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 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) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
96 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
103 # if defined(BUGGY_MSC6)
104 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
105 # pragma optimize("a",off)
106 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
107 # pragma optimize("w",on )
108 # endif /* BUGGY_MSC6 */
112 #define STATIC static
116 typedef struct RExC_state_t {
117 U32 flags; /* RXf_* are we folding, multilining? */
118 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
119 char *precomp; /* uncompiled string. */
120 REGEXP *rx_sv; /* The SV that is the regexp. */
121 regexp *rx; /* perl core regexp structure */
122 regexp_internal *rxi; /* internal data for regexp object pprivate field */
123 char *start; /* Start of input for compile */
124 char *end; /* End of input for compile */
125 char *parse; /* Input-scan pointer. */
126 I32 whilem_seen; /* number of WHILEM in this expr */
127 regnode *emit_start; /* Start of emitted-code area */
128 regnode *emit_bound; /* First regnode outside of the allocated space */
129 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
130 implies compiling, so don't emit */
131 regnode emit_dummy; /* placeholder for emit to point to */
132 I32 naughty; /* How bad is this pattern? */
133 I32 sawback; /* Did we see \1, ...? */
135 I32 size; /* Code size. */
136 I32 npar; /* Capture buffer count, (OPEN). */
137 I32 cpar; /* Capture buffer count, (CLOSE). */
138 I32 nestroot; /* root parens we are in - used by accept */
141 regnode **open_parens; /* pointers to open parens */
142 regnode **close_parens; /* pointers to close parens */
143 regnode *opend; /* END node in program */
144 I32 utf8; /* whether the pattern is utf8 or not */
145 I32 orig_utf8; /* whether the pattern was originally in utf8 */
146 /* XXX use this for future optimisation of case
147 * where pattern must be upgraded to utf8. */
148 I32 uni_semantics; /* If a d charset modifier should use unicode
149 rules, even if the pattern is not in
151 HV *paren_names; /* Paren names */
153 regnode **recurse; /* Recurse regops */
154 I32 recurse_count; /* Number of recurse regops */
157 I32 override_recoding;
158 I32 in_multi_char_class;
159 struct reg_code_block *code_blocks; /* positions of literal (?{})
161 int num_code_blocks; /* size of code_blocks[] */
162 int code_index; /* next code_blocks[] slot */
164 char *starttry; /* -Dr: where regtry was called. */
165 #define RExC_starttry (pRExC_state->starttry)
167 SV *runtime_code_qr; /* qr with the runtime code blocks */
169 const char *lastparse;
171 AV *paren_name_list; /* idx -> name */
172 #define RExC_lastparse (pRExC_state->lastparse)
173 #define RExC_lastnum (pRExC_state->lastnum)
174 #define RExC_paren_name_list (pRExC_state->paren_name_list)
178 #define RExC_flags (pRExC_state->flags)
179 #define RExC_pm_flags (pRExC_state->pm_flags)
180 #define RExC_precomp (pRExC_state->precomp)
181 #define RExC_rx_sv (pRExC_state->rx_sv)
182 #define RExC_rx (pRExC_state->rx)
183 #define RExC_rxi (pRExC_state->rxi)
184 #define RExC_start (pRExC_state->start)
185 #define RExC_end (pRExC_state->end)
186 #define RExC_parse (pRExC_state->parse)
187 #define RExC_whilem_seen (pRExC_state->whilem_seen)
188 #ifdef RE_TRACK_PATTERN_OFFSETS
189 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
191 #define RExC_emit (pRExC_state->emit)
192 #define RExC_emit_dummy (pRExC_state->emit_dummy)
193 #define RExC_emit_start (pRExC_state->emit_start)
194 #define RExC_emit_bound (pRExC_state->emit_bound)
195 #define RExC_naughty (pRExC_state->naughty)
196 #define RExC_sawback (pRExC_state->sawback)
197 #define RExC_seen (pRExC_state->seen)
198 #define RExC_size (pRExC_state->size)
199 #define RExC_npar (pRExC_state->npar)
200 #define RExC_nestroot (pRExC_state->nestroot)
201 #define RExC_extralen (pRExC_state->extralen)
202 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
203 #define RExC_utf8 (pRExC_state->utf8)
204 #define RExC_uni_semantics (pRExC_state->uni_semantics)
205 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
206 #define RExC_open_parens (pRExC_state->open_parens)
207 #define RExC_close_parens (pRExC_state->close_parens)
208 #define RExC_opend (pRExC_state->opend)
209 #define RExC_paren_names (pRExC_state->paren_names)
210 #define RExC_recurse (pRExC_state->recurse)
211 #define RExC_recurse_count (pRExC_state->recurse_count)
212 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
213 #define RExC_contains_locale (pRExC_state->contains_locale)
214 #define RExC_override_recoding (pRExC_state->override_recoding)
215 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
218 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
219 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
220 ((*s) == '{' && regcurly(s, FALSE)))
223 #undef SPSTART /* dratted cpp namespace... */
226 * Flags to be passed up and down.
228 #define WORST 0 /* Worst case. */
229 #define HASWIDTH 0x01 /* Known to match non-null strings. */
231 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
232 * character. (There needs to be a case: in the switch statement in regexec.c
233 * for any node marked SIMPLE.) Note that this is not the same thing as
236 #define SPSTART 0x04 /* Starts with * or + */
237 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
238 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
239 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
241 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
243 /* whether trie related optimizations are enabled */
244 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
245 #define TRIE_STUDY_OPT
246 #define FULL_TRIE_STUDY
252 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
253 #define PBITVAL(paren) (1 << ((paren) & 7))
254 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
255 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
256 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
258 #define REQUIRE_UTF8 STMT_START { \
260 *flagp = RESTART_UTF8; \
265 /* This converts the named class defined in regcomp.h to its equivalent class
266 * number defined in handy.h. */
267 #define namedclass_to_classnum(class) ((int) ((class) / 2))
268 #define classnum_to_namedclass(classnum) ((classnum) * 2)
270 /* About scan_data_t.
272 During optimisation we recurse through the regexp program performing
273 various inplace (keyhole style) optimisations. In addition study_chunk
274 and scan_commit populate this data structure with information about
275 what strings MUST appear in the pattern. We look for the longest
276 string that must appear at a fixed location, and we look for the
277 longest string that may appear at a floating location. So for instance
282 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
283 strings (because they follow a .* construct). study_chunk will identify
284 both FOO and BAR as being the longest fixed and floating strings respectively.
286 The strings can be composites, for instance
290 will result in a composite fixed substring 'foo'.
292 For each string some basic information is maintained:
294 - offset or min_offset
295 This is the position the string must appear at, or not before.
296 It also implicitly (when combined with minlenp) tells us how many
297 characters must match before the string we are searching for.
298 Likewise when combined with minlenp and the length of the string it
299 tells us how many characters must appear after the string we have
303 Only used for floating strings. This is the rightmost point that
304 the string can appear at. If set to I32 max it indicates that the
305 string can occur infinitely far to the right.
308 A pointer to the minimum number of characters of the pattern that the
309 string was found inside. This is important as in the case of positive
310 lookahead or positive lookbehind we can have multiple patterns
315 The minimum length of the pattern overall is 3, the minimum length
316 of the lookahead part is 3, but the minimum length of the part that
317 will actually match is 1. So 'FOO's minimum length is 3, but the
318 minimum length for the F is 1. This is important as the minimum length
319 is used to determine offsets in front of and behind the string being
320 looked for. Since strings can be composites this is the length of the
321 pattern at the time it was committed with a scan_commit. Note that
322 the length is calculated by study_chunk, so that the minimum lengths
323 are not known until the full pattern has been compiled, thus the
324 pointer to the value.
328 In the case of lookbehind the string being searched for can be
329 offset past the start point of the final matching string.
330 If this value was just blithely removed from the min_offset it would
331 invalidate some of the calculations for how many chars must match
332 before or after (as they are derived from min_offset and minlen and
333 the length of the string being searched for).
334 When the final pattern is compiled and the data is moved from the
335 scan_data_t structure into the regexp structure the information
336 about lookbehind is factored in, with the information that would
337 have been lost precalculated in the end_shift field for the
340 The fields pos_min and pos_delta are used to store the minimum offset
341 and the delta to the maximum offset at the current point in the pattern.
345 typedef struct scan_data_t {
346 /*I32 len_min; unused */
347 /*I32 len_delta; unused */
351 I32 last_end; /* min value, <0 unless valid. */
354 SV **longest; /* Either &l_fixed, or &l_float. */
355 SV *longest_fixed; /* longest fixed string found in pattern */
356 I32 offset_fixed; /* offset where it starts */
357 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
358 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
359 SV *longest_float; /* longest floating string found in pattern */
360 I32 offset_float_min; /* earliest point in string it can appear */
361 I32 offset_float_max; /* latest point in string it can appear */
362 I32 *minlen_float; /* pointer to the minlen relevant to the string */
363 I32 lookbehind_float; /* is the position of the string modified by LB */
367 struct regnode_charclass_class *start_class;
370 /* The below is perhaps overboard, but this allows us to save a test at the
371 * expense of a mask. This is because on both EBCDIC and ASCII machines, 'A'
372 * and 'a' differ by a single bit; the same with the upper and lower case of
373 * all other ASCII-range alphabetics. On ASCII platforms, they are 32 apart;
374 * on EBCDIC, they are 64. This uses an exclusive 'or' to find that bit and
375 * then inverts it to form a mask, with just a single 0, in the bit position
376 * where the upper- and lowercase differ. XXX There are about 40 other
377 * instances in the Perl core where this micro-optimization could be used.
378 * Should decide if maintenance cost is worse, before changing those
380 * Returns a boolean as to whether or not 'v' is either a lowercase or
381 * uppercase instance of 'c', where 'c' is in [A-Za-z]. If 'c' is a
382 * compile-time constant, the generated code is better than some optimizing
383 * compilers figure out, amounting to a mask and test. The results are
384 * meaningless if 'c' is not one of [A-Za-z] */
385 #define isARG2_lower_or_UPPER_ARG1(c, v) \
386 (((v) & ~('A' ^ 'a')) == ((c) & ~('A' ^ 'a')))
389 * Forward declarations for pregcomp()'s friends.
392 static const scan_data_t zero_scan_data =
393 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
395 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
396 #define SF_BEFORE_SEOL 0x0001
397 #define SF_BEFORE_MEOL 0x0002
398 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
399 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
402 # define SF_FIX_SHIFT_EOL (0+2)
403 # define SF_FL_SHIFT_EOL (0+4)
405 # define SF_FIX_SHIFT_EOL (+2)
406 # define SF_FL_SHIFT_EOL (+4)
409 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
410 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
412 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
413 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
414 #define SF_IS_INF 0x0040
415 #define SF_HAS_PAR 0x0080
416 #define SF_IN_PAR 0x0100
417 #define SF_HAS_EVAL 0x0200
418 #define SCF_DO_SUBSTR 0x0400
419 #define SCF_DO_STCLASS_AND 0x0800
420 #define SCF_DO_STCLASS_OR 0x1000
421 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
422 #define SCF_WHILEM_VISITED_POS 0x2000
424 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
425 #define SCF_SEEN_ACCEPT 0x8000
426 #define SCF_TRIE_DOING_RESTUDY 0x10000
428 #define UTF cBOOL(RExC_utf8)
430 /* The enums for all these are ordered so things work out correctly */
431 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
432 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
433 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
434 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
435 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
436 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
437 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
439 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
441 #define OOB_NAMEDCLASS -1
443 /* There is no code point that is out-of-bounds, so this is problematic. But
444 * its only current use is to initialize a variable that is always set before
446 #define OOB_UNICODE 0xDEADBEEF
448 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
449 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
452 /* length of regex to show in messages that don't mark a position within */
453 #define RegexLengthToShowInErrorMessages 127
456 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
457 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
458 * op/pragma/warn/regcomp.
460 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
461 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
463 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
466 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
467 * arg. Show regex, up to a maximum length. If it's too long, chop and add
470 #define _FAIL(code) STMT_START { \
471 const char *ellipses = ""; \
472 IV len = RExC_end - RExC_precomp; \
475 SAVEFREESV(RExC_rx_sv); \
476 if (len > RegexLengthToShowInErrorMessages) { \
477 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
478 len = RegexLengthToShowInErrorMessages - 10; \
484 #define FAIL(msg) _FAIL( \
485 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
486 msg, (int)len, RExC_precomp, ellipses))
488 #define FAIL2(msg,arg) _FAIL( \
489 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
490 arg, (int)len, RExC_precomp, ellipses))
493 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
495 #define Simple_vFAIL(m) STMT_START { \
496 const IV offset = RExC_parse - RExC_precomp; \
497 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
498 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
502 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
504 #define vFAIL(m) STMT_START { \
506 SAVEFREESV(RExC_rx_sv); \
511 * Like Simple_vFAIL(), but accepts two arguments.
513 #define Simple_vFAIL2(m,a1) STMT_START { \
514 const IV offset = RExC_parse - RExC_precomp; \
515 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
516 (int)offset, RExC_precomp, RExC_precomp + offset); \
520 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
522 #define vFAIL2(m,a1) STMT_START { \
524 SAVEFREESV(RExC_rx_sv); \
525 Simple_vFAIL2(m, a1); \
530 * Like Simple_vFAIL(), but accepts three arguments.
532 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
533 const IV offset = RExC_parse - RExC_precomp; \
534 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
535 (int)offset, RExC_precomp, RExC_precomp + offset); \
539 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
541 #define vFAIL3(m,a1,a2) STMT_START { \
543 SAVEFREESV(RExC_rx_sv); \
544 Simple_vFAIL3(m, a1, a2); \
548 * Like Simple_vFAIL(), but accepts four arguments.
550 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
551 const IV offset = RExC_parse - RExC_precomp; \
552 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
553 (int)offset, RExC_precomp, RExC_precomp + offset); \
556 #define vFAIL4(m,a1,a2,a3) STMT_START { \
558 SAVEFREESV(RExC_rx_sv); \
559 Simple_vFAIL4(m, a1, a2, a3); \
562 /* m is not necessarily a "literal string", in this macro */
563 #define reg_warn_non_literal_string(loc, m) STMT_START { \
564 const IV offset = loc - RExC_precomp; \
565 Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
566 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
569 #define ckWARNreg(loc,m) STMT_START { \
570 const IV offset = loc - RExC_precomp; \
571 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
572 (int)offset, RExC_precomp, RExC_precomp + offset); \
575 #define vWARN_dep(loc, m) STMT_START { \
576 const IV offset = loc - RExC_precomp; \
577 Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
578 (int)offset, RExC_precomp, RExC_precomp + offset); \
581 #define ckWARNdep(loc,m) STMT_START { \
582 const IV offset = loc - RExC_precomp; \
583 Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
585 (int)offset, RExC_precomp, RExC_precomp + offset); \
588 #define ckWARNregdep(loc,m) STMT_START { \
589 const IV offset = loc - RExC_precomp; \
590 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
592 (int)offset, RExC_precomp, RExC_precomp + offset); \
595 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
596 const IV offset = loc - RExC_precomp; \
597 Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
599 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
602 #define ckWARN2reg(loc, m, a1) STMT_START { \
603 const IV offset = loc - RExC_precomp; \
604 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
605 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
608 #define vWARN3(loc, m, a1, a2) STMT_START { \
609 const IV offset = loc - RExC_precomp; \
610 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
611 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
614 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
615 const IV offset = loc - RExC_precomp; \
616 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
617 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
620 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
621 const IV offset = loc - RExC_precomp; \
622 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
623 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
626 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
627 const IV offset = loc - RExC_precomp; \
628 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
629 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
632 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
633 const IV offset = loc - RExC_precomp; \
634 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
635 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
639 /* Allow for side effects in s */
640 #define REGC(c,s) STMT_START { \
641 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
644 /* Macros for recording node offsets. 20001227 mjd@plover.com
645 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
646 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
647 * Element 0 holds the number n.
648 * Position is 1 indexed.
650 #ifndef RE_TRACK_PATTERN_OFFSETS
651 #define Set_Node_Offset_To_R(node,byte)
652 #define Set_Node_Offset(node,byte)
653 #define Set_Cur_Node_Offset
654 #define Set_Node_Length_To_R(node,len)
655 #define Set_Node_Length(node,len)
656 #define Set_Node_Cur_Length(node,start)
657 #define Node_Offset(n)
658 #define Node_Length(n)
659 #define Set_Node_Offset_Length(node,offset,len)
660 #define ProgLen(ri) ri->u.proglen
661 #define SetProgLen(ri,x) ri->u.proglen = x
663 #define ProgLen(ri) ri->u.offsets[0]
664 #define SetProgLen(ri,x) ri->u.offsets[0] = x
665 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
667 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
668 __LINE__, (int)(node), (int)(byte))); \
670 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
672 RExC_offsets[2*(node)-1] = (byte); \
677 #define Set_Node_Offset(node,byte) \
678 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
679 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
681 #define Set_Node_Length_To_R(node,len) STMT_START { \
683 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
684 __LINE__, (int)(node), (int)(len))); \
686 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
688 RExC_offsets[2*(node)] = (len); \
693 #define Set_Node_Length(node,len) \
694 Set_Node_Length_To_R((node)-RExC_emit_start, len)
695 #define Set_Node_Cur_Length(node, start) \
696 Set_Node_Length(node, RExC_parse - start)
698 /* Get offsets and lengths */
699 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
700 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
702 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
703 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
704 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
708 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
709 #define EXPERIMENTAL_INPLACESCAN
710 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
712 #define DEBUG_STUDYDATA(str,data,depth) \
713 DEBUG_OPTIMISE_MORE_r(if(data){ \
714 PerlIO_printf(Perl_debug_log, \
715 "%*s" str "Pos:%"IVdf"/%"IVdf \
716 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
717 (int)(depth)*2, "", \
718 (IV)((data)->pos_min), \
719 (IV)((data)->pos_delta), \
720 (UV)((data)->flags), \
721 (IV)((data)->whilem_c), \
722 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
723 is_inf ? "INF " : "" \
725 if ((data)->last_found) \
726 PerlIO_printf(Perl_debug_log, \
727 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
728 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
729 SvPVX_const((data)->last_found), \
730 (IV)((data)->last_end), \
731 (IV)((data)->last_start_min), \
732 (IV)((data)->last_start_max), \
733 ((data)->longest && \
734 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
735 SvPVX_const((data)->longest_fixed), \
736 (IV)((data)->offset_fixed), \
737 ((data)->longest && \
738 (data)->longest==&((data)->longest_float)) ? "*" : "", \
739 SvPVX_const((data)->longest_float), \
740 (IV)((data)->offset_float_min), \
741 (IV)((data)->offset_float_max) \
743 PerlIO_printf(Perl_debug_log,"\n"); \
746 /* Mark that we cannot extend a found fixed substring at this point.
747 Update the longest found anchored substring and the longest found
748 floating substrings if needed. */
751 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
753 const STRLEN l = CHR_SVLEN(data->last_found);
754 const STRLEN old_l = CHR_SVLEN(*data->longest);
755 GET_RE_DEBUG_FLAGS_DECL;
757 PERL_ARGS_ASSERT_SCAN_COMMIT;
759 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
760 SvSetMagicSV(*data->longest, data->last_found);
761 if (*data->longest == data->longest_fixed) {
762 data->offset_fixed = l ? data->last_start_min : data->pos_min;
763 if (data->flags & SF_BEFORE_EOL)
765 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
767 data->flags &= ~SF_FIX_BEFORE_EOL;
768 data->minlen_fixed=minlenp;
769 data->lookbehind_fixed=0;
771 else { /* *data->longest == data->longest_float */
772 data->offset_float_min = l ? data->last_start_min : data->pos_min;
773 data->offset_float_max = (l
774 ? data->last_start_max
775 : (data->pos_delta == I32_MAX ? I32_MAX : data->pos_min + data->pos_delta));
776 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
777 data->offset_float_max = I32_MAX;
778 if (data->flags & SF_BEFORE_EOL)
780 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
782 data->flags &= ~SF_FL_BEFORE_EOL;
783 data->minlen_float=minlenp;
784 data->lookbehind_float=0;
787 SvCUR_set(data->last_found, 0);
789 SV * const sv = data->last_found;
790 if (SvUTF8(sv) && SvMAGICAL(sv)) {
791 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
797 data->flags &= ~SF_BEFORE_EOL;
798 DEBUG_STUDYDATA("commit: ",data,0);
801 /* These macros set, clear and test whether the synthetic start class ('ssc',
802 * given by the parameter) matches an empty string (EOS). This uses the
803 * 'next_off' field in the node, to save a bit in the flags field. The ssc
804 * stands alone, so there is never a next_off, so this field is otherwise
805 * unused. The EOS information is used only for compilation, but theoretically
806 * it could be passed on to the execution code. This could be used to store
807 * more than one bit of information, but only this one is currently used. */
808 #define SET_SSC_EOS(node) STMT_START { (node)->next_off = TRUE; } STMT_END
809 #define CLEAR_SSC_EOS(node) STMT_START { (node)->next_off = FALSE; } STMT_END
810 #define TEST_SSC_EOS(node) cBOOL((node)->next_off)
812 /* Can match anything (initialization) */
814 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
816 PERL_ARGS_ASSERT_CL_ANYTHING;
818 ANYOF_BITMAP_SETALL(cl);
819 cl->flags = ANYOF_UNICODE_ALL;
822 /* If any portion of the regex is to operate under locale rules,
823 * initialization includes it. The reason this isn't done for all regexes
824 * is that the optimizer was written under the assumption that locale was
825 * all-or-nothing. Given the complexity and lack of documentation in the
826 * optimizer, and that there are inadequate test cases for locale, so many
827 * parts of it may not work properly, it is safest to avoid locale unless
829 if (RExC_contains_locale) {
830 ANYOF_CLASS_SETALL(cl); /* /l uses class */
831 cl->flags |= ANYOF_LOCALE|ANYOF_CLASS|ANYOF_LOC_FOLD;
834 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
838 /* Can match anything (initialization) */
840 S_cl_is_anything(const struct regnode_charclass_class *cl)
844 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
846 for (value = 0; value < ANYOF_MAX; value += 2)
847 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
849 if (!(cl->flags & ANYOF_UNICODE_ALL))
851 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
856 /* Can match anything (initialization) */
858 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
860 PERL_ARGS_ASSERT_CL_INIT;
862 Zero(cl, 1, struct regnode_charclass_class);
864 cl_anything(pRExC_state, cl);
865 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
868 /* These two functions currently do the exact same thing */
869 #define cl_init_zero cl_init
871 /* 'AND' a given class with another one. Can create false positives. 'cl'
872 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
873 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
875 S_cl_and(struct regnode_charclass_class *cl,
876 const struct regnode_charclass_class *and_with)
878 PERL_ARGS_ASSERT_CL_AND;
880 assert(PL_regkind[and_with->type] == ANYOF);
882 /* I (khw) am not sure all these restrictions are necessary XXX */
883 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
884 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
885 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
886 && !(and_with->flags & ANYOF_LOC_FOLD)
887 && !(cl->flags & ANYOF_LOC_FOLD)) {
890 if (and_with->flags & ANYOF_INVERT)
891 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
892 cl->bitmap[i] &= ~and_with->bitmap[i];
894 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
895 cl->bitmap[i] &= and_with->bitmap[i];
896 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
898 if (and_with->flags & ANYOF_INVERT) {
900 /* Here, the and'ed node is inverted. Get the AND of the flags that
901 * aren't affected by the inversion. Those that are affected are
902 * handled individually below */
903 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
904 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
905 cl->flags |= affected_flags;
907 /* We currently don't know how to deal with things that aren't in the
908 * bitmap, but we know that the intersection is no greater than what
909 * is already in cl, so let there be false positives that get sorted
910 * out after the synthetic start class succeeds, and the node is
911 * matched for real. */
913 /* The inversion of these two flags indicate that the resulting
914 * intersection doesn't have them */
915 if (and_with->flags & ANYOF_UNICODE_ALL) {
916 cl->flags &= ~ANYOF_UNICODE_ALL;
918 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
919 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
922 else { /* and'd node is not inverted */
923 U8 outside_bitmap_but_not_utf8; /* Temp variable */
925 if (! ANYOF_NONBITMAP(and_with)) {
927 /* Here 'and_with' doesn't match anything outside the bitmap
928 * (except possibly ANYOF_UNICODE_ALL), which means the
929 * intersection can't either, except for ANYOF_UNICODE_ALL, in
930 * which case we don't know what the intersection is, but it's no
931 * greater than what cl already has, so can just leave it alone,
932 * with possible false positives */
933 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
934 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
935 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
938 else if (! ANYOF_NONBITMAP(cl)) {
940 /* Here, 'and_with' does match something outside the bitmap, and cl
941 * doesn't have a list of things to match outside the bitmap. If
942 * cl can match all code points above 255, the intersection will
943 * be those above-255 code points that 'and_with' matches. If cl
944 * can't match all Unicode code points, it means that it can't
945 * match anything outside the bitmap (since the 'if' that got us
946 * into this block tested for that), so we leave the bitmap empty.
948 if (cl->flags & ANYOF_UNICODE_ALL) {
949 ARG_SET(cl, ARG(and_with));
951 /* and_with's ARG may match things that don't require UTF8.
952 * And now cl's will too, in spite of this being an 'and'. See
953 * the comments below about the kludge */
954 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
958 /* Here, both 'and_with' and cl match something outside the
959 * bitmap. Currently we do not do the intersection, so just match
960 * whatever cl had at the beginning. */
964 /* Take the intersection of the two sets of flags. However, the
965 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
966 * kludge around the fact that this flag is not treated like the others
967 * which are initialized in cl_anything(). The way the optimizer works
968 * is that the synthetic start class (SSC) is initialized to match
969 * anything, and then the first time a real node is encountered, its
970 * values are AND'd with the SSC's with the result being the values of
971 * the real node. However, there are paths through the optimizer where
972 * the AND never gets called, so those initialized bits are set
973 * inappropriately, which is not usually a big deal, as they just cause
974 * false positives in the SSC, which will just mean a probably
975 * imperceptible slow down in execution. However this bit has a
976 * higher false positive consequence in that it can cause utf8.pm,
977 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
978 * bigger slowdown and also causes significant extra memory to be used.
979 * In order to prevent this, the code now takes a different tack. The
980 * bit isn't set unless some part of the regular expression needs it,
981 * but once set it won't get cleared. This means that these extra
982 * modules won't get loaded unless there was some path through the
983 * pattern that would have required them anyway, and so any false
984 * positives that occur by not ANDing them out when they could be
985 * aren't as severe as they would be if we treated this bit like all
987 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
988 & ANYOF_NONBITMAP_NON_UTF8;
989 cl->flags &= and_with->flags;
990 cl->flags |= outside_bitmap_but_not_utf8;
994 /* 'OR' a given class with another one. Can create false positives. 'cl'
995 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
996 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
998 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
1000 PERL_ARGS_ASSERT_CL_OR;
1002 if (or_with->flags & ANYOF_INVERT) {
1004 /* Here, the or'd node is to be inverted. This means we take the
1005 * complement of everything not in the bitmap, but currently we don't
1006 * know what that is, so give up and match anything */
1007 if (ANYOF_NONBITMAP(or_with)) {
1008 cl_anything(pRExC_state, cl);
1011 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
1012 * <= (B1 | !B2) | (CL1 | !CL2)
1013 * which is wasteful if CL2 is small, but we ignore CL2:
1014 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
1015 * XXXX Can we handle case-fold? Unclear:
1016 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
1017 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
1019 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1020 && !(or_with->flags & ANYOF_LOC_FOLD)
1021 && !(cl->flags & ANYOF_LOC_FOLD) ) {
1024 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1025 cl->bitmap[i] |= ~or_with->bitmap[i];
1026 } /* XXXX: logic is complicated otherwise */
1028 cl_anything(pRExC_state, cl);
1031 /* And, we can just take the union of the flags that aren't affected
1032 * by the inversion */
1033 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
1035 /* For the remaining flags:
1036 ANYOF_UNICODE_ALL and inverted means to not match anything above
1037 255, which means that the union with cl should just be
1038 what cl has in it, so can ignore this flag
1039 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
1040 is 127-255 to match them, but then invert that, so the
1041 union with cl should just be what cl has in it, so can
1044 } else { /* 'or_with' is not inverted */
1045 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
1046 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
1047 && (!(or_with->flags & ANYOF_LOC_FOLD)
1048 || (cl->flags & ANYOF_LOC_FOLD)) ) {
1051 /* OR char bitmap and class bitmap separately */
1052 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
1053 cl->bitmap[i] |= or_with->bitmap[i];
1054 if (or_with->flags & ANYOF_CLASS) {
1055 ANYOF_CLASS_OR(or_with, cl);
1058 else { /* XXXX: logic is complicated, leave it along for a moment. */
1059 cl_anything(pRExC_state, cl);
1062 if (ANYOF_NONBITMAP(or_with)) {
1064 /* Use the added node's outside-the-bit-map match if there isn't a
1065 * conflict. If there is a conflict (both nodes match something
1066 * outside the bitmap, but what they match outside is not the same
1067 * pointer, and hence not easily compared until XXX we extend
1068 * inversion lists this far), give up and allow the start class to
1069 * match everything outside the bitmap. If that stuff is all above
1070 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1071 if (! ANYOF_NONBITMAP(cl)) {
1072 ARG_SET(cl, ARG(or_with));
1074 else if (ARG(cl) != ARG(or_with)) {
1076 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1077 cl_anything(pRExC_state, cl);
1080 cl->flags |= ANYOF_UNICODE_ALL;
1085 /* Take the union */
1086 cl->flags |= or_with->flags;
1090 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1091 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1092 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1093 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1098 dump_trie(trie,widecharmap,revcharmap)
1099 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1100 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1102 These routines dump out a trie in a somewhat readable format.
1103 The _interim_ variants are used for debugging the interim
1104 tables that are used to generate the final compressed
1105 representation which is what dump_trie expects.
1107 Part of the reason for their existence is to provide a form
1108 of documentation as to how the different representations function.
1113 Dumps the final compressed table form of the trie to Perl_debug_log.
1114 Used for debugging make_trie().
1118 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1119 AV *revcharmap, U32 depth)
1122 SV *sv=sv_newmortal();
1123 int colwidth= widecharmap ? 6 : 4;
1125 GET_RE_DEBUG_FLAGS_DECL;
1127 PERL_ARGS_ASSERT_DUMP_TRIE;
1129 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1130 (int)depth * 2 + 2,"",
1131 "Match","Base","Ofs" );
1133 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1134 SV ** const tmp = av_fetch( revcharmap, state, 0);
1136 PerlIO_printf( Perl_debug_log, "%*s",
1138 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1139 PL_colors[0], PL_colors[1],
1140 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1141 PERL_PV_ESCAPE_FIRSTCHAR
1146 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1147 (int)depth * 2 + 2,"");
1149 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1150 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1151 PerlIO_printf( Perl_debug_log, "\n");
1153 for( state = 1 ; state < trie->statecount ; state++ ) {
1154 const U32 base = trie->states[ state ].trans.base;
1156 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1158 if ( trie->states[ state ].wordnum ) {
1159 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1161 PerlIO_printf( Perl_debug_log, "%6s", "" );
1164 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1169 while( ( base + ofs < trie->uniquecharcount ) ||
1170 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1171 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1174 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1176 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1177 if ( ( base + ofs >= trie->uniquecharcount ) &&
1178 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1179 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1181 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1183 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1185 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1189 PerlIO_printf( Perl_debug_log, "]");
1192 PerlIO_printf( Perl_debug_log, "\n" );
1194 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1195 for (word=1; word <= trie->wordcount; word++) {
1196 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1197 (int)word, (int)(trie->wordinfo[word].prev),
1198 (int)(trie->wordinfo[word].len));
1200 PerlIO_printf(Perl_debug_log, "\n" );
1203 Dumps a fully constructed but uncompressed trie in list form.
1204 List tries normally only are used for construction when the number of
1205 possible chars (trie->uniquecharcount) is very high.
1206 Used for debugging make_trie().
1209 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1210 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1214 SV *sv=sv_newmortal();
1215 int colwidth= widecharmap ? 6 : 4;
1216 GET_RE_DEBUG_FLAGS_DECL;
1218 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1220 /* print out the table precompression. */
1221 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1222 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1223 "------:-----+-----------------\n" );
1225 for( state=1 ; state < next_alloc ; state ++ ) {
1228 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1229 (int)depth * 2 + 2,"", (UV)state );
1230 if ( ! trie->states[ state ].wordnum ) {
1231 PerlIO_printf( Perl_debug_log, "%5s| ","");
1233 PerlIO_printf( Perl_debug_log, "W%4x| ",
1234 trie->states[ state ].wordnum
1237 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1238 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1240 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1242 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1243 PL_colors[0], PL_colors[1],
1244 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1245 PERL_PV_ESCAPE_FIRSTCHAR
1247 TRIE_LIST_ITEM(state,charid).forid,
1248 (UV)TRIE_LIST_ITEM(state,charid).newstate
1251 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1252 (int)((depth * 2) + 14), "");
1255 PerlIO_printf( Perl_debug_log, "\n");
1260 Dumps a fully constructed but uncompressed trie in table form.
1261 This is the normal DFA style state transition table, with a few
1262 twists to facilitate compression later.
1263 Used for debugging make_trie().
1266 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1267 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1272 SV *sv=sv_newmortal();
1273 int colwidth= widecharmap ? 6 : 4;
1274 GET_RE_DEBUG_FLAGS_DECL;
1276 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1279 print out the table precompression so that we can do a visual check
1280 that they are identical.
1283 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1285 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1286 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1288 PerlIO_printf( Perl_debug_log, "%*s",
1290 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1291 PL_colors[0], PL_colors[1],
1292 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1293 PERL_PV_ESCAPE_FIRSTCHAR
1299 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1301 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1302 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1305 PerlIO_printf( Perl_debug_log, "\n" );
1307 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1309 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1310 (int)depth * 2 + 2,"",
1311 (UV)TRIE_NODENUM( state ) );
1313 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1314 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1316 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1318 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1320 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1321 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1323 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1324 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1332 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1333 startbranch: the first branch in the whole branch sequence
1334 first : start branch of sequence of branch-exact nodes.
1335 May be the same as startbranch
1336 last : Thing following the last branch.
1337 May be the same as tail.
1338 tail : item following the branch sequence
1339 count : words in the sequence
1340 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1341 depth : indent depth
1343 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1345 A trie is an N'ary tree where the branches are determined by digital
1346 decomposition of the key. IE, at the root node you look up the 1st character and
1347 follow that branch repeat until you find the end of the branches. Nodes can be
1348 marked as "accepting" meaning they represent a complete word. Eg:
1352 would convert into the following structure. Numbers represent states, letters
1353 following numbers represent valid transitions on the letter from that state, if
1354 the number is in square brackets it represents an accepting state, otherwise it
1355 will be in parenthesis.
1357 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1361 (1) +-i->(6)-+-s->[7]
1363 +-s->(3)-+-h->(4)-+-e->[5]
1365 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1367 This shows that when matching against the string 'hers' we will begin at state 1
1368 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1369 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1370 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1371 single traverse. We store a mapping from accepting to state to which word was
1372 matched, and then when we have multiple possibilities we try to complete the
1373 rest of the regex in the order in which they occured in the alternation.
1375 The only prior NFA like behaviour that would be changed by the TRIE support is
1376 the silent ignoring of duplicate alternations which are of the form:
1378 / (DUPE|DUPE) X? (?{ ... }) Y /x
1380 Thus EVAL blocks following a trie may be called a different number of times with
1381 and without the optimisation. With the optimisations dupes will be silently
1382 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1383 the following demonstrates:
1385 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1387 which prints out 'word' three times, but
1389 'words'=~/(word|word|word)(?{ print $1 })S/
1391 which doesnt print it out at all. This is due to other optimisations kicking in.
1393 Example of what happens on a structural level:
1395 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1397 1: CURLYM[1] {1,32767}(18)
1408 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1409 and should turn into:
1411 1: CURLYM[1] {1,32767}(18)
1413 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1421 Cases where tail != last would be like /(?foo|bar)baz/:
1431 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1432 and would end up looking like:
1435 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1442 d = uvuni_to_utf8_flags(d, uv, 0);
1444 is the recommended Unicode-aware way of saying
1449 #define TRIE_STORE_REVCHAR(val) \
1452 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1453 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1454 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1455 SvCUR_set(zlopp, kapow - flrbbbbb); \
1458 av_push(revcharmap, zlopp); \
1460 char ooooff = (char)val; \
1461 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1465 #define TRIE_READ_CHAR STMT_START { \
1468 /* if it is UTF then it is either already folded, or does not need folding */ \
1469 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1471 else if (folder == PL_fold_latin1) { \
1472 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1473 if ( foldlen > 0 ) { \
1474 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1480 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, FOLD_FLAGS_FULL); \
1481 skiplen = UNISKIP(uvc); \
1482 foldlen -= skiplen; \
1483 scan = foldbuf + skiplen; \
1486 /* raw data, will be folded later if needed */ \
1494 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1495 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1496 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1497 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1499 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1500 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1501 TRIE_LIST_CUR( state )++; \
1504 #define TRIE_LIST_NEW(state) STMT_START { \
1505 Newxz( trie->states[ state ].trans.list, \
1506 4, reg_trie_trans_le ); \
1507 TRIE_LIST_CUR( state ) = 1; \
1508 TRIE_LIST_LEN( state ) = 4; \
1511 #define TRIE_HANDLE_WORD(state) STMT_START { \
1512 U16 dupe= trie->states[ state ].wordnum; \
1513 regnode * const noper_next = regnext( noper ); \
1516 /* store the word for dumping */ \
1518 if (OP(noper) != NOTHING) \
1519 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1521 tmp = newSVpvn_utf8( "", 0, UTF ); \
1522 av_push( trie_words, tmp ); \
1526 trie->wordinfo[curword].prev = 0; \
1527 trie->wordinfo[curword].len = wordlen; \
1528 trie->wordinfo[curword].accept = state; \
1530 if ( noper_next < tail ) { \
1532 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1533 trie->jump[curword] = (U16)(noper_next - convert); \
1535 jumper = noper_next; \
1537 nextbranch= regnext(cur); \
1541 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1542 /* chain, so that when the bits of chain are later */\
1543 /* linked together, the dups appear in the chain */\
1544 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1545 trie->wordinfo[dupe].prev = curword; \
1547 /* we haven't inserted this word yet. */ \
1548 trie->states[ state ].wordnum = curword; \
1553 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1554 ( ( base + charid >= ucharcount \
1555 && base + charid < ubound \
1556 && state == trie->trans[ base - ucharcount + charid ].check \
1557 && trie->trans[ base - ucharcount + charid ].next ) \
1558 ? trie->trans[ base - ucharcount + charid ].next \
1559 : ( state==1 ? special : 0 ) \
1563 #define MADE_JUMP_TRIE 2
1564 #define MADE_EXACT_TRIE 4
1567 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1570 /* first pass, loop through and scan words */
1571 reg_trie_data *trie;
1572 HV *widecharmap = NULL;
1573 AV *revcharmap = newAV();
1575 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1580 regnode *jumper = NULL;
1581 regnode *nextbranch = NULL;
1582 regnode *convert = NULL;
1583 U32 *prev_states; /* temp array mapping each state to previous one */
1584 /* we just use folder as a flag in utf8 */
1585 const U8 * folder = NULL;
1588 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1589 AV *trie_words = NULL;
1590 /* along with revcharmap, this only used during construction but both are
1591 * useful during debugging so we store them in the struct when debugging.
1594 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1595 STRLEN trie_charcount=0;
1597 SV *re_trie_maxbuff;
1598 GET_RE_DEBUG_FLAGS_DECL;
1600 PERL_ARGS_ASSERT_MAKE_TRIE;
1602 PERL_UNUSED_ARG(depth);
1609 case EXACTFU_TRICKYFOLD:
1610 case EXACTFU: folder = PL_fold_latin1; break;
1611 case EXACTF: folder = PL_fold; break;
1612 case EXACTFL: folder = PL_fold_locale; break;
1613 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1616 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1618 trie->startstate = 1;
1619 trie->wordcount = word_count;
1620 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1621 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1623 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1624 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1625 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1628 trie_words = newAV();
1631 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1632 if (!SvIOK(re_trie_maxbuff)) {
1633 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1635 DEBUG_TRIE_COMPILE_r({
1636 PerlIO_printf( Perl_debug_log,
1637 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1638 (int)depth * 2 + 2, "",
1639 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1640 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1644 /* Find the node we are going to overwrite */
1645 if ( first == startbranch && OP( last ) != BRANCH ) {
1646 /* whole branch chain */
1649 /* branch sub-chain */
1650 convert = NEXTOPER( first );
1653 /* -- First loop and Setup --
1655 We first traverse the branches and scan each word to determine if it
1656 contains widechars, and how many unique chars there are, this is
1657 important as we have to build a table with at least as many columns as we
1660 We use an array of integers to represent the character codes 0..255
1661 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1662 native representation of the character value as the key and IV's for the
1665 *TODO* If we keep track of how many times each character is used we can
1666 remap the columns so that the table compression later on is more
1667 efficient in terms of memory by ensuring the most common value is in the
1668 middle and the least common are on the outside. IMO this would be better
1669 than a most to least common mapping as theres a decent chance the most
1670 common letter will share a node with the least common, meaning the node
1671 will not be compressible. With a middle is most common approach the worst
1672 case is when we have the least common nodes twice.
1676 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1677 regnode *noper = NEXTOPER( cur );
1678 const U8 *uc = (U8*)STRING( noper );
1679 const U8 *e = uc + STR_LEN( noper );
1681 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1683 const U8 *scan = (U8*)NULL;
1684 U32 wordlen = 0; /* required init */
1686 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1688 if (OP(noper) == NOTHING) {
1689 regnode *noper_next= regnext(noper);
1690 if (noper_next != tail && OP(noper_next) == flags) {
1692 uc= (U8*)STRING(noper);
1693 e= uc + STR_LEN(noper);
1694 trie->minlen= STR_LEN(noper);
1701 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1702 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1703 regardless of encoding */
1704 if (OP( noper ) == EXACTFU_SS) {
1705 /* false positives are ok, so just set this */
1706 TRIE_BITMAP_SET(trie,0xDF);
1709 for ( ; uc < e ; uc += len ) {
1710 TRIE_CHARCOUNT(trie)++;
1715 U8 folded= folder[ (U8) uvc ];
1716 if ( !trie->charmap[ folded ] ) {
1717 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1718 TRIE_STORE_REVCHAR( folded );
1721 if ( !trie->charmap[ uvc ] ) {
1722 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1723 TRIE_STORE_REVCHAR( uvc );
1726 /* store the codepoint in the bitmap, and its folded
1728 TRIE_BITMAP_SET(trie, uvc);
1730 /* store the folded codepoint */
1731 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1734 /* store first byte of utf8 representation of
1735 variant codepoints */
1736 if (! UNI_IS_INVARIANT(uvc)) {
1737 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1740 set_bit = 0; /* We've done our bit :-) */
1745 widecharmap = newHV();
1747 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1750 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1752 if ( !SvTRUE( *svpp ) ) {
1753 sv_setiv( *svpp, ++trie->uniquecharcount );
1754 TRIE_STORE_REVCHAR(uvc);
1758 if( cur == first ) {
1759 trie->minlen = chars;
1760 trie->maxlen = chars;
1761 } else if (chars < trie->minlen) {
1762 trie->minlen = chars;
1763 } else if (chars > trie->maxlen) {
1764 trie->maxlen = chars;
1766 if (OP( noper ) == EXACTFU_SS) {
1767 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1768 if (trie->minlen > 1)
1771 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1772 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1773 * - We assume that any such sequence might match a 2 byte string */
1774 if (trie->minlen > 2 )
1778 } /* end first pass */
1779 DEBUG_TRIE_COMPILE_r(
1780 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1781 (int)depth * 2 + 2,"",
1782 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1783 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1784 (int)trie->minlen, (int)trie->maxlen )
1788 We now know what we are dealing with in terms of unique chars and
1789 string sizes so we can calculate how much memory a naive
1790 representation using a flat table will take. If it's over a reasonable
1791 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1792 conservative but potentially much slower representation using an array
1795 At the end we convert both representations into the same compressed
1796 form that will be used in regexec.c for matching with. The latter
1797 is a form that cannot be used to construct with but has memory
1798 properties similar to the list form and access properties similar
1799 to the table form making it both suitable for fast searches and
1800 small enough that its feasable to store for the duration of a program.
1802 See the comment in the code where the compressed table is produced
1803 inplace from the flat tabe representation for an explanation of how
1804 the compression works.
1809 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1812 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1814 Second Pass -- Array Of Lists Representation
1816 Each state will be represented by a list of charid:state records
1817 (reg_trie_trans_le) the first such element holds the CUR and LEN
1818 points of the allocated array. (See defines above).
1820 We build the initial structure using the lists, and then convert
1821 it into the compressed table form which allows faster lookups
1822 (but cant be modified once converted).
1825 STRLEN transcount = 1;
1827 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1828 "%*sCompiling trie using list compiler\n",
1829 (int)depth * 2 + 2, ""));
1831 trie->states = (reg_trie_state *)
1832 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1833 sizeof(reg_trie_state) );
1837 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1839 regnode *noper = NEXTOPER( cur );
1840 U8 *uc = (U8*)STRING( noper );
1841 const U8 *e = uc + STR_LEN( noper );
1842 U32 state = 1; /* required init */
1843 U16 charid = 0; /* sanity init */
1844 U8 *scan = (U8*)NULL; /* sanity init */
1845 STRLEN foldlen = 0; /* required init */
1846 U32 wordlen = 0; /* required init */
1847 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1850 if (OP(noper) == NOTHING) {
1851 regnode *noper_next= regnext(noper);
1852 if (noper_next != tail && OP(noper_next) == flags) {
1854 uc= (U8*)STRING(noper);
1855 e= uc + STR_LEN(noper);
1859 if (OP(noper) != NOTHING) {
1860 for ( ; uc < e ; uc += len ) {
1865 charid = trie->charmap[ uvc ];
1867 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1871 charid=(U16)SvIV( *svpp );
1874 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1881 if ( !trie->states[ state ].trans.list ) {
1882 TRIE_LIST_NEW( state );
1884 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1885 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1886 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1891 newstate = next_alloc++;
1892 prev_states[newstate] = state;
1893 TRIE_LIST_PUSH( state, charid, newstate );
1898 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1902 TRIE_HANDLE_WORD(state);
1904 } /* end second pass */
1906 /* next alloc is the NEXT state to be allocated */
1907 trie->statecount = next_alloc;
1908 trie->states = (reg_trie_state *)
1909 PerlMemShared_realloc( trie->states,
1911 * sizeof(reg_trie_state) );
1913 /* and now dump it out before we compress it */
1914 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1915 revcharmap, next_alloc,
1919 trie->trans = (reg_trie_trans *)
1920 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1927 for( state=1 ; state < next_alloc ; state ++ ) {
1931 DEBUG_TRIE_COMPILE_MORE_r(
1932 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1936 if (trie->states[state].trans.list) {
1937 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1941 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1942 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1943 if ( forid < minid ) {
1945 } else if ( forid > maxid ) {
1949 if ( transcount < tp + maxid - minid + 1) {
1951 trie->trans = (reg_trie_trans *)
1952 PerlMemShared_realloc( trie->trans,
1954 * sizeof(reg_trie_trans) );
1955 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1957 base = trie->uniquecharcount + tp - minid;
1958 if ( maxid == minid ) {
1960 for ( ; zp < tp ; zp++ ) {
1961 if ( ! trie->trans[ zp ].next ) {
1962 base = trie->uniquecharcount + zp - minid;
1963 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1964 trie->trans[ zp ].check = state;
1970 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1971 trie->trans[ tp ].check = state;
1976 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1977 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1978 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1979 trie->trans[ tid ].check = state;
1981 tp += ( maxid - minid + 1 );
1983 Safefree(trie->states[ state ].trans.list);
1986 DEBUG_TRIE_COMPILE_MORE_r(
1987 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1990 trie->states[ state ].trans.base=base;
1992 trie->lasttrans = tp + 1;
1996 Second Pass -- Flat Table Representation.
1998 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1999 We know that we will need Charcount+1 trans at most to store the data
2000 (one row per char at worst case) So we preallocate both structures
2001 assuming worst case.
2003 We then construct the trie using only the .next slots of the entry
2006 We use the .check field of the first entry of the node temporarily to
2007 make compression both faster and easier by keeping track of how many non
2008 zero fields are in the node.
2010 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2013 There are two terms at use here: state as a TRIE_NODEIDX() which is a
2014 number representing the first entry of the node, and state as a
2015 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
2016 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
2017 are 2 entrys per node. eg:
2025 The table is internally in the right hand, idx form. However as we also
2026 have to deal with the states array which is indexed by nodenum we have to
2027 use TRIE_NODENUM() to convert.
2030 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2031 "%*sCompiling trie using table compiler\n",
2032 (int)depth * 2 + 2, ""));
2034 trie->trans = (reg_trie_trans *)
2035 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2036 * trie->uniquecharcount + 1,
2037 sizeof(reg_trie_trans) );
2038 trie->states = (reg_trie_state *)
2039 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2040 sizeof(reg_trie_state) );
2041 next_alloc = trie->uniquecharcount + 1;
2044 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2046 regnode *noper = NEXTOPER( cur );
2047 const U8 *uc = (U8*)STRING( noper );
2048 const U8 *e = uc + STR_LEN( noper );
2050 U32 state = 1; /* required init */
2052 U16 charid = 0; /* sanity init */
2053 U32 accept_state = 0; /* sanity init */
2054 U8 *scan = (U8*)NULL; /* sanity init */
2056 STRLEN foldlen = 0; /* required init */
2057 U32 wordlen = 0; /* required init */
2059 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
2061 if (OP(noper) == NOTHING) {
2062 regnode *noper_next= regnext(noper);
2063 if (noper_next != tail && OP(noper_next) == flags) {
2065 uc= (U8*)STRING(noper);
2066 e= uc + STR_LEN(noper);
2070 if ( OP(noper) != NOTHING ) {
2071 for ( ; uc < e ; uc += len ) {
2076 charid = trie->charmap[ uvc ];
2078 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2079 charid = svpp ? (U16)SvIV(*svpp) : 0;
2083 if ( !trie->trans[ state + charid ].next ) {
2084 trie->trans[ state + charid ].next = next_alloc;
2085 trie->trans[ state ].check++;
2086 prev_states[TRIE_NODENUM(next_alloc)]
2087 = TRIE_NODENUM(state);
2088 next_alloc += trie->uniquecharcount;
2090 state = trie->trans[ state + charid ].next;
2092 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2094 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2097 accept_state = TRIE_NODENUM( state );
2098 TRIE_HANDLE_WORD(accept_state);
2100 } /* end second pass */
2102 /* and now dump it out before we compress it */
2103 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2105 next_alloc, depth+1));
2109 * Inplace compress the table.*
2111 For sparse data sets the table constructed by the trie algorithm will
2112 be mostly 0/FAIL transitions or to put it another way mostly empty.
2113 (Note that leaf nodes will not contain any transitions.)
2115 This algorithm compresses the tables by eliminating most such
2116 transitions, at the cost of a modest bit of extra work during lookup:
2118 - Each states[] entry contains a .base field which indicates the
2119 index in the state[] array wheres its transition data is stored.
2121 - If .base is 0 there are no valid transitions from that node.
2123 - If .base is nonzero then charid is added to it to find an entry in
2126 -If trans[states[state].base+charid].check!=state then the
2127 transition is taken to be a 0/Fail transition. Thus if there are fail
2128 transitions at the front of the node then the .base offset will point
2129 somewhere inside the previous nodes data (or maybe even into a node
2130 even earlier), but the .check field determines if the transition is
2134 The following process inplace converts the table to the compressed
2135 table: We first do not compress the root node 1,and mark all its
2136 .check pointers as 1 and set its .base pointer as 1 as well. This
2137 allows us to do a DFA construction from the compressed table later,
2138 and ensures that any .base pointers we calculate later are greater
2141 - We set 'pos' to indicate the first entry of the second node.
2143 - We then iterate over the columns of the node, finding the first and
2144 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2145 and set the .check pointers accordingly, and advance pos
2146 appropriately and repreat for the next node. Note that when we copy
2147 the next pointers we have to convert them from the original
2148 NODEIDX form to NODENUM form as the former is not valid post
2151 - If a node has no transitions used we mark its base as 0 and do not
2152 advance the pos pointer.
2154 - If a node only has one transition we use a second pointer into the
2155 structure to fill in allocated fail transitions from other states.
2156 This pointer is independent of the main pointer and scans forward
2157 looking for null transitions that are allocated to a state. When it
2158 finds one it writes the single transition into the "hole". If the
2159 pointer doesnt find one the single transition is appended as normal.
2161 - Once compressed we can Renew/realloc the structures to release the
2164 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2165 specifically Fig 3.47 and the associated pseudocode.
2169 const U32 laststate = TRIE_NODENUM( next_alloc );
2172 trie->statecount = laststate;
2174 for ( state = 1 ; state < laststate ; state++ ) {
2176 const U32 stateidx = TRIE_NODEIDX( state );
2177 const U32 o_used = trie->trans[ stateidx ].check;
2178 U32 used = trie->trans[ stateidx ].check;
2179 trie->trans[ stateidx ].check = 0;
2181 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2182 if ( flag || trie->trans[ stateidx + charid ].next ) {
2183 if ( trie->trans[ stateidx + charid ].next ) {
2185 for ( ; zp < pos ; zp++ ) {
2186 if ( ! trie->trans[ zp ].next ) {
2190 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2191 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2192 trie->trans[ zp ].check = state;
2193 if ( ++zp > pos ) pos = zp;
2200 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2202 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2203 trie->trans[ pos ].check = state;
2208 trie->lasttrans = pos + 1;
2209 trie->states = (reg_trie_state *)
2210 PerlMemShared_realloc( trie->states, laststate
2211 * sizeof(reg_trie_state) );
2212 DEBUG_TRIE_COMPILE_MORE_r(
2213 PerlIO_printf( Perl_debug_log,
2214 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2215 (int)depth * 2 + 2,"",
2216 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2219 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2222 } /* end table compress */
2224 DEBUG_TRIE_COMPILE_MORE_r(
2225 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2226 (int)depth * 2 + 2, "",
2227 (UV)trie->statecount,
2228 (UV)trie->lasttrans)
2230 /* resize the trans array to remove unused space */
2231 trie->trans = (reg_trie_trans *)
2232 PerlMemShared_realloc( trie->trans, trie->lasttrans
2233 * sizeof(reg_trie_trans) );
2235 { /* Modify the program and insert the new TRIE node */
2236 U8 nodetype =(U8)(flags & 0xFF);
2240 regnode *optimize = NULL;
2241 #ifdef RE_TRACK_PATTERN_OFFSETS
2244 U32 mjd_nodelen = 0;
2245 #endif /* RE_TRACK_PATTERN_OFFSETS */
2246 #endif /* DEBUGGING */
2248 This means we convert either the first branch or the first Exact,
2249 depending on whether the thing following (in 'last') is a branch
2250 or not and whther first is the startbranch (ie is it a sub part of
2251 the alternation or is it the whole thing.)
2252 Assuming its a sub part we convert the EXACT otherwise we convert
2253 the whole branch sequence, including the first.
2255 /* Find the node we are going to overwrite */
2256 if ( first != startbranch || OP( last ) == BRANCH ) {
2257 /* branch sub-chain */
2258 NEXT_OFF( first ) = (U16)(last - first);
2259 #ifdef RE_TRACK_PATTERN_OFFSETS
2261 mjd_offset= Node_Offset((convert));
2262 mjd_nodelen= Node_Length((convert));
2265 /* whole branch chain */
2267 #ifdef RE_TRACK_PATTERN_OFFSETS
2270 const regnode *nop = NEXTOPER( convert );
2271 mjd_offset= Node_Offset((nop));
2272 mjd_nodelen= Node_Length((nop));
2276 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2277 (int)depth * 2 + 2, "",
2278 (UV)mjd_offset, (UV)mjd_nodelen)
2281 /* But first we check to see if there is a common prefix we can
2282 split out as an EXACT and put in front of the TRIE node. */
2283 trie->startstate= 1;
2284 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2286 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2290 const U32 base = trie->states[ state ].trans.base;
2292 if ( trie->states[state].wordnum )
2295 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2296 if ( ( base + ofs >= trie->uniquecharcount ) &&
2297 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2298 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2300 if ( ++count > 1 ) {
2301 SV **tmp = av_fetch( revcharmap, ofs, 0);
2302 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2303 if ( state == 1 ) break;
2305 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2307 PerlIO_printf(Perl_debug_log,
2308 "%*sNew Start State=%"UVuf" Class: [",
2309 (int)depth * 2 + 2, "",
2312 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2313 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2315 TRIE_BITMAP_SET(trie,*ch);
2317 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2319 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2323 TRIE_BITMAP_SET(trie,*ch);
2325 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2326 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2332 SV **tmp = av_fetch( revcharmap, idx, 0);
2334 char *ch = SvPV( *tmp, len );
2336 SV *sv=sv_newmortal();
2337 PerlIO_printf( Perl_debug_log,
2338 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2339 (int)depth * 2 + 2, "",
2341 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2342 PL_colors[0], PL_colors[1],
2343 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2344 PERL_PV_ESCAPE_FIRSTCHAR
2349 OP( convert ) = nodetype;
2350 str=STRING(convert);
2353 STR_LEN(convert) += len;
2359 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2364 trie->prefixlen = (state-1);
2366 regnode *n = convert+NODE_SZ_STR(convert);
2367 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2368 trie->startstate = state;
2369 trie->minlen -= (state - 1);
2370 trie->maxlen -= (state - 1);
2372 /* At least the UNICOS C compiler choked on this
2373 * being argument to DEBUG_r(), so let's just have
2376 #ifdef PERL_EXT_RE_BUILD
2382 regnode *fix = convert;
2383 U32 word = trie->wordcount;
2385 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2386 while( ++fix < n ) {
2387 Set_Node_Offset_Length(fix, 0, 0);
2390 SV ** const tmp = av_fetch( trie_words, word, 0 );
2392 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2393 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2395 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2403 NEXT_OFF(convert) = (U16)(tail - convert);
2404 DEBUG_r(optimize= n);
2410 if ( trie->maxlen ) {
2411 NEXT_OFF( convert ) = (U16)(tail - convert);
2412 ARG_SET( convert, data_slot );
2413 /* Store the offset to the first unabsorbed branch in
2414 jump[0], which is otherwise unused by the jump logic.
2415 We use this when dumping a trie and during optimisation. */
2417 trie->jump[0] = (U16)(nextbranch - convert);
2419 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2420 * and there is a bitmap
2421 * and the first "jump target" node we found leaves enough room
2422 * then convert the TRIE node into a TRIEC node, with the bitmap
2423 * embedded inline in the opcode - this is hypothetically faster.
2425 if ( !trie->states[trie->startstate].wordnum
2427 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2429 OP( convert ) = TRIEC;
2430 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2431 PerlMemShared_free(trie->bitmap);
2434 OP( convert ) = TRIE;
2436 /* store the type in the flags */
2437 convert->flags = nodetype;
2441 + regarglen[ OP( convert ) ];
2443 /* XXX We really should free up the resource in trie now,
2444 as we won't use them - (which resources?) dmq */
2446 /* needed for dumping*/
2447 DEBUG_r(if (optimize) {
2448 regnode *opt = convert;
2450 while ( ++opt < optimize) {
2451 Set_Node_Offset_Length(opt,0,0);
2454 Try to clean up some of the debris left after the
2457 while( optimize < jumper ) {
2458 mjd_nodelen += Node_Length((optimize));
2459 OP( optimize ) = OPTIMIZED;
2460 Set_Node_Offset_Length(optimize,0,0);
2463 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2465 } /* end node insert */
2467 /* Finish populating the prev field of the wordinfo array. Walk back
2468 * from each accept state until we find another accept state, and if
2469 * so, point the first word's .prev field at the second word. If the
2470 * second already has a .prev field set, stop now. This will be the
2471 * case either if we've already processed that word's accept state,
2472 * or that state had multiple words, and the overspill words were
2473 * already linked up earlier.
2480 for (word=1; word <= trie->wordcount; word++) {
2482 if (trie->wordinfo[word].prev)
2484 state = trie->wordinfo[word].accept;
2486 state = prev_states[state];
2489 prev = trie->states[state].wordnum;
2493 trie->wordinfo[word].prev = prev;
2495 Safefree(prev_states);
2499 /* and now dump out the compressed format */
2500 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2502 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2504 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2505 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2507 SvREFCNT_dec_NN(revcharmap);
2511 : trie->startstate>1
2517 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2519 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2521 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2522 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2525 We find the fail state for each state in the trie, this state is the longest proper
2526 suffix of the current state's 'word' that is also a proper prefix of another word in our
2527 trie. State 1 represents the word '' and is thus the default fail state. This allows
2528 the DFA not to have to restart after its tried and failed a word at a given point, it
2529 simply continues as though it had been matching the other word in the first place.
2531 'abcdgu'=~/abcdefg|cdgu/
2532 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2533 fail, which would bring us to the state representing 'd' in the second word where we would
2534 try 'g' and succeed, proceeding to match 'cdgu'.
2536 /* add a fail transition */
2537 const U32 trie_offset = ARG(source);
2538 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2540 const U32 ucharcount = trie->uniquecharcount;
2541 const U32 numstates = trie->statecount;
2542 const U32 ubound = trie->lasttrans + ucharcount;
2546 U32 base = trie->states[ 1 ].trans.base;
2549 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2550 GET_RE_DEBUG_FLAGS_DECL;
2552 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2554 PERL_UNUSED_ARG(depth);
2558 ARG_SET( stclass, data_slot );
2559 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2560 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2561 aho->trie=trie_offset;
2562 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2563 Copy( trie->states, aho->states, numstates, reg_trie_state );
2564 Newxz( q, numstates, U32);
2565 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2568 /* initialize fail[0..1] to be 1 so that we always have
2569 a valid final fail state */
2570 fail[ 0 ] = fail[ 1 ] = 1;
2572 for ( charid = 0; charid < ucharcount ; charid++ ) {
2573 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2575 q[ q_write ] = newstate;
2576 /* set to point at the root */
2577 fail[ q[ q_write++ ] ]=1;
2580 while ( q_read < q_write) {
2581 const U32 cur = q[ q_read++ % numstates ];
2582 base = trie->states[ cur ].trans.base;
2584 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2585 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2587 U32 fail_state = cur;
2590 fail_state = fail[ fail_state ];
2591 fail_base = aho->states[ fail_state ].trans.base;
2592 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2594 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2595 fail[ ch_state ] = fail_state;
2596 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2598 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2600 q[ q_write++ % numstates] = ch_state;
2604 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2605 when we fail in state 1, this allows us to use the
2606 charclass scan to find a valid start char. This is based on the principle
2607 that theres a good chance the string being searched contains lots of stuff
2608 that cant be a start char.
2610 fail[ 0 ] = fail[ 1 ] = 0;
2611 DEBUG_TRIE_COMPILE_r({
2612 PerlIO_printf(Perl_debug_log,
2613 "%*sStclass Failtable (%"UVuf" states): 0",
2614 (int)(depth * 2), "", (UV)numstates
2616 for( q_read=1; q_read<numstates; q_read++ ) {
2617 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2619 PerlIO_printf(Perl_debug_log, "\n");
2622 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2627 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2628 * These need to be revisited when a newer toolchain becomes available.
2630 #if defined(__sparc64__) && defined(__GNUC__)
2631 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2632 # undef SPARC64_GCC_WORKAROUND
2633 # define SPARC64_GCC_WORKAROUND 1
2637 #define DEBUG_PEEP(str,scan,depth) \
2638 DEBUG_OPTIMISE_r({if (scan){ \
2639 SV * const mysv=sv_newmortal(); \
2640 regnode *Next = regnext(scan); \
2641 regprop(RExC_rx, mysv, scan); \
2642 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2643 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2644 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2648 /* The below joins as many adjacent EXACTish nodes as possible into a single
2649 * one. The regop may be changed if the node(s) contain certain sequences that
2650 * require special handling. The joining is only done if:
2651 * 1) there is room in the current conglomerated node to entirely contain the
2653 * 2) they are the exact same node type
2655 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2656 * these get optimized out
2658 * If a node is to match under /i (folded), the number of characters it matches
2659 * can be different than its character length if it contains a multi-character
2660 * fold. *min_subtract is set to the total delta of the input nodes.
2662 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2663 * and contains LATIN SMALL LETTER SHARP S
2665 * This is as good a place as any to discuss the design of handling these
2666 * multi-character fold sequences. It's been wrong in Perl for a very long
2667 * time. There are three code points in Unicode whose multi-character folds
2668 * were long ago discovered to mess things up. The previous designs for
2669 * dealing with these involved assigning a special node for them. This
2670 * approach doesn't work, as evidenced by this example:
2671 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2672 * Both these fold to "sss", but if the pattern is parsed to create a node that
2673 * would match just the \xDF, it won't be able to handle the case where a
2674 * successful match would have to cross the node's boundary. The new approach
2675 * that hopefully generally solves the problem generates an EXACTFU_SS node
2678 * It turns out that there are problems with all multi-character folds, and not
2679 * just these three. Now the code is general, for all such cases, but the
2680 * three still have some special handling. The approach taken is:
2681 * 1) This routine examines each EXACTFish node that could contain multi-
2682 * character fold sequences. It returns in *min_subtract how much to
2683 * subtract from the the actual length of the string to get a real minimum
2684 * match length; it is 0 if there are no multi-char folds. This delta is
2685 * used by the caller to adjust the min length of the match, and the delta
2686 * between min and max, so that the optimizer doesn't reject these
2687 * possibilities based on size constraints.
2688 * 2) Certain of these sequences require special handling by the trie code,
2689 * so, if found, this code changes the joined node type to special ops:
2690 * EXACTFU_TRICKYFOLD and EXACTFU_SS.
2691 * 3) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
2692 * is used for an EXACTFU node that contains at least one "ss" sequence in
2693 * it. For non-UTF-8 patterns and strings, this is the only case where
2694 * there is a possible fold length change. That means that a regular
2695 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
2696 * with length changes, and so can be processed faster. regexec.c takes
2697 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
2698 * pre-folded by regcomp.c. This saves effort in regex matching.
2699 * However, the pre-folding isn't done for non-UTF8 patterns because the
2700 * fold of the MICRO SIGN requires UTF-8, and we don't want to slow things
2701 * down by forcing the pattern into UTF8 unless necessary. Also what
2702 * EXACTF and EXACTFL nodes fold to isn't known until runtime. The fold
2703 * possibilities for the non-UTF8 patterns are quite simple, except for
2704 * the sharp s. All the ones that don't involve a UTF-8 target string are
2705 * members of a fold-pair, and arrays are set up for all of them so that
2706 * the other member of the pair can be found quickly. Code elsewhere in
2707 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
2708 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
2709 * described in the next item.
2710 * 4) A problem remains for the sharp s in EXACTF and EXACTFA nodes when the
2711 * pattern isn't in UTF-8. (BTW, there cannot be an EXACTF node with a
2712 * UTF-8 pattern.) An assumption that the optimizer part of regexec.c
2713 * (probably unwittingly, in Perl_regexec_flags()) makes is that a
2714 * character in the pattern corresponds to at most a single character in
2715 * the target string. (And I do mean character, and not byte here, unlike
2716 * other parts of the documentation that have never been updated to
2717 * account for multibyte Unicode.) sharp s in EXACTF nodes can match the
2718 * two character string 'ss'; in EXACTFA nodes it can match
2719 * "\x{17F}\x{17F}". These violate the assumption, and they are the only
2720 * instances where it is violated. I'm reluctant to try to change the
2721 * assumption, as the code involved is impenetrable to me (khw), so
2722 * instead the code here punts. This routine examines (when the pattern
2723 * isn't UTF-8) EXACTF and EXACTFA nodes for the sharp s, and returns a
2724 * boolean indicating whether or not the node contains a sharp s. When it
2725 * is true, the caller sets a flag that later causes the optimizer in this
2726 * file to not set values for the floating and fixed string lengths, and
2727 * thus avoids the optimizer code in regexec.c that makes the invalid
2728 * assumption. Thus, there is no optimization based on string lengths for
2729 * non-UTF8-pattern EXACTF and EXACTFA nodes that contain the sharp s.
2730 * (The reason the assumption is wrong only in these two cases is that all
2731 * other non-UTF-8 folds are 1-1; and, for UTF-8 patterns, we pre-fold all
2732 * other folds to their expanded versions. We can't prefold sharp s to
2733 * 'ss' in EXACTF nodes because we don't know at compile time if it
2734 * actually matches 'ss' or not. It will match iff the target string is
2735 * in UTF-8, unlike the EXACTFU nodes, where it always matches; and
2736 * EXACTFA and EXACTFL where it never does. In an EXACTFA node in a UTF-8
2737 * pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the problem;
2738 * but in a non-UTF8 pattern, folding it to that above-Latin1 string would
2739 * require the pattern to be forced into UTF-8, the overhead of which we
2743 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2744 if (PL_regkind[OP(scan)] == EXACT) \
2745 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2748 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan, UV *min_subtract, bool *has_exactf_sharp_s, U32 flags,regnode *val, U32 depth) {
2749 /* Merge several consecutive EXACTish nodes into one. */
2750 regnode *n = regnext(scan);
2752 regnode *next = scan + NODE_SZ_STR(scan);
2756 regnode *stop = scan;
2757 GET_RE_DEBUG_FLAGS_DECL;
2759 PERL_UNUSED_ARG(depth);
2762 PERL_ARGS_ASSERT_JOIN_EXACT;
2763 #ifndef EXPERIMENTAL_INPLACESCAN
2764 PERL_UNUSED_ARG(flags);
2765 PERL_UNUSED_ARG(val);
2767 DEBUG_PEEP("join",scan,depth);
2769 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2770 * EXACT ones that are mergeable to the current one. */
2772 && (PL_regkind[OP(n)] == NOTHING
2773 || (stringok && OP(n) == OP(scan)))
2775 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2778 if (OP(n) == TAIL || n > next)
2780 if (PL_regkind[OP(n)] == NOTHING) {
2781 DEBUG_PEEP("skip:",n,depth);
2782 NEXT_OFF(scan) += NEXT_OFF(n);
2783 next = n + NODE_STEP_REGNODE;
2790 else if (stringok) {
2791 const unsigned int oldl = STR_LEN(scan);
2792 regnode * const nnext = regnext(n);
2794 /* XXX I (khw) kind of doubt that this works on platforms where
2795 * U8_MAX is above 255 because of lots of other assumptions */
2796 /* Don't join if the sum can't fit into a single node */
2797 if (oldl + STR_LEN(n) > U8_MAX)
2800 DEBUG_PEEP("merg",n,depth);
2803 NEXT_OFF(scan) += NEXT_OFF(n);
2804 STR_LEN(scan) += STR_LEN(n);
2805 next = n + NODE_SZ_STR(n);
2806 /* Now we can overwrite *n : */
2807 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2815 #ifdef EXPERIMENTAL_INPLACESCAN
2816 if (flags && !NEXT_OFF(n)) {
2817 DEBUG_PEEP("atch", val, depth);
2818 if (reg_off_by_arg[OP(n)]) {
2819 ARG_SET(n, val - n);
2822 NEXT_OFF(n) = val - n;
2830 *has_exactf_sharp_s = FALSE;
2832 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2833 * can now analyze for sequences of problematic code points. (Prior to
2834 * this final joining, sequences could have been split over boundaries, and
2835 * hence missed). The sequences only happen in folding, hence for any
2836 * non-EXACT EXACTish node */
2837 if (OP(scan) != EXACT) {
2838 const U8 * const s0 = (U8*) STRING(scan);
2840 const U8 * const s_end = s0 + STR_LEN(scan);
2842 /* One pass is made over the node's string looking for all the
2843 * possibilities. to avoid some tests in the loop, there are two main
2844 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2848 /* Examine the string for a multi-character fold sequence. UTF-8
2849 * patterns have all characters pre-folded by the time this code is
2851 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
2852 length sequence we are looking for is 2 */
2855 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
2856 if (! len) { /* Not a multi-char fold: get next char */
2861 /* Nodes with 'ss' require special handling, except for EXACTFL
2862 * and EXACTFA for which there is no multi-char fold to this */
2863 if (len == 2 && *s == 's' && *(s+1) == 's'
2864 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2867 OP(scan) = EXACTFU_SS;
2870 else if (len == 6 /* len is the same in both ASCII and EBCDIC
2872 && (memEQ(s, GREEK_SMALL_LETTER_IOTA_UTF8
2873 COMBINING_DIAERESIS_UTF8
2874 COMBINING_ACUTE_ACCENT_UTF8,
2876 || memEQ(s, GREEK_SMALL_LETTER_UPSILON_UTF8
2877 COMBINING_DIAERESIS_UTF8
2878 COMBINING_ACUTE_ACCENT_UTF8,
2883 /* These two folds require special handling by trie's, so
2884 * change the node type to indicate this. If EXACTFA and
2885 * EXACTFL were ever to be handled by trie's, this would
2886 * have to be changed. If this node has already been
2887 * changed to EXACTFU_SS in this loop, leave it as is. (I
2888 * (khw) think it doesn't matter in regexec.c for UTF
2889 * patterns, but no need to change it */
2890 if (OP(scan) == EXACTFU) {
2891 OP(scan) = EXACTFU_TRICKYFOLD;
2895 else { /* Here is a generic multi-char fold. */
2896 const U8* multi_end = s + len;
2898 /* Count how many characters in it. In the case of /l and
2899 * /aa, no folds which contain ASCII code points are
2900 * allowed, so check for those, and skip if found. (In
2901 * EXACTFL, no folds are allowed to any Latin1 code point,
2902 * not just ASCII. But there aren't any of these
2903 * currently, nor ever likely, so don't take the time to
2904 * test for them. The code that generates the
2905 * is_MULTI_foo() macros croaks should one actually get put
2906 * into Unicode .) */
2907 if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2908 count = utf8_length(s, multi_end);
2912 while (s < multi_end) {
2915 goto next_iteration;
2925 /* The delta is how long the sequence is minus 1 (1 is how long
2926 * the character that folds to the sequence is) */
2927 *min_subtract += count - 1;
2931 else if (OP(scan) == EXACTFA) {
2933 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
2934 * fold to the ASCII range (and there are no existing ones in the
2935 * upper latin1 range). But, as outlined in the comments preceding
2936 * this function, we need to flag any occurrences of the sharp s */
2938 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
2939 *has_exactf_sharp_s = TRUE;
2946 else if (OP(scan) != EXACTFL) {
2948 /* Non-UTF-8 pattern, not EXACTFA nor EXACTFL node. Look for the
2949 * multi-char folds that are all Latin1. (This code knows that
2950 * there are no current multi-char folds possible with EXACTFL,
2951 * relying on fold_grind.t to catch any errors if the very unlikely
2952 * event happens that some get added in future Unicode versions.)
2953 * As explained in the comments preceding this function, we look
2954 * also for the sharp s in EXACTF nodes; it can be in the final
2955 * position. Otherwise we can stop looking 1 byte earlier because
2956 * have to find at least two characters for a multi-fold */
2957 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2960 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
2961 if (! len) { /* Not a multi-char fold. */
2962 if (*s == LATIN_SMALL_LETTER_SHARP_S && OP(scan) == EXACTF)
2964 *has_exactf_sharp_s = TRUE;
2971 && isARG2_lower_or_UPPER_ARG1('s', *s)
2972 && isARG2_lower_or_UPPER_ARG1('s', *(s+1)))
2975 /* EXACTF nodes need to know that the minimum length
2976 * changed so that a sharp s in the string can match this
2977 * ss in the pattern, but they remain EXACTF nodes, as they
2978 * won't match this unless the target string is is UTF-8,
2979 * which we don't know until runtime */
2980 if (OP(scan) != EXACTF) {
2981 OP(scan) = EXACTFU_SS;
2985 *min_subtract += len - 1;
2992 /* Allow dumping but overwriting the collection of skipped
2993 * ops and/or strings with fake optimized ops */
2994 n = scan + NODE_SZ_STR(scan);
3002 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3006 /* REx optimizer. Converts nodes into quicker variants "in place".
3007 Finds fixed substrings. */
3009 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3010 to the position after last scanned or to NULL. */
3012 #define INIT_AND_WITHP \
3013 assert(!and_withp); \
3014 Newx(and_withp,1,struct regnode_charclass_class); \
3015 SAVEFREEPV(and_withp)
3017 /* this is a chain of data about sub patterns we are processing that
3018 need to be handled separately/specially in study_chunk. Its so
3019 we can simulate recursion without losing state. */
3021 typedef struct scan_frame {
3022 regnode *last; /* last node to process in this frame */
3023 regnode *next; /* next node to process when last is reached */
3024 struct scan_frame *prev; /*previous frame*/
3025 I32 stop; /* what stopparen do we use */
3029 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
3032 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3033 I32 *minlenp, I32 *deltap,
3038 struct regnode_charclass_class *and_withp,
3039 U32 flags, U32 depth)
3040 /* scanp: Start here (read-write). */
3041 /* deltap: Write maxlen-minlen here. */
3042 /* last: Stop before this one. */
3043 /* data: string data about the pattern */
3044 /* stopparen: treat close N as END */
3045 /* recursed: which subroutines have we recursed into */
3046 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3049 I32 min = 0; /* There must be at least this number of characters to match */
3051 regnode *scan = *scanp, *next;
3053 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3054 int is_inf_internal = 0; /* The studied chunk is infinite */
3055 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3056 scan_data_t data_fake;
3057 SV *re_trie_maxbuff = NULL;
3058 regnode *first_non_open = scan;
3059 I32 stopmin = I32_MAX;
3060 scan_frame *frame = NULL;
3061 GET_RE_DEBUG_FLAGS_DECL;
3063 PERL_ARGS_ASSERT_STUDY_CHUNK;
3066 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3070 while (first_non_open && OP(first_non_open) == OPEN)
3071 first_non_open=regnext(first_non_open);
3076 while ( scan && OP(scan) != END && scan < last ){
3077 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3078 node length to get a real minimum (because
3079 the folded version may be shorter) */
3080 bool has_exactf_sharp_s = FALSE;
3081 /* Peephole optimizer: */
3082 DEBUG_STUDYDATA("Peep:", data,depth);
3083 DEBUG_PEEP("Peep",scan,depth);
3085 /* Its not clear to khw or hv why this is done here, and not in the
3086 * clauses that deal with EXACT nodes. khw's guess is that it's
3087 * because of a previous design */
3088 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3090 /* Follow the next-chain of the current node and optimize
3091 away all the NOTHINGs from it. */
3092 if (OP(scan) != CURLYX) {
3093 const int max = (reg_off_by_arg[OP(scan)]
3095 /* I32 may be smaller than U16 on CRAYs! */
3096 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3097 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3101 /* Skip NOTHING and LONGJMP. */
3102 while ((n = regnext(n))
3103 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3104 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3105 && off + noff < max)
3107 if (reg_off_by_arg[OP(scan)])
3110 NEXT_OFF(scan) = off;
3115 /* The principal pseudo-switch. Cannot be a switch, since we
3116 look into several different things. */
3117 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3118 || OP(scan) == IFTHEN) {
3119 next = regnext(scan);
3121 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3123 if (OP(next) == code || code == IFTHEN) {
3124 /* NOTE - There is similar code to this block below for handling
3125 TRIE nodes on a re-study. If you change stuff here check there
3127 I32 max1 = 0, min1 = I32_MAX, num = 0;
3128 struct regnode_charclass_class accum;
3129 regnode * const startbranch=scan;
3131 if (flags & SCF_DO_SUBSTR)
3132 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3133 if (flags & SCF_DO_STCLASS)
3134 cl_init_zero(pRExC_state, &accum);
3136 while (OP(scan) == code) {
3137 I32 deltanext, minnext, f = 0, fake;
3138 struct regnode_charclass_class this_class;
3141 data_fake.flags = 0;
3143 data_fake.whilem_c = data->whilem_c;
3144 data_fake.last_closep = data->last_closep;
3147 data_fake.last_closep = &fake;
3149 data_fake.pos_delta = delta;
3150 next = regnext(scan);
3151 scan = NEXTOPER(scan);
3153 scan = NEXTOPER(scan);
3154 if (flags & SCF_DO_STCLASS) {
3155 cl_init(pRExC_state, &this_class);
3156 data_fake.start_class = &this_class;
3157 f = SCF_DO_STCLASS_AND;
3159 if (flags & SCF_WHILEM_VISITED_POS)
3160 f |= SCF_WHILEM_VISITED_POS;
3162 /* we suppose the run is continuous, last=next...*/
3163 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3165 stopparen, recursed, NULL, f,depth+1);
3168 if (deltanext == I32_MAX) {
3169 is_inf = is_inf_internal = 1;
3171 } else if (max1 < minnext + deltanext)
3172 max1 = minnext + deltanext;
3174 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3176 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3177 if ( stopmin > minnext)
3178 stopmin = min + min1;
3179 flags &= ~SCF_DO_SUBSTR;
3181 data->flags |= SCF_SEEN_ACCEPT;
3184 if (data_fake.flags & SF_HAS_EVAL)
3185 data->flags |= SF_HAS_EVAL;
3186 data->whilem_c = data_fake.whilem_c;
3188 if (flags & SCF_DO_STCLASS)
3189 cl_or(pRExC_state, &accum, &this_class);
3191 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3193 if (flags & SCF_DO_SUBSTR) {
3194 data->pos_min += min1;
3195 if (data->pos_delta >= I32_MAX - (max1 - min1))
3196 data->pos_delta = I32_MAX;
3198 data->pos_delta += max1 - min1;
3199 if (max1 != min1 || is_inf)
3200 data->longest = &(data->longest_float);
3203 if (delta == I32_MAX || I32_MAX - delta - (max1 - min1) < 0)
3206 delta += max1 - min1;
3207 if (flags & SCF_DO_STCLASS_OR) {
3208 cl_or(pRExC_state, data->start_class, &accum);
3210 cl_and(data->start_class, and_withp);
3211 flags &= ~SCF_DO_STCLASS;
3214 else if (flags & SCF_DO_STCLASS_AND) {
3216 cl_and(data->start_class, &accum);
3217 flags &= ~SCF_DO_STCLASS;
3220 /* Switch to OR mode: cache the old value of
3221 * data->start_class */
3223 StructCopy(data->start_class, and_withp,
3224 struct regnode_charclass_class);
3225 flags &= ~SCF_DO_STCLASS_AND;
3226 StructCopy(&accum, data->start_class,
3227 struct regnode_charclass_class);
3228 flags |= SCF_DO_STCLASS_OR;
3229 SET_SSC_EOS(data->start_class);
3233 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3236 Assuming this was/is a branch we are dealing with: 'scan' now
3237 points at the item that follows the branch sequence, whatever
3238 it is. We now start at the beginning of the sequence and look
3245 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3247 If we can find such a subsequence we need to turn the first
3248 element into a trie and then add the subsequent branch exact
3249 strings to the trie.
3253 1. patterns where the whole set of branches can be converted.
3255 2. patterns where only a subset can be converted.
3257 In case 1 we can replace the whole set with a single regop
3258 for the trie. In case 2 we need to keep the start and end
3261 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3262 becomes BRANCH TRIE; BRANCH X;
3264 There is an additional case, that being where there is a
3265 common prefix, which gets split out into an EXACT like node
3266 preceding the TRIE node.
3268 If x(1..n)==tail then we can do a simple trie, if not we make
3269 a "jump" trie, such that when we match the appropriate word
3270 we "jump" to the appropriate tail node. Essentially we turn
3271 a nested if into a case structure of sorts.
3276 if (!re_trie_maxbuff) {
3277 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3278 if (!SvIOK(re_trie_maxbuff))
3279 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3281 if ( SvIV(re_trie_maxbuff)>=0 ) {
3283 regnode *first = (regnode *)NULL;
3284 regnode *last = (regnode *)NULL;
3285 regnode *tail = scan;
3290 SV * const mysv = sv_newmortal(); /* for dumping */
3292 /* var tail is used because there may be a TAIL
3293 regop in the way. Ie, the exacts will point to the
3294 thing following the TAIL, but the last branch will
3295 point at the TAIL. So we advance tail. If we
3296 have nested (?:) we may have to move through several
3300 while ( OP( tail ) == TAIL ) {
3301 /* this is the TAIL generated by (?:) */
3302 tail = regnext( tail );
3306 DEBUG_TRIE_COMPILE_r({
3307 regprop(RExC_rx, mysv, tail );
3308 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3309 (int)depth * 2 + 2, "",
3310 "Looking for TRIE'able sequences. Tail node is: ",
3311 SvPV_nolen_const( mysv )
3317 Step through the branches
3318 cur represents each branch,
3319 noper is the first thing to be matched as part of that branch
3320 noper_next is the regnext() of that node.
3322 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3323 via a "jump trie" but we also support building with NOJUMPTRIE,
3324 which restricts the trie logic to structures like /FOO|BAR/.
3326 If noper is a trieable nodetype then the branch is a possible optimization
3327 target. If we are building under NOJUMPTRIE then we require that noper_next
3328 is the same as scan (our current position in the regex program).
3330 Once we have two or more consecutive such branches we can create a
3331 trie of the EXACT's contents and stitch it in place into the program.
3333 If the sequence represents all of the branches in the alternation we
3334 replace the entire thing with a single TRIE node.
3336 Otherwise when it is a subsequence we need to stitch it in place and
3337 replace only the relevant branches. This means the first branch has
3338 to remain as it is used by the alternation logic, and its next pointer,
3339 and needs to be repointed at the item on the branch chain following
3340 the last branch we have optimized away.
3342 This could be either a BRANCH, in which case the subsequence is internal,
3343 or it could be the item following the branch sequence in which case the
3344 subsequence is at the end (which does not necessarily mean the first node
3345 is the start of the alternation).
3347 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3350 ----------------+-----------
3354 EXACTFU_SS | EXACTFU
3355 EXACTFU_TRICKYFOLD | EXACTFU
3360 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3361 ( EXACT == (X) ) ? EXACT : \
3362 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3365 /* dont use tail as the end marker for this traverse */
3366 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3367 regnode * const noper = NEXTOPER( cur );
3368 U8 noper_type = OP( noper );
3369 U8 noper_trietype = TRIE_TYPE( noper_type );
3370 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3371 regnode * const noper_next = regnext( noper );
3372 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3373 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3376 DEBUG_TRIE_COMPILE_r({
3377 regprop(RExC_rx, mysv, cur);
3378 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3379 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3381 regprop(RExC_rx, mysv, noper);
3382 PerlIO_printf( Perl_debug_log, " -> %s",
3383 SvPV_nolen_const(mysv));
3386 regprop(RExC_rx, mysv, noper_next );
3387 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3388 SvPV_nolen_const(mysv));
3390 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3391 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3392 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3396 /* Is noper a trieable nodetype that can be merged with the
3397 * current trie (if there is one)? */
3401 ( noper_trietype == NOTHING)
3402 || ( trietype == NOTHING )
3403 || ( trietype == noper_trietype )
3406 && noper_next == tail
3410 /* Handle mergable triable node
3411 * Either we are the first node in a new trieable sequence,
3412 * in which case we do some bookkeeping, otherwise we update
3413 * the end pointer. */
3416 if ( noper_trietype == NOTHING ) {
3417 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3418 regnode * const noper_next = regnext( noper );
3419 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3420 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3423 if ( noper_next_trietype ) {
3424 trietype = noper_next_trietype;
3425 } else if (noper_next_type) {
3426 /* a NOTHING regop is 1 regop wide. We need at least two
3427 * for a trie so we can't merge this in */
3431 trietype = noper_trietype;
3434 if ( trietype == NOTHING )
3435 trietype = noper_trietype;
3440 } /* end handle mergable triable node */
3442 /* handle unmergable node -
3443 * noper may either be a triable node which can not be tried
3444 * together with the current trie, or a non triable node */
3446 /* If last is set and trietype is not NOTHING then we have found
3447 * at least two triable branch sequences in a row of a similar
3448 * trietype so we can turn them into a trie. If/when we
3449 * allow NOTHING to start a trie sequence this condition will be
3450 * required, and it isn't expensive so we leave it in for now. */
3451 if ( trietype && trietype != NOTHING )
3452 make_trie( pRExC_state,
3453 startbranch, first, cur, tail, count,
3454 trietype, depth+1 );
3455 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3459 && noper_next == tail
3462 /* noper is triable, so we can start a new trie sequence */
3465 trietype = noper_trietype;
3467 /* if we already saw a first but the current node is not triable then we have
3468 * to reset the first information. */
3473 } /* end handle unmergable node */
3474 } /* loop over branches */
3475 DEBUG_TRIE_COMPILE_r({
3476 regprop(RExC_rx, mysv, cur);
3477 PerlIO_printf( Perl_debug_log,
3478 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3479 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3482 if ( last && trietype ) {
3483 if ( trietype != NOTHING ) {
3484 /* the last branch of the sequence was part of a trie,
3485 * so we have to construct it here outside of the loop
3487 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3488 #ifdef TRIE_STUDY_OPT
3489 if ( ((made == MADE_EXACT_TRIE &&
3490 startbranch == first)
3491 || ( first_non_open == first )) &&
3493 flags |= SCF_TRIE_RESTUDY;
3494 if ( startbranch == first
3497 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3502 /* at this point we know whatever we have is a NOTHING sequence/branch
3503 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3505 if ( startbranch == first ) {
3507 /* the entire thing is a NOTHING sequence, something like this:
3508 * (?:|) So we can turn it into a plain NOTHING op. */
3509 DEBUG_TRIE_COMPILE_r({
3510 regprop(RExC_rx, mysv, cur);
3511 PerlIO_printf( Perl_debug_log,
3512 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3513 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3516 OP(startbranch)= NOTHING;
3517 NEXT_OFF(startbranch)= tail - startbranch;
3518 for ( opt= startbranch + 1; opt < tail ; opt++ )
3522 } /* end if ( last) */
3523 } /* TRIE_MAXBUF is non zero */
3528 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3529 scan = NEXTOPER(NEXTOPER(scan));
3530 } else /* single branch is optimized. */
3531 scan = NEXTOPER(scan);
3533 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3534 scan_frame *newframe = NULL;
3539 if (OP(scan) != SUSPEND) {
3540 /* set the pointer */
3541 if (OP(scan) == GOSUB) {
3543 RExC_recurse[ARG2L(scan)] = scan;
3544 start = RExC_open_parens[paren-1];
3545 end = RExC_close_parens[paren-1];
3548 start = RExC_rxi->program + 1;
3552 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3553 SAVEFREEPV(recursed);
3555 if (!PAREN_TEST(recursed,paren+1)) {
3556 PAREN_SET(recursed,paren+1);
3557 Newx(newframe,1,scan_frame);
3559 if (flags & SCF_DO_SUBSTR) {
3560 SCAN_COMMIT(pRExC_state,data,minlenp);
3561 data->longest = &(data->longest_float);
3563 is_inf = is_inf_internal = 1;
3564 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3565 cl_anything(pRExC_state, data->start_class);
3566 flags &= ~SCF_DO_STCLASS;
3569 Newx(newframe,1,scan_frame);
3572 end = regnext(scan);
3577 SAVEFREEPV(newframe);
3578 newframe->next = regnext(scan);
3579 newframe->last = last;
3580 newframe->stop = stopparen;
3581 newframe->prev = frame;
3591 else if (OP(scan) == EXACT) {
3592 I32 l = STR_LEN(scan);
3595 const U8 * const s = (U8*)STRING(scan);
3596 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3597 l = utf8_length(s, s + l);
3599 uc = *((U8*)STRING(scan));
3602 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3603 /* The code below prefers earlier match for fixed
3604 offset, later match for variable offset. */
3605 if (data->last_end == -1) { /* Update the start info. */
3606 data->last_start_min = data->pos_min;
3607 data->last_start_max = is_inf
3608 ? I32_MAX : data->pos_min + data->pos_delta;
3610 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3612 SvUTF8_on(data->last_found);
3614 SV * const sv = data->last_found;
3615 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3616 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3617 if (mg && mg->mg_len >= 0)
3618 mg->mg_len += utf8_length((U8*)STRING(scan),
3619 (U8*)STRING(scan)+STR_LEN(scan));
3621 data->last_end = data->pos_min + l;
3622 data->pos_min += l; /* As in the first entry. */
3623 data->flags &= ~SF_BEFORE_EOL;
3625 if (flags & SCF_DO_STCLASS_AND) {
3626 /* Check whether it is compatible with what we know already! */
3630 /* If compatible, we or it in below. It is compatible if is
3631 * in the bitmp and either 1) its bit or its fold is set, or 2)
3632 * it's for a locale. Even if there isn't unicode semantics
3633 * here, at runtime there may be because of matching against a
3634 * utf8 string, so accept a possible false positive for
3635 * latin1-range folds */
3637 (!(data->start_class->flags & ANYOF_LOCALE)
3638 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3639 && (!(data->start_class->flags & ANYOF_LOC_FOLD)
3640 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3645 ANYOF_CLASS_ZERO(data->start_class);
3646 ANYOF_BITMAP_ZERO(data->start_class);
3648 ANYOF_BITMAP_SET(data->start_class, uc);
3649 else if (uc >= 0x100) {
3652 /* Some Unicode code points fold to the Latin1 range; as
3653 * XXX temporary code, instead of figuring out if this is
3654 * one, just assume it is and set all the start class bits
3655 * that could be some such above 255 code point's fold
3656 * which will generate fals positives. As the code
3657 * elsewhere that does compute the fold settles down, it
3658 * can be extracted out and re-used here */
3659 for (i = 0; i < 256; i++){
3660 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3661 ANYOF_BITMAP_SET(data->start_class, i);
3665 CLEAR_SSC_EOS(data->start_class);
3667 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3669 else if (flags & SCF_DO_STCLASS_OR) {
3670 /* false positive possible if the class is case-folded */
3672 ANYOF_BITMAP_SET(data->start_class, uc);
3674 data->start_class->flags |= ANYOF_UNICODE_ALL;
3675 CLEAR_SSC_EOS(data->start_class);
3676 cl_and(data->start_class, and_withp);
3678 flags &= ~SCF_DO_STCLASS;
3680 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3681 I32 l = STR_LEN(scan);
3682 UV uc = *((U8*)STRING(scan));
3684 /* Search for fixed substrings supports EXACT only. */
3685 if (flags & SCF_DO_SUBSTR) {
3687 SCAN_COMMIT(pRExC_state, data, minlenp);
3690 const U8 * const s = (U8 *)STRING(scan);
3691 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3692 l = utf8_length(s, s + l);
3694 if (has_exactf_sharp_s) {
3695 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3697 min += l - min_subtract;
3699 delta += min_subtract;
3700 if (flags & SCF_DO_SUBSTR) {
3701 data->pos_min += l - min_subtract;
3702 if (data->pos_min < 0) {
3705 data->pos_delta += min_subtract;
3707 data->longest = &(data->longest_float);
3710 if (flags & SCF_DO_STCLASS_AND) {
3711 /* Check whether it is compatible with what we know already! */
3714 (!(data->start_class->flags & ANYOF_LOCALE)
3715 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3716 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3720 ANYOF_CLASS_ZERO(data->start_class);
3721 ANYOF_BITMAP_ZERO(data->start_class);
3723 ANYOF_BITMAP_SET(data->start_class, uc);
3724 CLEAR_SSC_EOS(data->start_class);
3725 if (OP(scan) == EXACTFL) {
3726 /* XXX This set is probably no longer necessary, and
3727 * probably wrong as LOCALE now is on in the initial
3729 data->start_class->flags |= ANYOF_LOCALE|ANYOF_LOC_FOLD;
3733 /* Also set the other member of the fold pair. In case
3734 * that unicode semantics is called for at runtime, use
3735 * the full latin1 fold. (Can't do this for locale,
3736 * because not known until runtime) */
3737 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3739 /* All other (EXACTFL handled above) folds except under
3740 * /iaa that include s, S, and sharp_s also may include
3742 if (OP(scan) != EXACTFA) {
3743 if (uc == 's' || uc == 'S') {
3744 ANYOF_BITMAP_SET(data->start_class,
3745 LATIN_SMALL_LETTER_SHARP_S);
3747 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3748 ANYOF_BITMAP_SET(data->start_class, 's');
3749 ANYOF_BITMAP_SET(data->start_class, 'S');
3754 else if (uc >= 0x100) {
3756 for (i = 0; i < 256; i++){
3757 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3758 ANYOF_BITMAP_SET(data->start_class, i);
3763 else if (flags & SCF_DO_STCLASS_OR) {
3764 if (data->start_class->flags & ANYOF_LOC_FOLD) {
3765 /* false positive possible if the class is case-folded.
3766 Assume that the locale settings are the same... */
3768 ANYOF_BITMAP_SET(data->start_class, uc);
3769 if (OP(scan) != EXACTFL) {
3771 /* And set the other member of the fold pair, but
3772 * can't do that in locale because not known until
3774 ANYOF_BITMAP_SET(data->start_class,
3775 PL_fold_latin1[uc]);
3777 /* All folds except under /iaa that include s, S,
3778 * and sharp_s also may include the others */
3779 if (OP(scan) != EXACTFA) {
3780 if (uc == 's' || uc == 'S') {
3781 ANYOF_BITMAP_SET(data->start_class,
3782 LATIN_SMALL_LETTER_SHARP_S);
3784 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3785 ANYOF_BITMAP_SET(data->start_class, 's');
3786 ANYOF_BITMAP_SET(data->start_class, 'S');
3791 CLEAR_SSC_EOS(data->start_class);
3793 cl_and(data->start_class, and_withp);
3795 flags &= ~SCF_DO_STCLASS;
3797 else if (REGNODE_VARIES(OP(scan))) {
3798 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3799 I32 f = flags, pos_before = 0;
3800 regnode * const oscan = scan;
3801 struct regnode_charclass_class this_class;
3802 struct regnode_charclass_class *oclass = NULL;
3803 I32 next_is_eval = 0;
3805 switch (PL_regkind[OP(scan)]) {
3806 case WHILEM: /* End of (?:...)* . */
3807 scan = NEXTOPER(scan);
3810 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3811 next = NEXTOPER(scan);
3812 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3814 maxcount = REG_INFTY;
3815 next = regnext(scan);
3816 scan = NEXTOPER(scan);
3820 if (flags & SCF_DO_SUBSTR)
3825 if (flags & SCF_DO_STCLASS) {
3827 maxcount = REG_INFTY;
3828 next = regnext(scan);
3829 scan = NEXTOPER(scan);
3832 is_inf = is_inf_internal = 1;
3833 scan = regnext(scan);
3834 if (flags & SCF_DO_SUBSTR) {
3835 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3836 data->longest = &(data->longest_float);
3838 goto optimize_curly_tail;
3840 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3841 && (scan->flags == stopparen))
3846 mincount = ARG1(scan);
3847 maxcount = ARG2(scan);
3849 next = regnext(scan);
3850 if (OP(scan) == CURLYX) {
3851 I32 lp = (data ? *(data->last_closep) : 0);
3852 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3854 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3855 next_is_eval = (OP(scan) == EVAL);
3857 if (flags & SCF_DO_SUBSTR) {
3858 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3859 pos_before = data->pos_min;
3863 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3865 data->flags |= SF_IS_INF;
3867 if (flags & SCF_DO_STCLASS) {
3868 cl_init(pRExC_state, &this_class);
3869 oclass = data->start_class;
3870 data->start_class = &this_class;
3871 f |= SCF_DO_STCLASS_AND;
3872 f &= ~SCF_DO_STCLASS_OR;
3874 /* Exclude from super-linear cache processing any {n,m}
3875 regops for which the combination of input pos and regex
3876 pos is not enough information to determine if a match
3879 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3880 regex pos at the \s*, the prospects for a match depend not
3881 only on the input position but also on how many (bar\s*)
3882 repeats into the {4,8} we are. */
3883 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3884 f &= ~SCF_WHILEM_VISITED_POS;
3886 /* This will finish on WHILEM, setting scan, or on NULL: */
3887 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3888 last, data, stopparen, recursed, NULL,
3890 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3892 if (flags & SCF_DO_STCLASS)
3893 data->start_class = oclass;
3894 if (mincount == 0 || minnext == 0) {
3895 if (flags & SCF_DO_STCLASS_OR) {
3896 cl_or(pRExC_state, data->start_class, &this_class);
3898 else if (flags & SCF_DO_STCLASS_AND) {
3899 /* Switch to OR mode: cache the old value of
3900 * data->start_class */
3902 StructCopy(data->start_class, and_withp,
3903 struct regnode_charclass_class);
3904 flags &= ~SCF_DO_STCLASS_AND;
3905 StructCopy(&this_class, data->start_class,
3906 struct regnode_charclass_class);
3907 flags |= SCF_DO_STCLASS_OR;
3908 SET_SSC_EOS(data->start_class);
3910 } else { /* Non-zero len */
3911 if (flags & SCF_DO_STCLASS_OR) {
3912 cl_or(pRExC_state, data->start_class, &this_class);
3913 cl_and(data->start_class, and_withp);
3915 else if (flags & SCF_DO_STCLASS_AND)
3916 cl_and(data->start_class, &this_class);
3917 flags &= ~SCF_DO_STCLASS;
3919 if (!scan) /* It was not CURLYX, but CURLY. */
3921 if (!(flags & SCF_TRIE_DOING_RESTUDY)
3922 /* ? quantifier ok, except for (?{ ... }) */
3923 && (next_is_eval || !(mincount == 0 && maxcount == 1))
3924 && (minnext == 0) && (deltanext == 0)
3925 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3926 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3928 /* Fatal warnings may leak the regexp without this: */
3929 SAVEFREESV(RExC_rx_sv);
3930 ckWARNreg(RExC_parse,
3931 "Quantifier unexpected on zero-length expression");
3932 (void)ReREFCNT_inc(RExC_rx_sv);
3935 min += minnext * mincount;
3936 is_inf_internal |= deltanext == I32_MAX
3937 || (maxcount == REG_INFTY && minnext + deltanext > 0);
3938 is_inf |= is_inf_internal;
3942 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3944 /* Try powerful optimization CURLYX => CURLYN. */
3945 if ( OP(oscan) == CURLYX && data
3946 && data->flags & SF_IN_PAR
3947 && !(data->flags & SF_HAS_EVAL)
3948 && !deltanext && minnext == 1 ) {
3949 /* Try to optimize to CURLYN. */
3950 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3951 regnode * const nxt1 = nxt;
3958 if (!REGNODE_SIMPLE(OP(nxt))
3959 && !(PL_regkind[OP(nxt)] == EXACT
3960 && STR_LEN(nxt) == 1))
3966 if (OP(nxt) != CLOSE)
3968 if (RExC_open_parens) {
3969 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3970 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3972 /* Now we know that nxt2 is the only contents: */
3973 oscan->flags = (U8)ARG(nxt);
3975 OP(nxt1) = NOTHING; /* was OPEN. */
3978 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3979 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3980 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3981 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3982 OP(nxt + 1) = OPTIMIZED; /* was count. */
3983 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3988 /* Try optimization CURLYX => CURLYM. */
3989 if ( OP(oscan) == CURLYX && data
3990 && !(data->flags & SF_HAS_PAR)
3991 && !(data->flags & SF_HAS_EVAL)
3992 && !deltanext /* atom is fixed width */
3993 && minnext != 0 /* CURLYM can't handle zero width */
3994 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3996 /* XXXX How to optimize if data == 0? */
3997 /* Optimize to a simpler form. */
3998 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4002 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4003 && (OP(nxt2) != WHILEM))
4005 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4006 /* Need to optimize away parenths. */
4007 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4008 /* Set the parenth number. */
4009 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4011 oscan->flags = (U8)ARG(nxt);
4012 if (RExC_open_parens) {
4013 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4014 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4016 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4017 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4020 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4021 OP(nxt + 1) = OPTIMIZED; /* was count. */
4022 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4023 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4026 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4027 regnode *nnxt = regnext(nxt1);
4029 if (reg_off_by_arg[OP(nxt1)])
4030 ARG_SET(nxt1, nxt2 - nxt1);
4031 else if (nxt2 - nxt1 < U16_MAX)
4032 NEXT_OFF(nxt1) = nxt2 - nxt1;
4034 OP(nxt) = NOTHING; /* Cannot beautify */
4039 /* Optimize again: */
4040 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4041 NULL, stopparen, recursed, NULL, 0,depth+1);
4046 else if ((OP(oscan) == CURLYX)
4047 && (flags & SCF_WHILEM_VISITED_POS)
4048 /* See the comment on a similar expression above.
4049 However, this time it's not a subexpression
4050 we care about, but the expression itself. */
4051 && (maxcount == REG_INFTY)
4052 && data && ++data->whilem_c < 16) {
4053 /* This stays as CURLYX, we can put the count/of pair. */
4054 /* Find WHILEM (as in regexec.c) */
4055 regnode *nxt = oscan + NEXT_OFF(oscan);
4057 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4059 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4060 | (RExC_whilem_seen << 4)); /* On WHILEM */
4062 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4064 if (flags & SCF_DO_SUBSTR) {
4065 SV *last_str = NULL;
4066 int counted = mincount != 0;
4068 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4069 #if defined(SPARC64_GCC_WORKAROUND)
4072 const char *s = NULL;
4075 if (pos_before >= data->last_start_min)
4078 b = data->last_start_min;
4081 s = SvPV_const(data->last_found, l);
4082 old = b - data->last_start_min;
4085 I32 b = pos_before >= data->last_start_min
4086 ? pos_before : data->last_start_min;
4088 const char * const s = SvPV_const(data->last_found, l);
4089 I32 old = b - data->last_start_min;
4093 old = utf8_hop((U8*)s, old) - (U8*)s;
4095 /* Get the added string: */
4096 last_str = newSVpvn_utf8(s + old, l, UTF);
4097 if (deltanext == 0 && pos_before == b) {
4098 /* What was added is a constant string */
4100 SvGROW(last_str, (mincount * l) + 1);
4101 repeatcpy(SvPVX(last_str) + l,
4102 SvPVX_const(last_str), l, mincount - 1);
4103 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4104 /* Add additional parts. */
4105 SvCUR_set(data->last_found,
4106 SvCUR(data->last_found) - l);
4107 sv_catsv(data->last_found, last_str);
4109 SV * sv = data->last_found;
4111 SvUTF8(sv) && SvMAGICAL(sv) ?
4112 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4113 if (mg && mg->mg_len >= 0)
4114 mg->mg_len += CHR_SVLEN(last_str) - l;
4116 data->last_end += l * (mincount - 1);
4119 /* start offset must point into the last copy */
4120 data->last_start_min += minnext * (mincount - 1);
4121 data->last_start_max += is_inf ? I32_MAX
4122 : (maxcount - 1) * (minnext + data->pos_delta);
4125 /* It is counted once already... */
4126 data->pos_min += minnext * (mincount - counted);
4128 PerlIO_printf(Perl_debug_log, "counted=%d deltanext=%d I32_MAX=%d minnext=%d maxcount=%d mincount=%d\n",
4129 counted, deltanext, I32_MAX, minnext, maxcount, mincount);
4130 if (deltanext != I32_MAX)
4131 PerlIO_printf(Perl_debug_log, "LHS=%d RHS=%d\n", -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount, I32_MAX - data->pos_delta);
4133 if (deltanext == I32_MAX || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= I32_MAX - data->pos_delta)
4134 data->pos_delta = I32_MAX;
4136 data->pos_delta += - counted * deltanext +
4137 (minnext + deltanext) * maxcount - minnext * mincount;
4138 if (mincount != maxcount) {
4139 /* Cannot extend fixed substrings found inside
4141 SCAN_COMMIT(pRExC_state,data,minlenp);
4142 if (mincount && last_str) {
4143 SV * const sv = data->last_found;
4144 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4145 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4149 sv_setsv(sv, last_str);
4150 data->last_end = data->pos_min;
4151 data->last_start_min =
4152 data->pos_min - CHR_SVLEN(last_str);
4153 data->last_start_max = is_inf
4155 : data->pos_min + data->pos_delta
4156 - CHR_SVLEN(last_str);
4158 data->longest = &(data->longest_float);
4160 SvREFCNT_dec(last_str);
4162 if (data && (fl & SF_HAS_EVAL))
4163 data->flags |= SF_HAS_EVAL;
4164 optimize_curly_tail:
4165 if (OP(oscan) != CURLYX) {
4166 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4168 NEXT_OFF(oscan) += NEXT_OFF(next);
4171 default: /* REF, and CLUMP only? */
4172 if (flags & SCF_DO_SUBSTR) {
4173 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4174 data->longest = &(data->longest_float);
4176 is_inf = is_inf_internal = 1;
4177 if (flags & SCF_DO_STCLASS_OR)
4178 cl_anything(pRExC_state, data->start_class);
4179 flags &= ~SCF_DO_STCLASS;
4183 else if (OP(scan) == LNBREAK) {
4184 if (flags & SCF_DO_STCLASS) {
4186 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4187 if (flags & SCF_DO_STCLASS_AND) {
4188 for (value = 0; value < 256; value++)
4189 if (!is_VERTWS_cp(value))
4190 ANYOF_BITMAP_CLEAR(data->start_class, value);
4193 for (value = 0; value < 256; value++)
4194 if (is_VERTWS_cp(value))
4195 ANYOF_BITMAP_SET(data->start_class, value);
4197 if (flags & SCF_DO_STCLASS_OR)
4198 cl_and(data->start_class, and_withp);
4199 flags &= ~SCF_DO_STCLASS;
4202 delta++; /* Because of the 2 char string cr-lf */
4203 if (flags & SCF_DO_SUBSTR) {
4204 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4206 data->pos_delta += 1;
4207 data->longest = &(data->longest_float);
4210 else if (REGNODE_SIMPLE(OP(scan))) {
4213 if (flags & SCF_DO_SUBSTR) {
4214 SCAN_COMMIT(pRExC_state,data,minlenp);
4218 if (flags & SCF_DO_STCLASS) {
4220 CLEAR_SSC_EOS(data->start_class); /* No match on empty */
4222 /* Some of the logic below assumes that switching
4223 locale on will only add false positives. */
4224 switch (PL_regkind[OP(scan)]) {
4230 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan));
4233 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4234 cl_anything(pRExC_state, data->start_class);
4237 if (OP(scan) == SANY)
4239 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4240 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4241 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4242 cl_anything(pRExC_state, data->start_class);
4244 if (flags & SCF_DO_STCLASS_AND || !value)
4245 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4248 if (flags & SCF_DO_STCLASS_AND)
4249 cl_and(data->start_class,
4250 (struct regnode_charclass_class*)scan);
4252 cl_or(pRExC_state, data->start_class,
4253 (struct regnode_charclass_class*)scan);
4261 classnum = FLAGS(scan);
4262 if (flags & SCF_DO_STCLASS_AND) {
4263 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4264 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum) + 1);
4265 for (value = 0; value < loop_max; value++) {
4266 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4267 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4273 if (data->start_class->flags & ANYOF_LOCALE) {
4274 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum));
4278 /* Even if under locale, set the bits for non-locale
4279 * in case it isn't a true locale-node. This will
4280 * create false positives if it truly is locale */
4281 for (value = 0; value < loop_max; value++) {
4282 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4283 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4295 classnum = FLAGS(scan);
4296 if (flags & SCF_DO_STCLASS_AND) {
4297 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4298 ANYOF_CLASS_CLEAR(data->start_class, classnum_to_namedclass(classnum));
4299 for (value = 0; value < loop_max; value++) {
4300 if (_generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4301 ANYOF_BITMAP_CLEAR(data->start_class, UNI_TO_NATIVE(value));
4307 if (data->start_class->flags & ANYOF_LOCALE) {
4308 ANYOF_CLASS_SET(data->start_class, classnum_to_namedclass(classnum) + 1);
4312 /* Even if under locale, set the bits for non-locale in
4313 * case it isn't a true locale-node. This will create
4314 * false positives if it truly is locale */
4315 for (value = 0; value < loop_max; value++) {
4316 if (! _generic_isCC(UNI_TO_NATIVE(value), classnum)) {
4317 ANYOF_BITMAP_SET(data->start_class, UNI_TO_NATIVE(value));
4320 if (PL_regkind[OP(scan)] == NPOSIXD) {
4321 data->start_class->flags |= ANYOF_NON_UTF8_LATIN1_ALL;
4327 if (flags & SCF_DO_STCLASS_OR)
4328 cl_and(data->start_class, and_withp);
4329 flags &= ~SCF_DO_STCLASS;
4332 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4333 data->flags |= (OP(scan) == MEOL
4336 SCAN_COMMIT(pRExC_state, data, minlenp);
4339 else if ( PL_regkind[OP(scan)] == BRANCHJ
4340 /* Lookbehind, or need to calculate parens/evals/stclass: */
4341 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4342 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4343 if ( OP(scan) == UNLESSM &&
4345 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4346 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4349 regnode *upto= regnext(scan);
4351 SV * const mysv_val=sv_newmortal();
4352 DEBUG_STUDYDATA("OPFAIL",data,depth);
4354 /*DEBUG_PARSE_MSG("opfail");*/
4355 regprop(RExC_rx, mysv_val, upto);
4356 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4357 SvPV_nolen_const(mysv_val),
4358 (IV)REG_NODE_NUM(upto),
4363 NEXT_OFF(scan) = upto - scan;
4364 for (opt= scan + 1; opt < upto ; opt++)
4365 OP(opt) = OPTIMIZED;
4369 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4370 || OP(scan) == UNLESSM )
4372 /* Negative Lookahead/lookbehind
4373 In this case we can't do fixed string optimisation.
4376 I32 deltanext, minnext, fake = 0;
4378 struct regnode_charclass_class intrnl;
4381 data_fake.flags = 0;
4383 data_fake.whilem_c = data->whilem_c;
4384 data_fake.last_closep = data->last_closep;
4387 data_fake.last_closep = &fake;
4388 data_fake.pos_delta = delta;
4389 if ( flags & SCF_DO_STCLASS && !scan->flags
4390 && OP(scan) == IFMATCH ) { /* Lookahead */
4391 cl_init(pRExC_state, &intrnl);
4392 data_fake.start_class = &intrnl;
4393 f |= SCF_DO_STCLASS_AND;
4395 if (flags & SCF_WHILEM_VISITED_POS)
4396 f |= SCF_WHILEM_VISITED_POS;
4397 next = regnext(scan);
4398 nscan = NEXTOPER(NEXTOPER(scan));
4399 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4400 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4403 FAIL("Variable length lookbehind not implemented");
4405 else if (minnext > (I32)U8_MAX) {
4406 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4408 scan->flags = (U8)minnext;
4411 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4413 if (data_fake.flags & SF_HAS_EVAL)
4414 data->flags |= SF_HAS_EVAL;
4415 data->whilem_c = data_fake.whilem_c;
4417 if (f & SCF_DO_STCLASS_AND) {
4418 if (flags & SCF_DO_STCLASS_OR) {
4419 /* OR before, AND after: ideally we would recurse with
4420 * data_fake to get the AND applied by study of the
4421 * remainder of the pattern, and then derecurse;
4422 * *** HACK *** for now just treat as "no information".
4423 * See [perl #56690].
4425 cl_init(pRExC_state, data->start_class);
4427 /* AND before and after: combine and continue */
4428 const int was = TEST_SSC_EOS(data->start_class);
4430 cl_and(data->start_class, &intrnl);
4432 SET_SSC_EOS(data->start_class);
4436 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4438 /* Positive Lookahead/lookbehind
4439 In this case we can do fixed string optimisation,
4440 but we must be careful about it. Note in the case of
4441 lookbehind the positions will be offset by the minimum
4442 length of the pattern, something we won't know about
4443 until after the recurse.
4445 I32 deltanext, fake = 0;
4447 struct regnode_charclass_class intrnl;
4449 /* We use SAVEFREEPV so that when the full compile
4450 is finished perl will clean up the allocated
4451 minlens when it's all done. This way we don't
4452 have to worry about freeing them when we know
4453 they wont be used, which would be a pain.
4456 Newx( minnextp, 1, I32 );
4457 SAVEFREEPV(minnextp);
4460 StructCopy(data, &data_fake, scan_data_t);
4461 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4464 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4465 data_fake.last_found=newSVsv(data->last_found);
4469 data_fake.last_closep = &fake;
4470 data_fake.flags = 0;
4471 data_fake.pos_delta = delta;
4473 data_fake.flags |= SF_IS_INF;
4474 if ( flags & SCF_DO_STCLASS && !scan->flags
4475 && OP(scan) == IFMATCH ) { /* Lookahead */
4476 cl_init(pRExC_state, &intrnl);
4477 data_fake.start_class = &intrnl;
4478 f |= SCF_DO_STCLASS_AND;
4480 if (flags & SCF_WHILEM_VISITED_POS)
4481 f |= SCF_WHILEM_VISITED_POS;
4482 next = regnext(scan);
4483 nscan = NEXTOPER(NEXTOPER(scan));
4485 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4486 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4489 FAIL("Variable length lookbehind not implemented");
4491 else if (*minnextp > (I32)U8_MAX) {
4492 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4494 scan->flags = (U8)*minnextp;
4499 if (f & SCF_DO_STCLASS_AND) {
4500 const int was = TEST_SSC_EOS(data.start_class);
4502 cl_and(data->start_class, &intrnl);
4504 SET_SSC_EOS(data->start_class);
4507 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4509 if (data_fake.flags & SF_HAS_EVAL)
4510 data->flags |= SF_HAS_EVAL;
4511 data->whilem_c = data_fake.whilem_c;
4512 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4513 if (RExC_rx->minlen<*minnextp)
4514 RExC_rx->minlen=*minnextp;
4515 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4516 SvREFCNT_dec_NN(data_fake.last_found);
4518 if ( data_fake.minlen_fixed != minlenp )
4520 data->offset_fixed= data_fake.offset_fixed;
4521 data->minlen_fixed= data_fake.minlen_fixed;
4522 data->lookbehind_fixed+= scan->flags;
4524 if ( data_fake.minlen_float != minlenp )
4526 data->minlen_float= data_fake.minlen_float;
4527 data->offset_float_min=data_fake.offset_float_min;
4528 data->offset_float_max=data_fake.offset_float_max;
4529 data->lookbehind_float+= scan->flags;
4536 else if (OP(scan) == OPEN) {
4537 if (stopparen != (I32)ARG(scan))
4540 else if (OP(scan) == CLOSE) {
4541 if (stopparen == (I32)ARG(scan)) {
4544 if ((I32)ARG(scan) == is_par) {
4545 next = regnext(scan);
4547 if ( next && (OP(next) != WHILEM) && next < last)
4548 is_par = 0; /* Disable optimization */
4551 *(data->last_closep) = ARG(scan);
4553 else if (OP(scan) == EVAL) {
4555 data->flags |= SF_HAS_EVAL;
4557 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4558 if (flags & SCF_DO_SUBSTR) {
4559 SCAN_COMMIT(pRExC_state,data,minlenp);
4560 flags &= ~SCF_DO_SUBSTR;
4562 if (data && OP(scan)==ACCEPT) {
4563 data->flags |= SCF_SEEN_ACCEPT;
4568 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4570 if (flags & SCF_DO_SUBSTR) {
4571 SCAN_COMMIT(pRExC_state,data,minlenp);
4572 data->longest = &(data->longest_float);
4574 is_inf = is_inf_internal = 1;
4575 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4576 cl_anything(pRExC_state, data->start_class);
4577 flags &= ~SCF_DO_STCLASS;
4579 else if (OP(scan) == GPOS) {
4580 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4581 !(delta || is_inf || (data && data->pos_delta)))
4583 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4584 RExC_rx->extflags |= RXf_ANCH_GPOS;
4585 if (RExC_rx->gofs < (U32)min)
4586 RExC_rx->gofs = min;
4588 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4592 #ifdef TRIE_STUDY_OPT
4593 #ifdef FULL_TRIE_STUDY
4594 else if (PL_regkind[OP(scan)] == TRIE) {
4595 /* NOTE - There is similar code to this block above for handling
4596 BRANCH nodes on the initial study. If you change stuff here
4598 regnode *trie_node= scan;
4599 regnode *tail= regnext(scan);
4600 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4601 I32 max1 = 0, min1 = I32_MAX;
4602 struct regnode_charclass_class accum;
4604 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4605 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4606 if (flags & SCF_DO_STCLASS)
4607 cl_init_zero(pRExC_state, &accum);
4613 const regnode *nextbranch= NULL;
4616 for ( word=1 ; word <= trie->wordcount ; word++)
4618 I32 deltanext=0, minnext=0, f = 0, fake;
4619 struct regnode_charclass_class this_class;
4621 data_fake.flags = 0;
4623 data_fake.whilem_c = data->whilem_c;
4624 data_fake.last_closep = data->last_closep;
4627 data_fake.last_closep = &fake;
4628 data_fake.pos_delta = delta;
4629 if (flags & SCF_DO_STCLASS) {
4630 cl_init(pRExC_state, &this_class);
4631 data_fake.start_class = &this_class;
4632 f = SCF_DO_STCLASS_AND;
4634 if (flags & SCF_WHILEM_VISITED_POS)
4635 f |= SCF_WHILEM_VISITED_POS;
4637 if (trie->jump[word]) {
4639 nextbranch = trie_node + trie->jump[0];
4640 scan= trie_node + trie->jump[word];
4641 /* We go from the jump point to the branch that follows
4642 it. Note this means we need the vestigal unused branches
4643 even though they arent otherwise used.
4645 minnext = study_chunk(pRExC_state, &scan, minlenp,
4646 &deltanext, (regnode *)nextbranch, &data_fake,
4647 stopparen, recursed, NULL, f,depth+1);
4649 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4650 nextbranch= regnext((regnode*)nextbranch);
4652 if (min1 > (I32)(minnext + trie->minlen))
4653 min1 = minnext + trie->minlen;
4654 if (deltanext == I32_MAX) {
4655 is_inf = is_inf_internal = 1;
4657 } else if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4658 max1 = minnext + deltanext + trie->maxlen;
4660 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4662 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4663 if ( stopmin > min + min1)
4664 stopmin = min + min1;
4665 flags &= ~SCF_DO_SUBSTR;
4667 data->flags |= SCF_SEEN_ACCEPT;
4670 if (data_fake.flags & SF_HAS_EVAL)
4671 data->flags |= SF_HAS_EVAL;
4672 data->whilem_c = data_fake.whilem_c;
4674 if (flags & SCF_DO_STCLASS)
4675 cl_or(pRExC_state, &accum, &this_class);
4678 if (flags & SCF_DO_SUBSTR) {
4679 data->pos_min += min1;
4680 data->pos_delta += max1 - min1;
4681 if (max1 != min1 || is_inf)
4682 data->longest = &(data->longest_float);
4685 delta += max1 - min1;
4686 if (flags & SCF_DO_STCLASS_OR) {
4687 cl_or(pRExC_state, data->start_class, &accum);
4689 cl_and(data->start_class, and_withp);
4690 flags &= ~SCF_DO_STCLASS;
4693 else if (flags & SCF_DO_STCLASS_AND) {
4695 cl_and(data->start_class, &accum);
4696 flags &= ~SCF_DO_STCLASS;
4699 /* Switch to OR mode: cache the old value of
4700 * data->start_class */
4702 StructCopy(data->start_class, and_withp,
4703 struct regnode_charclass_class);
4704 flags &= ~SCF_DO_STCLASS_AND;
4705 StructCopy(&accum, data->start_class,
4706 struct regnode_charclass_class);
4707 flags |= SCF_DO_STCLASS_OR;
4708 SET_SSC_EOS(data->start_class);
4715 else if (PL_regkind[OP(scan)] == TRIE) {
4716 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4719 min += trie->minlen;
4720 delta += (trie->maxlen - trie->minlen);
4721 flags &= ~SCF_DO_STCLASS; /* xxx */
4722 if (flags & SCF_DO_SUBSTR) {
4723 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4724 data->pos_min += trie->minlen;
4725 data->pos_delta += (trie->maxlen - trie->minlen);
4726 if (trie->maxlen != trie->minlen)
4727 data->longest = &(data->longest_float);
4729 if (trie->jump) /* no more substrings -- for now /grr*/
4730 flags &= ~SCF_DO_SUBSTR;
4732 #endif /* old or new */
4733 #endif /* TRIE_STUDY_OPT */
4735 /* Else: zero-length, ignore. */
4736 scan = regnext(scan);
4741 stopparen = frame->stop;
4742 frame = frame->prev;
4743 goto fake_study_recurse;
4748 DEBUG_STUDYDATA("pre-fin:",data,depth);
4751 *deltap = is_inf_internal ? I32_MAX : delta;
4752 if (flags & SCF_DO_SUBSTR && is_inf)
4753 data->pos_delta = I32_MAX - data->pos_min;
4754 if (is_par > (I32)U8_MAX)
4756 if (is_par && pars==1 && data) {
4757 data->flags |= SF_IN_PAR;
4758 data->flags &= ~SF_HAS_PAR;
4760 else if (pars && data) {
4761 data->flags |= SF_HAS_PAR;
4762 data->flags &= ~SF_IN_PAR;
4764 if (flags & SCF_DO_STCLASS_OR)
4765 cl_and(data->start_class, and_withp);
4766 if (flags & SCF_TRIE_RESTUDY)
4767 data->flags |= SCF_TRIE_RESTUDY;
4769 DEBUG_STUDYDATA("post-fin:",data,depth);
4771 return min < stopmin ? min : stopmin;
4775 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4777 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4779 PERL_ARGS_ASSERT_ADD_DATA;
4781 Renewc(RExC_rxi->data,
4782 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4783 char, struct reg_data);
4785 Renew(RExC_rxi->data->what, count + n, U8);
4787 Newx(RExC_rxi->data->what, n, U8);
4788 RExC_rxi->data->count = count + n;
4789 Copy(s, RExC_rxi->data->what + count, n, U8);
4793 /*XXX: todo make this not included in a non debugging perl */
4794 #ifndef PERL_IN_XSUB_RE
4796 Perl_reginitcolors(pTHX)
4799 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4801 char *t = savepv(s);
4805 t = strchr(t, '\t');
4811 PL_colors[i] = t = (char *)"";
4816 PL_colors[i++] = (char *)"";
4823 #ifdef TRIE_STUDY_OPT
4824 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
4827 (data.flags & SCF_TRIE_RESTUDY) \
4835 #define CHECK_RESTUDY_GOTO_butfirst
4839 * pregcomp - compile a regular expression into internal code
4841 * Decides which engine's compiler to call based on the hint currently in
4845 #ifndef PERL_IN_XSUB_RE
4847 /* return the currently in-scope regex engine (or the default if none) */
4849 regexp_engine const *
4850 Perl_current_re_engine(pTHX)
4854 if (IN_PERL_COMPILETIME) {
4855 HV * const table = GvHV(PL_hintgv);
4859 return &PL_core_reg_engine;
4860 ptr = hv_fetchs(table, "regcomp", FALSE);
4861 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4862 return &PL_core_reg_engine;
4863 return INT2PTR(regexp_engine*,SvIV(*ptr));
4867 if (!PL_curcop->cop_hints_hash)
4868 return &PL_core_reg_engine;
4869 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4870 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4871 return &PL_core_reg_engine;
4872 return INT2PTR(regexp_engine*,SvIV(ptr));
4878 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4881 regexp_engine const *eng = current_re_engine();
4882 GET_RE_DEBUG_FLAGS_DECL;
4884 PERL_ARGS_ASSERT_PREGCOMP;
4886 /* Dispatch a request to compile a regexp to correct regexp engine. */
4888 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4891 return CALLREGCOMP_ENG(eng, pattern, flags);
4895 /* public(ish) entry point for the perl core's own regex compiling code.
4896 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4897 * pattern rather than a list of OPs, and uses the internal engine rather
4898 * than the current one */
4901 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4903 SV *pat = pattern; /* defeat constness! */
4904 PERL_ARGS_ASSERT_RE_COMPILE;
4905 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4906 #ifdef PERL_IN_XSUB_RE
4909 &PL_core_reg_engine,
4911 NULL, NULL, rx_flags, 0);
4915 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
4916 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
4917 * point to the realloced string and length.
4919 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
4923 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
4924 char **pat_p, STRLEN *plen_p, int num_code_blocks)
4926 U8 *const src = (U8*)*pat_p;
4929 STRLEN s = 0, d = 0;
4931 GET_RE_DEBUG_FLAGS_DECL;
4933 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
4934 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
4936 Newx(dst, *plen_p * 2 + 1, U8);
4938 while (s < *plen_p) {
4939 const UV uv = NATIVE_TO_ASCII(src[s]);
4940 if (UNI_IS_INVARIANT(uv))
4941 dst[d] = (U8)UTF_TO_NATIVE(uv);
4943 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
4944 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
4946 if (n < num_code_blocks) {
4947 if (!do_end && pRExC_state->code_blocks[n].start == s) {
4948 pRExC_state->code_blocks[n].start = d;
4949 assert(dst[d] == '(');
4952 else if (do_end && pRExC_state->code_blocks[n].end == s) {
4953 pRExC_state->code_blocks[n].end = d;
4954 assert(dst[d] == ')');
4964 *pat_p = (char*) dst;
4966 RExC_orig_utf8 = RExC_utf8 = 1;
4971 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
4972 * while recording any code block indices, and handling overloading,
4973 * nested qr// objects etc. If pat is null, it will allocate a new
4974 * string, or just return the first arg, if there's only one.
4976 * Returns the malloced/updated pat.
4977 * patternp and pat_count is the array of SVs to be concatted;
4978 * oplist is the optional list of ops that generated the SVs;
4979 * recompile_p is a pointer to a boolean that will be set if
4980 * the regex will need to be recompiled.
4981 * delim, if non-null is an SV that will be inserted between each element
4985 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
4986 SV *pat, SV ** const patternp, int pat_count,
4987 OP *oplist, bool *recompile_p, SV *delim)
4991 bool use_delim = FALSE;
4992 bool alloced = FALSE;
4994 /* if we know we have at least two args, create an empty string,
4995 * then concatenate args to that. For no args, return an empty string */
4996 if (!pat && pat_count != 1) {
4997 pat = newSVpvn("", 0);
5002 for (svp = patternp; svp < patternp + pat_count; svp++) {
5005 STRLEN orig_patlen = 0;
5007 SV *msv = use_delim ? delim : *svp;
5009 /* if we've got a delimiter, we go round the loop twice for each
5010 * svp slot (except the last), using the delimiter the second
5019 if (SvTYPE(msv) == SVt_PVAV) {
5020 /* we've encountered an interpolated array within
5021 * the pattern, e.g. /...@a..../. Expand the list of elements,
5022 * then recursively append elements.
5023 * The code in this block is based on S_pushav() */
5025 AV *const av = (AV*)msv;
5026 const I32 maxarg = AvFILL(av) + 1;
5030 assert(oplist->op_type == OP_PADAV
5031 || oplist->op_type == OP_RV2AV);
5032 oplist = oplist->op_sibling;;
5035 if (SvRMAGICAL(av)) {
5038 Newx(array, maxarg, SV*);
5040 for (i=0; i < (U32)maxarg; i++) {
5041 SV ** const svp = av_fetch(av, i, FALSE);
5042 array[i] = svp ? *svp : &PL_sv_undef;
5046 array = AvARRAY(av);
5048 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5049 array, maxarg, NULL, recompile_p,
5051 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5057 /* we make the assumption here that each op in the list of
5058 * op_siblings maps to one SV pushed onto the stack,
5059 * except for code blocks, with have both an OP_NULL and
5061 * This allows us to match up the list of SVs against the
5062 * list of OPs to find the next code block.
5064 * Note that PUSHMARK PADSV PADSV ..
5066 * PADRANGE PADSV PADSV ..
5067 * so the alignment still works. */
5070 if (oplist->op_type == OP_NULL
5071 && (oplist->op_flags & OPf_SPECIAL))
5073 assert(n < pRExC_state->num_code_blocks);
5074 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5075 pRExC_state->code_blocks[n].block = oplist;
5076 pRExC_state->code_blocks[n].src_regex = NULL;
5079 oplist = oplist->op_sibling; /* skip CONST */
5082 oplist = oplist->op_sibling;;
5085 /* apply magic and QR overloading to arg */
5088 if (SvROK(msv) && SvAMAGIC(msv)) {
5089 SV *sv = AMG_CALLunary(msv, regexp_amg);
5093 if (SvTYPE(sv) != SVt_REGEXP)
5094 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5099 /* try concatenation overload ... */
5100 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5101 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5104 /* overloading involved: all bets are off over literal
5105 * code. Pretend we haven't seen it */
5106 pRExC_state->num_code_blocks -= n;
5110 /* ... or failing that, try "" overload */
5111 while (SvAMAGIC(msv)
5112 && (sv = AMG_CALLunary(msv, string_amg))
5116 && SvRV(msv) == SvRV(sv))
5121 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5125 /* this is a partially unrolled
5126 * sv_catsv_nomg(pat, msv);
5127 * that allows us to adjust code block indices if
5130 char *dst = SvPV_force_nomg(pat, dlen);
5132 if (SvUTF8(msv) && !SvUTF8(pat)) {
5133 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
5134 sv_setpvn(pat, dst, dlen);
5137 sv_catsv_nomg(pat, msv);
5144 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5147 /* extract any code blocks within any embedded qr//'s */
5148 if (rx && SvTYPE(rx) == SVt_REGEXP
5149 && RX_ENGINE((REGEXP*)rx)->op_comp)
5152 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
5153 if (ri->num_code_blocks) {
5155 /* the presence of an embedded qr// with code means
5156 * we should always recompile: the text of the
5157 * qr// may not have changed, but it may be a
5158 * different closure than last time */
5160 Renew(pRExC_state->code_blocks,
5161 pRExC_state->num_code_blocks + ri->num_code_blocks,
5162 struct reg_code_block);
5163 pRExC_state->num_code_blocks += ri->num_code_blocks;
5165 for (i=0; i < ri->num_code_blocks; i++) {
5166 struct reg_code_block *src, *dst;
5167 STRLEN offset = orig_patlen
5168 + ReANY((REGEXP *)rx)->pre_prefix;
5169 assert(n < pRExC_state->num_code_blocks);
5170 src = &ri->code_blocks[i];
5171 dst = &pRExC_state->code_blocks[n];
5172 dst->start = src->start + offset;
5173 dst->end = src->end + offset;
5174 dst->block = src->block;
5175 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5184 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
5193 /* see if there are any run-time code blocks in the pattern.
5194 * False positives are allowed */
5197 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5198 char *pat, STRLEN plen)
5203 for (s = 0; s < plen; s++) {
5204 if (n < pRExC_state->num_code_blocks
5205 && s == pRExC_state->code_blocks[n].start)
5207 s = pRExC_state->code_blocks[n].end;
5211 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5213 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
5215 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
5222 /* Handle run-time code blocks. We will already have compiled any direct
5223 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5224 * copy of it, but with any literal code blocks blanked out and
5225 * appropriate chars escaped; then feed it into
5227 * eval "qr'modified_pattern'"
5231 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5235 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
5237 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5238 * and merge them with any code blocks of the original regexp.
5240 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5241 * instead, just save the qr and return FALSE; this tells our caller that
5242 * the original pattern needs upgrading to utf8.
5246 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5247 char *pat, STRLEN plen)
5251 GET_RE_DEBUG_FLAGS_DECL;
5253 if (pRExC_state->runtime_code_qr) {
5254 /* this is the second time we've been called; this should
5255 * only happen if the main pattern got upgraded to utf8
5256 * during compilation; re-use the qr we compiled first time
5257 * round (which should be utf8 too)
5259 qr = pRExC_state->runtime_code_qr;
5260 pRExC_state->runtime_code_qr = NULL;
5261 assert(RExC_utf8 && SvUTF8(qr));
5267 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5271 /* determine how many extra chars we need for ' and \ escaping */
5272 for (s = 0; s < plen; s++) {
5273 if (pat[s] == '\'' || pat[s] == '\\')
5277 Newx(newpat, newlen, char);
5279 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5281 for (s = 0; s < plen; s++) {
5282 if (n < pRExC_state->num_code_blocks
5283 && s == pRExC_state->code_blocks[n].start)
5285 /* blank out literal code block */
5286 assert(pat[s] == '(');
5287 while (s <= pRExC_state->code_blocks[n].end) {
5295 if (pat[s] == '\'' || pat[s] == '\\')
5300 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5304 PerlIO_printf(Perl_debug_log,
5305 "%sre-parsing pattern for runtime code:%s %s\n",
5306 PL_colors[4],PL_colors[5],newpat);
5309 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5315 PUSHSTACKi(PERLSI_REQUIRE);
5316 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
5317 * parsing qr''; normally only q'' does this. It also alters
5319 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
5320 SvREFCNT_dec_NN(sv);
5325 SV * const errsv = ERRSV;
5326 if (SvTRUE_NN(errsv))
5328 Safefree(pRExC_state->code_blocks);
5329 /* use croak_sv ? */
5330 Perl_croak_nocontext("%s", SvPV_nolen_const(errsv));
5333 assert(SvROK(qr_ref));
5335 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5336 /* the leaving below frees the tmp qr_ref.
5337 * Give qr a life of its own */
5345 if (!RExC_utf8 && SvUTF8(qr)) {
5346 /* first time through; the pattern got upgraded; save the
5347 * qr for the next time through */
5348 assert(!pRExC_state->runtime_code_qr);
5349 pRExC_state->runtime_code_qr = qr;
5354 /* extract any code blocks within the returned qr// */
5357 /* merge the main (r1) and run-time (r2) code blocks into one */
5359 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
5360 struct reg_code_block *new_block, *dst;
5361 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5364 if (!r2->num_code_blocks) /* we guessed wrong */
5366 SvREFCNT_dec_NN(qr);
5371 r1->num_code_blocks + r2->num_code_blocks,
5372 struct reg_code_block);
5375 while ( i1 < r1->num_code_blocks
5376 || i2 < r2->num_code_blocks)
5378 struct reg_code_block *src;
5381 if (i1 == r1->num_code_blocks) {
5382 src = &r2->code_blocks[i2++];
5385 else if (i2 == r2->num_code_blocks)
5386 src = &r1->code_blocks[i1++];
5387 else if ( r1->code_blocks[i1].start
5388 < r2->code_blocks[i2].start)
5390 src = &r1->code_blocks[i1++];
5391 assert(src->end < r2->code_blocks[i2].start);
5394 assert( r1->code_blocks[i1].start
5395 > r2->code_blocks[i2].start);
5396 src = &r2->code_blocks[i2++];
5398 assert(src->end < r1->code_blocks[i1].start);
5401 assert(pat[src->start] == '(');
5402 assert(pat[src->end] == ')');
5403 dst->start = src->start;
5404 dst->end = src->end;
5405 dst->block = src->block;
5406 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5410 r1->num_code_blocks += r2->num_code_blocks;
5411 Safefree(r1->code_blocks);
5412 r1->code_blocks = new_block;
5415 SvREFCNT_dec_NN(qr);
5421 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest, SV** rx_utf8, SV** rx_substr, I32* rx_end_shift, I32 lookbehind, I32 offset, I32 *minlen, STRLEN longest_length, bool eol, bool meol)
5423 /* This is the common code for setting up the floating and fixed length
5424 * string data extracted from Perl_re_op_compile() below. Returns a boolean
5425 * as to whether succeeded or not */
5429 if (! (longest_length
5430 || (eol /* Can't have SEOL and MULTI */
5431 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5433 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5434 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5439 /* copy the information about the longest from the reg_scan_data
5440 over to the program. */
5441 if (SvUTF8(sv_longest)) {
5442 *rx_utf8 = sv_longest;
5445 *rx_substr = sv_longest;
5448 /* end_shift is how many chars that must be matched that
5449 follow this item. We calculate it ahead of time as once the
5450 lookbehind offset is added in we lose the ability to correctly
5452 ml = minlen ? *(minlen) : (I32)longest_length;
5453 *rx_end_shift = ml - offset
5454 - longest_length + (SvTAIL(sv_longest) != 0)
5457 t = (eol/* Can't have SEOL and MULTI */
5458 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5459 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5465 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5466 * regular expression into internal code.
5467 * The pattern may be passed either as:
5468 * a list of SVs (patternp plus pat_count)
5469 * a list of OPs (expr)
5470 * If both are passed, the SV list is used, but the OP list indicates
5471 * which SVs are actually pre-compiled code blocks
5473 * The SVs in the list have magic and qr overloading applied to them (and
5474 * the list may be modified in-place with replacement SVs in the latter
5477 * If the pattern hasn't changed from old_re, then old_re will be
5480 * eng is the current engine. If that engine has an op_comp method, then
5481 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5482 * do the initial concatenation of arguments and pass on to the external
5485 * If is_bare_re is not null, set it to a boolean indicating whether the
5486 * arg list reduced (after overloading) to a single bare regex which has
5487 * been returned (i.e. /$qr/).
5489 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5491 * pm_flags contains the PMf_* flags, typically based on those from the
5492 * pm_flags field of the related PMOP. Currently we're only interested in
5493 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5495 * We can't allocate space until we know how big the compiled form will be,
5496 * but we can't compile it (and thus know how big it is) until we've got a
5497 * place to put the code. So we cheat: we compile it twice, once with code
5498 * generation turned off and size counting turned on, and once "for real".
5499 * This also means that we don't allocate space until we are sure that the
5500 * thing really will compile successfully, and we never have to move the
5501 * code and thus invalidate pointers into it. (Note that it has to be in
5502 * one piece because free() must be able to free it all.) [NB: not true in perl]
5504 * Beware that the optimization-preparation code in here knows about some
5505 * of the structure of the compiled regexp. [I'll say.]
5509 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5510 OP *expr, const regexp_engine* eng, REGEXP *old_re,
5511 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5516 regexp_internal *ri;
5524 SV *code_blocksv = NULL;
5525 SV** new_patternp = patternp;
5527 /* these are all flags - maybe they should be turned
5528 * into a single int with different bit masks */
5529 I32 sawlookahead = 0;
5534 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5536 bool runtime_code = 0;
5538 RExC_state_t RExC_state;
5539 RExC_state_t * const pRExC_state = &RExC_state;
5540 #ifdef TRIE_STUDY_OPT
5542 RExC_state_t copyRExC_state;
5544 GET_RE_DEBUG_FLAGS_DECL;
5546 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5548 DEBUG_r(if (!PL_colorset) reginitcolors());
5550 #ifndef PERL_IN_XSUB_RE
5551 /* Initialize these here instead of as-needed, as is quick and avoids
5552 * having to test them each time otherwise */
5553 if (! PL_AboveLatin1) {
5554 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5555 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5556 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5558 PL_L1Posix_ptrs[_CC_ALPHANUMERIC]
5559 = _new_invlist_C_array(L1PosixAlnum_invlist);
5560 PL_Posix_ptrs[_CC_ALPHANUMERIC]
5561 = _new_invlist_C_array(PosixAlnum_invlist);
5563 PL_L1Posix_ptrs[_CC_ALPHA]
5564 = _new_invlist_C_array(L1PosixAlpha_invlist);
5565 PL_Posix_ptrs[_CC_ALPHA] = _new_invlist_C_array(PosixAlpha_invlist);
5567 PL_Posix_ptrs[_CC_BLANK] = _new_invlist_C_array(PosixBlank_invlist);
5568 PL_XPosix_ptrs[_CC_BLANK] = _new_invlist_C_array(XPosixBlank_invlist);
5570 /* Cased is the same as Alpha in the ASCII range */
5571 PL_L1Posix_ptrs[_CC_CASED] = _new_invlist_C_array(L1Cased_invlist);
5572 PL_Posix_ptrs[_CC_CASED] = _new_invlist_C_array(PosixAlpha_invlist);
5574 PL_Posix_ptrs[_CC_CNTRL] = _new_invlist_C_array(PosixCntrl_invlist);
5575 PL_XPosix_ptrs[_CC_CNTRL] = _new_invlist_C_array(XPosixCntrl_invlist);
5577 PL_Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5578 PL_L1Posix_ptrs[_CC_DIGIT] = _new_invlist_C_array(PosixDigit_invlist);
5580 PL_L1Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(L1PosixGraph_invlist);
5581 PL_Posix_ptrs[_CC_GRAPH] = _new_invlist_C_array(PosixGraph_invlist);
5583 PL_L1Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(L1PosixLower_invlist);
5584 PL_Posix_ptrs[_CC_LOWER] = _new_invlist_C_array(PosixLower_invlist);
5586 PL_L1Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(L1PosixPrint_invlist);
5587 PL_Posix_ptrs[_CC_PRINT] = _new_invlist_C_array(PosixPrint_invlist);
5589 PL_L1Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(L1PosixPunct_invlist);
5590 PL_Posix_ptrs[_CC_PUNCT] = _new_invlist_C_array(PosixPunct_invlist);
5592 PL_Posix_ptrs[_CC_SPACE] = _new_invlist_C_array(PerlSpace_invlist);
5593 PL_XPosix_ptrs[_CC_SPACE] = _new_invlist_C_array(XPerlSpace_invlist);
5594 PL_Posix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(PosixSpace_invlist);
5595 PL_XPosix_ptrs[_CC_PSXSPC] = _new_invlist_C_array(XPosixSpace_invlist);
5597 PL_L1Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(L1PosixUpper_invlist);
5598 PL_Posix_ptrs[_CC_UPPER] = _new_invlist_C_array(PosixUpper_invlist);
5600 PL_XPosix_ptrs[_CC_VERTSPACE] = _new_invlist_C_array(VertSpace_invlist);
5602 PL_Posix_ptrs[_CC_WORDCHAR] = _new_invlist_C_array(PosixWord_invlist);
5603 PL_L1Posix_ptrs[_CC_WORDCHAR]
5604 = _new_invlist_C_array(L1PosixWord_invlist);
5606 PL_Posix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(PosixXDigit_invlist);
5607 PL_XPosix_ptrs[_CC_XDIGIT] = _new_invlist_C_array(XPosixXDigit_invlist);
5609 PL_HasMultiCharFold = _new_invlist_C_array(_Perl_Multi_Char_Folds_invlist);
5613 pRExC_state->code_blocks = NULL;
5614 pRExC_state->num_code_blocks = 0;
5617 *is_bare_re = FALSE;
5619 if (expr && (expr->op_type == OP_LIST ||
5620 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5621 /* allocate code_blocks if needed */
5625 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling)
5626 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5627 ncode++; /* count of DO blocks */
5629 pRExC_state->num_code_blocks = ncode;
5630 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5635 /* compile-time pattern with just OP_CONSTs and DO blocks */
5640 /* find how many CONSTs there are */
5643 if (expr->op_type == OP_CONST)
5646 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5647 if (o->op_type == OP_CONST)
5651 /* fake up an SV array */
5653 assert(!new_patternp);
5654 Newx(new_patternp, n, SV*);
5655 SAVEFREEPV(new_patternp);
5659 if (expr->op_type == OP_CONST)
5660 new_patternp[n] = cSVOPx_sv(expr);
5662 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5663 if (o->op_type == OP_CONST)
5664 new_patternp[n++] = cSVOPo_sv;
5669 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5670 "Assembling pattern from %d elements%s\n", pat_count,
5671 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5673 /* set expr to the first arg op */
5675 if (pRExC_state->num_code_blocks
5676 && expr->op_type != OP_CONST)
5678 expr = cLISTOPx(expr)->op_first;
5679 assert( expr->op_type == OP_PUSHMARK
5680 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
5681 || expr->op_type == OP_PADRANGE);
5682 expr = expr->op_sibling;
5685 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
5686 expr, &recompile, NULL);
5688 /* handle bare (possibly after overloading) regex: foo =~ $re */
5693 if (SvTYPE(re) == SVt_REGEXP) {
5697 Safefree(pRExC_state->code_blocks);
5698 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5699 "Precompiled pattern%s\n",
5700 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
5706 exp = SvPV_nomg(pat, plen);
5708 if (!eng->op_comp) {
5709 if ((SvUTF8(pat) && IN_BYTES)
5710 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5712 /* make a temporary copy; either to convert to bytes,
5713 * or to avoid repeating get-magic / overloaded stringify */
5714 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5715 (IN_BYTES ? 0 : SvUTF8(pat)));
5717 Safefree(pRExC_state->code_blocks);
5718 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5721 /* ignore the utf8ness if the pattern is 0 length */
5722 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5723 RExC_uni_semantics = 0;
5724 RExC_contains_locale = 0;
5725 pRExC_state->runtime_code_qr = NULL;
5728 SV *dsv= sv_newmortal();
5729 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
5730 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5731 PL_colors[4],PL_colors[5],s);
5735 /* we jump here if we upgrade the pattern to utf8 and have to
5738 if ((pm_flags & PMf_USE_RE_EVAL)
5739 /* this second condition covers the non-regex literal case,
5740 * i.e. $foo =~ '(?{})'. */
5741 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
5743 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
5745 /* return old regex if pattern hasn't changed */
5746 /* XXX: note in the below we have to check the flags as well as the pattern.
5748 * Things get a touch tricky as we have to compare the utf8 flag independently
5749 * from the compile flags.
5754 && !!RX_UTF8(old_re) == !!RExC_utf8
5755 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
5756 && RX_PRECOMP(old_re)
5757 && RX_PRELEN(old_re) == plen
5758 && memEQ(RX_PRECOMP(old_re), exp, plen)
5759 && !runtime_code /* with runtime code, always recompile */ )
5761 Safefree(pRExC_state->code_blocks);
5765 rx_flags = orig_rx_flags;
5767 if (initial_charset == REGEX_LOCALE_CHARSET) {
5768 RExC_contains_locale = 1;
5770 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5772 /* Set to use unicode semantics if the pattern is in utf8 and has the
5773 * 'depends' charset specified, as it means unicode when utf8 */
5774 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5778 RExC_flags = rx_flags;
5779 RExC_pm_flags = pm_flags;
5782 if (TAINTING_get && TAINT_get)
5783 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5785 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5786 /* whoops, we have a non-utf8 pattern, whilst run-time code
5787 * got compiled as utf8. Try again with a utf8 pattern */
5788 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5789 pRExC_state->num_code_blocks);
5790 goto redo_first_pass;
5793 assert(!pRExC_state->runtime_code_qr);
5798 RExC_in_lookbehind = 0;
5799 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5801 RExC_override_recoding = 0;
5802 RExC_in_multi_char_class = 0;
5804 /* First pass: determine size, legality. */
5807 RExC_end = exp + plen;
5812 RExC_emit = &RExC_emit_dummy;
5813 RExC_whilem_seen = 0;
5814 RExC_open_parens = NULL;
5815 RExC_close_parens = NULL;
5817 RExC_paren_names = NULL;
5819 RExC_paren_name_list = NULL;
5821 RExC_recurse = NULL;
5822 RExC_recurse_count = 0;
5823 pRExC_state->code_index = 0;
5825 #if 0 /* REGC() is (currently) a NOP at the first pass.
5826 * Clever compilers notice this and complain. --jhi */
5827 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5830 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5832 RExC_lastparse=NULL;
5834 /* reg may croak on us, not giving us a chance to free
5835 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
5836 need it to survive as long as the regexp (qr/(?{})/).
5837 We must check that code_blocksv is not already set, because we may
5838 have jumped back to restart the sizing pass. */
5839 if (pRExC_state->code_blocks && !code_blocksv) {
5840 code_blocksv = newSV_type(SVt_PV);
5841 SAVEFREESV(code_blocksv);
5842 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
5843 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
5845 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5846 /* It's possible to write a regexp in ascii that represents Unicode
5847 codepoints outside of the byte range, such as via \x{100}. If we
5848 detect such a sequence we have to convert the entire pattern to utf8
5849 and then recompile, as our sizing calculation will have been based
5850 on 1 byte == 1 character, but we will need to use utf8 to encode
5851 at least some part of the pattern, and therefore must convert the whole
5854 if (flags & RESTART_UTF8) {
5855 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
5856 pRExC_state->num_code_blocks);
5857 goto redo_first_pass;
5859 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
5862 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
5865 PerlIO_printf(Perl_debug_log,
5866 "Required size %"IVdf" nodes\n"
5867 "Starting second pass (creation)\n",
5870 RExC_lastparse=NULL;
5873 /* The first pass could have found things that force Unicode semantics */
5874 if ((RExC_utf8 || RExC_uni_semantics)
5875 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5877 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5880 /* Small enough for pointer-storage convention?
5881 If extralen==0, this means that we will not need long jumps. */
5882 if (RExC_size >= 0x10000L && RExC_extralen)
5883 RExC_size += RExC_extralen;
5886 if (RExC_whilem_seen > 15)
5887 RExC_whilem_seen = 15;
5889 /* Allocate space and zero-initialize. Note, the two step process
5890 of zeroing when in debug mode, thus anything assigned has to
5891 happen after that */
5892 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5894 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5895 char, regexp_internal);
5896 if ( r == NULL || ri == NULL )
5897 FAIL("Regexp out of space");
5899 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5900 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5902 /* bulk initialize base fields with 0. */
5903 Zero(ri, sizeof(regexp_internal), char);
5906 /* non-zero initialization begins here */
5909 r->extflags = rx_flags;
5910 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
5912 if (pm_flags & PMf_IS_QR) {
5913 ri->code_blocks = pRExC_state->code_blocks;
5914 ri->num_code_blocks = pRExC_state->num_code_blocks;
5919 for (n = 0; n < pRExC_state->num_code_blocks; n++)
5920 if (pRExC_state->code_blocks[n].src_regex)
5921 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
5922 SAVEFREEPV(pRExC_state->code_blocks);
5926 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5927 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5929 /* The caret is output if there are any defaults: if not all the STD
5930 * flags are set, or if no character set specifier is needed */
5932 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5934 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5935 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5936 >> RXf_PMf_STD_PMMOD_SHIFT);
5937 const char *fptr = STD_PAT_MODS; /*"msix"*/
5939 /* Allocate for the worst case, which is all the std flags are turned
5940 * on. If more precision is desired, we could do a population count of
5941 * the flags set. This could be done with a small lookup table, or by
5942 * shifting, masking and adding, or even, when available, assembly
5943 * language for a machine-language population count.
5944 * We never output a minus, as all those are defaults, so are
5945 * covered by the caret */
5946 const STRLEN wraplen = plen + has_p + has_runon
5947 + has_default /* If needs a caret */
5949 /* If needs a character set specifier */
5950 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5951 + (sizeof(STD_PAT_MODS) - 1)
5952 + (sizeof("(?:)") - 1);
5954 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
5955 r->xpv_len_u.xpvlenu_pv = p;
5957 SvFLAGS(rx) |= SVf_UTF8;
5960 /* If a default, cover it using the caret */
5962 *p++= DEFAULT_PAT_MOD;
5966 const char* const name = get_regex_charset_name(r->extflags, &len);
5967 Copy(name, p, len, char);
5971 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5974 while((ch = *fptr++)) {
5982 Copy(RExC_precomp, p, plen, char);
5983 assert ((RX_WRAPPED(rx) - p) < 16);
5984 r->pre_prefix = p - RX_WRAPPED(rx);
5990 SvCUR_set(rx, p - RX_WRAPPED(rx));
5994 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5996 if (RExC_seen & REG_SEEN_RECURSE) {
5997 Newxz(RExC_open_parens, RExC_npar,regnode *);
5998 SAVEFREEPV(RExC_open_parens);
5999 Newxz(RExC_close_parens,RExC_npar,regnode *);
6000 SAVEFREEPV(RExC_close_parens);
6003 /* Useful during FAIL. */
6004 #ifdef RE_TRACK_PATTERN_OFFSETS
6005 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6006 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6007 "%s %"UVuf" bytes for offset annotations.\n",
6008 ri->u.offsets ? "Got" : "Couldn't get",
6009 (UV)((2*RExC_size+1) * sizeof(U32))));
6011 SetProgLen(ri,RExC_size);
6016 /* Second pass: emit code. */
6017 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6018 RExC_pm_flags = pm_flags;
6020 RExC_end = exp + plen;
6023 RExC_emit_start = ri->program;
6024 RExC_emit = ri->program;
6025 RExC_emit_bound = ri->program + RExC_size + 1;
6026 pRExC_state->code_index = 0;
6028 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
6029 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6031 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6033 /* XXXX To minimize changes to RE engine we always allocate
6034 3-units-long substrs field. */
6035 Newx(r->substrs, 1, struct reg_substr_data);
6036 if (RExC_recurse_count) {
6037 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6038 SAVEFREEPV(RExC_recurse);
6042 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6043 Zero(r->substrs, 1, struct reg_substr_data);
6045 #ifdef TRIE_STUDY_OPT
6047 StructCopy(&zero_scan_data, &data, scan_data_t);
6048 copyRExC_state = RExC_state;
6051 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6053 RExC_state = copyRExC_state;
6054 if (seen & REG_TOP_LEVEL_BRANCHES)
6055 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6057 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6058 StructCopy(&zero_scan_data, &data, scan_data_t);
6061 StructCopy(&zero_scan_data, &data, scan_data_t);
6064 /* Dig out information for optimizations. */
6065 r->extflags = RExC_flags; /* was pm_op */
6066 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6069 SvUTF8_on(rx); /* Unicode in it? */
6070 ri->regstclass = NULL;
6071 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6072 r->intflags |= PREGf_NAUGHTY;
6073 scan = ri->program + 1; /* First BRANCH. */
6075 /* testing for BRANCH here tells us whether there is "must appear"
6076 data in the pattern. If there is then we can use it for optimisations */
6077 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6079 STRLEN longest_float_length, longest_fixed_length;
6080 struct regnode_charclass_class ch_class; /* pointed to by data */
6082 I32 last_close = 0; /* pointed to by data */
6083 regnode *first= scan;
6084 regnode *first_next= regnext(first);
6086 * Skip introductions and multiplicators >= 1
6087 * so that we can extract the 'meat' of the pattern that must
6088 * match in the large if() sequence following.
6089 * NOTE that EXACT is NOT covered here, as it is normally
6090 * picked up by the optimiser separately.
6092 * This is unfortunate as the optimiser isnt handling lookahead
6093 * properly currently.
6096 while ((OP(first) == OPEN && (sawopen = 1)) ||
6097 /* An OR of *one* alternative - should not happen now. */
6098 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6099 /* for now we can't handle lookbehind IFMATCH*/
6100 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6101 (OP(first) == PLUS) ||
6102 (OP(first) == MINMOD) ||
6103 /* An {n,m} with n>0 */
6104 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6105 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6108 * the only op that could be a regnode is PLUS, all the rest
6109 * will be regnode_1 or regnode_2.
6111 * (yves doesn't think this is true)
6113 if (OP(first) == PLUS)
6116 if (OP(first) == MINMOD)
6118 first += regarglen[OP(first)];
6120 first = NEXTOPER(first);
6121 first_next= regnext(first);
6124 /* Starting-point info. */
6126 DEBUG_PEEP("first:",first,0);
6127 /* Ignore EXACT as we deal with it later. */
6128 if (PL_regkind[OP(first)] == EXACT) {
6129 if (OP(first) == EXACT)
6130 NOOP; /* Empty, get anchored substr later. */
6132 ri->regstclass = first;
6135 else if (PL_regkind[OP(first)] == TRIE &&
6136 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6139 /* this can happen only on restudy */
6140 if ( OP(first) == TRIE ) {
6141 struct regnode_1 *trieop = (struct regnode_1 *)
6142 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6143 StructCopy(first,trieop,struct regnode_1);
6144 trie_op=(regnode *)trieop;
6146 struct regnode_charclass *trieop = (struct regnode_charclass *)
6147 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6148 StructCopy(first,trieop,struct regnode_charclass);
6149 trie_op=(regnode *)trieop;
6152 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6153 ri->regstclass = trie_op;
6156 else if (REGNODE_SIMPLE(OP(first)))
6157 ri->regstclass = first;
6158 else if (PL_regkind[OP(first)] == BOUND ||
6159 PL_regkind[OP(first)] == NBOUND)
6160 ri->regstclass = first;
6161 else if (PL_regkind[OP(first)] == BOL) {
6162 r->extflags |= (OP(first) == MBOL
6164 : (OP(first) == SBOL
6167 first = NEXTOPER(first);
6170 else if (OP(first) == GPOS) {
6171 r->extflags |= RXf_ANCH_GPOS;
6172 first = NEXTOPER(first);
6175 else if ((!sawopen || !RExC_sawback) &&
6176 (OP(first) == STAR &&
6177 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6178 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6180 /* turn .* into ^.* with an implied $*=1 */
6182 (OP(NEXTOPER(first)) == REG_ANY)
6185 r->extflags |= type;
6186 r->intflags |= PREGf_IMPLICIT;
6187 first = NEXTOPER(first);
6190 if (sawplus && !sawminmod && !sawlookahead && (!sawopen || !RExC_sawback)
6191 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6192 /* x+ must match at the 1st pos of run of x's */
6193 r->intflags |= PREGf_SKIP;
6195 /* Scan is after the zeroth branch, first is atomic matcher. */
6196 #ifdef TRIE_STUDY_OPT
6199 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6200 (IV)(first - scan + 1))
6204 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6205 (IV)(first - scan + 1))
6211 * If there's something expensive in the r.e., find the
6212 * longest literal string that must appear and make it the
6213 * regmust. Resolve ties in favor of later strings, since
6214 * the regstart check works with the beginning of the r.e.
6215 * and avoiding duplication strengthens checking. Not a
6216 * strong reason, but sufficient in the absence of others.
6217 * [Now we resolve ties in favor of the earlier string if
6218 * it happens that c_offset_min has been invalidated, since the
6219 * earlier string may buy us something the later one won't.]
6222 data.longest_fixed = newSVpvs("");
6223 data.longest_float = newSVpvs("");
6224 data.last_found = newSVpvs("");
6225 data.longest = &(data.longest_fixed);
6226 ENTER_with_name("study_chunk");
6227 SAVEFREESV(data.longest_fixed);
6228 SAVEFREESV(data.longest_float);
6229 SAVEFREESV(data.last_found);
6231 if (!ri->regstclass) {
6232 cl_init(pRExC_state, &ch_class);
6233 data.start_class = &ch_class;
6234 stclass_flag = SCF_DO_STCLASS_AND;
6235 } else /* XXXX Check for BOUND? */
6237 data.last_closep = &last_close;
6239 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6240 &data, -1, NULL, NULL,
6241 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
6242 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6246 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
6249 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6250 && data.last_start_min == 0 && data.last_end > 0
6251 && !RExC_seen_zerolen
6252 && !(RExC_seen & REG_SEEN_VERBARG)
6253 && !((RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6254 r->extflags |= RXf_CHECK_ALL;
6255 scan_commit(pRExC_state, &data,&minlen,0);
6257 longest_float_length = CHR_SVLEN(data.longest_float);
6259 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6260 && data.offset_fixed == data.offset_float_min
6261 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6262 && S_setup_longest (aTHX_ pRExC_state,
6266 &(r->float_end_shift),
6267 data.lookbehind_float,
6268 data.offset_float_min,
6270 longest_float_length,
6271 cBOOL(data.flags & SF_FL_BEFORE_EOL),
6272 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
6274 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6275 r->float_max_offset = data.offset_float_max;
6276 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6277 r->float_max_offset -= data.lookbehind_float;
6278 SvREFCNT_inc_simple_void_NN(data.longest_float);
6281 r->float_substr = r->float_utf8 = NULL;
6282 longest_float_length = 0;
6285 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6287 if (S_setup_longest (aTHX_ pRExC_state,
6289 &(r->anchored_utf8),
6290 &(r->anchored_substr),
6291 &(r->anchored_end_shift),
6292 data.lookbehind_fixed,
6295 longest_fixed_length,
6296 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
6297 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
6299 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6300 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
6303 r->anchored_substr = r->anchored_utf8 = NULL;
6304 longest_fixed_length = 0;
6306 LEAVE_with_name("study_chunk");
6309 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6310 ri->regstclass = NULL;
6312 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6314 && ! TEST_SSC_EOS(data.start_class)
6315 && !cl_is_anything(data.start_class))
6317 const U32 n = add_data(pRExC_state, 1, "f");
6318 OP(data.start_class) = ANYOF_SYNTHETIC;
6320 Newx(RExC_rxi->data->data[n], 1,
6321 struct regnode_charclass_class);
6322 StructCopy(data.start_class,
6323 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6324 struct regnode_charclass_class);
6325 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6326 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6327 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6328 regprop(r, sv, (regnode*)data.start_class);
6329 PerlIO_printf(Perl_debug_log,
6330 "synthetic stclass \"%s\".\n",
6331 SvPVX_const(sv));});
6334 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6335 if (longest_fixed_length > longest_float_length) {
6336 r->check_end_shift = r->anchored_end_shift;
6337 r->check_substr = r->anchored_substr;
6338 r->check_utf8 = r->anchored_utf8;
6339 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6340 if (r->extflags & RXf_ANCH_SINGLE)
6341 r->extflags |= RXf_NOSCAN;
6344 r->check_end_shift = r->float_end_shift;
6345 r->check_substr = r->float_substr;
6346 r->check_utf8 = r->float_utf8;
6347 r->check_offset_min = r->float_min_offset;
6348 r->check_offset_max = r->float_max_offset;
6350 if ((r->check_substr || r->check_utf8) ) {
6351 r->extflags |= RXf_USE_INTUIT;
6352 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6353 r->extflags |= RXf_INTUIT_TAIL;
6355 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6356 if ( (STRLEN)minlen < longest_float_length )
6357 minlen= longest_float_length;
6358 if ( (STRLEN)minlen < longest_fixed_length )
6359 minlen= longest_fixed_length;
6363 /* Several toplevels. Best we can is to set minlen. */
6365 struct regnode_charclass_class ch_class;
6368 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6370 scan = ri->program + 1;
6371 cl_init(pRExC_state, &ch_class);
6372 data.start_class = &ch_class;
6373 data.last_closep = &last_close;
6376 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6377 &data, -1, NULL, NULL,
6378 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS
6379 |(restudied ? SCF_TRIE_DOING_RESTUDY : 0),
6382 CHECK_RESTUDY_GOTO_butfirst(NOOP);
6384 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6385 = r->float_substr = r->float_utf8 = NULL;
6387 if (! TEST_SSC_EOS(data.start_class)
6388 && !cl_is_anything(data.start_class))
6390 const U32 n = add_data(pRExC_state, 1, "f");
6391 OP(data.start_class) = ANYOF_SYNTHETIC;
6393 Newx(RExC_rxi->data->data[n], 1,
6394 struct regnode_charclass_class);
6395 StructCopy(data.start_class,
6396 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6397 struct regnode_charclass_class);
6398 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6399 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6400 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6401 regprop(r, sv, (regnode*)data.start_class);
6402 PerlIO_printf(Perl_debug_log,
6403 "synthetic stclass \"%s\".\n",
6404 SvPVX_const(sv));});
6408 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6409 the "real" pattern. */
6411 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6412 (IV)minlen, (IV)r->minlen);
6414 r->minlenret = minlen;
6415 if (r->minlen < minlen)
6418 if (RExC_seen & REG_SEEN_GPOS)
6419 r->extflags |= RXf_GPOS_SEEN;
6420 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6421 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the lookbehind */
6422 if (pRExC_state->num_code_blocks)
6423 r->extflags |= RXf_EVAL_SEEN;
6424 if (RExC_seen & REG_SEEN_CANY)
6425 r->extflags |= RXf_CANY_SEEN;
6426 if (RExC_seen & REG_SEEN_VERBARG)
6428 r->intflags |= PREGf_VERBARG_SEEN;
6429 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
6431 if (RExC_seen & REG_SEEN_CUTGROUP)
6432 r->intflags |= PREGf_CUTGROUP_SEEN;
6433 if (pm_flags & PMf_USE_RE_EVAL)
6434 r->intflags |= PREGf_USE_RE_EVAL;
6435 if (RExC_paren_names)
6436 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6438 RXp_PAREN_NAMES(r) = NULL;
6441 regnode *first = ri->program + 1;
6443 regnode *next = NEXTOPER(first);
6446 if (PL_regkind[fop] == NOTHING && nop == END)
6447 r->extflags |= RXf_NULL;
6448 else if (PL_regkind[fop] == BOL && nop == END)
6449 r->extflags |= RXf_START_ONLY;
6450 else if (fop == PLUS && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE && OP(regnext(first)) == END)
6451 r->extflags |= RXf_WHITE;
6452 else if ( r->extflags & RXf_SPLIT && fop == EXACT && STR_LEN(first) == 1 && *(STRING(first)) == ' ' && OP(regnext(first)) == END )
6453 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6457 if (RExC_paren_names) {
6458 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6459 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6462 ri->name_list_idx = 0;
6464 if (RExC_recurse_count) {
6465 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6466 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6467 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6470 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6471 /* assume we don't need to swap parens around before we match */
6474 PerlIO_printf(Perl_debug_log,"Final program:\n");
6477 #ifdef RE_TRACK_PATTERN_OFFSETS
6478 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6479 const U32 len = ri->u.offsets[0];
6481 GET_RE_DEBUG_FLAGS_DECL;
6482 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6483 for (i = 1; i <= len; i++) {
6484 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6485 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6486 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6488 PerlIO_printf(Perl_debug_log, "\n");
6493 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
6494 * by setting the regexp SV to readonly-only instead. If the
6495 * pattern's been recompiled, the USEDness should remain. */
6496 if (old_re && SvREADONLY(old_re))
6504 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6507 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6509 PERL_UNUSED_ARG(value);
6511 if (flags & RXapif_FETCH) {
6512 return reg_named_buff_fetch(rx, key, flags);
6513 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6514 Perl_croak_no_modify();
6516 } else if (flags & RXapif_EXISTS) {
6517 return reg_named_buff_exists(rx, key, flags)
6520 } else if (flags & RXapif_REGNAMES) {
6521 return reg_named_buff_all(rx, flags);
6522 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6523 return reg_named_buff_scalar(rx, flags);
6525 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6531 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6534 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6535 PERL_UNUSED_ARG(lastkey);
6537 if (flags & RXapif_FIRSTKEY)
6538 return reg_named_buff_firstkey(rx, flags);
6539 else if (flags & RXapif_NEXTKEY)
6540 return reg_named_buff_nextkey(rx, flags);
6542 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6548 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6551 AV *retarray = NULL;
6553 struct regexp *const rx = ReANY(r);
6555 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6557 if (flags & RXapif_ALL)
6560 if (rx && RXp_PAREN_NAMES(rx)) {
6561 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6564 SV* sv_dat=HeVAL(he_str);
6565 I32 *nums=(I32*)SvPVX(sv_dat);
6566 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6567 if ((I32)(rx->nparens) >= nums[i]
6568 && rx->offs[nums[i]].start != -1
6569 && rx->offs[nums[i]].end != -1)
6572 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6577 ret = newSVsv(&PL_sv_undef);
6580 av_push(retarray, ret);
6583 return newRV_noinc(MUTABLE_SV(retarray));
6590 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6593 struct regexp *const rx = ReANY(r);
6595 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6597 if (rx && RXp_PAREN_NAMES(rx)) {
6598 if (flags & RXapif_ALL) {
6599 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6601 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6603 SvREFCNT_dec_NN(sv);
6615 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6617 struct regexp *const rx = ReANY(r);
6619 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6621 if ( rx && RXp_PAREN_NAMES(rx) ) {
6622 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6624 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6631 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6633 struct regexp *const rx = ReANY(r);
6634 GET_RE_DEBUG_FLAGS_DECL;
6636 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6638 if (rx && RXp_PAREN_NAMES(rx)) {
6639 HV *hv = RXp_PAREN_NAMES(rx);
6641 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6644 SV* sv_dat = HeVAL(temphe);
6645 I32 *nums = (I32*)SvPVX(sv_dat);
6646 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6647 if ((I32)(rx->lastparen) >= nums[i] &&
6648 rx->offs[nums[i]].start != -1 &&
6649 rx->offs[nums[i]].end != -1)
6655 if (parno || flags & RXapif_ALL) {
6656 return newSVhek(HeKEY_hek(temphe));
6664 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6669 struct regexp *const rx = ReANY(r);
6671 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6673 if (rx && RXp_PAREN_NAMES(rx)) {
6674 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6675 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6676 } else if (flags & RXapif_ONE) {
6677 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6678 av = MUTABLE_AV(SvRV(ret));
6679 length = av_len(av);
6680 SvREFCNT_dec_NN(ret);
6681 return newSViv(length + 1);
6683 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6687 return &PL_sv_undef;
6691 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6693 struct regexp *const rx = ReANY(r);
6696 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6698 if (rx && RXp_PAREN_NAMES(rx)) {
6699 HV *hv= RXp_PAREN_NAMES(rx);
6701 (void)hv_iterinit(hv);
6702 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6705 SV* sv_dat = HeVAL(temphe);
6706 I32 *nums = (I32*)SvPVX(sv_dat);
6707 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6708 if ((I32)(rx->lastparen) >= nums[i] &&
6709 rx->offs[nums[i]].start != -1 &&
6710 rx->offs[nums[i]].end != -1)
6716 if (parno || flags & RXapif_ALL) {
6717 av_push(av, newSVhek(HeKEY_hek(temphe)));
6722 return newRV_noinc(MUTABLE_SV(av));
6726 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6729 struct regexp *const rx = ReANY(r);
6735 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6737 if ( n == RX_BUFF_IDX_CARET_PREMATCH
6738 || n == RX_BUFF_IDX_CARET_FULLMATCH
6739 || n == RX_BUFF_IDX_CARET_POSTMATCH
6742 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6744 /* on something like
6747 * the KEEPCOPY is set on the PMOP rather than the regex */
6748 if (PL_curpm && r == PM_GETRE(PL_curpm))
6749 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6758 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6759 /* no need to distinguish between them any more */
6760 n = RX_BUFF_IDX_FULLMATCH;
6762 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6763 && rx->offs[0].start != -1)
6765 /* $`, ${^PREMATCH} */
6766 i = rx->offs[0].start;
6770 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6771 && rx->offs[0].end != -1)
6773 /* $', ${^POSTMATCH} */
6774 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6775 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6778 if ( 0 <= n && n <= (I32)rx->nparens &&
6779 (s1 = rx->offs[n].start) != -1 &&
6780 (t1 = rx->offs[n].end) != -1)
6782 /* $&, ${^MATCH}, $1 ... */
6784 s = rx->subbeg + s1 - rx->suboffset;
6789 assert(s >= rx->subbeg);
6790 assert(rx->sublen >= (s - rx->subbeg) + i );
6792 #if NO_TAINT_SUPPORT
6793 sv_setpvn(sv, s, i);
6795 const int oldtainted = TAINT_get;
6797 sv_setpvn(sv, s, i);
6798 TAINT_set(oldtainted);
6800 if ( (rx->extflags & RXf_CANY_SEEN)
6801 ? (RXp_MATCH_UTF8(rx)
6802 && (!i || is_utf8_string((U8*)s, i)))
6803 : (RXp_MATCH_UTF8(rx)) )
6810 if (RXp_MATCH_TAINTED(rx)) {
6811 if (SvTYPE(sv) >= SVt_PVMG) {
6812 MAGIC* const mg = SvMAGIC(sv);
6815 SvMAGIC_set(sv, mg->mg_moremagic);
6817 if ((mgt = SvMAGIC(sv))) {
6818 mg->mg_moremagic = mgt;
6819 SvMAGIC_set(sv, mg);
6830 sv_setsv(sv,&PL_sv_undef);
6836 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6837 SV const * const value)
6839 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6841 PERL_UNUSED_ARG(rx);
6842 PERL_UNUSED_ARG(paren);
6843 PERL_UNUSED_ARG(value);
6846 Perl_croak_no_modify();
6850 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6853 struct regexp *const rx = ReANY(r);
6857 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6859 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
6860 || paren == RX_BUFF_IDX_CARET_FULLMATCH
6861 || paren == RX_BUFF_IDX_CARET_POSTMATCH
6864 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
6866 /* on something like
6869 * the KEEPCOPY is set on the PMOP rather than the regex */
6870 if (PL_curpm && r == PM_GETRE(PL_curpm))
6871 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
6877 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6879 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6880 case RX_BUFF_IDX_PREMATCH: /* $` */
6881 if (rx->offs[0].start != -1) {
6882 i = rx->offs[0].start;
6891 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6892 case RX_BUFF_IDX_POSTMATCH: /* $' */
6893 if (rx->offs[0].end != -1) {
6894 i = rx->sublen - rx->offs[0].end;
6896 s1 = rx->offs[0].end;
6903 default: /* $& / ${^MATCH}, $1, $2, ... */
6904 if (paren <= (I32)rx->nparens &&
6905 (s1 = rx->offs[paren].start) != -1 &&
6906 (t1 = rx->offs[paren].end) != -1)
6912 if (ckWARN(WARN_UNINITIALIZED))
6913 report_uninit((const SV *)sv);
6918 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6919 const char * const s = rx->subbeg - rx->suboffset + s1;
6924 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6931 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6933 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6934 PERL_UNUSED_ARG(rx);
6938 return newSVpvs("Regexp");
6941 /* Scans the name of a named buffer from the pattern.
6942 * If flags is REG_RSN_RETURN_NULL returns null.
6943 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6944 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6945 * to the parsed name as looked up in the RExC_paren_names hash.
6946 * If there is an error throws a vFAIL().. type exception.
6949 #define REG_RSN_RETURN_NULL 0
6950 #define REG_RSN_RETURN_NAME 1
6951 #define REG_RSN_RETURN_DATA 2
6954 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6956 char *name_start = RExC_parse;
6958 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6960 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6961 /* skip IDFIRST by using do...while */
6964 RExC_parse += UTF8SKIP(RExC_parse);
6965 } while (isWORDCHAR_utf8((U8*)RExC_parse));
6969 } while (isWORDCHAR(*RExC_parse));
6971 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6972 vFAIL("Group name must start with a non-digit word character");
6976 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6977 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6978 if ( flags == REG_RSN_RETURN_NAME)
6980 else if (flags==REG_RSN_RETURN_DATA) {
6983 if ( ! sv_name ) /* should not happen*/
6984 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6985 if (RExC_paren_names)
6986 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6988 sv_dat = HeVAL(he_str);
6990 vFAIL("Reference to nonexistent named group");
6994 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6995 (unsigned long) flags);
6997 assert(0); /* NOT REACHED */
7002 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7003 int rem=(int)(RExC_end - RExC_parse); \
7012 if (RExC_lastparse!=RExC_parse) \
7013 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
7016 iscut ? "..." : "<" \
7019 PerlIO_printf(Perl_debug_log,"%16s",""); \
7022 num = RExC_size + 1; \
7024 num=REG_NODE_NUM(RExC_emit); \
7025 if (RExC_lastnum!=num) \
7026 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7028 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7029 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7030 (int)((depth*2)), "", \
7034 RExC_lastparse=RExC_parse; \
7039 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7040 DEBUG_PARSE_MSG((funcname)); \
7041 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7043 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7044 DEBUG_PARSE_MSG((funcname)); \
7045 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7048 /* This section of code defines the inversion list object and its methods. The
7049 * interfaces are highly subject to change, so as much as possible is static to
7050 * this file. An inversion list is here implemented as a malloc'd C UV array
7051 * as an SVt_INVLIST scalar.
7053 * An inversion list for Unicode is an array of code points, sorted by ordinal
7054 * number. The zeroth element is the first code point in the list. The 1th
7055 * element is the first element beyond that not in the list. In other words,
7056 * the first range is
7057 * invlist[0]..(invlist[1]-1)
7058 * The other ranges follow. Thus every element whose index is divisible by two
7059 * marks the beginning of a range that is in the list, and every element not
7060 * divisible by two marks the beginning of a range not in the list. A single
7061 * element inversion list that contains the single code point N generally
7062 * consists of two elements
7065 * (The exception is when N is the highest representable value on the
7066 * machine, in which case the list containing just it would be a single
7067 * element, itself. By extension, if the last range in the list extends to
7068 * infinity, then the first element of that range will be in the inversion list
7069 * at a position that is divisible by two, and is the final element in the
7071 * Taking the complement (inverting) an inversion list is quite simple, if the
7072 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7073 * This implementation reserves an element at the beginning of each inversion
7074 * list to always contain 0; there is an additional flag in the header which
7075 * indicates if the list begins at the 0, or is offset to begin at the next
7078 * More about inversion lists can be found in "Unicode Demystified"
7079 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7080 * More will be coming when functionality is added later.
7082 * The inversion list data structure is currently implemented as an SV pointing
7083 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7084 * array of UV whose memory management is automatically handled by the existing
7085 * facilities for SV's.
7087 * Some of the methods should always be private to the implementation, and some
7088 * should eventually be made public */
7090 /* The header definitions are in F<inline_invlist.c> */
7092 PERL_STATIC_INLINE UV*
7093 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7095 /* Returns a pointer to the first element in the inversion list's array.
7096 * This is called upon initialization of an inversion list. Where the
7097 * array begins depends on whether the list has the code point U+0000 in it
7098 * or not. The other parameter tells it whether the code that follows this
7099 * call is about to put a 0 in the inversion list or not. The first
7100 * element is either the element reserved for 0, if TRUE, or the element
7101 * after it, if FALSE */
7103 bool* offset = get_invlist_offset_addr(invlist);
7104 UV* zero_addr = (UV *) SvPVX(invlist);
7106 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7109 assert(! _invlist_len(invlist));
7113 /* 1^1 = 0; 1^0 = 1 */
7114 *offset = 1 ^ will_have_0;
7115 return zero_addr + *offset;
7118 PERL_STATIC_INLINE UV*
7119 S_invlist_array(pTHX_ SV* const invlist)
7121 /* Returns the pointer to the inversion list's array. Every time the
7122 * length changes, this needs to be called in case malloc or realloc moved
7125 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7127 /* Must not be empty. If these fail, you probably didn't check for <len>
7128 * being non-zero before trying to get the array */
7129 assert(_invlist_len(invlist));
7131 /* The very first element always contains zero, The array begins either
7132 * there, or if the inversion list is offset, at the element after it.
7133 * The offset header field determines which; it contains 0 or 1 to indicate
7134 * how much additionally to add */
7135 assert(0 == *(SvPVX(invlist)));
7136 return ((UV *) SvPVX(invlist) + *get_invlist_offset_addr(invlist));
7139 PERL_STATIC_INLINE void
7140 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
7142 /* Sets the current number of elements stored in the inversion list.
7143 * Updates SvCUR correspondingly */
7145 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7147 assert(SvTYPE(invlist) == SVt_INVLIST);
7152 : TO_INTERNAL_SIZE(len + offset));
7153 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
7156 PERL_STATIC_INLINE IV*
7157 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7159 /* Return the address of the IV that is reserved to hold the cached index
7162 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7164 assert(SvTYPE(invlist) == SVt_INVLIST);
7166 return &(((XINVLIST*) SvANY(invlist))->prev_index);
7169 PERL_STATIC_INLINE IV
7170 S_invlist_previous_index(pTHX_ SV* const invlist)
7172 /* Returns cached index of previous search */
7174 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7176 return *get_invlist_previous_index_addr(invlist);
7179 PERL_STATIC_INLINE void
7180 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7182 /* Caches <index> for later retrieval */
7184 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7186 assert(index == 0 || index < (int) _invlist_len(invlist));
7188 *get_invlist_previous_index_addr(invlist) = index;
7191 PERL_STATIC_INLINE UV
7192 S_invlist_max(pTHX_ SV* const invlist)
7194 /* Returns the maximum number of elements storable in the inversion list's
7195 * array, without having to realloc() */
7197 PERL_ARGS_ASSERT_INVLIST_MAX;
7199 assert(SvTYPE(invlist) == SVt_INVLIST);
7201 /* Assumes worst case, in which the 0 element is not counted in the
7202 * inversion list, so subtracts 1 for that */
7203 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
7204 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
7205 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
7208 #ifndef PERL_IN_XSUB_RE
7210 Perl__new_invlist(pTHX_ IV initial_size)
7213 /* Return a pointer to a newly constructed inversion list, with enough
7214 * space to store 'initial_size' elements. If that number is negative, a
7215 * system default is used instead */
7219 if (initial_size < 0) {
7223 /* Allocate the initial space */
7224 new_list = newSV_type(SVt_INVLIST);
7226 /* First 1 is in case the zero element isn't in the list; second 1 is for
7228 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
7229 invlist_set_len(new_list, 0, 0);
7231 /* Force iterinit() to be used to get iteration to work */
7232 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
7234 *get_invlist_previous_index_addr(new_list) = 0;
7241 S__new_invlist_C_array(pTHX_ const UV* const list)
7243 /* Return a pointer to a newly constructed inversion list, initialized to
7244 * point to <list>, which has to be in the exact correct inversion list
7245 * form, including internal fields. Thus this is a dangerous routine that
7246 * should not be used in the wrong hands. The passed in 'list' contains
7247 * several header fields at the beginning that are not part of the
7248 * inversion list body proper */
7250 const STRLEN length = (STRLEN) list[0];
7251 const UV version_id = list[1];
7252 const bool offset = cBOOL(list[2]);
7253 #define HEADER_LENGTH 3
7254 /* If any of the above changes in any way, you must change HEADER_LENGTH
7255 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
7256 * perl -E 'say int(rand 2**31-1)'
7258 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
7259 data structure type, so that one being
7260 passed in can be validated to be an
7261 inversion list of the correct vintage.
7264 SV* invlist = newSV_type(SVt_INVLIST);
7266 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7268 if (version_id != INVLIST_VERSION_ID) {
7269 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7272 /* The generated array passed in includes header elements that aren't part
7273 * of the list proper, so start it just after them */
7274 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
7276 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7277 shouldn't touch it */
7279 *(get_invlist_offset_addr(invlist)) = offset;
7281 /* The 'length' passed to us is the physical number of elements in the
7282 * inversion list. But if there is an offset the logical number is one
7284 invlist_set_len(invlist, length - offset, offset);
7286 invlist_set_previous_index(invlist, 0);
7288 /* Initialize the iteration pointer. */
7289 invlist_iterfinish(invlist);
7295 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7297 /* Grow the maximum size of an inversion list */
7299 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7301 assert(SvTYPE(invlist) == SVt_INVLIST);
7303 /* Add one to account for the zero element at the beginning which may not
7304 * be counted by the calling parameters */
7305 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
7308 PERL_STATIC_INLINE void
7309 S_invlist_trim(pTHX_ SV* const invlist)
7311 PERL_ARGS_ASSERT_INVLIST_TRIM;
7313 assert(SvTYPE(invlist) == SVt_INVLIST);
7315 /* Change the length of the inversion list to how many entries it currently
7317 SvPV_shrink_to_cur((SV *) invlist);
7320 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7323 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7325 /* Subject to change or removal. Append the range from 'start' to 'end' at
7326 * the end of the inversion list. The range must be above any existing
7330 UV max = invlist_max(invlist);
7331 UV len = _invlist_len(invlist);
7334 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7336 if (len == 0) { /* Empty lists must be initialized */
7337 offset = start != 0;
7338 array = _invlist_array_init(invlist, ! offset);
7341 /* Here, the existing list is non-empty. The current max entry in the
7342 * list is generally the first value not in the set, except when the
7343 * set extends to the end of permissible values, in which case it is
7344 * the first entry in that final set, and so this call is an attempt to
7345 * append out-of-order */
7347 UV final_element = len - 1;
7348 array = invlist_array(invlist);
7349 if (array[final_element] > start
7350 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7352 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",
7353 array[final_element], start,
7354 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7357 /* Here, it is a legal append. If the new range begins with the first
7358 * value not in the set, it is extending the set, so the new first
7359 * value not in the set is one greater than the newly extended range.
7361 offset = *get_invlist_offset_addr(invlist);
7362 if (array[final_element] == start) {
7363 if (end != UV_MAX) {
7364 array[final_element] = end + 1;
7367 /* But if the end is the maximum representable on the machine,
7368 * just let the range that this would extend to have no end */
7369 invlist_set_len(invlist, len - 1, offset);
7375 /* Here the new range doesn't extend any existing set. Add it */
7377 len += 2; /* Includes an element each for the start and end of range */
7379 /* If wll overflow the existing space, extend, which may cause the array to
7382 invlist_extend(invlist, len);
7384 /* Have to set len here to avoid assert failure in invlist_array() */
7385 invlist_set_len(invlist, len, offset);
7387 array = invlist_array(invlist);
7390 invlist_set_len(invlist, len, offset);
7393 /* The next item on the list starts the range, the one after that is
7394 * one past the new range. */
7395 array[len - 2] = start;
7396 if (end != UV_MAX) {
7397 array[len - 1] = end + 1;
7400 /* But if the end is the maximum representable on the machine, just let
7401 * the range have no end */
7402 invlist_set_len(invlist, len - 1, offset);
7406 #ifndef PERL_IN_XSUB_RE
7409 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7411 /* Searches the inversion list for the entry that contains the input code
7412 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7413 * return value is the index into the list's array of the range that
7418 IV high = _invlist_len(invlist);
7419 const IV highest_element = high - 1;
7422 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7424 /* If list is empty, return failure. */
7429 /* (We can't get the array unless we know the list is non-empty) */
7430 array = invlist_array(invlist);
7432 mid = invlist_previous_index(invlist);
7433 assert(mid >=0 && mid <= highest_element);
7435 /* <mid> contains the cache of the result of the previous call to this
7436 * function (0 the first time). See if this call is for the same result,
7437 * or if it is for mid-1. This is under the theory that calls to this
7438 * function will often be for related code points that are near each other.
7439 * And benchmarks show that caching gives better results. We also test
7440 * here if the code point is within the bounds of the list. These tests
7441 * replace others that would have had to be made anyway to make sure that
7442 * the array bounds were not exceeded, and these give us extra information
7443 * at the same time */
7444 if (cp >= array[mid]) {
7445 if (cp >= array[highest_element]) {
7446 return highest_element;
7449 /* Here, array[mid] <= cp < array[highest_element]. This means that
7450 * the final element is not the answer, so can exclude it; it also
7451 * means that <mid> is not the final element, so can refer to 'mid + 1'
7453 if (cp < array[mid + 1]) {
7459 else { /* cp < aray[mid] */
7460 if (cp < array[0]) { /* Fail if outside the array */
7464 if (cp >= array[mid - 1]) {
7469 /* Binary search. What we are looking for is <i> such that
7470 * array[i] <= cp < array[i+1]
7471 * The loop below converges on the i+1. Note that there may not be an
7472 * (i+1)th element in the array, and things work nonetheless */
7473 while (low < high) {
7474 mid = (low + high) / 2;
7475 assert(mid <= highest_element);
7476 if (array[mid] <= cp) { /* cp >= array[mid] */
7479 /* We could do this extra test to exit the loop early.
7480 if (cp < array[low]) {
7485 else { /* cp < array[mid] */
7492 invlist_set_previous_index(invlist, high);
7497 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7499 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7500 * but is used when the swash has an inversion list. This makes this much
7501 * faster, as it uses a binary search instead of a linear one. This is
7502 * intimately tied to that function, and perhaps should be in utf8.c,
7503 * except it is intimately tied to inversion lists as well. It assumes
7504 * that <swatch> is all 0's on input */
7507 const IV len = _invlist_len(invlist);
7511 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7513 if (len == 0) { /* Empty inversion list */
7517 array = invlist_array(invlist);
7519 /* Find which element it is */
7520 i = _invlist_search(invlist, start);
7522 /* We populate from <start> to <end> */
7523 while (current < end) {
7526 /* The inversion list gives the results for every possible code point
7527 * after the first one in the list. Only those ranges whose index is
7528 * even are ones that the inversion list matches. For the odd ones,
7529 * and if the initial code point is not in the list, we have to skip
7530 * forward to the next element */
7531 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7533 if (i >= len) { /* Finished if beyond the end of the array */
7537 if (current >= end) { /* Finished if beyond the end of what we
7539 if (LIKELY(end < UV_MAX)) {
7543 /* We get here when the upper bound is the maximum
7544 * representable on the machine, and we are looking for just
7545 * that code point. Have to special case it */
7547 goto join_end_of_list;
7550 assert(current >= start);
7552 /* The current range ends one below the next one, except don't go past
7555 upper = (i < len && array[i] < end) ? array[i] : end;
7557 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7558 * for each code point in it */
7559 for (; current < upper; current++) {
7560 const STRLEN offset = (STRLEN)(current - start);
7561 swatch[offset >> 3] |= 1 << (offset & 7);
7566 /* Quit if at the end of the list */
7569 /* But first, have to deal with the highest possible code point on
7570 * the platform. The previous code assumes that <end> is one
7571 * beyond where we want to populate, but that is impossible at the
7572 * platform's infinity, so have to handle it specially */
7573 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7575 const STRLEN offset = (STRLEN)(end - start);
7576 swatch[offset >> 3] |= 1 << (offset & 7);
7581 /* Advance to the next range, which will be for code points not in the
7590 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** output)
7592 /* Take the union of two inversion lists and point <output> to it. *output
7593 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7594 * the reference count to that list will be decremented. The first list,
7595 * <a>, may be NULL, in which case a copy of the second list is returned.
7596 * If <complement_b> is TRUE, the union is taken of the complement
7597 * (inversion) of <b> instead of b itself.
7599 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7600 * Richard Gillam, published by Addison-Wesley, and explained at some
7601 * length there. The preface says to incorporate its examples into your
7602 * code at your own risk.
7604 * The algorithm is like a merge sort.
7606 * XXX A potential performance improvement is to keep track as we go along
7607 * if only one of the inputs contributes to the result, meaning the other
7608 * is a subset of that one. In that case, we can skip the final copy and
7609 * return the larger of the input lists, but then outside code might need
7610 * to keep track of whether to free the input list or not */
7612 const UV* array_a; /* a's array */
7614 UV len_a; /* length of a's array */
7617 SV* u; /* the resulting union */
7621 UV i_a = 0; /* current index into a's array */
7625 /* running count, as explained in the algorithm source book; items are
7626 * stopped accumulating and are output when the count changes to/from 0.
7627 * The count is incremented when we start a range that's in the set, and
7628 * decremented when we start a range that's not in the set. So its range
7629 * is 0 to 2. Only when the count is zero is something not in the set.
7633 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7636 /* If either one is empty, the union is the other one */
7637 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7644 *output = invlist_clone(b);
7646 _invlist_invert(*output);
7648 } /* else *output already = b; */
7651 else if ((len_b = _invlist_len(b)) == 0) {
7656 /* The complement of an empty list is a list that has everything in it,
7657 * so the union with <a> includes everything too */
7662 *output = _new_invlist(1);
7663 _append_range_to_invlist(*output, 0, UV_MAX);
7665 else if (*output != a) {
7666 *output = invlist_clone(a);
7668 /* else *output already = a; */
7672 /* Here both lists exist and are non-empty */
7673 array_a = invlist_array(a);
7674 array_b = invlist_array(b);
7676 /* If are to take the union of 'a' with the complement of b, set it
7677 * up so are looking at b's complement. */
7680 /* To complement, we invert: if the first element is 0, remove it. To
7681 * do this, we just pretend the array starts one later */
7682 if (array_b[0] == 0) {
7688 /* But if the first element is not zero, we pretend the list starts
7689 * at the 0 that is always stored immediately before the array. */
7695 /* Size the union for the worst case: that the sets are completely
7697 u = _new_invlist(len_a + len_b);
7699 /* Will contain U+0000 if either component does */
7700 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7701 || (len_b > 0 && array_b[0] == 0));
7703 /* Go through each list item by item, stopping when exhausted one of
7705 while (i_a < len_a && i_b < len_b) {
7706 UV cp; /* The element to potentially add to the union's array */
7707 bool cp_in_set; /* is it in the the input list's set or not */
7709 /* We need to take one or the other of the two inputs for the union.
7710 * Since we are merging two sorted lists, we take the smaller of the
7711 * next items. In case of a tie, we take the one that is in its set
7712 * first. If we took one not in the set first, it would decrement the
7713 * count, possibly to 0 which would cause it to be output as ending the
7714 * range, and the next time through we would take the same number, and
7715 * output it again as beginning the next range. By doing it the
7716 * opposite way, there is no possibility that the count will be
7717 * momentarily decremented to 0, and thus the two adjoining ranges will
7718 * be seamlessly merged. (In a tie and both are in the set or both not
7719 * in the set, it doesn't matter which we take first.) */
7720 if (array_a[i_a] < array_b[i_b]
7721 || (array_a[i_a] == array_b[i_b]
7722 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7724 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7728 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7729 cp = array_b[i_b++];
7732 /* Here, have chosen which of the two inputs to look at. Only output
7733 * if the running count changes to/from 0, which marks the
7734 * beginning/end of a range in that's in the set */
7737 array_u[i_u++] = cp;
7744 array_u[i_u++] = cp;
7749 /* Here, we are finished going through at least one of the lists, which
7750 * means there is something remaining in at most one. We check if the list
7751 * that hasn't been exhausted is positioned such that we are in the middle
7752 * of a range in its set or not. (i_a and i_b point to the element beyond
7753 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7754 * is potentially more to output.
7755 * There are four cases:
7756 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7757 * in the union is entirely from the non-exhausted set.
7758 * 2) Both were in their sets, count is 2. Nothing further should
7759 * be output, as everything that remains will be in the exhausted
7760 * list's set, hence in the union; decrementing to 1 but not 0 insures
7762 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7763 * Nothing further should be output because the union includes
7764 * everything from the exhausted set. Not decrementing ensures that.
7765 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7766 * decrementing to 0 insures that we look at the remainder of the
7767 * non-exhausted set */
7768 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7769 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7774 /* The final length is what we've output so far, plus what else is about to
7775 * be output. (If 'count' is non-zero, then the input list we exhausted
7776 * has everything remaining up to the machine's limit in its set, and hence
7777 * in the union, so there will be no further output. */
7780 /* At most one of the subexpressions will be non-zero */
7781 len_u += (len_a - i_a) + (len_b - i_b);
7784 /* Set result to final length, which can change the pointer to array_u, so
7786 if (len_u != _invlist_len(u)) {
7787 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
7789 array_u = invlist_array(u);
7792 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7793 * the other) ended with everything above it not in its set. That means
7794 * that the remaining part of the union is precisely the same as the
7795 * non-exhausted list, so can just copy it unchanged. (If both list were
7796 * exhausted at the same time, then the operations below will be both 0.)
7799 IV copy_count; /* At most one will have a non-zero copy count */
7800 if ((copy_count = len_a - i_a) > 0) {
7801 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7803 else if ((copy_count = len_b - i_b) > 0) {
7804 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7808 /* We may be removing a reference to one of the inputs */
7809 if (a == *output || b == *output) {
7810 assert(! invlist_is_iterating(*output));
7811 SvREFCNT_dec_NN(*output);
7819 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, const bool complement_b, SV** i)
7821 /* Take the intersection of two inversion lists and point <i> to it. *i
7822 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
7823 * the reference count to that list will be decremented.
7824 * If <complement_b> is TRUE, the result will be the intersection of <a>
7825 * and the complement (or inversion) of <b> instead of <b> directly.
7827 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7828 * Richard Gillam, published by Addison-Wesley, and explained at some
7829 * length there. The preface says to incorporate its examples into your
7830 * code at your own risk. In fact, it had bugs
7832 * The algorithm is like a merge sort, and is essentially the same as the
7836 const UV* array_a; /* a's array */
7838 UV len_a; /* length of a's array */
7841 SV* r; /* the resulting intersection */
7845 UV i_a = 0; /* current index into a's array */
7849 /* running count, as explained in the algorithm source book; items are
7850 * stopped accumulating and are output when the count changes to/from 2.
7851 * The count is incremented when we start a range that's in the set, and
7852 * decremented when we start a range that's not in the set. So its range
7853 * is 0 to 2. Only when the count is 2 is something in the intersection.
7857 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7860 /* Special case if either one is empty */
7861 len_a = (a == NULL) ? 0 : _invlist_len(a);
7862 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7864 if (len_a != 0 && complement_b) {
7866 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7867 * be empty. Here, also we are using 'b's complement, which hence
7868 * must be every possible code point. Thus the intersection is
7875 *i = invlist_clone(a);
7877 /* else *i is already 'a' */
7881 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7882 * intersection must be empty */
7889 *i = _new_invlist(0);
7893 /* Here both lists exist and are non-empty */
7894 array_a = invlist_array(a);
7895 array_b = invlist_array(b);
7897 /* If are to take the intersection of 'a' with the complement of b, set it
7898 * up so are looking at b's complement. */
7901 /* To complement, we invert: if the first element is 0, remove it. To
7902 * do this, we just pretend the array starts one later */
7903 if (array_b[0] == 0) {
7909 /* But if the first element is not zero, we pretend the list starts
7910 * at the 0 that is always stored immediately before the array. */
7916 /* Size the intersection for the worst case: that the intersection ends up
7917 * fragmenting everything to be completely disjoint */
7918 r= _new_invlist(len_a + len_b);
7920 /* Will contain U+0000 iff both components do */
7921 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7922 && len_b > 0 && array_b[0] == 0);
7924 /* Go through each list item by item, stopping when exhausted one of
7926 while (i_a < len_a && i_b < len_b) {
7927 UV cp; /* The element to potentially add to the intersection's
7929 bool cp_in_set; /* Is it in the input list's set or not */
7931 /* We need to take one or the other of the two inputs for the
7932 * intersection. Since we are merging two sorted lists, we take the
7933 * smaller of the next items. In case of a tie, we take the one that
7934 * is not in its set first (a difference from the union algorithm). If
7935 * we took one in the set first, it would increment the count, possibly
7936 * to 2 which would cause it to be output as starting a range in the
7937 * intersection, and the next time through we would take that same
7938 * number, and output it again as ending the set. By doing it the
7939 * opposite of this, there is no possibility that the count will be
7940 * momentarily incremented to 2. (In a tie and both are in the set or
7941 * both not in the set, it doesn't matter which we take first.) */
7942 if (array_a[i_a] < array_b[i_b]
7943 || (array_a[i_a] == array_b[i_b]
7944 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7946 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7950 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7954 /* Here, have chosen which of the two inputs to look at. Only output
7955 * if the running count changes to/from 2, which marks the
7956 * beginning/end of a range that's in the intersection */
7960 array_r[i_r++] = cp;
7965 array_r[i_r++] = cp;
7971 /* Here, we are finished going through at least one of the lists, which
7972 * means there is something remaining in at most one. We check if the list
7973 * that has been exhausted is positioned such that we are in the middle
7974 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7975 * the ones we care about.) There are four cases:
7976 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7977 * nothing left in the intersection.
7978 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7979 * above 2. What should be output is exactly that which is in the
7980 * non-exhausted set, as everything it has is also in the intersection
7981 * set, and everything it doesn't have can't be in the intersection
7982 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7983 * gets incremented to 2. Like the previous case, the intersection is
7984 * everything that remains in the non-exhausted set.
7985 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7986 * remains 1. And the intersection has nothing more. */
7987 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7988 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7993 /* The final length is what we've output so far plus what else is in the
7994 * intersection. At most one of the subexpressions below will be non-zero */
7997 len_r += (len_a - i_a) + (len_b - i_b);
8000 /* Set result to final length, which can change the pointer to array_r, so
8002 if (len_r != _invlist_len(r)) {
8003 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8005 array_r = invlist_array(r);
8008 /* Finish outputting any remaining */
8009 if (count >= 2) { /* At most one will have a non-zero copy count */
8011 if ((copy_count = len_a - i_a) > 0) {
8012 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8014 else if ((copy_count = len_b - i_b) > 0) {
8015 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8019 /* We may be removing a reference to one of the inputs */
8020 if (a == *i || b == *i) {
8021 assert(! invlist_is_iterating(*i));
8022 SvREFCNT_dec_NN(*i);
8030 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8032 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8033 * set. A pointer to the inversion list is returned. This may actually be
8034 * a new list, in which case the passed in one has been destroyed. The
8035 * passed in inversion list can be NULL, in which case a new one is created
8036 * with just the one range in it */
8041 if (invlist == NULL) {
8042 invlist = _new_invlist(2);
8046 len = _invlist_len(invlist);
8049 /* If comes after the final entry actually in the list, can just append it
8052 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
8053 && start >= invlist_array(invlist)[len - 1]))
8055 _append_range_to_invlist(invlist, start, end);
8059 /* Here, can't just append things, create and return a new inversion list
8060 * which is the union of this range and the existing inversion list */
8061 range_invlist = _new_invlist(2);
8062 _append_range_to_invlist(range_invlist, start, end);
8064 _invlist_union(invlist, range_invlist, &invlist);
8066 /* The temporary can be freed */
8067 SvREFCNT_dec_NN(range_invlist);
8074 PERL_STATIC_INLINE SV*
8075 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8076 return _add_range_to_invlist(invlist, cp, cp);
8079 #ifndef PERL_IN_XSUB_RE
8081 Perl__invlist_invert(pTHX_ SV* const invlist)
8083 /* Complement the input inversion list. This adds a 0 if the list didn't
8084 * have a zero; removes it otherwise. As described above, the data
8085 * structure is set up so that this is very efficient */
8087 PERL_ARGS_ASSERT__INVLIST_INVERT;
8089 assert(! invlist_is_iterating(invlist));
8091 /* The inverse of matching nothing is matching everything */
8092 if (_invlist_len(invlist) == 0) {
8093 _append_range_to_invlist(invlist, 0, UV_MAX);
8097 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
8101 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8103 /* Complement the input inversion list (which must be a Unicode property,
8104 * all of which don't match above the Unicode maximum code point.) And
8105 * Perl has chosen to not have the inversion match above that either. This
8106 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8112 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8114 _invlist_invert(invlist);
8116 len = _invlist_len(invlist);
8118 if (len != 0) { /* If empty do nothing */
8119 array = invlist_array(invlist);
8120 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8121 /* Add 0x110000. First, grow if necessary */
8123 if (invlist_max(invlist) < len) {
8124 invlist_extend(invlist, len);
8125 array = invlist_array(invlist);
8127 invlist_set_len(invlist, len, *get_invlist_offset_addr(invlist));
8128 array[len - 1] = PERL_UNICODE_MAX + 1;
8130 else { /* Remove the 0x110000 */
8131 invlist_set_len(invlist, len - 1, *get_invlist_offset_addr(invlist));
8139 PERL_STATIC_INLINE SV*
8140 S_invlist_clone(pTHX_ SV* const invlist)
8143 /* Return a new inversion list that is a copy of the input one, which is
8146 /* Need to allocate extra space to accommodate Perl's addition of a
8147 * trailing NUL to SvPV's, since it thinks they are always strings */
8148 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8149 STRLEN physical_length = SvCUR(invlist);
8150 bool offset = *(get_invlist_offset_addr(invlist));
8152 PERL_ARGS_ASSERT_INVLIST_CLONE;
8154 *(get_invlist_offset_addr(new_invlist)) = offset;
8155 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
8156 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
8161 PERL_STATIC_INLINE STRLEN*
8162 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8164 /* Return the address of the UV that contains the current iteration
8167 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8169 assert(SvTYPE(invlist) == SVt_INVLIST);
8171 return &(((XINVLIST*) SvANY(invlist))->iterator);
8174 PERL_STATIC_INLINE void
8175 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8177 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8179 *get_invlist_iter_addr(invlist) = 0;
8182 PERL_STATIC_INLINE void
8183 S_invlist_iterfinish(pTHX_ SV* invlist)
8185 /* Terminate iterator for invlist. This is to catch development errors.
8186 * Any iteration that is interrupted before completed should call this
8187 * function. Functions that add code points anywhere else but to the end
8188 * of an inversion list assert that they are not in the middle of an
8189 * iteration. If they were, the addition would make the iteration
8190 * problematical: if the iteration hadn't reached the place where things
8191 * were being added, it would be ok */
8193 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
8195 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
8199 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8201 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8202 * This call sets in <*start> and <*end>, the next range in <invlist>.
8203 * Returns <TRUE> if successful and the next call will return the next
8204 * range; <FALSE> if was already at the end of the list. If the latter,
8205 * <*start> and <*end> are unchanged, and the next call to this function
8206 * will start over at the beginning of the list */
8208 STRLEN* pos = get_invlist_iter_addr(invlist);
8209 UV len = _invlist_len(invlist);
8212 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8215 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
8219 array = invlist_array(invlist);
8221 *start = array[(*pos)++];
8227 *end = array[(*pos)++] - 1;
8233 PERL_STATIC_INLINE bool
8234 S_invlist_is_iterating(pTHX_ SV* const invlist)
8236 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8238 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8241 PERL_STATIC_INLINE UV
8242 S_invlist_highest(pTHX_ SV* const invlist)
8244 /* Returns the highest code point that matches an inversion list. This API
8245 * has an ambiguity, as it returns 0 under either the highest is actually
8246 * 0, or if the list is empty. If this distinction matters to you, check
8247 * for emptiness before calling this function */
8249 UV len = _invlist_len(invlist);
8252 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8258 array = invlist_array(invlist);
8260 /* The last element in the array in the inversion list always starts a
8261 * range that goes to infinity. That range may be for code points that are
8262 * matched in the inversion list, or it may be for ones that aren't
8263 * matched. In the latter case, the highest code point in the set is one
8264 * less than the beginning of this range; otherwise it is the final element
8265 * of this range: infinity */
8266 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8268 : array[len - 1] - 1;
8271 #ifndef PERL_IN_XSUB_RE
8273 Perl__invlist_contents(pTHX_ SV* const invlist)
8275 /* Get the contents of an inversion list into a string SV so that they can
8276 * be printed out. It uses the format traditionally done for debug tracing
8280 SV* output = newSVpvs("\n");
8282 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8284 assert(! invlist_is_iterating(invlist));
8286 invlist_iterinit(invlist);
8287 while (invlist_iternext(invlist, &start, &end)) {
8288 if (end == UV_MAX) {
8289 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8291 else if (end != start) {
8292 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8296 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8304 #ifndef PERL_IN_XSUB_RE
8306 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level, const char * const indent, SV* const invlist)
8308 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
8309 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
8310 * the string 'indent'. The output looks like this:
8311 [0] 0x000A .. 0x000D
8313 [4] 0x2028 .. 0x2029
8314 [6] 0x3104 .. INFINITY
8315 * This means that the first range of code points matched by the list are
8316 * 0xA through 0xD; the second range contains only the single code point
8317 * 0x85, etc. An inversion list is an array of UVs. Two array elements
8318 * are used to define each range (except if the final range extends to
8319 * infinity, only a single element is needed). The array index of the
8320 * first element for the corresponding range is given in brackets. */
8325 PERL_ARGS_ASSERT__INVLIST_DUMP;
8327 if (invlist_is_iterating(invlist)) {
8328 Perl_dump_indent(aTHX_ level, file,
8329 "%sCan't dump inversion list because is in middle of iterating\n",
8334 invlist_iterinit(invlist);
8335 while (invlist_iternext(invlist, &start, &end)) {
8336 if (end == UV_MAX) {
8337 Perl_dump_indent(aTHX_ level, file,
8338 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
8339 indent, (UV)count, start);
8341 else if (end != start) {
8342 Perl_dump_indent(aTHX_ level, file,
8343 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
8344 indent, (UV)count, start, end);
8347 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
8348 indent, (UV)count, start);
8355 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
8357 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
8359 /* Return a boolean as to if the two passed in inversion lists are
8360 * identical. The final argument, if TRUE, says to take the complement of
8361 * the second inversion list before doing the comparison */
8363 const UV* array_a = invlist_array(a);
8364 const UV* array_b = invlist_array(b);
8365 UV len_a = _invlist_len(a);
8366 UV len_b = _invlist_len(b);
8368 UV i = 0; /* current index into the arrays */
8369 bool retval = TRUE; /* Assume are identical until proven otherwise */
8371 PERL_ARGS_ASSERT__INVLISTEQ;
8373 /* If are to compare 'a' with the complement of b, set it
8374 * up so are looking at b's complement. */
8377 /* The complement of nothing is everything, so <a> would have to have
8378 * just one element, starting at zero (ending at infinity) */
8380 return (len_a == 1 && array_a[0] == 0);
8382 else if (array_b[0] == 0) {
8384 /* Otherwise, to complement, we invert. Here, the first element is
8385 * 0, just remove it. To do this, we just pretend the array starts
8393 /* But if the first element is not zero, we pretend the list starts
8394 * at the 0 that is always stored immediately before the array. */
8400 /* Make sure that the lengths are the same, as well as the final element
8401 * before looping through the remainder. (Thus we test the length, final,
8402 * and first elements right off the bat) */
8403 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8406 else for (i = 0; i < len_a - 1; i++) {
8407 if (array_a[i] != array_b[i]) {
8417 #undef HEADER_LENGTH
8418 #undef TO_INTERNAL_SIZE
8419 #undef FROM_INTERNAL_SIZE
8420 #undef INVLIST_VERSION_ID
8422 /* End of inversion list object */
8425 S_parse_lparen_question_flags(pTHX_ struct RExC_state_t *pRExC_state)
8427 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
8428 * constructs, and updates RExC_flags with them. On input, RExC_parse
8429 * should point to the first flag; it is updated on output to point to the
8430 * final ')' or ':'. There needs to be at least one flag, or this will
8433 /* for (?g), (?gc), and (?o) warnings; warning
8434 about (?c) will warn about (?g) -- japhy */
8436 #define WASTED_O 0x01
8437 #define WASTED_G 0x02
8438 #define WASTED_C 0x04
8439 #define WASTED_GC (WASTED_G|WASTED_C)
8440 I32 wastedflags = 0x00;
8441 U32 posflags = 0, negflags = 0;
8442 U32 *flagsp = &posflags;
8443 char has_charset_modifier = '\0';
8445 bool has_use_defaults = FALSE;
8446 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
8448 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
8450 /* '^' as an initial flag sets certain defaults */
8451 if (UCHARAT(RExC_parse) == '^') {
8453 has_use_defaults = TRUE;
8454 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8455 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8456 ? REGEX_UNICODE_CHARSET
8457 : REGEX_DEPENDS_CHARSET);
8460 cs = get_regex_charset(RExC_flags);
8461 if (cs == REGEX_DEPENDS_CHARSET
8462 && (RExC_utf8 || RExC_uni_semantics))
8464 cs = REGEX_UNICODE_CHARSET;
8467 while (*RExC_parse) {
8468 /* && strchr("iogcmsx", *RExC_parse) */
8469 /* (?g), (?gc) and (?o) are useless here
8470 and must be globally applied -- japhy */
8471 switch (*RExC_parse) {
8473 /* Code for the imsx flags */
8474 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8476 case LOCALE_PAT_MOD:
8477 if (has_charset_modifier) {
8478 goto excess_modifier;
8480 else if (flagsp == &negflags) {
8483 cs = REGEX_LOCALE_CHARSET;
8484 has_charset_modifier = LOCALE_PAT_MOD;
8485 RExC_contains_locale = 1;
8487 case UNICODE_PAT_MOD:
8488 if (has_charset_modifier) {
8489 goto excess_modifier;
8491 else if (flagsp == &negflags) {
8494 cs = REGEX_UNICODE_CHARSET;
8495 has_charset_modifier = UNICODE_PAT_MOD;
8497 case ASCII_RESTRICT_PAT_MOD:
8498 if (flagsp == &negflags) {
8501 if (has_charset_modifier) {
8502 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8503 goto excess_modifier;
8505 /* Doubled modifier implies more restricted */
8506 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8509 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8511 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8513 case DEPENDS_PAT_MOD:
8514 if (has_use_defaults) {
8515 goto fail_modifiers;
8517 else if (flagsp == &negflags) {
8520 else if (has_charset_modifier) {
8521 goto excess_modifier;
8524 /* The dual charset means unicode semantics if the
8525 * pattern (or target, not known until runtime) are
8526 * utf8, or something in the pattern indicates unicode
8528 cs = (RExC_utf8 || RExC_uni_semantics)
8529 ? REGEX_UNICODE_CHARSET
8530 : REGEX_DEPENDS_CHARSET;
8531 has_charset_modifier = DEPENDS_PAT_MOD;
8535 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
8536 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
8538 else if (has_charset_modifier == *(RExC_parse - 1)) {
8539 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
8542 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
8547 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
8549 case ONCE_PAT_MOD: /* 'o' */
8550 case GLOBAL_PAT_MOD: /* 'g' */
8551 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8552 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
8553 if (! (wastedflags & wflagbit) ) {
8554 wastedflags |= wflagbit;
8555 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8558 "Useless (%s%c) - %suse /%c modifier",
8559 flagsp == &negflags ? "?-" : "?",
8561 flagsp == &negflags ? "don't " : "",
8568 case CONTINUE_PAT_MOD: /* 'c' */
8569 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
8570 if (! (wastedflags & WASTED_C) ) {
8571 wastedflags |= WASTED_GC;
8572 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
8575 "Useless (%sc) - %suse /gc modifier",
8576 flagsp == &negflags ? "?-" : "?",
8577 flagsp == &negflags ? "don't " : ""
8582 case KEEPCOPY_PAT_MOD: /* 'p' */
8583 if (flagsp == &negflags) {
8585 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
8587 *flagsp |= RXf_PMf_KEEPCOPY;
8591 /* A flag is a default iff it is following a minus, so
8592 * if there is a minus, it means will be trying to
8593 * re-specify a default which is an error */
8594 if (has_use_defaults || flagsp == &negflags) {
8595 goto fail_modifiers;
8598 wastedflags = 0; /* reset so (?g-c) warns twice */
8602 RExC_flags |= posflags;
8603 RExC_flags &= ~negflags;
8604 set_regex_charset(&RExC_flags, cs);
8610 vFAIL3("Sequence (%.*s...) not recognized",
8611 RExC_parse-seqstart, seqstart);
8620 - reg - regular expression, i.e. main body or parenthesized thing
8622 * Caller must absorb opening parenthesis.
8624 * Combining parenthesis handling with the base level of regular expression
8625 * is a trifle forced, but the need to tie the tails of the branches to what
8626 * follows makes it hard to avoid.
8628 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8630 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8632 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8635 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
8636 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
8637 needs to be restarted.
8638 Otherwise would only return NULL if regbranch() returns NULL, which
8641 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8642 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
8643 * 2 is like 1, but indicates that nextchar() has been called to advance
8644 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
8645 * this flag alerts us to the need to check for that */
8648 regnode *ret; /* Will be the head of the group. */
8651 regnode *ender = NULL;
8654 U32 oregflags = RExC_flags;
8655 bool have_branch = 0;
8657 I32 freeze_paren = 0;
8658 I32 after_freeze = 0;
8660 char * parse_start = RExC_parse; /* MJD */
8661 char * const oregcomp_parse = RExC_parse;
8663 GET_RE_DEBUG_FLAGS_DECL;
8665 PERL_ARGS_ASSERT_REG;
8666 DEBUG_PARSE("reg ");
8668 *flagp = 0; /* Tentatively. */
8671 /* Make an OPEN node, if parenthesized. */
8674 /* Under /x, space and comments can be gobbled up between the '(' and
8675 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
8676 * intervening space, as the sequence is a token, and a token should be
8678 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
8680 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8681 char *start_verb = RExC_parse;
8682 STRLEN verb_len = 0;
8683 char *start_arg = NULL;
8684 unsigned char op = 0;
8686 int internal_argval = 0; /* internal_argval is only useful if !argok */
8688 if (has_intervening_patws && SIZE_ONLY) {
8689 ckWARNregdep(RExC_parse + 1, "In '(*VERB...)', splitting the initial '(*' is deprecated");
8691 while ( *RExC_parse && *RExC_parse != ')' ) {
8692 if ( *RExC_parse == ':' ) {
8693 start_arg = RExC_parse + 1;
8699 verb_len = RExC_parse - start_verb;
8702 while ( *RExC_parse && *RExC_parse != ')' )
8704 if ( *RExC_parse != ')' )
8705 vFAIL("Unterminated verb pattern argument");
8706 if ( RExC_parse == start_arg )
8709 if ( *RExC_parse != ')' )
8710 vFAIL("Unterminated verb pattern");
8713 switch ( *start_verb ) {
8714 case 'A': /* (*ACCEPT) */
8715 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8717 internal_argval = RExC_nestroot;
8720 case 'C': /* (*COMMIT) */
8721 if ( memEQs(start_verb,verb_len,"COMMIT") )
8724 case 'F': /* (*FAIL) */
8725 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8730 case ':': /* (*:NAME) */
8731 case 'M': /* (*MARK:NAME) */
8732 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8737 case 'P': /* (*PRUNE) */
8738 if ( memEQs(start_verb,verb_len,"PRUNE") )
8741 case 'S': /* (*SKIP) */
8742 if ( memEQs(start_verb,verb_len,"SKIP") )
8745 case 'T': /* (*THEN) */
8746 /* [19:06] <TimToady> :: is then */
8747 if ( memEQs(start_verb,verb_len,"THEN") ) {
8749 RExC_seen |= REG_SEEN_CUTGROUP;
8755 vFAIL3("Unknown verb pattern '%.*s'",
8756 verb_len, start_verb);
8759 if ( start_arg && internal_argval ) {
8760 vFAIL3("Verb pattern '%.*s' may not have an argument",
8761 verb_len, start_verb);
8762 } else if ( argok < 0 && !start_arg ) {
8763 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8764 verb_len, start_verb);
8766 ret = reganode(pRExC_state, op, internal_argval);
8767 if ( ! internal_argval && ! SIZE_ONLY ) {
8769 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8770 ARG(ret) = add_data( pRExC_state, 1, "S" );
8771 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8778 if (!internal_argval)
8779 RExC_seen |= REG_SEEN_VERBARG;
8780 } else if ( start_arg ) {
8781 vFAIL3("Verb pattern '%.*s' may not have an argument",
8782 verb_len, start_verb);
8784 ret = reg_node(pRExC_state, op);
8786 nextchar(pRExC_state);
8789 else if (*RExC_parse == '?') { /* (?...) */
8790 bool is_logical = 0;
8791 const char * const seqstart = RExC_parse;
8792 if (has_intervening_patws && SIZE_ONLY) {
8793 ckWARNregdep(RExC_parse + 1, "In '(?...)', splitting the initial '(?' is deprecated");
8797 paren = *RExC_parse++;
8798 ret = NULL; /* For look-ahead/behind. */
8801 case 'P': /* (?P...) variants for those used to PCRE/Python */
8802 paren = *RExC_parse++;
8803 if ( paren == '<') /* (?P<...>) named capture */
8805 else if (paren == '>') { /* (?P>name) named recursion */
8806 goto named_recursion;
8808 else if (paren == '=') { /* (?P=...) named backref */
8809 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8810 you change this make sure you change that */
8811 char* name_start = RExC_parse;
8813 SV *sv_dat = reg_scan_name(pRExC_state,
8814 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8815 if (RExC_parse == name_start || *RExC_parse != ')')
8816 vFAIL2("Sequence %.3s... not terminated",parse_start);
8819 num = add_data( pRExC_state, 1, "S" );
8820 RExC_rxi->data->data[num]=(void*)sv_dat;
8821 SvREFCNT_inc_simple_void(sv_dat);
8824 ret = reganode(pRExC_state,
8827 : (ASCII_FOLD_RESTRICTED)
8829 : (AT_LEAST_UNI_SEMANTICS)
8837 Set_Node_Offset(ret, parse_start+1);
8838 Set_Node_Cur_Length(ret, parse_start);
8840 nextchar(pRExC_state);
8844 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8846 case '<': /* (?<...) */
8847 if (*RExC_parse == '!')
8849 else if (*RExC_parse != '=')
8855 case '\'': /* (?'...') */
8856 name_start= RExC_parse;
8857 svname = reg_scan_name(pRExC_state,
8858 SIZE_ONLY ? /* reverse test from the others */
8859 REG_RSN_RETURN_NAME :
8860 REG_RSN_RETURN_NULL);
8861 if (RExC_parse == name_start) {
8863 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8866 if (*RExC_parse != paren)
8867 vFAIL2("Sequence (?%c... not terminated",
8868 paren=='>' ? '<' : paren);
8872 if (!svname) /* shouldn't happen */
8874 "panic: reg_scan_name returned NULL");
8875 if (!RExC_paren_names) {
8876 RExC_paren_names= newHV();
8877 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8879 RExC_paren_name_list= newAV();
8880 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8883 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8885 sv_dat = HeVAL(he_str);
8887 /* croak baby croak */
8889 "panic: paren_name hash element allocation failed");
8890 } else if ( SvPOK(sv_dat) ) {
8891 /* (?|...) can mean we have dupes so scan to check
8892 its already been stored. Maybe a flag indicating
8893 we are inside such a construct would be useful,
8894 but the arrays are likely to be quite small, so
8895 for now we punt -- dmq */
8896 IV count = SvIV(sv_dat);
8897 I32 *pv = (I32*)SvPVX(sv_dat);
8899 for ( i = 0 ; i < count ; i++ ) {
8900 if ( pv[i] == RExC_npar ) {
8906 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8907 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8908 pv[count] = RExC_npar;
8909 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8912 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8913 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8915 SvIV_set(sv_dat, 1);
8918 /* Yes this does cause a memory leak in debugging Perls */
8919 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8920 SvREFCNT_dec_NN(svname);
8923 /*sv_dump(sv_dat);*/
8925 nextchar(pRExC_state);
8927 goto capturing_parens;
8929 RExC_seen |= REG_SEEN_LOOKBEHIND;
8930 RExC_in_lookbehind++;
8932 case '=': /* (?=...) */
8933 RExC_seen_zerolen++;
8935 case '!': /* (?!...) */
8936 RExC_seen_zerolen++;
8937 if (*RExC_parse == ')') {
8938 ret=reg_node(pRExC_state, OPFAIL);
8939 nextchar(pRExC_state);
8943 case '|': /* (?|...) */
8944 /* branch reset, behave like a (?:...) except that
8945 buffers in alternations share the same numbers */
8947 after_freeze = freeze_paren = RExC_npar;
8949 case ':': /* (?:...) */
8950 case '>': /* (?>...) */
8952 case '$': /* (?$...) */
8953 case '@': /* (?@...) */
8954 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8956 case '#': /* (?#...) */
8957 /* XXX As soon as we disallow separating the '?' and '*' (by
8958 * spaces or (?#...) comment), it is believed that this case
8959 * will be unreachable and can be removed. See
8961 while (*RExC_parse && *RExC_parse != ')')
8963 if (*RExC_parse != ')')
8964 FAIL("Sequence (?#... not terminated");
8965 nextchar(pRExC_state);
8968 case '0' : /* (?0) */
8969 case 'R' : /* (?R) */
8970 if (*RExC_parse != ')')
8971 FAIL("Sequence (?R) not terminated");
8972 ret = reg_node(pRExC_state, GOSTART);
8973 *flagp |= POSTPONED;
8974 nextchar(pRExC_state);
8977 { /* named and numeric backreferences */
8979 case '&': /* (?&NAME) */
8980 parse_start = RExC_parse - 1;
8983 SV *sv_dat = reg_scan_name(pRExC_state,
8984 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8985 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8987 goto gen_recurse_regop;
8988 assert(0); /* NOT REACHED */
8990 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8992 vFAIL("Illegal pattern");
8994 goto parse_recursion;
8996 case '-': /* (?-1) */
8997 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8998 RExC_parse--; /* rewind to let it be handled later */
9002 case '1': case '2': case '3': case '4': /* (?1) */
9003 case '5': case '6': case '7': case '8': case '9':
9006 num = atoi(RExC_parse);
9007 parse_start = RExC_parse - 1; /* MJD */
9008 if (*RExC_parse == '-')
9010 while (isDIGIT(*RExC_parse))
9012 if (*RExC_parse!=')')
9013 vFAIL("Expecting close bracket");
9016 if ( paren == '-' ) {
9018 Diagram of capture buffer numbering.
9019 Top line is the normal capture buffer numbers
9020 Bottom line is the negative indexing as from
9024 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
9028 num = RExC_npar + num;
9031 vFAIL("Reference to nonexistent group");
9033 } else if ( paren == '+' ) {
9034 num = RExC_npar + num - 1;
9037 ret = reganode(pRExC_state, GOSUB, num);
9039 if (num > (I32)RExC_rx->nparens) {
9041 vFAIL("Reference to nonexistent group");
9043 ARG2L_SET( ret, RExC_recurse_count++);
9045 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9046 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
9050 RExC_seen |= REG_SEEN_RECURSE;
9051 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
9052 Set_Node_Offset(ret, parse_start); /* MJD */
9054 *flagp |= POSTPONED;
9055 nextchar(pRExC_state);
9057 } /* named and numeric backreferences */
9058 assert(0); /* NOT REACHED */
9060 case '?': /* (??...) */
9062 if (*RExC_parse != '{') {
9064 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9067 *flagp |= POSTPONED;
9068 paren = *RExC_parse++;
9070 case '{': /* (?{...}) */
9073 struct reg_code_block *cb;
9075 RExC_seen_zerolen++;
9077 if ( !pRExC_state->num_code_blocks
9078 || pRExC_state->code_index >= pRExC_state->num_code_blocks
9079 || pRExC_state->code_blocks[pRExC_state->code_index].start
9080 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
9083 if (RExC_pm_flags & PMf_USE_RE_EVAL)
9084 FAIL("panic: Sequence (?{...}): no code block found\n");
9085 FAIL("Eval-group not allowed at runtime, use re 'eval'");
9087 /* this is a pre-compiled code block (?{...}) */
9088 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
9089 RExC_parse = RExC_start + cb->end;
9092 if (cb->src_regex) {
9093 n = add_data(pRExC_state, 2, "rl");
9094 RExC_rxi->data->data[n] =
9095 (void*)SvREFCNT_inc((SV*)cb->src_regex);
9096 RExC_rxi->data->data[n+1] = (void*)o;
9099 n = add_data(pRExC_state, 1,
9100 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
9101 RExC_rxi->data->data[n] = (void*)o;
9104 pRExC_state->code_index++;
9105 nextchar(pRExC_state);
9109 ret = reg_node(pRExC_state, LOGICAL);
9110 eval = reganode(pRExC_state, EVAL, n);
9113 /* for later propagation into (??{}) return value */
9114 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
9116 REGTAIL(pRExC_state, ret, eval);
9117 /* deal with the length of this later - MJD */
9120 ret = reganode(pRExC_state, EVAL, n);
9121 Set_Node_Length(ret, RExC_parse - parse_start + 1);
9122 Set_Node_Offset(ret, parse_start);
9125 case '(': /* (?(?{...})...) and (?(?=...)...) */
9128 if (RExC_parse[0] == '?') { /* (?(?...)) */
9129 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
9130 || RExC_parse[1] == '<'
9131 || RExC_parse[1] == '{') { /* Lookahead or eval. */
9135 ret = reg_node(pRExC_state, LOGICAL);
9139 tail = reg(pRExC_state, 1, &flag, depth+1);
9140 if (flag & RESTART_UTF8) {
9141 *flagp = RESTART_UTF8;
9144 REGTAIL(pRExC_state, ret, tail);
9148 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
9149 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
9151 char ch = RExC_parse[0] == '<' ? '>' : '\'';
9152 char *name_start= RExC_parse++;
9154 SV *sv_dat=reg_scan_name(pRExC_state,
9155 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9156 if (RExC_parse == name_start || *RExC_parse != ch)
9157 vFAIL2("Sequence (?(%c... not terminated",
9158 (ch == '>' ? '<' : ch));
9161 num = add_data( pRExC_state, 1, "S" );
9162 RExC_rxi->data->data[num]=(void*)sv_dat;
9163 SvREFCNT_inc_simple_void(sv_dat);
9165 ret = reganode(pRExC_state,NGROUPP,num);
9166 goto insert_if_check_paren;
9168 else if (RExC_parse[0] == 'D' &&
9169 RExC_parse[1] == 'E' &&
9170 RExC_parse[2] == 'F' &&
9171 RExC_parse[3] == 'I' &&
9172 RExC_parse[4] == 'N' &&
9173 RExC_parse[5] == 'E')
9175 ret = reganode(pRExC_state,DEFINEP,0);
9178 goto insert_if_check_paren;
9180 else if (RExC_parse[0] == 'R') {
9183 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9184 parno = atoi(RExC_parse++);
9185 while (isDIGIT(*RExC_parse))
9187 } else if (RExC_parse[0] == '&') {
9190 sv_dat = reg_scan_name(pRExC_state,
9191 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9192 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
9194 ret = reganode(pRExC_state,INSUBP,parno);
9195 goto insert_if_check_paren;
9197 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
9200 parno = atoi(RExC_parse++);
9202 while (isDIGIT(*RExC_parse))
9204 ret = reganode(pRExC_state, GROUPP, parno);
9206 insert_if_check_paren:
9207 if ((c = *nextchar(pRExC_state)) != ')')
9208 vFAIL("Switch condition not recognized");
9210 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
9211 br = regbranch(pRExC_state, &flags, 1,depth+1);
9213 if (flags & RESTART_UTF8) {
9214 *flagp = RESTART_UTF8;
9217 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9220 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
9221 c = *nextchar(pRExC_state);
9226 vFAIL("(?(DEFINE)....) does not allow branches");
9227 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
9228 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
9229 if (flags & RESTART_UTF8) {
9230 *flagp = RESTART_UTF8;
9233 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
9236 REGTAIL(pRExC_state, ret, lastbr);
9239 c = *nextchar(pRExC_state);
9244 vFAIL("Switch (?(condition)... contains too many branches");
9245 ender = reg_node(pRExC_state, TAIL);
9246 REGTAIL(pRExC_state, br, ender);
9248 REGTAIL(pRExC_state, lastbr, ender);
9249 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
9252 REGTAIL(pRExC_state, ret, ender);
9253 RExC_size++; /* XXX WHY do we need this?!!
9254 For large programs it seems to be required
9255 but I can't figure out why. -- dmq*/
9259 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
9262 case '[': /* (?[ ... ]) */
9263 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
9266 RExC_parse--; /* for vFAIL to print correctly */
9267 vFAIL("Sequence (? incomplete");
9269 default: /* e.g., (?i) */
9272 parse_lparen_question_flags(pRExC_state);
9273 if (UCHARAT(RExC_parse) != ':') {
9274 nextchar(pRExC_state);
9279 nextchar(pRExC_state);
9289 ret = reganode(pRExC_state, OPEN, parno);
9292 RExC_nestroot = parno;
9293 if (RExC_seen & REG_SEEN_RECURSE
9294 && !RExC_open_parens[parno-1])
9296 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9297 "Setting open paren #%"IVdf" to %d\n",
9298 (IV)parno, REG_NODE_NUM(ret)));
9299 RExC_open_parens[parno-1]= ret;
9302 Set_Node_Length(ret, 1); /* MJD */
9303 Set_Node_Offset(ret, RExC_parse); /* MJD */
9311 /* Pick up the branches, linking them together. */
9312 parse_start = RExC_parse; /* MJD */
9313 br = regbranch(pRExC_state, &flags, 1,depth+1);
9315 /* branch_len = (paren != 0); */
9318 if (flags & RESTART_UTF8) {
9319 *flagp = RESTART_UTF8;
9322 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9324 if (*RExC_parse == '|') {
9325 if (!SIZE_ONLY && RExC_extralen) {
9326 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9329 reginsert(pRExC_state, BRANCH, br, depth+1);
9330 Set_Node_Length(br, paren != 0);
9331 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9335 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9337 else if (paren == ':') {
9338 *flagp |= flags&SIMPLE;
9340 if (is_open) { /* Starts with OPEN. */
9341 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9343 else if (paren != '?') /* Not Conditional */
9345 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9347 while (*RExC_parse == '|') {
9348 if (!SIZE_ONLY && RExC_extralen) {
9349 ender = reganode(pRExC_state, LONGJMP,0);
9350 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9353 RExC_extralen += 2; /* Account for LONGJMP. */
9354 nextchar(pRExC_state);
9356 if (RExC_npar > after_freeze)
9357 after_freeze = RExC_npar;
9358 RExC_npar = freeze_paren;
9360 br = regbranch(pRExC_state, &flags, 0, depth+1);
9363 if (flags & RESTART_UTF8) {
9364 *flagp = RESTART_UTF8;
9367 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
9369 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9371 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9374 if (have_branch || paren != ':') {
9375 /* Make a closing node, and hook it on the end. */
9378 ender = reg_node(pRExC_state, TAIL);
9381 ender = reganode(pRExC_state, CLOSE, parno);
9382 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9383 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9384 "Setting close paren #%"IVdf" to %d\n",
9385 (IV)parno, REG_NODE_NUM(ender)));
9386 RExC_close_parens[parno-1]= ender;
9387 if (RExC_nestroot == parno)
9390 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9391 Set_Node_Length(ender,1); /* MJD */
9397 *flagp &= ~HASWIDTH;
9400 ender = reg_node(pRExC_state, SUCCEED);
9403 ender = reg_node(pRExC_state, END);
9405 assert(!RExC_opend); /* there can only be one! */
9410 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9411 SV * const mysv_val1=sv_newmortal();
9412 SV * const mysv_val2=sv_newmortal();
9413 DEBUG_PARSE_MSG("lsbr");
9414 regprop(RExC_rx, mysv_val1, lastbr);
9415 regprop(RExC_rx, mysv_val2, ender);
9416 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9417 SvPV_nolen_const(mysv_val1),
9418 (IV)REG_NODE_NUM(lastbr),
9419 SvPV_nolen_const(mysv_val2),
9420 (IV)REG_NODE_NUM(ender),
9421 (IV)(ender - lastbr)
9424 REGTAIL(pRExC_state, lastbr, ender);
9426 if (have_branch && !SIZE_ONLY) {
9429 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9431 /* Hook the tails of the branches to the closing node. */
9432 for (br = ret; br; br = regnext(br)) {
9433 const U8 op = PL_regkind[OP(br)];
9435 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9436 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9439 else if (op == BRANCHJ) {
9440 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9441 /* for now we always disable this optimisation * /
9442 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9448 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9449 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9450 SV * const mysv_val1=sv_newmortal();
9451 SV * const mysv_val2=sv_newmortal();
9452 DEBUG_PARSE_MSG("NADA");
9453 regprop(RExC_rx, mysv_val1, ret);
9454 regprop(RExC_rx, mysv_val2, ender);
9455 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9456 SvPV_nolen_const(mysv_val1),
9457 (IV)REG_NODE_NUM(ret),
9458 SvPV_nolen_const(mysv_val2),
9459 (IV)REG_NODE_NUM(ender),
9464 if (OP(ender) == TAIL) {
9469 for ( opt= br + 1; opt < ender ; opt++ )
9471 NEXT_OFF(br)= ender - br;
9479 static const char parens[] = "=!<,>";
9481 if (paren && (p = strchr(parens, paren))) {
9482 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9483 int flag = (p - parens) > 1;
9486 node = SUSPEND, flag = 0;
9487 reginsert(pRExC_state, node,ret, depth+1);
9488 Set_Node_Cur_Length(ret, parse_start);
9489 Set_Node_Offset(ret, parse_start + 1);
9491 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9495 /* Check for proper termination. */
9497 /* restore original flags, but keep (?p) */
9498 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
9499 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9500 RExC_parse = oregcomp_parse;
9501 vFAIL("Unmatched (");
9504 else if (!paren && RExC_parse < RExC_end) {
9505 if (*RExC_parse == ')') {
9507 vFAIL("Unmatched )");
9510 FAIL("Junk on end of regexp"); /* "Can't happen". */
9511 assert(0); /* NOTREACHED */
9514 if (RExC_in_lookbehind) {
9515 RExC_in_lookbehind--;
9517 if (after_freeze > RExC_npar)
9518 RExC_npar = after_freeze;
9523 - regbranch - one alternative of an | operator
9525 * Implements the concatenation operator.
9527 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9531 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9535 regnode *chain = NULL;
9537 I32 flags = 0, c = 0;
9538 GET_RE_DEBUG_FLAGS_DECL;
9540 PERL_ARGS_ASSERT_REGBRANCH;
9542 DEBUG_PARSE("brnc");
9547 if (!SIZE_ONLY && RExC_extralen)
9548 ret = reganode(pRExC_state, BRANCHJ,0);
9550 ret = reg_node(pRExC_state, BRANCH);
9551 Set_Node_Length(ret, 1);
9555 if (!first && SIZE_ONLY)
9556 RExC_extralen += 1; /* BRANCHJ */
9558 *flagp = WORST; /* Tentatively. */
9561 nextchar(pRExC_state);
9562 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9564 latest = regpiece(pRExC_state, &flags,depth+1);
9565 if (latest == NULL) {
9566 if (flags & TRYAGAIN)
9568 if (flags & RESTART_UTF8) {
9569 *flagp = RESTART_UTF8;
9572 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
9574 else if (ret == NULL)
9576 *flagp |= flags&(HASWIDTH|POSTPONED);
9577 if (chain == NULL) /* First piece. */
9578 *flagp |= flags&SPSTART;
9581 REGTAIL(pRExC_state, chain, latest);
9586 if (chain == NULL) { /* Loop ran zero times. */
9587 chain = reg_node(pRExC_state, NOTHING);
9592 *flagp |= flags&SIMPLE;
9599 - regpiece - something followed by possible [*+?]
9601 * Note that the branching code sequences used for ? and the general cases
9602 * of * and + are somewhat optimized: they use the same NOTHING node as
9603 * both the endmarker for their branch list and the body of the last branch.
9604 * It might seem that this node could be dispensed with entirely, but the
9605 * endmarker role is not redundant.
9607 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
9609 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
9613 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9620 const char * const origparse = RExC_parse;
9622 I32 max = REG_INFTY;
9623 #ifdef RE_TRACK_PATTERN_OFFSETS
9626 const char *maxpos = NULL;
9628 /* Save the original in case we change the emitted regop to a FAIL. */
9629 regnode * const orig_emit = RExC_emit;
9631 GET_RE_DEBUG_FLAGS_DECL;
9633 PERL_ARGS_ASSERT_REGPIECE;
9635 DEBUG_PARSE("piec");
9637 ret = regatom(pRExC_state, &flags,depth+1);
9639 if (flags & (TRYAGAIN|RESTART_UTF8))
9640 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
9642 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
9648 if (op == '{' && regcurly(RExC_parse, FALSE)) {
9650 #ifdef RE_TRACK_PATTERN_OFFSETS
9651 parse_start = RExC_parse; /* MJD */
9653 next = RExC_parse + 1;
9654 while (isDIGIT(*next) || *next == ',') {
9663 if (*next == '}') { /* got one */
9667 min = atoi(RExC_parse);
9671 maxpos = RExC_parse;
9673 if (!max && *maxpos != '0')
9674 max = REG_INFTY; /* meaning "infinity" */
9675 else if (max >= REG_INFTY)
9676 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9678 nextchar(pRExC_state);
9679 if (max < min) { /* If can't match, warn and optimize to fail
9682 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9684 /* We can't back off the size because we have to reserve
9685 * enough space for all the things we are about to throw
9686 * away, but we can shrink it by the ammount we are about
9688 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9691 RExC_emit = orig_emit;
9693 ret = reg_node(pRExC_state, OPFAIL);
9698 if ((flags&SIMPLE)) {
9699 RExC_naughty += 2 + RExC_naughty / 2;
9700 reginsert(pRExC_state, CURLY, ret, depth+1);
9701 Set_Node_Offset(ret, parse_start+1); /* MJD */
9702 Set_Node_Cur_Length(ret, parse_start);
9705 regnode * const w = reg_node(pRExC_state, WHILEM);
9708 REGTAIL(pRExC_state, ret, w);
9709 if (!SIZE_ONLY && RExC_extralen) {
9710 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9711 reginsert(pRExC_state, NOTHING,ret, depth+1);
9712 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9714 reginsert(pRExC_state, CURLYX,ret, depth+1);
9716 Set_Node_Offset(ret, parse_start+1);
9717 Set_Node_Length(ret,
9718 op == '{' ? (RExC_parse - parse_start) : 1);
9720 if (!SIZE_ONLY && RExC_extralen)
9721 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9722 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9724 RExC_whilem_seen++, RExC_extralen += 3;
9725 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9734 ARG1_SET(ret, (U16)min);
9735 ARG2_SET(ret, (U16)max);
9747 #if 0 /* Now runtime fix should be reliable. */
9749 /* if this is reinstated, don't forget to put this back into perldiag:
9751 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9753 (F) The part of the regexp subject to either the * or + quantifier
9754 could match an empty string. The {#} shows in the regular
9755 expression about where the problem was discovered.
9759 if (!(flags&HASWIDTH) && op != '?')
9760 vFAIL("Regexp *+ operand could be empty");
9763 #ifdef RE_TRACK_PATTERN_OFFSETS
9764 parse_start = RExC_parse;
9766 nextchar(pRExC_state);
9768 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9770 if (op == '*' && (flags&SIMPLE)) {
9771 reginsert(pRExC_state, STAR, ret, depth+1);
9775 else if (op == '*') {
9779 else if (op == '+' && (flags&SIMPLE)) {
9780 reginsert(pRExC_state, PLUS, ret, depth+1);
9784 else if (op == '+') {
9788 else if (op == '?') {
9793 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9794 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
9795 ckWARN3reg(RExC_parse,
9796 "%.*s matches null string many times",
9797 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9799 (void)ReREFCNT_inc(RExC_rx_sv);
9802 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9803 nextchar(pRExC_state);
9804 reginsert(pRExC_state, MINMOD, ret, depth+1);
9805 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9808 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9810 nextchar(pRExC_state);
9811 ender = reg_node(pRExC_state, SUCCEED);
9812 REGTAIL(pRExC_state, ret, ender);
9813 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9815 ender = reg_node(pRExC_state, TAIL);
9816 REGTAIL(pRExC_state, ret, ender);
9819 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9821 vFAIL("Nested quantifiers");
9828 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class,
9829 const bool strict /* Apply stricter parsing rules? */
9833 /* This is expected to be called by a parser routine that has recognized '\N'
9834 and needs to handle the rest. RExC_parse is expected to point at the first
9835 char following the N at the time of the call. On successful return,
9836 RExC_parse has been updated to point to just after the sequence identified
9837 by this routine, and <*flagp> has been updated.
9839 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9842 \N may begin either a named sequence, or if outside a character class, mean
9843 to match a non-newline. For non single-quoted regexes, the tokenizer has
9844 attempted to decide which, and in the case of a named sequence, converted it
9845 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9846 where c1... are the characters in the sequence. For single-quoted regexes,
9847 the tokenizer passes the \N sequence through unchanged; this code will not
9848 attempt to determine this nor expand those, instead raising a syntax error.
9849 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9850 or there is no '}', it signals that this \N occurrence means to match a
9853 Only the \N{U+...} form should occur in a character class, for the same
9854 reason that '.' inside a character class means to just match a period: it
9855 just doesn't make sense.
9857 The function raises an error (via vFAIL), and doesn't return for various
9858 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9859 success; it returns FALSE otherwise. Returns FALSE, setting *flagp to
9860 RESTART_UTF8 if the sizing scan needs to be restarted. Such a restart is
9861 only possible if node_p is non-NULL.
9864 If <valuep> is non-null, it means the caller can accept an input sequence
9865 consisting of a just a single code point; <*valuep> is set to that value
9866 if the input is such.
9868 If <node_p> is non-null it signifies that the caller can accept any other
9869 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9871 1) \N means not-a-NL: points to a newly created REG_ANY node;
9872 2) \N{}: points to a new NOTHING node;
9873 3) otherwise: points to a new EXACT node containing the resolved
9875 Note that FALSE is returned for single code point sequences if <valuep> is
9879 char * endbrace; /* '}' following the name */
9881 char *endchar; /* Points to '.' or '}' ending cur char in the input
9883 bool has_multiple_chars; /* true if the input stream contains a sequence of
9884 more than one character */
9886 GET_RE_DEBUG_FLAGS_DECL;
9888 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9892 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9894 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9895 * modifier. The other meaning does not */
9896 p = (RExC_flags & RXf_PMf_EXTENDED)
9897 ? regwhite( pRExC_state, RExC_parse )
9900 /* Disambiguate between \N meaning a named character versus \N meaning
9901 * [^\n]. The former is assumed when it can't be the latter. */
9902 if (*p != '{' || regcurly(p, FALSE)) {
9905 /* no bare \N in a charclass */
9906 if (in_char_class) {
9907 vFAIL("\\N in a character class must be a named character: \\N{...}");
9911 nextchar(pRExC_state);
9912 *node_p = reg_node(pRExC_state, REG_ANY);
9913 *flagp |= HASWIDTH|SIMPLE;
9916 Set_Node_Length(*node_p, 1); /* MJD */
9920 /* Here, we have decided it should be a named character or sequence */
9922 /* The test above made sure that the next real character is a '{', but
9923 * under the /x modifier, it could be separated by space (or a comment and
9924 * \n) and this is not allowed (for consistency with \x{...} and the
9925 * tokenizer handling of \N{NAME}). */
9926 if (*RExC_parse != '{') {
9927 vFAIL("Missing braces on \\N{}");
9930 RExC_parse++; /* Skip past the '{' */
9932 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9933 || ! (endbrace == RExC_parse /* nothing between the {} */
9934 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9935 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9937 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9938 vFAIL("\\N{NAME} must be resolved by the lexer");
9941 if (endbrace == RExC_parse) { /* empty: \N{} */
9944 *node_p = reg_node(pRExC_state,NOTHING);
9946 else if (in_char_class) {
9947 if (SIZE_ONLY && in_char_class) {
9949 RExC_parse++; /* Position after the "}" */
9950 vFAIL("Zero length \\N{}");
9953 ckWARNreg(RExC_parse,
9954 "Ignoring zero length \\N{} in character class");
9962 nextchar(pRExC_state);
9966 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9967 RExC_parse += 2; /* Skip past the 'U+' */
9969 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9971 /* Code points are separated by dots. If none, there is only one code
9972 * point, and is terminated by the brace */
9973 has_multiple_chars = (endchar < endbrace);
9975 if (valuep && (! has_multiple_chars || in_char_class)) {
9976 /* We only pay attention to the first char of
9977 multichar strings being returned in char classes. I kinda wonder
9978 if this makes sense as it does change the behaviour
9979 from earlier versions, OTOH that behaviour was broken
9980 as well. XXX Solution is to recharacterize as
9981 [rest-of-class]|multi1|multi2... */
9983 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9984 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9985 | PERL_SCAN_DISALLOW_PREFIX
9986 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9988 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9990 /* The tokenizer should have guaranteed validity, but it's possible to
9991 * bypass it by using single quoting, so check */
9992 if (length_of_hex == 0
9993 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9995 RExC_parse += length_of_hex; /* Includes all the valid */
9996 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9997 ? UTF8SKIP(RExC_parse)
9999 /* Guard against malformed utf8 */
10000 if (RExC_parse >= endchar) {
10001 RExC_parse = endchar;
10003 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10006 if (in_char_class && has_multiple_chars) {
10008 RExC_parse = endbrace;
10009 vFAIL("\\N{} in character class restricted to one character");
10012 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
10016 RExC_parse = endbrace + 1;
10018 else if (! node_p || ! has_multiple_chars) {
10020 /* Here, the input is legal, but not according to the caller's
10021 * options. We fail without advancing the parse, so that the
10022 * caller can try again */
10028 /* What is done here is to convert this to a sub-pattern of the form
10029 * (?:\x{char1}\x{char2}...)
10030 * and then call reg recursively. That way, it retains its atomicness,
10031 * while not having to worry about special handling that some code
10032 * points may have. toke.c has converted the original Unicode values
10033 * to native, so that we can just pass on the hex values unchanged. We
10034 * do have to set a flag to keep recoding from happening in the
10037 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
10039 char *orig_end = RExC_end;
10042 while (RExC_parse < endbrace) {
10044 /* Convert to notation the rest of the code understands */
10045 sv_catpv(substitute_parse, "\\x{");
10046 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
10047 sv_catpv(substitute_parse, "}");
10049 /* Point to the beginning of the next character in the sequence. */
10050 RExC_parse = endchar + 1;
10051 endchar = RExC_parse + strcspn(RExC_parse, ".}");
10053 sv_catpv(substitute_parse, ")");
10055 RExC_parse = SvPV(substitute_parse, len);
10057 /* Don't allow empty number */
10059 vFAIL("Invalid hexadecimal number in \\N{U+...}");
10061 RExC_end = RExC_parse + len;
10063 /* The values are Unicode, and therefore not subject to recoding */
10064 RExC_override_recoding = 1;
10066 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
10067 if (flags & RESTART_UTF8) {
10068 *flagp = RESTART_UTF8;
10071 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
10074 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10076 RExC_parse = endbrace;
10077 RExC_end = orig_end;
10078 RExC_override_recoding = 0;
10080 nextchar(pRExC_state);
10090 * It returns the code point in utf8 for the value in *encp.
10091 * value: a code value in the source encoding
10092 * encp: a pointer to an Encode object
10094 * If the result from Encode is not a single character,
10095 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
10098 S_reg_recode(pTHX_ const char value, SV **encp)
10101 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
10102 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
10103 const STRLEN newlen = SvCUR(sv);
10104 UV uv = UNICODE_REPLACEMENT;
10106 PERL_ARGS_ASSERT_REG_RECODE;
10110 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
10113 if (!newlen || numlen != newlen) {
10114 uv = UNICODE_REPLACEMENT;
10120 PERL_STATIC_INLINE U8
10121 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
10125 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
10131 op = get_regex_charset(RExC_flags);
10132 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
10133 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
10134 been, so there is no hole */
10137 return op + EXACTF;
10140 PERL_STATIC_INLINE void
10141 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
10143 /* This knows the details about sizing an EXACTish node, setting flags for
10144 * it (by setting <*flagp>, and potentially populating it with a single
10147 * If <len> (the length in bytes) is non-zero, this function assumes that
10148 * the node has already been populated, and just does the sizing. In this
10149 * case <code_point> should be the final code point that has already been
10150 * placed into the node. This value will be ignored except that under some
10151 * circumstances <*flagp> is set based on it.
10153 * If <len> is zero, the function assumes that the node is to contain only
10154 * the single character given by <code_point> and calculates what <len>
10155 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
10156 * additionally will populate the node's STRING with <code_point>, if <len>
10157 * is 0. In both cases <*flagp> is appropriately set
10159 * It knows that under FOLD, the Latin Sharp S and UTF characters above
10160 * 255, must be folded (the former only when the rules indicate it can
10163 bool len_passed_in = cBOOL(len != 0);
10164 U8 character[UTF8_MAXBYTES_CASE+1];
10166 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
10168 if (! len_passed_in) {
10170 if (FOLD && (! LOC || code_point > 255)) {
10171 _to_uni_fold_flags(NATIVE_TO_UNI(code_point),
10174 FOLD_FLAGS_FULL | ((LOC)
10175 ? FOLD_FLAGS_LOCALE
10176 : (ASCII_FOLD_RESTRICTED)
10177 ? FOLD_FLAGS_NOMIX_ASCII
10181 uvchr_to_utf8( character, code_point);
10182 len = UTF8SKIP(character);
10186 || code_point != LATIN_SMALL_LETTER_SHARP_S
10187 || ASCII_FOLD_RESTRICTED
10188 || ! AT_LEAST_UNI_SEMANTICS)
10190 *character = (U8) code_point;
10195 *(character + 1) = 's';
10201 RExC_size += STR_SZ(len);
10204 RExC_emit += STR_SZ(len);
10205 STR_LEN(node) = len;
10206 if (! len_passed_in) {
10207 Copy((char *) character, STRING(node), len, char);
10211 *flagp |= HASWIDTH;
10213 /* A single character node is SIMPLE, except for the special-cased SHARP S
10215 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
10216 && (code_point != LATIN_SMALL_LETTER_SHARP_S
10217 || ! FOLD || ! DEPENDS_SEMANTICS))
10224 - regatom - the lowest level
10226 Try to identify anything special at the start of the pattern. If there
10227 is, then handle it as required. This may involve generating a single regop,
10228 such as for an assertion; or it may involve recursing, such as to
10229 handle a () structure.
10231 If the string doesn't start with something special then we gobble up
10232 as much literal text as we can.
10234 Once we have been able to handle whatever type of thing started the
10235 sequence, we return.
10237 Note: we have to be careful with escapes, as they can be both literal
10238 and special, and in the case of \10 and friends, context determines which.
10240 A summary of the code structure is:
10242 switch (first_byte) {
10243 cases for each special:
10244 handle this special;
10247 switch (2nd byte) {
10248 cases for each unambiguous special:
10249 handle this special;
10251 cases for each ambigous special/literal:
10253 if (special) handle here
10255 default: // unambiguously literal:
10258 default: // is a literal char
10261 create EXACTish node for literal;
10262 while (more input and node isn't full) {
10263 switch (input_byte) {
10264 cases for each special;
10265 make sure parse pointer is set so that the next call to
10266 regatom will see this special first
10267 goto loopdone; // EXACTish node terminated by prev. char
10269 append char to EXACTISH node;
10271 get next input byte;
10275 return the generated node;
10277 Specifically there are two separate switches for handling
10278 escape sequences, with the one for handling literal escapes requiring
10279 a dummy entry for all of the special escapes that are actually handled
10282 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
10284 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10286 Otherwise does not return NULL.
10290 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10293 regnode *ret = NULL;
10295 char *parse_start = RExC_parse;
10299 GET_RE_DEBUG_FLAGS_DECL;
10301 *flagp = WORST; /* Tentatively. */
10303 DEBUG_PARSE("atom");
10305 PERL_ARGS_ASSERT_REGATOM;
10308 switch ((U8)*RExC_parse) {
10310 RExC_seen_zerolen++;
10311 nextchar(pRExC_state);
10312 if (RExC_flags & RXf_PMf_MULTILINE)
10313 ret = reg_node(pRExC_state, MBOL);
10314 else if (RExC_flags & RXf_PMf_SINGLELINE)
10315 ret = reg_node(pRExC_state, SBOL);
10317 ret = reg_node(pRExC_state, BOL);
10318 Set_Node_Length(ret, 1); /* MJD */
10321 nextchar(pRExC_state);
10323 RExC_seen_zerolen++;
10324 if (RExC_flags & RXf_PMf_MULTILINE)
10325 ret = reg_node(pRExC_state, MEOL);
10326 else if (RExC_flags & RXf_PMf_SINGLELINE)
10327 ret = reg_node(pRExC_state, SEOL);
10329 ret = reg_node(pRExC_state, EOL);
10330 Set_Node_Length(ret, 1); /* MJD */
10333 nextchar(pRExC_state);
10334 if (RExC_flags & RXf_PMf_SINGLELINE)
10335 ret = reg_node(pRExC_state, SANY);
10337 ret = reg_node(pRExC_state, REG_ANY);
10338 *flagp |= HASWIDTH|SIMPLE;
10340 Set_Node_Length(ret, 1); /* MJD */
10344 char * const oregcomp_parse = ++RExC_parse;
10345 ret = regclass(pRExC_state, flagp,depth+1,
10346 FALSE, /* means parse the whole char class */
10347 TRUE, /* allow multi-char folds */
10348 FALSE, /* don't silence non-portable warnings. */
10350 if (*RExC_parse != ']') {
10351 RExC_parse = oregcomp_parse;
10352 vFAIL("Unmatched [");
10355 if (*flagp & RESTART_UTF8)
10357 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10360 nextchar(pRExC_state);
10361 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10365 nextchar(pRExC_state);
10366 ret = reg(pRExC_state, 2, &flags,depth+1);
10368 if (flags & TRYAGAIN) {
10369 if (RExC_parse == RExC_end) {
10370 /* Make parent create an empty node if needed. */
10371 *flagp |= TRYAGAIN;
10376 if (flags & RESTART_UTF8) {
10377 *flagp = RESTART_UTF8;
10380 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"", (UV) flags);
10382 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10386 if (flags & TRYAGAIN) {
10387 *flagp |= TRYAGAIN;
10390 vFAIL("Internal urp");
10391 /* Supposed to be caught earlier. */
10394 if (!regcurly(RExC_parse, FALSE)) {
10403 vFAIL("Quantifier follows nothing");
10408 This switch handles escape sequences that resolve to some kind
10409 of special regop and not to literal text. Escape sequnces that
10410 resolve to literal text are handled below in the switch marked
10413 Every entry in this switch *must* have a corresponding entry
10414 in the literal escape switch. However, the opposite is not
10415 required, as the default for this switch is to jump to the
10416 literal text handling code.
10418 switch ((U8)*++RExC_parse) {
10420 /* Special Escapes */
10422 RExC_seen_zerolen++;
10423 ret = reg_node(pRExC_state, SBOL);
10425 goto finish_meta_pat;
10427 ret = reg_node(pRExC_state, GPOS);
10428 RExC_seen |= REG_SEEN_GPOS;
10430 goto finish_meta_pat;
10432 RExC_seen_zerolen++;
10433 ret = reg_node(pRExC_state, KEEPS);
10435 /* XXX:dmq : disabling in-place substitution seems to
10436 * be necessary here to avoid cases of memory corruption, as
10437 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10439 RExC_seen |= REG_SEEN_LOOKBEHIND;
10440 goto finish_meta_pat;
10442 ret = reg_node(pRExC_state, SEOL);
10444 RExC_seen_zerolen++; /* Do not optimize RE away */
10445 goto finish_meta_pat;
10447 ret = reg_node(pRExC_state, EOS);
10449 RExC_seen_zerolen++; /* Do not optimize RE away */
10450 goto finish_meta_pat;
10452 ret = reg_node(pRExC_state, CANY);
10453 RExC_seen |= REG_SEEN_CANY;
10454 *flagp |= HASWIDTH|SIMPLE;
10455 goto finish_meta_pat;
10457 ret = reg_node(pRExC_state, CLUMP);
10458 *flagp |= HASWIDTH;
10459 goto finish_meta_pat;
10465 arg = ANYOF_WORDCHAR;
10469 RExC_seen_zerolen++;
10470 RExC_seen |= REG_SEEN_LOOKBEHIND;
10471 op = BOUND + get_regex_charset(RExC_flags);
10472 if (op > BOUNDA) { /* /aa is same as /a */
10475 ret = reg_node(pRExC_state, op);
10476 FLAGS(ret) = get_regex_charset(RExC_flags);
10478 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10479 ckWARNdep(RExC_parse, "\"\\b{\" is deprecated; use \"\\b\\{\" or \"\\b[{]\" instead");
10481 goto finish_meta_pat;
10483 RExC_seen_zerolen++;
10484 RExC_seen |= REG_SEEN_LOOKBEHIND;
10485 op = NBOUND + get_regex_charset(RExC_flags);
10486 if (op > NBOUNDA) { /* /aa is same as /a */
10489 ret = reg_node(pRExC_state, op);
10490 FLAGS(ret) = get_regex_charset(RExC_flags);
10492 if (! SIZE_ONLY && (U8) *(RExC_parse + 1) == '{') {
10493 ckWARNdep(RExC_parse, "\"\\B{\" is deprecated; use \"\\B\\{\" or \"\\B[{]\" instead");
10495 goto finish_meta_pat;
10505 ret = reg_node(pRExC_state, LNBREAK);
10506 *flagp |= HASWIDTH|SIMPLE;
10507 goto finish_meta_pat;
10515 goto join_posix_op_known;
10521 arg = ANYOF_VERTWS;
10523 goto join_posix_op_known;
10533 op = POSIXD + get_regex_charset(RExC_flags);
10534 if (op > POSIXA) { /* /aa is same as /a */
10538 join_posix_op_known:
10541 op += NPOSIXD - POSIXD;
10544 ret = reg_node(pRExC_state, op);
10546 FLAGS(ret) = namedclass_to_classnum(arg);
10549 *flagp |= HASWIDTH|SIMPLE;
10553 nextchar(pRExC_state);
10554 Set_Node_Length(ret, 2); /* MJD */
10560 char* parse_start = RExC_parse - 2;
10565 ret = regclass(pRExC_state, flagp,depth+1,
10566 TRUE, /* means just parse this element */
10567 FALSE, /* don't allow multi-char folds */
10568 FALSE, /* don't silence non-portable warnings.
10569 It would be a bug if these returned
10572 /* regclass() can only return RESTART_UTF8 if multi-char folds
10575 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
10580 Set_Node_Offset(ret, parse_start + 2);
10581 Set_Node_Cur_Length(ret, parse_start);
10582 nextchar(pRExC_state);
10586 /* Handle \N and \N{NAME} with multiple code points here and not
10587 * below because it can be multicharacter. join_exact() will join
10588 * them up later on. Also this makes sure that things like
10589 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10590 * The options to the grok function call causes it to fail if the
10591 * sequence is just a single code point. We then go treat it as
10592 * just another character in the current EXACT node, and hence it
10593 * gets uniform treatment with all the other characters. The
10594 * special treatment for quantifiers is not needed for such single
10595 * character sequences */
10597 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE,
10598 FALSE /* not strict */ )) {
10599 if (*flagp & RESTART_UTF8)
10605 case 'k': /* Handle \k<NAME> and \k'NAME' */
10608 char ch= RExC_parse[1];
10609 if (ch != '<' && ch != '\'' && ch != '{') {
10611 vFAIL2("Sequence %.2s... not terminated",parse_start);
10613 /* this pretty much dupes the code for (?P=...) in reg(), if
10614 you change this make sure you change that */
10615 char* name_start = (RExC_parse += 2);
10617 SV *sv_dat = reg_scan_name(pRExC_state,
10618 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10619 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10620 if (RExC_parse == name_start || *RExC_parse != ch)
10621 vFAIL2("Sequence %.3s... not terminated",parse_start);
10624 num = add_data( pRExC_state, 1, "S" );
10625 RExC_rxi->data->data[num]=(void*)sv_dat;
10626 SvREFCNT_inc_simple_void(sv_dat);
10630 ret = reganode(pRExC_state,
10633 : (ASCII_FOLD_RESTRICTED)
10635 : (AT_LEAST_UNI_SEMANTICS)
10641 *flagp |= HASWIDTH;
10643 /* override incorrect value set in reganode MJD */
10644 Set_Node_Offset(ret, parse_start+1);
10645 Set_Node_Cur_Length(ret, parse_start);
10646 nextchar(pRExC_state);
10652 case '1': case '2': case '3': case '4':
10653 case '5': case '6': case '7': case '8': case '9':
10656 bool isg = *RExC_parse == 'g';
10661 if (*RExC_parse == '{') {
10665 if (*RExC_parse == '-') {
10669 if (hasbrace && !isDIGIT(*RExC_parse)) {
10670 if (isrel) RExC_parse--;
10672 goto parse_named_seq;
10674 num = atoi(RExC_parse);
10675 if (isg && num == 0) {
10676 if (*RExC_parse == '0') {
10677 vFAIL("Reference to invalid group 0");
10680 vFAIL("Unterminated \\g... pattern");
10684 num = RExC_npar - num;
10686 vFAIL("Reference to nonexistent or unclosed group");
10688 if (!isg && num > 9 && num >= RExC_npar && *RExC_parse != '8' && *RExC_parse != '9')
10689 /* Probably a character specified in octal, e.g. \35 */
10692 #ifdef RE_TRACK_PATTERN_OFFSETS
10693 char * const parse_start = RExC_parse - 1; /* MJD */
10695 while (isDIGIT(*RExC_parse))
10698 if (*RExC_parse != '}')
10699 vFAIL("Unterminated \\g{...} pattern");
10703 if (num > (I32)RExC_rx->nparens)
10704 vFAIL("Reference to nonexistent group");
10707 ret = reganode(pRExC_state,
10710 : (ASCII_FOLD_RESTRICTED)
10712 : (AT_LEAST_UNI_SEMANTICS)
10718 *flagp |= HASWIDTH;
10720 /* override incorrect value set in reganode MJD */
10721 Set_Node_Offset(ret, parse_start+1);
10722 Set_Node_Cur_Length(ret, parse_start);
10724 nextchar(pRExC_state);
10729 if (RExC_parse >= RExC_end)
10730 FAIL("Trailing \\");
10733 /* Do not generate "unrecognized" warnings here, we fall
10734 back into the quick-grab loop below */
10741 if (RExC_flags & RXf_PMf_EXTENDED) {
10742 if ( reg_skipcomment( pRExC_state ) )
10749 parse_start = RExC_parse - 1;
10758 #define MAX_NODE_STRING_SIZE 127
10759 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10761 U8 upper_parse = MAX_NODE_STRING_SIZE;
10763 U8 node_type = compute_EXACTish(pRExC_state);
10764 bool next_is_quantifier;
10765 char * oldp = NULL;
10767 /* We can convert EXACTF nodes to EXACTFU if they contain only
10768 * characters that match identically regardless of the target
10769 * string's UTF8ness. The reason to do this is that EXACTF is not
10770 * trie-able, EXACTFU is. (We don't need to figure this out until
10772 bool maybe_exactfu = node_type == EXACTF && PASS2;
10774 /* If a folding node contains only code points that don't
10775 * participate in folds, it can be changed into an EXACT node,
10776 * which allows the optimizer more things to look for */
10779 ret = reg_node(pRExC_state, node_type);
10781 /* In pass1, folded, we use a temporary buffer instead of the
10782 * actual node, as the node doesn't exist yet */
10783 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10789 /* We do the EXACTFish to EXACT node only if folding, and not if in
10790 * locale, as whether a character folds or not isn't known until
10791 * runtime. (And we don't need to figure this out until pass 2) */
10792 maybe_exact = FOLD && ! LOC && PASS2;
10794 /* XXX The node can hold up to 255 bytes, yet this only goes to
10795 * 127. I (khw) do not know why. Keeping it somewhat less than
10796 * 255 allows us to not have to worry about overflow due to
10797 * converting to utf8 and fold expansion, but that value is
10798 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10799 * split up by this limit into a single one using the real max of
10800 * 255. Even at 127, this breaks under rare circumstances. If
10801 * folding, we do not want to split a node at a character that is a
10802 * non-final in a multi-char fold, as an input string could just
10803 * happen to want to match across the node boundary. The join
10804 * would solve that problem if the join actually happens. But a
10805 * series of more than two nodes in a row each of 127 would cause
10806 * the first join to succeed to get to 254, but then there wouldn't
10807 * be room for the next one, which could at be one of those split
10808 * multi-char folds. I don't know of any fool-proof solution. One
10809 * could back off to end with only a code point that isn't such a
10810 * non-final, but it is possible for there not to be any in the
10812 for (p = RExC_parse - 1;
10813 len < upper_parse && p < RExC_end;
10818 if (RExC_flags & RXf_PMf_EXTENDED)
10819 p = regwhite( pRExC_state, p );
10830 /* Literal Escapes Switch
10832 This switch is meant to handle escape sequences that
10833 resolve to a literal character.
10835 Every escape sequence that represents something
10836 else, like an assertion or a char class, is handled
10837 in the switch marked 'Special Escapes' above in this
10838 routine, but also has an entry here as anything that
10839 isn't explicitly mentioned here will be treated as
10840 an unescaped equivalent literal.
10843 switch ((U8)*++p) {
10844 /* These are all the special escapes. */
10845 case 'A': /* Start assertion */
10846 case 'b': case 'B': /* Word-boundary assertion*/
10847 case 'C': /* Single char !DANGEROUS! */
10848 case 'd': case 'D': /* digit class */
10849 case 'g': case 'G': /* generic-backref, pos assertion */
10850 case 'h': case 'H': /* HORIZWS */
10851 case 'k': case 'K': /* named backref, keep marker */
10852 case 'p': case 'P': /* Unicode property */
10853 case 'R': /* LNBREAK */
10854 case 's': case 'S': /* space class */
10855 case 'v': case 'V': /* VERTWS */
10856 case 'w': case 'W': /* word class */
10857 case 'X': /* eXtended Unicode "combining character sequence" */
10858 case 'z': case 'Z': /* End of line/string assertion */
10862 /* Anything after here is an escape that resolves to a
10863 literal. (Except digits, which may or may not)
10869 case 'N': /* Handle a single-code point named character. */
10870 /* The options cause it to fail if a multiple code
10871 * point sequence. Handle those in the switch() above
10873 RExC_parse = p + 1;
10874 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10875 flagp, depth, FALSE,
10876 FALSE /* not strict */ ))
10878 if (*flagp & RESTART_UTF8)
10879 FAIL("panic: grok_bslash_N set RESTART_UTF8");
10880 RExC_parse = p = oldp;
10884 if (ender > 0xff) {
10901 ender = ASCII_TO_NATIVE('\033');
10905 ender = ASCII_TO_NATIVE('\007');
10911 const char* error_msg;
10913 bool valid = grok_bslash_o(&p,
10916 TRUE, /* out warnings */
10917 FALSE, /* not strict */
10918 TRUE, /* Output warnings
10923 RExC_parse = p; /* going to die anyway; point
10924 to exact spot of failure */
10928 if (PL_encoding && ender < 0x100) {
10929 goto recode_encoding;
10931 if (ender > 0xff) {
10938 UV result = UV_MAX; /* initialize to erroneous
10940 const char* error_msg;
10942 bool valid = grok_bslash_x(&p,
10945 TRUE, /* out warnings */
10946 FALSE, /* not strict */
10947 TRUE, /* Output warnings
10952 RExC_parse = p; /* going to die anyway; point
10953 to exact spot of failure */
10958 if (PL_encoding && ender < 0x100) {
10959 goto recode_encoding;
10961 if (ender > 0xff) {
10968 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10970 case '8': case '9': /* must be a backreference */
10973 case '1': case '2': case '3':case '4':
10974 case '5': case '6': case '7':
10975 /* When we parse backslash escapes there is ambiguity between
10976 * backreferences and octal escapes. Any escape from \1 - \9 is
10977 * a backreference, any multi-digit escape which does not start with
10978 * 0 and which when evaluated as decimal could refer to an already
10979 * parsed capture buffer is a backslash. Anything else is octal.
10981 * Note this implies that \118 could be interpreted as 118 OR as
10982 * "\11" . "8" depending on whether there were 118 capture buffers
10983 * defined already in the pattern.
10985 if ( !isDIGIT(p[1]) || atoi(p) <= RExC_npar )
10986 { /* Not to be treated as an octal constant, go
10993 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10995 ender = grok_oct(p, &numlen, &flags, NULL);
10996 if (ender > 0xff) {
11000 if (SIZE_ONLY /* like \08, \178 */
11003 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
11005 reg_warn_non_literal_string(
11007 form_short_octal_warning(p, numlen));
11010 if (PL_encoding && ender < 0x100)
11011 goto recode_encoding;
11014 if (! RExC_override_recoding) {
11015 SV* enc = PL_encoding;
11016 ender = reg_recode((const char)(U8)ender, &enc);
11017 if (!enc && SIZE_ONLY)
11018 ckWARNreg(p, "Invalid escape in the specified encoding");
11024 FAIL("Trailing \\");
11027 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
11028 /* Include any { following the alpha to emphasize
11029 * that it could be part of an escape at some point
11031 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
11032 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
11034 goto normal_default;
11035 } /* End of switch on '\' */
11037 default: /* A literal character */
11040 && RExC_flags & RXf_PMf_EXTENDED
11041 && ckWARN_d(WARN_DEPRECATED)
11042 && is_PATWS_non_low(p, UTF))
11044 vWARN_dep(p + ((UTF) ? UTF8SKIP(p) : 1),
11045 "Escape literal pattern white space under /x");
11049 if (UTF8_IS_START(*p) && UTF) {
11051 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
11052 &numlen, UTF8_ALLOW_DEFAULT);
11058 } /* End of switch on the literal */
11060 /* Here, have looked at the literal character and <ender>
11061 * contains its ordinal, <p> points to the character after it
11064 if ( RExC_flags & RXf_PMf_EXTENDED)
11065 p = regwhite( pRExC_state, p );
11067 /* If the next thing is a quantifier, it applies to this
11068 * character only, which means that this character has to be in
11069 * its own node and can't just be appended to the string in an
11070 * existing node, so if there are already other characters in
11071 * the node, close the node with just them, and set up to do
11072 * this character again next time through, when it will be the
11073 * only thing in its new node */
11074 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
11082 const STRLEN unilen = reguni(pRExC_state, ender, s);
11088 /* The loop increments <len> each time, as all but this
11089 * path (and one other) through it add a single byte to
11090 * the EXACTish node. But this one has changed len to
11091 * be the correct final value, so subtract one to
11092 * cancel out the increment that follows */
11096 REGC((char)ender, s++);
11101 /* See comments for join_exact() as to why we fold this
11102 * non-UTF at compile time */
11103 || (node_type == EXACTFU
11104 && ender == LATIN_SMALL_LETTER_SHARP_S)))
11106 if (IS_IN_SOME_FOLD_L1(ender)) {
11107 maybe_exact = FALSE;
11109 /* See if the character's fold differs between /d and
11110 * /u. This includes the multi-char fold SHARP S to
11113 && (PL_fold[ender] != PL_fold_latin1[ender]
11114 || ender == LATIN_SMALL_LETTER_SHARP_S
11116 && isARG2_lower_or_UPPER_ARG1('s', ender)
11117 && isARG2_lower_or_UPPER_ARG1('s', *(s-1)))))
11119 maybe_exactfu = FALSE;
11122 *(s++) = (char) ender;
11126 /* Prime the casefolded buffer. Locale rules, which apply
11127 * only to code points < 256, aren't known until execution,
11128 * so for them, just output the original character using
11129 * utf8. If we start to fold non-UTF patterns, be sure to
11130 * update join_exact() */
11131 if (LOC && ender < 256) {
11132 if (UNI_IS_INVARIANT(ender)) {
11136 *s = UTF8_TWO_BYTE_HI(ender);
11137 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
11142 UV folded = _to_uni_fold_flags(
11147 | ((LOC) ? FOLD_FLAGS_LOCALE
11148 : (ASCII_FOLD_RESTRICTED)
11149 ? FOLD_FLAGS_NOMIX_ASCII
11153 /* If this node only contains non-folding code points
11154 * so far, see if this new one is also non-folding */
11156 if (folded != ender) {
11157 maybe_exact = FALSE;
11160 /* Here the fold is the original; we have
11161 * to check further to see if anything
11163 if (! PL_utf8_foldable) {
11164 SV* swash = swash_init("utf8",
11166 &PL_sv_undef, 1, 0);
11168 _get_swash_invlist(swash);
11169 SvREFCNT_dec_NN(swash);
11171 if (_invlist_contains_cp(PL_utf8_foldable,
11174 maybe_exact = FALSE;
11182 /* The loop increments <len> each time, as all but this
11183 * path (and one other) through it add a single byte to the
11184 * EXACTish node. But this one has changed len to be the
11185 * correct final value, so subtract one to cancel out the
11186 * increment that follows */
11187 len += foldlen - 1;
11190 if (next_is_quantifier) {
11192 /* Here, the next input is a quantifier, and to get here,
11193 * the current character is the only one in the node.
11194 * Also, here <len> doesn't include the final byte for this
11200 } /* End of loop through literal characters */
11202 /* Here we have either exhausted the input or ran out of room in
11203 * the node. (If we encountered a character that can't be in the
11204 * node, transfer is made directly to <loopdone>, and so we
11205 * wouldn't have fallen off the end of the loop.) In the latter
11206 * case, we artificially have to split the node into two, because
11207 * we just don't have enough space to hold everything. This
11208 * creates a problem if the final character participates in a
11209 * multi-character fold in the non-final position, as a match that
11210 * should have occurred won't, due to the way nodes are matched,
11211 * and our artificial boundary. So back off until we find a non-
11212 * problematic character -- one that isn't at the beginning or
11213 * middle of such a fold. (Either it doesn't participate in any
11214 * folds, or appears only in the final position of all the folds it
11215 * does participate in.) A better solution with far fewer false
11216 * positives, and that would fill the nodes more completely, would
11217 * be to actually have available all the multi-character folds to
11218 * test against, and to back-off only far enough to be sure that
11219 * this node isn't ending with a partial one. <upper_parse> is set
11220 * further below (if we need to reparse the node) to include just
11221 * up through that final non-problematic character that this code
11222 * identifies, so when it is set to less than the full node, we can
11223 * skip the rest of this */
11224 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
11226 const STRLEN full_len = len;
11228 assert(len >= MAX_NODE_STRING_SIZE);
11230 /* Here, <s> points to the final byte of the final character.
11231 * Look backwards through the string until find a non-
11232 * problematic character */
11236 /* These two have no multi-char folds to non-UTF characters
11238 if (ASCII_FOLD_RESTRICTED || LOC) {
11242 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
11246 if (! PL_NonL1NonFinalFold) {
11247 PL_NonL1NonFinalFold = _new_invlist_C_array(
11248 NonL1_Perl_Non_Final_Folds_invlist);
11251 /* Point to the first byte of the final character */
11252 s = (char *) utf8_hop((U8 *) s, -1);
11254 while (s >= s0) { /* Search backwards until find
11255 non-problematic char */
11256 if (UTF8_IS_INVARIANT(*s)) {
11258 /* There are no ascii characters that participate
11259 * in multi-char folds under /aa. In EBCDIC, the
11260 * non-ascii invariants are all control characters,
11261 * so don't ever participate in any folds. */
11262 if (ASCII_FOLD_RESTRICTED
11263 || ! IS_NON_FINAL_FOLD(*s))
11268 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
11270 /* No Latin1 characters participate in multi-char
11271 * folds under /l */
11273 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
11279 else if (! _invlist_contains_cp(
11280 PL_NonL1NonFinalFold,
11281 valid_utf8_to_uvchr((U8 *) s, NULL)))
11286 /* Here, the current character is problematic in that
11287 * it does occur in the non-final position of some
11288 * fold, so try the character before it, but have to
11289 * special case the very first byte in the string, so
11290 * we don't read outside the string */
11291 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
11292 } /* End of loop backwards through the string */
11294 /* If there were only problematic characters in the string,
11295 * <s> will point to before s0, in which case the length
11296 * should be 0, otherwise include the length of the
11297 * non-problematic character just found */
11298 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
11301 /* Here, have found the final character, if any, that is
11302 * non-problematic as far as ending the node without splitting
11303 * it across a potential multi-char fold. <len> contains the
11304 * number of bytes in the node up-to and including that
11305 * character, or is 0 if there is no such character, meaning
11306 * the whole node contains only problematic characters. In
11307 * this case, give up and just take the node as-is. We can't
11312 /* If the node ends in an 's' we make sure it stays EXACTF,
11313 * as if it turns into an EXACTFU, it could later get
11314 * joined with another 's' that would then wrongly match
11316 if (maybe_exactfu && isARG2_lower_or_UPPER_ARG1('s', ender))
11318 maybe_exactfu = FALSE;
11322 /* Here, the node does contain some characters that aren't
11323 * problematic. If one such is the final character in the
11324 * node, we are done */
11325 if (len == full_len) {
11328 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
11330 /* If the final character is problematic, but the
11331 * penultimate is not, back-off that last character to
11332 * later start a new node with it */
11337 /* Here, the final non-problematic character is earlier
11338 * in the input than the penultimate character. What we do
11339 * is reparse from the beginning, going up only as far as
11340 * this final ok one, thus guaranteeing that the node ends
11341 * in an acceptable character. The reason we reparse is
11342 * that we know how far in the character is, but we don't
11343 * know how to correlate its position with the input parse.
11344 * An alternate implementation would be to build that
11345 * correlation as we go along during the original parse,
11346 * but that would entail extra work for every node, whereas
11347 * this code gets executed only when the string is too
11348 * large for the node, and the final two characters are
11349 * problematic, an infrequent occurrence. Yet another
11350 * possible strategy would be to save the tail of the
11351 * string, and the next time regatom is called, initialize
11352 * with that. The problem with this is that unless you
11353 * back off one more character, you won't be guaranteed
11354 * regatom will get called again, unless regbranch,
11355 * regpiece ... are also changed. If you do back off that
11356 * extra character, so that there is input guaranteed to
11357 * force calling regatom, you can't handle the case where
11358 * just the first character in the node is acceptable. I
11359 * (khw) decided to try this method which doesn't have that
11360 * pitfall; if performance issues are found, we can do a
11361 * combination of the current approach plus that one */
11367 } /* End of verifying node ends with an appropriate char */
11369 loopdone: /* Jumped to when encounters something that shouldn't be in
11372 /* I (khw) don't know if you can get here with zero length, but the
11373 * old code handled this situation by creating a zero-length EXACT
11374 * node. Might as well be NOTHING instead */
11380 /* If 'maybe_exact' is still set here, means there are no
11381 * code points in the node that participate in folds;
11382 * similarly for 'maybe_exactfu' and code points that match
11383 * differently depending on UTF8ness of the target string
11388 else if (maybe_exactfu) {
11392 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11395 RExC_parse = p - 1;
11396 Set_Node_Cur_Length(ret, parse_start);
11397 nextchar(pRExC_state);
11399 /* len is STRLEN which is unsigned, need to copy to signed */
11402 vFAIL("Internal disaster");
11405 } /* End of label 'defchar:' */
11407 } /* End of giant switch on input character */
11413 S_regwhite( RExC_state_t *pRExC_state, char *p )
11415 const char *e = RExC_end;
11417 PERL_ARGS_ASSERT_REGWHITE;
11422 else if (*p == '#') {
11425 if (*p++ == '\n') {
11431 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11440 S_regpatws( RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
11442 /* Returns the next non-pattern-white space, non-comment character (the
11443 * latter only if 'recognize_comment is true) in the string p, which is
11444 * ended by RExC_end. If there is no line break ending a comment,
11445 * RExC_seen has added the REG_SEEN_RUN_ON_COMMENT flag; */
11446 const char *e = RExC_end;
11448 PERL_ARGS_ASSERT_REGPATWS;
11452 if ((len = is_PATWS_safe(p, e, UTF))) {
11455 else if (recognize_comment && *p == '#') {
11459 if (is_LNBREAK_safe(p, e, UTF)) {
11465 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11473 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11474 Character classes ([:foo:]) can also be negated ([:^foo:]).
11475 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11476 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11477 but trigger failures because they are currently unimplemented. */
11479 #define POSIXCC_DONE(c) ((c) == ':')
11480 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11481 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11483 PERL_STATIC_INLINE I32
11484 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
11487 I32 namedclass = OOB_NAMEDCLASS;
11489 PERL_ARGS_ASSERT_REGPPOSIXCC;
11491 if (value == '[' && RExC_parse + 1 < RExC_end &&
11492 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11493 POSIXCC(UCHARAT(RExC_parse)))
11495 const char c = UCHARAT(RExC_parse);
11496 char* const s = RExC_parse++;
11498 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11500 if (RExC_parse == RExC_end) {
11503 /* Try to give a better location for the error (than the end of
11504 * the string) by looking for the matching ']' */
11506 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11509 vFAIL2("Unmatched '%c' in POSIX class", c);
11511 /* Grandfather lone [:, [=, [. */
11515 const char* const t = RExC_parse++; /* skip over the c */
11518 if (UCHARAT(RExC_parse) == ']') {
11519 const char *posixcc = s + 1;
11520 RExC_parse++; /* skip over the ending ] */
11523 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11524 const I32 skip = t - posixcc;
11526 /* Initially switch on the length of the name. */
11529 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
11530 this is the Perl \w
11532 namedclass = ANYOF_WORDCHAR;
11535 /* Names all of length 5. */
11536 /* alnum alpha ascii blank cntrl digit graph lower
11537 print punct space upper */
11538 /* Offset 4 gives the best switch position. */
11539 switch (posixcc[4]) {
11541 if (memEQ(posixcc, "alph", 4)) /* alpha */
11542 namedclass = ANYOF_ALPHA;
11545 if (memEQ(posixcc, "spac", 4)) /* space */
11546 namedclass = ANYOF_PSXSPC;
11549 if (memEQ(posixcc, "grap", 4)) /* graph */
11550 namedclass = ANYOF_GRAPH;
11553 if (memEQ(posixcc, "asci", 4)) /* ascii */
11554 namedclass = ANYOF_ASCII;
11557 if (memEQ(posixcc, "blan", 4)) /* blank */
11558 namedclass = ANYOF_BLANK;
11561 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11562 namedclass = ANYOF_CNTRL;
11565 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11566 namedclass = ANYOF_ALPHANUMERIC;
11569 if (memEQ(posixcc, "lowe", 4)) /* lower */
11570 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
11571 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11572 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
11575 if (memEQ(posixcc, "digi", 4)) /* digit */
11576 namedclass = ANYOF_DIGIT;
11577 else if (memEQ(posixcc, "prin", 4)) /* print */
11578 namedclass = ANYOF_PRINT;
11579 else if (memEQ(posixcc, "punc", 4)) /* punct */
11580 namedclass = ANYOF_PUNCT;
11585 if (memEQ(posixcc, "xdigit", 6))
11586 namedclass = ANYOF_XDIGIT;
11590 if (namedclass == OOB_NAMEDCLASS)
11591 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11594 /* The #defines are structured so each complement is +1 to
11595 * the normal one */
11599 assert (posixcc[skip] == ':');
11600 assert (posixcc[skip+1] == ']');
11601 } else if (!SIZE_ONLY) {
11602 /* [[=foo=]] and [[.foo.]] are still future. */
11604 /* adjust RExC_parse so the warning shows after
11605 the class closes */
11606 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11608 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11611 /* Maternal grandfather:
11612 * "[:" ending in ":" but not in ":]" */
11614 vFAIL("Unmatched '[' in POSIX class");
11617 /* Grandfather lone [:, [=, [. */
11627 S_could_it_be_a_POSIX_class(pTHX_ RExC_state_t *pRExC_state)
11629 /* This applies some heuristics at the current parse position (which should
11630 * be at a '[') to see if what follows might be intended to be a [:posix:]
11631 * class. It returns true if it really is a posix class, of course, but it
11632 * also can return true if it thinks that what was intended was a posix
11633 * class that didn't quite make it.
11635 * It will return true for
11637 * [:alphanumerics] (as long as the ] isn't followed immediately by a
11638 * ')' indicating the end of the (?[
11639 * [:any garbage including %^&$ punctuation:]
11641 * This is designed to be called only from S_handle_regex_sets; it could be
11642 * easily adapted to be called from the spot at the beginning of regclass()
11643 * that checks to see in a normal bracketed class if the surrounding []
11644 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
11645 * change long-standing behavior, so I (khw) didn't do that */
11646 char* p = RExC_parse + 1;
11647 char first_char = *p;
11649 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
11651 assert(*(p - 1) == '[');
11653 if (! POSIXCC(first_char)) {
11658 while (p < RExC_end && isWORDCHAR(*p)) p++;
11660 if (p >= RExC_end) {
11664 if (p - RExC_parse > 2 /* Got at least 1 word character */
11665 && (*p == first_char
11666 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
11671 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
11674 && p - RExC_parse > 2 /* [:] evaluates to colon;
11675 [::] is a bad posix class. */
11676 && first_char == *(p - 1));
11680 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist, I32 *flagp, U32 depth,
11681 char * const oregcomp_parse)
11683 /* Handle the (?[...]) construct to do set operations */
11686 UV start, end; /* End points of code point ranges */
11688 char *save_end, *save_parse;
11693 const bool save_fold = FOLD;
11695 GET_RE_DEBUG_FLAGS_DECL;
11697 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
11700 vFAIL("(?[...]) not valid in locale");
11702 RExC_uni_semantics = 1;
11704 /* This will return only an ANYOF regnode, or (unlikely) something smaller
11705 * (such as EXACT). Thus we can skip most everything if just sizing. We
11706 * call regclass to handle '[]' so as to not have to reinvent its parsing
11707 * rules here (throwing away the size it computes each time). And, we exit
11708 * upon an unescaped ']' that isn't one ending a regclass. To do both
11709 * these things, we need to realize that something preceded by a backslash
11710 * is escaped, so we have to keep track of backslashes */
11712 UV depth = 0; /* how many nested (?[...]) constructs */
11714 Perl_ck_warner_d(aTHX_
11715 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
11716 "The regex_sets feature is experimental" REPORT_LOCATION,
11717 (int) (RExC_parse - RExC_precomp) , RExC_precomp, RExC_parse);
11719 while (RExC_parse < RExC_end) {
11720 SV* current = NULL;
11721 RExC_parse = regpatws(pRExC_state, RExC_parse,
11722 TRUE); /* means recognize comments */
11723 switch (*RExC_parse) {
11725 if (RExC_parse[1] == '[') depth++, RExC_parse++;
11730 /* Skip the next byte (which could cause us to end up in
11731 * the middle of a UTF-8 character, but since none of those
11732 * are confusable with anything we currently handle in this
11733 * switch (invariants all), it's safe. We'll just hit the
11734 * default: case next time and keep on incrementing until
11735 * we find one of the invariants we do handle. */
11740 /* If this looks like it is a [:posix:] class, leave the
11741 * parse pointer at the '[' to fool regclass() into
11742 * thinking it is part of a '[[:posix:]]'. That function
11743 * will use strict checking to force a syntax error if it
11744 * doesn't work out to a legitimate class */
11745 bool is_posix_class
11746 = could_it_be_a_POSIX_class(pRExC_state);
11747 if (! is_posix_class) {
11751 /* regclass() can only return RESTART_UTF8 if multi-char
11752 folds are allowed. */
11753 if (!regclass(pRExC_state, flagp,depth+1,
11754 is_posix_class, /* parse the whole char
11755 class only if not a
11757 FALSE, /* don't allow multi-char folds */
11758 TRUE, /* silence non-portable warnings. */
11760 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11763 /* function call leaves parse pointing to the ']', except
11764 * if we faked it */
11765 if (is_posix_class) {
11769 SvREFCNT_dec(current); /* In case it returned something */
11774 if (depth--) break;
11776 if (RExC_parse < RExC_end
11777 && *RExC_parse == ')')
11779 node = reganode(pRExC_state, ANYOF, 0);
11780 RExC_size += ANYOF_SKIP;
11781 nextchar(pRExC_state);
11782 Set_Node_Length(node,
11783 RExC_parse - oregcomp_parse + 1); /* MJD */
11792 FAIL("Syntax error in (?[...])");
11795 /* Pass 2 only after this. Everything in this construct is a
11796 * metacharacter. Operands begin with either a '\' (for an escape
11797 * sequence), or a '[' for a bracketed character class. Any other
11798 * character should be an operator, or parenthesis for grouping. Both
11799 * types of operands are handled by calling regclass() to parse them. It
11800 * is called with a parameter to indicate to return the computed inversion
11801 * list. The parsing here is implemented via a stack. Each entry on the
11802 * stack is a single character representing one of the operators, or the
11803 * '('; or else a pointer to an operand inversion list. */
11805 #define IS_OPERAND(a) (! SvIOK(a))
11807 /* The stack starts empty. It is a syntax error if the first thing parsed
11808 * is a binary operator; everything else is pushed on the stack. When an
11809 * operand is parsed, the top of the stack is examined. If it is a binary
11810 * operator, the item before it should be an operand, and both are replaced
11811 * by the result of doing that operation on the new operand and the one on
11812 * the stack. Thus a sequence of binary operands is reduced to a single
11813 * one before the next one is parsed.
11815 * A unary operator may immediately follow a binary in the input, for
11818 * When an operand is parsed and the top of the stack is a unary operator,
11819 * the operation is performed, and then the stack is rechecked to see if
11820 * this new operand is part of a binary operation; if so, it is handled as
11823 * A '(' is simply pushed on the stack; it is valid only if the stack is
11824 * empty, or the top element of the stack is an operator or another '('
11825 * (for which the parenthesized expression will become an operand). By the
11826 * time the corresponding ')' is parsed everything in between should have
11827 * been parsed and evaluated to a single operand (or else is a syntax
11828 * error), and is handled as a regular operand */
11830 sv_2mortal((SV *)(stack = newAV()));
11832 while (RExC_parse < RExC_end) {
11833 I32 top_index = av_tindex(stack);
11835 SV* current = NULL;
11837 /* Skip white space */
11838 RExC_parse = regpatws(pRExC_state, RExC_parse,
11839 TRUE); /* means recognize comments */
11840 if (RExC_parse >= RExC_end) {
11841 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
11843 if ((curchar = UCHARAT(RExC_parse)) == ']') {
11850 if (av_tindex(stack) >= 0 /* This makes sure that we can
11851 safely subtract 1 from
11852 RExC_parse in the next clause.
11853 If we have something on the
11854 stack, we have parsed something
11856 && UCHARAT(RExC_parse - 1) == '('
11857 && RExC_parse < RExC_end)
11859 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
11860 * This happens when we have some thing like
11862 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
11864 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
11866 * Here we would be handling the interpolated
11867 * '$thai_or_lao'. We handle this by a recursive call to
11868 * ourselves which returns the inversion list the
11869 * interpolated expression evaluates to. We use the flags
11870 * from the interpolated pattern. */
11871 U32 save_flags = RExC_flags;
11872 const char * const save_parse = ++RExC_parse;
11874 parse_lparen_question_flags(pRExC_state);
11876 if (RExC_parse == save_parse /* Makes sure there was at
11877 least one flag (or this
11878 embedding wasn't compiled)
11880 || RExC_parse >= RExC_end - 4
11881 || UCHARAT(RExC_parse) != ':'
11882 || UCHARAT(++RExC_parse) != '('
11883 || UCHARAT(++RExC_parse) != '?'
11884 || UCHARAT(++RExC_parse) != '[')
11887 /* In combination with the above, this moves the
11888 * pointer to the point just after the first erroneous
11889 * character (or if there are no flags, to where they
11890 * should have been) */
11891 if (RExC_parse >= RExC_end - 4) {
11892 RExC_parse = RExC_end;
11894 else if (RExC_parse != save_parse) {
11895 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11897 vFAIL("Expecting '(?flags:(?[...'");
11900 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
11901 depth+1, oregcomp_parse);
11903 /* Here, 'current' contains the embedded expression's
11904 * inversion list, and RExC_parse points to the trailing
11905 * ']'; the next character should be the ')' which will be
11906 * paired with the '(' that has been put on the stack, so
11907 * the whole embedded expression reduces to '(operand)' */
11910 RExC_flags = save_flags;
11911 goto handle_operand;
11916 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
11917 vFAIL("Unexpected character");
11920 /* regclass() can only return RESTART_UTF8 if multi-char
11921 folds are allowed. */
11922 if (!regclass(pRExC_state, flagp,depth+1,
11923 TRUE, /* means parse just the next thing */
11924 FALSE, /* don't allow multi-char folds */
11925 FALSE, /* don't silence non-portable warnings. */
11927 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11929 /* regclass() will return with parsing just the \ sequence,
11930 * leaving the parse pointer at the next thing to parse */
11932 goto handle_operand;
11934 case '[': /* Is a bracketed character class */
11936 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
11938 if (! is_posix_class) {
11942 /* regclass() can only return RESTART_UTF8 if multi-char
11943 folds are allowed. */
11944 if(!regclass(pRExC_state, flagp,depth+1,
11945 is_posix_class, /* parse the whole char class
11946 only if not a posix class */
11947 FALSE, /* don't allow multi-char folds */
11948 FALSE, /* don't silence non-portable warnings. */
11950 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf"",
11952 /* function call leaves parse pointing to the ']', except if we
11954 if (is_posix_class) {
11958 goto handle_operand;
11967 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
11968 || ! IS_OPERAND(*top_ptr))
11971 vFAIL2("Unexpected binary operator '%c' with no preceding operand", curchar);
11973 av_push(stack, newSVuv(curchar));
11977 av_push(stack, newSVuv(curchar));
11981 if (top_index >= 0) {
11982 top_ptr = av_fetch(stack, top_index, FALSE);
11984 if (IS_OPERAND(*top_ptr)) {
11986 vFAIL("Unexpected '(' with no preceding operator");
11989 av_push(stack, newSVuv(curchar));
11996 || ! (current = av_pop(stack))
11997 || ! IS_OPERAND(current)
11998 || ! (lparen = av_pop(stack))
11999 || IS_OPERAND(lparen)
12000 || SvUV(lparen) != '(')
12002 SvREFCNT_dec(current);
12004 vFAIL("Unexpected ')'");
12007 SvREFCNT_dec_NN(lparen);
12014 /* Here, we have an operand to process, in 'current' */
12016 if (top_index < 0) { /* Just push if stack is empty */
12017 av_push(stack, current);
12020 SV* top = av_pop(stack);
12022 char current_operator;
12024 if (IS_OPERAND(top)) {
12025 SvREFCNT_dec_NN(top);
12026 SvREFCNT_dec_NN(current);
12027 vFAIL("Operand with no preceding operator");
12029 current_operator = (char) SvUV(top);
12030 switch (current_operator) {
12031 case '(': /* Push the '(' back on followed by the new
12033 av_push(stack, top);
12034 av_push(stack, current);
12035 SvREFCNT_inc(top); /* Counters the '_dec' done
12036 just after the 'break', so
12037 it doesn't get wrongly freed
12042 _invlist_invert(current);
12044 /* Unlike binary operators, the top of the stack,
12045 * now that this unary one has been popped off, may
12046 * legally be an operator, and we now have operand
12049 SvREFCNT_dec_NN(top);
12050 goto handle_operand;
12053 prev = av_pop(stack);
12054 _invlist_intersection(prev,
12057 av_push(stack, current);
12062 prev = av_pop(stack);
12063 _invlist_union(prev, current, ¤t);
12064 av_push(stack, current);
12068 prev = av_pop(stack);;
12069 _invlist_subtract(prev, current, ¤t);
12070 av_push(stack, current);
12073 case '^': /* The union minus the intersection */
12079 prev = av_pop(stack);
12080 _invlist_union(prev, current, &u);
12081 _invlist_intersection(prev, current, &i);
12082 /* _invlist_subtract will overwrite current
12083 without freeing what it already contains */
12085 _invlist_subtract(u, i, ¤t);
12086 av_push(stack, current);
12087 SvREFCNT_dec_NN(i);
12088 SvREFCNT_dec_NN(u);
12089 SvREFCNT_dec_NN(element);
12094 Perl_croak(aTHX_ "panic: Unexpected item on '(?[ ])' stack");
12096 SvREFCNT_dec_NN(top);
12097 SvREFCNT_dec(prev);
12101 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12104 if (av_tindex(stack) < 0 /* Was empty */
12105 || ((final = av_pop(stack)) == NULL)
12106 || ! IS_OPERAND(final)
12107 || av_tindex(stack) >= 0) /* More left on stack */
12109 vFAIL("Incomplete expression within '(?[ ])'");
12112 /* Here, 'final' is the resultant inversion list from evaluating the
12113 * expression. Return it if so requested */
12114 if (return_invlist) {
12115 *return_invlist = final;
12119 /* Otherwise generate a resultant node, based on 'final'. regclass() is
12120 * expecting a string of ranges and individual code points */
12121 invlist_iterinit(final);
12122 result_string = newSVpvs("");
12123 while (invlist_iternext(final, &start, &end)) {
12124 if (start == end) {
12125 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
12128 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
12133 save_parse = RExC_parse;
12134 RExC_parse = SvPV(result_string, len);
12135 save_end = RExC_end;
12136 RExC_end = RExC_parse + len;
12138 /* We turn off folding around the call, as the class we have constructed
12139 * already has all folding taken into consideration, and we don't want
12140 * regclass() to add to that */
12141 RExC_flags &= ~RXf_PMf_FOLD;
12142 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
12144 node = regclass(pRExC_state, flagp,depth+1,
12145 FALSE, /* means parse the whole char class */
12146 FALSE, /* don't allow multi-char folds */
12147 TRUE, /* silence non-portable warnings. The above may very
12148 well have generated non-portable code points, but
12149 they're valid on this machine */
12152 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
12155 RExC_flags |= RXf_PMf_FOLD;
12157 RExC_parse = save_parse + 1;
12158 RExC_end = save_end;
12159 SvREFCNT_dec_NN(final);
12160 SvREFCNT_dec_NN(result_string);
12162 nextchar(pRExC_state);
12163 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
12168 /* The names of properties whose definitions are not known at compile time are
12169 * stored in this SV, after a constant heading. So if the length has been
12170 * changed since initialization, then there is a run-time definition. */
12171 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
12174 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
12175 const bool stop_at_1, /* Just parse the next thing, don't
12176 look for a full character class */
12177 bool allow_multi_folds,
12178 const bool silence_non_portable, /* Don't output warnings
12181 SV** ret_invlist) /* Return an inversion list, not a node */
12183 /* parse a bracketed class specification. Most of these will produce an
12184 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
12185 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
12186 * under /i with multi-character folds: it will be rewritten following the
12187 * paradigm of this example, where the <multi-fold>s are characters which
12188 * fold to multiple character sequences:
12189 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
12190 * gets effectively rewritten as:
12191 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
12192 * reg() gets called (recursively) on the rewritten version, and this
12193 * function will return what it constructs. (Actually the <multi-fold>s
12194 * aren't physically removed from the [abcdefghi], it's just that they are
12195 * ignored in the recursion by means of a flag:
12196 * <RExC_in_multi_char_class>.)
12198 * ANYOF nodes contain a bit map for the first 256 characters, with the
12199 * corresponding bit set if that character is in the list. For characters
12200 * above 255, a range list or swash is used. There are extra bits for \w,
12201 * etc. in locale ANYOFs, as what these match is not determinable at
12204 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
12205 * to be restarted. This can only happen if ret_invlist is non-NULL.
12209 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
12211 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
12214 IV namedclass = OOB_NAMEDCLASS;
12215 char *rangebegin = NULL;
12216 bool need_class = 0;
12218 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
12219 than just initialized. */
12220 SV* properties = NULL; /* Code points that match \p{} \P{} */
12221 SV* posixes = NULL; /* Code points that match classes like, [:word:],
12222 extended beyond the Latin1 range */
12223 UV element_count = 0; /* Number of distinct elements in the class.
12224 Optimizations may be possible if this is tiny */
12225 AV * multi_char_matches = NULL; /* Code points that fold to more than one
12226 character; used under /i */
12228 char * stop_ptr = RExC_end; /* where to stop parsing */
12229 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
12231 const bool strict = cBOOL(ret_invlist); /* Apply strict parsing rules? */
12233 /* Unicode properties are stored in a swash; this holds the current one
12234 * being parsed. If this swash is the only above-latin1 component of the
12235 * character class, an optimization is to pass it directly on to the
12236 * execution engine. Otherwise, it is set to NULL to indicate that there
12237 * are other things in the class that have to be dealt with at execution
12239 SV* swash = NULL; /* Code points that match \p{} \P{} */
12241 /* Set if a component of this character class is user-defined; just passed
12242 * on to the engine */
12243 bool has_user_defined_property = FALSE;
12245 /* inversion list of code points this node matches only when the target
12246 * string is in UTF-8. (Because is under /d) */
12247 SV* depends_list = NULL;
12249 /* inversion list of code points this node matches. For much of the
12250 * function, it includes only those that match regardless of the utf8ness
12251 * of the target string */
12252 SV* cp_list = NULL;
12255 /* In a range, counts how many 0-2 of the ends of it came from literals,
12256 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
12257 UV literal_endpoint = 0;
12259 bool invert = FALSE; /* Is this class to be complemented */
12261 /* Is there any thing like \W or [:^digit:] that matches above the legal
12262 * Unicode range? */
12263 bool runtime_posix_matches_above_Unicode = FALSE;
12265 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
12266 case we need to change the emitted regop to an EXACT. */
12267 const char * orig_parse = RExC_parse;
12268 const I32 orig_size = RExC_size;
12269 GET_RE_DEBUG_FLAGS_DECL;
12271 PERL_ARGS_ASSERT_REGCLASS;
12273 PERL_UNUSED_ARG(depth);
12276 DEBUG_PARSE("clas");
12278 /* Assume we are going to generate an ANYOF node. */
12279 ret = reganode(pRExC_state, ANYOF, 0);
12282 RExC_size += ANYOF_SKIP;
12283 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
12286 ANYOF_FLAGS(ret) = 0;
12288 RExC_emit += ANYOF_SKIP;
12290 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
12292 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
12293 initial_listsv_len = SvCUR(listsv);
12294 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
12298 RExC_parse = regpatws(pRExC_state, RExC_parse,
12299 FALSE /* means don't recognize comments */);
12302 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
12305 allow_multi_folds = FALSE;
12308 RExC_parse = regpatws(pRExC_state, RExC_parse,
12309 FALSE /* means don't recognize comments */);
12313 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
12314 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
12315 const char *s = RExC_parse;
12316 const char c = *s++;
12318 while (isWORDCHAR(*s))
12320 if (*s && c == *s && s[1] == ']') {
12321 SAVEFREESV(RExC_rx_sv);
12323 "POSIX syntax [%c %c] belongs inside character classes",
12325 (void)ReREFCNT_inc(RExC_rx_sv);
12329 /* If the caller wants us to just parse a single element, accomplish this
12330 * by faking the loop ending condition */
12331 if (stop_at_1 && RExC_end > RExC_parse) {
12332 stop_ptr = RExC_parse + 1;
12335 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
12336 if (UCHARAT(RExC_parse) == ']')
12337 goto charclassloop;
12341 if (RExC_parse >= stop_ptr) {
12346 RExC_parse = regpatws(pRExC_state, RExC_parse,
12347 FALSE /* means don't recognize comments */);
12350 if (UCHARAT(RExC_parse) == ']') {
12356 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
12357 save_value = value;
12358 save_prevvalue = prevvalue;
12361 rangebegin = RExC_parse;
12365 value = utf8n_to_uvchr((U8*)RExC_parse,
12366 RExC_end - RExC_parse,
12367 &numlen, UTF8_ALLOW_DEFAULT);
12368 RExC_parse += numlen;
12371 value = UCHARAT(RExC_parse++);
12374 && RExC_parse < RExC_end
12375 && POSIXCC(UCHARAT(RExC_parse)))
12377 namedclass = regpposixcc(pRExC_state, value, strict);
12379 else if (value == '\\') {
12381 value = utf8n_to_uvchr((U8*)RExC_parse,
12382 RExC_end - RExC_parse,
12383 &numlen, UTF8_ALLOW_DEFAULT);
12384 RExC_parse += numlen;
12387 value = UCHARAT(RExC_parse++);
12389 /* Some compilers cannot handle switching on 64-bit integer
12390 * values, therefore value cannot be an UV. Yes, this will
12391 * be a problem later if we want switch on Unicode.
12392 * A similar issue a little bit later when switching on
12393 * namedclass. --jhi */
12395 /* If the \ is escaping white space when white space is being
12396 * skipped, it means that that white space is wanted literally, and
12397 * is already in 'value'. Otherwise, need to translate the escape
12398 * into what it signifies. */
12399 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
12401 case 'w': namedclass = ANYOF_WORDCHAR; break;
12402 case 'W': namedclass = ANYOF_NWORDCHAR; break;
12403 case 's': namedclass = ANYOF_SPACE; break;
12404 case 'S': namedclass = ANYOF_NSPACE; break;
12405 case 'd': namedclass = ANYOF_DIGIT; break;
12406 case 'D': namedclass = ANYOF_NDIGIT; break;
12407 case 'v': namedclass = ANYOF_VERTWS; break;
12408 case 'V': namedclass = ANYOF_NVERTWS; break;
12409 case 'h': namedclass = ANYOF_HORIZWS; break;
12410 case 'H': namedclass = ANYOF_NHORIZWS; break;
12411 case 'N': /* Handle \N{NAME} in class */
12413 /* We only pay attention to the first char of
12414 multichar strings being returned. I kinda wonder
12415 if this makes sense as it does change the behaviour
12416 from earlier versions, OTOH that behaviour was broken
12418 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
12419 TRUE, /* => charclass */
12422 if (*flagp & RESTART_UTF8)
12423 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12433 /* We will handle any undefined properties ourselves */
12434 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
12436 if (RExC_parse >= RExC_end)
12437 vFAIL2("Empty \\%c{}", (U8)value);
12438 if (*RExC_parse == '{') {
12439 const U8 c = (U8)value;
12440 e = strchr(RExC_parse++, '}');
12442 vFAIL2("Missing right brace on \\%c{}", c);
12443 while (isSPACE(UCHARAT(RExC_parse)))
12445 if (e == RExC_parse)
12446 vFAIL2("Empty \\%c{}", c);
12447 n = e - RExC_parse;
12448 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
12459 if (UCHARAT(RExC_parse) == '^') {
12462 /* toggle. (The rhs xor gets the single bit that
12463 * differs between P and p; the other xor inverts just
12465 value ^= 'P' ^ 'p';
12467 while (isSPACE(UCHARAT(RExC_parse))) {
12472 /* Try to get the definition of the property into
12473 * <invlist>. If /i is in effect, the effective property
12474 * will have its name be <__NAME_i>. The design is
12475 * discussed in commit
12476 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
12477 Newx(name, n + sizeof("_i__\n"), char);
12479 sprintf(name, "%s%.*s%s\n",
12480 (FOLD) ? "__" : "",
12486 /* Look up the property name, and get its swash and
12487 * inversion list, if the property is found */
12489 SvREFCNT_dec_NN(swash);
12491 swash = _core_swash_init("utf8", name, &PL_sv_undef,
12494 NULL, /* No inversion list */
12497 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
12499 SvREFCNT_dec_NN(swash);
12503 /* Here didn't find it. It could be a user-defined
12504 * property that will be available at run-time. If we
12505 * accept only compile-time properties, is an error;
12506 * otherwise add it to the list for run-time look up */
12508 RExC_parse = e + 1;
12509 vFAIL3("Property '%.*s' is unknown", (int) n, name);
12511 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
12512 (value == 'p' ? '+' : '!'),
12514 has_user_defined_property = TRUE;
12516 /* We don't know yet, so have to assume that the
12517 * property could match something in the Latin1 range,
12518 * hence something that isn't utf8. Note that this
12519 * would cause things in <depends_list> to match
12520 * inappropriately, except that any \p{}, including
12521 * this one forces Unicode semantics, which means there
12522 * is <no depends_list> */
12523 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
12527 /* Here, did get the swash and its inversion list. If
12528 * the swash is from a user-defined property, then this
12529 * whole character class should be regarded as such */
12530 has_user_defined_property =
12532 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
12534 /* Invert if asking for the complement */
12535 if (value == 'P') {
12536 _invlist_union_complement_2nd(properties,
12540 /* The swash can't be used as-is, because we've
12541 * inverted things; delay removing it to here after
12542 * have copied its invlist above */
12543 SvREFCNT_dec_NN(swash);
12547 _invlist_union(properties, invlist, &properties);
12552 RExC_parse = e + 1;
12553 namedclass = ANYOF_UNIPROP; /* no official name, but it's
12556 /* \p means they want Unicode semantics */
12557 RExC_uni_semantics = 1;
12560 case 'n': value = '\n'; break;
12561 case 'r': value = '\r'; break;
12562 case 't': value = '\t'; break;
12563 case 'f': value = '\f'; break;
12564 case 'b': value = '\b'; break;
12565 case 'e': value = ASCII_TO_NATIVE('\033');break;
12566 case 'a': value = ASCII_TO_NATIVE('\007');break;
12568 RExC_parse--; /* function expects to be pointed at the 'o' */
12570 const char* error_msg;
12571 bool valid = grok_bslash_o(&RExC_parse,
12574 SIZE_ONLY, /* warnings in pass
12577 silence_non_portable,
12583 if (PL_encoding && value < 0x100) {
12584 goto recode_encoding;
12588 RExC_parse--; /* function expects to be pointed at the 'x' */
12590 const char* error_msg;
12591 bool valid = grok_bslash_x(&RExC_parse,
12594 TRUE, /* Output warnings */
12596 silence_non_portable,
12602 if (PL_encoding && value < 0x100)
12603 goto recode_encoding;
12606 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
12608 case '0': case '1': case '2': case '3': case '4':
12609 case '5': case '6': case '7':
12611 /* Take 1-3 octal digits */
12612 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12613 numlen = (strict) ? 4 : 3;
12614 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
12615 RExC_parse += numlen;
12618 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
12619 vFAIL("Need exactly 3 octal digits");
12621 else if (! SIZE_ONLY /* like \08, \178 */
12623 && RExC_parse < RExC_end
12624 && isDIGIT(*RExC_parse)
12625 && ckWARN(WARN_REGEXP))
12627 SAVEFREESV(RExC_rx_sv);
12628 reg_warn_non_literal_string(
12630 form_short_octal_warning(RExC_parse, numlen));
12631 (void)ReREFCNT_inc(RExC_rx_sv);
12634 if (PL_encoding && value < 0x100)
12635 goto recode_encoding;
12639 if (! RExC_override_recoding) {
12640 SV* enc = PL_encoding;
12641 value = reg_recode((const char)(U8)value, &enc);
12644 vFAIL("Invalid escape in the specified encoding");
12646 else if (SIZE_ONLY) {
12647 ckWARNreg(RExC_parse,
12648 "Invalid escape in the specified encoding");
12654 /* Allow \_ to not give an error */
12655 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
12657 vFAIL2("Unrecognized escape \\%c in character class",
12661 SAVEFREESV(RExC_rx_sv);
12662 ckWARN2reg(RExC_parse,
12663 "Unrecognized escape \\%c in character class passed through",
12665 (void)ReREFCNT_inc(RExC_rx_sv);
12669 } /* End of switch on char following backslash */
12670 } /* end of handling backslash escape sequences */
12673 literal_endpoint++;
12676 /* Here, we have the current token in 'value' */
12678 /* What matches in a locale is not known until runtime. This includes
12679 * what the Posix classes (like \w, [:space:]) match. Room must be
12680 * reserved (one time per class) to store such classes, either if Perl
12681 * is compiled so that locale nodes always should have this space, or
12682 * if there is such class info to be stored. The space will contain a
12683 * bit for each named class that is to be matched against. This isn't
12684 * needed for \p{} and pseudo-classes, as they are not affected by
12685 * locale, and hence are dealt with separately */
12688 && (ANYOF_LOCALE == ANYOF_CLASS
12689 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
12693 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12696 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
12697 ANYOF_CLASS_ZERO(ret);
12699 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
12702 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
12704 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
12705 * literal, as is the character that began the false range, i.e.
12706 * the 'a' in the examples */
12709 const int w = (RExC_parse >= rangebegin)
12710 ? RExC_parse - rangebegin
12713 vFAIL4("False [] range \"%*.*s\"", w, w, rangebegin);
12716 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
12717 ckWARN4reg(RExC_parse,
12718 "False [] range \"%*.*s\"",
12720 (void)ReREFCNT_inc(RExC_rx_sv);
12721 cp_list = add_cp_to_invlist(cp_list, '-');
12722 cp_list = add_cp_to_invlist(cp_list, prevvalue);
12726 range = 0; /* this was not a true range */
12727 element_count += 2; /* So counts for three values */
12731 U8 classnum = namedclass_to_classnum(namedclass);
12732 if (namedclass >= ANYOF_MAX) { /* If a special class */
12733 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
12735 /* Here, should be \h, \H, \v, or \V. Neither /d nor
12736 * /l make a difference in what these match. There
12737 * would be problems if these characters had folds
12738 * other than themselves, as cp_list is subject to
12740 if (classnum != _CC_VERTSPACE) {
12741 assert( namedclass == ANYOF_HORIZWS
12742 || namedclass == ANYOF_NHORIZWS);
12744 /* It turns out that \h is just a synonym for
12746 classnum = _CC_BLANK;
12749 _invlist_union_maybe_complement_2nd(
12751 PL_XPosix_ptrs[classnum],
12752 cBOOL(namedclass % 2), /* Complement if odd
12753 (NHORIZWS, NVERTWS)
12758 else if (classnum == _CC_ASCII) {
12761 ANYOF_CLASS_SET(ret, namedclass);
12764 #endif /* Not isascii(); just use the hard-coded definition for it */
12765 _invlist_union_maybe_complement_2nd(
12768 cBOOL(namedclass % 2), /* Complement if odd
12772 else { /* Garden variety class */
12774 /* The ascii range inversion list */
12775 SV* ascii_source = PL_Posix_ptrs[classnum];
12777 /* The full Latin1 range inversion list */
12778 SV* l1_source = PL_L1Posix_ptrs[classnum];
12780 /* This code is structured into two major clauses. The
12781 * first is for classes whose complete definitions may not
12782 * already be known. It not, the Latin1 definition
12783 * (guaranteed to already known) is used plus code is
12784 * generated to load the rest at run-time (only if needed).
12785 * If the complete definition is known, it drops down to
12786 * the second clause, where the complete definition is
12789 if (classnum < _FIRST_NON_SWASH_CC) {
12791 /* Here, the class has a swash, which may or not
12792 * already be loaded */
12794 /* The name of the property to use to match the full
12795 * eXtended Unicode range swash for this character
12797 const char *Xname = swash_property_names[classnum];
12799 /* If returning the inversion list, we can't defer
12800 * getting this until runtime */
12801 if (ret_invlist && ! PL_utf8_swash_ptrs[classnum]) {
12802 PL_utf8_swash_ptrs[classnum] =
12803 _core_swash_init("utf8", Xname, &PL_sv_undef,
12806 NULL, /* No inversion list */
12807 NULL /* No flags */
12809 assert(PL_utf8_swash_ptrs[classnum]);
12811 if ( ! PL_utf8_swash_ptrs[classnum]) {
12812 if (namedclass % 2 == 0) { /* A non-complemented
12814 /* If not /a matching, there are code points we
12815 * don't know at compile time. Arrange for the
12816 * unknown matches to be loaded at run-time, if
12818 if (! AT_LEAST_ASCII_RESTRICTED) {
12819 Perl_sv_catpvf(aTHX_ listsv, "+utf8::%s\n",
12822 if (LOC) { /* Under locale, set run-time
12824 ANYOF_CLASS_SET(ret, namedclass);
12827 /* Add the current class's code points to
12828 * the running total */
12829 _invlist_union(posixes,
12830 (AT_LEAST_ASCII_RESTRICTED)
12836 else { /* A complemented class */
12837 if (AT_LEAST_ASCII_RESTRICTED) {
12838 /* Under /a should match everything above
12839 * ASCII, plus the complement of the set's
12841 _invlist_union_complement_2nd(posixes,
12846 /* Arrange for the unknown matches to be
12847 * loaded at run-time, if needed */
12848 Perl_sv_catpvf(aTHX_ listsv, "!utf8::%s\n",
12850 runtime_posix_matches_above_Unicode = TRUE;
12852 ANYOF_CLASS_SET(ret, namedclass);
12856 /* We want to match everything in
12857 * Latin1, except those things that
12858 * l1_source matches */
12859 SV* scratch_list = NULL;
12860 _invlist_subtract(PL_Latin1, l1_source,
12863 /* Add the list from this class to the
12866 posixes = scratch_list;
12869 _invlist_union(posixes,
12872 SvREFCNT_dec_NN(scratch_list);
12874 if (DEPENDS_SEMANTICS) {
12876 |= ANYOF_NON_UTF8_LATIN1_ALL;
12881 goto namedclass_done;
12884 /* Here, there is a swash loaded for the class. If no
12885 * inversion list for it yet, get it */
12886 if (! PL_XPosix_ptrs[classnum]) {
12887 PL_XPosix_ptrs[classnum]
12888 = _swash_to_invlist(PL_utf8_swash_ptrs[classnum]);
12892 /* Here there is an inversion list already loaded for the
12895 if (namedclass % 2 == 0) { /* A non-complemented class,
12896 like ANYOF_PUNCT */
12898 /* For non-locale, just add it to any existing list
12900 _invlist_union(posixes,
12901 (AT_LEAST_ASCII_RESTRICTED)
12903 : PL_XPosix_ptrs[classnum],
12906 else { /* Locale */
12907 SV* scratch_list = NULL;
12909 /* For above Latin1 code points, we use the full
12911 _invlist_intersection(PL_AboveLatin1,
12912 PL_XPosix_ptrs[classnum],
12914 /* And set the output to it, adding instead if
12915 * there already is an output. Checking if
12916 * 'posixes' is NULL first saves an extra clone.
12917 * Its reference count will be decremented at the
12918 * next union, etc, or if this is the only
12919 * instance, at the end of the routine */
12921 posixes = scratch_list;
12924 _invlist_union(posixes, scratch_list, &posixes);
12925 SvREFCNT_dec_NN(scratch_list);
12928 #ifndef HAS_ISBLANK
12929 if (namedclass != ANYOF_BLANK) {
12931 /* Set this class in the node for runtime
12933 ANYOF_CLASS_SET(ret, namedclass);
12934 #ifndef HAS_ISBLANK
12937 /* No isblank(), use the hard-coded ASCII-range
12938 * blanks, adding them to the running total. */
12940 _invlist_union(posixes, ascii_source, &posixes);
12945 else { /* A complemented class, like ANYOF_NPUNCT */
12947 _invlist_union_complement_2nd(
12949 (AT_LEAST_ASCII_RESTRICTED)
12951 : PL_XPosix_ptrs[classnum],
12953 /* Under /d, everything in the upper half of the
12954 * Latin1 range matches this complement */
12955 if (DEPENDS_SEMANTICS) {
12956 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
12959 else { /* Locale */
12960 SV* scratch_list = NULL;
12961 _invlist_subtract(PL_AboveLatin1,
12962 PL_XPosix_ptrs[classnum],
12965 posixes = scratch_list;
12968 _invlist_union(posixes, scratch_list, &posixes);
12969 SvREFCNT_dec_NN(scratch_list);
12971 #ifndef HAS_ISBLANK
12972 if (namedclass != ANYOF_NBLANK) {
12974 ANYOF_CLASS_SET(ret, namedclass);
12975 #ifndef HAS_ISBLANK
12978 /* Get the list of all code points in Latin1
12979 * that are not ASCII blanks, and add them to
12980 * the running total */
12981 _invlist_subtract(PL_Latin1, ascii_source,
12983 _invlist_union(posixes, scratch_list, &posixes);
12984 SvREFCNT_dec_NN(scratch_list);
12991 continue; /* Go get next character */
12993 } /* end of namedclass \blah */
12995 /* Here, we have a single value. If 'range' is set, it is the ending
12996 * of a range--check its validity. Later, we will handle each
12997 * individual code point in the range. If 'range' isn't set, this
12998 * could be the beginning of a range, so check for that by looking
12999 * ahead to see if the next real character to be processed is the range
13000 * indicator--the minus sign */
13003 RExC_parse = regpatws(pRExC_state, RExC_parse,
13004 FALSE /* means don't recognize comments */);
13008 if (prevvalue > value) /* b-a */ {
13009 const int w = RExC_parse - rangebegin;
13010 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
13011 range = 0; /* not a valid range */
13015 prevvalue = value; /* save the beginning of the potential range */
13016 if (! stop_at_1 /* Can't be a range if parsing just one thing */
13017 && *RExC_parse == '-')
13019 char* next_char_ptr = RExC_parse + 1;
13020 if (skip_white) { /* Get the next real char after the '-' */
13021 next_char_ptr = regpatws(pRExC_state,
13023 FALSE); /* means don't recognize
13027 /* If the '-' is at the end of the class (just before the ']',
13028 * it is a literal minus; otherwise it is a range */
13029 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
13030 RExC_parse = next_char_ptr;
13032 /* a bad range like \w-, [:word:]- ? */
13033 if (namedclass > OOB_NAMEDCLASS) {
13034 if (strict || ckWARN(WARN_REGEXP)) {
13036 RExC_parse >= rangebegin ?
13037 RExC_parse - rangebegin : 0;
13039 vFAIL4("False [] range \"%*.*s\"",
13044 "False [] range \"%*.*s\"",
13049 cp_list = add_cp_to_invlist(cp_list, '-');
13053 range = 1; /* yeah, it's a range! */
13054 continue; /* but do it the next time */
13059 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
13062 /* non-Latin1 code point implies unicode semantics. Must be set in
13063 * pass1 so is there for the whole of pass 2 */
13065 RExC_uni_semantics = 1;
13068 /* Ready to process either the single value, or the completed range.
13069 * For single-valued non-inverted ranges, we consider the possibility
13070 * of multi-char folds. (We made a conscious decision to not do this
13071 * for the other cases because it can often lead to non-intuitive
13072 * results. For example, you have the peculiar case that:
13073 * "s s" =~ /^[^\xDF]+$/i => Y
13074 * "ss" =~ /^[^\xDF]+$/i => N
13076 * See [perl #89750] */
13077 if (FOLD && allow_multi_folds && value == prevvalue) {
13078 if (value == LATIN_SMALL_LETTER_SHARP_S
13079 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
13082 /* Here <value> is indeed a multi-char fold. Get what it is */
13084 U8 foldbuf[UTF8_MAXBYTES_CASE];
13087 UV folded = _to_uni_fold_flags(
13092 | ((LOC) ? FOLD_FLAGS_LOCALE
13093 : (ASCII_FOLD_RESTRICTED)
13094 ? FOLD_FLAGS_NOMIX_ASCII
13098 /* Here, <folded> should be the first character of the
13099 * multi-char fold of <value>, with <foldbuf> containing the
13100 * whole thing. But, if this fold is not allowed (because of
13101 * the flags), <fold> will be the same as <value>, and should
13102 * be processed like any other character, so skip the special
13104 if (folded != value) {
13106 /* Skip if we are recursed, currently parsing the class
13107 * again. Otherwise add this character to the list of
13108 * multi-char folds. */
13109 if (! RExC_in_multi_char_class) {
13110 AV** this_array_ptr;
13112 STRLEN cp_count = utf8_length(foldbuf,
13113 foldbuf + foldlen);
13114 SV* multi_fold = sv_2mortal(newSVpvn("", 0));
13116 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
13119 if (! multi_char_matches) {
13120 multi_char_matches = newAV();
13123 /* <multi_char_matches> is actually an array of arrays.
13124 * There will be one or two top-level elements: [2],
13125 * and/or [3]. The [2] element is an array, each
13126 * element thereof is a character which folds to TWO
13127 * characters; [3] is for folds to THREE characters.
13128 * (Unicode guarantees a maximum of 3 characters in any
13129 * fold.) When we rewrite the character class below,
13130 * we will do so such that the longest folds are
13131 * written first, so that it prefers the longest
13132 * matching strings first. This is done even if it
13133 * turns out that any quantifier is non-greedy, out of
13134 * programmer laziness. Tom Christiansen has agreed
13135 * that this is ok. This makes the test for the
13136 * ligature 'ffi' come before the test for 'ff' */
13137 if (av_exists(multi_char_matches, cp_count)) {
13138 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13140 this_array = *this_array_ptr;
13143 this_array = newAV();
13144 av_store(multi_char_matches, cp_count,
13147 av_push(this_array, multi_fold);
13150 /* This element should not be processed further in this
13153 value = save_value;
13154 prevvalue = save_prevvalue;
13160 /* Deal with this element of the class */
13163 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
13165 SV* this_range = _new_invlist(1);
13166 _append_range_to_invlist(this_range, prevvalue, value);
13168 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
13169 * If this range was specified using something like 'i-j', we want
13170 * to include only the 'i' and the 'j', and not anything in
13171 * between, so exclude non-ASCII, non-alphabetics from it.
13172 * However, if the range was specified with something like
13173 * [\x89-\x91] or [\x89-j], all code points within it should be
13174 * included. literal_endpoint==2 means both ends of the range used
13175 * a literal character, not \x{foo} */
13176 if (literal_endpoint == 2
13177 && (prevvalue >= 'a' && value <= 'z')
13178 || (prevvalue >= 'A' && value <= 'Z'))
13180 _invlist_intersection(this_range, PL_Posix_ptrs[_CC_ALPHA],
13183 _invlist_union(cp_list, this_range, &cp_list);
13184 literal_endpoint = 0;
13188 range = 0; /* this range (if it was one) is done now */
13189 } /* End of loop through all the text within the brackets */
13191 /* If anything in the class expands to more than one character, we have to
13192 * deal with them by building up a substitute parse string, and recursively
13193 * calling reg() on it, instead of proceeding */
13194 if (multi_char_matches) {
13195 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
13198 char *save_end = RExC_end;
13199 char *save_parse = RExC_parse;
13200 bool first_time = TRUE; /* First multi-char occurrence doesn't get
13205 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
13206 because too confusing */
13208 sv_catpv(substitute_parse, "(?:");
13212 /* Look at the longest folds first */
13213 for (cp_count = av_len(multi_char_matches); cp_count > 0; cp_count--) {
13215 if (av_exists(multi_char_matches, cp_count)) {
13216 AV** this_array_ptr;
13219 this_array_ptr = (AV**) av_fetch(multi_char_matches,
13221 while ((this_sequence = av_pop(*this_array_ptr)) !=
13224 if (! first_time) {
13225 sv_catpv(substitute_parse, "|");
13227 first_time = FALSE;
13229 sv_catpv(substitute_parse, SvPVX(this_sequence));
13234 /* If the character class contains anything else besides these
13235 * multi-character folds, have to include it in recursive parsing */
13236 if (element_count) {
13237 sv_catpv(substitute_parse, "|[");
13238 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
13239 sv_catpv(substitute_parse, "]");
13242 sv_catpv(substitute_parse, ")");
13245 /* This is a way to get the parse to skip forward a whole named
13246 * sequence instead of matching the 2nd character when it fails the
13248 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
13252 RExC_parse = SvPV(substitute_parse, len);
13253 RExC_end = RExC_parse + len;
13254 RExC_in_multi_char_class = 1;
13255 RExC_emit = (regnode *)orig_emit;
13257 ret = reg(pRExC_state, 1, ®_flags, depth+1);
13259 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
13261 RExC_parse = save_parse;
13262 RExC_end = save_end;
13263 RExC_in_multi_char_class = 0;
13264 SvREFCNT_dec_NN(multi_char_matches);
13268 /* If the character class contains only a single element, it may be
13269 * optimizable into another node type which is smaller and runs faster.
13270 * Check if this is the case for this class */
13271 if (element_count == 1 && ! ret_invlist) {
13275 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
13276 [:digit:] or \p{foo} */
13278 /* All named classes are mapped into POSIXish nodes, with its FLAG
13279 * argument giving which class it is */
13280 switch ((I32)namedclass) {
13281 case ANYOF_UNIPROP:
13284 /* These don't depend on the charset modifiers. They always
13285 * match under /u rules */
13286 case ANYOF_NHORIZWS:
13287 case ANYOF_HORIZWS:
13288 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
13291 case ANYOF_NVERTWS:
13296 /* The actual POSIXish node for all the rest depends on the
13297 * charset modifier. The ones in the first set depend only on
13298 * ASCII or, if available on this platform, locale */
13302 op = (LOC) ? POSIXL : POSIXA;
13313 /* under /a could be alpha */
13315 if (ASCII_RESTRICTED) {
13316 namedclass = ANYOF_ALPHA + (namedclass % 2);
13324 /* The rest have more possibilities depending on the charset.
13325 * We take advantage of the enum ordering of the charset
13326 * modifiers to get the exact node type, */
13328 op = POSIXD + get_regex_charset(RExC_flags);
13329 if (op > POSIXA) { /* /aa is same as /a */
13332 #ifndef HAS_ISBLANK
13334 && (namedclass == ANYOF_BLANK
13335 || namedclass == ANYOF_NBLANK))
13342 /* The odd numbered ones are the complements of the
13343 * next-lower even number one */
13344 if (namedclass % 2 == 1) {
13348 arg = namedclass_to_classnum(namedclass);
13352 else if (value == prevvalue) {
13354 /* Here, the class consists of just a single code point */
13357 if (! LOC && value == '\n') {
13358 op = REG_ANY; /* Optimize [^\n] */
13359 *flagp |= HASWIDTH|SIMPLE;
13363 else if (value < 256 || UTF) {
13365 /* Optimize a single value into an EXACTish node, but not if it
13366 * would require converting the pattern to UTF-8. */
13367 op = compute_EXACTish(pRExC_state);
13369 } /* Otherwise is a range */
13370 else if (! LOC) { /* locale could vary these */
13371 if (prevvalue == '0') {
13372 if (value == '9') {
13379 /* Here, we have changed <op> away from its initial value iff we found
13380 * an optimization */
13383 /* Throw away this ANYOF regnode, and emit the calculated one,
13384 * which should correspond to the beginning, not current, state of
13386 const char * cur_parse = RExC_parse;
13387 RExC_parse = (char *)orig_parse;
13391 /* To get locale nodes to not use the full ANYOF size would
13392 * require moving the code above that writes the portions
13393 * of it that aren't in other nodes to after this point.
13394 * e.g. ANYOF_CLASS_SET */
13395 RExC_size = orig_size;
13399 RExC_emit = (regnode *)orig_emit;
13400 if (PL_regkind[op] == POSIXD) {
13402 op += NPOSIXD - POSIXD;
13407 ret = reg_node(pRExC_state, op);
13409 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
13413 *flagp |= HASWIDTH|SIMPLE;
13415 else if (PL_regkind[op] == EXACT) {
13416 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13419 RExC_parse = (char *) cur_parse;
13421 SvREFCNT_dec(posixes);
13422 SvREFCNT_dec(cp_list);
13429 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
13431 /* If folding, we calculate all characters that could fold to or from the
13432 * ones already on the list */
13433 if (FOLD && cp_list) {
13434 UV start, end; /* End points of code point ranges */
13436 SV* fold_intersection = NULL;
13438 /* If the highest code point is within Latin1, we can use the
13439 * compiled-in Alphas list, and not have to go out to disk. This
13440 * yields two false positives, the masculine and feminine ordinal
13441 * indicators, which are weeded out below using the
13442 * IS_IN_SOME_FOLD_L1() macro */
13443 if (invlist_highest(cp_list) < 256) {
13444 _invlist_intersection(PL_L1Posix_ptrs[_CC_ALPHA], cp_list,
13445 &fold_intersection);
13449 /* Here, there are non-Latin1 code points, so we will have to go
13450 * fetch the list of all the characters that participate in folds
13452 if (! PL_utf8_foldable) {
13453 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13454 &PL_sv_undef, 1, 0);
13455 PL_utf8_foldable = _get_swash_invlist(swash);
13456 SvREFCNT_dec_NN(swash);
13459 /* This is a hash that for a particular fold gives all characters
13460 * that are involved in it */
13461 if (! PL_utf8_foldclosures) {
13463 /* If we were unable to find any folds, then we likely won't be
13464 * able to find the closures. So just create an empty list.
13465 * Folding will effectively be restricted to the non-Unicode
13466 * rules hard-coded into Perl. (This case happens legitimately
13467 * during compilation of Perl itself before the Unicode tables
13468 * are generated) */
13469 if (_invlist_len(PL_utf8_foldable) == 0) {
13470 PL_utf8_foldclosures = newHV();
13473 /* If the folds haven't been read in, call a fold function
13475 if (! PL_utf8_tofold) {
13476 U8 dummy[UTF8_MAXBYTES+1];
13478 /* This string is just a short named one above \xff */
13479 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
13480 assert(PL_utf8_tofold); /* Verify that worked */
13482 PL_utf8_foldclosures =
13483 _swash_inversion_hash(PL_utf8_tofold);
13487 /* Only the characters in this class that participate in folds need
13488 * be checked. Get the intersection of this class and all the
13489 * possible characters that are foldable. This can quickly narrow
13490 * down a large class */
13491 _invlist_intersection(PL_utf8_foldable, cp_list,
13492 &fold_intersection);
13495 /* Now look at the foldable characters in this class individually */
13496 invlist_iterinit(fold_intersection);
13497 while (invlist_iternext(fold_intersection, &start, &end)) {
13500 /* Locale folding for Latin1 characters is deferred until runtime */
13501 if (LOC && start < 256) {
13505 /* Look at every character in the range */
13506 for (j = start; j <= end; j++) {
13508 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
13514 /* We have the latin1 folding rules hard-coded here so that
13515 * an innocent-looking character class, like /[ks]/i won't
13516 * have to go out to disk to find the possible matches.
13517 * XXX It would be better to generate these via regen, in
13518 * case a new version of the Unicode standard adds new
13519 * mappings, though that is not really likely, and may be
13520 * caught by the default: case of the switch below. */
13522 if (IS_IN_SOME_FOLD_L1(j)) {
13524 /* ASCII is always matched; non-ASCII is matched only
13525 * under Unicode rules */
13526 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
13528 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
13532 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
13536 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
13537 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
13539 /* Certain Latin1 characters have matches outside
13540 * Latin1. To get here, <j> is one of those
13541 * characters. None of these matches is valid for
13542 * ASCII characters under /aa, which is why the 'if'
13543 * just above excludes those. These matches only
13544 * happen when the target string is utf8. The code
13545 * below adds the single fold closures for <j> to the
13546 * inversion list. */
13551 add_cp_to_invlist(cp_list, KELVIN_SIGN);
13555 cp_list = add_cp_to_invlist(cp_list,
13556 LATIN_SMALL_LETTER_LONG_S);
13559 cp_list = add_cp_to_invlist(cp_list,
13560 GREEK_CAPITAL_LETTER_MU);
13561 cp_list = add_cp_to_invlist(cp_list,
13562 GREEK_SMALL_LETTER_MU);
13564 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
13565 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
13567 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
13569 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
13570 cp_list = add_cp_to_invlist(cp_list,
13571 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
13573 case LATIN_SMALL_LETTER_SHARP_S:
13574 cp_list = add_cp_to_invlist(cp_list,
13575 LATIN_CAPITAL_LETTER_SHARP_S);
13577 case 'F': case 'f':
13578 case 'I': case 'i':
13579 case 'L': case 'l':
13580 case 'T': case 't':
13581 case 'A': case 'a':
13582 case 'H': case 'h':
13583 case 'J': case 'j':
13584 case 'N': case 'n':
13585 case 'W': case 'w':
13586 case 'Y': case 'y':
13587 /* These all are targets of multi-character
13588 * folds from code points that require UTF8 to
13589 * express, so they can't match unless the
13590 * target string is in UTF-8, so no action here
13591 * is necessary, as regexec.c properly handles
13592 * the general case for UTF-8 matching and
13593 * multi-char folds */
13596 /* Use deprecated warning to increase the
13597 * chances of this being output */
13598 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
13605 /* Here is an above Latin1 character. We don't have the rules
13606 * hard-coded for it. First, get its fold. This is the simple
13607 * fold, as the multi-character folds have been handled earlier
13608 * and separated out */
13609 _to_uni_fold_flags(j, foldbuf, &foldlen,
13611 ? FOLD_FLAGS_LOCALE
13612 : (ASCII_FOLD_RESTRICTED)
13613 ? FOLD_FLAGS_NOMIX_ASCII
13616 /* Single character fold of above Latin1. Add everything in
13617 * its fold closure to the list that this node should match.
13618 * The fold closures data structure is a hash with the keys
13619 * being the UTF-8 of every character that is folded to, like
13620 * 'k', and the values each an array of all code points that
13621 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
13622 * Multi-character folds are not included */
13623 if ((listp = hv_fetch(PL_utf8_foldclosures,
13624 (char *) foldbuf, foldlen, FALSE)))
13626 AV* list = (AV*) *listp;
13628 for (k = 0; k <= av_len(list); k++) {
13629 SV** c_p = av_fetch(list, k, FALSE);
13632 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
13636 /* /aa doesn't allow folds between ASCII and non-; /l
13637 * doesn't allow them between above and below 256 */
13638 if ((ASCII_FOLD_RESTRICTED
13639 && (isASCII(c) != isASCII(j)))
13640 || (LOC && c < 256)) {
13644 /* Folds involving non-ascii Latin1 characters
13645 * under /d are added to a separate list */
13646 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
13648 cp_list = add_cp_to_invlist(cp_list, c);
13651 depends_list = add_cp_to_invlist(depends_list, c);
13657 SvREFCNT_dec_NN(fold_intersection);
13660 /* And combine the result (if any) with any inversion list from posix
13661 * classes. The lists are kept separate up to now because we don't want to
13662 * fold the classes (folding of those is automatically handled by the swash
13663 * fetching code) */
13665 if (! DEPENDS_SEMANTICS) {
13667 _invlist_union(cp_list, posixes, &cp_list);
13668 SvREFCNT_dec_NN(posixes);
13675 /* Under /d, we put into a separate list the Latin1 things that
13676 * match only when the target string is utf8 */
13677 SV* nonascii_but_latin1_properties = NULL;
13678 _invlist_intersection(posixes, PL_Latin1,
13679 &nonascii_but_latin1_properties);
13680 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
13681 &nonascii_but_latin1_properties);
13682 _invlist_subtract(posixes, nonascii_but_latin1_properties,
13685 _invlist_union(cp_list, posixes, &cp_list);
13686 SvREFCNT_dec_NN(posixes);
13692 if (depends_list) {
13693 _invlist_union(depends_list, nonascii_but_latin1_properties,
13695 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
13698 depends_list = nonascii_but_latin1_properties;
13703 /* And combine the result (if any) with any inversion list from properties.
13704 * The lists are kept separate up to now so that we can distinguish the two
13705 * in regards to matching above-Unicode. A run-time warning is generated
13706 * if a Unicode property is matched against a non-Unicode code point. But,
13707 * we allow user-defined properties to match anything, without any warning,
13708 * and we also suppress the warning if there is a portion of the character
13709 * class that isn't a Unicode property, and which matches above Unicode, \W
13710 * or [\x{110000}] for example.
13711 * (Note that in this case, unlike the Posix one above, there is no
13712 * <depends_list>, because having a Unicode property forces Unicode
13715 bool warn_super = ! has_user_defined_property;
13718 /* If it matters to the final outcome, see if a non-property
13719 * component of the class matches above Unicode. If so, the
13720 * warning gets suppressed. This is true even if just a single
13721 * such code point is specified, as though not strictly correct if
13722 * another such code point is matched against, the fact that they
13723 * are using above-Unicode code points indicates they should know
13724 * the issues involved */
13726 bool non_prop_matches_above_Unicode =
13727 runtime_posix_matches_above_Unicode
13728 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
13730 non_prop_matches_above_Unicode =
13731 ! non_prop_matches_above_Unicode;
13733 warn_super = ! non_prop_matches_above_Unicode;
13736 _invlist_union(properties, cp_list, &cp_list);
13737 SvREFCNT_dec_NN(properties);
13740 cp_list = properties;
13744 OP(ret) = ANYOF_WARN_SUPER;
13748 /* Here, we have calculated what code points should be in the character
13751 * Now we can see about various optimizations. Fold calculation (which we
13752 * did above) needs to take place before inversion. Otherwise /[^k]/i
13753 * would invert to include K, which under /i would match k, which it
13754 * shouldn't. Therefore we can't invert folded locale now, as it won't be
13755 * folded until runtime */
13757 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
13758 * at compile time. Besides not inverting folded locale now, we can't
13759 * invert if there are things such as \w, which aren't known until runtime
13762 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
13764 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13766 _invlist_invert(cp_list);
13768 /* Any swash can't be used as-is, because we've inverted things */
13770 SvREFCNT_dec_NN(swash);
13774 /* Clear the invert flag since have just done it here */
13779 *ret_invlist = cp_list;
13780 SvREFCNT_dec(swash);
13782 /* Discard the generated node */
13784 RExC_size = orig_size;
13787 RExC_emit = orig_emit;
13792 /* If we didn't do folding, it's because some information isn't available
13793 * until runtime; set the run-time fold flag for these. (We don't have to
13794 * worry about properties folding, as that is taken care of by the swash
13798 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
13801 /* Some character classes are equivalent to other nodes. Such nodes take
13802 * up less room and generally fewer operations to execute than ANYOF nodes.
13803 * Above, we checked for and optimized into some such equivalents for
13804 * certain common classes that are easy to test. Getting to this point in
13805 * the code means that the class didn't get optimized there. Since this
13806 * code is only executed in Pass 2, it is too late to save space--it has
13807 * been allocated in Pass 1, and currently isn't given back. But turning
13808 * things into an EXACTish node can allow the optimizer to join it to any
13809 * adjacent such nodes. And if the class is equivalent to things like /./,
13810 * expensive run-time swashes can be avoided. Now that we have more
13811 * complete information, we can find things necessarily missed by the
13812 * earlier code. I (khw) am not sure how much to look for here. It would
13813 * be easy, but perhaps too slow, to check any candidates against all the
13814 * node types they could possibly match using _invlistEQ(). */
13819 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
13820 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
13823 U8 op = END; /* The optimzation node-type */
13824 const char * cur_parse= RExC_parse;
13826 invlist_iterinit(cp_list);
13827 if (! invlist_iternext(cp_list, &start, &end)) {
13829 /* Here, the list is empty. This happens, for example, when a
13830 * Unicode property is the only thing in the character class, and
13831 * it doesn't match anything. (perluniprops.pod notes such
13834 *flagp |= HASWIDTH|SIMPLE;
13836 else if (start == end) { /* The range is a single code point */
13837 if (! invlist_iternext(cp_list, &start, &end)
13839 /* Don't do this optimization if it would require changing
13840 * the pattern to UTF-8 */
13841 && (start < 256 || UTF))
13843 /* Here, the list contains a single code point. Can optimize
13844 * into an EXACT node */
13853 /* A locale node under folding with one code point can be
13854 * an EXACTFL, as its fold won't be calculated until
13860 /* Here, we are generally folding, but there is only one
13861 * code point to match. If we have to, we use an EXACT
13862 * node, but it would be better for joining with adjacent
13863 * nodes in the optimization pass if we used the same
13864 * EXACTFish node that any such are likely to be. We can
13865 * do this iff the code point doesn't participate in any
13866 * folds. For example, an EXACTF of a colon is the same as
13867 * an EXACT one, since nothing folds to or from a colon. */
13869 if (IS_IN_SOME_FOLD_L1(value)) {
13874 if (! PL_utf8_foldable) {
13875 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
13876 &PL_sv_undef, 1, 0);
13877 PL_utf8_foldable = _get_swash_invlist(swash);
13878 SvREFCNT_dec_NN(swash);
13880 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
13885 /* If we haven't found the node type, above, it means we
13886 * can use the prevailing one */
13888 op = compute_EXACTish(pRExC_state);
13893 else if (start == 0) {
13894 if (end == UV_MAX) {
13896 *flagp |= HASWIDTH|SIMPLE;
13899 else if (end == '\n' - 1
13900 && invlist_iternext(cp_list, &start, &end)
13901 && start == '\n' + 1 && end == UV_MAX)
13904 *flagp |= HASWIDTH|SIMPLE;
13908 invlist_iterfinish(cp_list);
13911 RExC_parse = (char *)orig_parse;
13912 RExC_emit = (regnode *)orig_emit;
13914 ret = reg_node(pRExC_state, op);
13916 RExC_parse = (char *)cur_parse;
13918 if (PL_regkind[op] == EXACT) {
13919 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
13922 SvREFCNT_dec_NN(cp_list);
13927 /* Here, <cp_list> contains all the code points we can determine at
13928 * compile time that match under all conditions. Go through it, and
13929 * for things that belong in the bitmap, put them there, and delete from
13930 * <cp_list>. While we are at it, see if everything above 255 is in the
13931 * list, and if so, set a flag to speed up execution */
13932 ANYOF_BITMAP_ZERO(ret);
13935 /* This gets set if we actually need to modify things */
13936 bool change_invlist = FALSE;
13940 /* Start looking through <cp_list> */
13941 invlist_iterinit(cp_list);
13942 while (invlist_iternext(cp_list, &start, &end)) {
13946 if (end == UV_MAX && start <= 256) {
13947 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
13950 /* Quit if are above what we should change */
13955 change_invlist = TRUE;
13957 /* Set all the bits in the range, up to the max that we are doing */
13958 high = (end < 255) ? end : 255;
13959 for (i = start; i <= (int) high; i++) {
13960 if (! ANYOF_BITMAP_TEST(ret, i)) {
13961 ANYOF_BITMAP_SET(ret, i);
13965 invlist_iterfinish(cp_list);
13967 /* Done with loop; remove any code points that are in the bitmap from
13969 if (change_invlist) {
13970 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
13973 /* If have completely emptied it, remove it completely */
13974 if (_invlist_len(cp_list) == 0) {
13975 SvREFCNT_dec_NN(cp_list);
13981 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
13984 /* Here, the bitmap has been populated with all the Latin1 code points that
13985 * always match. Can now add to the overall list those that match only
13986 * when the target string is UTF-8 (<depends_list>). */
13987 if (depends_list) {
13989 _invlist_union(cp_list, depends_list, &cp_list);
13990 SvREFCNT_dec_NN(depends_list);
13993 cp_list = depends_list;
13997 /* If there is a swash and more than one element, we can't use the swash in
13998 * the optimization below. */
13999 if (swash && element_count > 1) {
14000 SvREFCNT_dec_NN(swash);
14005 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14007 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
14010 /* av[0] stores the character class description in its textual form:
14011 * used later (regexec.c:Perl_regclass_swash()) to initialize the
14012 * appropriate swash, and is also useful for dumping the regnode.
14013 * av[1] if NULL, is a placeholder to later contain the swash computed
14014 * from av[0]. But if no further computation need be done, the
14015 * swash is stored there now.
14016 * av[2] stores the cp_list inversion list for use in addition or
14017 * instead of av[0]; used only if av[1] is NULL
14018 * av[3] is set if any component of the class is from a user-defined
14019 * property; used only if av[1] is NULL */
14020 AV * const av = newAV();
14023 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
14024 ? SvREFCNT_inc(listsv) : &PL_sv_undef);
14026 av_store(av, 1, swash);
14027 SvREFCNT_dec_NN(cp_list);
14030 av_store(av, 1, NULL);
14032 av_store(av, 2, cp_list);
14033 av_store(av, 3, newSVuv(has_user_defined_property));
14037 rv = newRV_noinc(MUTABLE_SV(av));
14038 n = add_data(pRExC_state, 1, "s");
14039 RExC_rxi->data->data[n] = (void*)rv;
14043 *flagp |= HASWIDTH|SIMPLE;
14046 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
14049 /* reg_skipcomment()
14051 Absorbs an /x style # comments from the input stream.
14052 Returns true if there is more text remaining in the stream.
14053 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
14054 terminates the pattern without including a newline.
14056 Note its the callers responsibility to ensure that we are
14057 actually in /x mode
14062 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
14066 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
14068 while (RExC_parse < RExC_end)
14069 if (*RExC_parse++ == '\n') {
14074 /* we ran off the end of the pattern without ending
14075 the comment, so we have to add an \n when wrapping */
14076 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
14084 Advances the parse position, and optionally absorbs
14085 "whitespace" from the inputstream.
14087 Without /x "whitespace" means (?#...) style comments only,
14088 with /x this means (?#...) and # comments and whitespace proper.
14090 Returns the RExC_parse point from BEFORE the scan occurs.
14092 This is the /x friendly way of saying RExC_parse++.
14096 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
14098 char* const retval = RExC_parse++;
14100 PERL_ARGS_ASSERT_NEXTCHAR;
14103 if (RExC_end - RExC_parse >= 3
14104 && *RExC_parse == '('
14105 && RExC_parse[1] == '?'
14106 && RExC_parse[2] == '#')
14108 while (*RExC_parse != ')') {
14109 if (RExC_parse == RExC_end)
14110 FAIL("Sequence (?#... not terminated");
14116 if (RExC_flags & RXf_PMf_EXTENDED) {
14117 if (isSPACE(*RExC_parse)) {
14121 else if (*RExC_parse == '#') {
14122 if ( reg_skipcomment( pRExC_state ) )
14131 - reg_node - emit a node
14133 STATIC regnode * /* Location. */
14134 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
14138 regnode * const ret = RExC_emit;
14139 GET_RE_DEBUG_FLAGS_DECL;
14141 PERL_ARGS_ASSERT_REG_NODE;
14144 SIZE_ALIGN(RExC_size);
14148 if (RExC_emit >= RExC_emit_bound)
14149 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14150 op, RExC_emit, RExC_emit_bound);
14152 NODE_ALIGN_FILL(ret);
14154 FILL_ADVANCE_NODE(ptr, op);
14155 #ifdef RE_TRACK_PATTERN_OFFSETS
14156 if (RExC_offsets) { /* MJD */
14157 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
14158 "reg_node", __LINE__,
14160 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
14161 ? "Overwriting end of array!\n" : "OK",
14162 (UV)(RExC_emit - RExC_emit_start),
14163 (UV)(RExC_parse - RExC_start),
14164 (UV)RExC_offsets[0]));
14165 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
14173 - reganode - emit a node with an argument
14175 STATIC regnode * /* Location. */
14176 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
14180 regnode * const ret = RExC_emit;
14181 GET_RE_DEBUG_FLAGS_DECL;
14183 PERL_ARGS_ASSERT_REGANODE;
14186 SIZE_ALIGN(RExC_size);
14191 assert(2==regarglen[op]+1);
14193 Anything larger than this has to allocate the extra amount.
14194 If we changed this to be:
14196 RExC_size += (1 + regarglen[op]);
14198 then it wouldn't matter. Its not clear what side effect
14199 might come from that so its not done so far.
14204 if (RExC_emit >= RExC_emit_bound)
14205 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
14206 op, RExC_emit, RExC_emit_bound);
14208 NODE_ALIGN_FILL(ret);
14210 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
14211 #ifdef RE_TRACK_PATTERN_OFFSETS
14212 if (RExC_offsets) { /* MJD */
14213 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14217 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
14218 "Overwriting end of array!\n" : "OK",
14219 (UV)(RExC_emit - RExC_emit_start),
14220 (UV)(RExC_parse - RExC_start),
14221 (UV)RExC_offsets[0]));
14222 Set_Cur_Node_Offset;
14230 - reguni - emit (if appropriate) a Unicode character
14233 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
14237 PERL_ARGS_ASSERT_REGUNI;
14239 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
14243 - reginsert - insert an operator in front of already-emitted operand
14245 * Means relocating the operand.
14248 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
14254 const int offset = regarglen[(U8)op];
14255 const int size = NODE_STEP_REGNODE + offset;
14256 GET_RE_DEBUG_FLAGS_DECL;
14258 PERL_ARGS_ASSERT_REGINSERT;
14259 PERL_UNUSED_ARG(depth);
14260 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
14261 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
14270 if (RExC_open_parens) {
14272 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
14273 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
14274 if ( RExC_open_parens[paren] >= opnd ) {
14275 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
14276 RExC_open_parens[paren] += size;
14278 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
14280 if ( RExC_close_parens[paren] >= opnd ) {
14281 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
14282 RExC_close_parens[paren] += size;
14284 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
14289 while (src > opnd) {
14290 StructCopy(--src, --dst, regnode);
14291 #ifdef RE_TRACK_PATTERN_OFFSETS
14292 if (RExC_offsets) { /* MJD 20010112 */
14293 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
14297 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
14298 ? "Overwriting end of array!\n" : "OK",
14299 (UV)(src - RExC_emit_start),
14300 (UV)(dst - RExC_emit_start),
14301 (UV)RExC_offsets[0]));
14302 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
14303 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
14309 place = opnd; /* Op node, where operand used to be. */
14310 #ifdef RE_TRACK_PATTERN_OFFSETS
14311 if (RExC_offsets) { /* MJD */
14312 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
14316 (UV)(place - RExC_emit_start) > RExC_offsets[0]
14317 ? "Overwriting end of array!\n" : "OK",
14318 (UV)(place - RExC_emit_start),
14319 (UV)(RExC_parse - RExC_start),
14320 (UV)RExC_offsets[0]));
14321 Set_Node_Offset(place, RExC_parse);
14322 Set_Node_Length(place, 1);
14325 src = NEXTOPER(place);
14326 FILL_ADVANCE_NODE(place, op);
14327 Zero(src, offset, regnode);
14331 - regtail - set the next-pointer at the end of a node chain of p to val.
14332 - SEE ALSO: regtail_study
14334 /* TODO: All three parms should be const */
14336 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14340 GET_RE_DEBUG_FLAGS_DECL;
14342 PERL_ARGS_ASSERT_REGTAIL;
14344 PERL_UNUSED_ARG(depth);
14350 /* Find last node. */
14353 regnode * const temp = regnext(scan);
14355 SV * const mysv=sv_newmortal();
14356 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
14357 regprop(RExC_rx, mysv, scan);
14358 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
14359 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
14360 (temp == NULL ? "->" : ""),
14361 (temp == NULL ? PL_reg_name[OP(val)] : "")
14369 if (reg_off_by_arg[OP(scan)]) {
14370 ARG_SET(scan, val - scan);
14373 NEXT_OFF(scan) = val - scan;
14379 - regtail_study - set the next-pointer at the end of a node chain of p to val.
14380 - Look for optimizable sequences at the same time.
14381 - currently only looks for EXACT chains.
14383 This is experimental code. The idea is to use this routine to perform
14384 in place optimizations on branches and groups as they are constructed,
14385 with the long term intention of removing optimization from study_chunk so
14386 that it is purely analytical.
14388 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
14389 to control which is which.
14392 /* TODO: All four parms should be const */
14395 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
14400 #ifdef EXPERIMENTAL_INPLACESCAN
14403 GET_RE_DEBUG_FLAGS_DECL;
14405 PERL_ARGS_ASSERT_REGTAIL_STUDY;
14411 /* Find last node. */
14415 regnode * const temp = regnext(scan);
14416 #ifdef EXPERIMENTAL_INPLACESCAN
14417 if (PL_regkind[OP(scan)] == EXACT) {
14418 bool has_exactf_sharp_s; /* Unexamined in this routine */
14419 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
14424 switch (OP(scan)) {
14430 case EXACTFU_TRICKYFOLD:
14432 if( exact == PSEUDO )
14434 else if ( exact != OP(scan) )
14443 SV * const mysv=sv_newmortal();
14444 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
14445 regprop(RExC_rx, mysv, scan);
14446 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
14447 SvPV_nolen_const(mysv),
14448 REG_NODE_NUM(scan),
14449 PL_reg_name[exact]);
14456 SV * const mysv_val=sv_newmortal();
14457 DEBUG_PARSE_MSG("");
14458 regprop(RExC_rx, mysv_val, val);
14459 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
14460 SvPV_nolen_const(mysv_val),
14461 (IV)REG_NODE_NUM(val),
14465 if (reg_off_by_arg[OP(scan)]) {
14466 ARG_SET(scan, val - scan);
14469 NEXT_OFF(scan) = val - scan;
14477 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
14482 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
14487 for (bit=0; bit<32; bit++) {
14488 if (flags & (1<<bit)) {
14489 if (!set++ && lead)
14490 PerlIO_printf(Perl_debug_log, "%s",lead);
14491 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
14496 PerlIO_printf(Perl_debug_log, "\n");
14498 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14503 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
14509 for (bit=0; bit<32; bit++) {
14510 if (flags & (1<<bit)) {
14511 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
14514 if (!set++ && lead)
14515 PerlIO_printf(Perl_debug_log, "%s",lead);
14516 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
14519 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
14520 if (!set++ && lead) {
14521 PerlIO_printf(Perl_debug_log, "%s",lead);
14524 case REGEX_UNICODE_CHARSET:
14525 PerlIO_printf(Perl_debug_log, "UNICODE");
14527 case REGEX_LOCALE_CHARSET:
14528 PerlIO_printf(Perl_debug_log, "LOCALE");
14530 case REGEX_ASCII_RESTRICTED_CHARSET:
14531 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
14533 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
14534 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
14537 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
14543 PerlIO_printf(Perl_debug_log, "\n");
14545 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
14551 Perl_regdump(pTHX_ const regexp *r)
14555 SV * const sv = sv_newmortal();
14556 SV *dsv= sv_newmortal();
14557 RXi_GET_DECL(r,ri);
14558 GET_RE_DEBUG_FLAGS_DECL;
14560 PERL_ARGS_ASSERT_REGDUMP;
14562 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
14564 /* Header fields of interest. */
14565 if (r->anchored_substr) {
14566 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
14567 RE_SV_DUMPLEN(r->anchored_substr), 30);
14568 PerlIO_printf(Perl_debug_log,
14569 "anchored %s%s at %"IVdf" ",
14570 s, RE_SV_TAIL(r->anchored_substr),
14571 (IV)r->anchored_offset);
14572 } else if (r->anchored_utf8) {
14573 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
14574 RE_SV_DUMPLEN(r->anchored_utf8), 30);
14575 PerlIO_printf(Perl_debug_log,
14576 "anchored utf8 %s%s at %"IVdf" ",
14577 s, RE_SV_TAIL(r->anchored_utf8),
14578 (IV)r->anchored_offset);
14580 if (r->float_substr) {
14581 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
14582 RE_SV_DUMPLEN(r->float_substr), 30);
14583 PerlIO_printf(Perl_debug_log,
14584 "floating %s%s at %"IVdf"..%"UVuf" ",
14585 s, RE_SV_TAIL(r->float_substr),
14586 (IV)r->float_min_offset, (UV)r->float_max_offset);
14587 } else if (r->float_utf8) {
14588 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
14589 RE_SV_DUMPLEN(r->float_utf8), 30);
14590 PerlIO_printf(Perl_debug_log,
14591 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
14592 s, RE_SV_TAIL(r->float_utf8),
14593 (IV)r->float_min_offset, (UV)r->float_max_offset);
14595 if (r->check_substr || r->check_utf8)
14596 PerlIO_printf(Perl_debug_log,
14598 (r->check_substr == r->float_substr
14599 && r->check_utf8 == r->float_utf8
14600 ? "(checking floating" : "(checking anchored"));
14601 if (r->extflags & RXf_NOSCAN)
14602 PerlIO_printf(Perl_debug_log, " noscan");
14603 if (r->extflags & RXf_CHECK_ALL)
14604 PerlIO_printf(Perl_debug_log, " isall");
14605 if (r->check_substr || r->check_utf8)
14606 PerlIO_printf(Perl_debug_log, ") ");
14608 if (ri->regstclass) {
14609 regprop(r, sv, ri->regstclass);
14610 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
14612 if (r->extflags & RXf_ANCH) {
14613 PerlIO_printf(Perl_debug_log, "anchored");
14614 if (r->extflags & RXf_ANCH_BOL)
14615 PerlIO_printf(Perl_debug_log, "(BOL)");
14616 if (r->extflags & RXf_ANCH_MBOL)
14617 PerlIO_printf(Perl_debug_log, "(MBOL)");
14618 if (r->extflags & RXf_ANCH_SBOL)
14619 PerlIO_printf(Perl_debug_log, "(SBOL)");
14620 if (r->extflags & RXf_ANCH_GPOS)
14621 PerlIO_printf(Perl_debug_log, "(GPOS)");
14622 PerlIO_putc(Perl_debug_log, ' ');
14624 if (r->extflags & RXf_GPOS_SEEN)
14625 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
14626 if (r->intflags & PREGf_SKIP)
14627 PerlIO_printf(Perl_debug_log, "plus ");
14628 if (r->intflags & PREGf_IMPLICIT)
14629 PerlIO_printf(Perl_debug_log, "implicit ");
14630 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
14631 if (r->extflags & RXf_EVAL_SEEN)
14632 PerlIO_printf(Perl_debug_log, "with eval ");
14633 PerlIO_printf(Perl_debug_log, "\n");
14635 regdump_extflags("r->extflags: ",r->extflags);
14636 regdump_intflags("r->intflags: ",r->intflags);
14639 PERL_ARGS_ASSERT_REGDUMP;
14640 PERL_UNUSED_CONTEXT;
14641 PERL_UNUSED_ARG(r);
14642 #endif /* DEBUGGING */
14646 - regprop - printable representation of opcode
14648 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
14651 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
14652 if (flags & ANYOF_INVERT) \
14653 /*make sure the invert info is in each */ \
14654 sv_catpvs(sv, "^"); \
14660 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
14666 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
14667 static const char * const anyofs[] = {
14668 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
14669 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
14670 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
14671 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
14672 || _CC_PSXSPC != 13 || _CC_CNTRL != 14 || _CC_ASCII != 15 \
14673 || _CC_VERTSPACE != 16
14674 #error Need to adjust order of anyofs[]
14711 RXi_GET_DECL(prog,progi);
14712 GET_RE_DEBUG_FLAGS_DECL;
14714 PERL_ARGS_ASSERT_REGPROP;
14718 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
14719 /* It would be nice to FAIL() here, but this may be called from
14720 regexec.c, and it would be hard to supply pRExC_state. */
14721 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
14722 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
14724 k = PL_regkind[OP(o)];
14727 sv_catpvs(sv, " ");
14728 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
14729 * is a crude hack but it may be the best for now since
14730 * we have no flag "this EXACTish node was UTF-8"
14732 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
14733 PERL_PV_ESCAPE_UNI_DETECT |
14734 PERL_PV_ESCAPE_NONASCII |
14735 PERL_PV_PRETTY_ELLIPSES |
14736 PERL_PV_PRETTY_LTGT |
14737 PERL_PV_PRETTY_NOCLEAR
14739 } else if (k == TRIE) {
14740 /* print the details of the trie in dumpuntil instead, as
14741 * progi->data isn't available here */
14742 const char op = OP(o);
14743 const U32 n = ARG(o);
14744 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
14745 (reg_ac_data *)progi->data->data[n] :
14747 const reg_trie_data * const trie
14748 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
14750 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
14751 DEBUG_TRIE_COMPILE_r(
14752 Perl_sv_catpvf(aTHX_ sv,
14753 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
14754 (UV)trie->startstate,
14755 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
14756 (UV)trie->wordcount,
14759 (UV)TRIE_CHARCOUNT(trie),
14760 (UV)trie->uniquecharcount
14763 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
14764 sv_catpvs(sv, "[");
14765 (void) put_latin1_charclass_innards(sv, IS_ANYOF_TRIE(op)
14767 : TRIE_BITMAP(trie));
14768 sv_catpvs(sv, "]");
14771 } else if (k == CURLY) {
14772 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
14773 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
14774 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
14776 else if (k == WHILEM && o->flags) /* Ordinal/of */
14777 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
14778 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
14779 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
14780 if ( RXp_PAREN_NAMES(prog) ) {
14781 if ( k != REF || (OP(o) < NREF)) {
14782 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
14783 SV **name= av_fetch(list, ARG(o), 0 );
14785 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14788 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
14789 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
14790 I32 *nums=(I32*)SvPVX(sv_dat);
14791 SV **name= av_fetch(list, nums[0], 0 );
14794 for ( n=0; n<SvIVX(sv_dat); n++ ) {
14795 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
14796 (n ? "," : ""), (IV)nums[n]);
14798 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
14802 } else if (k == GOSUB)
14803 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
14804 else if (k == VERB) {
14806 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
14807 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
14808 } else if (k == LOGICAL)
14809 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
14810 else if (k == ANYOF) {
14811 const U8 flags = ANYOF_FLAGS(o);
14815 if (flags & ANYOF_LOCALE)
14816 sv_catpvs(sv, "{loc}");
14817 if (flags & ANYOF_LOC_FOLD)
14818 sv_catpvs(sv, "{i}");
14819 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
14820 if (flags & ANYOF_INVERT)
14821 sv_catpvs(sv, "^");
14823 /* output what the standard cp 0-255 bitmap matches */
14824 do_sep = put_latin1_charclass_innards(sv, ANYOF_BITMAP(o));
14826 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14827 /* output any special charclass tests (used entirely under use locale) */
14828 if (ANYOF_CLASS_TEST_ANY_SET(o)) {
14830 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++) {
14831 if (ANYOF_CLASS_TEST(o,i)) {
14832 sv_catpv(sv, anyofs[i]);
14838 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
14840 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
14841 sv_catpvs(sv, "{non-utf8-latin1-all}");
14844 /* output information about the unicode matching */
14845 if (flags & ANYOF_UNICODE_ALL)
14846 sv_catpvs(sv, "{unicode_all}");
14847 else if (ANYOF_NONBITMAP(o)) {
14848 SV *lv; /* Set if there is something outside the bit map. */
14850 bool byte_output = FALSE; /* If something in the bitmap has been
14853 if (flags & ANYOF_NONBITMAP_NON_UTF8) {
14854 sv_catpvs(sv, "{outside bitmap}");
14857 sv_catpvs(sv, "{utf8}");
14860 /* Get the stuff that wasn't in the bitmap */
14861 sw = regclass_swash(prog, o, FALSE, &lv, NULL);
14862 if (lv && lv != &PL_sv_undef) {
14863 char *s = savesvpv(lv);
14864 char * const origs = s;
14866 while (*s && *s != '\n')
14870 const char * const t = ++s;
14873 sv_catpvs(sv, " ");
14879 /* Truncate very long output */
14880 if (s - origs > 256) {
14881 Perl_sv_catpvf(aTHX_ sv,
14883 (int) (s - origs - 1),
14889 else if (*s == '\t') {
14903 SvREFCNT_dec_NN(lv);
14907 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
14909 else if (k == POSIXD || k == NPOSIXD) {
14910 U8 index = FLAGS(o) * 2;
14911 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
14912 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
14915 sv_catpv(sv, anyofs[index]);
14918 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
14919 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
14921 PERL_UNUSED_CONTEXT;
14922 PERL_UNUSED_ARG(sv);
14923 PERL_UNUSED_ARG(o);
14924 PERL_UNUSED_ARG(prog);
14925 #endif /* DEBUGGING */
14929 Perl_re_intuit_string(pTHX_ REGEXP * const r)
14930 { /* Assume that RE_INTUIT is set */
14932 struct regexp *const prog = ReANY(r);
14933 GET_RE_DEBUG_FLAGS_DECL;
14935 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
14936 PERL_UNUSED_CONTEXT;
14940 const char * const s = SvPV_nolen_const(prog->check_substr
14941 ? prog->check_substr : prog->check_utf8);
14943 if (!PL_colorset) reginitcolors();
14944 PerlIO_printf(Perl_debug_log,
14945 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
14947 prog->check_substr ? "" : "utf8 ",
14948 PL_colors[5],PL_colors[0],
14951 (strlen(s) > 60 ? "..." : ""));
14954 return prog->check_substr ? prog->check_substr : prog->check_utf8;
14960 handles refcounting and freeing the perl core regexp structure. When
14961 it is necessary to actually free the structure the first thing it
14962 does is call the 'free' method of the regexp_engine associated to
14963 the regexp, allowing the handling of the void *pprivate; member
14964 first. (This routine is not overridable by extensions, which is why
14965 the extensions free is called first.)
14967 See regdupe and regdupe_internal if you change anything here.
14969 #ifndef PERL_IN_XSUB_RE
14971 Perl_pregfree(pTHX_ REGEXP *r)
14977 Perl_pregfree2(pTHX_ REGEXP *rx)
14980 struct regexp *const r = ReANY(rx);
14981 GET_RE_DEBUG_FLAGS_DECL;
14983 PERL_ARGS_ASSERT_PREGFREE2;
14985 if (r->mother_re) {
14986 ReREFCNT_dec(r->mother_re);
14988 CALLREGFREE_PVT(rx); /* free the private data */
14989 SvREFCNT_dec(RXp_PAREN_NAMES(r));
14990 Safefree(r->xpv_len_u.xpvlenu_pv);
14993 SvREFCNT_dec(r->anchored_substr);
14994 SvREFCNT_dec(r->anchored_utf8);
14995 SvREFCNT_dec(r->float_substr);
14996 SvREFCNT_dec(r->float_utf8);
14997 Safefree(r->substrs);
14999 RX_MATCH_COPY_FREE(rx);
15000 #ifdef PERL_ANY_COW
15001 SvREFCNT_dec(r->saved_copy);
15004 SvREFCNT_dec(r->qr_anoncv);
15005 rx->sv_u.svu_rx = 0;
15010 This is a hacky workaround to the structural issue of match results
15011 being stored in the regexp structure which is in turn stored in
15012 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
15013 could be PL_curpm in multiple contexts, and could require multiple
15014 result sets being associated with the pattern simultaneously, such
15015 as when doing a recursive match with (??{$qr})
15017 The solution is to make a lightweight copy of the regexp structure
15018 when a qr// is returned from the code executed by (??{$qr}) this
15019 lightweight copy doesn't actually own any of its data except for
15020 the starp/end and the actual regexp structure itself.
15026 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
15028 struct regexp *ret;
15029 struct regexp *const r = ReANY(rx);
15030 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
15032 PERL_ARGS_ASSERT_REG_TEMP_COPY;
15035 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
15037 SvOK_off((SV *)ret_x);
15039 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
15040 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
15041 made both spots point to the same regexp body.) */
15042 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
15043 assert(!SvPVX(ret_x));
15044 ret_x->sv_u.svu_rx = temp->sv_any;
15045 temp->sv_any = NULL;
15046 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
15047 SvREFCNT_dec_NN(temp);
15048 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
15049 ing below will not set it. */
15050 SvCUR_set(ret_x, SvCUR(rx));
15053 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
15054 sv_force_normal(sv) is called. */
15056 ret = ReANY(ret_x);
15058 SvFLAGS(ret_x) |= SvUTF8(rx);
15059 /* We share the same string buffer as the original regexp, on which we
15060 hold a reference count, incremented when mother_re is set below.
15061 The string pointer is copied here, being part of the regexp struct.
15063 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
15064 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
15066 const I32 npar = r->nparens+1;
15067 Newx(ret->offs, npar, regexp_paren_pair);
15068 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15071 Newx(ret->substrs, 1, struct reg_substr_data);
15072 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15074 SvREFCNT_inc_void(ret->anchored_substr);
15075 SvREFCNT_inc_void(ret->anchored_utf8);
15076 SvREFCNT_inc_void(ret->float_substr);
15077 SvREFCNT_inc_void(ret->float_utf8);
15079 /* check_substr and check_utf8, if non-NULL, point to either their
15080 anchored or float namesakes, and don't hold a second reference. */
15082 RX_MATCH_COPIED_off(ret_x);
15083 #ifdef PERL_ANY_COW
15084 ret->saved_copy = NULL;
15086 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
15087 SvREFCNT_inc_void(ret->qr_anoncv);
15093 /* regfree_internal()
15095 Free the private data in a regexp. This is overloadable by
15096 extensions. Perl takes care of the regexp structure in pregfree(),
15097 this covers the *pprivate pointer which technically perl doesn't
15098 know about, however of course we have to handle the
15099 regexp_internal structure when no extension is in use.
15101 Note this is called before freeing anything in the regexp
15106 Perl_regfree_internal(pTHX_ REGEXP * const rx)
15109 struct regexp *const r = ReANY(rx);
15110 RXi_GET_DECL(r,ri);
15111 GET_RE_DEBUG_FLAGS_DECL;
15113 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
15119 SV *dsv= sv_newmortal();
15120 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
15121 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
15122 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
15123 PL_colors[4],PL_colors[5],s);
15126 #ifdef RE_TRACK_PATTERN_OFFSETS
15128 Safefree(ri->u.offsets); /* 20010421 MJD */
15130 if (ri->code_blocks) {
15132 for (n = 0; n < ri->num_code_blocks; n++)
15133 SvREFCNT_dec(ri->code_blocks[n].src_regex);
15134 Safefree(ri->code_blocks);
15138 int n = ri->data->count;
15141 /* If you add a ->what type here, update the comment in regcomp.h */
15142 switch (ri->data->what[n]) {
15148 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
15151 Safefree(ri->data->data[n]);
15157 { /* Aho Corasick add-on structure for a trie node.
15158 Used in stclass optimization only */
15160 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
15162 refcount = --aho->refcount;
15165 PerlMemShared_free(aho->states);
15166 PerlMemShared_free(aho->fail);
15167 /* do this last!!!! */
15168 PerlMemShared_free(ri->data->data[n]);
15169 PerlMemShared_free(ri->regstclass);
15175 /* trie structure. */
15177 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
15179 refcount = --trie->refcount;
15182 PerlMemShared_free(trie->charmap);
15183 PerlMemShared_free(trie->states);
15184 PerlMemShared_free(trie->trans);
15186 PerlMemShared_free(trie->bitmap);
15188 PerlMemShared_free(trie->jump);
15189 PerlMemShared_free(trie->wordinfo);
15190 /* do this last!!!! */
15191 PerlMemShared_free(ri->data->data[n]);
15196 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
15199 Safefree(ri->data->what);
15200 Safefree(ri->data);
15206 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
15207 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
15208 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
15211 re_dup - duplicate a regexp.
15213 This routine is expected to clone a given regexp structure. It is only
15214 compiled under USE_ITHREADS.
15216 After all of the core data stored in struct regexp is duplicated
15217 the regexp_engine.dupe method is used to copy any private data
15218 stored in the *pprivate pointer. This allows extensions to handle
15219 any duplication it needs to do.
15221 See pregfree() and regfree_internal() if you change anything here.
15223 #if defined(USE_ITHREADS)
15224 #ifndef PERL_IN_XSUB_RE
15226 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
15230 const struct regexp *r = ReANY(sstr);
15231 struct regexp *ret = ReANY(dstr);
15233 PERL_ARGS_ASSERT_RE_DUP_GUTS;
15235 npar = r->nparens+1;
15236 Newx(ret->offs, npar, regexp_paren_pair);
15237 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
15239 if (ret->substrs) {
15240 /* Do it this way to avoid reading from *r after the StructCopy().
15241 That way, if any of the sv_dup_inc()s dislodge *r from the L1
15242 cache, it doesn't matter. */
15243 const bool anchored = r->check_substr
15244 ? r->check_substr == r->anchored_substr
15245 : r->check_utf8 == r->anchored_utf8;
15246 Newx(ret->substrs, 1, struct reg_substr_data);
15247 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
15249 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
15250 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
15251 ret->float_substr = sv_dup_inc(ret->float_substr, param);
15252 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
15254 /* check_substr and check_utf8, if non-NULL, point to either their
15255 anchored or float namesakes, and don't hold a second reference. */
15257 if (ret->check_substr) {
15259 assert(r->check_utf8 == r->anchored_utf8);
15260 ret->check_substr = ret->anchored_substr;
15261 ret->check_utf8 = ret->anchored_utf8;
15263 assert(r->check_substr == r->float_substr);
15264 assert(r->check_utf8 == r->float_utf8);
15265 ret->check_substr = ret->float_substr;
15266 ret->check_utf8 = ret->float_utf8;
15268 } else if (ret->check_utf8) {
15270 ret->check_utf8 = ret->anchored_utf8;
15272 ret->check_utf8 = ret->float_utf8;
15277 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
15278 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
15281 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
15283 if (RX_MATCH_COPIED(dstr))
15284 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
15286 ret->subbeg = NULL;
15287 #ifdef PERL_ANY_COW
15288 ret->saved_copy = NULL;
15291 /* Whether mother_re be set or no, we need to copy the string. We
15292 cannot refrain from copying it when the storage points directly to
15293 our mother regexp, because that's
15294 1: a buffer in a different thread
15295 2: something we no longer hold a reference on
15296 so we need to copy it locally. */
15297 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
15298 ret->mother_re = NULL;
15300 #endif /* PERL_IN_XSUB_RE */
15305 This is the internal complement to regdupe() which is used to copy
15306 the structure pointed to by the *pprivate pointer in the regexp.
15307 This is the core version of the extension overridable cloning hook.
15308 The regexp structure being duplicated will be copied by perl prior
15309 to this and will be provided as the regexp *r argument, however
15310 with the /old/ structures pprivate pointer value. Thus this routine
15311 may override any copying normally done by perl.
15313 It returns a pointer to the new regexp_internal structure.
15317 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
15320 struct regexp *const r = ReANY(rx);
15321 regexp_internal *reti;
15323 RXi_GET_DECL(r,ri);
15325 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
15329 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
15330 Copy(ri->program, reti->program, len+1, regnode);
15332 reti->num_code_blocks = ri->num_code_blocks;
15333 if (ri->code_blocks) {
15335 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
15336 struct reg_code_block);
15337 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
15338 struct reg_code_block);
15339 for (n = 0; n < ri->num_code_blocks; n++)
15340 reti->code_blocks[n].src_regex = (REGEXP*)
15341 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
15344 reti->code_blocks = NULL;
15346 reti->regstclass = NULL;
15349 struct reg_data *d;
15350 const int count = ri->data->count;
15353 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
15354 char, struct reg_data);
15355 Newx(d->what, count, U8);
15358 for (i = 0; i < count; i++) {
15359 d->what[i] = ri->data->what[i];
15360 switch (d->what[i]) {
15361 /* see also regcomp.h and regfree_internal() */
15362 case 'a': /* actually an AV, but the dup function is identical. */
15366 case 'u': /* actually an HV, but the dup function is identical. */
15367 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
15370 /* This is cheating. */
15371 Newx(d->data[i], 1, struct regnode_charclass_class);
15372 StructCopy(ri->data->data[i], d->data[i],
15373 struct regnode_charclass_class);
15374 reti->regstclass = (regnode*)d->data[i];
15377 /* Trie stclasses are readonly and can thus be shared
15378 * without duplication. We free the stclass in pregfree
15379 * when the corresponding reg_ac_data struct is freed.
15381 reti->regstclass= ri->regstclass;
15385 ((reg_trie_data*)ri->data->data[i])->refcount++;
15390 d->data[i] = ri->data->data[i];
15393 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
15402 reti->name_list_idx = ri->name_list_idx;
15404 #ifdef RE_TRACK_PATTERN_OFFSETS
15405 if (ri->u.offsets) {
15406 Newx(reti->u.offsets, 2*len+1, U32);
15407 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
15410 SetProgLen(reti,len);
15413 return (void*)reti;
15416 #endif /* USE_ITHREADS */
15418 #ifndef PERL_IN_XSUB_RE
15421 - regnext - dig the "next" pointer out of a node
15424 Perl_regnext(pTHX_ regnode *p)
15432 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
15433 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
15436 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
15445 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
15448 STRLEN l1 = strlen(pat1);
15449 STRLEN l2 = strlen(pat2);
15452 const char *message;
15454 PERL_ARGS_ASSERT_RE_CROAK2;
15460 Copy(pat1, buf, l1 , char);
15461 Copy(pat2, buf + l1, l2 , char);
15462 buf[l1 + l2] = '\n';
15463 buf[l1 + l2 + 1] = '\0';
15465 /* ANSI variant takes additional second argument */
15466 va_start(args, pat2);
15470 msv = vmess(buf, &args);
15472 message = SvPV_const(msv,l1);
15475 Copy(message, buf, l1 , char);
15476 buf[l1-1] = '\0'; /* Overwrite \n */
15477 Perl_croak(aTHX_ "%s", buf);
15480 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
15482 #ifndef PERL_IN_XSUB_RE
15484 Perl_save_re_context(pTHX)
15488 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
15490 const REGEXP * const rx = PM_GETRE(PL_curpm);
15493 for (i = 1; i <= RX_NPARENS(rx); i++) {
15494 char digits[TYPE_CHARS(long)];
15495 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
15496 GV *const *const gvp
15497 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
15500 GV * const gv = *gvp;
15501 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
15513 S_put_byte(pTHX_ SV *sv, int c)
15515 PERL_ARGS_ASSERT_PUT_BYTE;
15517 /* Our definition of isPRINT() ignores locales, so only bytes that are
15518 not part of UTF-8 are considered printable. I assume that the same
15519 holds for UTF-EBCDIC.
15520 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
15521 which Wikipedia says:
15523 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
15524 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
15525 identical, to the ASCII delete (DEL) or rubout control character. ...
15526 it is typically mapped to hexadecimal code 9F, in order to provide a
15527 unique character mapping in both directions)
15529 So the old condition can be simplified to !isPRINT(c) */
15532 case '\r': Perl_sv_catpvf(aTHX_ sv, "\\r"); break;
15533 case '\n': Perl_sv_catpvf(aTHX_ sv, "\\n"); break;
15534 case '\t': Perl_sv_catpvf(aTHX_ sv, "\\t"); break;
15535 case '\f': Perl_sv_catpvf(aTHX_ sv, "\\f"); break;
15536 case '\a': Perl_sv_catpvf(aTHX_ sv, "\\a"); break;
15539 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
15544 const char string = c;
15545 if (c == '-' || c == ']' || c == '\\' || c == '^')
15546 sv_catpvs(sv, "\\");
15547 sv_catpvn(sv, &string, 1);
15552 S_put_latin1_charclass_innards(pTHX_ SV *sv, char *bitmap)
15554 /* Appends to 'sv' a displayable version of the innards of the bracketed
15555 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
15556 * output anything */
15559 int rangestart = -1;
15560 bool has_output_anything = FALSE;
15562 PERL_ARGS_ASSERT_PUT_LATIN1_CHARCLASS_INNARDS;
15564 for (i = 0; i <= 256; i++) {
15565 if (i < 256 && BITMAP_TEST((U8 *) bitmap,i)) {
15566 if (rangestart == -1)
15568 } else if (rangestart != -1) {
15570 if (i <= rangestart + 3) { /* Individual chars in short ranges */
15571 for (; rangestart < i; rangestart++)
15572 put_byte(sv, rangestart);
15575 || ! isALPHANUMERIC(rangestart)
15576 || ! isALPHANUMERIC(j)
15577 || isDIGIT(rangestart) != isDIGIT(j)
15578 || isUPPER(rangestart) != isUPPER(j)
15579 || isLOWER(rangestart) != isLOWER(j)
15581 /* This final test should get optimized out except
15582 * on EBCDIC platforms, where it causes ranges that
15583 * cross discontinuities like i/j to be shown as hex
15584 * instead of the misleading, e.g. H-K (since that
15585 * range includes more than H, I, J, K). */
15586 || (j - rangestart)
15587 != NATIVE_TO_ASCII(j) - NATIVE_TO_ASCII(rangestart))
15589 Perl_sv_catpvf(aTHX_ sv, "\\x{%02x}-\\x{%02x}",
15591 (j < 256) ? j : 255);
15593 else { /* Here, the ends of the range are both digits, or both
15594 uppercase, or both lowercase; and there's no
15595 discontinuity in the range (which could happen on EBCDIC
15597 put_byte(sv, rangestart);
15598 sv_catpvs(sv, "-");
15602 has_output_anything = TRUE;
15606 return has_output_anything;
15609 #define CLEAR_OPTSTART \
15610 if (optstart) STMT_START { \
15611 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
15615 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
15617 STATIC const regnode *
15618 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
15619 const regnode *last, const regnode *plast,
15620 SV* sv, I32 indent, U32 depth)
15623 U8 op = PSEUDO; /* Arbitrary non-END op. */
15624 const regnode *next;
15625 const regnode *optstart= NULL;
15627 RXi_GET_DECL(r,ri);
15628 GET_RE_DEBUG_FLAGS_DECL;
15630 PERL_ARGS_ASSERT_DUMPUNTIL;
15632 #ifdef DEBUG_DUMPUNTIL
15633 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
15634 last ? last-start : 0,plast ? plast-start : 0);
15637 if (plast && plast < last)
15640 while (PL_regkind[op] != END && (!last || node < last)) {
15641 /* While that wasn't END last time... */
15644 if (op == CLOSE || op == WHILEM)
15646 next = regnext((regnode *)node);
15649 if (OP(node) == OPTIMIZED) {
15650 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
15657 regprop(r, sv, node);
15658 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
15659 (int)(2*indent + 1), "", SvPVX_const(sv));
15661 if (OP(node) != OPTIMIZED) {
15662 if (next == NULL) /* Next ptr. */
15663 PerlIO_printf(Perl_debug_log, " (0)");
15664 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
15665 PerlIO_printf(Perl_debug_log, " (FAIL)");
15667 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
15668 (void)PerlIO_putc(Perl_debug_log, '\n');
15672 if (PL_regkind[(U8)op] == BRANCHJ) {
15675 const regnode *nnode = (OP(next) == LONGJMP
15676 ? regnext((regnode *)next)
15678 if (last && nnode > last)
15680 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
15683 else if (PL_regkind[(U8)op] == BRANCH) {
15685 DUMPUNTIL(NEXTOPER(node), next);
15687 else if ( PL_regkind[(U8)op] == TRIE ) {
15688 const regnode *this_trie = node;
15689 const char op = OP(node);
15690 const U32 n = ARG(node);
15691 const reg_ac_data * const ac = op>=AHOCORASICK ?
15692 (reg_ac_data *)ri->data->data[n] :
15694 const reg_trie_data * const trie =
15695 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
15697 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
15699 const regnode *nextbranch= NULL;
15702 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
15703 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
15705 PerlIO_printf(Perl_debug_log, "%*s%s ",
15706 (int)(2*(indent+3)), "",
15707 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
15708 PL_colors[0], PL_colors[1],
15709 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
15710 PERL_PV_PRETTY_ELLIPSES |
15711 PERL_PV_PRETTY_LTGT
15716 U16 dist= trie->jump[word_idx+1];
15717 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
15718 (UV)((dist ? this_trie + dist : next) - start));
15721 nextbranch= this_trie + trie->jump[0];
15722 DUMPUNTIL(this_trie + dist, nextbranch);
15724 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
15725 nextbranch= regnext((regnode *)nextbranch);
15727 PerlIO_printf(Perl_debug_log, "\n");
15730 if (last && next > last)
15735 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
15736 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
15737 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
15739 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
15741 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
15743 else if ( op == PLUS || op == STAR) {
15744 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
15746 else if (PL_regkind[(U8)op] == ANYOF) {
15747 /* arglen 1 + class block */
15748 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
15749 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
15750 node = NEXTOPER(node);
15752 else if (PL_regkind[(U8)op] == EXACT) {
15753 /* Literal string, where present. */
15754 node += NODE_SZ_STR(node) - 1;
15755 node = NEXTOPER(node);
15758 node = NEXTOPER(node);
15759 node += regarglen[(U8)op];
15761 if (op == CURLYX || op == OPEN)
15765 #ifdef DEBUG_DUMPUNTIL
15766 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
15771 #endif /* DEBUGGING */
15775 * c-indentation-style: bsd
15776 * c-basic-offset: 4
15777 * indent-tabs-mode: nil
15780 * ex: set ts=8 sts=4 sw=4 et: