExC_npar ; i++ )
3641 PerlIO_printf(Perl_debug_log,"%d",
3642 PAREN_TEST(RExC_study_chunk_recursed +
3643 (j * RExC_study_chunk_recursed_bytes), i)
3646 PerlIO_printf(Perl_debug_log,"]");
3649 PerlIO_printf(Perl_debug_log,"\n");
3652 DEBUG_STUDYDATA("Peep:", data, depth);
3653 DEBUG_PEEP("Peep", scan, depth);
3656 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3657 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3658 * by a different invocation of reg() -- Yves
3660 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3662 /* Follow the next-chain of the current node and optimize
3663 away all the NOTHINGs from it. */
3664 if (OP(scan) != CURLYX) {
3665 const int max = (reg_off_by_arg[OP(scan)]
3667 /* I32 may be smaller than U16 on CRAYs! */
3668 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3669 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3673 /* Skip NOTHING and LONGJMP. */
3674 while ((n = regnext(n))
3675 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3676 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3677 && off + noff < max)
3679 if (reg_off_by_arg[OP(scan)])
3682 NEXT_OFF(scan) = off;
3687 /* The principal pseudo-switch. Cannot be a switch, since we
3688 look into several different things. */
3689 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3690 || OP(scan) == IFTHEN) {
3691 next = regnext(scan);
3693 /* demq: the op(next)==code check is to see if we have
3694 * "branch-branch" AFAICT */
3696 if (OP(next) == code || code == IFTHEN) {
3697 /* NOTE - There is similar code to this block below for
3698 * handling TRIE nodes on a re-study. If you change stuff here
3699 * check there too. */
3700 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3702 regnode * const startbranch=scan;
3704 if (flags & SCF_DO_SUBSTR) {
3705 /* Cannot merge strings after this. */
3706 scan_commit(pRExC_state, data, minlenp, is_inf);
3709 if (flags & SCF_DO_STCLASS)
3710 ssc_init_zero(pRExC_state, &accum);
3712 while (OP(scan) == code) {
3713 SSize_t deltanext, minnext, fake;
3715 regnode_ssc this_class;
3718 data_fake.flags = 0;
3720 data_fake.whilem_c = data->whilem_c;
3721 data_fake.last_closep = data->last_closep;
3724 data_fake.last_closep = &fake;
3726 data_fake.pos_delta = delta;
3727 next = regnext(scan);
3728 scan = NEXTOPER(scan);
3730 scan = NEXTOPER(scan);
3731 if (flags & SCF_DO_STCLASS) {
3732 ssc_init(pRExC_state, &this_class);
3733 data_fake.start_class = &this_class;
3734 f = SCF_DO_STCLASS_AND;
3736 if (flags & SCF_WHILEM_VISITED_POS)
3737 f |= SCF_WHILEM_VISITED_POS;
3739 /* we suppose the run is continuous, last=next...*/
3740 minnext = study_chunk(pRExC_state, &scan, minlenp,
3741 &deltanext, next, &data_fake, stopparen,
3742 recursed_depth, NULL, f,depth+1);
3745 if (deltanext == SSize_t_MAX) {
3746 is_inf = is_inf_internal = 1;
3748 } else if (max1 < minnext + deltanext)
3749 max1 = minnext + deltanext;
3751 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3753 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3754 if ( stopmin > minnext)
3755 stopmin = min + min1;
3756 flags &= ~SCF_DO_SUBSTR;
3758 data->flags |= SCF_SEEN_ACCEPT;
3761 if (data_fake.flags & SF_HAS_EVAL)
3762 data->flags |= SF_HAS_EVAL;
3763 data->whilem_c = data_fake.whilem_c;
3765 if (flags & SCF_DO_STCLASS)
3766 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3768 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3770 if (flags & SCF_DO_SUBSTR) {
3771 data->pos_min += min1;
3772 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3773 data->pos_delta = SSize_t_MAX;
3775 data->pos_delta += max1 - min1;
3776 if (max1 != min1 || is_inf)
3777 data->longest = &(data->longest_float);
3780 if (delta == SSize_t_MAX
3781 || SSize_t_MAX - delta - (max1 - min1) < 0)
3782 delta = SSize_t_MAX;
3784 delta += max1 - min1;
3785 if (flags & SCF_DO_STCLASS_OR) {
3786 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
3788 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
3789 flags &= ~SCF_DO_STCLASS;
3792 else if (flags & SCF_DO_STCLASS_AND) {
3794 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
3795 flags &= ~SCF_DO_STCLASS;
3798 /* Switch to OR mode: cache the old value of
3799 * data->start_class */
3801 StructCopy(data->start_class, and_withp, regnode_ssc);
3802 flags &= ~SCF_DO_STCLASS_AND;
3803 StructCopy(&accum, data->start_class, regnode_ssc);
3804 flags |= SCF_DO_STCLASS_OR;
3808 if (PERL_ENABLE_TRIE_OPTIMISATION &&
3809 OP( startbranch ) == BRANCH )
3813 Assuming this was/is a branch we are dealing with: 'scan'
3814 now points at the item that follows the branch sequence,
3815 whatever it is. We now start at the beginning of the
3816 sequence and look for subsequences of
3822 which would be constructed from a pattern like
3825 If we can find such a subsequence we need to turn the first
3826 element into a trie and then add the subsequent branch exact
3827 strings to the trie.
3831 1. patterns where the whole set of branches can be
3834 2. patterns where only a subset can be converted.
3836 In case 1 we can replace the whole set with a single regop
3837 for the trie. In case 2 we need to keep the start and end
3840 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3841 becomes BRANCH TRIE; BRANCH X;
3843 There is an additional case, that being where there is a
3844 common prefix, which gets split out into an EXACT like node
3845 preceding the TRIE node.
3847 If x(1..n)==tail then we can do a simple trie, if not we make
3848 a "jump" trie, such that when we match the appropriate word
3849 we "jump" to the appropriate tail node. Essentially we turn
3850 a nested if into a case structure of sorts.
3855 if (!re_trie_maxbuff) {
3856 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3857 if (!SvIOK(re_trie_maxbuff))
3858 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3860 if ( SvIV(re_trie_maxbuff)>=0 ) {
3862 regnode *first = (regnode *)NULL;
3863 regnode *last = (regnode *)NULL;
3864 regnode *tail = scan;
3869 SV * const mysv = sv_newmortal(); /* for dumping */
3871 /* var tail is used because there may be a TAIL
3872 regop in the way. Ie, the exacts will point to the
3873 thing following the TAIL, but the last branch will
3874 point at the TAIL. So we advance tail. If we
3875 have nested (?:) we may have to move through several
3879 while ( OP( tail ) == TAIL ) {
3880 /* this is the TAIL generated by (?:) */
3881 tail = regnext( tail );
3885 DEBUG_TRIE_COMPILE_r({
3886 regprop(RExC_rx, mysv, tail, NULL);
3887 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3888 (int)depth * 2 + 2, "",
3889 "Looking for TRIE'able sequences. Tail node is: ",
3890 SvPV_nolen_const( mysv )
3896 Step through the branches
3897 cur represents each branch,
3898 noper is the first thing to be matched as part
3900 noper_next is the regnext() of that node.
3902 We normally handle a case like this
3903 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
3904 support building with NOJUMPTRIE, which restricts
3905 the trie logic to structures like /FOO|BAR/.
3907 If noper is a trieable nodetype then the branch is
3908 a possible optimization target. If we are building
3909 under NOJUMPTRIE then we require that noper_next is
3910 the same as scan (our current position in the regex
3913 Once we have two or more consecutive such branches
3914 we can create a trie of the EXACT's contents and
3915 stitch it in place into the program.
3917 If the sequence represents all of the branches in
3918 the alternation we replace the entire thing with a
3921 Otherwise when it is a subsequence we need to
3922 stitch it in place and replace only the relevant
3923 branches. This means the first branch has to remain
3924 as it is used by the alternation logic, and its
3925 next pointer, and needs to be repointed at the item
3926 on the branch chain following the last branch we
3927 have optimized away.
3929 This could be either a BRANCH, in which case the
3930 subsequence is internal, or it could be the item
3931 following the branch sequence in which case the
3932 subsequence is at the end (which does not
3933 necessarily mean the first node is the start of the
3936 TRIE_TYPE(X) is a define which maps the optype to a
3940 ----------------+-----------
3944 EXACTFU_SS | EXACTFU
3949 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3950 ( EXACT == (X) ) ? EXACT : \
3951 ( EXACTFU == (X) || EXACTFU_SS == (X) ) ? EXACTFU : \
3952 ( EXACTFA == (X) ) ? EXACTFA : \
3955 /* dont use tail as the end marker for this traverse */
3956 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3957 regnode * const noper = NEXTOPER( cur );
3958 U8 noper_type = OP( noper );
3959 U8 noper_trietype = TRIE_TYPE( noper_type );
3960 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3961 regnode * const noper_next = regnext( noper );
3962 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3963 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3966 DEBUG_TRIE_COMPILE_r({
3967 regprop(RExC_rx, mysv, cur, NULL);
3968 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3969 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3971 regprop(RExC_rx, mysv, noper, NULL);
3972 PerlIO_printf( Perl_debug_log, " -> %s",
3973 SvPV_nolen_const(mysv));
3976 regprop(RExC_rx, mysv, noper_next, NULL);
3977 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3978 SvPV_nolen_const(mysv));
3980 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3981 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3982 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3986 /* Is noper a trieable nodetype that can be merged
3987 * with the current trie (if there is one)? */
3991 ( noper_trietype == NOTHING)
3992 || ( trietype == NOTHING )
3993 || ( trietype == noper_trietype )
3996 && noper_next == tail
4000 /* Handle mergable triable node Either we are
4001 * the first node in a new trieable sequence,
4002 * in which case we do some bookkeeping,
4003 * otherwise we update the end pointer. */
4006 if ( noper_trietype == NOTHING ) {
4007 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4008 regnode * const noper_next = regnext( noper );
4009 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4010 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4013 if ( noper_next_trietype ) {
4014 trietype = noper_next_trietype;
4015 } else if (noper_next_type) {
4016 /* a NOTHING regop is 1 regop wide.
4017 * We need at least two for a trie
4018 * so we can't merge this in */
4022 trietype = noper_trietype;
4025 if ( trietype == NOTHING )
4026 trietype = noper_trietype;
4031 } /* end handle mergable triable node */
4033 /* handle unmergable node -
4034 * noper may either be a triable node which can
4035 * not be tried together with the current trie,
4036 * or a non triable node */
4038 /* If last is set and trietype is not
4039 * NOTHING then we have found at least two
4040 * triable branch sequences in a row of a
4041 * similar trietype so we can turn them
4042 * into a trie. If/when we allow NOTHING to
4043 * start a trie sequence this condition
4044 * will be required, and it isn't expensive
4045 * so we leave it in for now. */
4046 if ( trietype && trietype != NOTHING )
4047 make_trie( pRExC_state,
4048 startbranch, first, cur, tail,
4049 count, trietype, depth+1 );
4050 last = NULL; /* note: we clear/update
4051 first, trietype etc below,
4052 so we dont do it here */
4056 && noper_next == tail
4059 /* noper is triable, so we can start a new
4063 trietype = noper_trietype;
4065 /* if we already saw a first but the
4066 * current node is not triable then we have
4067 * to reset the first information. */
4072 } /* end handle unmergable node */
4073 } /* loop over branches */
4074 DEBUG_TRIE_COMPILE_r({
4075 regprop(RExC_rx, mysv, cur, NULL);
4076 PerlIO_printf( Perl_debug_log,
4077 "%*s- %s (%d) <SCAN FINISHED>\n",
4079 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4082 if ( last && trietype ) {
4083 if ( trietype != NOTHING ) {
4084 /* the last branch of the sequence was part of
4085 * a trie, so we have to construct it here
4086 * outside of the loop */
4087 made= make_trie( pRExC_state, startbranch,
4088 first, scan, tail, count,
4089 trietype, depth+1 );
4090 #ifdef TRIE_STUDY_OPT
4091 if ( ((made == MADE_EXACT_TRIE &&
4092 startbranch == first)
4093 || ( first_non_open == first )) &&
4095 flags |= SCF_TRIE_RESTUDY;
4096 if ( startbranch == first
4099 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4104 /* at this point we know whatever we have is a
4105 * NOTHING sequence/branch AND if 'startbranch'
4106 * is 'first' then we can turn the whole thing
4109 if ( startbranch == first ) {
4111 /* the entire thing is a NOTHING sequence,
4112 * something like this: (?:|) So we can
4113 * turn it into a plain NOTHING op. */
4114 DEBUG_TRIE_COMPILE_r({
4115 regprop(RExC_rx, mysv, cur, NULL);
4116 PerlIO_printf( Perl_debug_log,
4117 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4118 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
4121 OP(startbranch)= NOTHING;
4122 NEXT_OFF(startbranch)= tail - startbranch;
4123 for ( opt= startbranch + 1; opt < tail ; opt++ )
4127 } /* end if ( last) */
4128 } /* TRIE_MAXBUF is non zero */
4133 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4134 scan = NEXTOPER(NEXTOPER(scan));
4135 } else /* single branch is optimized. */
4136 scan = NEXTOPER(scan);
4138 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4139 scan_frame *newframe = NULL;
4143 U32 my_recursed_depth= recursed_depth;
4145 if (OP(scan) != SUSPEND) {
4146 /* set the pointer */
4147 if (OP(scan) == GOSUB) {
4149 RExC_recurse[ARG2L(scan)] = scan;
4150 start = RExC_open_parens[paren-1];
4151 end = RExC_close_parens[paren-1];
4154 start = RExC_rxi->program + 1;
4159 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4161 if (!recursed_depth) {
4162 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4164 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4165 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4166 RExC_study_chunk_recursed_bytes, U8);
4168 /* we havent recursed into this paren yet, so recurse into it */
4169 DEBUG_STUDYDATA("set:", data,depth);
4170 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4171 my_recursed_depth= recursed_depth + 1;
4172 Newx(newframe,1,scan_frame);
4174 DEBUG_STUDYDATA("inf:", data,depth);
4175 /* some form of infinite recursion, assume infinite length
4177 if (flags & SCF_DO_SUBSTR) {
4178 scan_commit(pRExC_state, data, minlenp, is_inf);
4179 data->longest = &(data->longest_float);
4181 is_inf = is_inf_internal = 1;
4182 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4183 ssc_anything(data->start_class);
4184 flags &= ~SCF_DO_STCLASS;
4187 Newx(newframe,1,scan_frame);
4190 end = regnext(scan);
4195 SAVEFREEPV(newframe);
4196 newframe->next = regnext(scan);
4197 newframe->last = last;
4198 newframe->stop = stopparen;
4199 newframe->prev = frame;
4200 newframe->prev_recursed_depth = recursed_depth;
4202 DEBUG_STUDYDATA("frame-new:",data,depth);
4203 DEBUG_PEEP("fnew", scan, depth);
4210 recursed_depth= my_recursed_depth;
4215 else if (OP(scan) == EXACT) {
4216 SSize_t l = STR_LEN(scan);
4219 const U8 * const s = (U8*)STRING(scan);
4220 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4221 l = utf8_length(s, s + l);
4223 uc = *((U8*)STRING(scan));
4226 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4227 /* The code below prefers earlier match for fixed
4228 offset, later match for variable offset. */
4229 if (data->last_end == -1) { /* Update the start info. */
4230 data->last_start_min = data->pos_min;
4231 data->last_start_max = is_inf
4232 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4234 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4236 SvUTF8_on(data->last_found);
4238 SV * const sv = data->last_found;
4239 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4240 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4241 if (mg && mg->mg_len >= 0)
4242 mg->mg_len += utf8_length((U8*)STRING(scan),
4243 (U8*)STRING(scan)+STR_LEN(scan));
4245 data->last_end = data->pos_min + l;
4246 data->pos_min += l; /* As in the first entry. */
4247 data->flags &= ~SF_BEFORE_EOL;
4250 /* ANDing the code point leaves at most it, and not in locale, and
4251 * can't match null string */
4252 if (flags & SCF_DO_STCLASS_AND) {
4253 ssc_cp_and(data->start_class, uc);
4254 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4255 ssc_clear_locale(data->start_class);
4257 else if (flags & SCF_DO_STCLASS_OR) {
4258 ssc_add_cp(data->start_class, uc);
4259 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4261 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4262 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4264 flags &= ~SCF_DO_STCLASS;
4266 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT!, so is
4268 SSize_t l = STR_LEN(scan);
4269 UV uc = *((U8*)STRING(scan));
4270 SV* EXACTF_invlist = _new_invlist(4); /* Start out big enough for 2
4271 separate code points */
4272 const U8 * s = (U8*)STRING(scan);
4274 /* Search for fixed substrings supports EXACT only. */
4275 if (flags & SCF_DO_SUBSTR) {
4277 scan_commit(pRExC_state, data, minlenp, is_inf);
4280 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4281 l = utf8_length(s, s + l);
4283 if (unfolded_multi_char) {
4284 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4286 min += l - min_subtract;
4288 delta += min_subtract;
4289 if (flags & SCF_DO_SUBSTR) {
4290 data->pos_min += l - min_subtract;
4291 if (data->pos_min < 0) {
4294 data->pos_delta += min_subtract;
4296 data->longest = &(data->longest_float);
4300 if (OP(scan) != EXACTFL && flags & SCF_DO_STCLASS_AND) {
4301 ssc_clear_locale(data->start_class);
4306 /* We punt and assume can match anything if the node begins
4307 * with a multi-character fold. Things are complicated. For
4308 * example, /ffi/i could match any of:
4309 * "\N{LATIN SMALL LIGATURE FFI}"
4310 * "\N{LATIN SMALL LIGATURE FF}I"
4311 * "F\N{LATIN SMALL LIGATURE FI}"
4312 * plus several other things; and making sure we have all the
4313 * possibilities is hard. */
4314 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + STR_LEN(scan))) {
4316 _add_range_to_invlist(EXACTF_invlist, 0, UV_MAX);
4320 /* Any Latin1 range character can potentially match any
4321 * other depending on the locale */
4322 if (OP(scan) == EXACTFL) {
4323 _invlist_union(EXACTF_invlist, PL_Latin1,
4327 /* But otherwise, it matches at least itself. We can
4328 * quickly tell if it has a distinct fold, and if so,
4329 * it matches that as well */
4330 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4331 if (IS_IN_SOME_FOLD_L1(uc)) {
4332 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist,
4333 PL_fold_latin1[uc]);
4337 /* Some characters match above-Latin1 ones under /i. This
4338 * is true of EXACTFL ones when the locale is UTF-8 */
4339 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
4340 && (! isASCII(uc) || (OP(scan) != EXACTFA
4341 && OP(scan) != EXACTFA_NO_TRIE)))
4343 add_above_Latin1_folds(pRExC_state,
4349 else { /* Pattern is UTF-8 */
4350 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
4351 STRLEN foldlen = UTF8SKIP(s);
4352 const U8* e = s + STR_LEN(scan);
4355 /* The only code points that aren't folded in a UTF EXACTFish
4356 * node are are the problematic ones in EXACTFL nodes */
4357 if (OP(scan) == EXACTFL
4358 && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc))
4360 /* We need to check for the possibility that this EXACTFL
4361 * node begins with a multi-char fold. Therefore we fold
4362 * the first few characters of it so that we can make that
4367 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
4369 *(d++) = (U8) toFOLD(*s);
4374 to_utf8_fold(s, d, &len);
4380 /* And set up so the code below that looks in this folded
4381 * buffer instead of the node's string */
4383 foldlen = UTF8SKIP(folded);
4387 /* When we reach here 's' points to the fold of the first
4388 * character(s) of the node; and 'e' points to far enough along
4389 * the folded string to be just past any possible multi-char
4390 * fold. 'foldlen' is the length in bytes of the first
4393 * Unlike the non-UTF-8 case, the macro for determining if a
4394 * string is a multi-char fold requires all the characters to
4395 * already be folded. This is because of all the complications
4396 * if not. Note that they are folded anyway, except in EXACTFL
4397 * nodes. Like the non-UTF case above, we punt if the node
4398 * begins with a multi-char fold */
4400 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
4402 _add_range_to_invlist(EXACTF_invlist, 0, UV_MAX);
4404 else { /* Single char fold */
4406 /* It matches all the things that fold to it, which are
4407 * found in PL_utf8_foldclosures (including itself) */
4408 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, uc);
4409 if (! PL_utf8_foldclosures) {
4410 _load_PL_utf8_foldclosures();
4412 if ((listp = hv_fetch(PL_utf8_foldclosures,
4413 (char *) s, foldlen, FALSE)))
4415 AV* list = (AV*) *listp;
4417 for (k = 0; k <= av_tindex(list); k++) {
4418 SV** c_p = av_fetch(list, k, FALSE);
4424 /* /aa doesn't allow folds between ASCII and non- */
4425 if ((OP(scan) == EXACTFA || OP(scan) == EXACTFA_NO_TRIE)
4426 && isASCII(c) != isASCII(uc))
4431 EXACTF_invlist = add_cp_to_invlist(EXACTF_invlist, c);
4436 if (flags & SCF_DO_STCLASS_AND) {
4437 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4438 ANYOF_POSIXL_ZERO(data->start_class);
4439 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4441 else if (flags & SCF_DO_STCLASS_OR) {
4442 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4443 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4445 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4446 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4448 flags &= ~SCF_DO_STCLASS;
4449 SvREFCNT_dec(EXACTF_invlist);
4451 else if (REGNODE_VARIES(OP(scan))) {
4452 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4453 I32 fl = 0, f = flags;
4454 regnode * const oscan = scan;
4455 regnode_ssc this_class;
4456 regnode_ssc *oclass = NULL;
4457 I32 next_is_eval = 0;
4459 switch (PL_regkind[OP(scan)]) {
4460 case WHILEM: /* End of (?:...)* . */
4461 scan = NEXTOPER(scan);
4464 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4465 next = NEXTOPER(scan);
4466 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
4468 maxcount = REG_INFTY;
4469 next = regnext(scan);
4470 scan = NEXTOPER(scan);
4474 if (flags & SCF_DO_SUBSTR)
4479 if (flags & SCF_DO_STCLASS) {
4481 maxcount = REG_INFTY;
4482 next = regnext(scan);
4483 scan = NEXTOPER(scan);
4486 if (flags & SCF_DO_SUBSTR) {
4487 scan_commit(pRExC_state, data, minlenp, is_inf);
4488 /* Cannot extend fixed substrings */
4489 data->longest = &(data->longest_float);
4491 is_inf = is_inf_internal = 1;
4492 scan = regnext(scan);
4493 goto optimize_curly_tail;
4495 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4496 && (scan->flags == stopparen))
4501 mincount = ARG1(scan);
4502 maxcount = ARG2(scan);
4504 next = regnext(scan);
4505 if (OP(scan) == CURLYX) {
4506 I32 lp = (data ? *(data->last_closep) : 0);
4507 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4509 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4510 next_is_eval = (OP(scan) == EVAL);
4512 if (flags & SCF_DO_SUBSTR) {
4514 scan_commit(pRExC_state, data, minlenp, is_inf);
4515 /* Cannot extend fixed substrings */
4516 pos_before = data->pos_min;
4520 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4522 data->flags |= SF_IS_INF;
4524 if (flags & SCF_DO_STCLASS) {
4525 ssc_init(pRExC_state, &this_class);
4526 oclass = data->start_class;
4527 data->start_class = &this_class;
4528 f |= SCF_DO_STCLASS_AND;
4529 f &= ~SCF_DO_STCLASS_OR;
4531 /* Exclude from super-linear cache processing any {n,m}
4532 regops for which the combination of input pos and regex
4533 pos is not enough information to determine if a match
4536 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4537 regex pos at the \s*, the prospects for a match depend not
4538 only on the input position but also on how many (bar\s*)
4539 repeats into the {4,8} we are. */
4540 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4541 f &= ~SCF_WHILEM_VISITED_POS;
4543 /* This will finish on WHILEM, setting scan, or on NULL: */
4544 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4545 last, data, stopparen, recursed_depth, NULL,
4547 ? (f & ~SCF_DO_SUBSTR)
4551 if (flags & SCF_DO_STCLASS)
4552 data->start_class = oclass;
4553 if (mincount == 0 || minnext == 0) {
4554 if (flags & SCF_DO_STCLASS_OR) {
4555 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4557 else if (flags & SCF_DO_STCLASS_AND) {
4558 /* Switch to OR mode: cache the old value of
4559 * data->start_class */
4561 StructCopy(data->start_class, and_withp, regnode_ssc);
4562 flags &= ~SCF_DO_STCLASS_AND;
4563 StructCopy(&this_class, data->start_class, regnode_ssc);
4564 flags |= SCF_DO_STCLASS_OR;
4565 ANYOF_FLAGS(data->start_class) |= ANYOF_EMPTY_STRING;
4567 } else { /* Non-zero len */
4568 if (flags & SCF_DO_STCLASS_OR) {
4569 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4570 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4572 else if (flags & SCF_DO_STCLASS_AND)
4573 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4574 flags &= ~SCF_DO_STCLASS;
4576 if (!scan) /* It was not CURLYX, but CURLY. */
4578 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4579 /* ? quantifier ok, except for (?{ ... }) */
4580 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4581 && (minnext == 0) && (deltanext == 0)
4582 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4583 && maxcount <= REG_INFTY/3) /* Complement check for big
4586 /* Fatal warnings may leak the regexp without this: */
4587 SAVEFREESV(RExC_rx_sv);
4588 ckWARNreg(RExC_parse,
4589 "Quantifier unexpected on zero-length expression");
4590 (void)ReREFCNT_inc(RExC_rx_sv);
4593 min += minnext * mincount;
4594 is_inf_internal |= deltanext == SSize_t_MAX
4595 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4596 is_inf |= is_inf_internal;
4598 delta = SSize_t_MAX;
4600 delta += (minnext + deltanext) * maxcount
4601 - minnext * mincount;
4603 /* Try powerful optimization CURLYX => CURLYN. */
4604 if ( OP(oscan) == CURLYX && data
4605 && data->flags & SF_IN_PAR
4606 && !(data->flags & SF_HAS_EVAL)
4607 && !deltanext && minnext == 1 ) {
4608 /* Try to optimize to CURLYN. */
4609 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4610 regnode * const nxt1 = nxt;
4617 if (!REGNODE_SIMPLE(OP(nxt))
4618 && !(PL_regkind[OP(nxt)] == EXACT
4619 && STR_LEN(nxt) == 1))
4625 if (OP(nxt) != CLOSE)
4627 if (RExC_open_parens) {
4628 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4629 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4631 /* Now we know that nxt2 is the only contents: */
4632 oscan->flags = (U8)ARG(nxt);
4634 OP(nxt1) = NOTHING; /* was OPEN. */
4637 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4638 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4639 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4640 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4641 OP(nxt + 1) = OPTIMIZED; /* was count. */
4642 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4647 /* Try optimization CURLYX => CURLYM. */
4648 if ( OP(oscan) == CURLYX && data
4649 && !(data->flags & SF_HAS_PAR)
4650 && !(data->flags & SF_HAS_EVAL)
4651 && !deltanext /* atom is fixed width */
4652 && minnext != 0 /* CURLYM can't handle zero width */
4654 /* Nor characters whose fold at run-time may be
4655 * multi-character */
4656 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4658 /* XXXX How to optimize if data == 0? */
4659 /* Optimize to a simpler form. */
4660 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4664 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4665 && (OP(nxt2) != WHILEM))
4667 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4668 /* Need to optimize away parenths. */
4669 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4670 /* Set the parenth number. */
4671 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4673 oscan->flags = (U8)ARG(nxt);
4674 if (RExC_open_parens) {
4675 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4676 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4678 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4679 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4682 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4683 OP(nxt + 1) = OPTIMIZED; /* was count. */
4684 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4685 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4688 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4689 regnode *nnxt = regnext(nxt1);
4691 if (reg_off_by_arg[OP(nxt1)])
4692 ARG_SET(nxt1, nxt2 - nxt1);
4693 else if (nxt2 - nxt1 < U16_MAX)
4694 NEXT_OFF(nxt1) = nxt2 - nxt1;
4696 OP(nxt) = NOTHING; /* Cannot beautify */
4701 /* Optimize again: */
4702 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4703 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4708 else if ((OP(oscan) == CURLYX)
4709 && (flags & SCF_WHILEM_VISITED_POS)
4710 /* See the comment on a similar expression above.
4711 However, this time it's not a subexpression
4712 we care about, but the expression itself. */
4713 && (maxcount == REG_INFTY)
4714 && data && ++data->whilem_c < 16) {
4715 /* This stays as CURLYX, we can put the count/of pair. */
4716 /* Find WHILEM (as in regexec.c) */
4717 regnode *nxt = oscan + NEXT_OFF(oscan);
4719 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4721 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4722 | (RExC_whilem_seen << 4)); /* On WHILEM */
4724 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4726 if (flags & SCF_DO_SUBSTR) {
4727 SV *last_str = NULL;
4728 STRLEN last_chrs = 0;
4729 int counted = mincount != 0;
4731 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4733 SSize_t b = pos_before >= data->last_start_min
4734 ? pos_before : data->last_start_min;
4736 const char * const s = SvPV_const(data->last_found, l);
4737 SSize_t old = b - data->last_start_min;
4740 old = utf8_hop((U8*)s, old) - (U8*)s;
4742 /* Get the added string: */
4743 last_str = newSVpvn_utf8(s + old, l, UTF);
4744 last_chrs = UTF ? utf8_length((U8*)(s + old),
4745 (U8*)(s + old + l)) : l;
4746 if (deltanext == 0 && pos_before == b) {
4747 /* What was added is a constant string */
4750 SvGROW(last_str, (mincount * l) + 1);
4751 repeatcpy(SvPVX(last_str) + l,
4752 SvPVX_const(last_str), l,
4754 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4755 /* Add additional parts. */
4756 SvCUR_set(data->last_found,
4757 SvCUR(data->last_found) - l);
4758 sv_catsv(data->last_found, last_str);
4760 SV * sv = data->last_found;
4762 SvUTF8(sv) && SvMAGICAL(sv) ?
4763 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4764 if (mg && mg->mg_len >= 0)
4765 mg->mg_len += last_chrs * (mincount-1);
4767 last_chrs *= mincount;
4768 data->last_end += l * (mincount - 1);
4771 /* start offset must point into the last copy */
4772 data->last_start_min += minnext * (mincount - 1);
4773 data->last_start_max += is_inf ? SSize_t_MAX
4774 : (maxcount - 1) * (minnext + data->pos_delta);
4777 /* It is counted once already... */
4778 data->pos_min += minnext * (mincount - counted);
4780 PerlIO_printf(Perl_debug_log, "counted=%"UVuf" deltanext=%"UVuf
4781 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
4782 " maxcount=%"UVuf" mincount=%"UVuf"\n",
4783 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4785 if (deltanext != SSize_t_MAX)
4786 PerlIO_printf(Perl_debug_log, "LHS=%"UVuf" RHS=%"UVuf"\n",
4787 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4788 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4790 if (deltanext == SSize_t_MAX
4791 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4792 data->pos_delta = SSize_t_MAX;
4794 data->pos_delta += - counted * deltanext +
4795 (minnext + deltanext) * maxcount - minnext * mincount;
4796 if (mincount != maxcount) {
4797 /* Cannot extend fixed substrings found inside
4799 scan_commit(pRExC_state, data, minlenp, is_inf);
4800 if (mincount && last_str) {
4801 SV * const sv = data->last_found;
4802 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4803 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4807 sv_setsv(sv, last_str);
4808 data->last_end = data->pos_min;
4809 data->last_start_min = data->pos_min - last_chrs;
4810 data->last_start_max = is_inf
4812 : data->pos_min + data->pos_delta - last_chrs;
4814 data->longest = &(data->longest_float);
4816 SvREFCNT_dec(last_str);
4818 if (data && (fl & SF_HAS_EVAL))
4819 data->flags |= SF_HAS_EVAL;
4820 optimize_curly_tail:
4821 if (OP(oscan) != CURLYX) {
4822 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4824 NEXT_OFF(oscan) += NEXT_OFF(next);
4830 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4835 if (flags & SCF_DO_SUBSTR) {
4836 /* Cannot expect anything... */
4837 scan_commit(pRExC_state, data, minlenp, is_inf);
4838 data->longest = &(data->longest_float);
4840 is_inf = is_inf_internal = 1;
4841 if (flags & SCF_DO_STCLASS_OR) {
4842 if (OP(scan) == CLUMP) {
4843 /* Actually is any start char, but very few code points
4844 * aren't start characters */
4845 ssc_match_all_cp(data->start_class);
4848 ssc_anything(data->start_class);
4851 flags &= ~SCF_DO_STCLASS;
4855 else if (OP(scan) == LNBREAK) {
4856 if (flags & SCF_DO_STCLASS) {
4857 if (flags & SCF_DO_STCLASS_AND) {
4858 ssc_intersection(data->start_class,
4859 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
4860 ssc_clear_locale(data->start_class);
4861 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4863 else if (flags & SCF_DO_STCLASS_OR) {
4864 ssc_union(data->start_class,
4865 PL_XPosix_ptrs[_CC_VERTSPACE],
4867 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4869 /* See commit msg for
4870 * 749e076fceedeb708a624933726e7989f2302f6a */
4871 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4873 flags &= ~SCF_DO_STCLASS;
4876 delta++; /* Because of the 2 char string cr-lf */
4877 if (flags & SCF_DO_SUBSTR) {
4878 /* Cannot expect anything... */
4879 scan_commit(pRExC_state, data, minlenp, is_inf);
4881 data->pos_delta += 1;
4882 data->longest = &(data->longest_float);
4885 else if (REGNODE_SIMPLE(OP(scan))) {
4887 if (flags & SCF_DO_SUBSTR) {
4888 scan_commit(pRExC_state, data, minlenp, is_inf);
4892 if (flags & SCF_DO_STCLASS) {
4894 SV* my_invlist = sv_2mortal(_new_invlist(0));
4897 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4898 ANYOF_FLAGS(data->start_class) &= ~ANYOF_EMPTY_STRING;
4900 /* Some of the logic below assumes that switching
4901 locale on will only add false positives. */
4906 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
4911 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4912 ssc_match_all_cp(data->start_class);
4917 SV* REG_ANY_invlist = _new_invlist(2);
4918 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
4920 if (flags & SCF_DO_STCLASS_OR) {
4921 ssc_union(data->start_class,
4923 TRUE /* TRUE => invert, hence all but \n
4927 else if (flags & SCF_DO_STCLASS_AND) {
4928 ssc_intersection(data->start_class,
4930 TRUE /* TRUE => invert */
4932 ssc_clear_locale(data->start_class);
4934 SvREFCNT_dec_NN(REG_ANY_invlist);
4939 if (flags & SCF_DO_STCLASS_AND)
4940 ssc_and(pRExC_state, data->start_class,
4941 (regnode_charclass *) scan);
4943 ssc_or(pRExC_state, data->start_class,
4944 (regnode_charclass *) scan);
4952 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
4953 if (flags & SCF_DO_STCLASS_AND) {
4954 bool was_there = cBOOL(
4955 ANYOF_POSIXL_TEST(data->start_class,
4957 ANYOF_POSIXL_ZERO(data->start_class);
4958 if (was_there) { /* Do an AND */
4959 ANYOF_POSIXL_SET(data->start_class, namedclass);
4961 /* No individual code points can now match */
4962 data->start_class->invlist
4963 = sv_2mortal(_new_invlist(0));
4966 int complement = namedclass + ((invert) ? -1 : 1);
4968 assert(flags & SCF_DO_STCLASS_OR);
4970 /* If the complement of this class was already there,
4971 * the result is that they match all code points,
4972 * (\d + \D == everything). Remove the classes from
4973 * future consideration. Locale is not relevant in
4975 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
4976 ssc_match_all_cp(data->start_class);
4977 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
4978 ANYOF_POSIXL_CLEAR(data->start_class, complement);
4980 else { /* The usual case; just add this class to the
4982 ANYOF_POSIXL_SET(data->start_class, namedclass);
4987 case NPOSIXA: /* For these, we always know the exact set of
4992 if (FLAGS(scan) == _CC_ASCII) {
4993 my_invlist = PL_XPosix_ptrs[_CC_ASCII];
4996 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
4997 PL_XPosix_ptrs[_CC_ASCII],
5008 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5010 /* NPOSIXD matches all upper Latin1 code points unless the
5011 * target string being matched is UTF-8, which is
5012 * unknowable until match time. Since we are going to
5013 * invert, we want to get rid of all of them so that the
5014 * inversion will match all */
5015 if (OP(scan) == NPOSIXD) {
5016 _invlist_subtract(my_invlist, PL_UpperLatin1,
5022 if (flags & SCF_DO_STCLASS_AND) {
5023 ssc_intersection(data->start_class, my_invlist, invert);
5024 ssc_clear_locale(data->start_class);
5027 assert(flags & SCF_DO_STCLASS_OR);
5028 ssc_union(data->start_class, my_invlist, invert);
5031 if (flags & SCF_DO_STCLASS_OR)
5032 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5033 flags &= ~SCF_DO_STCLASS;
5036 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5037 data->flags |= (OP(scan) == MEOL
5040 scan_commit(pRExC_state, data, minlenp, is_inf);
5043 else if ( PL_regkind[OP(scan)] == BRANCHJ
5044 /* Lookbehind, or need to calculate parens/evals/stclass: */
5045 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5046 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
5047 if ( OP(scan) == UNLESSM &&
5049 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
5050 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
5053 regnode *upto= regnext(scan);
5055 SV * const mysv_val=sv_newmortal();
5056 DEBUG_STUDYDATA("OPFAIL",data,depth);
5058 /*DEBUG_PARSE_MSG("opfail");*/
5059 regprop(RExC_rx, mysv_val, upto, NULL);
5060 PerlIO_printf(Perl_debug_log,
5061 "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
5062 SvPV_nolen_const(mysv_val),
5063 (IV)REG_NODE_NUM(upto),
5068 NEXT_OFF(scan) = upto - scan;
5069 for (opt= scan + 1; opt < upto ; opt++)
5070 OP(opt) = OPTIMIZED;
5074 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5075 || OP(scan) == UNLESSM )
5077 /* Negative Lookahead/lookbehind
5078 In this case we can't do fixed string optimisation.
5081 SSize_t deltanext, minnext, fake = 0;
5086 data_fake.flags = 0;
5088 data_fake.whilem_c = data->whilem_c;
5089 data_fake.last_closep = data->last_closep;
5092 data_fake.last_closep = &fake;
5093 data_fake.pos_delta = delta;
5094 if ( flags & SCF_DO_STCLASS && !scan->flags
5095 && OP(scan) == IFMATCH ) { /* Lookahead */
5096 ssc_init(pRExC_state, &intrnl);
5097 data_fake.start_class = &intrnl;
5098 f |= SCF_DO_STCLASS_AND;
5100 if (flags & SCF_WHILEM_VISITED_POS)
5101 f |= SCF_WHILEM_VISITED_POS;
5102 next = regnext(scan);
5103 nscan = NEXTOPER(NEXTOPER(scan));
5104 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5105 last, &data_fake, stopparen,
5106 recursed_depth, NULL, f, depth+1);
5109 FAIL("Variable length lookbehind not implemented");
5111 else if (minnext > (I32)U8_MAX) {
5112 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5115 scan->flags = (U8)minnext;
5118 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5120 if (data_fake.flags & SF_HAS_EVAL)
5121 data->flags |= SF_HAS_EVAL;
5122 data->whilem_c = data_fake.whilem_c;
5124 if (f & SCF_DO_STCLASS_AND) {
5125 if (flags & SCF_DO_STCLASS_OR) {
5126 /* OR before, AND after: ideally we would recurse with
5127 * data_fake to get the AND applied by study of the
5128 * remainder of the pattern, and then derecurse;
5129 * *** HACK *** for now just treat as "no information".
5130 * See [perl #56690].
5132 ssc_init(pRExC_state, data->start_class);
5134 /* AND before and after: combine and continue */
5135 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5139 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5141 /* Positive Lookahead/lookbehind
5142 In this case we can do fixed string optimisation,
5143 but we must be careful about it. Note in the case of
5144 lookbehind the positions will be offset by the minimum
5145 length of the pattern, something we won't know about
5146 until after the recurse.
5148 SSize_t deltanext, fake = 0;
5152 /* We use SAVEFREEPV so that when the full compile
5153 is finished perl will clean up the allocated
5154 minlens when it's all done. This way we don't
5155 have to worry about freeing them when we know
5156 they wont be used, which would be a pain.
5159 Newx( minnextp, 1, SSize_t );
5160 SAVEFREEPV(minnextp);
5163 StructCopy(data, &data_fake, scan_data_t);
5164 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5167 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5168 data_fake.last_found=newSVsv(data->last_found);
5172 data_fake.last_closep = &fake;
5173 data_fake.flags = 0;
5174 data_fake.pos_delta = delta;
5176 data_fake.flags |= SF_IS_INF;
5177 if ( flags & SCF_DO_STCLASS && !scan->flags
5178 && OP(scan) == IFMATCH ) { /* Lookahead */
5179 ssc_init(pRExC_state, &intrnl);
5180 data_fake.start_class = &intrnl;
5181 f |= SCF_DO_STCLASS_AND;
5183 if (flags & SCF_WHILEM_VISITED_POS)
5184 f |= SCF_WHILEM_VISITED_POS;
5185 next = regnext(scan);
5186 nscan = NEXTOPER(NEXTOPER(scan));
5188 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5189 &deltanext, last, &data_fake,
5190 stopparen, recursed_depth, NULL,
5194 FAIL("Variable length lookbehind not implemented");
5196 else if (*minnextp > (I32)U8_MAX) {
5197 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5200 scan->flags = (U8)*minnextp;
5205 if (f & SCF_DO_STCLASS_AND) {
5206 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5209 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5211 if (data_fake.flags & SF_HAS_EVAL)
5212 data->flags |= SF_HAS_EVAL;
5213 data->whilem_c = data_fake.whilem_c;
5214 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5215 if (RExC_rx->minlen<*minnextp)
5216 RExC_rx->minlen=*minnextp;
5217 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5218 SvREFCNT_dec_NN(data_fake.last_found);
5220 if ( data_fake.minlen_fixed != minlenp )
5222 data->offset_fixed= data_fake.offset_fixed;
5223 data->minlen_fixed= data_fake.minlen_fixed;
5224 data->lookbehind_fixed+= scan->flags;
5226 if ( data_fake.minlen_float != minlenp )
5228 data->minlen_float= data_fake.minlen_float;
5229 data->offset_float_min=data_fake.offset_float_min;
5230 data->offset_float_max=data_fake.offset_float_max;
5231 data->lookbehind_float+= scan->flags;
5238 else if (OP(scan) == OPEN) {
5239 if (stopparen != (I32)ARG(scan))
5242 else if (OP(scan) == CLOSE) {
5243 if (stopparen == (I32)ARG(scan)) {
5246 if ((I32)ARG(scan) == is_par) {
5247 next = regnext(scan);
5249 if ( next && (OP(next) != WHILEM) && next < last)
5250 is_par = 0; /* Disable optimization */
5253 *(data->last_closep) = ARG(scan);
5255 else if (OP(scan) == EVAL) {
5257 data->flags |= SF_HAS_EVAL;
5259 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5260 if (flags & SCF_DO_SUBSTR) {
5261 scan_commit(pRExC_state, data, minlenp, is_inf);
5262 flags &= ~SCF_DO_SUBSTR;
5264 if (data && OP(scan)==ACCEPT) {
5265 data->flags |= SCF_SEEN_ACCEPT;
5270 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5272 if (flags & SCF_DO_SUBSTR) {
5273 scan_commit(pRExC_state, data, minlenp, is_inf);
5274 data->longest = &(data->longest_float);
5276 is_inf = is_inf_internal = 1;
5277 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5278 ssc_anything(data->start_class);
5279 flags &= ~SCF_DO_STCLASS;
5281 else if (OP(scan) == GPOS) {
5282 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5283 !(delta || is_inf || (data && data->pos_delta)))
5285 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5286 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5287 if (RExC_rx->gofs < (STRLEN)min)
5288 RExC_rx->gofs = min;
5290 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5294 #ifdef TRIE_STUDY_OPT
5295 #ifdef FULL_TRIE_STUDY
5296 else if (PL_regkind[OP(scan)] == TRIE) {
5297 /* NOTE - There is similar code to this block above for handling
5298 BRANCH nodes on the initial study. If you change stuff here
5300 regnode *trie_node= scan;
5301 regnode *tail= regnext(scan);
5302 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5303 SSize_t max1 = 0, min1 = SSize_t_MAX;
5306 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5307 /* Cannot merge strings after this. */
5308 scan_commit(pRExC_state, data, minlenp, is_inf);
5310 if (flags & SCF_DO_STCLASS)
5311 ssc_init_zero(pRExC_state, &accum);
5317 const regnode *nextbranch= NULL;
5320 for ( word=1 ; word <= trie->wordcount ; word++)
5322 SSize_t deltanext=0, minnext=0, f = 0, fake;
5323 regnode_ssc this_class;
5325 data_fake.flags = 0;
5327 data_fake.whilem_c = data->whilem_c;
5328 data_fake.last_closep = data->last_closep;
5331 data_fake.last_closep = &fake;
5332 data_fake.pos_delta = delta;
5333 if (flags & SCF_DO_STCLASS) {
5334 ssc_init(pRExC_state, &this_class);
5335 data_fake.start_class = &this_class;
5336 f = SCF_DO_STCLASS_AND;
5338 if (flags & SCF_WHILEM_VISITED_POS)
5339 f |= SCF_WHILEM_VISITED_POS;
5341 if (trie->jump[word]) {
5343 nextbranch = trie_node + trie->jump[0];
5344 scan= trie_node + trie->jump[word];
5345 /* We go from the jump point to the branch that follows
5346 it. Note this means we need the vestigal unused
5347 branches even though they arent otherwise used. */
5348 minnext = study_chunk(pRExC_state, &scan, minlenp,
5349 &deltanext, (regnode *)nextbranch, &data_fake,
5350 stopparen, recursed_depth, NULL, f,depth+1);
5352 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5353 nextbranch= regnext((regnode*)nextbranch);
5355 if (min1 > (SSize_t)(minnext + trie->minlen))
5356 min1 = minnext + trie->minlen;
5357 if (deltanext == SSize_t_MAX) {
5358 is_inf = is_inf_internal = 1;
5360 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5361 max1 = minnext + deltanext + trie->maxlen;
5363 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5365 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5366 if ( stopmin > min + min1)
5367 stopmin = min + min1;
5368 flags &= ~SCF_DO_SUBSTR;
5370 data->flags |= SCF_SEEN_ACCEPT;
5373 if (data_fake.flags & SF_HAS_EVAL)
5374 data->flags |= SF_HAS_EVAL;
5375 data->whilem_c = data_fake.whilem_c;
5377 if (flags & SCF_DO_STCLASS)
5378 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5381 if (flags & SCF_DO_SUBSTR) {
5382 data->pos_min += min1;
5383 data->pos_delta += max1 - min1;
5384 if (max1 != min1 || is_inf)
5385 data->longest = &(data->longest_float);
5388 delta += max1 - min1;
5389 if (flags & SCF_DO_STCLASS_OR) {
5390 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5392 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5393 flags &= ~SCF_DO_STCLASS;
5396 else if (flags & SCF_DO_STCLASS_AND) {
5398 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5399 flags &= ~SCF_DO_STCLASS;
5402 /* Switch to OR mode: cache the old value of
5403 * data->start_class */
5405 StructCopy(data->start_class, and_withp, regnode_ssc);
5406 flags &= ~SCF_DO_STCLASS_AND;
5407 StructCopy(&accum, data->start_class, regnode_ssc);
5408 flags |= SCF_DO_STCLASS_OR;
5415 else if (PL_regkind[OP(scan)] == TRIE) {
5416 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5419 min += trie->minlen;
5420 delta += (trie->maxlen - trie->minlen);
5421 flags &= ~SCF_DO_STCLASS; /* xxx */
5422 if (flags & SCF_DO_SUBSTR) {
5423 /* Cannot expect anything... */
5424 scan_commit(pRExC_state, data, minlenp, is_inf);
5425 data->pos_min += trie->minlen;
5426 data->pos_delta += (trie->maxlen - trie->minlen);
5427 if (trie->maxlen != trie->minlen)
5428 data->longest = &(data->longest_float);
5430 if (trie->jump) /* no more substrings -- for now /grr*/
5431 flags &= ~SCF_DO_SUBSTR;
5433 #endif /* old or new */
5434 #endif /* TRIE_STUDY_OPT */
5436 /* Else: zero-length, ignore. */
5437 scan = regnext(scan);
5439 /* If we are exiting a recursion we can unset its recursed bit
5440 * and allow ourselves to enter it again - no danger of an
5441 * infinite loop there.
5442 if (stopparen > -1 && recursed) {
5443 DEBUG_STUDYDATA("unset:", data,depth);
5444 PAREN_UNSET( recursed, stopparen);
5448 DEBUG_STUDYDATA("frame-end:",data,depth);
5449 DEBUG_PEEP("fend", scan, depth);
5450 /* restore previous context */
5453 stopparen = frame->stop;
5454 recursed_depth = frame->prev_recursed_depth;
5457 frame = frame->prev;
5458 goto fake_study_recurse;
5463 DEBUG_STUDYDATA("pre-fin:",data,depth);
5466 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5468 if (flags & SCF_DO_SUBSTR && is_inf)
5469 data->pos_delta = SSize_t_MAX - data->pos_min;
5470 if (is_par > (I32)U8_MAX)
5472 if (is_par && pars==1 && data) {
5473 data->flags |= SF_IN_PAR;
5474 data->flags &= ~SF_HAS_PAR;
5476 else if (pars && data) {
5477 data->flags |= SF_HAS_PAR;
5478 data->flags &= ~SF_IN_PAR;
5480 if (flags & SCF_DO_STCLASS_OR)
5481 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5482 if (flags & SCF_TRIE_RESTUDY)
5483 data->flags |= SCF_TRIE_RESTUDY;
5485 DEBUG_STUDYDATA("post-fin:",data,depth);
5488 SSize_t final_minlen= min < stopmin ? min : stopmin;
5490 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) && (RExC_maxlen < final_minlen + delta)) {
5491 RExC_maxlen = final_minlen + delta;
5493 return final_minlen;
5499 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5501 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5503 PERL_ARGS_ASSERT_ADD_DATA;
5505 Renewc(RExC_rxi->data,
5506 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5507 char, struct reg_data);
5509 Renew(RExC_rxi->data->what, count + n, U8);
5511 Newx(RExC_rxi->data->what, n, U8);
5512 RExC_rxi->data->count = count + n;
5513 Copy(s, RExC_rxi->data->what + count, n, U8);
5517 /*XXX: todo make this not included in a non debugging perl, but appears to be
5518 * used anyway there, in 'use re' */
5519 #ifndef PERL_IN_XSUB_RE
5521 Perl_reginitcolors(pTHX)
5523 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5525 char *t = savepv(s);
5529 t = strchr(t, '\t');
5535 PL_colors[i] = t = (char *)"";
5540 PL_colors[i++] = (char *)"";
5547 #ifdef TRIE_STUDY_OPT
5548 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5551 (data.flags & SCF_TRIE_RESTUDY) \
5559 #define CHECK_RESTUDY_GOTO_butfirst
5563 * pregcomp - compile a regular expression into internal code
5565 * Decides which engine's compiler to call based on the hint currently in
5569 #ifndef PERL_IN_XSUB_RE
5571 /* return the currently in-scope regex engine (or the default if none) */
5573 regexp_engine const *
5574 Perl_current_re_engine(pTHX)
5576 if (IN_PERL_COMPILETIME) {
5577 HV * const table = GvHV(PL_hintgv);
5580 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5581 return &PL_core_reg_engine;
5582 ptr = hv_fetchs(table, "regcomp", FALSE);
5583 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5584 return &PL_core_reg_engine;
5585 return INT2PTR(regexp_engine*,SvIV(*ptr));
5589 if (!PL_curcop->cop_hints_hash)
5590 return &PL_core_reg_engine;
5591 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5592 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5593 return &PL_core_reg_engine;
5594 return INT2PTR(regexp_engine*,SvIV(ptr));
5600 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5602 regexp_engine const *eng = current_re_engine();
5603 GET_RE_DEBUG_FLAGS_DECL;
5605 PERL_ARGS_ASSERT_PREGCOMP;
5607 /* Dispatch a request to compile a regexp to correct regexp engine. */
5609 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5612 return CALLREGCOMP_ENG(eng, pattern, flags);
5616 /* public(ish) entry point for the perl core's own regex compiling code.
5617 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5618 * pattern rather than a list of OPs, and uses the internal engine rather
5619 * than the current one */
5622 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5624 SV *pat = pattern; /* defeat constness! */
5625 PERL_ARGS_ASSERT_RE_COMPILE;
5626 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5627 #ifdef PERL_IN_XSUB_RE
5630 &PL_core_reg_engine,
5632 NULL, NULL, rx_flags, 0);
5636 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5637 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5638 * point to the realloced string and length.
5640 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5644 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5645 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5647 U8 *const src = (U8*)*pat_p;
5650 STRLEN s = 0, d = 0;
5652 GET_RE_DEBUG_FLAGS_DECL;
5654 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5655 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5657 Newx(dst, *plen_p * 2 + 1, U8);
5659 while (s < *plen_p) {
5660 if (NATIVE_BYTE_IS_INVARIANT(src[s]))
5663 dst[d++] = UTF8_EIGHT_BIT_HI(src[s]);
5664 dst[d] = UTF8_EIGHT_BIT_LO(src[s]);
5666 if (n < num_code_blocks) {
5667 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5668 pRExC_state->code_blocks[n].start = d;
5669 assert(dst[d] == '(');
5672 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5673 pRExC_state->code_blocks[n].end = d;
5674 assert(dst[d] == ')');
5684 *pat_p = (char*) dst;
5686 RExC_orig_utf8 = RExC_utf8 = 1;
5691 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5692 * while recording any code block indices, and handling overloading,
5693 * nested qr// objects etc. If pat is null, it will allocate a new
5694 * string, or just return the first arg, if there's only one.
5696 * Returns the malloced/updated pat.
5697 * patternp and pat_count is the array of SVs to be concatted;
5698 * oplist is the optional list of ops that generated the SVs;
5699 * recompile_p is a pointer to a boolean that will be set if
5700 * the regex will need to be recompiled.
5701 * delim, if non-null is an SV that will be inserted between each element
5705 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5706 SV *pat, SV ** const patternp, int pat_count,
5707 OP *oplist, bool *recompile_p, SV *delim)
5711 bool use_delim = FALSE;
5712 bool alloced = FALSE;
5714 /* if we know we have at least two args, create an empty string,
5715 * then concatenate args to that. For no args, return an empty string */
5716 if (!pat && pat_count != 1) {
5722 for (svp = patternp; svp < patternp + pat_count; svp++) {
5725 STRLEN orig_patlen = 0;
5727 SV *msv = use_delim ? delim : *svp;
5728 if (!msv) msv = &PL_sv_undef;
5730 /* if we've got a delimiter, we go round the loop twice for each
5731 * svp slot (except the last), using the delimiter the second
5740 if (SvTYPE(msv) == SVt_PVAV) {
5741 /* we've encountered an interpolated array within
5742 * the pattern, e.g. /...@a..../. Expand the list of elements,
5743 * then recursively append elements.
5744 * The code in this block is based on S_pushav() */
5746 AV *const av = (AV*)msv;
5747 const SSize_t maxarg = AvFILL(av) + 1;
5751 assert(oplist->op_type == OP_PADAV
5752 || oplist->op_type == OP_RV2AV);
5753 oplist = OP_SIBLING(oplist);
5756 if (SvRMAGICAL(av)) {
5759 Newx(array, maxarg, SV*);
5761 for (i=0; i < maxarg; i++) {
5762 SV ** const svp = av_fetch(av, i, FALSE);
5763 array[i] = svp ? *svp : &PL_sv_undef;
5767 array = AvARRAY(av);
5769 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5770 array, maxarg, NULL, recompile_p,
5772 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5778 /* we make the assumption here that each op in the list of
5779 * op_siblings maps to one SV pushed onto the stack,
5780 * except for code blocks, with have both an OP_NULL and
5782 * This allows us to match up the list of SVs against the
5783 * list of OPs to find the next code block.
5785 * Note that PUSHMARK PADSV PADSV ..
5787 * PADRANGE PADSV PADSV ..
5788 * so the alignment still works. */
5791 if (oplist->op_type == OP_NULL
5792 && (oplist->op_flags & OPf_SPECIAL))
5794 assert(n < pRExC_state->num_code_blocks);
5795 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5796 pRExC_state->code_blocks[n].block = oplist;
5797 pRExC_state->code_blocks[n].src_regex = NULL;
5800 oplist = OP_SIBLING(oplist); /* skip CONST */
5803 oplist = OP_SIBLING(oplist);;
5806 /* apply magic and QR overloading to arg */
5809 if (SvROK(msv) && SvAMAGIC(msv)) {
5810 SV *sv = AMG_CALLunary(msv, regexp_amg);
5814 if (SvTYPE(sv) != SVt_REGEXP)
5815 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5820 /* try concatenation overload ... */
5821 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5822 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5825 /* overloading involved: all bets are off over literal
5826 * code. Pretend we haven't seen it */
5827 pRExC_state->num_code_blocks -= n;
5831 /* ... or failing that, try "" overload */
5832 while (SvAMAGIC(msv)
5833 && (sv = AMG_CALLunary(msv, string_amg))
5837 && SvRV(msv) == SvRV(sv))
5842 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5846 /* this is a partially unrolled
5847 * sv_catsv_nomg(pat, msv);
5848 * that allows us to adjust code block indices if
5851 char *dst = SvPV_force_nomg(pat, dlen);
5853 if (SvUTF8(msv) && !SvUTF8(pat)) {
5854 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5855 sv_setpvn(pat, dst, dlen);
5858 sv_catsv_nomg(pat, msv);
5865 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5868 /* extract any code blocks within any embedded qr//'s */
5869 if (rx && SvTYPE(rx) == SVt_REGEXP
5870 && RX_ENGINE((REGEXP*)rx)->op_comp)
5873 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5874 if (ri->num_code_blocks) {
5876 /* the presence of an embedded qr// with code means
5877 * we should always recompile: the text of the
5878 * qr// may not have changed, but it may be a
5879 * different closure than last time */
5881 Renew(pRExC_state->code_blocks,
5882 pRExC_state->num_code_blocks + ri->num_code_blocks,
5883 struct reg_code_block);
5884 pRExC_state->num_code_blocks += ri->num_code_blocks;
5886 for (i=0; i < ri->num_code_blocks; i++) {
5887 struct reg_code_block *src, *dst;
5888 STRLEN offset = orig_patlen
5889 + ReANY((REGEXP *)rx)->pre_prefix;
5890 assert(n < pRExC_state->num_code_blocks);
5891 src = &ri->code_blocks[i];
5892 dst = &pRExC_state->code_blocks[n];
5893 dst->start = src->start + offset;
5894 dst->end = src->end + offset;
5895 dst->block = src->block;
5896 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5905 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5914 /* see if there are any run-time code blocks in the pattern.
5915 * False positives are allowed */
5918 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5919 char *pat, STRLEN plen)
5924 PERL_UNUSED_CONTEXT;
5926 for (s = 0; s < plen; s++) {
5927 if (n < pRExC_state->num_code_blocks
5928 && s == pRExC_state->code_blocks[n].start)
5930 s = pRExC_state->code_blocks[n].end;
5934 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5936 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5938 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5945 /* Handle run-time code blocks. We will already have compiled any direct
5946 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5947 * copy of it, but with any literal code blocks blanked out and
5948 * appropriate chars escaped; then feed it into
5950 * eval "qr'modified_pattern'"
5954 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5958 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5960 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5961 * and merge them with any code blocks of the original regexp.
5963 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5964 * instead, just save the qr and return FALSE; this tells our caller that
5965 * the original pattern needs upgrading to utf8.
5969 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5970 char *pat, STRLEN plen)
5974 GET_RE_DEBUG_FLAGS_DECL;
5976 if (pRExC_state->runtime_code_qr) {
5977 /* this is the second time we've been called; this should
5978 * only happen if the main pattern got upgraded to utf8
5979 * during compilation; re-use the qr we compiled first time
5980 * round (which should be utf8 too)
5982 qr = pRExC_state->runtime_code_qr;
5983 pRExC_state->runtime_code_qr = NULL;
5984 assert(RExC_utf8 && SvUTF8(qr));
5990 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5994 /* determine how many extra chars we need for ' and \ escaping */
5995 for (s = 0; s < plen; s++) {
5996 if (pat[s] == '\'' || pat[s] == '\\')
6000 Newx(newpat, newlen, char);
6002 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6004 for (s = 0; s < plen; s++) {
6005 if (n < pRExC_state->num_code_blocks
6006 && s == pRExC_state->code_blocks[n].start)
6008 /* blank out literal code block */
6009 assert(pat[s] == '(');
6010 while (s <= pRExC_state->code_blocks[n].end) {
6018 if (pat[s] == '\'' || pat[s] == '\\')
6023 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6027 PerlIO_printf(Perl_debug_log,
6028 "%sre-parsing pattern for runtime code:%s %s\n",
6029 PL_colors[4],PL_colors[5],newpat);
6032 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6038 PUSHSTACKi(PERLSI_REQUIRE);
6039 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6040 * parsing qr''; normally only q'' does this. It also alters
6042 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6043 SvREFCNT_dec_NN(sv);
6048 SV * const errsv = ERRSV;
6049 if (SvTRUE_NN(errsv))
6051 Safefree(pRExC_state->code_blocks);
6052 /* use croak_sv ? */
6053 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6056 assert(SvROK(qr_ref));
6058 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6059 /* the leaving below frees the tmp qr_ref.
6060 * Give qr a life of its own */
6068 if (!RExC_utf8 && SvUTF8(qr)) {
6069 /* first time through; the pattern got upgraded; save the
6070 * qr for the next time through */
6071 assert(!pRExC_state->runtime_code_qr);
6072 pRExC_state->runtime_code_qr = qr;
6077 /* extract any code blocks within the returned qr// */
6080 /* merge the main (r1) and run-time (r2) code blocks into one */
6082 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6083 struct reg_code_block *new_block, *dst;
6084 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6087 if (!r2->num_code_blocks) /* we guessed wrong */
6089 SvREFCNT_dec_NN(qr);
6094 r1->num_code_blocks + r2->num_code_blocks,
6095 struct reg_code_block);
6098 while ( i1 < r1->num_code_blocks
6099 || i2 < r2->num_code_blocks)
6101 struct reg_code_block *src;
6104 if (i1 == r1->num_code_blocks) {
6105 src = &r2->code_blocks[i2++];
6108 else if (i2 == r2->num_code_blocks)
6109 src = &r1->code_blocks[i1++];
6110 else if ( r1->code_blocks[i1].start
6111 < r2->code_blocks[i2].start)
6113 src = &r1->code_blocks[i1++];
6114 assert(src->end < r2->code_blocks[i2].start);
6117 assert( r1->code_blocks[i1].start
6118 > r2->code_blocks[i2].start);
6119 src = &r2->code_blocks[i2++];
6121 assert(src->end < r1->code_blocks[i1].start);
6124 assert(pat[src->start] == '(');
6125 assert(pat[src->end] == ')');
6126 dst->start = src->start;
6127 dst->end = src->end;
6128 dst->block = src->block;
6129 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6133 r1->num_code_blocks += r2->num_code_blocks;
6134 Safefree(r1->code_blocks);
6135 r1->code_blocks = new_block;
6138 SvREFCNT_dec_NN(qr);
6144 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6145 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6146 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6147 STRLEN longest_length, bool eol, bool meol)
6149 /* This is the common code for setting up the floating and fixed length
6150 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6151 * as to whether succeeded or not */
6156 if (! (longest_length
6157 || (eol /* Can't have SEOL and MULTI */
6158 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6160 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6161 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6166 /* copy the information about the longest from the reg_scan_data
6167 over to the program. */
6168 if (SvUTF8(sv_longest)) {
6169 *rx_utf8 = sv_longest;
6172 *rx_substr = sv_longest;
6175 /* end_shift is how many chars that must be matched that
6176 follow this item. We calculate it ahead of time as once the
6177 lookbehind offset is added in we lose the ability to correctly
6179 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6180 *rx_end_shift = ml - offset
6181 - longest_length + (SvTAIL(sv_longest) != 0)
6184 t = (eol/* Can't have SEOL and MULTI */
6185 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6186 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6192 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6193 * regular expression into internal code.
6194 * The pattern may be passed either as:
6195 * a list of SVs (patternp plus pat_count)
6196 * a list of OPs (expr)
6197 * If both are passed, the SV list is used, but the OP list indicates
6198 * which SVs are actually pre-compiled code blocks
6200 * The SVs in the list have magic and qr overloading applied to them (and
6201 * the list may be modified in-place with replacement SVs in the latter
6204 * If the pattern hasn't changed from old_re, then old_re will be
6207 * eng is the current engine. If that engine has an op_comp method, then
6208 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6209 * do the initial concatenation of arguments and pass on to the external
6212 * If is_bare_re is not null, set it to a boolean indicating whether the
6213 * arg list reduced (after overloading) to a single bare regex which has
6214 * been returned (i.e. /$qr/).
6216 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6218 * pm_flags contains the PMf_* flags, typically based on those from the
6219 * pm_flags field of the related PMOP. Currently we're only interested in
6220 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6222 * We can't allocate space until we know how big the compiled form will be,
6223 * but we can't compile it (and thus know how big it is) until we've got a
6224 * place to put the code. So we cheat: we compile it twice, once with code
6225 * generation turned off and size counting turned on, and once "for real".
6226 * This also means that we don't allocate space until we are sure that the
6227 * thing really will compile successfully, and we never have to move the
6228 * code and thus invalidate pointers into it. (Note that it has to be in
6229 * one piece because free() must be able to free it all.) [NB: not true in perl]
6231 * Beware that the optimization-preparation code in here knows about some
6232 * of the structure of the compiled regexp. [I'll say.]
6236 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6237 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6238 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6242 regexp_internal *ri;
6250 SV *code_blocksv = NULL;
6251 SV** new_patternp = patternp;
6253 /* these are all flags - maybe they should be turned
6254 * into a single int with different bit masks */
6255 I32 sawlookahead = 0;
6260 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6262 bool runtime_code = 0;
6264 RExC_state_t RExC_state;
6265 RExC_state_t * const pRExC_state = &RExC_state;
6266 #ifdef TRIE_STUDY_OPT
6268 RExC_state_t copyRExC_state;
6270 GET_RE_DEBUG_FLAGS_DECL;
6272 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6274 DEBUG_r(if (!PL_colorset) reginitcolors());
6276 #ifndef PERL_IN_XSUB_RE
6277 /* Initialize these here instead of as-needed, as is quick and avoids
6278 * having to test them each time otherwise */
6279 if (! PL_AboveLatin1) {
6280 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6281 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6282 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6283 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6284 PL_HasMultiCharFold =
6285 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6289 pRExC_state->code_blocks = NULL;
6290 pRExC_state->num_code_blocks = 0;
6293 *is_bare_re = FALSE;
6295 if (expr && (expr->op_type == OP_LIST ||
6296 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6297 /* allocate code_blocks if needed */
6301 for (o = cLISTOPx(expr)->op_first; o; o = OP_SIBLING(o))
6302 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6303 ncode++; /* count of DO blocks */
6305 pRExC_state->num_code_blocks = ncode;
6306 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6311 /* compile-time pattern with just OP_CONSTs and DO blocks */
6316 /* find how many CONSTs there are */
6319 if (expr->op_type == OP_CONST)
6322 for (o = cLISTOPx(expr)->op_first; o; o = OP_SIBLING(o)) {
6323 if (o->op_type == OP_CONST)
6327 /* fake up an SV array */
6329 assert(!new_patternp);
6330 Newx(new_patternp, n, SV*);
6331 SAVEFREEPV(new_patternp);
6335 if (expr->op_type == OP_CONST)
6336 new_patternp[n] = cSVOPx_sv(expr);
6338 for (o = cLISTOPx(expr)->op_first; o; o = OP_SIBLING(o)) {
6339 if (o->op_type == OP_CONST)
6340 new_patternp[n++] = cSVOPo_sv;
6345 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6346 "Assembling pattern from %d elements%s\n", pat_count,
6347 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6349 /* set expr to the first arg op */
6351 if (pRExC_state->num_code_blocks
6352 && expr->op_type != OP_CONST)
6354 expr = cLISTOPx(expr)->op_first;
6355 assert( expr->op_type == OP_PUSHMARK
6356 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6357 || expr->op_type == OP_PADRANGE);
6358 expr = OP_SIBLING(expr);
6361 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6362 expr, &recompile, NULL);
6364 /* handle bare (possibly after overloading) regex: foo =~ $re */
6369 if (SvTYPE(re) == SVt_REGEXP) {
6373 Safefree(pRExC_state->code_blocks);
6374 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6375 "Precompiled pattern%s\n",
6376 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6382 exp = SvPV_nomg(pat, plen);
6384 if (!eng->op_comp) {
6385 if ((SvUTF8(pat) && IN_BYTES)
6386 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6388 /* make a temporary copy; either to convert to bytes,
6389 * or to avoid repeating get-magic / overloaded stringify */
6390 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6391 (IN_BYTES ? 0 : SvUTF8(pat)));
6393 Safefree(pRExC_state->code_blocks);
6394 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6397 /* ignore the utf8ness if the pattern is 0 length */
6398 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6399 RExC_uni_semantics = 0;
6400 RExC_contains_locale = 0;
6401 RExC_contains_i = 0;
6402 pRExC_state->runtime_code_qr = NULL;
6405 SV *dsv= sv_newmortal();
6406 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6407 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6408 PL_colors[4],PL_colors[5],s);
6412 /* we jump here if we upgrade the pattern to utf8 and have to
6415 if ((pm_flags & PMf_USE_RE_EVAL)
6416 /* this second condition covers the non-regex literal case,
6417 * i.e. $foo =~ '(?{})'. */
6418 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6420 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6422 /* return old regex if pattern hasn't changed */
6423 /* XXX: note in the below we have to check the flags as well as the
6426 * Things get a touch tricky as we have to compare the utf8 flag
6427 * independently from the compile flags. */
6431 && !!RX_UTF8(old_re) == !!RExC_utf8
6432 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6433 && RX_PRECOMP(old_re)
6434 && RX_PRELEN(old_re) == plen
6435 && memEQ(RX_PRECOMP(old_re), exp, plen)
6436 && !runtime_code /* with runtime code, always recompile */ )
6438 Safefree(pRExC_state->code_blocks);
6442 rx_flags = orig_rx_flags;
6444 if (rx_flags & PMf_FOLD) {
6445 RExC_contains_i = 1;
6447 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6449 /* Set to use unicode semantics if the pattern is in utf8 and has the
6450 * 'depends' charset specified, as it means unicode when utf8 */
6451 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6455 RExC_flags = rx_flags;
6456 RExC_pm_flags = pm_flags;
6459 if (TAINTING_get && TAINT_get)
6460 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6462 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6463 /* whoops, we have a non-utf8 pattern, whilst run-time code
6464 * got compiled as utf8. Try again with a utf8 pattern */
6465 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6466 pRExC_state->num_code_blocks);
6467 goto redo_first_pass;
6470 assert(!pRExC_state->runtime_code_qr);
6476 RExC_in_lookbehind = 0;
6477 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6479 RExC_override_recoding = 0;
6480 RExC_in_multi_char_class = 0;
6482 /* First pass: determine size, legality. */
6485 RExC_end = exp + plen;
6490 RExC_emit = (regnode *) &RExC_emit_dummy;
6491 RExC_whilem_seen = 0;
6492 RExC_open_parens = NULL;
6493 RExC_close_parens = NULL;
6495 RExC_paren_names = NULL;
6497 RExC_paren_name_list = NULL;
6499 RExC_recurse = NULL;
6500 RExC_study_chunk_recursed = NULL;
6501 RExC_study_chunk_recursed_bytes= 0;
6502 RExC_recurse_count = 0;
6503 pRExC_state->code_index = 0;
6505 #if 0 /* REGC() is (currently) a NOP at the first pass.
6506 * Clever compilers notice this and complain. --jhi */
6507 REGC((U8)REG_MAGIC, (char*)RExC_emit);
6510 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6512 RExC_lastparse=NULL;
6514 /* reg may croak on us, not giving us a chance to free
6515 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6516 need it to survive as long as the regexp (qr/(?{})/).
6517 We must check that code_blocksv is not already set, because we may
6518 have jumped back to restart the sizing pass. */
6519 if (pRExC_state->code_blocks && !code_blocksv) {
6520 code_blocksv = newSV_type(SVt_PV);
6521 SAVEFREESV(code_blocksv);
6522 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6523 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6525 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6526 /* It's possible to write a regexp in ascii that represents Unicode
6527 codepoints outside of the byte range, such as via \x{100}. If we
6528 detect such a sequence we have to convert the entire pattern to utf8
6529 and then recompile, as our sizing calculation will have been based
6530 on 1 byte == 1 character, but we will need to use utf8 to encode
6531 at least some part of the pattern, and therefore must convert the whole
6534 if (flags & RESTART_UTF8) {
6535 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6536 pRExC_state->num_code_blocks);
6537 goto redo_first_pass;
6539 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6542 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6545 PerlIO_printf(Perl_debug_log,
6546 "Required size %"IVdf" nodes\n"
6547 "Starting second pass (creation)\n",
6550 RExC_lastparse=NULL;
6553 /* The first pass could have found things that force Unicode semantics */
6554 if ((RExC_utf8 || RExC_uni_semantics)
6555 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6557 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6560 /* Small enough for pointer-storage convention?
6561 If extralen==0, this means that we will not need long jumps. */
6562 if (RExC_size >= 0x10000L && RExC_extralen)
6563 RExC_size += RExC_extralen;
6566 if (RExC_whilem_seen > 15)
6567 RExC_whilem_seen = 15;
6569 /* Allocate space and zero-initialize. Note, the two step process
6570 of zeroing when in debug mode, thus anything assigned has to
6571 happen after that */
6572 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6574 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6575 char, regexp_internal);
6576 if ( r == NULL || ri == NULL )
6577 FAIL("Regexp out of space");
6579 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6580 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6583 /* bulk initialize base fields with 0. */
6584 Zero(ri, sizeof(regexp_internal), char);
6587 /* non-zero initialization begins here */
6590 r->extflags = rx_flags;
6591 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6593 if (pm_flags & PMf_IS_QR) {
6594 ri->code_blocks = pRExC_state->code_blocks;
6595 ri->num_code_blocks = pRExC_state->num_code_blocks;
6600 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6601 if (pRExC_state->code_blocks[n].src_regex)
6602 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6603 SAVEFREEPV(pRExC_state->code_blocks);
6607 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6608 bool has_charset = (get_regex_charset(r->extflags)
6609 != REGEX_DEPENDS_CHARSET);
6611 /* The caret is output if there are any defaults: if not all the STD
6612 * flags are set, or if no character set specifier is needed */
6614 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6616 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6617 == REG_RUN_ON_COMMENT_SEEN);
6618 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6619 >> RXf_PMf_STD_PMMOD_SHIFT);
6620 const char *fptr = STD_PAT_MODS; /*"msix"*/
6622 /* Allocate for the worst case, which is all the std flags are turned
6623 * on. If more precision is desired, we could do a population count of
6624 * the flags set. This could be done with a small lookup table, or by
6625 * shifting, masking and adding, or even, when available, assembly
6626 * language for a machine-language population count.
6627 * We never output a minus, as all those are defaults, so are
6628 * covered by the caret */
6629 const STRLEN wraplen = plen + has_p + has_runon
6630 + has_default /* If needs a caret */
6632 /* If needs a character set specifier */
6633 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6634 + (sizeof(STD_PAT_MODS) - 1)
6635 + (sizeof("(?:)") - 1);
6637 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6638 r->xpv_len_u.xpvlenu_pv = p;
6640 SvFLAGS(rx) |= SVf_UTF8;
6643 /* If a default, cover it using the caret */
6645 *p++= DEFAULT_PAT_MOD;
6649 const char* const name = get_regex_charset_name(r->extflags, &len);
6650 Copy(name, p, len, char);
6654 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6657 while((ch = *fptr++)) {
6665 Copy(RExC_precomp, p, plen, char);
6666 assert ((RX_WRAPPED(rx) - p) < 16);
6667 r->pre_prefix = p - RX_WRAPPED(rx);
6673 SvCUR_set(rx, p - RX_WRAPPED(rx));
6677 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6679 /* setup various meta data about recursion, this all requires
6680 * RExC_npar to be correctly set, and a bit later on we clear it */
6681 if (RExC_seen & REG_RECURSE_SEEN) {
6682 Newxz(RExC_open_parens, RExC_npar,regnode *);
6683 SAVEFREEPV(RExC_open_parens);
6684 Newxz(RExC_close_parens,RExC_npar,regnode *);
6685 SAVEFREEPV(RExC_close_parens);
6687 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6688 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6689 * So its 1 if there are no parens. */
6690 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6691 ((RExC_npar & 0x07) != 0);
6692 Newx(RExC_study_chunk_recursed,
6693 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6694 SAVEFREEPV(RExC_study_chunk_recursed);
6697 /* Useful during FAIL. */
6698 #ifdef RE_TRACK_PATTERN_OFFSETS
6699 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6700 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6701 "%s %"UVuf" bytes for offset annotations.\n",
6702 ri->u.offsets ? "Got" : "Couldn't get",
6703 (UV)((2*RExC_size+1) * sizeof(U32))));
6705 SetProgLen(ri,RExC_size);
6710 /* Second pass: emit code. */
6711 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6712 RExC_pm_flags = pm_flags;
6714 RExC_end = exp + plen;
6717 RExC_emit_start = ri->program;
6718 RExC_emit = ri->program;
6719 RExC_emit_bound = ri->program + RExC_size + 1;
6720 pRExC_state->code_index = 0;
6722 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6723 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6725 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6727 /* XXXX To minimize changes to RE engine we always allocate
6728 3-units-long substrs field. */
6729 Newx(r->substrs, 1, struct reg_substr_data);
6730 if (RExC_recurse_count) {
6731 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6732 SAVEFREEPV(RExC_recurse);
6736 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6737 Zero(r->substrs, 1, struct reg_substr_data);
6738 if (RExC_study_chunk_recursed)
6739 Zero(RExC_study_chunk_recursed,
6740 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6742 #ifdef TRIE_STUDY_OPT
6744 StructCopy(&zero_scan_data, &data, scan_data_t);
6745 copyRExC_state = RExC_state;
6748 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6750 RExC_state = copyRExC_state;
6751 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6752 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6754 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6755 StructCopy(&zero_scan_data, &data, scan_data_t);
6758 StructCopy(&zero_scan_data, &data, scan_data_t);
6761 /* Dig out information for optimizations. */
6762 r->extflags = RExC_flags; /* was pm_op */
6763 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6766 SvUTF8_on(rx); /* Unicode in it? */
6767 ri->regstclass = NULL;
6768 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6769 r->intflags |= PREGf_NAUGHTY;
6770 scan = ri->program + 1; /* First BRANCH. */
6772 /* testing for BRANCH here tells us whether there is "must appear"
6773 data in the pattern. If there is then we can use it for optimisations */
6774 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6777 STRLEN longest_float_length, longest_fixed_length;
6778 regnode_ssc ch_class; /* pointed to by data */
6780 SSize_t last_close = 0; /* pointed to by data */
6781 regnode *first= scan;
6782 regnode *first_next= regnext(first);
6784 * Skip introductions and multiplicators >= 1
6785 * so that we can extract the 'meat' of the pattern that must
6786 * match in the large if() sequence following.
6787 * NOTE that EXACT is NOT covered here, as it is normally
6788 * picked up by the optimiser separately.
6790 * This is unfortunate as the optimiser isnt handling lookahead
6791 * properly currently.
6794 while ((OP(first) == OPEN && (sawopen = 1)) ||
6795 /* An OR of *one* alternative - should not happen now. */
6796 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6797 /* for now we can't handle lookbehind IFMATCH*/
6798 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6799 (OP(first) == PLUS) ||
6800 (OP(first) == MINMOD) ||
6801 /* An {n,m} with n>0 */
6802 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6803 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6806 * the only op that could be a regnode is PLUS, all the rest
6807 * will be regnode_1 or regnode_2.
6809 * (yves doesn't think this is true)
6811 if (OP(first) == PLUS)
6814 if (OP(first) == MINMOD)
6816 first += regarglen[OP(first)];
6818 first = NEXTOPER(first);
6819 first_next= regnext(first);
6822 /* Starting-point info. */
6824 DEBUG_PEEP("first:",first,0);
6825 /* Ignore EXACT as we deal with it later. */
6826 if (PL_regkind[OP(first)] == EXACT) {
6827 if (OP(first) == EXACT)
6828 NOOP; /* Empty, get anchored substr later. */
6830 ri->regstclass = first;
6833 else if (PL_regkind[OP(first)] == TRIE &&
6834 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6836 /* this can happen only on restudy */
6837 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
6840 else if (REGNODE_SIMPLE(OP(first)))
6841 ri->regstclass = first;
6842 else if (PL_regkind[OP(first)] == BOUND ||
6843 PL_regkind[OP(first)] == NBOUND)
6844 ri->regstclass = first;
6845 else if (PL_regkind[OP(first)] == BOL) {
6846 r->intflags |= (OP(first) == MBOL
6848 : (OP(first) == SBOL
6851 first = NEXTOPER(first);
6854 else if (OP(first) == GPOS) {
6855 r->intflags |= PREGf_ANCH_GPOS;
6856 first = NEXTOPER(first);
6859 else if ((!sawopen || !RExC_sawback) &&
6861 (OP(first) == STAR &&
6862 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6863 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
6865 /* turn .* into ^.* with an implied $*=1 */
6867 (OP(NEXTOPER(first)) == REG_ANY)
6870 r->intflags |= (type | PREGf_IMPLICIT);
6871 first = NEXTOPER(first);
6874 if (sawplus && !sawminmod && !sawlookahead
6875 && (!sawopen || !RExC_sawback)
6876 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6877 /* x+ must match at the 1st pos of run of x's */
6878 r->intflags |= PREGf_SKIP;
6880 /* Scan is after the zeroth branch, first is atomic matcher. */
6881 #ifdef TRIE_STUDY_OPT
6884 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6885 (IV)(first - scan + 1))
6889 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6890 (IV)(first - scan + 1))
6896 * If there's something expensive in the r.e., find the
6897 * longest literal string that must appear and make it the
6898 * regmust. Resolve ties in favor of later strings, since
6899 * the regstart check works with the beginning of the r.e.
6900 * and avoiding duplication strengthens checking. Not a
6901 * strong reason, but sufficient in the absence of others.
6902 * [Now we resolve ties in favor of the earlier string if
6903 * it happens that c_offset_min has been invalidated, since the
6904 * earlier string may buy us something the later one won't.]
6907 data.longest_fixed = newSVpvs("");
6908 data.longest_float = newSVpvs("");
6909 data.last_found = newSVpvs("");
6910 data.longest = &(data.longest_fixed);
6911 ENTER_with_name("study_chunk");
6912 SAVEFREESV(data.longest_fixed);
6913 SAVEFREESV(data.longest_float);
6914 SAVEFREESV(data.last_found);
6916 if (!ri->regstclass) {
6917 ssc_init(pRExC_state, &ch_class);
6918 data.start_class = &ch_class;
6919 stclass_flag = SCF_DO_STCLASS_AND;
6920 } else /* XXXX Check for BOUND? */
6922 data.last_closep = &last_close;
6925 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
6926 scan + RExC_size, /* Up to end */
6928 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6929 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6933 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6936 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6937 && data.last_start_min == 0 && data.last_end > 0
6938 && !RExC_seen_zerolen
6939 && !(RExC_seen & REG_VERBARG_SEEN)
6940 && !(RExC_seen & REG_GPOS_SEEN)
6942 r->extflags |= RXf_CHECK_ALL;
6944 scan_commit(pRExC_state, &data,&minlen,0);
6946 longest_float_length = CHR_SVLEN(data.longest_float);
6948 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6949 && data.offset_fixed == data.offset_float_min
6950 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6951 && S_setup_longest (aTHX_ pRExC_state,
6955 &(r->float_end_shift),
6956 data.lookbehind_float,
6957 data.offset_float_min,
6959 longest_float_length,
6960 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6961 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6963 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6964 r->float_max_offset = data.offset_float_max;
6965 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
6966 r->float_max_offset -= data.lookbehind_float;
6967 SvREFCNT_inc_simple_void_NN(data.longest_float);
6970 r->float_substr = r->float_utf8 = NULL;
6971 longest_float_length = 0;
6974 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6976 if (S_setup_longest (aTHX_ pRExC_state,
6978 &(r->anchored_utf8),
6979 &(r->anchored_substr),
6980 &(r->anchored_end_shift),
6981 data.lookbehind_fixed,
6984 longest_fixed_length,
6985 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6986 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6988 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6989 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6992 r->anchored_substr = r->anchored_utf8 = NULL;
6993 longest_fixed_length = 0;
6995 LEAVE_with_name("study_chunk");
6998 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6999 ri->regstclass = NULL;
7001 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7003 && ! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
7004 && !ssc_is_anything(data.start_class))
7006 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7008 ssc_finalize(pRExC_state, data.start_class);
7010 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7011 StructCopy(data.start_class,
7012 (regnode_ssc*)RExC_rxi->data->data[n],
7014 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7015 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7016 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7017 regprop(r, sv, (regnode*)data.start_class, NULL);
7018 PerlIO_printf(Perl_debug_log,
7019 "synthetic stclass \"%s\".\n",
7020 SvPVX_const(sv));});
7021 data.start_class = NULL;
7024 /* A temporary algorithm prefers floated substr to fixed one to dig
7026 if (longest_fixed_length > longest_float_length) {
7027 r->substrs->check_ix = 0;
7028 r->check_end_shift = r->anchored_end_shift;
7029 r->check_substr = r->anchored_substr;
7030 r->check_utf8 = r->anchored_utf8;
7031 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7032 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7033 r->intflags |= PREGf_NOSCAN;
7036 r->substrs->check_ix = 1;
7037 r->check_end_shift = r->float_end_shift;
7038 r->check_substr = r->float_substr;
7039 r->check_utf8 = r->float_utf8;
7040 r->check_offset_min = r->float_min_offset;
7041 r->check_offset_max = r->float_max_offset;
7043 if ((r->check_substr || r->check_utf8) ) {
7044 r->extflags |= RXf_USE_INTUIT;
7045 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7046 r->extflags |= RXf_INTUIT_TAIL;
7048 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7050 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7051 if ( (STRLEN)minlen < longest_float_length )
7052 minlen= longest_float_length;
7053 if ( (STRLEN)minlen < longest_fixed_length )
7054 minlen= longest_fixed_length;
7058 /* Several toplevels. Best we can is to set minlen. */
7060 regnode_ssc ch_class;
7061 SSize_t last_close = 0;
7063 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7065 scan = ri->program + 1;
7066 ssc_init(pRExC_state, &ch_class);
7067 data.start_class = &ch_class;
7068 data.last_closep = &last_close;
7071 minlen = study_chunk(pRExC_state,
7072 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7073 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7074 ? SCF_TRIE_DOING_RESTUDY
7078 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7080 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7081 = r->float_substr = r->float_utf8 = NULL;
7083 if (! (ANYOF_FLAGS(data.start_class) & ANYOF_EMPTY_STRING)
7084 && ! ssc_is_anything(data.start_class))
7086 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7088 ssc_finalize(pRExC_state, data.start_class);
7090 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7091 StructCopy(data.start_class,
7092 (regnode_ssc*)RExC_rxi->data->data[n],
7094 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7095 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7096 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7097 regprop(r, sv, (regnode*)data.start_class, NULL);
7098 PerlIO_printf(Perl_debug_log,
7099 "synthetic stclass \"%s\".\n",
7100 SvPVX_const(sv));});
7101 data.start_class = NULL;
7105 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7106 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7107 r->maxlen = REG_INFTY;
7110 r->maxlen = RExC_maxlen;
7113 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7114 the "real" pattern. */
7116 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%ld\n",
7117 (IV)minlen, (IV)r->minlen, RExC_maxlen);
7119 r->minlenret = minlen;
7120 if (r->minlen < minlen)
7123 if (RExC_seen & REG_GPOS_SEEN)
7124 r->intflags |= PREGf_GPOS_SEEN;
7125 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7126 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7128 if (pRExC_state->num_code_blocks)
7129 r->extflags |= RXf_EVAL_SEEN;
7130 if (RExC_seen & REG_CANY_SEEN)
7131 r->intflags |= PREGf_CANY_SEEN;
7132 if (RExC_seen & REG_VERBARG_SEEN)
7134 r->intflags |= PREGf_VERBARG_SEEN;
7135 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7137 if (RExC_seen & REG_CUTGROUP_SEEN)
7138 r->intflags |= PREGf_CUTGROUP_SEEN;
7139 if (pm_flags & PMf_USE_RE_EVAL)
7140 r->intflags |= PREGf_USE_RE_EVAL;
7141 if (RExC_paren_names)
7142 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7144 RXp_PAREN_NAMES(r) = NULL;
7146 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7147 * so it can be used in pp.c */
7148 if (r->intflags & PREGf_ANCH)
7149 r->extflags |= RXf_IS_ANCHORED;
7153 /* this is used to identify "special" patterns that might result
7154 * in Perl NOT calling the regex engine and instead doing the match "itself",
7155 * particularly special cases in split//. By having the regex compiler
7156 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7157 * we avoid weird issues with equivalent patterns resulting in different behavior,
7158 * AND we allow non Perl engines to get the same optimizations by the setting the
7159 * flags appropriately - Yves */
7160 regnode *first = ri->program + 1;
7162 regnode *next = NEXTOPER(first);
7165 if (PL_regkind[fop] == NOTHING && nop == END)
7166 r->extflags |= RXf_NULL;
7167 else if (PL_regkind[fop] == BOL && nop == END)
7168 r->extflags |= RXf_START_ONLY;
7169 else if (fop == PLUS
7170 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7171 && OP(regnext(first)) == END)
7172 r->extflags |= RXf_WHITE;
7173 else if ( r->extflags & RXf_SPLIT
7175 && STR_LEN(first) == 1
7176 && *(STRING(first)) == ' '
7177 && OP(regnext(first)) == END )
7178 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7182 if (RExC_contains_locale) {
7183 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7187 if (RExC_paren_names) {
7188 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7189 ri->data->data[ri->name_list_idx]
7190 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7193 ri->name_list_idx = 0;
7195 if (RExC_recurse_count) {
7196 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7197 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7198 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7201 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7202 /* assume we don't need to swap parens around before we match */
7206 PerlIO_printf(Perl_debug_log,"Final program:\n");
7209 #ifdef RE_TRACK_PATTERN_OFFSETS
7210 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7211 const STRLEN len = ri->u.offsets[0];
7213 GET_RE_DEBUG_FLAGS_DECL;
7214 PerlIO_printf(Perl_debug_log,
7215 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7216 for (i = 1; i <= len; i++) {
7217 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7218 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7219 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7221 PerlIO_printf(Perl_debug_log, "\n");
7226 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7227 * by setting the regexp SV to readonly-only instead. If the
7228 * pattern's been recompiled, the USEDness should remain. */
7229 if (old_re && SvREADONLY(old_re))
7237 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7240 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7242 PERL_UNUSED_ARG(value);
7244 if (flags & RXapif_FETCH) {
7245 return reg_named_buff_fetch(rx, key, flags);
7246 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7247 Perl_croak_no_modify();
7249 } else if (flags & RXapif_EXISTS) {
7250 return reg_named_buff_exists(rx, key, flags)
7253 } else if (flags & RXapif_REGNAMES) {
7254 return reg_named_buff_all(rx, flags);
7255 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7256 return reg_named_buff_scalar(rx, flags);
7258 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7264 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7267 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7268 PERL_UNUSED_ARG(lastkey);
7270 if (flags & RXapif_FIRSTKEY)
7271 return reg_named_buff_firstkey(rx, flags);
7272 else if (flags & RXapif_NEXTKEY)
7273 return reg_named_buff_nextkey(rx, flags);
7275 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7282 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7285 AV *retarray = NULL;
7287 struct regexp *const rx = ReANY(r);
7289 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7291 if (flags & RXapif_ALL)
7294 if (rx && RXp_PAREN_NAMES(rx)) {
7295 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7298 SV* sv_dat=HeVAL(he_str);
7299 I32 *nums=(I32*)SvPVX(sv_dat);
7300 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7301 if ((I32)(rx->nparens) >= nums[i]
7302 && rx->offs[nums[i]].start != -1
7303 && rx->offs[nums[i]].end != -1)
7306 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7311 ret = newSVsv(&PL_sv_undef);
7314 av_push(retarray, ret);
7317 return newRV_noinc(MUTABLE_SV(retarray));
7324 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7327 struct regexp *const rx = ReANY(r);
7329 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7331 if (rx && RXp_PAREN_NAMES(rx)) {
7332 if (flags & RXapif_ALL) {
7333 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7335 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7337 SvREFCNT_dec_NN(sv);
7349 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7351 struct regexp *const rx = ReANY(r);
7353 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7355 if ( rx && RXp_PAREN_NAMES(rx) ) {
7356 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7358 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7365 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7367 struct regexp *const rx = ReANY(r);
7368 GET_RE_DEBUG_FLAGS_DECL;
7370 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7372 if (rx && RXp_PAREN_NAMES(rx)) {
7373 HV *hv = RXp_PAREN_NAMES(rx);
7375 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7378 SV* sv_dat = HeVAL(temphe);
7379 I32 *nums = (I32*)SvPVX(sv_dat);
7380 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7381 if ((I32)(rx->lastparen) >= nums[i] &&
7382 rx->offs[nums[i]].start != -1 &&
7383 rx->offs[nums[i]].end != -1)
7389 if (parno || flags & RXapif_ALL) {
7390 return newSVhek(HeKEY_hek(temphe));
7398 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7403 struct regexp *const rx = ReANY(r);
7405 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7407 if (rx && RXp_PAREN_NAMES(rx)) {
7408 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7409 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7410 } else if (flags & RXapif_ONE) {
7411 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7412 av = MUTABLE_AV(SvRV(ret));
7413 length = av_tindex(av);
7414 SvREFCNT_dec_NN(ret);
7415 return newSViv(length + 1);
7417 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7422 return &PL_sv_undef;
7426 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7428 struct regexp *const rx = ReANY(r);
7431 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7433 if (rx && RXp_PAREN_NAMES(rx)) {
7434 HV *hv= RXp_PAREN_NAMES(rx);
7436 (void)hv_iterinit(hv);
7437 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7440 SV* sv_dat = HeVAL(temphe);
7441 I32 *nums = (I32*)SvPVX(sv_dat);
7442 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7443 if ((I32)(rx->lastparen) >= nums[i] &&
7444 rx->offs[nums[i]].start != -1 &&
7445 rx->offs[nums[i]].end != -1)
7451 if (parno || flags & RXapif_ALL) {
7452 av_push(av, newSVhek(HeKEY_hek(temphe)));
7457 return newRV_noinc(MUTABLE_SV(av));
7461 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7464 struct regexp *const rx = ReANY(r);
7470 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7472 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7473 || n == RX_BUFF_IDX_CARET_FULLMATCH
7474 || n == RX_BUFF_IDX_CARET_POSTMATCH
7477 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7479 /* on something like
7482 * the KEEPCOPY is set on the PMOP rather than the regex */
7483 if (PL_curpm && r == PM_GETRE(PL_curpm))
7484 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7493 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7494 /* no need to distinguish between them any more */
7495 n = RX_BUFF_IDX_FULLMATCH;
7497 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7498 && rx->offs[0].start != -1)
7500 /* $`, ${^PREMATCH} */
7501 i = rx->offs[0].start;
7505 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7506 && rx->offs[0].end != -1)
7508 /* $', ${^POSTMATCH} */
7509 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7510 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7513 if ( 0 <= n && n <= (I32)rx->nparens &&
7514 (s1 = rx->offs[n].start) != -1 &&
7515 (t1 = rx->offs[n].end) != -1)
7517 /* $&, ${^MATCH}, $1 ... */
7519 s = rx->subbeg + s1 - rx->suboffset;
7524 assert(s >= rx->subbeg);
7525 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7527 #ifdef NO_TAINT_SUPPORT
7528 sv_setpvn(sv, s, i);
7530 const int oldtainted = TAINT_get;
7532 sv_setpvn(sv, s, i);
7533 TAINT_set(oldtainted);
7535 if ( (rx->intflags & PREGf_CANY_SEEN)
7536 ? (RXp_MATCH_UTF8(rx)
7537 && (!i || is_utf8_string((U8*)s, i)))
7538 : (RXp_MATCH_UTF8(rx)) )
7545 if (RXp_MATCH_TAINTED(rx)) {
7546 if (SvTYPE(sv) >= SVt_PVMG) {
7547 MAGIC* const mg = SvMAGIC(sv);
7550 SvMAGIC_set(sv, mg->mg_moremagic);
7552 if ((mgt = SvMAGIC(sv))) {
7553 mg->mg_moremagic = mgt;
7554 SvMAGIC_set(sv, mg);
7565 sv_setsv(sv,&PL_sv_undef);
7571 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7572 SV const * const value)
7574 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7576 PERL_UNUSED_ARG(rx);
7577 PERL_UNUSED_ARG(paren);
7578 PERL_UNUSED_ARG(value);
7581 Perl_croak_no_modify();
7585 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7588 struct regexp *const rx = ReANY(r);
7592 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7594 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7595 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7596 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7599 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7601 /* on something like
7604 * the KEEPCOPY is set on the PMOP rather than the regex */
7605 if (PL_curpm && r == PM_GETRE(PL_curpm))
7606 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7612 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7614 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7615 case RX_BUFF_IDX_PREMATCH: /* $` */
7616 if (rx->offs[0].start != -1) {
7617 i = rx->offs[0].start;
7626 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7627 case RX_BUFF_IDX_POSTMATCH: /* $' */
7628 if (rx->offs[0].end != -1) {
7629 i = rx->sublen - rx->offs[0].end;
7631 s1 = rx->offs[0].end;
7638 default: /* $& / ${^MATCH}, $1, $2, ... */
7639 if (paren <= (I32)rx->nparens &&
7640 (s1 = rx->offs[paren].start) != -1 &&
7641 (t1 = rx->offs[paren].end) != -1)
7647 if (ckWARN(WARN_UNINITIALIZED))
7648 report_uninit((const SV *)sv);
7653 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7654 const char * const s = rx->subbeg - rx->suboffset + s1;
7659 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7666 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7668 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7669 PERL_UNUSED_ARG(rx);
7673 return newSVpvs("Regexp");
7676 /* Scans the name of a named buffer from the pattern.
7677 * If flags is REG_RSN_RETURN_NULL returns null.
7678 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7679 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7680 * to the parsed name as looked up in the RExC_paren_names hash.
7681 * If there is an error throws a vFAIL().. type exception.
7684 #define REG_RSN_RETURN_NULL 0
7685 #define REG_RSN_RETURN_NAME 1
7686 #define REG_RSN_RETURN_DATA 2
7689 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7691 char *name_start = RExC_parse;
7693 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7695 assert (RExC_parse <= RExC_end);
7696 if (RExC_parse == RExC_end) NOOP;
7697 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7698 /* skip IDFIRST by using do...while */
7701 RExC_parse += UTF8SKIP(RExC_parse);
7702 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7706 } while (isWORDCHAR(*RExC_parse));
7708 RExC_parse++; /* so the <- from the vFAIL is after the offending
7710 vFAIL("Group name must start with a non-digit word character");
7714 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7715 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7716 if ( flags == REG_RSN_RETURN_NAME)
7718 else if (flags==REG_RSN_RETURN_DATA) {
7721 if ( ! sv_name ) /* should not happen*/
7722 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7723 if (RExC_paren_names)
7724 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7726 sv_dat = HeVAL(he_str);
7728 vFAIL("Reference to nonexistent named group");
7732 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7733 (unsigned long) flags);
7735 assert(0); /* NOT REACHED */
7740 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7741 int rem=(int)(RExC_end - RExC_parse); \
7750 if (RExC_lastparse!=RExC_parse) \
7751 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7754 iscut ? "..." : "<" \
7757 PerlIO_printf(Perl_debug_log,"%16s",""); \
7760 num = RExC_size + 1; \
7762 num=REG_NODE_NUM(RExC_emit); \
7763 if (RExC_lastnum!=num) \
7764 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7766 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7767 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7768 (int)((depth*2)), "", \
7772 RExC_lastparse=RExC_parse; \
7777 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7778 DEBUG_PARSE_MSG((funcname)); \
7779 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7781 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7782 DEBUG_PARSE_MSG((funcname)); \
7783 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7786 /* This section of code defines the inversion list object and its methods. The
7787 * interfaces are highly subject to change, so as much as possible is static to
7788 * this file. An inversion list is here implemented as a malloc'd C UV array
7789 * as an SVt_INVLIST scalar.
7791 * An inversion list for Unicode is an array of code points, sorted by ordinal
7792 * number. The zeroth element is the first code point in the list. The 1th
7793 * element is the first element beyond that not in the list. In other words,
7794 * the first range is
7795 * invlist[0]..(invlist[1]-1)
7796 * The other ranges follow. Thus every element whose index is divisible by two
7797 * marks the beginning of a range that is in the list, and every element not
7798 * divisible by two marks the beginning of a range not in the list. A single
7799 * element inversion list that contains the single code point N generally
7800 * consists of two elements
7803 * (The exception is when N is the highest representable value on the
7804 * machine, in which case the list containing just it would be a single
7805 * element, itself. By extension, if the last range in the list extends to
7806 * infinity, then the first element of that range will be in the inversion list
7807 * at a position that is divisible by two, and is the final element in the
7809 * Taking the complement (inverting) an inversion list is quite simple, if the
7810 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7811 * This implementation reserves an element at the beginning of each inversion
7812 * list to always contain 0; there is an additional flag in the header which
7813 * indicates if the list begins at the 0, or is offset to begin at the next
7816 * More about inversion lists can be found in "Unicode Demystified"
7817 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7818 * More will be coming when functionality is added later.
7820 * The inversion list data structure is currently implemented as an SV pointing
7821 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7822 * array of UV whose memory management is automatically handled by the existing
7823 * facilities for SV's.
7825 * Some of the methods should always be private to the implementation, and some
7826 * should eventually be made public */
7828 /* The header definitions are in F<inline_invlist.c> */
7830 PERL_STATIC_INLINE UV*
7831 S__invlist_array_init(SV* const invlist, const bool will_have_0)
7833 /* Returns a pointer to the first element in the inversion list's array.
7834 * This is called upon initialization of an inversion list. Where the
7835 * array begins depends on whether the list has the code point U+0000 in it
7836 * or not. The other parameter tells it whether the code that follows this
7837 * call is about to put a 0 in the inversion list or not. The first
7838 * element is either the element reserved for 0, if TRUE, or the element
7839 * after it, if FALSE */
7841 bool* offset = get_invlist_offset_addr(invlist);
7842 UV* zero_addr = (UV *) SvPVX(invlist);
7844 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7847 assert(! _invlist_len(invlist));
7851 /* 1^1 = 0; 1^0 = 1 */
7852 *offset = 1 ^ will_have_0;
7853 return zero_addr + *offset;
7856 PERL_STATIC_INLINE UV*
7857 S_invlist_array(SV* const invlist)
7859 /* Returns the pointer to the inversion list's array. Every time the
7860 * length changes, this needs to be called in case malloc or realloc moved
7863 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7865 /* Must not be empty. If these fail, you probably didn't check for <len>
7866 * being non-zero before trying to get the array */
7867 assert(_invlist_len(invlist));
7869 /* The very first element always contains zero, The array begins either
7870 * there, or if the inversion list is offset, at the element after it.
7871 * The offset header field determines which; it contains 0 or 1 to indicate
7872 * how much additionally to add */
7873 assert(0 == *(SvPVX(invlist)));
7874 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7877 PERL_STATIC_INLINE void
7878 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7880 /* Sets the current number of elements stored in the inversion list.
7881 * Updates SvCUR correspondingly */
7882 PERL_UNUSED_CONTEXT;
7883 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7885 assert(SvTYPE(invlist) == SVt_INVLIST);
7890 : TO_INTERNAL_SIZE(len + offset));
7891 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7894 PERL_STATIC_INLINE IV*
7895 S_get_invlist_previous_index_addr(SV* invlist)
7897 /* Return the address of the IV that is reserved to hold the cached index
7899 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7901 assert(SvTYPE(invlist) == SVt_INVLIST);
7903 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7906 PERL_STATIC_INLINE IV
7907 S_invlist_previous_index(SV* const invlist)
7909 /* Returns cached index of previous search */
7911 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7913 return *get_invlist_previous_index_addr(invlist);
7916 PERL_STATIC_INLINE void
7917 S_invlist_set_previous_index(SV* const invlist, const IV index)
7919 /* Caches <index> for later retrieval */
7921 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7923 assert(index == 0 || index < (int) _invlist_len(invlist));
7925 *get_invlist_previous_index_addr(invlist) = index;
7928 PERL_STATIC_INLINE UV
7929 S_invlist_max(SV* const invlist)
7931 /* Returns the maximum number of elements storable in the inversion list's
7932 * array, without having to realloc() */
7934 PERL_ARGS_ASSERT_INVLIST_MAX;
7936 assert(SvTYPE(invlist) == SVt_INVLIST);
7938 /* Assumes worst case, in which the 0 element is not counted in the
7939 * inversion list, so subtracts 1 for that */
7940 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7941 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7942 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7945 #ifndef PERL_IN_XSUB_RE
7947 Perl__new_invlist(pTHX_ IV initial_size)
7950 /* Return a pointer to a newly constructed inversion list, with enough
7951 * space to store 'initial_size' elements. If that number is negative, a
7952 * system default is used instead */
7956 if (initial_size < 0) {
7960 /* Allocate the initial space */
7961 new_list = newSV_type(SVt_INVLIST);
7963 /* First 1 is in case the zero element isn't in the list; second 1 is for
7965 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7966 invlist_set_len(new_list, 0, 0);
7968 /* Force iterinit() to be used to get iteration to work */
7969 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7971 *get_invlist_previous_index_addr(new_list) = 0;
7977 Perl__new_invlist_C_array(pTHX_ const UV* const list)
7979 /* Return a pointer to a newly constructed inversion list, initialized to
7980 * point to <list>, which has to be in the exact correct inversion list
7981 * form, including internal fields. Thus this is a dangerous routine that
7982 * should not be used in the wrong hands. The passed in 'list' contains
7983 * several header fields at the beginning that are not part of the
7984 * inversion list body proper */
7986 const STRLEN length = (STRLEN) list[0];
7987 const UV version_id = list[1];
7988 const bool offset = cBOOL(list[2]);
7989 #define HEADER_LENGTH 3
7990 /* If any of the above changes in any way, you must change HEADER_LENGTH
7991 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7992 * perl -E 'say int(rand 2**31-1)'
7994 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7995 data structure type, so that one being
7996 passed in can be validated to be an
7997 inversion list of the correct vintage.
8000 SV* invlist = newSV_type(SVt_INVLIST);
8002 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8004 if (version_id != INVLIST_VERSION_ID) {
8005 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8008 /* The generated array passed in includes header elements that aren't part
8009 * of the list proper, so start it just after them */
8010 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8012 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8013 shouldn't touch it */
8015 *(get_invlist_offset_addr(invlist)) = offset;
8017 /* The 'length' passed to us is the physical number of elements in the
8018 * inversion list. But if there is an offset the logical number is one
8020 invlist_set_len(invlist, length - offset, offset);
8022 invlist_set_previous_index(invlist, 0);
8024 /* Initialize the iteration pointer. */
8025 invlist_iterfinish(invlist);
8027 SvREADONLY_on(invlist);
8031 #endif /* ifndef PERL_IN_XSUB_RE */
8034 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8036 /* Grow the maximum size of an inversion list */
8038 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8040 assert(SvTYPE(invlist) == SVt_INVLIST);
8042 /* Add one to account for the zero element at the beginning which may not
8043 * be counted by the calling parameters */
8044 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8047 PERL_STATIC_INLINE void
8048 S_invlist_trim(SV* const invlist)
8050 PERL_ARGS_ASSERT_INVLIST_TRIM;
8052 assert(SvTYPE(invlist) == SVt_INVLIST);
8054 /* Change the length of the inversion list to how many entries it currently
8056 SvPV_shrink_to_cur((SV *) invlist);
8060 S__append_range_to_invlist(pTHX_ SV* const invlist,
8061 const UV start, const UV end)
8063 /* Subject to change or removal. Append the range from 'start' to 'end' at
8064 * the end of the inversion list. The range must be above any existing
8068 UV max = invlist_max(invlist);
8069 UV len = _invlist_len(invlist);
8072 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8074 if (len == 0) { /* Empty lists must be initialized */
8075 offset = start != 0;
8076 array = _invlist_array_init(invlist, ! offset);
8079 /* Here, the existing list is non-empty. The current max entry in the
8080 * list is generally the first value not in the set, except when the
8081 * set extends to the end of permissible values, in which case it is
8082 * the first entry in that final set, and so this call is an attempt to
8083 * append out-of-order */
8085 UV final_element = len - 1;
8086 array = invlist_array(invlist);
8087 if (array[final_element] > start
8088 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8090 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",
8091 array[final_element], start,
8092 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8095 /* Here, it is a legal append. If the new range begins with the first
8096 * value not in the set, it is extending the set, so the new first
8097 * value not in the set is one greater than the newly extended range.
8099 offset = *get_invlist_offset_addr(invlist);
8100 if (array[final_element] == start) {
8101 if (end != UV_MAX) {
8102 array[final_element] = end + 1;
8105 /* But if the end is the maximum representable on the machine,
8106 * just let the range that this would extend to have no end */
8107 invlist_set_len(invlist, len - 1, offset);
8113 /* Here the new range doesn't extend any existing set. Add it */
8115 len += 2; /* Includes an element each for the start and end of range */
8117 /* If wll overflow the existing space, extend, which may cause the array to
8120 invlist_extend(invlist, len);
8122 /* Have to set len here to avoid assert failure in invlist_array() */
8123 invlist_set_len(invlist, len, offset);
8125 array = invlist_array(invlist);
8128 invlist_set_len(invlist, len, offset);
8131 /* The next item on the list starts the range, the one after that is
8132 * one past the new range. */
8133 array[len - 2] = start;
8134 if (end != UV_MAX) {
8135 array[len - 1] = end + 1;
8138 /* But if the end is the maximum representable on the machine, just let
8139 * the range have no end */
8140 invlist_set_len(invlist, len - 1, offset);
8144 #ifndef PERL_IN_XSUB_RE
8147 Perl__invlist_search(SV* const invlist, const UV cp)
8149 /* Searches the inversion list for the entry that contains the input code
8150 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8151 * return value is the index into the list's array of the range that
8156 IV high = _invlist_len(invlist);
8157 const IV highest_element = high - 1;
8160 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8162 /* If list is empty, return failure. */
8167 /* (We can't get the array unless we know the list is non-empty) */
8168 array = invlist_array(invlist);
8170 mid = invlist_previous_index(invlist);
8171 assert(mid >=0 && mid <= highest_element);
8173 /* <mid> contains the cache of the result of the previous call to this
8174 * function (0 the first time). See if this call is for the same result,
8175 * or if it is for mid-1. This is under the theory that calls to this
8176 * function will often be for related code points that are near each other.
8177 * And benchmarks show that caching gives better results. We also test
8178 * here if the code point is within the bounds of the list. These tests
8179 * replace others that would have had to be made anyway to make sure that
8180 * the array bounds were not exceeded, and these give us extra information
8181 * at the same time */
8182 if (cp >= array[mid]) {
8183 if (cp >= array[highest_element]) {
8184 return highest_element;
8187 /* Here, array[mid] <= cp < array[highest_element]. This means that
8188 * the final element is not the answer, so can exclude it; it also
8189 * means that <mid> is not the final element, so can refer to 'mid + 1'
8191 if (cp < array[mid + 1]) {
8197 else { /* cp < aray[mid] */
8198 if (cp < array[0]) { /* Fail if outside the array */
8202 if (cp >= array[mid - 1]) {
8207 /* Binary search. What we are looking for is <i> such that
8208 * array[i] <= cp < array[i+1]
8209 * The loop below converges on the i+1. Note that there may not be an
8210 * (i+1)th element in the array, and things work nonetheless */
8211 while (low < high) {
8212 mid = (low + high) / 2;
8213 assert(mid <= highest_element);
8214 if (array[mid] <= cp) { /* cp >= array[mid] */
8217 /* We could do this extra test to exit the loop early.
8218 if (cp < array[low]) {
8223 else { /* cp < array[mid] */
8230 invlist_set_previous_index(invlist, high);
8235 Perl__invlist_populate_swatch(SV* const invlist,
8236 const UV start, const UV end, U8* swatch)
8238 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8239 * but is used when the swash has an inversion list. This makes this much
8240 * faster, as it uses a binary search instead of a linear one. This is
8241 * intimately tied to that function, and perhaps should be in utf8.c,
8242 * except it is intimately tied to inversion lists as well. It assumes
8243 * that <swatch> is all 0's on input */
8246 const IV len = _invlist_len(invlist);
8250 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8252 if (len == 0) { /* Empty inversion list */
8256 array = invlist_array(invlist);
8258 /* Find which element it is */
8259 i = _invlist_search(invlist, start);
8261 /* We populate from <start> to <end> */
8262 while (current < end) {
8265 /* The inversion list gives the results for every possible code point
8266 * after the first one in the list. Only those ranges whose index is
8267 * even are ones that the inversion list matches. For the odd ones,
8268 * and if the initial code point is not in the list, we have to skip
8269 * forward to the next element */
8270 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8272 if (i >= len) { /* Finished if beyond the end of the array */
8276 if (current >= end) { /* Finished if beyond the end of what we
8278 if (LIKELY(end < UV_MAX)) {
8282 /* We get here when the upper bound is the maximum
8283 * representable on the machine, and we are looking for just
8284 * that code point. Have to special case it */
8286 goto join_end_of_list;
8289 assert(current >= start);
8291 /* The current range ends one below the next one, except don't go past
8294 upper = (i < len && array[i] < end) ? array[i] : end;
8296 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8297 * for each code point in it */
8298 for (; current < upper; current++) {
8299 const STRLEN offset = (STRLEN)(current - start);
8300 swatch[offset >> 3] |= 1 << (offset & 7);
8305 /* Quit if at the end of the list */
8308 /* But first, have to deal with the highest possible code point on
8309 * the platform. The previous code assumes that <end> is one
8310 * beyond where we want to populate, but that is impossible at the
8311 * platform's infinity, so have to handle it specially */
8312 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8314 const STRLEN offset = (STRLEN)(end - start);
8315 swatch[offset >> 3] |= 1 << (offset & 7);
8320 /* Advance to the next range, which will be for code points not in the
8329 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8330 const bool complement_b, SV** output)
8332 /* Take the union of two inversion lists and point <output> to it. *output
8333 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8334 * the reference count to that list will be decremented if not already a
8335 * temporary (mortal); otherwise *output will be made correspondingly
8336 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8337 * second list is returned. If <complement_b> is TRUE, the union is taken
8338 * of the complement (inversion) of <b> instead of b itself.
8340 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8341 * Richard Gillam, published by Addison-Wesley, and explained at some
8342 * length there. The preface says to incorporate its examples into your
8343 * code at your own risk.
8345 * The algorithm is like a merge sort.
8347 * XXX A potential performance improvement is to keep track as we go along
8348 * if only one of the inputs contributes to the result, meaning the other
8349 * is a subset of that one. In that case, we can skip the final copy and
8350 * return the larger of the input lists, but then outside code might need
8351 * to keep track of whether to free the input list or not */
8353 const UV* array_a; /* a's array */
8355 UV len_a; /* length of a's array */
8358 SV* u; /* the resulting union */
8362 UV i_a = 0; /* current index into a's array */
8366 /* running count, as explained in the algorithm source book; items are
8367 * stopped accumulating and are output when the count changes to/from 0.
8368 * The count is incremented when we start a range that's in the set, and
8369 * decremented when we start a range that's not in the set. So its range
8370 * is 0 to 2. Only when the count is zero is something not in the set.
8374 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8377 /* If either one is empty, the union is the other one */
8378 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8379 bool make_temp = FALSE; /* Should we mortalize the result? */
8383 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8389 *output = invlist_clone(b);
8391 _invlist_invert(*output);
8393 } /* else *output already = b; */
8396 sv_2mortal(*output);
8400 else if ((len_b = _invlist_len(b)) == 0) {
8401 bool make_temp = FALSE;
8403 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8408 /* The complement of an empty list is a list that has everything in it,
8409 * so the union with <a> includes everything too */
8412 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8416 *output = _new_invlist(1);
8417 _append_range_to_invlist(*output, 0, UV_MAX);
8419 else if (*output != a) {
8420 *output = invlist_clone(a);
8422 /* else *output already = a; */
8425 sv_2mortal(*output);
8430 /* Here both lists exist and are non-empty */
8431 array_a = invlist_array(a);
8432 array_b = invlist_array(b);
8434 /* If are to take the union of 'a' with the complement of b, set it
8435 * up so are looking at b's complement. */
8438 /* To complement, we invert: if the first element is 0, remove it. To
8439 * do this, we just pretend the array starts one later */
8440 if (array_b[0] == 0) {
8446 /* But if the first element is not zero, we pretend the list starts
8447 * at the 0 that is always stored immediately before the array. */
8453 /* Size the union for the worst case: that the sets are completely
8455 u = _new_invlist(len_a + len_b);
8457 /* Will contain U+0000 if either component does */
8458 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8459 || (len_b > 0 && array_b[0] == 0));
8461 /* Go through each list item by item, stopping when exhausted one of
8463 while (i_a < len_a && i_b < len_b) {
8464 UV cp; /* The element to potentially add to the union's array */
8465 bool cp_in_set; /* is it in the the input list's set or not */
8467 /* We need to take one or the other of the two inputs for the union.
8468 * Since we are merging two sorted lists, we take the smaller of the
8469 * next items. In case of a tie, we take the one that is in its set
8470 * first. If we took one not in the set first, it would decrement the
8471 * count, possibly to 0 which would cause it to be output as ending the
8472 * range, and the next time through we would take the same number, and
8473 * output it again as beginning the next range. By doing it the
8474 * opposite way, there is no possibility that the count will be
8475 * momentarily decremented to 0, and thus the two adjoining ranges will
8476 * be seamlessly merged. (In a tie and both are in the set or both not
8477 * in the set, it doesn't matter which we take first.) */
8478 if (array_a[i_a] < array_b[i_b]
8479 || (array_a[i_a] == array_b[i_b]
8480 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8482 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8486 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8487 cp = array_b[i_b++];
8490 /* Here, have chosen which of the two inputs to look at. Only output
8491 * if the running count changes to/from 0, which marks the
8492 * beginning/end of a range in that's in the set */
8495 array_u[i_u++] = cp;
8502 array_u[i_u++] = cp;
8507 /* Here, we are finished going through at least one of the lists, which
8508 * means there is something remaining in at most one. We check if the list
8509 * that hasn't been exhausted is positioned such that we are in the middle
8510 * of a range in its set or not. (i_a and i_b point to the element beyond
8511 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8512 * is potentially more to output.
8513 * There are four cases:
8514 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8515 * in the union is entirely from the non-exhausted set.
8516 * 2) Both were in their sets, count is 2. Nothing further should
8517 * be output, as everything that remains will be in the exhausted
8518 * list's set, hence in the union; decrementing to 1 but not 0 insures
8520 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8521 * Nothing further should be output because the union includes
8522 * everything from the exhausted set. Not decrementing ensures that.
8523 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8524 * decrementing to 0 insures that we look at the remainder of the
8525 * non-exhausted set */
8526 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8527 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8532 /* The final length is what we've output so far, plus what else is about to
8533 * be output. (If 'count' is non-zero, then the input list we exhausted
8534 * has everything remaining up to the machine's limit in its set, and hence
8535 * in the union, so there will be no further output. */
8538 /* At most one of the subexpressions will be non-zero */
8539 len_u += (len_a - i_a) + (len_b - i_b);
8542 /* Set result to final length, which can change the pointer to array_u, so
8544 if (len_u != _invlist_len(u)) {
8545 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8547 array_u = invlist_array(u);
8550 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8551 * the other) ended with everything above it not in its set. That means
8552 * that the remaining part of the union is precisely the same as the
8553 * non-exhausted list, so can just copy it unchanged. (If both list were
8554 * exhausted at the same time, then the operations below will be both 0.)
8557 IV copy_count; /* At most one will have a non-zero copy count */
8558 if ((copy_count = len_a - i_a) > 0) {
8559 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8561 else if ((copy_count = len_b - i_b) > 0) {
8562 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8566 /* We may be removing a reference to one of the inputs. If so, the output
8567 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8568 * count decremented) */
8569 if (a == *output || b == *output) {
8570 assert(! invlist_is_iterating(*output));
8571 if ((SvTEMP(*output))) {
8575 SvREFCNT_dec_NN(*output);
8585 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8586 const bool complement_b, SV** i)
8588 /* Take the intersection of two inversion lists and point <i> to it. *i
8589 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8590 * the reference count to that list will be decremented if not already a
8591 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8592 * The first list, <a>, may be NULL, in which case an empty list is
8593 * returned. If <complement_b> is TRUE, the result will be the
8594 * intersection of <a> and the complement (or inversion) of <b> instead of
8597 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8598 * Richard Gillam, published by Addison-Wesley, and explained at some
8599 * length there. The preface says to incorporate its examples into your
8600 * code at your own risk. In fact, it had bugs
8602 * The algorithm is like a merge sort, and is essentially the same as the
8606 const UV* array_a; /* a's array */
8608 UV len_a; /* length of a's array */
8611 SV* r; /* the resulting intersection */
8615 UV i_a = 0; /* current index into a's array */
8619 /* running count, as explained in the algorithm source book; items are
8620 * stopped accumulating and are output when the count changes to/from 2.
8621 * The count is incremented when we start a range that's in the set, and
8622 * decremented when we start a range that's not in the set. So its range
8623 * is 0 to 2. Only when the count is 2 is something in the intersection.
8627 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8630 /* Special case if either one is empty */
8631 len_a = (a == NULL) ? 0 : _invlist_len(a);
8632 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8633 bool make_temp = FALSE;
8635 if (len_a != 0 && complement_b) {
8637 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8638 * be empty. Here, also we are using 'b's complement, which hence
8639 * must be every possible code point. Thus the intersection is
8643 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8648 *i = invlist_clone(a);
8650 /* else *i is already 'a' */
8658 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8659 * intersection must be empty */
8661 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8666 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8670 *i = _new_invlist(0);
8678 /* Here both lists exist and are non-empty */
8679 array_a = invlist_array(a);
8680 array_b = invlist_array(b);
8682 /* If are to take the intersection of 'a' with the complement of b, set it
8683 * up so are looking at b's complement. */
8686 /* To complement, we invert: if the first element is 0, remove it. To
8687 * do this, we just pretend the array starts one later */
8688 if (array_b[0] == 0) {
8694 /* But if the first element is not zero, we pretend the list starts
8695 * at the 0 that is always stored immediately before the array. */
8701 /* Size the intersection for the worst case: that the intersection ends up
8702 * fragmenting everything to be completely disjoint */
8703 r= _new_invlist(len_a + len_b);
8705 /* Will contain U+0000 iff both components do */
8706 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8707 && len_b > 0 && array_b[0] == 0);
8709 /* Go through each list item by item, stopping when exhausted one of
8711 while (i_a < len_a && i_b < len_b) {
8712 UV cp; /* The element to potentially add to the intersection's
8714 bool cp_in_set; /* Is it in the input list's set or not */
8716 /* We need to take one or the other of the two inputs for the
8717 * intersection. Since we are merging two sorted lists, we take the
8718 * smaller of the next items. In case of a tie, we take the one that
8719 * is not in its set first (a difference from the union algorithm). If
8720 * we took one in the set first, it would increment the count, possibly
8721 * to 2 which would cause it to be output as starting a range in the
8722 * intersection, and the next time through we would take that same
8723 * number, and output it again as ending the set. By doing it the
8724 * opposite of this, there is no possibility that the count will be
8725 * momentarily incremented to 2. (In a tie and both are in the set or
8726 * both not in the set, it doesn't matter which we take first.) */
8727 if (array_a[i_a] < array_b[i_b]
8728 || (array_a[i_a] == array_b[i_b]
8729 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8731 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8735 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8739 /* Here, have chosen which of the two inputs to look at. Only output
8740 * if the running count changes to/from 2, which marks the
8741 * beginning/end of a range that's in the intersection */
8745 array_r[i_r++] = cp;
8750 array_r[i_r++] = cp;
8756 /* Here, we are finished going through at least one of the lists, which
8757 * means there is something remaining in at most one. We check if the list
8758 * that has been exhausted is positioned such that we are in the middle
8759 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8760 * the ones we care about.) There are four cases:
8761 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8762 * nothing left in the intersection.
8763 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8764 * above 2. What should be output is exactly that which is in the
8765 * non-exhausted set, as everything it has is also in the intersection
8766 * set, and everything it doesn't have can't be in the intersection
8767 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8768 * gets incremented to 2. Like the previous case, the intersection is
8769 * everything that remains in the non-exhausted set.
8770 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8771 * remains 1. And the intersection has nothing more. */
8772 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8773 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8778 /* The final length is what we've output so far plus what else is in the
8779 * intersection. At most one of the subexpressions below will be non-zero
8783 len_r += (len_a - i_a) + (len_b - i_b);
8786 /* Set result to final length, which can change the pointer to array_r, so
8788 if (len_r != _invlist_len(r)) {
8789 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8791 array_r = invlist_array(r);
8794 /* Finish outputting any remaining */
8795 if (count >= 2) { /* At most one will have a non-zero copy count */
8797 if ((copy_count = len_a - i_a) > 0) {
8798 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8800 else if ((copy_count = len_b - i_b) > 0) {
8801 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8805 /* We may be removing a reference to one of the inputs. If so, the output
8806 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8807 * count decremented) */
8808 if (a == *i || b == *i) {
8809 assert(! invlist_is_iterating(*i));
8814 SvREFCNT_dec_NN(*i);
8824 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8826 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8827 * set. A pointer to the inversion list is returned. This may actually be
8828 * a new list, in which case the passed in one has been destroyed. The
8829 * passed in inversion list can be NULL, in which case a new one is created
8830 * with just the one range in it */
8835 if (invlist == NULL) {
8836 invlist = _new_invlist(2);
8840 len = _invlist_len(invlist);
8843 /* If comes after the final entry actually in the list, can just append it
8846 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8847 && start >= invlist_array(invlist)[len - 1]))
8849 _append_range_to_invlist(invlist, start, end);
8853 /* Here, can't just append things, create and return a new inversion list
8854 * which is the union of this range and the existing inversion list */
8855 range_invlist = _new_invlist(2);
8856 _append_range_to_invlist(range_invlist, start, end);
8858 _invlist_union(invlist, range_invlist, &invlist);
8860 /* The temporary can be freed */
8861 SvREFCNT_dec_NN(range_invlist);
8867 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
8868 UV** other_elements_ptr)
8870 /* Create and return an inversion list whose contents are to be populated
8871 * by the caller. The caller gives the number of elements (in 'size') and
8872 * the very first element ('element0'). This function will set
8873 * '*other_elements_ptr' to an array of UVs, where the remaining elements
8876 * Obviously there is some trust involved that the caller will properly
8877 * fill in the other elements of the array.
8879 * (The first element needs to be passed in, as the underlying code does
8880 * things differently depending on whether it is zero or non-zero) */
8882 SV* invlist = _new_invlist(size);
8885 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
8887 _append_range_to_invlist(invlist, element0, element0);
8888 offset = *get_invlist_offset_addr(invlist);
8890 invlist_set_len(invlist, size, offset);
8891 *other_elements_ptr = invlist_array(invlist) + 1;
8897 PERL_STATIC_INLINE SV*
8898 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8899 return _add_range_to_invlist(invlist, cp, cp);
8902 #ifndef PERL_IN_XSUB_RE
8904 Perl__invlist_invert(pTHX_ SV* const invlist)
8906 /* Complement the input inversion list. This adds a 0 if the list didn't
8907 * have a zero; removes it otherwise. As described above, the data
8908 * structure is set up so that this is very efficient */
8910 PERL_ARGS_ASSERT__INVLIST_INVERT;
8912 assert(! invlist_is_iterating(invlist));
8914 /* The inverse of matching nothing is matching everything */
8915 if (_invlist_len(invlist) == 0) {
8916 _append_range_to_invlist(invlist, 0, UV_MAX);
8920 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8925 PERL_STATIC_INLINE SV*
8926 S_invlist_clone(pTHX_ SV* const invlist)
8929 /* Return a new inversion list that is a copy of the input one, which is
8930 * unchanged. The new list will not be mortal even if the old one was. */
8932 /* Need to allocate extra space to accommodate Perl's addition of a
8933 * trailing NUL to SvPV's, since it thinks they are always strings */
8934 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8935 STRLEN physical_length = SvCUR(invlist);
8936 bool offset = *(get_invlist_offset_addr(invlist));
8938 PERL_ARGS_ASSERT_INVLIST_CLONE;
8940 *(get_invlist_offset_addr(new_invlist)) = offset;
8941 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8942 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8947 PERL_STATIC_INLINE STRLEN*
8948 S_get_invlist_iter_addr(SV* invlist)
8950 /* Return the address of the UV that contains the current iteration
8953 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8955 assert(SvTYPE(invlist) == SVt_INVLIST);
8957 return &(((XINVLIST*) SvANY(invlist))->iterator);
8960 PERL_STATIC_INLINE void
8961 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
8963 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8965 *get_invlist_iter_addr(invlist) = 0;
8968 PERL_STATIC_INLINE void
8969 S_invlist_iterfinish(SV* invlist)
8971 /* Terminate iterator for invlist. This is to catch development errors.
8972 * Any iteration that is interrupted before completed should call this
8973 * function. Functions that add code points anywhere else but to the end
8974 * of an inversion list assert that they are not in the middle of an
8975 * iteration. If they were, the addition would make the iteration
8976 * problematical: if the iteration hadn't reached the place where things
8977 * were being added, it would be ok */
8979 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8981 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8985 S_invlist_iternext(SV* invlist, UV* start, UV* end)
8987 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8988 * This call sets in <*start> and <*end>, the next range in <invlist>.
8989 * Returns <TRUE> if successful and the next call will return the next
8990 * range; <FALSE> if was already at the end of the list. If the latter,
8991 * <*start> and <*end> are unchanged, and the next call to this function
8992 * will start over at the beginning of the list */
8994 STRLEN* pos = get_invlist_iter_addr(invlist);
8995 UV len = _invlist_len(invlist);
8998 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9001 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9005 array = invlist_array(invlist);
9007 *start = array[(*pos)++];
9013 *end = array[(*pos)++] - 1;
9019 PERL_STATIC_INLINE bool
9020 S_invlist_is_iterating(SV* const invlist)
9022 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
9024 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
9027 PERL_STATIC_INLINE UV
9028 S_invlist_highest(SV* const invlist)
9030 /* Returns the highest code point that matches an inversion list. This API
9031 * has an ambiguity, as it returns 0 under either the highest is actually
9032 * 0, or if the list is empty. If this distinction matters to you, check
9033 * for emptiness before calling this function */
9035 UV len = _invlist_len(invlist);
9038 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9044 array = invlist_array(invlist);
9046 /* The last element in the array in the inversion list always starts a
9047 * range that goes to infinity. That range may be for code points that are
9048 * matched in the inversion list, or it may be for ones that aren't
9049 * matched. In the latter case, the highest code point in the set is one
9050 * less than the beginning of this range; otherwise it is the final element
9051 * of this range: infinity */
9052 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9054 : array[len - 1] - 1;
9057 #ifndef PERL_IN_XSUB_RE
9059 Perl__invlist_contents(pTHX_ SV* const invlist)
9061 /* Get the contents of an inversion list into a string SV so that they can
9062 * be printed out. It uses the format traditionally done for debug tracing
9066 SV* output = newSVpvs("\n");
9068 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9070 assert(! invlist_is_iterating(invlist));
9072 invlist_iterinit(invlist);
9073 while (invlist_iternext(invlist, &start, &end)) {
9074 if (end == UV_MAX) {
9075 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9077 else if (end != start) {
9078 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9082 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9090 #ifndef PERL_IN_XSUB_RE
9092 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9093 const char * const indent, SV* const invlist)
9095 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9096 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9097 * the string 'indent'. The output looks like this:
9098 [0] 0x000A .. 0x000D
9100 [4] 0x2028 .. 0x2029
9101 [6] 0x3104 .. INFINITY
9102 * This means that the first range of code points matched by the list are
9103 * 0xA through 0xD; the second range contains only the single code point
9104 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9105 * are used to define each range (except if the final range extends to
9106 * infinity, only a single element is needed). The array index of the
9107 * first element for the corresponding range is given in brackets. */
9112 PERL_ARGS_ASSERT__INVLIST_DUMP;
9114 if (invlist_is_iterating(invlist)) {
9115 Perl_dump_indent(aTHX_ level, file,
9116 "%sCan't dump inversion list because is in middle of iterating\n",
9121 invlist_iterinit(invlist);
9122 while (invlist_iternext(invlist, &start, &end)) {
9123 if (end == UV_MAX) {
9124 Perl_dump_indent(aTHX_ level, file,
9125 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9126 indent, (UV)count, start);
9128 else if (end != start) {
9129 Perl_dump_indent(aTHX_ level, file,
9130 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9131 indent, (UV)count, start, end);
9134 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9135 indent, (UV)count, start);
9142 Perl__load_PL_utf8_foldclosures (pTHX)
9144 assert(! PL_utf8_foldclosures);
9146 /* If the folds haven't been read in, call a fold function
9148 if (! PL_utf8_tofold) {
9149 U8 dummy[UTF8_MAXBYTES_CASE+1];
9151 /* This string is just a short named one above \xff */
9152 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
9153 assert(PL_utf8_tofold); /* Verify that worked */
9155 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
9159 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9161 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9163 /* Return a boolean as to if the two passed in inversion lists are
9164 * identical. The final argument, if TRUE, says to take the complement of
9165 * the second inversion list before doing the comparison */
9167 const UV* array_a = invlist_array(a);
9168 const UV* array_b = invlist_array(b);
9169 UV len_a = _invlist_len(a);
9170 UV len_b = _invlist_len(b);
9172 UV i = 0; /* current index into the arrays */
9173 bool retval = TRUE; /* Assume are identical until proven otherwise */
9175 PERL_ARGS_ASSERT__INVLISTEQ;
9177 /* If are to compare 'a' with the complement of b, set it
9178 * up so are looking at b's complement. */
9181 /* The complement of nothing is everything, so <a> would have to have
9182 * just one element, starting at zero (ending at infinity) */
9184 return (len_a == 1 && array_a[0] == 0);
9186 else if (array_b[0] == 0) {
9188 /* Otherwise, to complement, we invert. Here, the first element is
9189 * 0, just remove it. To do this, we just pretend the array starts
9197 /* But if the first element is not zero, we pretend the list starts
9198 * at the 0 that is always stored immediately before the array. */
9204 /* Make sure that the lengths are the same, as well as the final element
9205 * before looping through the remainder. (Thus we test the length, final,
9206 * and first elements right off the bat) */
9207 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9210 else for (i = 0; i < len_a - 1; i++) {
9211 if (array_a[i] != array_b[i]) {
9221 #undef HEADER_LENGTH
9222 #undef TO_INTERNAL_SIZE
9223 #undef FROM_INTERNAL_SIZE
9224 #undef INVLIST_VERSION_ID
9226 /* End of inversion list object */
9229 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9231 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9232 * constructs, and updates RExC_flags with them. On input, RExC_parse
9233 * should point to the first flag; it is updated on output to point to the
9234 * final ')' or ':'. There needs to be at least one flag, or this will
9237 /* for (?g), (?gc), and (?o) warnings; warning
9238 about (?c) will warn about (?g) -- japhy */
9240 #define WASTED_O 0x01
9241 #define WASTED_G 0x02
9242 #define WASTED_C 0x04
9243 #define WASTED_GC (WASTED_G|WASTED_C)
9244 I32 wastedflags = 0x00;
9245 U32 posflags = 0, negflags = 0;
9246 U32 *flagsp = &posflags;
9247 char has_charset_modifier = '\0';
9249 bool has_use_defaults = FALSE;
9250 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9252 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9254 /* '^' as an initial flag sets certain defaults */
9255 if (UCHARAT(RExC_parse) == '^') {
9257 has_use_defaults = TRUE;
9258 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9259 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9260 ? REGEX_UNICODE_CHARSET
9261 : REGEX_DEPENDS_CHARSET);
9264 cs = get_regex_charset(RExC_flags);
9265 if (cs == REGEX_DEPENDS_CHARSET
9266 && (RExC_utf8 || RExC_uni_semantics))
9268 cs = REGEX_UNICODE_CHARSET;
9271 while (*RExC_parse) {
9272 /* && strchr("iogcmsx", *RExC_parse) */
9273 /* (?g), (?gc) and (?o) are useless here
9274 and must be globally applied -- japhy */
9275 switch (*RExC_parse) {
9277 /* Code for the imsx flags */
9278 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
9280 case LOCALE_PAT_MOD:
9281 if (has_charset_modifier) {
9282 goto excess_modifier;
9284 else if (flagsp == &negflags) {
9287 cs = REGEX_LOCALE_CHARSET;
9288 has_charset_modifier = LOCALE_PAT_MOD;
9290 case UNICODE_PAT_MOD:
9291 if (has_charset_modifier) {
9292 goto excess_modifier;
9294 else if (flagsp == &negflags) {
9297 cs = REGEX_UNICODE_CHARSET;
9298 has_charset_modifier = UNICODE_PAT_MOD;
9300 case ASCII_RESTRICT_PAT_MOD:
9301 if (flagsp == &negflags) {
9304 if (has_charset_modifier) {
9305 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9306 goto excess_modifier;
9308 /* Doubled modifier implies more restricted */
9309 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9312 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9314 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9316 case DEPENDS_PAT_MOD:
9317 if (has_use_defaults) {
9318 goto fail_modifiers;
9320 else if (flagsp == &negflags) {
9323 else if (has_charset_modifier) {
9324 goto excess_modifier;
9327 /* The dual charset means unicode semantics if the
9328 * pattern (or target, not known until runtime) are
9329 * utf8, or something in the pattern indicates unicode
9331 cs = (RExC_utf8 || RExC_uni_semantics)
9332 ? REGEX_UNICODE_CHARSET
9333 : REGEX_DEPENDS_CHARSET;
9334 has_charset_modifier = DEPENDS_PAT_MOD;
9338 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9339 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9341 else if (has_charset_modifier == *(RExC_parse - 1)) {
9342 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9346 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9351 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9354 case ONCE_PAT_MOD: /* 'o' */
9355 case GLOBAL_PAT_MOD: /* 'g' */
9356 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9357 const I32 wflagbit = *RExC_parse == 'o'
9360 if (! (wastedflags & wflagbit) ) {
9361 wastedflags |= wflagbit;
9362 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9365 "Useless (%s%c) - %suse /%c modifier",
9366 flagsp == &negflags ? "?-" : "?",
9368 flagsp == &negflags ? "don't " : "",
9375 case CONTINUE_PAT_MOD: /* 'c' */
9376 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9377 if (! (wastedflags & WASTED_C) ) {
9378 wastedflags |= WASTED_GC;
9379 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9382 "Useless (%sc) - %suse /gc modifier",
9383 flagsp == &negflags ? "?-" : "?",
9384 flagsp == &negflags ? "don't " : ""
9389 case KEEPCOPY_PAT_MOD: /* 'p' */
9390 if (flagsp == &negflags) {
9392 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9394 *flagsp |= RXf_PMf_KEEPCOPY;
9398 /* A flag is a default iff it is following a minus, so
9399 * if there is a minus, it means will be trying to
9400 * re-specify a default which is an error */
9401 if (has_use_defaults || flagsp == &negflags) {
9402 goto fail_modifiers;
9405 wastedflags = 0; /* reset so (?g-c) warns twice */
9409 RExC_flags |= posflags;
9410 RExC_flags &= ~negflags;
9411 set_regex_charset(&RExC_flags, cs);
9412 if (RExC_flags & RXf_PMf_FOLD) {
9413 RExC_contains_i = 1;
9419 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9420 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9421 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9422 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9431 - reg - regular expression, i.e. main body or parenthesized thing
9433 * Caller must absorb opening parenthesis.
9435 * Combining parenthesis handling with the base level of regular expression
9436 * is a trifle forced, but the need to tie the tails of the branches to what
9437 * follows makes it hard to avoid.
9439 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9441 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9443 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9446 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9447 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9448 needs to be restarted.
9449 Otherwise would only return NULL if regbranch() returns NULL, which
9452 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9453 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9454 * 2 is like 1, but indicates that nextchar() has been called to advance
9455 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9456 * this flag alerts us to the need to check for that */
9458 regnode *ret; /* Will be the head of the group. */
9461 regnode *ender = NULL;
9464 U32 oregflags = RExC_flags;
9465 bool have_branch = 0;
9467 I32 freeze_paren = 0;
9468 I32 after_freeze = 0;
9469 I32 num; /* numeric backreferences */
9471 char * parse_start = RExC_parse; /* MJD */
9472 char * const oregcomp_parse = RExC_parse;
9474 GET_RE_DEBUG_FLAGS_DECL;
9476 PERL_ARGS_ASSERT_REG;
9477 DEBUG_PARSE("reg ");
9479 *flagp = 0; /* Tentatively. */
9482 /* Make an OPEN node, if parenthesized. */
9485 /* Under /x, space and comments can be gobbled up between the '(' and
9486 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9487 * intervening space, as the sequence is a token, and a token should be
9489 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9491 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9492 char *start_verb = RExC_parse;
9493 STRLEN verb_len = 0;
9494 char *start_arg = NULL;
9495 unsigned char op = 0;
9497 int internal_argval = 0; /* internal_argval is only useful if
9500 if (has_intervening_patws) {
9502 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
9504 while ( *RExC_parse && *RExC_parse != ')' ) {
9505 if ( *RExC_parse == ':' ) {
9506 start_arg = RExC_parse + 1;
9512 verb_len = RExC_parse - start_verb;
9515 while ( *RExC_parse && *RExC_parse != ')' )
9517 if ( *RExC_parse != ')' )
9518 vFAIL("Unterminated verb pattern argument");
9519 if ( RExC_parse == start_arg )
9522 if ( *RExC_parse != ')' )
9523 vFAIL("Unterminated verb pattern");
9526 switch ( *start_verb ) {
9527 case 'A': /* (*ACCEPT) */
9528 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9530 internal_argval = RExC_nestroot;
9533 case 'C': /* (*COMMIT) */
9534 if ( memEQs(start_verb,verb_len,"COMMIT") )
9537 case 'F': /* (*FAIL) */
9538 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9543 case ':': /* (*:NAME) */
9544 case 'M': /* (*MARK:NAME) */
9545 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9550 case 'P': /* (*PRUNE) */
9551 if ( memEQs(start_verb,verb_len,"PRUNE") )
9554 case 'S': /* (*SKIP) */
9555 if ( memEQs(start_verb,verb_len,"SKIP") )
9558 case 'T': /* (*THEN) */
9559 /* [19:06] <TimToady> :: is then */
9560 if ( memEQs(start_verb,verb_len,"THEN") ) {
9562 RExC_seen |= REG_CUTGROUP_SEEN;
9567 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9569 "Unknown verb pattern '%"UTF8f"'",
9570 UTF8fARG(UTF, verb_len, start_verb));
9573 if ( start_arg && internal_argval ) {
9574 vFAIL3("Verb pattern '%.*s' may not have an argument",
9575 verb_len, start_verb);
9576 } else if ( argok < 0 && !start_arg ) {
9577 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9578 verb_len, start_verb);
9580 ret = reganode(pRExC_state, op, internal_argval);
9581 if ( ! internal_argval && ! SIZE_ONLY ) {
9583 SV *sv = newSVpvn( start_arg,
9584 RExC_parse - start_arg);
9585 ARG(ret) = add_data( pRExC_state,
9587 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9594 if (!internal_argval)
9595 RExC_seen |= REG_VERBARG_SEEN;
9596 } else if ( start_arg ) {
9597 vFAIL3("Verb pattern '%.*s' may not have an argument",
9598 verb_len, start_verb);
9600 ret = reg_node(pRExC_state, op);
9602 nextchar(pRExC_state);
9605 else if (*RExC_parse == '?') { /* (?...) */
9606 bool is_logical = 0;
9607 const char * const seqstart = RExC_parse;
9608 if (has_intervening_patws) {
9610 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
9614 paren = *RExC_parse++;
9615 ret = NULL; /* For look-ahead/behind. */
9618 case 'P': /* (?P...) variants for those used to PCRE/Python */
9619 paren = *RExC_parse++;
9620 if ( paren == '<') /* (?P<...>) named capture */
9622 else if (paren == '>') { /* (?P>name) named recursion */
9623 goto named_recursion;
9625 else if (paren == '=') { /* (?P=...) named backref */
9626 /* this pretty much dupes the code for \k<NAME> in
9627 * regatom(), if you change this make sure you change that
9629 char* name_start = RExC_parse;
9631 SV *sv_dat = reg_scan_name(pRExC_state,
9632 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9633 if (RExC_parse == name_start || *RExC_parse != ')')
9634 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9635 vFAIL2("Sequence %.3s... not terminated",parse_start);
9638 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9639 RExC_rxi->data->data[num]=(void*)sv_dat;
9640 SvREFCNT_inc_simple_void(sv_dat);
9643 ret = reganode(pRExC_state,
9646 : (ASCII_FOLD_RESTRICTED)
9648 : (AT_LEAST_UNI_SEMANTICS)
9656 Set_Node_Offset(ret, parse_start+1);
9657 Set_Node_Cur_Length(ret, parse_start);
9659 nextchar(pRExC_state);
9663 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9664 vFAIL3("Sequence (%.*s...) not recognized",
9665 RExC_parse-seqstart, seqstart);
9667 case '<': /* (?<...) */
9668 if (*RExC_parse == '!')
9670 else if (*RExC_parse != '=')
9676 case '\'': /* (?'...') */
9677 name_start= RExC_parse;
9678 svname = reg_scan_name(pRExC_state,
9679 SIZE_ONLY /* reverse test from the others */
9680 ? REG_RSN_RETURN_NAME
9681 : REG_RSN_RETURN_NULL);
9682 if (RExC_parse == name_start || *RExC_parse != paren)
9683 vFAIL2("Sequence (?%c... not terminated",
9684 paren=='>' ? '<' : paren);
9688 if (!svname) /* shouldn't happen */
9690 "panic: reg_scan_name returned NULL");
9691 if (!RExC_paren_names) {
9692 RExC_paren_names= newHV();
9693 sv_2mortal(MUTABLE_SV(RExC_paren_names));
9695 RExC_paren_name_list= newAV();
9696 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
9699 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
9701 sv_dat = HeVAL(he_str);
9703 /* croak baby croak */
9705 "panic: paren_name hash element allocation failed");
9706 } else if ( SvPOK(sv_dat) ) {
9707 /* (?|...) can mean we have dupes so scan to check
9708 its already been stored. Maybe a flag indicating
9709 we are inside such a construct would be useful,
9710 but the arrays are likely to be quite small, so
9711 for now we punt -- dmq */
9712 IV count = SvIV(sv_dat);
9713 I32 *pv = (I32*)SvPVX(sv_dat);
9715 for ( i = 0 ; i < count ; i++ ) {
9716 if ( pv[i] == RExC_npar ) {
9722 pv = (I32*)SvGROW(sv_dat,
9723 SvCUR(sv_dat) + sizeof(I32)+1);
9724 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
9725 pv[count] = RExC_npar;
9726 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
9729 (void)SvUPGRADE(sv_dat,SVt_PVNV);
9730 sv_setpvn(sv_dat, (char *)&(RExC_npar),
9733 SvIV_set(sv_dat, 1);
9736 /* Yes this does cause a memory leak in debugging Perls
9738 if (!av_store(RExC_paren_name_list,
9739 RExC_npar, SvREFCNT_inc(svname)))
9740 SvREFCNT_dec_NN(svname);
9743 /*sv_dump(sv_dat);*/
9745 nextchar(pRExC_state);
9747 goto capturing_parens;
9749 RExC_seen |= REG_LOOKBEHIND_SEEN;
9750 RExC_in_lookbehind++;
9753 case '=': /* (?=...) */
9754 RExC_seen_zerolen++;
9756 case '!': /* (?!...) */
9757 RExC_seen_zerolen++;
9758 if (*RExC_parse == ')') {
9759 ret=reg_node(pRExC_state, OPFAIL);
9760 nextchar(pRExC_state);
9764 case '|': /* (?|...) */
9765 /* branch reset, behave like a (?:...) except that
9766 buffers in alternations share the same numbers */
9768 after_freeze = freeze_paren = RExC_npar;
9770 case ':': /* (?:...) */
9771 case '>': /* (?>...) */
9773 case '$': /* (?$...) */
9774 case '@': /* (?@...) */
9775 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
9777 case '0' : /* (?0) */
9778 case 'R' : /* (?R) */
9779 if (*RExC_parse != ')')
9780 FAIL("Sequence (?R) not terminated");
9781 ret = reg_node(pRExC_state, GOSTART);
9782 RExC_seen |= REG_GOSTART_SEEN;
9783 *flagp |= POSTPONED;
9784 nextchar(pRExC_state);
9787 /* named and numeric backreferences */
9788 case '&': /* (?&NAME) */
9789 parse_start = RExC_parse - 1;
9792 SV *sv_dat = reg_scan_name(pRExC_state,
9793 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9794 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9796 if (RExC_parse == RExC_end || *RExC_parse != ')')
9797 vFAIL("Sequence (?&... not terminated");
9798 goto gen_recurse_regop;
9799 assert(0); /* NOT REACHED */
9801 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9803 vFAIL("Illegal pattern");
9805 goto parse_recursion;
9807 case '-': /* (?-1) */
9808 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
9809 RExC_parse--; /* rewind to let it be handled later */
9813 case '1': case '2': case '3': case '4': /* (?1) */
9814 case '5': case '6': case '7': case '8': case '9':
9817 num = atoi(RExC_parse);
9818 parse_start = RExC_parse - 1; /* MJD */
9819 if (*RExC_parse == '-')
9821 while (isDIGIT(*RExC_parse))
9823 if (*RExC_parse!=')')
9824 vFAIL("Expecting close bracket");
9827 if ( paren == '-' ) {
9829 Diagram of capture buffer numbering.
9830 Top line is the normal capture buffer numbers
9831 Bottom line is the negative indexing as from
9835 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9839 num = RExC_npar + num;
9842 vFAIL("Reference to nonexistent group");
9844 } else if ( paren == '+' ) {
9845 num = RExC_npar + num - 1;
9848 ret = reganode(pRExC_state, GOSUB, num);
9850 if (num > (I32)RExC_rx->nparens) {
9852 vFAIL("Reference to nonexistent group");
9854 ARG2L_SET( ret, RExC_recurse_count++);
9856 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9857 "Recurse #%"UVuf" to %"IVdf"\n",
9858 (UV)ARG(ret), (IV)ARG2L(ret)));
9862 RExC_seen |= REG_RECURSE_SEEN;
9863 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9864 Set_Node_Offset(ret, parse_start); /* MJD */
9866 *flagp |= POSTPONED;
9867 nextchar(pRExC_state);
9870 assert(0); /* NOT REACHED */
9872 case '?': /* (??...) */
9874 if (*RExC_parse != '{') {
9876 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9878 "Sequence (%"UTF8f"...) not recognized",
9879 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9882 *flagp |= POSTPONED;
9883 paren = *RExC_parse++;
9885 case '{': /* (?{...}) */
9888 struct reg_code_block *cb;
9890 RExC_seen_zerolen++;
9892 if ( !pRExC_state->num_code_blocks
9893 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9894 || pRExC_state->code_blocks[pRExC_state->code_index].start
9895 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9898 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9899 FAIL("panic: Sequence (?{...}): no code block found\n");
9900 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9902 /* this is a pre-compiled code block (?{...}) */
9903 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9904 RExC_parse = RExC_start + cb->end;
9907 if (cb->src_regex) {
9908 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
9909 RExC_rxi->data->data[n] =
9910 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9911 RExC_rxi->data->data[n+1] = (void*)o;
9914 n = add_data(pRExC_state,
9915 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
9916 RExC_rxi->data->data[n] = (void*)o;
9919 pRExC_state->code_index++;
9920 nextchar(pRExC_state);
9924 ret = reg_node(pRExC_state, LOGICAL);
9925 eval = reganode(pRExC_state, EVAL, n);
9928 /* for later propagation into (??{}) return value */
9929 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9931 REGTAIL(pRExC_state, ret, eval);
9932 /* deal with the length of this later - MJD */
9935 ret = reganode(pRExC_state, EVAL, n);
9936 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9937 Set_Node_Offset(ret, parse_start);
9940 case '(': /* (?(?{...})...) and (?(?=...)...) */
9943 if (RExC_parse[0] == '?') { /* (?(?...)) */
9944 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9945 || RExC_parse[1] == '<'
9946 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9950 ret = reg_node(pRExC_state, LOGICAL);
9954 tail = reg(pRExC_state, 1, &flag, depth+1);
9955 if (flag & RESTART_UTF8) {
9956 *flagp = RESTART_UTF8;
9959 REGTAIL(pRExC_state, ret, tail);
9963 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9964 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9966 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9967 char *name_start= RExC_parse++;
9969 SV *sv_dat=reg_scan_name(pRExC_state,
9970 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9971 if (RExC_parse == name_start || *RExC_parse != ch)
9972 vFAIL2("Sequence (?(%c... not terminated",
9973 (ch == '>' ? '<' : ch));
9976 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9977 RExC_rxi->data->data[num]=(void*)sv_dat;
9978 SvREFCNT_inc_simple_void(sv_dat);
9980 ret = reganode(pRExC_state,NGROUPP,num);
9981 goto insert_if_check_paren;
9983 else if (RExC_parse[0] == 'D' &&
9984 RExC_parse[1] == 'E' &&
9985 RExC_parse[2] == 'F' &&
9986 RExC_parse[3] == 'I' &&
9987 RExC_parse[4] == 'N' &&
9988 RExC_parse[5] == 'E')
9990 ret = reganode(pRExC_state,DEFINEP,0);
9993 goto insert_if_check_paren;
9995 else if (RExC_parse[0] == 'R') {
9998 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9999 parno = atoi(RExC_parse++);
10000 while (isDIGIT(*RExC_parse))
10002 } else if (RExC_parse[0] == '&') {
10005 sv_dat = reg_scan_name(pRExC_state,
10007 ? REG_RSN_RETURN_NULL
10008 : REG_RSN_RETURN_DATA);
10009 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10011 ret = reganode(pRExC_state,INSUBP,parno);
10012 goto insert_if_check_paren;
10014 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10018 parno = atoi(RExC_parse++);
10020 while (isDIGIT(*RExC_parse))
10022 ret = reganode(pRExC_state, GROUPP, parno);
10024 insert_if_check_paren:
10025 if (*(tmp = nextchar(pRExC_state)) != ')') {
10026 /* nextchar also skips comments, so undo its work
10027 * and skip over the the next character.
10030 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10031 vFAIL("Switch condition not recognized");
10034 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
10035 br = regbranch(pRExC_state, &flags, 1,depth+1);
10037 if (flags & RESTART_UTF8) {
10038 *flagp = RESTART_UTF8;
10041 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10044 REGTAIL(pRExC_state, br, reganode(pRExC_state,
10046 c = *nextchar(pRExC_state);
10047 if (flags&HASWIDTH)
10048 *flagp |= HASWIDTH;
10051 vFAIL("(?(DEFINE)....) does not allow branches");
10053 /* Fake one for optimizer. */
10054 lastbr = reganode(pRExC_state, IFTHEN, 0);
10056 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
10057 if (flags & RESTART_UTF8) {
10058 *flagp = RESTART_UTF8;
10061 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10064 REGTAIL(pRExC_state, ret, lastbr);
10065 if (flags&HASWIDTH)
10066 *flagp |= HASWIDTH;
10067 c = *nextchar(pRExC_state);
10072 vFAIL("Switch (?(condition)... contains too many branches");
10073 ender = reg_node(pRExC_state, TAIL);
10074 REGTAIL(pRExC_state, br, ender);
10076 REGTAIL(pRExC_state, lastbr, ender);
10077 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10080 REGTAIL(pRExC_state, ret, ender);
10081 RExC_size++; /* XXX WHY do we need this?!!
10082 For large programs it seems to be required
10083 but I can't figure out why. -- dmq*/
10087 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10088 vFAIL("Unknown switch condition (?(...))");
10091 case '[': /* (?[ ... ]) */
10092 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10095 RExC_parse--; /* for vFAIL to print correctly */
10096 vFAIL("Sequence (? incomplete");
10098 default: /* e.g., (?i) */
10101 parse_lparen_question_flags(pRExC_state);
10102 if (UCHARAT(RExC_parse) != ':') {
10103 nextchar(pRExC_state);
10108 nextchar(pRExC_state);
10118 ret = reganode(pRExC_state, OPEN, parno);
10120 if (!RExC_nestroot)
10121 RExC_nestroot = parno;
10122 if (RExC_seen & REG_RECURSE_SEEN
10123 && !RExC_open_parens[parno-1])
10125 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10126 "Setting open paren #%"IVdf" to %d\n",
10127 (IV)parno, REG_NODE_NUM(ret)));
10128 RExC_open_parens[parno-1]= ret;
10131 Set_Node_Length(ret, 1); /* MJD */
10132 Set_Node_Offset(ret, RExC_parse); /* MJD */
10140 /* Pick up the branches, linking them together. */
10141 parse_start = RExC_parse; /* MJD */
10142 br = regbranch(pRExC_state, &flags, 1,depth+1);
10144 /* branch_len = (paren != 0); */
10147 if (flags & RESTART_UTF8) {
10148 *flagp = RESTART_UTF8;
10151 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10153 if (*RExC_parse == '|') {
10154 if (!SIZE_ONLY && RExC_extralen) {
10155 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10158 reginsert(pRExC_state, BRANCH, br, depth+1);
10159 Set_Node_Length(br, paren != 0);
10160 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10164 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10166 else if (paren == ':') {
10167 *flagp |= flags&SIMPLE;
10169 if (is_open) { /* Starts with OPEN. */
10170 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10172 else if (paren != '?') /* Not Conditional */
10174 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10176 while (*RExC_parse == '|') {
10177 if (!SIZE_ONLY && RExC_extralen) {
10178 ender = reganode(pRExC_state, LONGJMP,0);
10180 /* Append to the previous. */
10181 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10184 RExC_extralen += 2; /* Account for LONGJMP. */
10185 nextchar(pRExC_state);
10186 if (freeze_paren) {
10187 if (RExC_npar > after_freeze)
10188 after_freeze = RExC_npar;
10189 RExC_npar = freeze_paren;
10191 br = regbranch(pRExC_state, &flags, 0, depth+1);
10194 if (flags & RESTART_UTF8) {
10195 *flagp = RESTART_UTF8;
10198 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10200 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10202 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10205 if (have_branch || paren != ':') {
10206 /* Make a closing node, and hook it on the end. */
10209 ender = reg_node(pRExC_state, TAIL);
10212 ender = reganode(pRExC_state, CLOSE, parno);
10213 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10214 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10215 "Setting close paren #%"IVdf" to %d\n",
10216 (IV)parno, REG_NODE_NUM(ender)));
10217 RExC_close_parens[parno-1]= ender;
10218 if (RExC_nestroot == parno)
10221 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10222 Set_Node_Length(ender,1); /* MJD */
10228 *flagp &= ~HASWIDTH;
10231 ender = reg_node(pRExC_state, SUCCEED);
10234 ender = reg_node(pRExC_state, END);
10236 assert(!RExC_opend); /* there can only be one! */
10237 RExC_opend = ender;
10241 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10242 SV * const mysv_val1=sv_newmortal();
10243 SV * const mysv_val2=sv_newmortal();
10244 DEBUG_PARSE_MSG("lsbr");
10245 regprop(RExC_rx, mysv_val1, lastbr, NULL);
10246 regprop(RExC_rx, mysv_val2, ender, NULL);
10247 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10248 SvPV_nolen_const(mysv_val1),
10249 (IV)REG_NODE_NUM(lastbr),
10250 SvPV_nolen_const(mysv_val2),
10251 (IV)REG_NODE_NUM(ender),
10252 (IV)(ender - lastbr)
10255 REGTAIL(pRExC_state, lastbr, ender);
10257 if (have_branch && !SIZE_ONLY) {
10258 char is_nothing= 1;
10260 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10262 /* Hook the tails of the branches to the closing node. */
10263 for (br = ret; br; br = regnext(br)) {
10264 const U8 op = PL_regkind[OP(br)];
10265 if (op == BRANCH) {
10266 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10267 if ( OP(NEXTOPER(br)) != NOTHING
10268 || regnext(NEXTOPER(br)) != ender)
10271 else if (op == BRANCHJ) {
10272 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10273 /* for now we always disable this optimisation * /
10274 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10275 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10281 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10282 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10283 SV * const mysv_val1=sv_newmortal();
10284 SV * const mysv_val2=sv_newmortal();
10285 DEBUG_PARSE_MSG("NADA");
10286 regprop(RExC_rx, mysv_val1, ret, NULL);
10287 regprop(RExC_rx, mysv_val2, ender, NULL);
10288 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10289 SvPV_nolen_const(mysv_val1),
10290 (IV)REG_NODE_NUM(ret),
10291 SvPV_nolen_const(mysv_val2),
10292 (IV)REG_NODE_NUM(ender),
10297 if (OP(ender) == TAIL) {
10302 for ( opt= br + 1; opt < ender ; opt++ )
10303 OP(opt)= OPTIMIZED;
10304 NEXT_OFF(br)= ender - br;
10312 static const char parens[] = "=!<,>";
10314 if (paren && (p = strchr(parens, paren))) {
10315 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10316 int flag = (p - parens) > 1;
10319 node = SUSPEND, flag = 0;
10320 reginsert(pRExC_state, node,ret, depth+1);
10321 Set_Node_Cur_Length(ret, parse_start);
10322 Set_Node_Offset(ret, parse_start + 1);
10324 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10328 /* Check for proper termination. */
10330 /* restore original flags, but keep (?p) */
10331 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10332 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10333 RExC_parse = oregcomp_parse;
10334 vFAIL("Unmatched (");
10337 else if (!paren && RExC_parse < RExC_end) {
10338 if (*RExC_parse == ')') {
10340 vFAIL("Unmatched )");
10343 FAIL("Junk on end of regexp"); /* "Can't happen". */
10344 assert(0); /* NOTREACHED */
10347 if (RExC_in_lookbehind) {
10348 RExC_in_lookbehind--;
10350 if (after_freeze > RExC_npar)
10351 RExC_npar = after_freeze;
10356 - regbranch - one alternative of an | operator
10358 * Implements the concatenation operator.
10360 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10364 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10367 regnode *chain = NULL;
10369 I32 flags = 0, c = 0;
10370 GET_RE_DEBUG_FLAGS_DECL;
10372 PERL_ARGS_ASSERT_REGBRANCH;
10374 DEBUG_PARSE("brnc");
10379 if (!SIZE_ONLY && RExC_extralen)
10380 ret = reganode(pRExC_state, BRANCHJ,0);
10382 ret = reg_node(pRExC_state, BRANCH);
10383 Set_Node_Length(ret, 1);
10387 if (!first && SIZE_ONLY)
10388 RExC_extralen += 1; /* BRANCHJ */
10390 *flagp = WORST; /* Tentatively. */
10393 nextchar(pRExC_state);
10394 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10395 flags &= ~TRYAGAIN;
10396 latest = regpiece(pRExC_state, &flags,depth+1);
10397 if (latest == NULL) {
10398 if (flags & TRYAGAIN)
10400 if (flags & RESTART_UTF8) {
10401 *flagp = RESTART_UTF8;
10404 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10406 else if (ret == NULL)
10408 *flagp |= flags&(HASWIDTH|POSTPONED);
10409 if (chain == NULL) /* First piece. */
10410 *flagp |= flags&SPSTART;
10413 REGTAIL(pRExC_state, chain, latest);
10418 if (chain == NULL) { /* Loop ran zero times. */
10419 chain = reg_node(pRExC_state, NOTHING);
10424 *flagp |= flags&SIMPLE;
10431 - regpiece - something followed by possible [*+?]
10433 * Note that the branching code sequences used for ? and the general cases
10434 * of * and + are somewhat optimized: they use the same NOTHING node as
10435 * both the endmarker for their branch list and the body of the last branch.
10436 * It might seem that this node could be dispensed with entirely, but the
10437 * endmarker role is not redundant.
10439 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10441 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10445 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10451 const char * const origparse = RExC_parse;
10453 I32 max = REG_INFTY;
10454 #ifdef RE_TRACK_PATTERN_OFFSETS
10457 const char *maxpos = NULL;
10459 /* Save the original in case we change the emitted regop to a FAIL. */
10460 regnode * const orig_emit = RExC_emit;
10462 GET_RE_DEBUG_FLAGS_DECL;
10464 PERL_ARGS_ASSERT_REGPIECE;
10466 DEBUG_PARSE("piec");
10468 ret = regatom(pRExC_state, &flags,depth+1);
10470 if (flags & (TRYAGAIN|RESTART_UTF8))
10471 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10473 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10479 if (op == '{' && regcurly(RExC_parse)) {
10481 #ifdef RE_TRACK_PATTERN_OFFSETS
10482 parse_start = RExC_parse; /* MJD */
10484 next = RExC_parse + 1;
10485 while (isDIGIT(*next) || *next == ',') {
10486 if (*next == ',') {
10494 if (*next == '}') { /* got one */
10498 min = atoi(RExC_parse);
10499 if (*maxpos == ',')
10502 maxpos = RExC_parse;
10503 max = atoi(maxpos);
10504 if (!max && *maxpos != '0')
10505 max = REG_INFTY; /* meaning "infinity" */
10506 else if (max >= REG_INFTY)
10507 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10509 nextchar(pRExC_state);
10510 if (max < min) { /* If can't match, warn and optimize to fail
10513 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10515 /* We can't back off the size because we have to reserve
10516 * enough space for all the things we are about to throw
10517 * away, but we can shrink it by the ammount we are about
10518 * to re-use here */
10519 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10522 RExC_emit = orig_emit;
10524 ret = reg_node(pRExC_state, OPFAIL);
10527 else if (min == max
10528 && RExC_parse < RExC_end
10529 && (*RExC_parse == '?' || *RExC_parse == '+'))
10532 ckWARN2reg(RExC_parse + 1,
10533 "Useless use of greediness modifier '%c'",
10536 /* Absorb the modifier, so later code doesn't see nor use
10538 nextchar(pRExC_state);
10542 if ((flags&SIMPLE)) {
10543 RExC_naughty += 2 + RExC_naughty / 2;
10544 reginsert(pRExC_state, CURLY, ret, depth+1);
10545 Set_Node_Offset(ret, parse_start+1); /* MJD */
10546 Set_Node_Cur_Length(ret, parse_start);
10549 regnode * const w = reg_node(pRExC_state, WHILEM);
10552 REGTAIL(pRExC_state, ret, w);
10553 if (!SIZE_ONLY && RExC_extralen) {
10554 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10555 reginsert(pRExC_state, NOTHING,ret, depth+1);
10556 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10558 reginsert(pRExC_state, CURLYX,ret, depth+1);
10560 Set_Node_Offset(ret, parse_start+1);
10561 Set_Node_Length(ret,
10562 op == '{' ? (RExC_parse - parse_start) : 1);
10564 if (!SIZE_ONLY && RExC_extralen)
10565 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10566 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10568 RExC_whilem_seen++, RExC_extralen += 3;
10569 RExC_naughty += 4 + RExC_naughty; /* compound interest */
10576 *flagp |= HASWIDTH;
10578 ARG1_SET(ret, (U16)min);
10579 ARG2_SET(ret, (U16)max);
10581 if (max == REG_INFTY)
10582 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10588 if (!ISMULT1(op)) {
10593 #if 0 /* Now runtime fix should be reliable. */
10595 /* if this is reinstated, don't forget to put this back into perldiag:
10597 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10599 (F) The part of the regexp subject to either the * or + quantifier
10600 could match an empty string. The {#} shows in the regular
10601 expression about where the problem was discovered.
10605 if (!(flags&HASWIDTH) && op != '?')
10606 vFAIL("Regexp *+ operand could be empty");
10609 #ifdef RE_TRACK_PATTERN_OFFSETS
10610 parse_start = RExC_parse;
10612 nextchar(pRExC_state);
10614 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10616 if (op == '*' && (flags&SIMPLE)) {
10617 reginsert(pRExC_state, STAR, ret, depth+1);
10620 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10622 else if (op == '*') {
10626 else if (op == '+' && (flags&SIMPLE)) {
10627 reginsert(pRExC_state, PLUS, ret, depth+1);
10630 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10632 else if (op == '+') {
10636 else if (op == '?') {
10641 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10642 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10643 ckWARN2reg(RExC_parse,
10644 "%"UTF8f" matches null string many times",
10645 UTF8fARG(UTF, (RExC_parse >= origparse
10646 ? RExC_parse - origparse
10649 (void)ReREFCNT_inc(RExC_rx_sv);
10652 if (RExC_parse < RExC_end && *RExC_parse == '?') {
10653 nextchar(pRExC_state);
10654 reginsert(pRExC_state, MINMOD, ret, depth+1);
10655 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
10658 if (RExC_parse < RExC_end && *RExC_parse == '+') {
10660 nextchar(pRExC_state);
10661 ender = reg_node(pRExC_state, SUCCEED);
10662 REGTAIL(pRExC_state, ret, ender);
10663 reginsert(pRExC_state, SUSPEND, ret, depth+1);
10665 ender = reg_node(pRExC_state, TAIL);
10666 REGTAIL(pRExC_state, ret, ender);
10669 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
10671 vFAIL("Nested quantifiers");
10678 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p,
10679 UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
10680 const bool strict /* Apply stricter parsing rules? */
10684 /* This is expected to be called by a parser routine that has recognized '\N'
10685 and needs to handle the rest. RExC_parse is expected to point at the first
10686 char following the N at the time of the call. On successful return,
10687 RExC_parse has been updated to point to just after the sequence identified
10688 by this routine, and <*flagp> has been updated.
10690 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
10693 \N may begin either a named sequence, or if outside a character class, mean
10694 to match a non-newline. For non single-quoted regexes, the tokenizer has
10695 attempted to decide which, and in the case of a named sequence, converted it
10696 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
10697 where c1... are the characters in the sequence. For single-quoted regexes,
10698 the tokenizer passes the \N sequence through unchanged; this code will not
10699 attempt to determine this nor expand those, instead raising a syntax error.
10700 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
10701 or there is no '}', it signals that this \N occurrence means to match a
10704 Only the \N{U+...} form should occur in a character class, for the same
10705 reason that '.' inside a character class means to just match a period: it
10706 just doesn't make sense.
10708 The function raises an error (via vFAIL), and doesn't return for various
10709 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
10710 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
10711 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
10712 only possible if node_p is non-NULL.
10715 If <valuep> is non-null, it means the caller can accept an input sequence
10716 consisting of a just a single code point; <*valuep> is set to that value
10717 if the input is such.
10719 If <node_p> is non-null it signifies that the caller can accept any other
10720 legal sequence (i.e., one that isn't just a single code point). <*node_p>
10722 1) \N means not-a-NL: points to a newly created REG_ANY node;
10723 2) \N{}: points to a new NOTHING node;
10724 3) otherwise: points to a new EXACT node containing the resolved
10726 Note that FALSE is returned for single code point sequences if <valuep> is
10730 char * endbrace; /* '}' following the name */
10732 char *endchar; /* Points to '.' or '}' ending cur char in the input
10734 bool has_multiple_chars; /* true if the input stream contains a sequence of
10735 more than one character */
10737 GET_RE_DEBUG_FLAGS_DECL;
10739 PERL_ARGS_ASSERT_GROK_BSLASH_N;
10741 GET_RE_DEBUG_FLAGS;
10743 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
10745 /* The [^\n] meaning of \N ignores spaces and comments under the /x
10746 * modifier. The other meaning does not, so use a temporary until we find
10747 * out which we are being called with */
10748 p = (RExC_flags & RXf_PMf_EXTENDED)
10749 ? regpatws(pRExC_state, RExC_parse,
10750 TRUE) /* means recognize comments */
10753 /* Disambiguate between \N meaning a named character versus \N meaning
10754 * [^\n]. The former is assumed when it can't be the latter. */
10755 if (*p != '{' || regcurly(p)) {
10758 /* no bare \N allowed in a charclass */
10759 if (in_char_class) {
10760 vFAIL("\\N in a character class must be a named character: \\N{...}");
10764 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
10766 nextchar(pRExC_state);
10767 *node_p = reg_node(pRExC_state, REG_ANY);
10768 *flagp |= HASWIDTH|SIMPLE;
10770 Set_Node_Length(*node_p, 1); /* MJD */
10774 /* Here, we have decided it should be a named character or sequence */
10776 /* The test above made sure that the next real character is a '{', but
10777 * under the /x modifier, it could be separated by space (or a comment and
10778 * \n) and this is not allowed (for consistency with \x{...} and the
10779 * tokenizer handling of \N{NAME}). */
10780 if (*RExC_parse != '{') {
10781 vFAIL("Missing braces on \\N{}");
10784 RExC_parse++; /* Skip past the '{' */
10786 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
10787 || ! (endbrace == RExC_parse /* nothing between the {} */
10788 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below
10790 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg)
10793 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
10794 vFAIL("\\N{NAME} must be resolved by the lexer");
10797 if (endbrace == RExC_parse) { /* empty: \N{} */
10800 *node_p = reg_node(pRExC_state,NOTHING);
10802 else if (in_char_class) {
10803 if (SIZE_ONLY && in_char_class) {
10805 RExC_parse++; /* Position after the "}" */
10806 vFAIL("Zero length \\N{}");
10809 ckWARNreg(RExC_parse,
10810 "Ignoring zero length \\N{} in character class");
10818 nextchar(pRExC_state);
10822 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
10823 RExC_parse += 2; /* Skip past the 'U+' */
10825 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10827 /* Code points are separated by dots. If none, there is only one code
10828 * point, and is terminated by the brace */
10829 has_multiple_chars = (endchar < endbrace);
10831 if (valuep && (! has_multiple_chars || in_char_class)) {
10832 /* We only pay attention to the first char of
10833 multichar strings being returned in char classes. I kinda wonder
10834 if this makes sense as it does change the behaviour
10835 from earlier versions, OTOH that behaviour was broken
10836 as well. XXX Solution is to recharacterize as
10837 [rest-of-class]|multi1|multi2... */
10839 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
10840 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
10841 | PERL_SCAN_DISALLOW_PREFIX
10842 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
10844 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
10846 /* The tokenizer should have guaranteed validity, but it's possible to
10847 * bypass it by using single quoting, so check */
10848 if (length_of_hex == 0
10849 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
10851 RExC_parse += length_of_hex; /* Includes all the valid */
10852 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
10853 ? UTF8SKIP(RExC_parse)
10855 /* Guard against malformed utf8 */
10856 if (RExC_parse >= endchar) {
10857 RExC_parse = endchar;
10859 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10862 if (in_char_class && has_multiple_chars) {
10864 RExC_parse = endbrace;
10865 vFAIL("\\N{} in character class restricted to one character");
10868 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10872 RExC_parse = endbrace + 1;
10874 else if (! node_p || ! has_multiple_chars) {
10876 /* Here, the input is legal, but not according to the caller's
10877 * options. We fail without advancing the parse, so that the
10878 * caller can try again */
10884 /* What is done here is to convert this to a sub-pattern of the form
10885 * (?:\x{char1}\x{char2}...)
10886 * and then call reg recursively. That way, it retains its atomicness,
10887 * while not having to worry about special handling that some code
10888 * points may have. toke.c has converted the original Unicode values
10889 * to native, so that we can just pass on the hex values unchanged. We
10890 * do have to set a flag to keep recoding from happening in the
10893 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10895 char *orig_end = RExC_end;
10898 while (RExC_parse < endbrace) {
10900 /* Convert to notation the rest of the code understands */
10901 sv_catpv(substitute_parse, "\\x{");
10902 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10903 sv_catpv(substitute_parse, "}");
10905 /* Point to the beginning of the next character in the sequence. */
10906 RExC_parse = endchar + 1;
10907 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10909 sv_catpv(substitute_parse, ")");
10911 RExC_parse = SvPV(substitute_parse, len);
10913 /* Don't allow empty number */
10915 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10917 RExC_end = RExC_parse + len;
10919 /* The values are Unicode, and therefore not subject to recoding */
10920 RExC_override_recoding = 1;
10922 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10923 if (flags & RESTART_UTF8) {
10924 *flagp = RESTART_UTF8;
10927 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10930 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10932 RExC_parse = endbrace;
10933 RExC_end = orig_end;
10934 RExC_override_recoding = 0;
10936 nextchar(pRExC_state);
10946 * It returns the code point in utf8 for the value in *encp.
10947 * value: a code value in the source encoding
10948 * encp: a pointer to an Encode object
10950 * If the result from Encode is not a single character,
10951 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10954 S_reg_recode(pTHX_ const char value, SV **encp)
10957 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10958 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10959 const STRLEN newlen = SvCUR(sv);
10960 UV uv = UNICODE_REPLACEMENT;
10962 PERL_ARGS_ASSERT_REG_RECODE;
10966 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10969 if (!newlen || numlen != newlen) {
10970 uv = UNICODE_REPLACEMENT;
10976 PERL_STATIC_INLINE U8
10977 S_compute_EXACTish(RExC_state_t *pRExC_state)
10981 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10987 op = get_regex_charset(RExC_flags);
10988 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10989 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10990 been, so there is no hole */
10993 return op + EXACTF;
10996 PERL_STATIC_INLINE void
10997 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
10998 regnode *node, I32* flagp, STRLEN len, UV code_point,
11001 /* This knows the details about sizing an EXACTish node, setting flags for
11002 * it (by setting <*flagp>, and potentially populating it with a single
11005 * If <len> (the length in bytes) is non-zero, this function assumes that
11006 * the node has already been populated, and just does the sizing. In this
11007 * case <code_point> should be the final code point that has already been
11008 * placed into the node. This value will be ignored except that under some
11009 * circumstances <*flagp> is set based on it.
11011 * If <len> is zero, the function assumes that the node is to contain only
11012 * the single character given by <code_point> and calculates what <len>
11013 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
11014 * additionally will populate the node's STRING with <code_point> or its
11017 * In both cases <*flagp> is appropriately set
11019 * It knows that under FOLD, the Latin Sharp S and UTF characters above
11020 * 255, must be folded (the former only when the rules indicate it can
11023 * When it does the populating, it looks at the flag 'downgradable'. If
11024 * true with a node that folds, it checks if the single code point
11025 * participates in a fold, and if not downgrades the node to an EXACT.
11026 * This helps the optimizer */
11028 bool len_passed_in = cBOOL(len != 0);
11029 U8 character[UTF8_MAXBYTES_CASE+1];
11031 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
11033 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
11034 * sizing difference, and is extra work that is thrown away */
11035 if (downgradable && ! PASS2) {
11036 downgradable = FALSE;
11039 if (! len_passed_in) {
11041 if (UNI_IS_INVARIANT(code_point)) {
11042 if (LOC || ! FOLD) { /* /l defers folding until runtime */
11043 *character = (U8) code_point;
11045 else { /* Here is /i and not /l (toFOLD() is defined on just
11046 ASCII, which isn't the same thing as INVARIANT on
11047 EBCDIC, but it works there, as the extra invariants
11048 fold to themselves) */
11049 *character = toFOLD((U8) code_point);
11051 && *character == code_point
11052 && ! HAS_NONLATIN1_FOLD_CLOSURE(code_point))
11059 else if (FOLD && (! LOC
11060 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
11061 { /* Folding, and ok to do so now */
11062 UV folded = _to_uni_fold_flags(
11066 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11067 ? FOLD_FLAGS_NOMIX_ASCII
11070 && folded == code_point
11071 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11076 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11078 /* Not folding this cp, and can output it directly */
11079 *character = UTF8_TWO_BYTE_HI(code_point);
11080 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11084 uvchr_to_utf8( character, code_point);
11085 len = UTF8SKIP(character);
11087 } /* Else pattern isn't UTF8. */
11089 *character = (U8) code_point;
11091 } /* Else is folded non-UTF8 */
11092 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11094 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11095 * comments at join_exact()); */
11096 *character = (U8) code_point;
11099 /* Can turn into an EXACT node if we know the fold at compile time,
11100 * and it folds to itself and doesn't particpate in other folds */
11103 && PL_fold_latin1[code_point] == code_point
11104 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11105 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11109 } /* else is Sharp s. May need to fold it */
11110 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11112 *(character + 1) = 's';
11116 *character = LATIN_SMALL_LETTER_SHARP_S;
11122 RExC_size += STR_SZ(len);
11125 RExC_emit += STR_SZ(len);
11126 STR_LEN(node) = len;
11127 if (! len_passed_in) {
11128 Copy((char *) character, STRING(node), len, char);
11132 *flagp |= HASWIDTH;
11134 /* A single character node is SIMPLE, except for the special-cased SHARP S
11136 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11137 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11138 || ! FOLD || ! DEPENDS_SEMANTICS))
11143 /* The OP may not be well defined in PASS1 */
11144 if (PASS2 && OP(node) == EXACTFL) {
11145 RExC_contains_locale = 1;
11150 /* return atoi(p), unless it's too big to sensibly be a backref,
11151 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11154 S_backref_value(char *p)
11158 for (;isDIGIT(*q); q++) {} /* calculate length of num */
11159 if (q - p == 0 || q - p > 9)
11166 - regatom - the lowest level
11168 Try to identify anything special at the start of the pattern. If there
11169 is, then handle it as required. This may involve generating a single regop,
11170 such as for an assertion; or it may involve recursing, such as to
11171 handle a () structure.
11173 If the string doesn't start with something special then we gobble up
11174 as much literal text as we can.
11176 Once we have been able to handle whatever type of thing started the
11177 sequence, we return.
11179 Note: we have to be careful with escapes, as they can be both literal
11180 and special, and in the case of \10 and friends, context determines which.
11182 A summary of the code structure is:
11184 switch (first_byte) {
11185 cases for each special:
11186 handle this special;
11189 switch (2nd byte) {
11190 cases for each unambiguous special:
11191 handle this special;
11193 cases for each ambigous special/literal:
11195 if (special) handle here
11197 default: // unambiguously literal:
11200 default: // is a literal char
11203 create EXACTish node for literal;
11204 while (more input and node isn't full) {
11205 switch (input_byte) {
11206 cases for each special;
11207 make sure parse pointer is set so that the next call to
11208 regatom will see this special first
11209 goto loopdone; // EXACTish node terminated by prev. char
11211 append char to EXACTISH node;
11213 get next input byte;
11217 return the generated node;
11219 Specifically there are two separate switches for handling
11220 escape sequences, with the one for handling literal escapes requiring
11221 a dummy entry for all of the special escapes that are actually handled
11224 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11226 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11228 Otherwise does not return NULL.
11232 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11234 regnode *ret = NULL;
11236 char *parse_start = RExC_parse;
11241 GET_RE_DEBUG_FLAGS_DECL;
11243 *flagp = WORST; /* Tentatively. */
11245 DEBUG_PARSE("atom");
11247 PERL_ARGS_ASSERT_REGATOM;
11250 switch ((U8)*RExC_parse) {
11252 RExC_seen_zerolen++;
11253 nextchar(pRExC_state);
11254 if (RExC_flags & RXf_PMf_MULTILINE)
11255 ret = reg_node(pRExC_state, MBOL);
11256 else if (RExC_flags & RXf_PMf_SINGLELINE)
11257 ret = reg_node(pRExC_state, SBOL);
11259 ret = reg_node(pRExC_state, BOL);
11260 Set_Node_Length(ret, 1); /* MJD */
11263 nextchar(pRExC_state);
11265 RExC_seen_zerolen++;
11266 if (RExC_flags & RXf_PMf_MULTILINE)
11267 ret = reg_node(pRExC_state, MEOL);
11268 else if (RExC_flags & RXf_PMf_SINGLELINE)
11269 ret = reg_node(pRExC_state, SEOL);
11271 ret = reg_node(pRExC_state, EOL);
11272 Set_Node_Length(ret, 1); /* MJD */
11275 nextchar(pRExC_state);
11276 if (RExC_flags & RXf_PMf_SINGLELINE)
11277 ret = reg_node(pRExC_state, SANY);
11279 ret = reg_node(pRExC_state, REG_ANY);
11280 *flagp |= HASWIDTH|SIMPLE;
11282 Set_Node_Length(ret, 1); /* MJD */
11286 char * const oregcomp_parse = ++RExC_parse;
11287 ret = regclass(pRExC_state, flagp,depth+1,
11288 FALSE, /* means parse the whole char class */
11289 TRUE, /* allow multi-char folds */
11290 FALSE, /* don't silence non-portable warnings. */
11292 if (*RExC_parse != ']') {
11293 RExC_parse = oregcomp_parse;
11294 vFAIL("Unmatched [");
11297 if (*flagp & RESTART_UTF8)
11299 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11302 nextchar(pRExC_state);
11303 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11307 nextchar(pRExC_state);
11308 ret = reg(pRExC_state, 2, &flags,depth+1);
11310 if (flags & TRYAGAIN) {
11311 if (RExC_parse == RExC_end) {
11312 /* Make parent create an empty node if needed. */
11313 *flagp |= TRYAGAIN;
11318 if (flags & RESTART_UTF8) {
11319 *flagp = RESTART_UTF8;
11322 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11325 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11329 if (flags & TRYAGAIN) {
11330 *flagp |= TRYAGAIN;
11333 vFAIL("Internal urp");
11334 /* Supposed to be caught earlier. */
11340 vFAIL("Quantifier follows nothing");
11345 This switch handles escape sequences that resolve to some kind
11346 of special regop and not to literal text. Escape sequnces that
11347 resolve to literal text are handled below in the switch marked
11350 Every entry in this switch *must* have a corresponding entry
11351 in the literal escape switch. However, the opposite is not
11352 required, as the default for this switch is to jump to the
11353 literal text handling code.
11355 switch ((U8)*++RExC_parse) {
11356 /* Special Escapes */
11358 RExC_seen_zerolen++;
11359 ret = reg_node(pRExC_state, SBOL);
11361 goto finish_meta_pat;
11363 ret = reg_node(pRExC_state, GPOS);
11364 RExC_seen |= REG_GPOS_SEEN;
11366 goto finish_meta_pat;
11368 RExC_seen_zerolen++;
11369 ret = reg_node(pRExC_state, KEEPS);
11371 /* XXX:dmq : disabling in-place substitution seems to
11372 * be necessary here to avoid cases of memory corruption, as
11373 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11375 RExC_seen |= REG_LOOKBEHIND_SEEN;
11376 goto finish_meta_pat;
11378 ret = reg_node(pRExC_state, SEOL);
11380 RExC_seen_zerolen++; /* Do not optimize RE away */
11381 goto finish_meta_pat;
11383 ret = reg_node(pRExC_state, EOS);
11385 RExC_seen_zerolen++; /* Do not optimize RE away */
11386 goto finish_meta_pat;
11388 ret = reg_node(pRExC_state, CANY);
11389 RExC_seen |= REG_CANY_SEEN;
11390 *flagp |= HASWIDTH|SIMPLE;
11392 ckWARNdep(RExC_parse+1, "\\C is deprecated");
11394 goto finish_meta_pat;
11396 ret = reg_node(pRExC_state, CLUMP);
11397 *flagp |= HASWIDTH;
11398 goto finish_meta_pat;
11404 arg = ANYOF_WORDCHAR;
11408 RExC_seen_zerolen++;
11409 RExC_seen |= REG_LOOKBEHIND_SEEN;
11410 op = BOUND + get_regex_charset(RExC_flags);
11411 if (op > BOUNDA) { /* /aa is same as /a */
11414 else if (op == BOUNDL) {
11415 RExC_contains_locale = 1;
11417 ret = reg_node(pRExC_state, op);
11418 FLAGS(ret) = get_regex_charset(RExC_flags);
11420 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11421 /* diag_listed_as: Use "%s" instead of "%s" */
11422 vFAIL("Use \"\\b\\{\" instead of \"\\b{\"");
11424 goto finish_meta_pat;
11426 RExC_seen_zerolen++;
11427 RExC_seen |= REG_LOOKBEHIND_SEEN;
11428 op = NBOUND + get_regex_charset(RExC_flags);
11429 if (op > NBOUNDA) { /* /aa is same as /a */
11432 else if (op == NBOUNDL) {
11433 RExC_contains_locale = 1;
11435 ret = reg_node(pRExC_state, op);
11436 FLAGS(ret) = get_regex_charset(RExC_flags);
11438 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
11439 /* diag_listed_as: Use "%s" instead of "%s" */
11440 vFAIL("Use \"\\B\\{\" instead of \"\\B{\"");
11442 goto finish_meta_pat;
11452 ret = reg_node(pRExC_state, LNBREAK);
11453 *flagp |= HASWIDTH|SIMPLE;
11454 goto finish_meta_pat;
11462 goto join_posix_op_known;
11468 arg = ANYOF_VERTWS;
11470 goto join_posix_op_known;
11480 op = POSIXD + get_regex_charset(RExC_flags);
11481 if (op > POSIXA) { /* /aa is same as /a */
11484 else if (op == POSIXL) {
11485 RExC_contains_locale = 1;
11488 join_posix_op_known:
11491 op += NPOSIXD - POSIXD;
11494 ret = reg_node(pRExC_state, op);
11496 FLAGS(ret) = namedclass_to_classnum(arg);
11499 *flagp |= HASWIDTH|SIMPLE;
11503 nextchar(pRExC_state);
11504 Set_Node_Length(ret, 2); /* MJD */
11510 char* parse_start = RExC_parse - 2;
11515 ret = regclass(pRExC_state, flagp,depth+1,
11516 TRUE, /* means just parse this element */
11517 FALSE, /* don't allow multi-char folds */
11518 FALSE, /* don't silence non-portable warnings.
11519 It would be a bug if these returned
11522 /* regclass() can only return RESTART_UTF8 if multi-char folds
11525 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11530 Set_Node_Offset(ret, parse_start + 2);
11531 Set_Node_Cur_Length(ret, parse_start);
11532 nextchar(pRExC_state);
11536 /* Handle \N and \N{NAME} with multiple code points here and not
11537 * below because it can be multicharacter. join_exact() will join
11538 * them up later on. Also this makes sure that things like
11539 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
11540 * The options to the grok function call causes it to fail if the
11541 * sequence is just a single code point. We then go treat it as
11542 * just another character in the current EXACT node, and hence it
11543 * gets uniform treatment with all the other characters. The
11544 * special treatment for quantifiers is not needed for such single
11545 * character sequences */
11547 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
11548 FALSE /* not strict */ )) {
11549 if (*flagp & RESTART_UTF8)
11555 case 'k': /* Handle \k<NAME> and \k'NAME' */
11558 char ch= RExC_parse[1];
11559 if (ch != '<' && ch != '\'' && ch != '{') {
11561 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11562 vFAIL2("Sequence %.2s... not terminated",parse_start);
11564 /* this pretty much dupes the code for (?P=...) in reg(), if
11565 you change this make sure you change that */
11566 char* name_start = (RExC_parse += 2);
11568 SV *sv_dat = reg_scan_name(pRExC_state,
11569 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
11570 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
11571 if (RExC_parse == name_start || *RExC_parse != ch)
11572 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
11573 vFAIL2("Sequence %.3s... not terminated",parse_start);
11576 num = add_data( pRExC_state, STR_WITH_LEN("S"));
11577 RExC_rxi->data->data[num]=(void*)sv_dat;
11578 SvREFCNT_inc_simple_void(sv_dat);
11582 ret = reganode(pRExC_state,
11585 : (ASCII_FOLD_RESTRICTED)
11587 : (AT_LEAST_UNI_SEMANTICS)
11593 *flagp |= HASWIDTH;
11595 /* override incorrect value set in reganode MJD */
11596 Set_Node_Offset(ret, parse_start+1);
11597 Set_Node_Cur_Length(ret, parse_start);
11598 nextchar(pRExC_state);
11604 case '1': case '2': case '3': case '4':
11605 case '5': case '6': case '7': case '8': case '9':
11610 if (*RExC_parse == 'g') {
11614 if (*RExC_parse == '{') {
11618 if (*RExC_parse == '-') {
11622 if (hasbrace && !isDIGIT(*RExC_parse)) {
11623 if (isrel) RExC_parse--;
11625 goto parse_named_seq;
11628 num = S_backref_value(RExC_parse);
11630 vFAIL("Reference to invalid group 0");
11631 else if (num == I32_MAX) {
11632 if (isDIGIT(*RExC_parse))
11633 vFAIL("Reference to nonexistent group");
11635 vFAIL("Unterminated \\g... pattern");
11639 num = RExC_npar - num;
11641 vFAIL("Reference to nonexistent or unclosed group");
11645 num = S_backref_value(RExC_parse);
11646 /* bare \NNN might be backref or octal - if it is larger than or equal
11647 * RExC_npar then it is assumed to be and octal escape.
11648 * Note RExC_npar is +1 from the actual number of parens*/
11649 if (num == I32_MAX || (num > 9 && num >= RExC_npar
11650 && *RExC_parse != '8' && *RExC_parse != '9'))
11652 /* Probably a character specified in octal, e.g. \35 */
11657 /* at this point RExC_parse definitely points to a backref
11660 #ifdef RE_TRACK_PATTERN_OFFSETS
11661 char * const parse_start = RExC_parse - 1; /* MJD */
11663 while (isDIGIT(*RExC_parse))
11666 if (*RExC_parse != '}')
11667 vFAIL("Unterminated \\g{...} pattern");
11671 if (num > (I32)RExC_rx->nparens)
11672 vFAIL("Reference to nonexistent group");
11675 ret = reganode(pRExC_state,
11678 : (ASCII_FOLD_RESTRICTED)
11680 : (AT_LEAST_UNI_SEMANTICS)
11686 *flagp |= HASWIDTH;
11688 /* override incorrect value set in reganode MJD */
11689 Set_Node_Offset(ret, parse_start+1);
11690 Set_Node_Cur_Length(ret, parse_start);
11692 nextchar(pRExC_state);
11697 if (RExC_parse >= RExC_end)
11698 FAIL("Trailing \\");
11701 /* Do not generate "unrecognized" warnings here, we fall
11702 back into the quick-grab loop below */
11709 if (RExC_flags & RXf_PMf_EXTENDED) {
11710 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
11711 if (RExC_parse < RExC_end)
11718 parse_start = RExC_parse - 1;
11727 #define MAX_NODE_STRING_SIZE 127
11728 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
11730 U8 upper_parse = MAX_NODE_STRING_SIZE;
11731 U8 node_type = compute_EXACTish(pRExC_state);
11732 bool next_is_quantifier;
11733 char * oldp = NULL;
11735 /* We can convert EXACTF nodes to EXACTFU if they contain only
11736 * characters that match identically regardless of the target
11737 * string's UTF8ness. The reason to do this is that EXACTF is not
11738 * trie-able, EXACTFU is.
11740 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
11741 * contain only above-Latin1 characters (hence must be in UTF8),
11742 * which don't participate in folds with Latin1-range characters,
11743 * as the latter's folds aren't known until runtime. (We don't
11744 * need to figure this out until pass 2) */
11745 bool maybe_exactfu = PASS2
11746 && (node_type == EXACTF || node_type == EXACTFL);
11748 /* If a folding node contains only code points that don't
11749 * participate in folds, it can be changed into an EXACT node,
11750 * which allows the optimizer more things to look for */
11753 ret = reg_node(pRExC_state, node_type);
11755 /* In pass1, folded, we use a temporary buffer instead of the
11756 * actual node, as the node doesn't exist yet */
11757 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
11763 /* We do the EXACTFish to EXACT node only if folding. (And we
11764 * don't need to figure this out until pass 2) */
11765 maybe_exact = FOLD && PASS2;
11767 /* XXX The node can hold up to 255 bytes, yet this only goes to
11768 * 127. I (khw) do not know why. Keeping it somewhat less than
11769 * 255 allows us to not have to worry about overflow due to
11770 * converting to utf8 and fold expansion, but that value is
11771 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
11772 * split up by this limit into a single one using the real max of
11773 * 255. Even at 127, this breaks under rare circumstances. If
11774 * folding, we do not want to split a node at a character that is a
11775 * non-final in a multi-char fold, as an input string could just
11776 * happen to want to match across the node boundary. The join
11777 * would solve that problem if the join actually happens. But a
11778 * series of more than two nodes in a row each of 127 would cause
11779 * the first join to succeed to get to 254, but then there wouldn't
11780 * be room for the next one, which could at be one of those split
11781 * multi-char folds. I don't know of any fool-proof solution. One
11782 * could back off to end with only a code point that isn't such a
11783 * non-final, but it is possible for there not to be any in the
11785 for (p = RExC_parse - 1;
11786 len < upper_parse && p < RExC_end;
11791 if (RExC_flags & RXf_PMf_EXTENDED)
11792 p = regpatws(pRExC_state, p,
11793 TRUE); /* means recognize comments */
11804 /* Literal Escapes Switch
11806 This switch is meant to handle escape sequences that
11807 resolve to a literal character.
11809 Every escape sequence that represents something
11810 else, like an assertion or a char class, is handled
11811 in the switch marked 'Special Escapes' above in this
11812 routine, but also has an entry here as anything that
11813 isn't explicitly mentioned here will be treated as
11814 an unescaped equivalent literal.
11817 switch ((U8)*++p) {
11818 /* These are all the special escapes. */
11819 case 'A': /* Start assertion */
11820 case 'b': case 'B': /* Word-boundary assertion*/
11821 case 'C': /* Single char !DANGEROUS! */
11822 case 'd': case 'D': /* digit class */
11823 case 'g': case 'G': /* generic-backref, pos assertion */
11824 case 'h': case 'H': /* HORIZWS */
11825 case 'k': case 'K': /* named backref, keep marker */
11826 case 'p': case 'P': /* Unicode property */
11827 case 'R': /* LNBREAK */
11828 case 's': case 'S': /* space class */
11829 case 'v': case 'V': /* VERTWS */
11830 case 'w': case 'W': /* word class */
11831 case 'X': /* eXtended Unicode "combining
11832 character sequence" */
11833 case 'z': case 'Z': /* End of line/string assertion */
11837 /* Anything after here is an escape that resolves to a
11838 literal. (Except digits, which may or may not)
11844 case 'N': /* Handle a single-code point named character. */
11845 /* The options cause it to fail if a multiple code
11846 * point sequence. Handle those in the switch() above
11848 RExC_parse = p + 1;
11849 if (! grok_bslash_N(pRExC_state, NULL, &ender,
11850 flagp, depth, FALSE,
11851 FALSE /* not strict */ ))
11853 if (*flagp & RESTART_UTF8)
11854 FAIL("panic: grok_bslash_N set RESTART_UTF8");
11855 RExC_parse = p = oldp;
11859 if (ender > 0xff) {
11876 ender = ASCII_TO_NATIVE('\033');
11886 const char* error_msg;
11888 bool valid = grok_bslash_o(&p,
11891 TRUE, /* out warnings */
11892 FALSE, /* not strict */
11893 TRUE, /* Output warnings
11898 RExC_parse = p; /* going to die anyway; point
11899 to exact spot of failure */
11903 if (PL_encoding && ender < 0x100) {
11904 goto recode_encoding;
11906 if (ender > 0xff) {
11913 UV result = UV_MAX; /* initialize to erroneous
11915 const char* error_msg;
11917 bool valid = grok_bslash_x(&p,
11920 TRUE, /* out warnings */
11921 FALSE, /* not strict */
11922 TRUE, /* Output warnings
11927 RExC_parse = p; /* going to die anyway; point
11928 to exact spot of failure */
11933 if (PL_encoding && ender < 0x100) {
11934 goto recode_encoding;
11936 if (ender > 0xff) {
11943 ender = grok_bslash_c(*p++, SIZE_ONLY);
11945 case '8': case '9': /* must be a backreference */
11948 case '1': case '2': case '3':case '4':
11949 case '5': case '6': case '7':
11950 /* When we parse backslash escapes there is ambiguity
11951 * between backreferences and octal escapes. Any escape
11952 * from \1 - \9 is a backreference, any multi-digit
11953 * escape which does not start with 0 and which when
11954 * evaluated as decimal could refer to an already
11955 * parsed capture buffer is a backslash. Anything else
11958 * Note this implies that \118 could be interpreted as
11959 * 118 OR as "\11" . "8" depending on whether there
11960 * were 118 capture buffers defined already in the
11963 /* NOTE, RExC_npar is 1 more than the actual number of
11964 * parens we have seen so far, hence the < RExC_npar below. */
11966 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
11967 { /* Not to be treated as an octal constant, go
11975 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11977 ender = grok_oct(p, &numlen, &flags, NULL);
11978 if (ender > 0xff) {
11982 if (SIZE_ONLY /* like \08, \178 */
11985 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11987 reg_warn_non_literal_string(
11989 form_short_octal_warning(p, numlen));
11992 if (PL_encoding && ender < 0x100)
11993 goto recode_encoding;
11996 if (! RExC_override_recoding) {
11997 SV* enc = PL_encoding;
11998 ender = reg_recode((const char)(U8)ender, &enc);
11999 if (!enc && SIZE_ONLY)
12000 ckWARNreg(p, "Invalid escape in the specified encoding");
12006 FAIL("Trailing \\");
12009 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
12010 /* Include any { following the alpha to emphasize
12011 * that it could be part of an escape at some point
12013 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
12014 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
12016 goto normal_default;
12017 } /* End of switch on '\' */
12020 /* Currently we don't warn when the lbrace is at the start
12021 * of a construct. This catches it in the middle of a
12022 * literal string, or when its the first thing after
12023 * something like "\b" */
12025 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
12027 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
12030 default: /* A literal character */
12032 if (UTF8_IS_START(*p) && UTF) {
12034 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
12035 &numlen, UTF8_ALLOW_DEFAULT);
12041 } /* End of switch on the literal */
12043 /* Here, have looked at the literal character and <ender>
12044 * contains its ordinal, <p> points to the character after it
12047 if ( RExC_flags & RXf_PMf_EXTENDED)
12048 p = regpatws(pRExC_state, p,
12049 TRUE); /* means recognize comments */
12051 /* If the next thing is a quantifier, it applies to this
12052 * character only, which means that this character has to be in
12053 * its own node and can't just be appended to the string in an
12054 * existing node, so if there are already other characters in
12055 * the node, close the node with just them, and set up to do
12056 * this character again next time through, when it will be the
12057 * only thing in its new node */
12058 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
12064 if (! FOLD /* The simple case, just append the literal */
12065 || (LOC /* Also don't fold for tricky chars under /l */
12066 && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)))
12069 const STRLEN unilen = reguni(pRExC_state, ender, s);
12075 /* The loop increments <len> each time, as all but this
12076 * path (and one other) through it add a single byte to
12077 * the EXACTish node. But this one has changed len to
12078 * be the correct final value, so subtract one to
12079 * cancel out the increment that follows */
12083 REGC((char)ender, s++);
12086 /* Can get here if folding only if is one of the /l
12087 * characters whose fold depends on the locale. The
12088 * occurrence of any of these indicate that we can't
12089 * simplify things */
12091 maybe_exact = FALSE;
12092 maybe_exactfu = FALSE;
12097 /* See comments for join_exact() as to why we fold this
12098 * non-UTF at compile time */
12099 || (node_type == EXACTFU
12100 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12102 /* Here, are folding and are not UTF-8 encoded; therefore
12103 * the character must be in the range 0-255, and is not /l
12104 * (Not /l because we already handled these under /l in
12105 * is_PROBLEMATIC_LOCALE_FOLD_cp */
12106 if (IS_IN_SOME_FOLD_L1(ender)) {
12107 maybe_exact = FALSE;
12109 /* See if the character's fold differs between /d and
12110 * /u. This includes the multi-char fold SHARP S to
12113 && (PL_fold[ender] != PL_fold_latin1[ender]
12114 || ender == LATIN_SMALL_LETTER_SHARP_S
12116 && isARG2_lower_or_UPPER_ARG1('s', ender)
12117 && isARG2_lower_or_UPPER_ARG1('s',
12120 maybe_exactfu = FALSE;
12124 /* Even when folding, we store just the input character, as
12125 * we have an array that finds its fold quickly */
12126 *(s++) = (char) ender;
12128 else { /* FOLD and UTF */
12129 /* Unlike the non-fold case, we do actually have to
12130 * calculate the results here in pass 1. This is for two
12131 * reasons, the folded length may be longer than the
12132 * unfolded, and we have to calculate how many EXACTish
12133 * nodes it will take; and we may run out of room in a node
12134 * in the middle of a potential multi-char fold, and have
12135 * to back off accordingly. (Hence we can't use REGC for
12136 * the simple case just below.) */
12139 if (isASCII(ender)) {
12140 folded = toFOLD(ender);
12141 *(s)++ = (U8) folded;
12146 folded = _to_uni_fold_flags(
12150 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12151 ? FOLD_FLAGS_NOMIX_ASCII
12155 /* The loop increments <len> each time, as all but this
12156 * path (and one other) through it add a single byte to
12157 * the EXACTish node. But this one has changed len to
12158 * be the correct final value, so subtract one to
12159 * cancel out the increment that follows */
12160 len += foldlen - 1;
12162 /* If this node only contains non-folding code points so
12163 * far, see if this new one is also non-folding */
12165 if (folded != ender) {
12166 maybe_exact = FALSE;
12169 /* Here the fold is the original; we have to check
12170 * further to see if anything folds to it */
12171 if (_invlist_contains_cp(PL_utf8_foldable,
12174 maybe_exact = FALSE;
12181 if (next_is_quantifier) {
12183 /* Here, the next input is a quantifier, and to get here,
12184 * the current character is the only one in the node.
12185 * Also, here <len> doesn't include the final byte for this
12191 } /* End of loop through literal characters */
12193 /* Here we have either exhausted the input or ran out of room in
12194 * the node. (If we encountered a character that can't be in the
12195 * node, transfer is made directly to <loopdone>, and so we
12196 * wouldn't have fallen off the end of the loop.) In the latter
12197 * case, we artificially have to split the node into two, because
12198 * we just don't have enough space to hold everything. This
12199 * creates a problem if the final character participates in a
12200 * multi-character fold in the non-final position, as a match that
12201 * should have occurred won't, due to the way nodes are matched,
12202 * and our artificial boundary. So back off until we find a non-
12203 * problematic character -- one that isn't at the beginning or
12204 * middle of such a fold. (Either it doesn't participate in any
12205 * folds, or appears only in the final position of all the folds it
12206 * does participate in.) A better solution with far fewer false
12207 * positives, and that would fill the nodes more completely, would
12208 * be to actually have available all the multi-character folds to
12209 * test against, and to back-off only far enough to be sure that
12210 * this node isn't ending with a partial one. <upper_parse> is set
12211 * further below (if we need to reparse the node) to include just
12212 * up through that final non-problematic character that this code
12213 * identifies, so when it is set to less than the full node, we can
12214 * skip the rest of this */
12215 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12217 const STRLEN full_len = len;
12219 assert(len >= MAX_NODE_STRING_SIZE);
12221 /* Here, <s> points to the final byte of the final character.
12222 * Look backwards through the string until find a non-
12223 * problematic character */
12227 /* This has no multi-char folds to non-UTF characters */
12228 if (ASCII_FOLD_RESTRICTED) {
12232 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12236 if (! PL_NonL1NonFinalFold) {
12237 PL_NonL1NonFinalFold = _new_invlist_C_array(
12238 NonL1_Perl_Non_Final_Folds_invlist);
12241 /* Point to the first byte of the final character */
12242 s = (char *) utf8_hop((U8 *) s, -1);
12244 while (s >= s0) { /* Search backwards until find
12245 non-problematic char */
12246 if (UTF8_IS_INVARIANT(*s)) {
12248 /* There are no ascii characters that participate
12249 * in multi-char folds under /aa. In EBCDIC, the
12250 * non-ascii invariants are all control characters,
12251 * so don't ever participate in any folds. */
12252 if (ASCII_FOLD_RESTRICTED
12253 || ! IS_NON_FINAL_FOLD(*s))
12258 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12259 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12265 else if (! _invlist_contains_cp(
12266 PL_NonL1NonFinalFold,
12267 valid_utf8_to_uvchr((U8 *) s, NULL)))
12272 /* Here, the current character is problematic in that
12273 * it does occur in the non-final position of some
12274 * fold, so try the character before it, but have to
12275 * special case the very first byte in the string, so
12276 * we don't read outside the string */
12277 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12278 } /* End of loop backwards through the string */
12280 /* If there were only problematic characters in the string,
12281 * <s> will point to before s0, in which case the length
12282 * should be 0, otherwise include the length of the
12283 * non-problematic character just found */
12284 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12287 /* Here, have found the final character, if any, that is
12288 * non-problematic as far as ending the node without splitting
12289 * it across a potential multi-char fold. <len> contains the
12290 * number of bytes in the node up-to and including that
12291 * character, or is 0 if there is no such character, meaning
12292 * the whole node contains only problematic characters. In
12293 * this case, give up and just take the node as-is. We can't
12298 /* If the node ends in an 's' we make sure it stays EXACTF,
12299 * as if it turns into an EXACTFU, it could later get
12300 * joined with another 's' that would then wrongly match
12302 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
12304 maybe_exactfu = FALSE;
12308 /* Here, the node does contain some characters that aren't
12309 * problematic. If one such is the final character in the
12310 * node, we are done */
12311 if (len == full_len) {
12314 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12316 /* If the final character is problematic, but the
12317 * penultimate is not, back-off that last character to
12318 * later start a new node with it */
12323 /* Here, the final non-problematic character is earlier
12324 * in the input than the penultimate character. What we do
12325 * is reparse from the beginning, going up only as far as
12326 * this final ok one, thus guaranteeing that the node ends
12327 * in an acceptable character. The reason we reparse is
12328 * that we know how far in the character is, but we don't
12329 * know how to correlate its position with the input parse.
12330 * An alternate implementation would be to build that
12331 * correlation as we go along during the original parse,
12332 * but that would entail extra work for every node, whereas
12333 * this code gets executed only when the string is too
12334 * large for the node, and the final two characters are
12335 * problematic, an infrequent occurrence. Yet another
12336 * possible strategy would be to save the tail of the
12337 * string, and the next time regatom is called, initialize
12338 * with that. The problem with this is that unless you
12339 * back off one more character, you won't be guaranteed
12340 * regatom will get called again, unless regbranch,
12341 * regpiece ... are also changed. If you do back off that
12342 * extra character, so that there is input guaranteed to
12343 * force calling regatom, you can't handle the case where
12344 * just the first character in the node is acceptable. I
12345 * (khw) decided to try this method which doesn't have that
12346 * pitfall; if performance issues are found, we can do a
12347 * combination of the current approach plus that one */
12353 } /* End of verifying node ends with an appropriate char */
12355 loopdone: /* Jumped to when encounters something that shouldn't be in
12358 /* I (khw) don't know if you can get here with zero length, but the
12359 * old code handled this situation by creating a zero-length EXACT
12360 * node. Might as well be NOTHING instead */
12366 /* If 'maybe_exact' is still set here, means there are no
12367 * code points in the node that participate in folds;
12368 * similarly for 'maybe_exactfu' and code points that match
12369 * differently depending on UTF8ness of the target string
12370 * (for /u), or depending on locale for /l */
12374 else if (maybe_exactfu) {
12378 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12379 FALSE /* Don't look to see if could
12380 be turned into an EXACT
12381 node, as we have already
12386 RExC_parse = p - 1;
12387 Set_Node_Cur_Length(ret, parse_start);
12388 nextchar(pRExC_state);
12390 /* len is STRLEN which is unsigned, need to copy to signed */
12393 vFAIL("Internal disaster");
12396 } /* End of label 'defchar:' */
12398 } /* End of giant switch on input character */
12404 S_regpatws(RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12406 /* Returns the next non-pattern-white space, non-comment character (the
12407 * latter only if 'recognize_comment is true) in the string p, which is
12408 * ended by RExC_end. See also reg_skipcomment */
12409 const char *e = RExC_end;
12411 PERL_ARGS_ASSERT_REGPATWS;
12415 if ((len = is_PATWS_safe(p, e, UTF))) {
12418 else if (recognize_comment && *p == '#') {
12419 p = reg_skipcomment(pRExC_state, p);
12428 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
12430 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
12431 * sets up the bitmap and any flags, removing those code points from the
12432 * inversion list, setting it to NULL should it become completely empty */
12434 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
12435 assert(PL_regkind[OP(node)] == ANYOF);
12437 ANYOF_BITMAP_ZERO(node);
12438 if (*invlist_ptr) {
12440 /* This gets set if we actually need to modify things */
12441 bool change_invlist = FALSE;
12445 /* Start looking through *invlist_ptr */
12446 invlist_iterinit(*invlist_ptr);
12447 while (invlist_iternext(*invlist_ptr, &start, &end)) {
12451 if (end == UV_MAX && start <= 256) {
12452 ANYOF_FLAGS(node) |= ANYOF_ABOVE_LATIN1_ALL;
12454 else if (end >= 256) {
12455 ANYOF_FLAGS(node) |= ANYOF_UTF8;
12458 /* Quit if are above what we should change */
12463 change_invlist = TRUE;
12465 /* Set all the bits in the range, up to the max that we are doing */
12466 high = (end < 255) ? end : 255;
12467 for (i = start; i <= (int) high; i++) {
12468 if (! ANYOF_BITMAP_TEST(node, i)) {
12469 ANYOF_BITMAP_SET(node, i);
12473 invlist_iterfinish(*invlist_ptr);
12475 /* Done with loop; remove any code points that are in the bitmap from
12476 * *invlist_ptr; similarly for code points above latin1 if we have a
12477 * flag to match all of them anyways */
12478 if (change_invlist) {
12479 _invlist_subtract(*invlist_ptr, PL_Latin1, invlist_ptr);
12481 if (ANYOF_FLAGS(node) & ANYOF_ABOVE_LATIN1_ALL) {
12482 _invlist_intersection(*invlist_ptr, PL_Latin1, invlist_ptr);
12485 /* If have completely emptied it, remove it completely */
12486 if (_invlist_len(*invlist_ptr) == 0) {
12487 SvREFCNT_dec_NN(*invlist_ptr);
12488 *invlist_ptr = NULL;
12493 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
12494 Character classes ([:foo:]) can also be negated ([:^foo:]).
12495 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
12496 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
12497 but trigger failures because they are currently unimplemented. */
12499 #define POSIXCC_DONE(c) ((c) == ':')
12500 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
12501 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
12503 PERL_STATIC_INLINE I32
12504 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
12506 I32 namedclass = OOB_NAMEDCLASS;
12508 PERL_ARGS_ASSERT_REGPPOSIXCC;
12510 if (value == '[' && RExC_parse + 1 < RExC_end &&
12511 /* I smell either [: or [= or [. -- POSIX has been here, right? */
12512 POSIXCC(UCHARAT(RExC_parse)))
12514 const char c = UCHARAT(RExC_parse);
12515 char* const s = RExC_parse++;
12517 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
12519 if (RExC_parse == RExC_end) {
12522 /* Try to give a better location for the error (than the end of
12523 * the string) by looking for the matching ']' */
12525 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
12528 vFAIL2("Unmatched '%c' in POSIX class", c);
12530 /* Grandfather lone [:, [=, [. */
12534 const char* const t = RExC_parse++; /* skip over the c */
12537 if (UCHARAT(RExC_parse) == ']') {
12538 const char *posixcc = s + 1;
12539 RExC_parse++; /* skip over the ending ] */
12542 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
12543 const I32 skip = t - posixcc;
12545 /* Initially switch on the length of the name. */
12548 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
12549 this is the Perl \w
12551 namedclass = ANYOF_WORDCHAR;
12554 /* Names all of length 5. */
12555 /* alnum alpha ascii blank cntrl digit graph lower
12556 print punct space upper */
12557 /* Offset 4 gives the best switch position. */
12558 switch (posixcc[4]) {
12560 if (memEQ(posixcc, "alph", 4)) /* alpha */
12561 namedclass = ANYOF_ALPHA;
12564 if (memEQ(posixcc, "spac", 4)) /* space */
12565 namedclass = ANYOF_PSXSPC;
12568 if (memEQ(posixcc, "grap", 4)) /* graph */
12569 namedclass = ANYOF_GRAPH;
12572 if (memEQ(posixcc, "asci", 4)) /* ascii */
12573 namedclass = ANYOF_ASCII;
12576 if (memEQ(posixcc, "blan", 4)) /* blank */
12577 namedclass = ANYOF_BLANK;
12580 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
12581 namedclass = ANYOF_CNTRL;
12584 if (memEQ(posixcc, "alnu", 4)) /* alnum */
12585 namedclass = ANYOF_ALPHANUMERIC;
12588 if (memEQ(posixcc, "lowe", 4)) /* lower */
12589 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
12590 else if (memEQ(posixcc, "uppe", 4)) /* upper */
12591 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
12594 if (memEQ(posixcc, "digi", 4)) /* digit */
12595 namedclass = ANYOF_DIGIT;
12596 else if (memEQ(posixcc, "prin", 4)) /* print */
12597 namedclass = ANYOF_PRINT;
12598 else if (memEQ(posixcc, "punc", 4)) /* punct */
12599 namedclass = ANYOF_PUNCT;
12604 if (memEQ(posixcc, "xdigit", 6))
12605 namedclass = ANYOF_XDIGIT;
12609 if (namedclass == OOB_NAMEDCLASS)
12611 "POSIX class [:%"UTF8f":] unknown",
12612 UTF8fARG(UTF, t - s - 1, s + 1));
12614 /* The #defines are structured so each complement is +1 to
12615 * the normal one */
12619 assert (posixcc[skip] == ':');
12620 assert (posixcc[skip+1] == ']');
12621 } else if (!SIZE_ONLY) {
12622 /* [[=foo=]] and [[.foo.]] are still future. */
12624 /* adjust RExC_parse so the warning shows after
12625 the class closes */
12626 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
12628 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
12631 /* Maternal grandfather:
12632 * "[:" ending in ":" but not in ":]" */
12634 vFAIL("Unmatched '[' in POSIX class");
12637 /* Grandfather lone [:, [=, [. */
12647 S_could_it_be_a_POSIX_class(RExC_state_t *pRExC_state)
12649 /* This applies some heuristics at the current parse position (which should
12650 * be at a '[') to see if what follows might be intended to be a [:posix:]
12651 * class. It returns true if it really is a posix class, of course, but it
12652 * also can return true if it thinks that what was intended was a posix
12653 * class that didn't quite make it.
12655 * It will return true for
12657 * [:alphanumerics] (as long as the ] isn't followed immediately by a
12658 * ')' indicating the end of the (?[
12659 * [:any garbage including %^&$ punctuation:]
12661 * This is designed to be called only from S_handle_regex_sets; it could be
12662 * easily adapted to be called from the spot at the beginning of regclass()
12663 * that checks to see in a normal bracketed class if the surrounding []
12664 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
12665 * change long-standing behavior, so I (khw) didn't do that */
12666 char* p = RExC_parse + 1;
12667 char first_char = *p;
12669 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
12671 assert(*(p - 1) == '[');
12673 if (! POSIXCC(first_char)) {
12678 while (p < RExC_end && isWORDCHAR(*p)) p++;
12680 if (p >= RExC_end) {
12684 if (p - RExC_parse > 2 /* Got at least 1 word character */
12685 && (*p == first_char
12686 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
12691 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
12694 && p - RExC_parse > 2 /* [:] evaluates to colon;
12695 [::] is a bad posix class. */
12696 && first_char == *(p - 1));
12700 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
12701 I32 *flagp, U32 depth,
12702 char * const oregcomp_parse)
12704 /* Handle the (?[...]) construct to do set operations */
12707 UV start, end; /* End points of code point ranges */
12709 char *save_end, *save_parse;
12714 const bool save_fold = FOLD;
12716 GET_RE_DEBUG_FLAGS_DECL;
12718 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
12721 vFAIL("(?[...]) not valid in locale");
12723 RExC_uni_semantics = 1;
12725 /* This will return only an ANYOF regnode, or (unlikely) something smaller
12726 * (such as EXACT). Thus we can skip most everything if just sizing. We
12727 * call regclass to handle '[]' so as to not have to reinvent its parsing
12728 * rules here (throwing away the size it computes each time). And, we exit
12729 * upon an unescaped ']' that isn't one ending a regclass. To do both
12730 * these things, we need to realize that something preceded by a backslash
12731 * is escaped, so we have to keep track of backslashes */
12733 UV depth = 0; /* how many nested (?[...]) constructs */
12735 Perl_ck_warner_d(aTHX_
12736 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
12737 "The regex_sets feature is experimental" REPORT_LOCATION,
12738 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
12740 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
12741 RExC_precomp + (RExC_parse - RExC_precomp)));
12743 while (RExC_parse < RExC_end) {
12744 SV* current = NULL;
12745 RExC_parse = regpatws(pRExC_state, RExC_parse,
12746 TRUE); /* means recognize comments */
12747 switch (*RExC_parse) {
12749 if (RExC_parse[1] == '[') depth++, RExC_parse++;
12754 /* Skip the next byte (which could cause us to end up in
12755 * the middle of a UTF-8 character, but since none of those
12756 * are confusable with anything we currently handle in this
12757 * switch (invariants all), it's safe. We'll just hit the
12758 * default: case next time and keep on incrementing until
12759 * we find one of the invariants we do handle. */
12764 /* If this looks like it is a [:posix:] class, leave the
12765 * parse pointer at the '[' to fool regclass() into
12766 * thinking it is part of a '[[:posix:]]'. That function
12767 * will use strict checking to force a syntax error if it
12768 * doesn't work out to a legitimate class */
12769 bool is_posix_class
12770 = could_it_be_a_POSIX_class(pRExC_state);
12771 if (! is_posix_class) {
12775 /* regclass() can only return RESTART_UTF8 if multi-char
12776 folds are allowed. */
12777 if (!regclass(pRExC_state, flagp,depth+1,
12778 is_posix_class, /* parse the whole char
12779 class only if not a
12781 FALSE, /* don't allow multi-char folds */
12782 TRUE, /* silence non-portable warnings. */
12784 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12787 /* function call leaves parse pointing to the ']', except
12788 * if we faked it */
12789 if (is_posix_class) {
12793 SvREFCNT_dec(current); /* In case it returned something */
12798 if (depth--) break;
12800 if (RExC_parse < RExC_end
12801 && *RExC_parse == ')')
12803 node = reganode(pRExC_state, ANYOF, 0);
12804 RExC_size += ANYOF_SKIP;
12805 nextchar(pRExC_state);
12806 Set_Node_Length(node,
12807 RExC_parse - oregcomp_parse + 1); /* MJD */
12816 FAIL("Syntax error in (?[...])");
12819 /* Pass 2 only after this. Everything in this construct is a
12820 * metacharacter. Operands begin with either a '\' (for an escape
12821 * sequence), or a '[' for a bracketed character class. Any other
12822 * character should be an operator, or parenthesis for grouping. Both
12823 * types of operands are handled by calling regclass() to parse them. It
12824 * is called with a parameter to indicate to return the computed inversion
12825 * list. The parsing here is implemented via a stack. Each entry on the
12826 * stack is a single character representing one of the operators, or the
12827 * '('; or else a pointer to an operand inversion list. */
12829 #define IS_OPERAND(a) (! SvIOK(a))
12831 /* The stack starts empty. It is a syntax error if the first thing parsed
12832 * is a binary operator; everything else is pushed on the stack. When an
12833 * operand is parsed, the top of the stack is examined. If it is a binary
12834 * operator, the item before it should be an operand, and both are replaced
12835 * by the result of doing that operation on the new operand and the one on
12836 * the stack. Thus a sequence of binary operands is reduced to a single
12837 * one before the next one is parsed.
12839 * A unary operator may immediately follow a binary in the input, for
12842 * When an operand is parsed and the top of the stack is a unary operator,
12843 * the operation is performed, and then the stack is rechecked to see if
12844 * this new operand is part of a binary operation; if so, it is handled as
12847 * A '(' is simply pushed on the stack; it is valid only if the stack is
12848 * empty, or the top element of the stack is an operator or another '('
12849 * (for which the parenthesized expression will become an operand). By the
12850 * time the corresponding ')' is parsed everything in between should have
12851 * been parsed and evaluated to a single operand (or else is a syntax
12852 * error), and is handled as a regular operand */
12854 sv_2mortal((SV *)(stack = newAV()));
12856 while (RExC_parse < RExC_end) {
12857 I32 top_index = av_tindex(stack);
12859 SV* current = NULL;
12861 /* Skip white space */
12862 RExC_parse = regpatws(pRExC_state, RExC_parse,
12863 TRUE /* means recognize comments */ );
12864 if (RExC_parse >= RExC_end) {
12865 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
12867 if ((curchar = UCHARAT(RExC_parse)) == ']') {
12874 if (av_tindex(stack) >= 0 /* This makes sure that we can
12875 safely subtract 1 from
12876 RExC_parse in the next clause.
12877 If we have something on the
12878 stack, we have parsed something
12880 && UCHARAT(RExC_parse - 1) == '('
12881 && RExC_parse < RExC_end)
12883 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
12884 * This happens when we have some thing like
12886 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
12888 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
12890 * Here we would be handling the interpolated
12891 * '$thai_or_lao'. We handle this by a recursive call to
12892 * ourselves which returns the inversion list the
12893 * interpolated expression evaluates to. We use the flags
12894 * from the interpolated pattern. */
12895 U32 save_flags = RExC_flags;
12896 const char * const save_parse = ++RExC_parse;
12898 parse_lparen_question_flags(pRExC_state);
12900 if (RExC_parse == save_parse /* Makes sure there was at
12901 least one flag (or this
12902 embedding wasn't compiled)
12904 || RExC_parse >= RExC_end - 4
12905 || UCHARAT(RExC_parse) != ':'
12906 || UCHARAT(++RExC_parse) != '('
12907 || UCHARAT(++RExC_parse) != '?'
12908 || UCHARAT(++RExC_parse) != '[')
12911 /* In combination with the above, this moves the
12912 * pointer to the point just after the first erroneous
12913 * character (or if there are no flags, to where they
12914 * should have been) */
12915 if (RExC_parse >= RExC_end - 4) {
12916 RExC_parse = RExC_end;
12918 else if (RExC_parse != save_parse) {
12919 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12921 vFAIL("Expecting '(?flags:(?[...'");
12924 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
12925 depth+1, oregcomp_parse);
12927 /* Here, 'current' contains the embedded expression's
12928 * inversion list, and RExC_parse points to the trailing
12929 * ']'; the next character should be the ')' which will be
12930 * paired with the '(' that has been put on the stack, so
12931 * the whole embedded expression reduces to '(operand)' */
12934 RExC_flags = save_flags;
12935 goto handle_operand;
12940 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12941 vFAIL("Unexpected character");
12944 /* regclass() can only return RESTART_UTF8 if multi-char
12945 folds are allowed. */
12946 if (!regclass(pRExC_state, flagp,depth+1,
12947 TRUE, /* means parse just the next thing */
12948 FALSE, /* don't allow multi-char folds */
12949 FALSE, /* don't silence non-portable warnings. */
12951 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12953 /* regclass() will return with parsing just the \ sequence,
12954 * leaving the parse pointer at the next thing to parse */
12956 goto handle_operand;
12958 case '[': /* Is a bracketed character class */
12960 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
12962 if (! is_posix_class) {
12966 /* regclass() can only return RESTART_UTF8 if multi-char
12967 folds are allowed. */
12968 if(!regclass(pRExC_state, flagp,depth+1,
12969 is_posix_class, /* parse the whole char class
12970 only if not a posix class */
12971 FALSE, /* don't allow multi-char folds */
12972 FALSE, /* don't silence non-portable warnings. */
12974 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
12976 /* function call leaves parse pointing to the ']', except if we
12978 if (is_posix_class) {
12982 goto handle_operand;
12991 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
12992 || ! IS_OPERAND(*top_ptr))
12995 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
12997 av_push(stack, newSVuv(curchar));
13001 av_push(stack, newSVuv(curchar));
13005 if (top_index >= 0) {
13006 top_ptr = av_fetch(stack, top_index, FALSE);
13008 if (IS_OPERAND(*top_ptr)) {
13010 vFAIL("Unexpected '(' with no preceding operator");
13013 av_push(stack, newSVuv(curchar));
13020 || ! (current = av_pop(stack))
13021 || ! IS_OPERAND(current)
13022 || ! (lparen = av_pop(stack))
13023 || IS_OPERAND(lparen)
13024 || SvUV(lparen) != '(')
13026 SvREFCNT_dec(current);
13028 vFAIL("Unexpected ')'");
13031 SvREFCNT_dec_NN(lparen);
13038 /* Here, we have an operand to process, in 'current' */
13040 if (top_index < 0) { /* Just push if stack is empty */
13041 av_push(stack, current);
13044 SV* top = av_pop(stack);
13046 char current_operator;
13048 if (IS_OPERAND(top)) {
13049 SvREFCNT_dec_NN(top);
13050 SvREFCNT_dec_NN(current);
13051 vFAIL("Operand with no preceding operator");
13053 current_operator = (char) SvUV(top);
13054 switch (current_operator) {
13055 case '(': /* Push the '(' back on followed by the new
13057 av_push(stack, top);
13058 av_push(stack, current);
13059 SvREFCNT_inc(top); /* Counters the '_dec' done
13060 just after the 'break', so
13061 it doesn't get wrongly freed
13066 _invlist_invert(current);
13068 /* Unlike binary operators, the top of the stack,
13069 * now that this unary one has been popped off, may
13070 * legally be an operator, and we now have operand
13073 SvREFCNT_dec_NN(top);
13074 goto handle_operand;
13077 prev = av_pop(stack);
13078 _invlist_intersection(prev,
13081 av_push(stack, current);
13086 prev = av_pop(stack);
13087 _invlist_union(prev, current, ¤t);
13088 av_push(stack, current);
13092 prev = av_pop(stack);;
13093 _invlist_subtract(prev, current, ¤t);
13094 av_push(stack, current);
13097 case '^': /* The union minus the intersection */
13103 prev = av_pop(stack);
13104 _invlist_union(prev, current, &u);
13105 _invlist_intersection(prev, current, &i);
13106 /* _invlist_subtract will overwrite current
13107 without freeing what it already contains */
13109 _invlist_subtract(u, i, ¤t);
13110 av_push(stack, current);
13111 SvREFCNT_dec_NN(i);
13112 SvREFCNT_dec_NN(u);
13113 SvREFCNT_dec_NN(element);
13118 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
13120 SvREFCNT_dec_NN(top);
13121 SvREFCNT_dec(prev);
13125 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13128 if (av_tindex(stack) < 0 /* Was empty */
13129 || ((final = av_pop(stack)) == NULL)
13130 || ! IS_OPERAND(final)
13131 || av_tindex(stack) >= 0) /* More left on stack */
13133 vFAIL("Incomplete expression within '(?[ ])'");
13136 /* Here, 'final' is the resultant inversion list from evaluating the
13137 * expression. Return it if so requested */
13138 if (return_invlist) {
13139 *return_invlist = final;
13143 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13144 * expecting a string of ranges and individual code points */
13145 invlist_iterinit(final);
13146 result_string = newSVpvs("");
13147 while (invlist_iternext(final, &start, &end)) {
13148 if (start == end) {
13149 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13152 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13157 save_parse = RExC_parse;
13158 RExC_parse = SvPV(result_string, len);
13159 save_end = RExC_end;
13160 RExC_end = RExC_parse + len;
13162 /* We turn off folding around the call, as the class we have constructed
13163 * already has all folding taken into consideration, and we don't want
13164 * regclass() to add to that */
13165 RExC_flags &= ~RXf_PMf_FOLD;
13166 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13168 node = regclass(pRExC_state, flagp,depth+1,
13169 FALSE, /* means parse the whole char class */
13170 FALSE, /* don't allow multi-char folds */
13171 TRUE, /* silence non-portable warnings. The above may very
13172 well have generated non-portable code points, but
13173 they're valid on this machine */
13176 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13179 RExC_flags |= RXf_PMf_FOLD;
13181 RExC_parse = save_parse + 1;
13182 RExC_end = save_end;
13183 SvREFCNT_dec_NN(final);
13184 SvREFCNT_dec_NN(result_string);
13186 nextchar(pRExC_state);
13187 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13193 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
13195 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
13196 * innocent-looking character class, like /[ks]/i won't have to go out to
13197 * disk to find the possible matches.
13199 * This should be called only for a Latin1-range code points, cp, which is
13200 * known to be involved in a simple fold with other code points above
13201 * Latin1. It would give false results if /aa has been specified.
13202 * Multi-char folds are outside the scope of this, and must be handled
13205 * XXX It would be better to generate these via regen, in case a new
13206 * version of the Unicode standard adds new mappings, though that is not
13207 * really likely, and may be caught by the default: case of the switch
13210 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
13212 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
13218 add_cp_to_invlist(*invlist, KELVIN_SIGN);
13222 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
13225 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
13226 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
13228 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13229 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13230 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
13232 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13233 *invlist = add_cp_to_invlist(*invlist,
13234 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13236 case LATIN_SMALL_LETTER_SHARP_S:
13237 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
13240 /* Use deprecated warning to increase the chances of this being
13242 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
13247 /* The names of properties whose definitions are not known at compile time are
13248 * stored in this SV, after a constant heading. So if the length has been
13249 * changed since initialization, then there is a run-time definition. */
13250 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
13251 (SvCUR(listsv) != initial_listsv_len)
13254 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
13255 const bool stop_at_1, /* Just parse the next thing, don't
13256 look for a full character class */
13257 bool allow_multi_folds,
13258 const bool silence_non_portable, /* Don't output warnings
13261 SV** ret_invlist) /* Return an inversion list, not a node */
13263 /* parse a bracketed class specification. Most of these will produce an
13264 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
13265 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
13266 * under /i with multi-character folds: it will be rewritten following the
13267 * paradigm of this example, where the <multi-fold>s are characters which
13268 * fold to multiple character sequences:
13269 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
13270 * gets effectively rewritten as:
13271 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
13272 * reg() gets called (recursively) on the rewritten version, and this
13273 * function will return what it constructs. (Actually the <multi-fold>s
13274 * aren't physically removed from the [abcdefghi], it's just that they are
13275 * ignored in the recursion by means of a flag:
13276 * <RExC_in_multi_char_class>.)
13278 * ANYOF nodes contain a bit map for the first 256 characters, with the
13279 * corresponding bit set if that character is in the list. For characters
13280 * above 255, a range list or swash is used. There are extra bits for \w,
13281 * etc. in locale ANYOFs, as what these match is not determinable at
13284 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
13285 * to be restarted. This can only happen if ret_invlist is non-NULL.
13288 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
13290 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
13293 IV namedclass = OOB_NAMEDCLASS;
13294 char *rangebegin = NULL;
13295 bool need_class = 0;
13297 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
13298 than just initialized. */
13299 SV* properties = NULL; /* Code points that match \p{} \P{} */
13300 SV* posixes = NULL; /* Code points that match classes like [:word:],
13301 extended beyond the Latin1 range. These have to
13302 be kept separate from other code points for much
13303 of this function because their handling is
13304 different under /i, and for most classes under
13306 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
13307 separate for a while from the non-complemented
13308 versions because of complications with /d
13310 UV element_count = 0; /* Number of distinct elements in the class.
13311 Optimizations may be possible if this is tiny */
13312 AV * multi_char_matches = NULL; /* Code points that fold to more than one
13313 character; used under /i */
13315 char * stop_ptr = RExC_end; /* where to stop parsing */
13316 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
13318 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
13320 /* Unicode properties are stored in a swash; this holds the current one
13321 * being parsed. If this swash is the only above-latin1 component of the
13322 * character class, an optimization is to pass it directly on to the
13323 * execution engine. Otherwise, it is set to NULL to indicate that there
13324 * are other things in the class that have to be dealt with at execution
13326 SV* swash = NULL; /* Code points that match \p{} \P{} */
13328 /* Set if a component of this character class is user-defined; just passed
13329 * on to the engine */
13330 bool has_user_defined_property = FALSE;
13332 /* inversion list of code points this node matches only when the target
13333 * string is in UTF-8. (Because is under /d) */
13334 SV* depends_list = NULL;
13336 /* Inversion list of code points this node matches regardless of things
13337 * like locale, folding, utf8ness of the target string */
13338 SV* cp_list = NULL;
13340 /* Like cp_list, but code points on this list need to be checked for things
13341 * that fold to/from them under /i */
13342 SV* cp_foldable_list = NULL;
13344 /* Like cp_list, but code points on this list are valid only when the
13345 * runtime locale is UTF-8 */
13346 SV* only_utf8_locale_list = NULL;
13349 /* In a range, counts how many 0-2 of the ends of it came from literals,
13350 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
13351 UV literal_endpoint = 0;
13353 bool invert = FALSE; /* Is this class to be complemented */
13355 bool warn_super = ALWAYS_WARN_SUPER;
13357 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
13358 case we need to change the emitted regop to an EXACT. */
13359 const char * orig_parse = RExC_parse;
13360 const SSize_t orig_size = RExC_size;
13361 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
13362 GET_RE_DEBUG_FLAGS_DECL;
13364 PERL_ARGS_ASSERT_REGCLASS;
13366 PERL_UNUSED_ARG(depth);
13369 DEBUG_PARSE("clas");
13371 /* Assume we are going to generate an ANYOF node. */
13372 ret = reganode(pRExC_state, ANYOF, 0);
13375 RExC_size += ANYOF_SKIP;
13376 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
13379 ANYOF_FLAGS(ret) = 0;
13381 RExC_emit += ANYOF_SKIP;
13382 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
13383 initial_listsv_len = SvCUR(listsv);
13384 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
13388 RExC_parse = regpatws(pRExC_state, RExC_parse,
13389 FALSE /* means don't recognize comments */ );
13392 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
13395 allow_multi_folds = FALSE;
13398 RExC_parse = regpatws(pRExC_state, RExC_parse,
13399 FALSE /* means don't recognize comments */ );
13403 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
13404 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
13405 const char *s = RExC_parse;
13406 const char c = *s++;
13408 while (isWORDCHAR(*s))
13410 if (*s && c == *s && s[1] == ']') {
13411 SAVEFREESV(RExC_rx_sv);
13413 "POSIX syntax [%c %c] belongs inside character classes",
13415 (void)ReREFCNT_inc(RExC_rx_sv);
13419 /* If the caller wants us to just parse a single element, accomplish this
13420 * by faking the loop ending condition */
13421 if (stop_at_1 && RExC_end > RExC_parse) {
13422 stop_ptr = RExC_parse + 1;
13425 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
13426 if (UCHARAT(RExC_parse) == ']')
13427 goto charclassloop;
13431 if (RExC_parse >= stop_ptr) {
13436 RExC_parse = regpatws(pRExC_state, RExC_parse,
13437 FALSE /* means don't recognize comments */ );
13440 if (UCHARAT(RExC_parse) == ']') {
13446 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
13447 save_value = value;
13448 save_prevvalue = prevvalue;
13451 rangebegin = RExC_parse;
13455 value = utf8n_to_uvchr((U8*)RExC_parse,
13456 RExC_end - RExC_parse,
13457 &numlen, UTF8_ALLOW_DEFAULT);
13458 RExC_parse += numlen;
13461 value = UCHARAT(RExC_parse++);
13464 && RExC_parse < RExC_end
13465 && POSIXCC(UCHARAT(RExC_parse)))
13467 namedclass = regpposixcc(pRExC_state, value, strict);
13469 else if (value == '\\') {
13471 value = utf8n_to_uvchr((U8*)RExC_parse,
13472 RExC_end - RExC_parse,
13473 &numlen, UTF8_ALLOW_DEFAULT);
13474 RExC_parse += numlen;
13477 value = UCHARAT(RExC_parse++);
13479 /* Some compilers cannot handle switching on 64-bit integer
13480 * values, therefore value cannot be an UV. Yes, this will
13481 * be a problem later if we want switch on Unicode.
13482 * A similar issue a little bit later when switching on
13483 * namedclass. --jhi */
13485 /* If the \ is escaping white space when white space is being
13486 * skipped, it means that that white space is wanted literally, and
13487 * is already in 'value'. Otherwise, need to translate the escape
13488 * into what it signifies. */
13489 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
13491 case 'w': namedclass = ANYOF_WORDCHAR; break;
13492 case 'W': namedclass = ANYOF_NWORDCHAR; break;
13493 case 's': namedclass = ANYOF_SPACE; break;
13494 case 'S': namedclass = ANYOF_NSPACE; break;
13495 case 'd': namedclass = ANYOF_DIGIT; break;
13496 case 'D': namedclass = ANYOF_NDIGIT; break;
13497 case 'v': namedclass = ANYOF_VERTWS; break;
13498 case 'V': namedclass = ANYOF_NVERTWS; break;
13499 case 'h': namedclass = ANYOF_HORIZWS; break;
13500 case 'H': namedclass = ANYOF_NHORIZWS; break;
13501 case 'N': /* Handle \N{NAME} in class */
13503 /* We only pay attention to the first char of
13504 multichar strings being returned. I kinda wonder
13505 if this makes sense as it does change the behaviour
13506 from earlier versions, OTOH that behaviour was broken
13508 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
13509 TRUE, /* => charclass */
13512 if (*flagp & RESTART_UTF8)
13513 FAIL("panic: grok_bslash_N set RESTART_UTF8");
13523 /* We will handle any undefined properties ourselves */
13524 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
13525 /* And we actually would prefer to get
13526 * the straight inversion list of the
13527 * swash, since we will be accessing it
13528 * anyway, to save a little time */
13529 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
13531 if (RExC_parse >= RExC_end)
13532 vFAIL2("Empty \\%c{}", (U8)value);
13533 if (*RExC_parse == '{') {
13534 const U8 c = (U8)value;
13535 e = strchr(RExC_parse++, '}');
13537 vFAIL2("Missing right brace on \\%c{}", c);
13538 while (isSPACE(*RExC_parse))
13540 if (e == RExC_parse)
13541 vFAIL2("Empty \\%c{}", c);
13542 n = e - RExC_parse;
13543 while (isSPACE(*(RExC_parse + n - 1)))
13554 if (UCHARAT(RExC_parse) == '^') {
13557 /* toggle. (The rhs xor gets the single bit that
13558 * differs between P and p; the other xor inverts just
13560 value ^= 'P' ^ 'p';
13562 while (isSPACE(*RExC_parse)) {
13567 /* Try to get the definition of the property into
13568 * <invlist>. If /i is in effect, the effective property
13569 * will have its name be <__NAME_i>. The design is
13570 * discussed in commit
13571 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
13572 name = savepv(Perl_form(aTHX_
13574 (FOLD) ? "__" : "",
13580 /* Look up the property name, and get its swash and
13581 * inversion list, if the property is found */
13583 SvREFCNT_dec_NN(swash);
13585 swash = _core_swash_init("utf8", name, &PL_sv_undef,
13588 NULL, /* No inversion list */
13591 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
13592 HV* curpkg = (IN_PERL_COMPILETIME)
13594 : CopSTASH(PL_curcop);
13596 SvREFCNT_dec_NN(swash);
13600 /* Here didn't find it. It could be a user-defined
13601 * property that will be available at run-time. If we
13602 * accept only compile-time properties, is an error;
13603 * otherwise add it to the list for run-time look up */
13605 RExC_parse = e + 1;
13607 "Property '%"UTF8f"' is unknown",
13608 UTF8fARG(UTF, n, name));
13611 /* If the property name doesn't already have a package
13612 * name, add the current one to it so that it can be
13613 * referred to outside it. [perl #121777] */
13614 if (curpkg && ! instr(name, "::")) {
13615 char* pkgname = HvNAME(curpkg);
13616 if (strNE(pkgname, "main")) {
13617 char* full_name = Perl_form(aTHX_
13621 n = strlen(full_name);
13623 name = savepvn(full_name, n);
13626 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
13627 (value == 'p' ? '+' : '!'),
13628 UTF8fARG(UTF, n, name));
13629 has_user_defined_property = TRUE;
13631 /* We don't know yet, so have to assume that the
13632 * property could match something in the Latin1 range,
13633 * hence something that isn't utf8. Note that this
13634 * would cause things in <depends_list> to match
13635 * inappropriately, except that any \p{}, including
13636 * this one forces Unicode semantics, which means there
13637 * is no <depends_list> */
13638 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
13642 /* Here, did get the swash and its inversion list. If
13643 * the swash is from a user-defined property, then this
13644 * whole character class should be regarded as such */
13645 if (swash_init_flags
13646 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
13648 has_user_defined_property = TRUE;
13651 /* We warn on matching an above-Unicode code point
13652 * if the match would return true, except don't
13653 * warn for \p{All}, which has exactly one element
13655 (_invlist_contains_cp(invlist, 0x110000)
13656 && (! (_invlist_len(invlist) == 1
13657 && *invlist_array(invlist) == 0)))
13663 /* Invert if asking for the complement */
13664 if (value == 'P') {
13665 _invlist_union_complement_2nd(properties,
13669 /* The swash can't be used as-is, because we've
13670 * inverted things; delay removing it to here after
13671 * have copied its invlist above */
13672 SvREFCNT_dec_NN(swash);
13676 _invlist_union(properties, invlist, &properties);
13681 RExC_parse = e + 1;
13682 namedclass = ANYOF_UNIPROP; /* no official name, but it's
13685 /* \p means they want Unicode semantics */
13686 RExC_uni_semantics = 1;
13689 case 'n': value = '\n'; break;
13690 case 'r': value = '\r'; break;
13691 case 't': value = '\t'; break;
13692 case 'f': value = '\f'; break;
13693 case 'b': value = '\b'; break;
13694 case 'e': value = ASCII_TO_NATIVE('\033');break;
13695 case 'a': value = '\a'; break;
13697 RExC_parse--; /* function expects to be pointed at the 'o' */
13699 const char* error_msg;
13700 bool valid = grok_bslash_o(&RExC_parse,
13703 SIZE_ONLY, /* warnings in pass
13706 silence_non_portable,
13712 if (PL_encoding && value < 0x100) {
13713 goto recode_encoding;
13717 RExC_parse--; /* function expects to be pointed at the 'x' */
13719 const char* error_msg;
13720 bool valid = grok_bslash_x(&RExC_parse,
13723 TRUE, /* Output warnings */
13725 silence_non_portable,
13731 if (PL_encoding && value < 0x100)
13732 goto recode_encoding;
13735 value = grok_bslash_c(*RExC_parse++, SIZE_ONLY);
13737 case '0': case '1': case '2': case '3': case '4':
13738 case '5': case '6': case '7':
13740 /* Take 1-3 octal digits */
13741 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
13742 numlen = (strict) ? 4 : 3;
13743 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
13744 RExC_parse += numlen;
13747 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13748 vFAIL("Need exactly 3 octal digits");
13750 else if (! SIZE_ONLY /* like \08, \178 */
13752 && RExC_parse < RExC_end
13753 && isDIGIT(*RExC_parse)
13754 && ckWARN(WARN_REGEXP))
13756 SAVEFREESV(RExC_rx_sv);
13757 reg_warn_non_literal_string(
13759 form_short_octal_warning(RExC_parse, numlen));
13760 (void)ReREFCNT_inc(RExC_rx_sv);
13763 if (PL_encoding && value < 0x100)
13764 goto recode_encoding;
13768 if (! RExC_override_recoding) {
13769 SV* enc = PL_encoding;
13770 value = reg_recode((const char)(U8)value, &enc);
13773 vFAIL("Invalid escape in the specified encoding");
13775 else if (SIZE_ONLY) {
13776 ckWARNreg(RExC_parse,
13777 "Invalid escape in the specified encoding");
13783 /* Allow \_ to not give an error */
13784 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
13786 vFAIL2("Unrecognized escape \\%c in character class",
13790 SAVEFREESV(RExC_rx_sv);
13791 ckWARN2reg(RExC_parse,
13792 "Unrecognized escape \\%c in character class passed through",
13794 (void)ReREFCNT_inc(RExC_rx_sv);
13798 } /* End of switch on char following backslash */
13799 } /* end of handling backslash escape sequences */
13802 literal_endpoint++;
13805 /* Here, we have the current token in 'value' */
13807 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
13810 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
13811 * literal, as is the character that began the false range, i.e.
13812 * the 'a' in the examples */
13815 const int w = (RExC_parse >= rangebegin)
13816 ? RExC_parse - rangebegin
13820 "False [] range \"%"UTF8f"\"",
13821 UTF8fARG(UTF, w, rangebegin));
13824 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
13825 ckWARN2reg(RExC_parse,
13826 "False [] range \"%"UTF8f"\"",
13827 UTF8fARG(UTF, w, rangebegin));
13828 (void)ReREFCNT_inc(RExC_rx_sv);
13829 cp_list = add_cp_to_invlist(cp_list, '-');
13830 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
13835 range = 0; /* this was not a true range */
13836 element_count += 2; /* So counts for three values */
13839 classnum = namedclass_to_classnum(namedclass);
13841 if (LOC && namedclass < ANYOF_POSIXL_MAX
13842 #ifndef HAS_ISASCII
13843 && classnum != _CC_ASCII
13846 /* What the Posix classes (like \w, [:space:]) match in locale
13847 * isn't knowable under locale until actual match time. Room
13848 * must be reserved (one time per outer bracketed class) to
13849 * store such classes. The space will contain a bit for each
13850 * named class that is to be matched against. This isn't
13851 * needed for \p{} and pseudo-classes, as they are not affected
13852 * by locale, and hence are dealt with separately */
13853 if (! need_class) {
13856 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13859 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
13861 ANYOF_FLAGS(ret) |= ANYOF_POSIXL;
13862 ANYOF_POSIXL_ZERO(ret);
13865 /* Coverity thinks it is possible for this to be negative; both
13866 * jhi and khw think it's not, but be safer */
13867 assert(! (ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13868 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
13870 /* See if it already matches the complement of this POSIX
13872 if ((ANYOF_FLAGS(ret) & ANYOF_POSIXL)
13873 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
13877 posixl_matches_all = TRUE;
13878 break; /* No need to continue. Since it matches both
13879 e.g., \w and \W, it matches everything, and the
13880 bracketed class can be optimized into qr/./s */
13883 /* Add this class to those that should be checked at runtime */
13884 ANYOF_POSIXL_SET(ret, namedclass);
13886 /* The above-Latin1 characters are not subject to locale rules.
13887 * Just add them, in the second pass, to the
13888 * unconditionally-matched list */
13890 SV* scratch_list = NULL;
13892 /* Get the list of the above-Latin1 code points this
13894 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
13895 PL_XPosix_ptrs[classnum],
13897 /* Odd numbers are complements, like
13898 * NDIGIT, NASCII, ... */
13899 namedclass % 2 != 0,
13901 /* Checking if 'cp_list' is NULL first saves an extra
13902 * clone. Its reference count will be decremented at the
13903 * next union, etc, or if this is the only instance, at the
13904 * end of the routine */
13906 cp_list = scratch_list;
13909 _invlist_union(cp_list, scratch_list, &cp_list);
13910 SvREFCNT_dec_NN(scratch_list);
13912 continue; /* Go get next character */
13915 else if (! SIZE_ONLY) {
13917 /* Here, not in pass1 (in that pass we skip calculating the
13918 * contents of this class), and is /l, or is a POSIX class for
13919 * which /l doesn't matter (or is a Unicode property, which is
13920 * skipped here). */
13921 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
13922 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
13924 /* Here, should be \h, \H, \v, or \V. None of /d, /i
13925 * nor /l make a difference in what these match,
13926 * therefore we just add what they match to cp_list. */
13927 if (classnum != _CC_VERTSPACE) {
13928 assert( namedclass == ANYOF_HORIZWS
13929 || namedclass == ANYOF_NHORIZWS);
13931 /* It turns out that \h is just a synonym for
13933 classnum = _CC_BLANK;
13936 _invlist_union_maybe_complement_2nd(
13938 PL_XPosix_ptrs[classnum],
13939 namedclass % 2 != 0, /* Complement if odd
13940 (NHORIZWS, NVERTWS)
13945 else { /* Garden variety class. If is NASCII, NDIGIT, ...
13946 complement and use nposixes */
13947 SV** posixes_ptr = namedclass % 2 == 0
13950 SV** source_ptr = &PL_XPosix_ptrs[classnum];
13951 _invlist_union_maybe_complement_2nd(
13954 namedclass % 2 != 0,
13957 continue; /* Go get next character */
13959 } /* end of namedclass \blah */
13961 /* Here, we have a single value. If 'range' is set, it is the ending
13962 * of a range--check its validity. Later, we will handle each
13963 * individual code point in the range. If 'range' isn't set, this
13964 * could be the beginning of a range, so check for that by looking
13965 * ahead to see if the next real character to be processed is the range
13966 * indicator--the minus sign */
13969 RExC_parse = regpatws(pRExC_state, RExC_parse,
13970 FALSE /* means don't recognize comments */ );
13974 if (prevvalue > value) /* b-a */ {
13975 const int w = RExC_parse - rangebegin;
13977 "Invalid [] range \"%"UTF8f"\"",
13978 UTF8fARG(UTF, w, rangebegin));
13979 range = 0; /* not a valid range */
13983 prevvalue = value; /* save the beginning of the potential range */
13984 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13985 && *RExC_parse == '-')
13987 char* next_char_ptr = RExC_parse + 1;
13988 if (skip_white) { /* Get the next real char after the '-' */
13989 next_char_ptr = regpatws(pRExC_state,
13991 FALSE); /* means don't recognize
13995 /* If the '-' is at the end of the class (just before the ']',
13996 * it is a literal minus; otherwise it is a range */
13997 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13998 RExC_parse = next_char_ptr;
14000 /* a bad range like \w-, [:word:]- ? */
14001 if (namedclass > OOB_NAMEDCLASS) {
14002 if (strict || ckWARN(WARN_REGEXP)) {
14004 RExC_parse >= rangebegin ?
14005 RExC_parse - rangebegin : 0;
14007 vFAIL4("False [] range \"%*.*s\"",
14012 "False [] range \"%*.*s\"",
14017 cp_list = add_cp_to_invlist(cp_list, '-');
14021 range = 1; /* yeah, it's a range! */
14022 continue; /* but do it the next time */
14027 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
14030 /* non-Latin1 code point implies unicode semantics. Must be set in
14031 * pass1 so is there for the whole of pass 2 */
14033 RExC_uni_semantics = 1;
14036 /* Ready to process either the single value, or the completed range.
14037 * For single-valued non-inverted ranges, we consider the possibility
14038 * of multi-char folds. (We made a conscious decision to not do this
14039 * for the other cases because it can often lead to non-intuitive
14040 * results. For example, you have the peculiar case that:
14041 * "s s" =~ /^[^\xDF]+$/i => Y
14042 * "ss" =~ /^[^\xDF]+$/i => N
14044 * See [perl #89750] */
14045 if (FOLD && allow_multi_folds && value == prevvalue) {
14046 if (value == LATIN_SMALL_LETTER_SHARP_S
14047 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
14050 /* Here <value> is indeed a multi-char fold. Get what it is */
14052 U8 foldbuf[UTF8_MAXBYTES_CASE];
14055 UV folded = _to_uni_fold_flags(
14059 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
14060 ? FOLD_FLAGS_NOMIX_ASCII
14064 /* Here, <folded> should be the first character of the
14065 * multi-char fold of <value>, with <foldbuf> containing the
14066 * whole thing. But, if this fold is not allowed (because of
14067 * the flags), <fold> will be the same as <value>, and should
14068 * be processed like any other character, so skip the special
14070 if (folded != value) {
14072 /* Skip if we are recursed, currently parsing the class
14073 * again. Otherwise add this character to the list of
14074 * multi-char folds. */
14075 if (! RExC_in_multi_char_class) {
14076 AV** this_array_ptr;
14078 STRLEN cp_count = utf8_length(foldbuf,
14079 foldbuf + foldlen);
14080 SV* multi_fold = sv_2mortal(newSVpvs(""));
14082 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
14085 if (! multi_char_matches) {
14086 multi_char_matches = newAV();
14089 /* <multi_char_matches> is actually an array of arrays.
14090 * There will be one or two top-level elements: [2],
14091 * and/or [3]. The [2] element is an array, each
14092 * element thereof is a character which folds to TWO
14093 * characters; [3] is for folds to THREE characters.
14094 * (Unicode guarantees a maximum of 3 characters in any
14095 * fold.) When we rewrite the character class below,
14096 * we will do so such that the longest folds are
14097 * written first, so that it prefers the longest
14098 * matching strings first. This is done even if it
14099 * turns out that any quantifier is non-greedy, out of
14100 * programmer laziness. Tom Christiansen has agreed
14101 * that this is ok. This makes the test for the
14102 * ligature 'ffi' come before the test for 'ff' */
14103 if (av_exists(multi_char_matches, cp_count)) {
14104 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14106 this_array = *this_array_ptr;
14109 this_array = newAV();
14110 av_store(multi_char_matches, cp_count,
14113 av_push(this_array, multi_fold);
14116 /* This element should not be processed further in this
14119 value = save_value;
14120 prevvalue = save_prevvalue;
14126 /* Deal with this element of the class */
14129 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
14132 SV* this_range = _new_invlist(1);
14133 _append_range_to_invlist(this_range, prevvalue, value);
14135 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
14136 * If this range was specified using something like 'i-j', we want
14137 * to include only the 'i' and the 'j', and not anything in
14138 * between, so exclude non-ASCII, non-alphabetics from it.
14139 * However, if the range was specified with something like
14140 * [\x89-\x91] or [\x89-j], all code points within it should be
14141 * included. literal_endpoint==2 means both ends of the range used
14142 * a literal character, not \x{foo} */
14143 if (literal_endpoint == 2
14144 && ((prevvalue >= 'a' && value <= 'z')
14145 || (prevvalue >= 'A' && value <= 'Z')))
14147 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ASCII],
14150 /* Since this above only contains ascii, the intersection of it
14151 * with anything will still yield only ascii */
14152 _invlist_intersection(this_range, PL_XPosix_ptrs[_CC_ALPHA],
14155 _invlist_union(cp_foldable_list, this_range, &cp_foldable_list);
14156 literal_endpoint = 0;
14160 range = 0; /* this range (if it was one) is done now */
14161 } /* End of loop through all the text within the brackets */
14163 /* If anything in the class expands to more than one character, we have to
14164 * deal with them by building up a substitute parse string, and recursively
14165 * calling reg() on it, instead of proceeding */
14166 if (multi_char_matches) {
14167 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
14170 char *save_end = RExC_end;
14171 char *save_parse = RExC_parse;
14172 bool first_time = TRUE; /* First multi-char occurrence doesn't get
14177 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
14178 because too confusing */
14180 sv_catpv(substitute_parse, "(?:");
14184 /* Look at the longest folds first */
14185 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
14187 if (av_exists(multi_char_matches, cp_count)) {
14188 AV** this_array_ptr;
14191 this_array_ptr = (AV**) av_fetch(multi_char_matches,
14193 while ((this_sequence = av_pop(*this_array_ptr)) !=
14196 if (! first_time) {
14197 sv_catpv(substitute_parse, "|");
14199 first_time = FALSE;
14201 sv_catpv(substitute_parse, SvPVX(this_sequence));
14206 /* If the character class contains anything else besides these
14207 * multi-character folds, have to include it in recursive parsing */
14208 if (element_count) {
14209 sv_catpv(substitute_parse, "|[");
14210 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
14211 sv_catpv(substitute_parse, "]");
14214 sv_catpv(substitute_parse, ")");
14217 /* This is a way to get the parse to skip forward a whole named
14218 * sequence instead of matching the 2nd character when it fails the
14220 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
14224 RExC_parse = SvPV(substitute_parse, len);
14225 RExC_end = RExC_parse + len;
14226 RExC_in_multi_char_class = 1;
14227 RExC_emit = (regnode *)orig_emit;
14229 ret = reg(pRExC_state, 1, ®_flags, depth+1);
14231 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
14233 RExC_parse = save_parse;
14234 RExC_end = save_end;
14235 RExC_in_multi_char_class = 0;
14236 SvREFCNT_dec_NN(multi_char_matches);
14240 /* Here, we've gone through the entire class and dealt with multi-char
14241 * folds. We are now in a position that we can do some checks to see if we
14242 * can optimize this ANYOF node into a simpler one, even in Pass 1.
14243 * Currently we only do two checks:
14244 * 1) is in the unlikely event that the user has specified both, eg. \w and
14245 * \W under /l, then the class matches everything. (This optimization
14246 * is done only to make the optimizer code run later work.)
14247 * 2) if the character class contains only a single element (including a
14248 * single range), we see if there is an equivalent node for it.
14249 * Other checks are possible */
14250 if (! ret_invlist /* Can't optimize if returning the constructed
14252 && (UNLIKELY(posixl_matches_all) || element_count == 1))
14257 if (UNLIKELY(posixl_matches_all)) {
14260 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
14261 \w or [:digit:] or \p{foo}
14264 /* All named classes are mapped into POSIXish nodes, with its FLAG
14265 * argument giving which class it is */
14266 switch ((I32)namedclass) {
14267 case ANYOF_UNIPROP:
14270 /* These don't depend on the charset modifiers. They always
14271 * match under /u rules */
14272 case ANYOF_NHORIZWS:
14273 case ANYOF_HORIZWS:
14274 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
14277 case ANYOF_NVERTWS:
14282 /* The actual POSIXish node for all the rest depends on the
14283 * charset modifier. The ones in the first set depend only on
14284 * ASCII or, if available on this platform, locale */
14288 op = (LOC) ? POSIXL : POSIXA;
14299 /* under /a could be alpha */
14301 if (ASCII_RESTRICTED) {
14302 namedclass = ANYOF_ALPHA + (namedclass % 2);
14310 /* The rest have more possibilities depending on the charset.
14311 * We take advantage of the enum ordering of the charset
14312 * modifiers to get the exact node type, */
14314 op = POSIXD + get_regex_charset(RExC_flags);
14315 if (op > POSIXA) { /* /aa is same as /a */
14320 /* The odd numbered ones are the complements of the
14321 * next-lower even number one */
14322 if (namedclass % 2 == 1) {
14326 arg = namedclass_to_classnum(namedclass);
14330 else if (value == prevvalue) {
14332 /* Here, the class consists of just a single code point */
14335 if (! LOC && value == '\n') {
14336 op = REG_ANY; /* Optimize [^\n] */
14337 *flagp |= HASWIDTH|SIMPLE;
14341 else if (value < 256 || UTF) {
14343 /* Optimize a single value into an EXACTish node, but not if it
14344 * would require converting the pattern to UTF-8. */
14345 op = compute_EXACTish(pRExC_state);
14347 } /* Otherwise is a range */
14348 else if (! LOC) { /* locale could vary these */
14349 if (prevvalue == '0') {
14350 if (value == '9') {
14355 else if (prevvalue == 'A') {
14358 && literal_endpoint == 2
14361 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
14365 else if (prevvalue == 'a') {
14368 && literal_endpoint == 2
14371 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
14377 /* Here, we have changed <op> away from its initial value iff we found
14378 * an optimization */
14381 /* Throw away this ANYOF regnode, and emit the calculated one,
14382 * which should correspond to the beginning, not current, state of
14384 const char * cur_parse = RExC_parse;
14385 RExC_parse = (char *)orig_parse;
14389 /* To get locale nodes to not use the full ANYOF size would
14390 * require moving the code above that writes the portions
14391 * of it that aren't in other nodes to after this point.
14392 * e.g. ANYOF_POSIXL_SET */
14393 RExC_size = orig_size;
14397 RExC_emit = (regnode *)orig_emit;
14398 if (PL_regkind[op] == POSIXD) {
14399 if (op == POSIXL) {
14400 RExC_contains_locale = 1;
14403 op += NPOSIXD - POSIXD;
14408 ret = reg_node(pRExC_state, op);
14410 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
14414 *flagp |= HASWIDTH|SIMPLE;
14416 else if (PL_regkind[op] == EXACT) {
14417 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14418 TRUE /* downgradable to EXACT */
14422 RExC_parse = (char *) cur_parse;
14424 SvREFCNT_dec(posixes);
14425 SvREFCNT_dec(nposixes);
14426 SvREFCNT_dec(cp_list);
14427 SvREFCNT_dec(cp_foldable_list);
14434 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
14436 /* If folding, we calculate all characters that could fold to or from the
14437 * ones already on the list */
14438 if (cp_foldable_list) {
14440 UV start, end; /* End points of code point ranges */
14442 SV* fold_intersection = NULL;
14445 /* Our calculated list will be for Unicode rules. For locale
14446 * matching, we have to keep a separate list that is consulted at
14447 * runtime only when the locale indicates Unicode rules. For
14448 * non-locale, we just use to the general list */
14450 use_list = &only_utf8_locale_list;
14453 use_list = &cp_list;
14456 /* Only the characters in this class that participate in folds need
14457 * be checked. Get the intersection of this class and all the
14458 * possible characters that are foldable. This can quickly narrow
14459 * down a large class */
14460 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
14461 &fold_intersection);
14463 /* The folds for all the Latin1 characters are hard-coded into this
14464 * program, but we have to go out to disk to get the others. */
14465 if (invlist_highest(cp_foldable_list) >= 256) {
14467 /* This is a hash that for a particular fold gives all
14468 * characters that are involved in it */
14469 if (! PL_utf8_foldclosures) {
14470 _load_PL_utf8_foldclosures();
14474 /* Now look at the foldable characters in this class individually */
14475 invlist_iterinit(fold_intersection);
14476 while (invlist_iternext(fold_intersection, &start, &end)) {
14479 /* Look at every character in the range */
14480 for (j = start; j <= end; j++) {
14481 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
14487 if (IS_IN_SOME_FOLD_L1(j)) {
14489 /* ASCII is always matched; non-ASCII is matched
14490 * only under Unicode rules (which could happen
14491 * under /l if the locale is a UTF-8 one */
14492 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
14493 *use_list = add_cp_to_invlist(*use_list,
14494 PL_fold_latin1[j]);
14498 add_cp_to_invlist(depends_list,
14499 PL_fold_latin1[j]);
14503 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
14504 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
14506 add_above_Latin1_folds(pRExC_state,
14513 /* Here is an above Latin1 character. We don't have the
14514 * rules hard-coded for it. First, get its fold. This is
14515 * the simple fold, as the multi-character folds have been
14516 * handled earlier and separated out */
14517 _to_uni_fold_flags(j, foldbuf, &foldlen,
14518 (ASCII_FOLD_RESTRICTED)
14519 ? FOLD_FLAGS_NOMIX_ASCII
14522 /* Single character fold of above Latin1. Add everything in
14523 * its fold closure to the list that this node should match.
14524 * The fold closures data structure is a hash with the keys
14525 * being the UTF-8 of every character that is folded to, like
14526 * 'k', and the values each an array of all code points that
14527 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
14528 * Multi-character folds are not included */
14529 if ((listp = hv_fetch(PL_utf8_foldclosures,
14530 (char *) foldbuf, foldlen, FALSE)))
14532 AV* list = (AV*) *listp;
14534 for (k = 0; k <= av_tindex(list); k++) {
14535 SV** c_p = av_fetch(list, k, FALSE);
14541 /* /aa doesn't allow folds between ASCII and non- */
14542 if ((ASCII_FOLD_RESTRICTED
14543 && (isASCII(c) != isASCII(j))))
14548 /* Folds under /l which cross the 255/256 boundary
14549 * are added to a separate list. (These are valid
14550 * only when the locale is UTF-8.) */
14551 if (c < 256 && LOC) {
14552 *use_list = add_cp_to_invlist(*use_list, c);
14556 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
14558 cp_list = add_cp_to_invlist(cp_list, c);
14561 /* Similarly folds involving non-ascii Latin1
14562 * characters under /d are added to their list */
14563 depends_list = add_cp_to_invlist(depends_list,
14570 SvREFCNT_dec_NN(fold_intersection);
14573 /* Now that we have finished adding all the folds, there is no reason
14574 * to keep the foldable list separate */
14575 _invlist_union(cp_list, cp_foldable_list, &cp_list);
14576 SvREFCNT_dec_NN(cp_foldable_list);
14579 /* And combine the result (if any) with any inversion list from posix
14580 * classes. The lists are kept separate up to now because we don't want to
14581 * fold the classes (folding of those is automatically handled by the swash
14582 * fetching code) */
14583 if (posixes || nposixes) {
14584 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
14585 /* Under /a and /aa, nothing above ASCII matches these */
14586 _invlist_intersection(posixes,
14587 PL_XPosix_ptrs[_CC_ASCII],
14591 if (DEPENDS_SEMANTICS) {
14592 /* Under /d, everything in the upper half of the Latin1 range
14593 * matches these complements */
14594 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_NON_ASCII_ALL;
14596 else if (AT_LEAST_ASCII_RESTRICTED) {
14597 /* Under /a and /aa, everything above ASCII matches these
14599 _invlist_union_complement_2nd(nposixes,
14600 PL_XPosix_ptrs[_CC_ASCII],
14604 _invlist_union(posixes, nposixes, &posixes);
14605 SvREFCNT_dec_NN(nposixes);
14608 posixes = nposixes;
14611 if (! DEPENDS_SEMANTICS) {
14613 _invlist_union(cp_list, posixes, &cp_list);
14614 SvREFCNT_dec_NN(posixes);
14621 /* Under /d, we put into a separate list the Latin1 things that
14622 * match only when the target string is utf8 */
14623 SV* nonascii_but_latin1_properties = NULL;
14624 _invlist_intersection(posixes, PL_UpperLatin1,
14625 &nonascii_but_latin1_properties);
14626 _invlist_subtract(posixes, nonascii_but_latin1_properties,
14629 _invlist_union(cp_list, posixes, &cp_list);
14630 SvREFCNT_dec_NN(posixes);
14636 if (depends_list) {
14637 _invlist_union(depends_list, nonascii_but_latin1_properties,
14639 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
14642 depends_list = nonascii_but_latin1_properties;
14647 /* And combine the result (if any) with any inversion list from properties.
14648 * The lists are kept separate up to now so that we can distinguish the two
14649 * in regards to matching above-Unicode. A run-time warning is generated
14650 * if a Unicode property is matched against a non-Unicode code point. But,
14651 * we allow user-defined properties to match anything, without any warning,
14652 * and we also suppress the warning if there is a portion of the character
14653 * class that isn't a Unicode property, and which matches above Unicode, \W
14654 * or [\x{110000}] for example.
14655 * (Note that in this case, unlike the Posix one above, there is no
14656 * <depends_list>, because having a Unicode property forces Unicode
14661 /* If it matters to the final outcome, see if a non-property
14662 * component of the class matches above Unicode. If so, the
14663 * warning gets suppressed. This is true even if just a single
14664 * such code point is specified, as though not strictly correct if
14665 * another such code point is matched against, the fact that they
14666 * are using above-Unicode code points indicates they should know
14667 * the issues involved */
14669 warn_super = ! (invert
14670 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
14673 _invlist_union(properties, cp_list, &cp_list);
14674 SvREFCNT_dec_NN(properties);
14677 cp_list = properties;
14681 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
14685 /* Here, we have calculated what code points should be in the character
14688 * Now we can see about various optimizations. Fold calculation (which we
14689 * did above) needs to take place before inversion. Otherwise /[^k]/i
14690 * would invert to include K, which under /i would match k, which it
14691 * shouldn't. Therefore we can't invert folded locale now, as it won't be
14692 * folded until runtime */
14694 /* If we didn't do folding, it's because some information isn't available
14695 * until runtime; set the run-time fold flag for these. (We don't have to
14696 * worry about properties folding, as that is taken care of by the swash
14697 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
14698 * locales, or the class matches at least one 0-255 range code point */
14700 if (only_utf8_locale_list) {
14701 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14703 else if (cp_list) { /* Look to see if there a 0-255 code point is in
14706 invlist_iterinit(cp_list);
14707 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
14708 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
14710 invlist_iterfinish(cp_list);
14714 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
14715 * at compile time. Besides not inverting folded locale now, we can't
14716 * invert if there are things such as \w, which aren't known until runtime
14720 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14722 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14724 _invlist_invert(cp_list);
14726 /* Any swash can't be used as-is, because we've inverted things */
14728 SvREFCNT_dec_NN(swash);
14732 /* Clear the invert flag since have just done it here */
14737 *ret_invlist = cp_list;
14738 SvREFCNT_dec(swash);
14740 /* Discard the generated node */
14742 RExC_size = orig_size;
14745 RExC_emit = orig_emit;
14750 /* Some character classes are equivalent to other nodes. Such nodes take
14751 * up less room and generally fewer operations to execute than ANYOF nodes.
14752 * Above, we checked for and optimized into some such equivalents for
14753 * certain common classes that are easy to test. Getting to this point in
14754 * the code means that the class didn't get optimized there. Since this
14755 * code is only executed in Pass 2, it is too late to save space--it has
14756 * been allocated in Pass 1, and currently isn't given back. But turning
14757 * things into an EXACTish node can allow the optimizer to join it to any
14758 * adjacent such nodes. And if the class is equivalent to things like /./,
14759 * expensive run-time swashes can be avoided. Now that we have more
14760 * complete information, we can find things necessarily missed by the
14761 * earlier code. I (khw) am not sure how much to look for here. It would
14762 * be easy, but perhaps too slow, to check any candidates against all the
14763 * node types they could possibly match using _invlistEQ(). */
14768 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
14769 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14771 /* We don't optimize if we are supposed to make sure all non-Unicode
14772 * code points raise a warning, as only ANYOF nodes have this check.
14774 && ! ((ANYOF_FLAGS(ret) & ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
14777 U8 op = END; /* The optimzation node-type */
14778 const char * cur_parse= RExC_parse;
14780 invlist_iterinit(cp_list);
14781 if (! invlist_iternext(cp_list, &start, &end)) {
14783 /* Here, the list is empty. This happens, for example, when a
14784 * Unicode property is the only thing in the character class, and
14785 * it doesn't match anything. (perluniprops.pod notes such
14788 *flagp |= HASWIDTH|SIMPLE;
14790 else if (start == end) { /* The range is a single code point */
14791 if (! invlist_iternext(cp_list, &start, &end)
14793 /* Don't do this optimization if it would require changing
14794 * the pattern to UTF-8 */
14795 && (start < 256 || UTF))
14797 /* Here, the list contains a single code point. Can optimize
14798 * into an EXACTish node */
14807 /* A locale node under folding with one code point can be
14808 * an EXACTFL, as its fold won't be calculated until
14814 /* Here, we are generally folding, but there is only one
14815 * code point to match. If we have to, we use an EXACT
14816 * node, but it would be better for joining with adjacent
14817 * nodes in the optimization pass if we used the same
14818 * EXACTFish node that any such are likely to be. We can
14819 * do this iff the code point doesn't participate in any
14820 * folds. For example, an EXACTF of a colon is the same as
14821 * an EXACT one, since nothing folds to or from a colon. */
14823 if (IS_IN_SOME_FOLD_L1(value)) {
14828 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
14833 /* If we haven't found the node type, above, it means we
14834 * can use the prevailing one */
14836 op = compute_EXACTish(pRExC_state);
14841 else if (start == 0) {
14842 if (end == UV_MAX) {
14844 *flagp |= HASWIDTH|SIMPLE;
14847 else if (end == '\n' - 1
14848 && invlist_iternext(cp_list, &start, &end)
14849 && start == '\n' + 1 && end == UV_MAX)
14852 *flagp |= HASWIDTH|SIMPLE;
14856 invlist_iterfinish(cp_list);
14859 RExC_parse = (char *)orig_parse;
14860 RExC_emit = (regnode *)orig_emit;
14862 ret = reg_node(pRExC_state, op);
14864 RExC_parse = (char *)cur_parse;
14866 if (PL_regkind[op] == EXACT) {
14867 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
14868 TRUE /* downgradable to EXACT */
14872 SvREFCNT_dec_NN(cp_list);
14877 /* Here, <cp_list> contains all the code points we can determine at
14878 * compile time that match under all conditions. Go through it, and
14879 * for things that belong in the bitmap, put them there, and delete from
14880 * <cp_list>. While we are at it, see if everything above 255 is in the
14881 * list, and if so, set a flag to speed up execution */
14883 populate_ANYOF_from_invlist(ret, &cp_list);
14886 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
14889 /* Here, the bitmap has been populated with all the Latin1 code points that
14890 * always match. Can now add to the overall list those that match only
14891 * when the target string is UTF-8 (<depends_list>). */
14892 if (depends_list) {
14894 _invlist_union(cp_list, depends_list, &cp_list);
14895 SvREFCNT_dec_NN(depends_list);
14898 cp_list = depends_list;
14900 ANYOF_FLAGS(ret) |= ANYOF_UTF8;
14903 /* If there is a swash and more than one element, we can't use the swash in
14904 * the optimization below. */
14905 if (swash && element_count > 1) {
14906 SvREFCNT_dec_NN(swash);
14910 set_ANYOF_arg(pRExC_state, ret, cp_list,
14911 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14913 only_utf8_locale_list,
14914 swash, has_user_defined_property);
14916 *flagp |= HASWIDTH|SIMPLE;
14918 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
14919 RExC_contains_locale = 1;
14925 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14928 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
14929 regnode* const node,
14931 SV* const runtime_defns,
14932 SV* const only_utf8_locale_list,
14934 const bool has_user_defined_property)
14936 /* Sets the arg field of an ANYOF-type node 'node', using information about
14937 * the node passed-in. If there is nothing outside the node's bitmap, the
14938 * arg is set to ANYOF_NONBITMAP_EMPTY. Otherwise, it sets the argument to
14939 * the count returned by add_data(), having allocated and stored an array,
14940 * av, that that count references, as follows:
14941 * av[0] stores the character class description in its textual form.
14942 * This is used later (regexec.c:Perl_regclass_swash()) to
14943 * initialize the appropriate swash, and is also useful for dumping
14944 * the regnode. This is set to &PL_sv_undef if the textual
14945 * description is not needed at run-time (as happens if the other
14946 * elements completely define the class)
14947 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
14948 * computed from av[0]. But if no further computation need be done,
14949 * the swash is stored here now (and av[0] is &PL_sv_undef).
14950 * av[2] stores the inversion list of code points that match only if the
14951 * current locale is UTF-8
14952 * av[3] stores the cp_list inversion list for use in addition or instead
14953 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
14954 * (Otherwise everything needed is already in av[0] and av[1])
14955 * av[4] is set if any component of the class is from a user-defined
14956 * property; used only if av[3] exists */
14960 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
14962 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
14963 assert(! (ANYOF_FLAGS(node)
14964 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8)));
14965 ARG_SET(node, ANYOF_NONBITMAP_EMPTY);
14968 AV * const av = newAV();
14971 assert(ANYOF_FLAGS(node)
14972 & (ANYOF_UTF8|ANYOF_NONBITMAP_NON_UTF8|ANYOF_LOC_FOLD));
14974 av_store(av, 0, (runtime_defns)
14975 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
14978 av_store(av, 1, swash);
14979 SvREFCNT_dec_NN(cp_list);
14982 av_store(av, 1, &PL_sv_undef);
14984 av_store(av, 3, cp_list);
14985 av_store(av, 4, newSVuv(has_user_defined_property));
14989 if (only_utf8_locale_list) {
14990 av_store(av, 2, only_utf8_locale_list);
14993 av_store(av, 2, &PL_sv_undef);
14996 rv = newRV_noinc(MUTABLE_SV(av));
14997 n = add_data(pRExC_state, STR_WITH_LEN("s"));
14998 RExC_rxi->data->data[n] = (void*)rv;
15004 /* reg_skipcomment()
15006 Absorbs an /x style # comment from the input stream,
15007 returning a pointer to the first character beyond the comment, or if the
15008 comment terminates the pattern without anything following it, this returns
15009 one past the final character of the pattern (in other words, RExC_end) and
15010 sets the REG_RUN_ON_COMMENT_SEEN flag.
15012 Note it's the callers responsibility to ensure that we are
15013 actually in /x mode
15017 PERL_STATIC_INLINE char*
15018 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
15020 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
15024 while (p < RExC_end) {
15025 if (*(++p) == '\n') {
15030 /* we ran off the end of the pattern without ending the comment, so we have
15031 * to add an \n when wrapping */
15032 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
15038 Advances the parse position, and optionally absorbs
15039 "whitespace" from the inputstream.
15041 Without /x "whitespace" means (?#...) style comments only,
15042 with /x this means (?#...) and # comments and whitespace proper.
15044 Returns the RExC_parse point from BEFORE the scan occurs.
15046 This is the /x friendly way of saying RExC_parse++.
15050 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
15052 char* const retval = RExC_parse++;
15054 PERL_ARGS_ASSERT_NEXTCHAR;
15057 if (RExC_end - RExC_parse >= 3
15058 && *RExC_parse == '('
15059 && RExC_parse[1] == '?'
15060 && RExC_parse[2] == '#')
15062 while (*RExC_parse != ')') {
15063 if (RExC_parse == RExC_end)
15064 FAIL("Sequence (?#... not terminated");
15070 if (RExC_flags & RXf_PMf_EXTENDED) {
15071 char * p = regpatws(pRExC_state, RExC_parse,
15072 TRUE); /* means recognize comments */
15073 if (p != RExC_parse) {
15083 - reg_node - emit a node
15085 STATIC regnode * /* Location. */
15086 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
15089 regnode * const ret = RExC_emit;
15090 GET_RE_DEBUG_FLAGS_DECL;
15092 PERL_ARGS_ASSERT_REG_NODE;
15095 SIZE_ALIGN(RExC_size);
15099 if (RExC_emit >= RExC_emit_bound)
15100 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15101 op, (void*)RExC_emit, (void*)RExC_emit_bound);
15103 NODE_ALIGN_FILL(ret);
15105 FILL_ADVANCE_NODE(ptr, op);
15106 #ifdef RE_TRACK_PATTERN_OFFSETS
15107 if (RExC_offsets) { /* MJD */
15109 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
15110 "reg_node", __LINE__,
15112 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
15113 ? "Overwriting end of array!\n" : "OK",
15114 (UV)(RExC_emit - RExC_emit_start),
15115 (UV)(RExC_parse - RExC_start),
15116 (UV)RExC_offsets[0]));
15117 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
15125 - reganode - emit a node with an argument
15127 STATIC regnode * /* Location. */
15128 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
15131 regnode * const ret = RExC_emit;
15132 GET_RE_DEBUG_FLAGS_DECL;
15134 PERL_ARGS_ASSERT_REGANODE;
15137 SIZE_ALIGN(RExC_size);
15142 assert(2==regarglen[op]+1);
15144 Anything larger than this has to allocate the extra amount.
15145 If we changed this to be:
15147 RExC_size += (1 + regarglen[op]);
15149 then it wouldn't matter. Its not clear what side effect
15150 might come from that so its not done so far.
15155 if (RExC_emit >= RExC_emit_bound)
15156 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
15157 op, (void*)RExC_emit, (void*)RExC_emit_bound);
15159 NODE_ALIGN_FILL(ret);
15161 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
15162 #ifdef RE_TRACK_PATTERN_OFFSETS
15163 if (RExC_offsets) { /* MJD */
15165 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15169 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
15170 "Overwriting end of array!\n" : "OK",
15171 (UV)(RExC_emit - RExC_emit_start),
15172 (UV)(RExC_parse - RExC_start),
15173 (UV)RExC_offsets[0]));
15174 Set_Cur_Node_Offset;
15182 - reguni - emit (if appropriate) a Unicode character
15184 PERL_STATIC_INLINE STRLEN
15185 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
15187 PERL_ARGS_ASSERT_REGUNI;
15189 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
15193 - reginsert - insert an operator in front of already-emitted operand
15195 * Means relocating the operand.
15198 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
15203 const int offset = regarglen[(U8)op];
15204 const int size = NODE_STEP_REGNODE + offset;
15205 GET_RE_DEBUG_FLAGS_DECL;
15207 PERL_ARGS_ASSERT_REGINSERT;
15208 PERL_UNUSED_CONTEXT;
15209 PERL_UNUSED_ARG(depth);
15210 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
15211 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
15220 if (RExC_open_parens) {
15222 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
15223 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
15224 if ( RExC_open_parens[paren] >= opnd ) {
15225 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
15226 RExC_open_parens[paren] += size;
15228 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
15230 if ( RExC_close_parens[paren] >= opnd ) {
15231 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
15232 RExC_close_parens[paren] += size;
15234 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
15239 while (src > opnd) {
15240 StructCopy(--src, --dst, regnode);
15241 #ifdef RE_TRACK_PATTERN_OFFSETS
15242 if (RExC_offsets) { /* MJD 20010112 */
15244 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
15248 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
15249 ? "Overwriting end of array!\n" : "OK",
15250 (UV)(src - RExC_emit_start),
15251 (UV)(dst - RExC_emit_start),
15252 (UV)RExC_offsets[0]));
15253 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
15254 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
15260 place = opnd; /* Op node, where operand used to be. */
15261 #ifdef RE_TRACK_PATTERN_OFFSETS
15262 if (RExC_offsets) { /* MJD */
15264 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
15268 (UV)(place - RExC_emit_start) > RExC_offsets[0]
15269 ? "Overwriting end of array!\n" : "OK",
15270 (UV)(place - RExC_emit_start),
15271 (UV)(RExC_parse - RExC_start),
15272 (UV)RExC_offsets[0]));
15273 Set_Node_Offset(place, RExC_parse);
15274 Set_Node_Length(place, 1);
15277 src = NEXTOPER(place);
15278 FILL_ADVANCE_NODE(place, op);
15279 Zero(src, offset, regnode);
15283 - regtail - set the next-pointer at the end of a node chain of p to val.
15284 - SEE ALSO: regtail_study
15286 /* TODO: All three parms should be const */
15288 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15289 const regnode *val,U32 depth)
15292 GET_RE_DEBUG_FLAGS_DECL;
15294 PERL_ARGS_ASSERT_REGTAIL;
15296 PERL_UNUSED_ARG(depth);
15302 /* Find last node. */
15305 regnode * const temp = regnext(scan);
15307 SV * const mysv=sv_newmortal();
15308 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
15309 regprop(RExC_rx, mysv, scan, NULL);
15310 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
15311 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
15312 (temp == NULL ? "->" : ""),
15313 (temp == NULL ? PL_reg_name[OP(val)] : "")
15321 if (reg_off_by_arg[OP(scan)]) {
15322 ARG_SET(scan, val - scan);
15325 NEXT_OFF(scan) = val - scan;
15331 - regtail_study - set the next-pointer at the end of a node chain of p to val.
15332 - Look for optimizable sequences at the same time.
15333 - currently only looks for EXACT chains.
15335 This is experimental code. The idea is to use this routine to perform
15336 in place optimizations on branches and groups as they are constructed,
15337 with the long term intention of removing optimization from study_chunk so
15338 that it is purely analytical.
15340 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
15341 to control which is which.
15344 /* TODO: All four parms should be const */
15347 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
15348 const regnode *val,U32 depth)
15353 #ifdef EXPERIMENTAL_INPLACESCAN
15356 GET_RE_DEBUG_FLAGS_DECL;
15358 PERL_ARGS_ASSERT_REGTAIL_STUDY;
15364 /* Find last node. */
15368 regnode * const temp = regnext(scan);
15369 #ifdef EXPERIMENTAL_INPLACESCAN
15370 if (PL_regkind[OP(scan)] == EXACT) {
15371 bool unfolded_multi_char; /* Unexamined in this routine */
15372 if (join_exact(pRExC_state, scan, &min,
15373 &unfolded_multi_char, 1, val, depth+1))
15378 switch (OP(scan)) {
15381 case EXACTFA_NO_TRIE:
15386 if( exact == PSEUDO )
15388 else if ( exact != OP(scan) )
15397 SV * const mysv=sv_newmortal();
15398 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
15399 regprop(RExC_rx, mysv, scan, NULL);
15400 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
15401 SvPV_nolen_const(mysv),
15402 REG_NODE_NUM(scan),
15403 PL_reg_name[exact]);
15410 SV * const mysv_val=sv_newmortal();
15411 DEBUG_PARSE_MSG("");
15412 regprop(RExC_rx, mysv_val, val, NULL);
15413 PerlIO_printf(Perl_debug_log,
15414 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
15415 SvPV_nolen_const(mysv_val),
15416 (IV)REG_NODE_NUM(val),
15420 if (reg_off_by_arg[OP(scan)]) {
15421 ARG_SET(scan, val - scan);
15424 NEXT_OFF(scan) = val - scan;
15432 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
15437 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
15442 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15444 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
15445 if (flags & (1<<bit)) {
15446 if (!set++ && lead)
15447 PerlIO_printf(Perl_debug_log, "%s",lead);
15448 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
15453 PerlIO_printf(Perl_debug_log, "\n");
15455 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15460 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
15466 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
15468 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
15469 if (flags & (1<<bit)) {
15470 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
15473 if (!set++ && lead)
15474 PerlIO_printf(Perl_debug_log, "%s",lead);
15475 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
15478 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
15479 if (!set++ && lead) {
15480 PerlIO_printf(Perl_debug_log, "%s",lead);
15483 case REGEX_UNICODE_CHARSET:
15484 PerlIO_printf(Perl_debug_log, "UNICODE");
15486 case REGEX_LOCALE_CHARSET:
15487 PerlIO_printf(Perl_debug_log, "LOCALE");
15489 case REGEX_ASCII_RESTRICTED_CHARSET:
15490 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
15492 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
15493 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
15496 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
15502 PerlIO_printf(Perl_debug_log, "\n");
15504 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
15510 Perl_regdump(pTHX_ const regexp *r)
15514 SV * const sv = sv_newmortal();
15515 SV *dsv= sv_newmortal();
15516 RXi_GET_DECL(r,ri);
15517 GET_RE_DEBUG_FLAGS_DECL;
15519 PERL_ARGS_ASSERT_REGDUMP;
15521 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
15523 /* Header fields of interest. */
15524 if (r->anchored_substr) {
15525 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
15526 RE_SV_DUMPLEN(r->anchored_substr), 30);
15527 PerlIO_printf(Perl_debug_log,
15528 "anchored %s%s at %"IVdf" ",
15529 s, RE_SV_TAIL(r->anchored_substr),
15530 (IV)r->anchored_offset);
15531 } else if (r->anchored_utf8) {
15532 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
15533 RE_SV_DUMPLEN(r->anchored_utf8), 30);
15534 PerlIO_printf(Perl_debug_log,
15535 "anchored utf8 %s%s at %"IVdf" ",
15536 s, RE_SV_TAIL(r->anchored_utf8),
15537 (IV)r->anchored_offset);
15539 if (r->float_substr) {
15540 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
15541 RE_SV_DUMPLEN(r->float_substr), 30);
15542 PerlIO_printf(Perl_debug_log,
15543 "floating %s%s at %"IVdf"..%"UVuf" ",
15544 s, RE_SV_TAIL(r->float_substr),
15545 (IV)r->float_min_offset, (UV)r->float_max_offset);
15546 } else if (r->float_utf8) {
15547 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
15548 RE_SV_DUMPLEN(r->float_utf8), 30);
15549 PerlIO_printf(Perl_debug_log,
15550 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
15551 s, RE_SV_TAIL(r->float_utf8),
15552 (IV)r->float_min_offset, (UV)r->float_max_offset);
15554 if (r->check_substr || r->check_utf8)
15555 PerlIO_printf(Perl_debug_log,
15557 (r->check_substr == r->float_substr
15558 && r->check_utf8 == r->float_utf8
15559 ? "(checking floating" : "(checking anchored"));
15560 if (r->intflags & PREGf_NOSCAN)
15561 PerlIO_printf(Perl_debug_log, " noscan");
15562 if (r->extflags & RXf_CHECK_ALL)
15563 PerlIO_printf(Perl_debug_log, " isall");
15564 if (r->check_substr || r->check_utf8)
15565 PerlIO_printf(Perl_debug_log, ") ");
15567 if (ri->regstclass) {
15568 regprop(r, sv, ri->regstclass, NULL);
15569 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
15571 if (r->intflags & PREGf_ANCH) {
15572 PerlIO_printf(Perl_debug_log, "anchored");
15573 if (r->intflags & PREGf_ANCH_BOL)
15574 PerlIO_printf(Perl_debug_log, "(BOL)");
15575 if (r->intflags & PREGf_ANCH_MBOL)
15576 PerlIO_printf(Perl_debug_log, "(MBOL)");
15577 if (r->intflags & PREGf_ANCH_SBOL)
15578 PerlIO_printf(Perl_debug_log, "(SBOL)");
15579 if (r->intflags & PREGf_ANCH_GPOS)
15580 PerlIO_printf(Perl_debug_log, "(GPOS)");
15581 PerlIO_putc(Perl_debug_log, ' ');
15583 if (r->intflags & PREGf_GPOS_SEEN)
15584 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
15585 if (r->intflags & PREGf_SKIP)
15586 PerlIO_printf(Perl_debug_log, "plus ");
15587 if (r->intflags & PREGf_IMPLICIT)
15588 PerlIO_printf(Perl_debug_log, "implicit ");
15589 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
15590 if (r->extflags & RXf_EVAL_SEEN)
15591 PerlIO_printf(Perl_debug_log, "with eval ");
15592 PerlIO_printf(Perl_debug_log, "\n");
15594 regdump_extflags("r->extflags: ",r->extflags);
15595 regdump_intflags("r->intflags: ",r->intflags);
15598 PERL_ARGS_ASSERT_REGDUMP;
15599 PERL_UNUSED_CONTEXT;
15600 PERL_UNUSED_ARG(r);
15601 #endif /* DEBUGGING */
15605 - regprop - printable representation of opcode, with run time support
15609 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo)
15615 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
15616 static const char * const anyofs[] = {
15617 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
15618 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
15619 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
15620 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
15621 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
15622 || _CC_VERTSPACE != 16
15623 #error Need to adjust order of anyofs[]
15660 RXi_GET_DECL(prog,progi);
15661 GET_RE_DEBUG_FLAGS_DECL;
15663 PERL_ARGS_ASSERT_REGPROP;
15667 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
15668 /* It would be nice to FAIL() here, but this may be called from
15669 regexec.c, and it would be hard to supply pRExC_state. */
15670 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
15671 (int)OP(o), (int)REGNODE_MAX);
15672 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
15674 k = PL_regkind[OP(o)];
15677 sv_catpvs(sv, " ");
15678 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
15679 * is a crude hack but it may be the best for now since
15680 * we have no flag "this EXACTish node was UTF-8"
15682 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
15683 PERL_PV_ESCAPE_UNI_DETECT |
15684 PERL_PV_ESCAPE_NONASCII |
15685 PERL_PV_PRETTY_ELLIPSES |
15686 PERL_PV_PRETTY_LTGT |
15687 PERL_PV_PRETTY_NOCLEAR
15689 } else if (k == TRIE) {
15690 /* print the details of the trie in dumpuntil instead, as
15691 * progi->data isn't available here */
15692 const char op = OP(o);
15693 const U32 n = ARG(o);
15694 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
15695 (reg_ac_data *)progi->data->data[n] :
15697 const reg_trie_data * const trie
15698 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
15700 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
15701 DEBUG_TRIE_COMPILE_r(
15702 Perl_sv_catpvf(aTHX_ sv,
15703 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
15704 (UV)trie->startstate,
15705 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
15706 (UV)trie->wordcount,
15709 (UV)TRIE_CHARCOUNT(trie),
15710 (UV)trie->uniquecharcount
15713 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
15714 sv_catpvs(sv, "[");
15715 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
15717 : TRIE_BITMAP(trie));
15718 sv_catpvs(sv, "]");
15721 } else if (k == CURLY) {
15722 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
15723 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
15724 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
15726 else if (k == WHILEM && o->flags) /* Ordinal/of */
15727 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
15728 else if (k == REF || k == OPEN || k == CLOSE
15729 || k == GROUPP || OP(o)==ACCEPT)
15731 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
15732 if ( RXp_PAREN_NAMES(prog) ) {
15733 if ( k != REF || (OP(o) < NREF)) {
15734 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
15735 SV **name= av_fetch(list, ARG(o), 0 );
15737 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15740 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
15741 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
15742 I32 *nums=(I32*)SvPVX(sv_dat);
15743 SV **name= av_fetch(list, nums[0], 0 );
15746 for ( n=0; n<SvIVX(sv_dat); n++ ) {
15747 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
15748 (n ? "," : ""), (IV)nums[n]);
15750 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
15754 if ( k == REF && reginfo) {
15755 U32 n = ARG(o); /* which paren pair */
15756 I32 ln = prog->offs[n].start;
15757 if (prog->lastparen < n || ln == -1)
15758 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
15759 else if (ln == prog->offs[n].end)
15760 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
15762 const char *s = reginfo->strbeg + ln;
15763 Perl_sv_catpvf(aTHX_ sv, ": ");
15764 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
15765 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
15768 } else if (k == GOSUB)
15769 /* Paren and offset */
15770 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
15771 else if (k == VERB) {
15773 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
15774 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
15775 } else if (k == LOGICAL)
15776 /* 2: embedded, otherwise 1 */
15777 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
15778 else if (k == ANYOF) {
15779 const U8 flags = ANYOF_FLAGS(o);
15783 if (flags & ANYOF_LOCALE_FLAGS)
15784 sv_catpvs(sv, "{loc}");
15785 if (flags & ANYOF_LOC_FOLD)
15786 sv_catpvs(sv, "{i}");
15787 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
15788 if (flags & ANYOF_INVERT)
15789 sv_catpvs(sv, "^");
15791 /* output what the standard cp 0-255 bitmap matches */
15792 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
15794 /* output any special charclass tests (used entirely under use
15796 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
15798 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
15799 if (ANYOF_POSIXL_TEST(o,i)) {
15800 sv_catpv(sv, anyofs[i]);
15806 if ((flags & (ANYOF_ABOVE_LATIN1_ALL
15808 |ANYOF_NONBITMAP_NON_UTF8
15812 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
15813 if (flags & ANYOF_INVERT)
15814 /*make sure the invert info is in each */
15815 sv_catpvs(sv, "^");
15818 if (flags & ANYOF_NON_UTF8_NON_ASCII_ALL) {
15819 sv_catpvs(sv, "{non-utf8-latin1-all}");
15822 /* output information about the unicode matching */
15823 if (flags & ANYOF_ABOVE_LATIN1_ALL)
15824 sv_catpvs(sv, "{unicode_all}");
15825 else if (ARG(o) != ANYOF_NONBITMAP_EMPTY) {
15826 SV *lv; /* Set if there is something outside the bit map. */
15827 bool byte_output = FALSE; /* If something in the bitmap has
15829 SV *only_utf8_locale;
15831 /* Get the stuff that wasn't in the bitmap */
15832 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
15833 &lv, &only_utf8_locale);
15834 if (lv && lv != &PL_sv_undef) {
15835 char *s = savesvpv(lv);
15836 char * const origs = s;
15838 while (*s && *s != '\n')
15842 const char * const t = ++s;
15844 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
15845 sv_catpvs(sv, "{outside bitmap}");
15848 sv_catpvs(sv, "{utf8}");
15852 sv_catpvs(sv, " ");
15858 /* Truncate very long output */
15859 if (s - origs > 256) {
15860 Perl_sv_catpvf(aTHX_ sv,
15862 (int) (s - origs - 1),
15868 else if (*s == '\t') {
15882 SvREFCNT_dec_NN(lv);
15885 if ((flags & ANYOF_LOC_FOLD)
15886 && only_utf8_locale
15887 && only_utf8_locale != &PL_sv_undef)
15890 int max_entries = 256;
15892 sv_catpvs(sv, "{utf8 locale}");
15893 invlist_iterinit(only_utf8_locale);
15894 while (invlist_iternext(only_utf8_locale,
15896 put_range(sv, start, end);
15898 if (max_entries < 0) {
15899 sv_catpvs(sv, "...");
15903 invlist_iterfinish(only_utf8_locale);
15908 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
15910 else if (k == POSIXD || k == NPOSIXD) {
15911 U8 index = FLAGS(o) * 2;
15912 if (index < C_ARRAY_LENGTH(anyofs)) {
15913 if (*anyofs[index] != '[') {
15916 sv_catpv(sv, anyofs[index]);
15917 if (*anyofs[index] != '[') {
15922 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
15925 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
15926 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
15928 PERL_UNUSED_CONTEXT;
15929 PERL_UNUSED_ARG(sv);
15930 PERL_UNUSED_ARG(o);
15931 PERL_UNUSED_ARG(prog);
15932 PERL_UNUSED_ARG(reginfo);
15933 #endif /* DEBUGGING */
15939 Perl_re_intuit_string(pTHX_ REGEXP * const r)
15940 { /* Assume that RE_INTUIT is set */
15941 struct regexp *const prog = ReANY(r);
15942 GET_RE_DEBUG_FLAGS_DECL;
15944 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
15945 PERL_UNUSED_CONTEXT;
15949 const char * const s = SvPV_nolen_const(prog->check_substr
15950 ? prog->check_substr : prog->check_utf8);
15952 if (!PL_colorset) reginitcolors();
15953 PerlIO_printf(Perl_debug_log,
15954 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
15956 prog->check_substr ? "" : "utf8 ",
15957 PL_colors[5],PL_colors[0],
15960 (strlen(s) > 60 ? "..." : ""));
15963 return prog->check_substr ? prog->check_substr : prog->check_utf8;
15969 handles refcounting and freeing the perl core regexp structure. When
15970 it is necessary to actually free the structure the first thing it
15971 does is call the 'free' method of the regexp_engine associated to
15972 the regexp, allowing the handling of the void *pprivate; member
15973 first. (This routine is not overridable by extensions, which is why
15974 the extensions free is called first.)
15976 See regdupe and regdupe_internal if you change anything here.
15978 #ifndef PERL_IN_XSUB_RE
15980 Perl_pregfree(pTHX_ REGEXP *r)
15986 Perl_pregfree2(pTHX_ REGEXP *rx)
15988 struct regexp *const r = ReANY(rx);
15989 GET_RE_DEBUG_FLAGS_DECL;
15991 PERL_ARGS_ASSERT_PREGFREE2;
15993 if (r->mother_re) {
15994 ReREFCNT_dec(r->mother_re);
15996 CALLREGFREE_PVT(rx); /* free the private data */
15997 SvREFCNT_dec(RXp_PAREN_NAMES(r));
15998 Safefree(r->xpv_len_u.xpvlenu_pv);
16001 SvREFCNT_dec(r->anchored_substr);
16002 SvREFCNT_dec(r->anchored_utf8);
16003 SvREFCNT_dec(r->float_substr);
16004 SvREFCNT_dec(r->float_utf8);
16005 Safefree(r->substrs);
16007 RX_MATCH_COPY_FREE(rx);
16008 #ifdef PERL_ANY_COW
16009 SvREFCNT_dec(r->saved_copy);
16012 SvREFCNT_dec(r->qr_anoncv);
16013 rx->sv_u.svu_rx = 0;
16018 This is a hacky workaround to the structural issue of match results
16019 being stored in the regexp structure which is in turn stored in
16020 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
16021 could be PL_curpm in multiple contexts, and could require multiple
16022 result sets being associated with the pattern simultaneously, such
16023 as when doing a recursive match with (??{$qr})
16025 The solution is to make a lightweight copy of the regexp structure
16026 when a qr// is returned from the code executed by (??{$qr}) this
16027 lightweight copy doesn't actually own any of its data except for
16028 the starp/end and the actual regexp structure itself.
16034 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
16036 struct regexp *ret;
16037 struct regexp *const r = ReANY(rx);
16038 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
16040 PERL_ARGS_ASSERT_REG_TEMP_COPY;
16043 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
16045 SvOK_off((SV *)ret_x);
16047 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
16048 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
16049 made both spots point to the same regexp body.) */
16050 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
16051 assert(!SvPVX(ret_x));
16052 ret_x->sv_u.svu_rx = temp->sv_any;
16053 temp->sv_any = NULL;
16054 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
16055 SvREFCNT_dec_NN(temp);
16056 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
16057 ing below will not set it. */
16058 SvCUR_set(ret_x, SvCUR(rx));
16061 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
16062 sv_force_normal(sv) is called. */
16064 ret = ReANY(ret_x);
16066 SvFLAGS(ret_x) |= SvUTF8(rx);
16067 /* We share the same string buffer as the original regexp, on which we
16068 hold a reference count, incremented when mother_re is set below.
16069 The string pointer is copied here, being part of the regexp struct.
16071 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
16072 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
16074 const I32 npar = r->nparens+1;
16075 Newx(ret->offs, npar, regexp_paren_pair);
16076 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16079 Newx(ret->substrs, 1, struct reg_substr_data);
16080 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16082 SvREFCNT_inc_void(ret->anchored_substr);
16083 SvREFCNT_inc_void(ret->anchored_utf8);
16084 SvREFCNT_inc_void(ret->float_substr);
16085 SvREFCNT_inc_void(ret->float_utf8);
16087 /* check_substr and check_utf8, if non-NULL, point to either their
16088 anchored or float namesakes, and don't hold a second reference. */
16090 RX_MATCH_COPIED_off(ret_x);
16091 #ifdef PERL_ANY_COW
16092 ret->saved_copy = NULL;
16094 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
16095 SvREFCNT_inc_void(ret->qr_anoncv);
16101 /* regfree_internal()
16103 Free the private data in a regexp. This is overloadable by
16104 extensions. Perl takes care of the regexp structure in pregfree(),
16105 this covers the *pprivate pointer which technically perl doesn't
16106 know about, however of course we have to handle the
16107 regexp_internal structure when no extension is in use.
16109 Note this is called before freeing anything in the regexp
16114 Perl_regfree_internal(pTHX_ REGEXP * const rx)
16116 struct regexp *const r = ReANY(rx);
16117 RXi_GET_DECL(r,ri);
16118 GET_RE_DEBUG_FLAGS_DECL;
16120 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
16126 SV *dsv= sv_newmortal();
16127 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
16128 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
16129 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
16130 PL_colors[4],PL_colors[5],s);
16133 #ifdef RE_TRACK_PATTERN_OFFSETS
16135 Safefree(ri->u.offsets); /* 20010421 MJD */
16137 if (ri->code_blocks) {
16139 for (n = 0; n < ri->num_code_blocks; n++)
16140 SvREFCNT_dec(ri->code_blocks[n].src_regex);
16141 Safefree(ri->code_blocks);
16145 int n = ri->data->count;
16148 /* If you add a ->what type here, update the comment in regcomp.h */
16149 switch (ri->data->what[n]) {
16155 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
16158 Safefree(ri->data->data[n]);
16164 { /* Aho Corasick add-on structure for a trie node.
16165 Used in stclass optimization only */
16167 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
16168 #ifdef USE_ITHREADS
16172 refcount = --aho->refcount;
16175 PerlMemShared_free(aho->states);
16176 PerlMemShared_free(aho->fail);
16177 /* do this last!!!! */
16178 PerlMemShared_free(ri->data->data[n]);
16179 /* we should only ever get called once, so
16180 * assert as much, and also guard the free
16181 * which /might/ happen twice. At the least
16182 * it will make code anlyzers happy and it
16183 * doesn't cost much. - Yves */
16184 assert(ri->regstclass);
16185 if (ri->regstclass) {
16186 PerlMemShared_free(ri->regstclass);
16187 ri->regstclass = 0;
16194 /* trie structure. */
16196 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
16197 #ifdef USE_ITHREADS
16201 refcount = --trie->refcount;
16204 PerlMemShared_free(trie->charmap);
16205 PerlMemShared_free(trie->states);
16206 PerlMemShared_free(trie->trans);
16208 PerlMemShared_free(trie->bitmap);
16210 PerlMemShared_free(trie->jump);
16211 PerlMemShared_free(trie->wordinfo);
16212 /* do this last!!!! */
16213 PerlMemShared_free(ri->data->data[n]);
16218 Perl_croak(aTHX_ "panic: regfree data code '%c'",
16219 ri->data->what[n]);
16222 Safefree(ri->data->what);
16223 Safefree(ri->data);
16229 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
16230 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
16231 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
16234 re_dup - duplicate a regexp.
16236 This routine is expected to clone a given regexp structure. It is only
16237 compiled under USE_ITHREADS.
16239 After all of the core data stored in struct regexp is duplicated
16240 the regexp_engine.dupe method is used to copy any private data
16241 stored in the *pprivate pointer. This allows extensions to handle
16242 any duplication it needs to do.
16244 See pregfree() and regfree_internal() if you change anything here.
16246 #if defined(USE_ITHREADS)
16247 #ifndef PERL_IN_XSUB_RE
16249 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
16253 const struct regexp *r = ReANY(sstr);
16254 struct regexp *ret = ReANY(dstr);
16256 PERL_ARGS_ASSERT_RE_DUP_GUTS;
16258 npar = r->nparens+1;
16259 Newx(ret->offs, npar, regexp_paren_pair);
16260 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
16262 if (ret->substrs) {
16263 /* Do it this way to avoid reading from *r after the StructCopy().
16264 That way, if any of the sv_dup_inc()s dislodge *r from the L1
16265 cache, it doesn't matter. */
16266 const bool anchored = r->check_substr
16267 ? r->check_substr == r->anchored_substr
16268 : r->check_utf8 == r->anchored_utf8;
16269 Newx(ret->substrs, 1, struct reg_substr_data);
16270 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
16272 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
16273 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
16274 ret->float_substr = sv_dup_inc(ret->float_substr, param);
16275 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
16277 /* check_substr and check_utf8, if non-NULL, point to either their
16278 anchored or float namesakes, and don't hold a second reference. */
16280 if (ret->check_substr) {
16282 assert(r->check_utf8 == r->anchored_utf8);
16283 ret->check_substr = ret->anchored_substr;
16284 ret->check_utf8 = ret->anchored_utf8;
16286 assert(r->check_substr == r->float_substr);
16287 assert(r->check_utf8 == r->float_utf8);
16288 ret->check_substr = ret->float_substr;
16289 ret->check_utf8 = ret->float_utf8;
16291 } else if (ret->check_utf8) {
16293 ret->check_utf8 = ret->anchored_utf8;
16295 ret->check_utf8 = ret->float_utf8;
16300 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
16301 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
16304 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
16306 if (RX_MATCH_COPIED(dstr))
16307 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
16309 ret->subbeg = NULL;
16310 #ifdef PERL_ANY_COW
16311 ret->saved_copy = NULL;
16314 /* Whether mother_re be set or no, we need to copy the string. We
16315 cannot refrain from copying it when the storage points directly to
16316 our mother regexp, because that's
16317 1: a buffer in a different thread
16318 2: something we no longer hold a reference on
16319 so we need to copy it locally. */
16320 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
16321 ret->mother_re = NULL;
16323 #endif /* PERL_IN_XSUB_RE */
16328 This is the internal complement to regdupe() which is used to copy
16329 the structure pointed to by the *pprivate pointer in the regexp.
16330 This is the core version of the extension overridable cloning hook.
16331 The regexp structure being duplicated will be copied by perl prior
16332 to this and will be provided as the regexp *r argument, however
16333 with the /old/ structures pprivate pointer value. Thus this routine
16334 may override any copying normally done by perl.
16336 It returns a pointer to the new regexp_internal structure.
16340 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
16343 struct regexp *const r = ReANY(rx);
16344 regexp_internal *reti;
16346 RXi_GET_DECL(r,ri);
16348 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
16352 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
16353 char, regexp_internal);
16354 Copy(ri->program, reti->program, len+1, regnode);
16356 reti->num_code_blocks = ri->num_code_blocks;
16357 if (ri->code_blocks) {
16359 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
16360 struct reg_code_block);
16361 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
16362 struct reg_code_block);
16363 for (n = 0; n < ri->num_code_blocks; n++)
16364 reti->code_blocks[n].src_regex = (REGEXP*)
16365 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
16368 reti->code_blocks = NULL;
16370 reti->regstclass = NULL;
16373 struct reg_data *d;
16374 const int count = ri->data->count;
16377 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
16378 char, struct reg_data);
16379 Newx(d->what, count, U8);
16382 for (i = 0; i < count; i++) {
16383 d->what[i] = ri->data->what[i];
16384 switch (d->what[i]) {
16385 /* see also regcomp.h and regfree_internal() */
16386 case 'a': /* actually an AV, but the dup function is identical. */
16390 case 'u': /* actually an HV, but the dup function is identical. */
16391 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
16394 /* This is cheating. */
16395 Newx(d->data[i], 1, regnode_ssc);
16396 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
16397 reti->regstclass = (regnode*)d->data[i];
16400 /* Trie stclasses are readonly and can thus be shared
16401 * without duplication. We free the stclass in pregfree
16402 * when the corresponding reg_ac_data struct is freed.
16404 reti->regstclass= ri->regstclass;
16408 ((reg_trie_data*)ri->data->data[i])->refcount++;
16413 d->data[i] = ri->data->data[i];
16416 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
16417 ri->data->what[i]);
16426 reti->name_list_idx = ri->name_list_idx;
16428 #ifdef RE_TRACK_PATTERN_OFFSETS
16429 if (ri->u.offsets) {
16430 Newx(reti->u.offsets, 2*len+1, U32);
16431 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
16434 SetProgLen(reti,len);
16437 return (void*)reti;
16440 #endif /* USE_ITHREADS */
16442 #ifndef PERL_IN_XSUB_RE
16445 - regnext - dig the "next" pointer out of a node
16448 Perl_regnext(pTHX_ regnode *p)
16455 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
16456 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16457 (int)OP(p), (int)REGNODE_MAX);
16460 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
16469 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
16472 STRLEN l1 = strlen(pat1);
16473 STRLEN l2 = strlen(pat2);
16476 const char *message;
16478 PERL_ARGS_ASSERT_RE_CROAK2;
16484 Copy(pat1, buf, l1 , char);
16485 Copy(pat2, buf + l1, l2 , char);
16486 buf[l1 + l2] = '\n';
16487 buf[l1 + l2 + 1] = '\0';
16488 va_start(args, pat2);
16489 msv = vmess(buf, &args);
16491 message = SvPV_const(msv,l1);
16494 Copy(message, buf, l1 , char);
16495 /* l1-1 to avoid \n */
16496 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
16499 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
16501 #ifndef PERL_IN_XSUB_RE
16503 Perl_save_re_context(pTHX)
16505 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
16507 const REGEXP * const rx = PM_GETRE(PL_curpm);
16510 for (i = 1; i <= RX_NPARENS(rx); i++) {
16511 char digits[TYPE_CHARS(long)];
16512 const STRLEN len = my_snprintf(digits, sizeof(digits),
16514 GV *const *const gvp
16515 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
16518 GV * const gv = *gvp;
16519 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
16531 S_put_byte(pTHX_ SV *sv, int c)
16533 PERL_ARGS_ASSERT_PUT_BYTE;
16537 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
16538 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
16539 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
16540 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
16541 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
16544 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
16549 const char string = c;
16550 if (c == '-' || c == ']' || c == '\\' || c == '^')
16551 sv_catpvs(sv, "\\");
16552 sv_catpvn(sv, &string, 1);
16557 S_put_range(pTHX_ SV *sv, UV start, UV end)
16560 /* Appends to 'sv' a displayable version of the range of code points from
16561 * 'start' to 'end'. It assumes that only ASCII printables are displayable
16562 * as-is (though some of these will be escaped by put_byte()). For the
16563 * time being, this subroutine only works for latin1 (< 256) code points */
16565 assert(start <= end);
16567 PERL_ARGS_ASSERT_PUT_RANGE;
16569 while (start <= end) {
16570 if (end - start < 3) { /* Individual chars in short ranges */
16571 for (; start <= end; start++) {
16572 put_byte(sv, start);
16577 /* For small ranges that include printable ASCII characters, it's more
16578 * legible to print those characters rather than hex values. For
16579 * larger ranges that include more than printables, it's probably
16580 * clearer to just give the start and end points of the range in hex,
16581 * and that's all we can do if there aren't any printables within the
16584 * On ASCII platforms the range of printables is contiguous. If the
16585 * entire range is printable, we print each character as such. If the
16586 * range is partially printable and partially not, it's less likely
16587 * that the individual printables are meaningful, especially if all or
16588 * almost all of them are in the range. But we err on the side of the
16589 * individual printables being meaningful by using the hex only if the
16590 * range contains all but 2 of the printables.
16592 * On EBCDIC platforms, the printables are scattered around so that the
16593 * maximum range length containing only them is about 10. Anything
16594 * longer we treat as hex; otherwise we examine the range character by
16595 * character to see */
16597 if (start < 256 && (((end < 255) ? end : 255) - start <= 10))
16599 if ((isPRINT_A(start) && isPRINT_A(end))
16600 || (end >= 0x7F && (isPRINT_A(start) && start > 0x21))
16601 || ((end < 0x7D && isPRINT_A(end)) && start < 0x20))
16604 /* If the range beginning isn't an ASCII printable, we find the
16605 * last such in the range, then split the output, so all the
16606 * non-printables are in one subrange; then process the remaining
16607 * portion as usual. If the entire range isn't printables, we
16608 * don't split, but drop down to print as hex */
16609 if (! isPRINT_A(start)) {
16610 UV temp_end = start + 1;
16611 while (temp_end <= end && ! isPRINT_A(temp_end)) {
16614 if (temp_end <= end) {
16615 put_range(sv, start, temp_end - 1);
16621 /* If the range beginning is a digit, output a subrange of just the
16622 * digits, then process the remaining portion as usual */
16623 if (isDIGIT_A(start)) {
16624 put_byte(sv, start);
16625 sv_catpvs(sv, "-");
16626 while (start <= end && isDIGIT_A(start)) start++;
16627 put_byte(sv, start - 1);
16631 /* Similarly for alphabetics. Because in both ASCII and EBCDIC,
16632 * the code points for upper and lower A-Z and a-z aren't
16633 * intermixed, the resulting subrange will consist solely of either
16634 * upper- or lower- alphabetics */
16635 if (isALPHA_A(start)) {
16636 put_byte(sv, start);
16637 sv_catpvs(sv, "-");
16638 while (start <= end && isALPHA_A(start)) start++;
16639 put_byte(sv, start - 1);
16643 /* We output any remaining printables as individual characters */
16644 if (isPUNCT_A(start) || isSPACE_A(start)) {
16645 while (start <= end && (isPUNCT_A(start) || isSPACE_A(start))) {
16646 put_byte(sv, start);
16653 /* Here is a control or non-ascii. Output the range or subrange as
16655 Perl_sv_catpvf(aTHX_ sv, "\\x{%02" UVXf "}-\\x{%02" UVXf "}",
16657 (end < 256) ? end : 255);
16663 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
16665 /* Appends to 'sv' a displayable version of the innards of the bracketed
16666 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
16667 * output anything */
16670 bool has_output_anything = FALSE;
16672 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
16674 for (i = 0; i < 256; i++) {
16675 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
16677 /* The character at index i should be output. Find the next
16678 * character that should NOT be output */
16680 for (j = i + 1; j <= 256; j++) {
16681 if (! BITMAP_TEST((U8 *) bitmap, j)) {
16686 /* Everything between them is a single range that should be output
16688 put_range(sv, i, j - 1);
16689 has_output_anything = TRUE;
16694 return has_output_anything;
16697 #define CLEAR_OPTSTART \
16698 if (optstart) STMT_START { \
16699 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
16700 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
16704 #define DUMPUNTIL(b,e) \
16706 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
16708 STATIC const regnode *
16709 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
16710 const regnode *last, const regnode *plast,
16711 SV* sv, I32 indent, U32 depth)
16714 U8 op = PSEUDO; /* Arbitrary non-END op. */
16715 const regnode *next;
16716 const regnode *optstart= NULL;
16718 RXi_GET_DECL(r,ri);
16719 GET_RE_DEBUG_FLAGS_DECL;
16721 PERL_ARGS_ASSERT_DUMPUNTIL;
16723 #ifdef DEBUG_DUMPUNTIL
16724 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
16725 last ? last-start : 0,plast ? plast-start : 0);
16728 if (plast && plast < last)
16731 while (PL_regkind[op] != END && (!last || node < last)) {
16733 /* While that wasn't END last time... */
16736 if (op == CLOSE || op == WHILEM)
16738 next = regnext((regnode *)node);
16741 if (OP(node) == OPTIMIZED) {
16742 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
16749 regprop(r, sv, node, NULL);
16750 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
16751 (int)(2*indent + 1), "", SvPVX_const(sv));
16753 if (OP(node) != OPTIMIZED) {
16754 if (next == NULL) /* Next ptr. */
16755 PerlIO_printf(Perl_debug_log, " (0)");
16756 else if (PL_regkind[(U8)op] == BRANCH
16757 && PL_regkind[OP(next)] != BRANCH )
16758 PerlIO_printf(Perl_debug_log, " (FAIL)");
16760 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
16761 (void)PerlIO_putc(Perl_debug_log, '\n');
16765 if (PL_regkind[(U8)op] == BRANCHJ) {
16768 const regnode *nnode = (OP(next) == LONGJMP
16769 ? regnext((regnode *)next)
16771 if (last && nnode > last)
16773 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
16776 else if (PL_regkind[(U8)op] == BRANCH) {
16778 DUMPUNTIL(NEXTOPER(node), next);
16780 else if ( PL_regkind[(U8)op] == TRIE ) {
16781 const regnode *this_trie = node;
16782 const char op = OP(node);
16783 const U32 n = ARG(node);
16784 const reg_ac_data * const ac = op>=AHOCORASICK ?
16785 (reg_ac_data *)ri->data->data[n] :
16787 const reg_trie_data * const trie =
16788 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
16790 AV *const trie_words
16791 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
16793 const regnode *nextbranch= NULL;
16796 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
16797 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
16799 PerlIO_printf(Perl_debug_log, "%*s%s ",
16800 (int)(2*(indent+3)), "",
16802 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
16803 SvCUR(*elem_ptr), 60,
16804 PL_colors[0], PL_colors[1],
16806 ? PERL_PV_ESCAPE_UNI
16808 | PERL_PV_PRETTY_ELLIPSES
16809 | PERL_PV_PRETTY_LTGT
16814 U16 dist= trie->jump[word_idx+1];
16815 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
16816 (UV)((dist ? this_trie + dist : next) - start));
16819 nextbranch= this_trie + trie->jump[0];
16820 DUMPUNTIL(this_trie + dist, nextbranch);
16822 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
16823 nextbranch= regnext((regnode *)nextbranch);
16825 PerlIO_printf(Perl_debug_log, "\n");
16828 if (last && next > last)
16833 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
16834 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
16835 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
16837 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
16839 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
16841 else if ( op == PLUS || op == STAR) {
16842 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
16844 else if (PL_regkind[(U8)op] == ANYOF) {
16845 /* arglen 1 + class block */
16846 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_POSIXL)
16847 ? ANYOF_POSIXL_SKIP
16849 node = NEXTOPER(node);
16851 else if (PL_regkind[(U8)op] == EXACT) {
16852 /* Literal string, where present. */
16853 node += NODE_SZ_STR(node) - 1;
16854 node = NEXTOPER(node);
16857 node = NEXTOPER(node);
16858 node += regarglen[(U8)op];
16860 if (op == CURLYX || op == OPEN)
16864 #ifdef DEBUG_DUMPUNTIL
16865 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
16870 #endif /* DEBUGGING */
16874 * c-indentation-style: bsd
16875 * c-basic-offset: 4
16876 * indent-tabs-mode: nil
16879 * ex: set ts=8 sts=4 sw=4 et: