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 "utf8_strings.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)
102 # if defined(BUGGY_MSC6)
103 /* MSC 6.00A breaks on op/regexp.t test 85 unless we turn this off */
104 # pragma optimize("a",off)
105 /* But MSC 6.00A is happy with 'w', for aliases only across function calls*/
106 # pragma optimize("w",on )
107 # endif /* BUGGY_MSC6 */
111 #define STATIC static
115 typedef struct RExC_state_t {
116 U32 flags; /* RXf_* are we folding, multilining? */
117 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
118 char *precomp; /* uncompiled string. */
119 REGEXP *rx_sv; /* The SV that is the regexp. */
120 regexp *rx; /* perl core regexp structure */
121 regexp_internal *rxi; /* internal data for regexp object pprivate field */
122 char *start; /* Start of input for compile */
123 char *end; /* End of input for compile */
124 char *parse; /* Input-scan pointer. */
125 I32 whilem_seen; /* number of WHILEM in this expr */
126 regnode *emit_start; /* Start of emitted-code area */
127 regnode *emit_bound; /* First regnode outside of the allocated space */
128 regnode *emit; /* Code-emit pointer; ®dummy = don't = compiling */
129 I32 naughty; /* How bad is this pattern? */
130 I32 sawback; /* Did we see \1, ...? */
132 I32 size; /* Code size. */
133 I32 npar; /* Capture buffer count, (OPEN). */
134 I32 cpar; /* Capture buffer count, (CLOSE). */
135 I32 nestroot; /* root parens we are in - used by accept */
138 regnode **open_parens; /* pointers to open parens */
139 regnode **close_parens; /* pointers to close parens */
140 regnode *opend; /* END node in program */
141 I32 utf8; /* whether the pattern is utf8 or not */
142 I32 orig_utf8; /* whether the pattern was originally in utf8 */
143 /* XXX use this for future optimisation of case
144 * where pattern must be upgraded to utf8. */
145 I32 uni_semantics; /* If a d charset modifier should use unicode
146 rules, even if the pattern is not in
148 HV *paren_names; /* Paren names */
150 regnode **recurse; /* Recurse regops */
151 I32 recurse_count; /* Number of recurse regops */
154 I32 override_recoding;
155 struct reg_code_block *code_blocks; /* positions of literal (?{})
157 int num_code_blocks; /* size of code_blocks[] */
158 int code_index; /* next code_blocks[] slot */
160 char *starttry; /* -Dr: where regtry was called. */
161 #define RExC_starttry (pRExC_state->starttry)
163 SV *runtime_code_qr; /* qr with the runtime code blocks */
165 const char *lastparse;
167 AV *paren_name_list; /* idx -> name */
168 #define RExC_lastparse (pRExC_state->lastparse)
169 #define RExC_lastnum (pRExC_state->lastnum)
170 #define RExC_paren_name_list (pRExC_state->paren_name_list)
174 #define RExC_flags (pRExC_state->flags)
175 #define RExC_pm_flags (pRExC_state->pm_flags)
176 #define RExC_precomp (pRExC_state->precomp)
177 #define RExC_rx_sv (pRExC_state->rx_sv)
178 #define RExC_rx (pRExC_state->rx)
179 #define RExC_rxi (pRExC_state->rxi)
180 #define RExC_start (pRExC_state->start)
181 #define RExC_end (pRExC_state->end)
182 #define RExC_parse (pRExC_state->parse)
183 #define RExC_whilem_seen (pRExC_state->whilem_seen)
184 #ifdef RE_TRACK_PATTERN_OFFSETS
185 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the others */
187 #define RExC_emit (pRExC_state->emit)
188 #define RExC_emit_start (pRExC_state->emit_start)
189 #define RExC_emit_bound (pRExC_state->emit_bound)
190 #define RExC_naughty (pRExC_state->naughty)
191 #define RExC_sawback (pRExC_state->sawback)
192 #define RExC_seen (pRExC_state->seen)
193 #define RExC_size (pRExC_state->size)
194 #define RExC_npar (pRExC_state->npar)
195 #define RExC_nestroot (pRExC_state->nestroot)
196 #define RExC_extralen (pRExC_state->extralen)
197 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
198 #define RExC_utf8 (pRExC_state->utf8)
199 #define RExC_uni_semantics (pRExC_state->uni_semantics)
200 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
201 #define RExC_open_parens (pRExC_state->open_parens)
202 #define RExC_close_parens (pRExC_state->close_parens)
203 #define RExC_opend (pRExC_state->opend)
204 #define RExC_paren_names (pRExC_state->paren_names)
205 #define RExC_recurse (pRExC_state->recurse)
206 #define RExC_recurse_count (pRExC_state->recurse_count)
207 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
208 #define RExC_contains_locale (pRExC_state->contains_locale)
209 #define RExC_override_recoding (pRExC_state->override_recoding)
212 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
213 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
214 ((*s) == '{' && regcurly(s)))
217 #undef SPSTART /* dratted cpp namespace... */
220 * Flags to be passed up and down.
222 #define WORST 0 /* Worst case. */
223 #define HASWIDTH 0x01 /* Known to match non-null strings. */
225 /* Simple enough to be STAR/PLUS operand; in an EXACT node must be a single
226 * character, and if utf8, must be invariant. Note that this is not the same
227 * thing as REGNODE_SIMPLE */
229 #define SPSTART 0x04 /* Starts with * or +. */
230 #define TRYAGAIN 0x08 /* Weeded out a declaration. */
231 #define POSTPONED 0x10 /* (?1),(?&name), (??{...}) or similar */
233 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
235 /* whether trie related optimizations are enabled */
236 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
237 #define TRIE_STUDY_OPT
238 #define FULL_TRIE_STUDY
244 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
245 #define PBITVAL(paren) (1 << ((paren) & 7))
246 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
247 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
248 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
250 /* If not already in utf8, do a longjmp back to the beginning */
251 #define UTF8_LONGJMP 42 /* Choose a value not likely to ever conflict */
252 #define REQUIRE_UTF8 STMT_START { \
253 if (! UTF) JMPENV_JUMP(UTF8_LONGJMP); \
256 /* About scan_data_t.
258 During optimisation we recurse through the regexp program performing
259 various inplace (keyhole style) optimisations. In addition study_chunk
260 and scan_commit populate this data structure with information about
261 what strings MUST appear in the pattern. We look for the longest
262 string that must appear at a fixed location, and we look for the
263 longest string that may appear at a floating location. So for instance
268 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
269 strings (because they follow a .* construct). study_chunk will identify
270 both FOO and BAR as being the longest fixed and floating strings respectively.
272 The strings can be composites, for instance
276 will result in a composite fixed substring 'foo'.
278 For each string some basic information is maintained:
280 - offset or min_offset
281 This is the position the string must appear at, or not before.
282 It also implicitly (when combined with minlenp) tells us how many
283 characters must match before the string we are searching for.
284 Likewise when combined with minlenp and the length of the string it
285 tells us how many characters must appear after the string we have
289 Only used for floating strings. This is the rightmost point that
290 the string can appear at. If set to I32 max it indicates that the
291 string can occur infinitely far to the right.
294 A pointer to the minimum length of the pattern that the string
295 was found inside. This is important as in the case of positive
296 lookahead or positive lookbehind we can have multiple patterns
301 The minimum length of the pattern overall is 3, the minimum length
302 of the lookahead part is 3, but the minimum length of the part that
303 will actually match is 1. So 'FOO's minimum length is 3, but the
304 minimum length for the F is 1. This is important as the minimum length
305 is used to determine offsets in front of and behind the string being
306 looked for. Since strings can be composites this is the length of the
307 pattern at the time it was committed with a scan_commit. Note that
308 the length is calculated by study_chunk, so that the minimum lengths
309 are not known until the full pattern has been compiled, thus the
310 pointer to the value.
314 In the case of lookbehind the string being searched for can be
315 offset past the start point of the final matching string.
316 If this value was just blithely removed from the min_offset it would
317 invalidate some of the calculations for how many chars must match
318 before or after (as they are derived from min_offset and minlen and
319 the length of the string being searched for).
320 When the final pattern is compiled and the data is moved from the
321 scan_data_t structure into the regexp structure the information
322 about lookbehind is factored in, with the information that would
323 have been lost precalculated in the end_shift field for the
326 The fields pos_min and pos_delta are used to store the minimum offset
327 and the delta to the maximum offset at the current point in the pattern.
331 typedef struct scan_data_t {
332 /*I32 len_min; unused */
333 /*I32 len_delta; unused */
337 I32 last_end; /* min value, <0 unless valid. */
340 SV **longest; /* Either &l_fixed, or &l_float. */
341 SV *longest_fixed; /* longest fixed string found in pattern */
342 I32 offset_fixed; /* offset where it starts */
343 I32 *minlen_fixed; /* pointer to the minlen relevant to the string */
344 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
345 SV *longest_float; /* longest floating string found in pattern */
346 I32 offset_float_min; /* earliest point in string it can appear */
347 I32 offset_float_max; /* latest point in string it can appear */
348 I32 *minlen_float; /* pointer to the minlen relevant to the string */
349 I32 lookbehind_float; /* is the position of the string modified by LB */
353 struct regnode_charclass_class *start_class;
357 * Forward declarations for pregcomp()'s friends.
360 static const scan_data_t zero_scan_data =
361 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
363 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
364 #define SF_BEFORE_SEOL 0x0001
365 #define SF_BEFORE_MEOL 0x0002
366 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
367 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
370 # define SF_FIX_SHIFT_EOL (0+2)
371 # define SF_FL_SHIFT_EOL (0+4)
373 # define SF_FIX_SHIFT_EOL (+2)
374 # define SF_FL_SHIFT_EOL (+4)
377 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
378 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
380 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
381 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
382 #define SF_IS_INF 0x0040
383 #define SF_HAS_PAR 0x0080
384 #define SF_IN_PAR 0x0100
385 #define SF_HAS_EVAL 0x0200
386 #define SCF_DO_SUBSTR 0x0400
387 #define SCF_DO_STCLASS_AND 0x0800
388 #define SCF_DO_STCLASS_OR 0x1000
389 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
390 #define SCF_WHILEM_VISITED_POS 0x2000
392 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
393 #define SCF_SEEN_ACCEPT 0x8000
395 #define UTF cBOOL(RExC_utf8)
397 /* The enums for all these are ordered so things work out correctly */
398 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
399 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_DEPENDS_CHARSET)
400 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
401 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) >= REGEX_UNICODE_CHARSET)
402 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_RESTRICTED_CHARSET)
403 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) >= REGEX_ASCII_RESTRICTED_CHARSET)
404 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
406 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
408 #define OOB_NAMEDCLASS -1
410 /* There is no code point that is out-of-bounds, so this is problematic. But
411 * its only current use is to initialize a variable that is always set before
413 #define OOB_UNICODE 0xDEADBEEF
415 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
416 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
419 /* length of regex to show in messages that don't mark a position within */
420 #define RegexLengthToShowInErrorMessages 127
423 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
424 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
425 * op/pragma/warn/regcomp.
427 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
428 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
430 #define REPORT_LOCATION " in regex; marked by " MARKER1 " in m/%.*s" MARKER2 "%s/"
433 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
434 * arg. Show regex, up to a maximum length. If it's too long, chop and add
437 #define _FAIL(code) STMT_START { \
438 const char *ellipses = ""; \
439 IV len = RExC_end - RExC_precomp; \
442 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
443 if (len > RegexLengthToShowInErrorMessages) { \
444 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
445 len = RegexLengthToShowInErrorMessages - 10; \
451 #define FAIL(msg) _FAIL( \
452 Perl_croak(aTHX_ "%s in regex m/%.*s%s/", \
453 msg, (int)len, RExC_precomp, ellipses))
455 #define FAIL2(msg,arg) _FAIL( \
456 Perl_croak(aTHX_ msg " in regex m/%.*s%s/", \
457 arg, (int)len, RExC_precomp, ellipses))
460 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
462 #define Simple_vFAIL(m) STMT_START { \
463 const IV offset = RExC_parse - RExC_precomp; \
464 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
465 m, (int)offset, RExC_precomp, RExC_precomp + offset); \
469 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
471 #define vFAIL(m) STMT_START { \
473 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
478 * Like Simple_vFAIL(), but accepts two arguments.
480 #define Simple_vFAIL2(m,a1) STMT_START { \
481 const IV offset = RExC_parse - RExC_precomp; \
482 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, \
483 (int)offset, RExC_precomp, RExC_precomp + offset); \
487 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
489 #define vFAIL2(m,a1) STMT_START { \
491 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
492 Simple_vFAIL2(m, a1); \
497 * Like Simple_vFAIL(), but accepts three arguments.
499 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
500 const IV offset = RExC_parse - RExC_precomp; \
501 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, \
502 (int)offset, RExC_precomp, RExC_precomp + offset); \
506 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
508 #define vFAIL3(m,a1,a2) STMT_START { \
510 SAVEDESTRUCTOR_X(clear_re,(void*)RExC_rx_sv); \
511 Simple_vFAIL3(m, a1, a2); \
515 * Like Simple_vFAIL(), but accepts four arguments.
517 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
518 const IV offset = RExC_parse - RExC_precomp; \
519 S_re_croak2(aTHX_ m, REPORT_LOCATION, a1, a2, a3, \
520 (int)offset, RExC_precomp, RExC_precomp + offset); \
523 #define ckWARNreg(loc,m) STMT_START { \
524 const IV offset = loc - RExC_precomp; \
525 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
526 (int)offset, RExC_precomp, RExC_precomp + offset); \
529 #define ckWARNregdep(loc,m) STMT_START { \
530 const IV offset = loc - RExC_precomp; \
531 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
533 (int)offset, RExC_precomp, RExC_precomp + offset); \
536 #define ckWARN2regdep(loc,m, a1) STMT_START { \
537 const IV offset = loc - RExC_precomp; \
538 Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
540 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
543 #define ckWARN2reg(loc, m, a1) STMT_START { \
544 const IV offset = loc - RExC_precomp; \
545 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
546 a1, (int)offset, RExC_precomp, RExC_precomp + offset); \
549 #define vWARN3(loc, m, a1, a2) STMT_START { \
550 const IV offset = loc - RExC_precomp; \
551 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
552 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
555 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
556 const IV offset = loc - RExC_precomp; \
557 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
558 a1, a2, (int)offset, RExC_precomp, RExC_precomp + offset); \
561 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
562 const IV offset = loc - RExC_precomp; \
563 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
564 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
567 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
568 const IV offset = loc - RExC_precomp; \
569 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
570 a1, a2, a3, (int)offset, RExC_precomp, RExC_precomp + offset); \
573 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
574 const IV offset = loc - RExC_precomp; \
575 Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
576 a1, a2, a3, a4, (int)offset, RExC_precomp, RExC_precomp + offset); \
580 /* Allow for side effects in s */
581 #define REGC(c,s) STMT_START { \
582 if (!SIZE_ONLY) *(s) = (c); else (void)(s); \
585 /* Macros for recording node offsets. 20001227 mjd@plover.com
586 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
587 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
588 * Element 0 holds the number n.
589 * Position is 1 indexed.
591 #ifndef RE_TRACK_PATTERN_OFFSETS
592 #define Set_Node_Offset_To_R(node,byte)
593 #define Set_Node_Offset(node,byte)
594 #define Set_Cur_Node_Offset
595 #define Set_Node_Length_To_R(node,len)
596 #define Set_Node_Length(node,len)
597 #define Set_Node_Cur_Length(node)
598 #define Node_Offset(n)
599 #define Node_Length(n)
600 #define Set_Node_Offset_Length(node,offset,len)
601 #define ProgLen(ri) ri->u.proglen
602 #define SetProgLen(ri,x) ri->u.proglen = x
604 #define ProgLen(ri) ri->u.offsets[0]
605 #define SetProgLen(ri,x) ri->u.offsets[0] = x
606 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
608 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
609 __LINE__, (int)(node), (int)(byte))); \
611 Perl_croak(aTHX_ "value of node is %d in Offset macro", (int)(node)); \
613 RExC_offsets[2*(node)-1] = (byte); \
618 #define Set_Node_Offset(node,byte) \
619 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
620 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
622 #define Set_Node_Length_To_R(node,len) STMT_START { \
624 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
625 __LINE__, (int)(node), (int)(len))); \
627 Perl_croak(aTHX_ "value of node is %d in Length macro", (int)(node)); \
629 RExC_offsets[2*(node)] = (len); \
634 #define Set_Node_Length(node,len) \
635 Set_Node_Length_To_R((node)-RExC_emit_start, len)
636 #define Set_Cur_Node_Length(len) Set_Node_Length(RExC_emit, len)
637 #define Set_Node_Cur_Length(node) \
638 Set_Node_Length(node, RExC_parse - parse_start)
640 /* Get offsets and lengths */
641 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
642 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
644 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
645 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
646 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
650 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
651 #define EXPERIMENTAL_INPLACESCAN
652 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
654 #define DEBUG_STUDYDATA(str,data,depth) \
655 DEBUG_OPTIMISE_MORE_r(if(data){ \
656 PerlIO_printf(Perl_debug_log, \
657 "%*s" str "Pos:%"IVdf"/%"IVdf \
658 " Flags: 0x%"UVXf" Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
659 (int)(depth)*2, "", \
660 (IV)((data)->pos_min), \
661 (IV)((data)->pos_delta), \
662 (UV)((data)->flags), \
663 (IV)((data)->whilem_c), \
664 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
665 is_inf ? "INF " : "" \
667 if ((data)->last_found) \
668 PerlIO_printf(Perl_debug_log, \
669 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
670 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
671 SvPVX_const((data)->last_found), \
672 (IV)((data)->last_end), \
673 (IV)((data)->last_start_min), \
674 (IV)((data)->last_start_max), \
675 ((data)->longest && \
676 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
677 SvPVX_const((data)->longest_fixed), \
678 (IV)((data)->offset_fixed), \
679 ((data)->longest && \
680 (data)->longest==&((data)->longest_float)) ? "*" : "", \
681 SvPVX_const((data)->longest_float), \
682 (IV)((data)->offset_float_min), \
683 (IV)((data)->offset_float_max) \
685 PerlIO_printf(Perl_debug_log,"\n"); \
688 static void clear_re(pTHX_ void *r);
690 /* Mark that we cannot extend a found fixed substring at this point.
691 Update the longest found anchored substring and the longest found
692 floating substrings if needed. */
695 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data, I32 *minlenp, int is_inf)
697 const STRLEN l = CHR_SVLEN(data->last_found);
698 const STRLEN old_l = CHR_SVLEN(*data->longest);
699 GET_RE_DEBUG_FLAGS_DECL;
701 PERL_ARGS_ASSERT_SCAN_COMMIT;
703 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
704 SvSetMagicSV(*data->longest, data->last_found);
705 if (*data->longest == data->longest_fixed) {
706 data->offset_fixed = l ? data->last_start_min : data->pos_min;
707 if (data->flags & SF_BEFORE_EOL)
709 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
711 data->flags &= ~SF_FIX_BEFORE_EOL;
712 data->minlen_fixed=minlenp;
713 data->lookbehind_fixed=0;
715 else { /* *data->longest == data->longest_float */
716 data->offset_float_min = l ? data->last_start_min : data->pos_min;
717 data->offset_float_max = (l
718 ? data->last_start_max
719 : data->pos_min + data->pos_delta);
720 if (is_inf || (U32)data->offset_float_max > (U32)I32_MAX)
721 data->offset_float_max = I32_MAX;
722 if (data->flags & SF_BEFORE_EOL)
724 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
726 data->flags &= ~SF_FL_BEFORE_EOL;
727 data->minlen_float=minlenp;
728 data->lookbehind_float=0;
731 SvCUR_set(data->last_found, 0);
733 SV * const sv = data->last_found;
734 if (SvUTF8(sv) && SvMAGICAL(sv)) {
735 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
741 data->flags &= ~SF_BEFORE_EOL;
742 DEBUG_STUDYDATA("commit: ",data,0);
745 /* Can match anything (initialization) */
747 S_cl_anything(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
749 PERL_ARGS_ASSERT_CL_ANYTHING;
751 ANYOF_BITMAP_SETALL(cl);
752 cl->flags = ANYOF_CLASS|ANYOF_EOS|ANYOF_UNICODE_ALL
753 |ANYOF_LOC_NONBITMAP_FOLD|ANYOF_NON_UTF8_LATIN1_ALL;
755 /* If any portion of the regex is to operate under locale rules,
756 * initialization includes it. The reason this isn't done for all regexes
757 * is that the optimizer was written under the assumption that locale was
758 * all-or-nothing. Given the complexity and lack of documentation in the
759 * optimizer, and that there are inadequate test cases for locale, so many
760 * parts of it may not work properly, it is safest to avoid locale unless
762 if (RExC_contains_locale) {
763 ANYOF_CLASS_SETALL(cl); /* /l uses class */
764 cl->flags |= ANYOF_LOCALE;
767 ANYOF_CLASS_ZERO(cl); /* Only /l uses class now */
771 /* Can match anything (initialization) */
773 S_cl_is_anything(const struct regnode_charclass_class *cl)
777 PERL_ARGS_ASSERT_CL_IS_ANYTHING;
779 for (value = 0; value <= ANYOF_MAX; value += 2)
780 if (ANYOF_CLASS_TEST(cl, value) && ANYOF_CLASS_TEST(cl, value + 1))
782 if (!(cl->flags & ANYOF_UNICODE_ALL))
784 if (!ANYOF_BITMAP_TESTALLSET((const void*)cl))
789 /* Can match anything (initialization) */
791 S_cl_init(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl)
793 PERL_ARGS_ASSERT_CL_INIT;
795 Zero(cl, 1, struct regnode_charclass_class);
797 cl_anything(pRExC_state, cl);
798 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
801 /* These two functions currently do the exact same thing */
802 #define cl_init_zero S_cl_init
804 /* 'AND' a given class with another one. Can create false positives. 'cl'
805 * should not be inverted. 'and_with->flags & ANYOF_CLASS' should be 0 if
806 * 'and_with' is a regnode_charclass instead of a regnode_charclass_class. */
808 S_cl_and(struct regnode_charclass_class *cl,
809 const struct regnode_charclass_class *and_with)
811 PERL_ARGS_ASSERT_CL_AND;
813 assert(and_with->type == ANYOF);
815 /* I (khw) am not sure all these restrictions are necessary XXX */
816 if (!(ANYOF_CLASS_TEST_ANY_SET(and_with))
817 && !(ANYOF_CLASS_TEST_ANY_SET(cl))
818 && (and_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
819 && !(and_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
820 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) {
823 if (and_with->flags & ANYOF_INVERT)
824 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
825 cl->bitmap[i] &= ~and_with->bitmap[i];
827 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
828 cl->bitmap[i] &= and_with->bitmap[i];
829 } /* XXXX: logic is complicated otherwise, leave it along for a moment. */
831 if (and_with->flags & ANYOF_INVERT) {
833 /* Here, the and'ed node is inverted. Get the AND of the flags that
834 * aren't affected by the inversion. Those that are affected are
835 * handled individually below */
836 U8 affected_flags = cl->flags & ~INVERSION_UNAFFECTED_FLAGS;
837 cl->flags &= (and_with->flags & INVERSION_UNAFFECTED_FLAGS);
838 cl->flags |= affected_flags;
840 /* We currently don't know how to deal with things that aren't in the
841 * bitmap, but we know that the intersection is no greater than what
842 * is already in cl, so let there be false positives that get sorted
843 * out after the synthetic start class succeeds, and the node is
844 * matched for real. */
846 /* The inversion of these two flags indicate that the resulting
847 * intersection doesn't have them */
848 if (and_with->flags & ANYOF_UNICODE_ALL) {
849 cl->flags &= ~ANYOF_UNICODE_ALL;
851 if (and_with->flags & ANYOF_NON_UTF8_LATIN1_ALL) {
852 cl->flags &= ~ANYOF_NON_UTF8_LATIN1_ALL;
855 else { /* and'd node is not inverted */
856 U8 outside_bitmap_but_not_utf8; /* Temp variable */
858 if (! ANYOF_NONBITMAP(and_with)) {
860 /* Here 'and_with' doesn't match anything outside the bitmap
861 * (except possibly ANYOF_UNICODE_ALL), which means the
862 * intersection can't either, except for ANYOF_UNICODE_ALL, in
863 * which case we don't know what the intersection is, but it's no
864 * greater than what cl already has, so can just leave it alone,
865 * with possible false positives */
866 if (! (and_with->flags & ANYOF_UNICODE_ALL)) {
867 ARG_SET(cl, ANYOF_NONBITMAP_EMPTY);
868 cl->flags &= ~ANYOF_NONBITMAP_NON_UTF8;
871 else if (! ANYOF_NONBITMAP(cl)) {
873 /* Here, 'and_with' does match something outside the bitmap, and cl
874 * doesn't have a list of things to match outside the bitmap. If
875 * cl can match all code points above 255, the intersection will
876 * be those above-255 code points that 'and_with' matches. If cl
877 * can't match all Unicode code points, it means that it can't
878 * match anything outside the bitmap (since the 'if' that got us
879 * into this block tested for that), so we leave the bitmap empty.
881 if (cl->flags & ANYOF_UNICODE_ALL) {
882 ARG_SET(cl, ARG(and_with));
884 /* and_with's ARG may match things that don't require UTF8.
885 * And now cl's will too, in spite of this being an 'and'. See
886 * the comments below about the kludge */
887 cl->flags |= and_with->flags & ANYOF_NONBITMAP_NON_UTF8;
891 /* Here, both 'and_with' and cl match something outside the
892 * bitmap. Currently we do not do the intersection, so just match
893 * whatever cl had at the beginning. */
897 /* Take the intersection of the two sets of flags. However, the
898 * ANYOF_NONBITMAP_NON_UTF8 flag is treated as an 'or'. This is a
899 * kludge around the fact that this flag is not treated like the others
900 * which are initialized in cl_anything(). The way the optimizer works
901 * is that the synthetic start class (SSC) is initialized to match
902 * anything, and then the first time a real node is encountered, its
903 * values are AND'd with the SSC's with the result being the values of
904 * the real node. However, there are paths through the optimizer where
905 * the AND never gets called, so those initialized bits are set
906 * inappropriately, which is not usually a big deal, as they just cause
907 * false positives in the SSC, which will just mean a probably
908 * imperceptible slow down in execution. However this bit has a
909 * higher false positive consequence in that it can cause utf8.pm,
910 * utf8_heavy.pl ... to be loaded when not necessary, which is a much
911 * bigger slowdown and also causes significant extra memory to be used.
912 * In order to prevent this, the code now takes a different tack. The
913 * bit isn't set unless some part of the regular expression needs it,
914 * but once set it won't get cleared. This means that these extra
915 * modules won't get loaded unless there was some path through the
916 * pattern that would have required them anyway, and so any false
917 * positives that occur by not ANDing them out when they could be
918 * aren't as severe as they would be if we treated this bit like all
920 outside_bitmap_but_not_utf8 = (cl->flags | and_with->flags)
921 & ANYOF_NONBITMAP_NON_UTF8;
922 cl->flags &= and_with->flags;
923 cl->flags |= outside_bitmap_but_not_utf8;
927 /* 'OR' a given class with another one. Can create false positives. 'cl'
928 * should not be inverted. 'or_with->flags & ANYOF_CLASS' should be 0 if
929 * 'or_with' is a regnode_charclass instead of a regnode_charclass_class. */
931 S_cl_or(const RExC_state_t *pRExC_state, struct regnode_charclass_class *cl, const struct regnode_charclass_class *or_with)
933 PERL_ARGS_ASSERT_CL_OR;
935 if (or_with->flags & ANYOF_INVERT) {
937 /* Here, the or'd node is to be inverted. This means we take the
938 * complement of everything not in the bitmap, but currently we don't
939 * know what that is, so give up and match anything */
940 if (ANYOF_NONBITMAP(or_with)) {
941 cl_anything(pRExC_state, cl);
944 * (B1 | CL1) | (!B2 & !CL2) = (B1 | !B2 & !CL2) | (CL1 | (!B2 & !CL2))
945 * <= (B1 | !B2) | (CL1 | !CL2)
946 * which is wasteful if CL2 is small, but we ignore CL2:
947 * (B1 | CL1) | (!B2 & !CL2) <= (B1 | CL1) | !B2 = (B1 | !B2) | CL1
948 * XXXX Can we handle case-fold? Unclear:
949 * (OK1(i) | OK1(i')) | !(OK1(i) | OK1(i')) =
950 * (OK1(i) | OK1(i')) | (!OK1(i) & !OK1(i'))
952 else if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
953 && !(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
954 && !(cl->flags & ANYOF_LOC_NONBITMAP_FOLD) ) {
957 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
958 cl->bitmap[i] |= ~or_with->bitmap[i];
959 } /* XXXX: logic is complicated otherwise */
961 cl_anything(pRExC_state, cl);
964 /* And, we can just take the union of the flags that aren't affected
965 * by the inversion */
966 cl->flags |= or_with->flags & INVERSION_UNAFFECTED_FLAGS;
968 /* For the remaining flags:
969 ANYOF_UNICODE_ALL and inverted means to not match anything above
970 255, which means that the union with cl should just be
971 what cl has in it, so can ignore this flag
972 ANYOF_NON_UTF8_LATIN1_ALL and inverted means if not utf8 and ord
973 is 127-255 to match them, but then invert that, so the
974 union with cl should just be what cl has in it, so can
977 } else { /* 'or_with' is not inverted */
978 /* (B1 | CL1) | (B2 | CL2) = (B1 | B2) | (CL1 | CL2)) */
979 if ( (or_with->flags & ANYOF_LOCALE) == (cl->flags & ANYOF_LOCALE)
980 && (!(or_with->flags & ANYOF_LOC_NONBITMAP_FOLD)
981 || (cl->flags & ANYOF_LOC_NONBITMAP_FOLD)) ) {
984 /* OR char bitmap and class bitmap separately */
985 for (i = 0; i < ANYOF_BITMAP_SIZE; i++)
986 cl->bitmap[i] |= or_with->bitmap[i];
987 if (ANYOF_CLASS_TEST_ANY_SET(or_with)) {
988 for (i = 0; i < ANYOF_CLASSBITMAP_SIZE; i++)
989 cl->classflags[i] |= or_with->classflags[i];
990 cl->flags |= ANYOF_CLASS;
993 else { /* XXXX: logic is complicated, leave it along for a moment. */
994 cl_anything(pRExC_state, cl);
997 if (ANYOF_NONBITMAP(or_with)) {
999 /* Use the added node's outside-the-bit-map match if there isn't a
1000 * conflict. If there is a conflict (both nodes match something
1001 * outside the bitmap, but what they match outside is not the same
1002 * pointer, and hence not easily compared until XXX we extend
1003 * inversion lists this far), give up and allow the start class to
1004 * match everything outside the bitmap. If that stuff is all above
1005 * 255, can just set UNICODE_ALL, otherwise caould be anything. */
1006 if (! ANYOF_NONBITMAP(cl)) {
1007 ARG_SET(cl, ARG(or_with));
1009 else if (ARG(cl) != ARG(or_with)) {
1011 if ((or_with->flags & ANYOF_NONBITMAP_NON_UTF8)) {
1012 cl_anything(pRExC_state, cl);
1015 cl->flags |= ANYOF_UNICODE_ALL;
1020 /* Take the union */
1021 cl->flags |= or_with->flags;
1025 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1026 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1027 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1028 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list ? (TRIE_LIST_CUR( idx ) - 1) : 0 )
1033 dump_trie(trie,widecharmap,revcharmap)
1034 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1035 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1037 These routines dump out a trie in a somewhat readable format.
1038 The _interim_ variants are used for debugging the interim
1039 tables that are used to generate the final compressed
1040 representation which is what dump_trie expects.
1042 Part of the reason for their existence is to provide a form
1043 of documentation as to how the different representations function.
1048 Dumps the final compressed table form of the trie to Perl_debug_log.
1049 Used for debugging make_trie().
1053 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1054 AV *revcharmap, U32 depth)
1057 SV *sv=sv_newmortal();
1058 int colwidth= widecharmap ? 6 : 4;
1060 GET_RE_DEBUG_FLAGS_DECL;
1062 PERL_ARGS_ASSERT_DUMP_TRIE;
1064 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1065 (int)depth * 2 + 2,"",
1066 "Match","Base","Ofs" );
1068 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1069 SV ** const tmp = av_fetch( revcharmap, state, 0);
1071 PerlIO_printf( Perl_debug_log, "%*s",
1073 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1074 PL_colors[0], PL_colors[1],
1075 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1076 PERL_PV_ESCAPE_FIRSTCHAR
1081 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1082 (int)depth * 2 + 2,"");
1084 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1085 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1086 PerlIO_printf( Perl_debug_log, "\n");
1088 for( state = 1 ; state < trie->statecount ; state++ ) {
1089 const U32 base = trie->states[ state ].trans.base;
1091 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|", (int)depth * 2 + 2,"", (UV)state);
1093 if ( trie->states[ state ].wordnum ) {
1094 PerlIO_printf( Perl_debug_log, " W%4X", trie->states[ state ].wordnum );
1096 PerlIO_printf( Perl_debug_log, "%6s", "" );
1099 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1104 while( ( base + ofs < trie->uniquecharcount ) ||
1105 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1106 && trie->trans[ base + ofs - trie->uniquecharcount ].check != state))
1109 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1111 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1112 if ( ( base + ofs >= trie->uniquecharcount ) &&
1113 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
1114 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
1116 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1118 (UV)trie->trans[ base + ofs - trie->uniquecharcount ].next );
1120 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1124 PerlIO_printf( Perl_debug_log, "]");
1127 PerlIO_printf( Perl_debug_log, "\n" );
1129 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=", (int)depth*2, "");
1130 for (word=1; word <= trie->wordcount; word++) {
1131 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1132 (int)word, (int)(trie->wordinfo[word].prev),
1133 (int)(trie->wordinfo[word].len));
1135 PerlIO_printf(Perl_debug_log, "\n" );
1138 Dumps a fully constructed but uncompressed trie in list form.
1139 List tries normally only are used for construction when the number of
1140 possible chars (trie->uniquecharcount) is very high.
1141 Used for debugging make_trie().
1144 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1145 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1149 SV *sv=sv_newmortal();
1150 int colwidth= widecharmap ? 6 : 4;
1151 GET_RE_DEBUG_FLAGS_DECL;
1153 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1155 /* print out the table precompression. */
1156 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1157 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1158 "------:-----+-----------------\n" );
1160 for( state=1 ; state < next_alloc ; state ++ ) {
1163 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1164 (int)depth * 2 + 2,"", (UV)state );
1165 if ( ! trie->states[ state ].wordnum ) {
1166 PerlIO_printf( Perl_debug_log, "%5s| ","");
1168 PerlIO_printf( Perl_debug_log, "W%4x| ",
1169 trie->states[ state ].wordnum
1172 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1173 SV ** const tmp = av_fetch( revcharmap, TRIE_LIST_ITEM(state,charid).forid, 0);
1175 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1177 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1178 PL_colors[0], PL_colors[1],
1179 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1180 PERL_PV_ESCAPE_FIRSTCHAR
1182 TRIE_LIST_ITEM(state,charid).forid,
1183 (UV)TRIE_LIST_ITEM(state,charid).newstate
1186 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1187 (int)((depth * 2) + 14), "");
1190 PerlIO_printf( Perl_debug_log, "\n");
1195 Dumps a fully constructed but uncompressed trie in table form.
1196 This is the normal DFA style state transition table, with a few
1197 twists to facilitate compression later.
1198 Used for debugging make_trie().
1201 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1202 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1207 SV *sv=sv_newmortal();
1208 int colwidth= widecharmap ? 6 : 4;
1209 GET_RE_DEBUG_FLAGS_DECL;
1211 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1214 print out the table precompression so that we can do a visual check
1215 that they are identical.
1218 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1220 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1221 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1223 PerlIO_printf( Perl_debug_log, "%*s",
1225 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1226 PL_colors[0], PL_colors[1],
1227 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1228 PERL_PV_ESCAPE_FIRSTCHAR
1234 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1236 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1237 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1240 PerlIO_printf( Perl_debug_log, "\n" );
1242 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1244 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1245 (int)depth * 2 + 2,"",
1246 (UV)TRIE_NODENUM( state ) );
1248 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1249 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1251 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1253 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1255 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1256 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n", (UV)trie->trans[ state ].check );
1258 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n", (UV)trie->trans[ state ].check,
1259 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1267 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1268 startbranch: the first branch in the whole branch sequence
1269 first : start branch of sequence of branch-exact nodes.
1270 May be the same as startbranch
1271 last : Thing following the last branch.
1272 May be the same as tail.
1273 tail : item following the branch sequence
1274 count : words in the sequence
1275 flags : currently the OP() type we will be building one of /EXACT(|F|Fl)/
1276 depth : indent depth
1278 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1280 A trie is an N'ary tree where the branches are determined by digital
1281 decomposition of the key. IE, at the root node you look up the 1st character and
1282 follow that branch repeat until you find the end of the branches. Nodes can be
1283 marked as "accepting" meaning they represent a complete word. Eg:
1287 would convert into the following structure. Numbers represent states, letters
1288 following numbers represent valid transitions on the letter from that state, if
1289 the number is in square brackets it represents an accepting state, otherwise it
1290 will be in parenthesis.
1292 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1296 (1) +-i->(6)-+-s->[7]
1298 +-s->(3)-+-h->(4)-+-e->[5]
1300 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1302 This shows that when matching against the string 'hers' we will begin at state 1
1303 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1304 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1305 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1306 single traverse. We store a mapping from accepting to state to which word was
1307 matched, and then when we have multiple possibilities we try to complete the
1308 rest of the regex in the order in which they occured in the alternation.
1310 The only prior NFA like behaviour that would be changed by the TRIE support is
1311 the silent ignoring of duplicate alternations which are of the form:
1313 / (DUPE|DUPE) X? (?{ ... }) Y /x
1315 Thus EVAL blocks following a trie may be called a different number of times with
1316 and without the optimisation. With the optimisations dupes will be silently
1317 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1318 the following demonstrates:
1320 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1322 which prints out 'word' three times, but
1324 'words'=~/(word|word|word)(?{ print $1 })S/
1326 which doesnt print it out at all. This is due to other optimisations kicking in.
1328 Example of what happens on a structural level:
1330 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1332 1: CURLYM[1] {1,32767}(18)
1343 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1344 and should turn into:
1346 1: CURLYM[1] {1,32767}(18)
1348 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1356 Cases where tail != last would be like /(?foo|bar)baz/:
1366 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1367 and would end up looking like:
1370 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1377 d = uvuni_to_utf8_flags(d, uv, 0);
1379 is the recommended Unicode-aware way of saying
1384 #define TRIE_STORE_REVCHAR(val) \
1387 SV *zlopp = newSV(7); /* XXX: optimize me */ \
1388 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
1389 unsigned const char *const kapow = uvuni_to_utf8(flrbbbbb, val); \
1390 SvCUR_set(zlopp, kapow - flrbbbbb); \
1393 av_push(revcharmap, zlopp); \
1395 char ooooff = (char)val; \
1396 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
1400 #define TRIE_READ_CHAR STMT_START { \
1403 /* if it is UTF then it is either already folded, or does not need folding */ \
1404 uvc = utf8n_to_uvuni( (const U8*) uc, UTF8_MAXLEN, &len, uniflags); \
1406 else if (folder == PL_fold_latin1) { \
1407 /* if we use this folder we have to obey unicode rules on latin-1 data */ \
1408 if ( foldlen > 0 ) { \
1409 uvc = utf8n_to_uvuni( (const U8*) scan, UTF8_MAXLEN, &len, uniflags ); \
1415 uvc = _to_fold_latin1( (U8) *uc, foldbuf, &foldlen, 1); \
1416 skiplen = UNISKIP(uvc); \
1417 foldlen -= skiplen; \
1418 scan = foldbuf + skiplen; \
1421 /* raw data, will be folded later if needed */ \
1429 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
1430 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
1431 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
1432 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
1434 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
1435 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
1436 TRIE_LIST_CUR( state )++; \
1439 #define TRIE_LIST_NEW(state) STMT_START { \
1440 Newxz( trie->states[ state ].trans.list, \
1441 4, reg_trie_trans_le ); \
1442 TRIE_LIST_CUR( state ) = 1; \
1443 TRIE_LIST_LEN( state ) = 4; \
1446 #define TRIE_HANDLE_WORD(state) STMT_START { \
1447 U16 dupe= trie->states[ state ].wordnum; \
1448 regnode * const noper_next = regnext( noper ); \
1451 /* store the word for dumping */ \
1453 if (OP(noper) != NOTHING) \
1454 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
1456 tmp = newSVpvn_utf8( "", 0, UTF ); \
1457 av_push( trie_words, tmp ); \
1461 trie->wordinfo[curword].prev = 0; \
1462 trie->wordinfo[curword].len = wordlen; \
1463 trie->wordinfo[curword].accept = state; \
1465 if ( noper_next < tail ) { \
1467 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, sizeof(U16) ); \
1468 trie->jump[curword] = (U16)(noper_next - convert); \
1470 jumper = noper_next; \
1472 nextbranch= regnext(cur); \
1476 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
1477 /* chain, so that when the bits of chain are later */\
1478 /* linked together, the dups appear in the chain */\
1479 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
1480 trie->wordinfo[dupe].prev = curword; \
1482 /* we haven't inserted this word yet. */ \
1483 trie->states[ state ].wordnum = curword; \
1488 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
1489 ( ( base + charid >= ucharcount \
1490 && base + charid < ubound \
1491 && state == trie->trans[ base - ucharcount + charid ].check \
1492 && trie->trans[ base - ucharcount + charid ].next ) \
1493 ? trie->trans[ base - ucharcount + charid ].next \
1494 : ( state==1 ? special : 0 ) \
1498 #define MADE_JUMP_TRIE 2
1499 #define MADE_EXACT_TRIE 4
1502 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch, regnode *first, regnode *last, regnode *tail, U32 word_count, U32 flags, U32 depth)
1505 /* first pass, loop through and scan words */
1506 reg_trie_data *trie;
1507 HV *widecharmap = NULL;
1508 AV *revcharmap = newAV();
1510 const U32 uniflags = UTF8_ALLOW_DEFAULT;
1515 regnode *jumper = NULL;
1516 regnode *nextbranch = NULL;
1517 regnode *convert = NULL;
1518 U32 *prev_states; /* temp array mapping each state to previous one */
1519 /* we just use folder as a flag in utf8 */
1520 const U8 * folder = NULL;
1523 const U32 data_slot = add_data( pRExC_state, 4, "tuuu" );
1524 AV *trie_words = NULL;
1525 /* along with revcharmap, this only used during construction but both are
1526 * useful during debugging so we store them in the struct when debugging.
1529 const U32 data_slot = add_data( pRExC_state, 2, "tu" );
1530 STRLEN trie_charcount=0;
1532 SV *re_trie_maxbuff;
1533 GET_RE_DEBUG_FLAGS_DECL;
1535 PERL_ARGS_ASSERT_MAKE_TRIE;
1537 PERL_UNUSED_ARG(depth);
1544 case EXACTFU_TRICKYFOLD:
1545 case EXACTFU: folder = PL_fold_latin1; break;
1546 case EXACTF: folder = PL_fold; break;
1547 case EXACTFL: folder = PL_fold_locale; break;
1548 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
1551 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
1553 trie->startstate = 1;
1554 trie->wordcount = word_count;
1555 RExC_rxi->data->data[ data_slot ] = (void*)trie;
1556 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
1558 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
1559 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
1560 trie->wordcount+1, sizeof(reg_trie_wordinfo));
1563 trie_words = newAV();
1566 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
1567 if (!SvIOK(re_trie_maxbuff)) {
1568 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
1570 DEBUG_TRIE_COMPILE_r({
1571 PerlIO_printf( Perl_debug_log,
1572 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
1573 (int)depth * 2 + 2, "",
1574 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
1575 REG_NODE_NUM(last), REG_NODE_NUM(tail),
1579 /* Find the node we are going to overwrite */
1580 if ( first == startbranch && OP( last ) != BRANCH ) {
1581 /* whole branch chain */
1584 /* branch sub-chain */
1585 convert = NEXTOPER( first );
1588 /* -- First loop and Setup --
1590 We first traverse the branches and scan each word to determine if it
1591 contains widechars, and how many unique chars there are, this is
1592 important as we have to build a table with at least as many columns as we
1595 We use an array of integers to represent the character codes 0..255
1596 (trie->charmap) and we use a an HV* to store Unicode characters. We use the
1597 native representation of the character value as the key and IV's for the
1600 *TODO* If we keep track of how many times each character is used we can
1601 remap the columns so that the table compression later on is more
1602 efficient in terms of memory by ensuring the most common value is in the
1603 middle and the least common are on the outside. IMO this would be better
1604 than a most to least common mapping as theres a decent chance the most
1605 common letter will share a node with the least common, meaning the node
1606 will not be compressible. With a middle is most common approach the worst
1607 case is when we have the least common nodes twice.
1611 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1612 regnode *noper = NEXTOPER( cur );
1613 const U8 *uc = (U8*)STRING( noper );
1614 const U8 *e = uc + STR_LEN( noper );
1616 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1618 const U8 *scan = (U8*)NULL;
1619 U32 wordlen = 0; /* required init */
1621 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the bitmap?*/
1623 if (OP(noper) == NOTHING) {
1624 regnode *noper_next= regnext(noper);
1625 if (noper_next != tail && OP(noper_next) == flags) {
1627 uc= (U8*)STRING(noper);
1628 e= uc + STR_LEN(noper);
1629 trie->minlen= STR_LEN(noper);
1636 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
1637 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
1638 regardless of encoding */
1639 if (OP( noper ) == EXACTFU_SS) {
1640 /* false positives are ok, so just set this */
1641 TRIE_BITMAP_SET(trie,0xDF);
1644 for ( ; uc < e ; uc += len ) {
1645 TRIE_CHARCOUNT(trie)++;
1650 U8 folded= folder[ (U8) uvc ];
1651 if ( !trie->charmap[ folded ] ) {
1652 trie->charmap[ folded ]=( ++trie->uniquecharcount );
1653 TRIE_STORE_REVCHAR( folded );
1656 if ( !trie->charmap[ uvc ] ) {
1657 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
1658 TRIE_STORE_REVCHAR( uvc );
1661 /* store the codepoint in the bitmap, and its folded
1663 TRIE_BITMAP_SET(trie, uvc);
1665 /* store the folded codepoint */
1666 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
1669 /* store first byte of utf8 representation of
1670 variant codepoints */
1671 if (! UNI_IS_INVARIANT(uvc)) {
1672 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
1675 set_bit = 0; /* We've done our bit :-) */
1680 widecharmap = newHV();
1682 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
1685 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
1687 if ( !SvTRUE( *svpp ) ) {
1688 sv_setiv( *svpp, ++trie->uniquecharcount );
1689 TRIE_STORE_REVCHAR(uvc);
1693 if( cur == first ) {
1694 trie->minlen = chars;
1695 trie->maxlen = chars;
1696 } else if (chars < trie->minlen) {
1697 trie->minlen = chars;
1698 } else if (chars > trie->maxlen) {
1699 trie->maxlen = chars;
1701 if (OP( noper ) == EXACTFU_SS) {
1702 /* XXX: workaround - 'ss' could match "\x{DF}" so minlen could be 1 and not 2*/
1703 if (trie->minlen > 1)
1706 if (OP( noper ) == EXACTFU_TRICKYFOLD) {
1707 /* XXX: workround - things like "\x{1FBE}\x{0308}\x{0301}" can match "\x{0390}"
1708 * - We assume that any such sequence might match a 2 byte string */
1709 if (trie->minlen > 2 )
1713 } /* end first pass */
1714 DEBUG_TRIE_COMPILE_r(
1715 PerlIO_printf( Perl_debug_log, "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
1716 (int)depth * 2 + 2,"",
1717 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
1718 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
1719 (int)trie->minlen, (int)trie->maxlen )
1723 We now know what we are dealing with in terms of unique chars and
1724 string sizes so we can calculate how much memory a naive
1725 representation using a flat table will take. If it's over a reasonable
1726 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
1727 conservative but potentially much slower representation using an array
1730 At the end we convert both representations into the same compressed
1731 form that will be used in regexec.c for matching with. The latter
1732 is a form that cannot be used to construct with but has memory
1733 properties similar to the list form and access properties similar
1734 to the table form making it both suitable for fast searches and
1735 small enough that its feasable to store for the duration of a program.
1737 See the comment in the code where the compressed table is produced
1738 inplace from the flat tabe representation for an explanation of how
1739 the compression works.
1744 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
1747 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1) > SvIV(re_trie_maxbuff) ) {
1749 Second Pass -- Array Of Lists Representation
1751 Each state will be represented by a list of charid:state records
1752 (reg_trie_trans_le) the first such element holds the CUR and LEN
1753 points of the allocated array. (See defines above).
1755 We build the initial structure using the lists, and then convert
1756 it into the compressed table form which allows faster lookups
1757 (but cant be modified once converted).
1760 STRLEN transcount = 1;
1762 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1763 "%*sCompiling trie using list compiler\n",
1764 (int)depth * 2 + 2, ""));
1766 trie->states = (reg_trie_state *)
1767 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1768 sizeof(reg_trie_state) );
1772 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1774 regnode *noper = NEXTOPER( cur );
1775 U8 *uc = (U8*)STRING( noper );
1776 const U8 *e = uc + STR_LEN( noper );
1777 U32 state = 1; /* required init */
1778 U16 charid = 0; /* sanity init */
1779 U8 *scan = (U8*)NULL; /* sanity init */
1780 STRLEN foldlen = 0; /* required init */
1781 U32 wordlen = 0; /* required init */
1782 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1785 if (OP(noper) == NOTHING) {
1786 regnode *noper_next= regnext(noper);
1787 if (noper_next != tail && OP(noper_next) == flags) {
1789 uc= (U8*)STRING(noper);
1790 e= uc + STR_LEN(noper);
1794 if (OP(noper) != NOTHING) {
1795 for ( ; uc < e ; uc += len ) {
1800 charid = trie->charmap[ uvc ];
1802 SV** const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
1806 charid=(U16)SvIV( *svpp );
1809 /* charid is now 0 if we dont know the char read, or nonzero if we do */
1816 if ( !trie->states[ state ].trans.list ) {
1817 TRIE_LIST_NEW( state );
1819 for ( check = 1; check <= TRIE_LIST_USED( state ); check++ ) {
1820 if ( TRIE_LIST_ITEM( state, check ).forid == charid ) {
1821 newstate = TRIE_LIST_ITEM( state, check ).newstate;
1826 newstate = next_alloc++;
1827 prev_states[newstate] = state;
1828 TRIE_LIST_PUSH( state, charid, newstate );
1833 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
1837 TRIE_HANDLE_WORD(state);
1839 } /* end second pass */
1841 /* next alloc is the NEXT state to be allocated */
1842 trie->statecount = next_alloc;
1843 trie->states = (reg_trie_state *)
1844 PerlMemShared_realloc( trie->states,
1846 * sizeof(reg_trie_state) );
1848 /* and now dump it out before we compress it */
1849 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
1850 revcharmap, next_alloc,
1854 trie->trans = (reg_trie_trans *)
1855 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
1862 for( state=1 ; state < next_alloc ; state ++ ) {
1866 DEBUG_TRIE_COMPILE_MORE_r(
1867 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
1871 if (trie->states[state].trans.list) {
1872 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
1876 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1877 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
1878 if ( forid < minid ) {
1880 } else if ( forid > maxid ) {
1884 if ( transcount < tp + maxid - minid + 1) {
1886 trie->trans = (reg_trie_trans *)
1887 PerlMemShared_realloc( trie->trans,
1889 * sizeof(reg_trie_trans) );
1890 Zero( trie->trans + (transcount / 2), transcount / 2 , reg_trie_trans );
1892 base = trie->uniquecharcount + tp - minid;
1893 if ( maxid == minid ) {
1895 for ( ; zp < tp ; zp++ ) {
1896 if ( ! trie->trans[ zp ].next ) {
1897 base = trie->uniquecharcount + zp - minid;
1898 trie->trans[ zp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1899 trie->trans[ zp ].check = state;
1905 trie->trans[ tp ].next = TRIE_LIST_ITEM( state, 1).newstate;
1906 trie->trans[ tp ].check = state;
1911 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
1912 const U32 tid = base - trie->uniquecharcount + TRIE_LIST_ITEM( state, idx ).forid;
1913 trie->trans[ tid ].next = TRIE_LIST_ITEM( state, idx ).newstate;
1914 trie->trans[ tid ].check = state;
1916 tp += ( maxid - minid + 1 );
1918 Safefree(trie->states[ state ].trans.list);
1921 DEBUG_TRIE_COMPILE_MORE_r(
1922 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
1925 trie->states[ state ].trans.base=base;
1927 trie->lasttrans = tp + 1;
1931 Second Pass -- Flat Table Representation.
1933 we dont use the 0 slot of either trans[] or states[] so we add 1 to each.
1934 We know that we will need Charcount+1 trans at most to store the data
1935 (one row per char at worst case) So we preallocate both structures
1936 assuming worst case.
1938 We then construct the trie using only the .next slots of the entry
1941 We use the .check field of the first entry of the node temporarily to
1942 make compression both faster and easier by keeping track of how many non
1943 zero fields are in the node.
1945 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
1948 There are two terms at use here: state as a TRIE_NODEIDX() which is a
1949 number representing the first entry of the node, and state as a
1950 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1) and
1951 TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3) if there
1952 are 2 entrys per node. eg:
1960 The table is internally in the right hand, idx form. However as we also
1961 have to deal with the states array which is indexed by nodenum we have to
1962 use TRIE_NODENUM() to convert.
1965 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
1966 "%*sCompiling trie using table compiler\n",
1967 (int)depth * 2 + 2, ""));
1969 trie->trans = (reg_trie_trans *)
1970 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
1971 * trie->uniquecharcount + 1,
1972 sizeof(reg_trie_trans) );
1973 trie->states = (reg_trie_state *)
1974 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
1975 sizeof(reg_trie_state) );
1976 next_alloc = trie->uniquecharcount + 1;
1979 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
1981 regnode *noper = NEXTOPER( cur );
1982 const U8 *uc = (U8*)STRING( noper );
1983 const U8 *e = uc + STR_LEN( noper );
1985 U32 state = 1; /* required init */
1987 U16 charid = 0; /* sanity init */
1988 U32 accept_state = 0; /* sanity init */
1989 U8 *scan = (U8*)NULL; /* sanity init */
1991 STRLEN foldlen = 0; /* required init */
1992 U32 wordlen = 0; /* required init */
1994 U8 foldbuf[ UTF8_MAXBYTES_CASE + 1 ];
1996 if (OP(noper) == NOTHING) {
1997 regnode *noper_next= regnext(noper);
1998 if (noper_next != tail && OP(noper_next) == flags) {
2000 uc= (U8*)STRING(noper);
2001 e= uc + STR_LEN(noper);
2005 if ( OP(noper) != NOTHING ) {
2006 for ( ; uc < e ; uc += len ) {
2011 charid = trie->charmap[ uvc ];
2013 SV* const * const svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 0);
2014 charid = svpp ? (U16)SvIV(*svpp) : 0;
2018 if ( !trie->trans[ state + charid ].next ) {
2019 trie->trans[ state + charid ].next = next_alloc;
2020 trie->trans[ state ].check++;
2021 prev_states[TRIE_NODENUM(next_alloc)]
2022 = TRIE_NODENUM(state);
2023 next_alloc += trie->uniquecharcount;
2025 state = trie->trans[ state + charid ].next;
2027 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2029 /* charid is now 0 if we dont know the char read, or nonzero if we do */
2032 accept_state = TRIE_NODENUM( state );
2033 TRIE_HANDLE_WORD(accept_state);
2035 } /* end second pass */
2037 /* and now dump it out before we compress it */
2038 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2040 next_alloc, depth+1));
2044 * Inplace compress the table.*
2046 For sparse data sets the table constructed by the trie algorithm will
2047 be mostly 0/FAIL transitions or to put it another way mostly empty.
2048 (Note that leaf nodes will not contain any transitions.)
2050 This algorithm compresses the tables by eliminating most such
2051 transitions, at the cost of a modest bit of extra work during lookup:
2053 - Each states[] entry contains a .base field which indicates the
2054 index in the state[] array wheres its transition data is stored.
2056 - If .base is 0 there are no valid transitions from that node.
2058 - If .base is nonzero then charid is added to it to find an entry in
2061 -If trans[states[state].base+charid].check!=state then the
2062 transition is taken to be a 0/Fail transition. Thus if there are fail
2063 transitions at the front of the node then the .base offset will point
2064 somewhere inside the previous nodes data (or maybe even into a node
2065 even earlier), but the .check field determines if the transition is
2069 The following process inplace converts the table to the compressed
2070 table: We first do not compress the root node 1,and mark all its
2071 .check pointers as 1 and set its .base pointer as 1 as well. This
2072 allows us to do a DFA construction from the compressed table later,
2073 and ensures that any .base pointers we calculate later are greater
2076 - We set 'pos' to indicate the first entry of the second node.
2078 - We then iterate over the columns of the node, finding the first and
2079 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2080 and set the .check pointers accordingly, and advance pos
2081 appropriately and repreat for the next node. Note that when we copy
2082 the next pointers we have to convert them from the original
2083 NODEIDX form to NODENUM form as the former is not valid post
2086 - If a node has no transitions used we mark its base as 0 and do not
2087 advance the pos pointer.
2089 - If a node only has one transition we use a second pointer into the
2090 structure to fill in allocated fail transitions from other states.
2091 This pointer is independent of the main pointer and scans forward
2092 looking for null transitions that are allocated to a state. When it
2093 finds one it writes the single transition into the "hole". If the
2094 pointer doesnt find one the single transition is appended as normal.
2096 - Once compressed we can Renew/realloc the structures to release the
2099 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2100 specifically Fig 3.47 and the associated pseudocode.
2104 const U32 laststate = TRIE_NODENUM( next_alloc );
2107 trie->statecount = laststate;
2109 for ( state = 1 ; state < laststate ; state++ ) {
2111 const U32 stateidx = TRIE_NODEIDX( state );
2112 const U32 o_used = trie->trans[ stateidx ].check;
2113 U32 used = trie->trans[ stateidx ].check;
2114 trie->trans[ stateidx ].check = 0;
2116 for ( charid = 0 ; used && charid < trie->uniquecharcount ; charid++ ) {
2117 if ( flag || trie->trans[ stateidx + charid ].next ) {
2118 if ( trie->trans[ stateidx + charid ].next ) {
2120 for ( ; zp < pos ; zp++ ) {
2121 if ( ! trie->trans[ zp ].next ) {
2125 trie->states[ state ].trans.base = zp + trie->uniquecharcount - charid ;
2126 trie->trans[ zp ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2127 trie->trans[ zp ].check = state;
2128 if ( ++zp > pos ) pos = zp;
2135 trie->states[ state ].trans.base = pos + trie->uniquecharcount - charid ;
2137 trie->trans[ pos ].next = SAFE_TRIE_NODENUM( trie->trans[ stateidx + charid ].next );
2138 trie->trans[ pos ].check = state;
2143 trie->lasttrans = pos + 1;
2144 trie->states = (reg_trie_state *)
2145 PerlMemShared_realloc( trie->states, laststate
2146 * sizeof(reg_trie_state) );
2147 DEBUG_TRIE_COMPILE_MORE_r(
2148 PerlIO_printf( Perl_debug_log,
2149 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2150 (int)depth * 2 + 2,"",
2151 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1 ),
2154 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2157 } /* end table compress */
2159 DEBUG_TRIE_COMPILE_MORE_r(
2160 PerlIO_printf(Perl_debug_log, "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2161 (int)depth * 2 + 2, "",
2162 (UV)trie->statecount,
2163 (UV)trie->lasttrans)
2165 /* resize the trans array to remove unused space */
2166 trie->trans = (reg_trie_trans *)
2167 PerlMemShared_realloc( trie->trans, trie->lasttrans
2168 * sizeof(reg_trie_trans) );
2170 { /* Modify the program and insert the new TRIE node */
2171 U8 nodetype =(U8)(flags & 0xFF);
2175 regnode *optimize = NULL;
2176 #ifdef RE_TRACK_PATTERN_OFFSETS
2179 U32 mjd_nodelen = 0;
2180 #endif /* RE_TRACK_PATTERN_OFFSETS */
2181 #endif /* DEBUGGING */
2183 This means we convert either the first branch or the first Exact,
2184 depending on whether the thing following (in 'last') is a branch
2185 or not and whther first is the startbranch (ie is it a sub part of
2186 the alternation or is it the whole thing.)
2187 Assuming its a sub part we convert the EXACT otherwise we convert
2188 the whole branch sequence, including the first.
2190 /* Find the node we are going to overwrite */
2191 if ( first != startbranch || OP( last ) == BRANCH ) {
2192 /* branch sub-chain */
2193 NEXT_OFF( first ) = (U16)(last - first);
2194 #ifdef RE_TRACK_PATTERN_OFFSETS
2196 mjd_offset= Node_Offset((convert));
2197 mjd_nodelen= Node_Length((convert));
2200 /* whole branch chain */
2202 #ifdef RE_TRACK_PATTERN_OFFSETS
2205 const regnode *nop = NEXTOPER( convert );
2206 mjd_offset= Node_Offset((nop));
2207 mjd_nodelen= Node_Length((nop));
2211 PerlIO_printf(Perl_debug_log, "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2212 (int)depth * 2 + 2, "",
2213 (UV)mjd_offset, (UV)mjd_nodelen)
2216 /* But first we check to see if there is a common prefix we can
2217 split out as an EXACT and put in front of the TRIE node. */
2218 trie->startstate= 1;
2219 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2221 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2225 const U32 base = trie->states[ state ].trans.base;
2227 if ( trie->states[state].wordnum )
2230 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2231 if ( ( base + ofs >= trie->uniquecharcount ) &&
2232 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2233 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2235 if ( ++count > 1 ) {
2236 SV **tmp = av_fetch( revcharmap, ofs, 0);
2237 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2238 if ( state == 1 ) break;
2240 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2242 PerlIO_printf(Perl_debug_log,
2243 "%*sNew Start State=%"UVuf" Class: [",
2244 (int)depth * 2 + 2, "",
2247 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2248 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2250 TRIE_BITMAP_SET(trie,*ch);
2252 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2254 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2258 TRIE_BITMAP_SET(trie,*ch);
2260 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2261 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2267 SV **tmp = av_fetch( revcharmap, idx, 0);
2269 char *ch = SvPV( *tmp, len );
2271 SV *sv=sv_newmortal();
2272 PerlIO_printf( Perl_debug_log,
2273 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2274 (int)depth * 2 + 2, "",
2276 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2277 PL_colors[0], PL_colors[1],
2278 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2279 PERL_PV_ESCAPE_FIRSTCHAR
2284 OP( convert ) = nodetype;
2285 str=STRING(convert);
2288 STR_LEN(convert) += len;
2294 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
2299 trie->prefixlen = (state-1);
2301 regnode *n = convert+NODE_SZ_STR(convert);
2302 NEXT_OFF(convert) = NODE_SZ_STR(convert);
2303 trie->startstate = state;
2304 trie->minlen -= (state - 1);
2305 trie->maxlen -= (state - 1);
2307 /* At least the UNICOS C compiler choked on this
2308 * being argument to DEBUG_r(), so let's just have
2311 #ifdef PERL_EXT_RE_BUILD
2317 regnode *fix = convert;
2318 U32 word = trie->wordcount;
2320 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
2321 while( ++fix < n ) {
2322 Set_Node_Offset_Length(fix, 0, 0);
2325 SV ** const tmp = av_fetch( trie_words, word, 0 );
2327 if ( STR_LEN(convert) <= SvCUR(*tmp) )
2328 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
2330 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
2338 NEXT_OFF(convert) = (U16)(tail - convert);
2339 DEBUG_r(optimize= n);
2345 if ( trie->maxlen ) {
2346 NEXT_OFF( convert ) = (U16)(tail - convert);
2347 ARG_SET( convert, data_slot );
2348 /* Store the offset to the first unabsorbed branch in
2349 jump[0], which is otherwise unused by the jump logic.
2350 We use this when dumping a trie and during optimisation. */
2352 trie->jump[0] = (U16)(nextbranch - convert);
2354 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
2355 * and there is a bitmap
2356 * and the first "jump target" node we found leaves enough room
2357 * then convert the TRIE node into a TRIEC node, with the bitmap
2358 * embedded inline in the opcode - this is hypothetically faster.
2360 if ( !trie->states[trie->startstate].wordnum
2362 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
2364 OP( convert ) = TRIEC;
2365 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
2366 PerlMemShared_free(trie->bitmap);
2369 OP( convert ) = TRIE;
2371 /* store the type in the flags */
2372 convert->flags = nodetype;
2376 + regarglen[ OP( convert ) ];
2378 /* XXX We really should free up the resource in trie now,
2379 as we won't use them - (which resources?) dmq */
2381 /* needed for dumping*/
2382 DEBUG_r(if (optimize) {
2383 regnode *opt = convert;
2385 while ( ++opt < optimize) {
2386 Set_Node_Offset_Length(opt,0,0);
2389 Try to clean up some of the debris left after the
2392 while( optimize < jumper ) {
2393 mjd_nodelen += Node_Length((optimize));
2394 OP( optimize ) = OPTIMIZED;
2395 Set_Node_Offset_Length(optimize,0,0);
2398 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
2400 } /* end node insert */
2402 /* Finish populating the prev field of the wordinfo array. Walk back
2403 * from each accept state until we find another accept state, and if
2404 * so, point the first word's .prev field at the second word. If the
2405 * second already has a .prev field set, stop now. This will be the
2406 * case either if we've already processed that word's accept state,
2407 * or that state had multiple words, and the overspill words were
2408 * already linked up earlier.
2415 for (word=1; word <= trie->wordcount; word++) {
2417 if (trie->wordinfo[word].prev)
2419 state = trie->wordinfo[word].accept;
2421 state = prev_states[state];
2424 prev = trie->states[state].wordnum;
2428 trie->wordinfo[word].prev = prev;
2430 Safefree(prev_states);
2434 /* and now dump out the compressed format */
2435 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
2437 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
2439 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
2440 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
2442 SvREFCNT_dec(revcharmap);
2446 : trie->startstate>1
2452 S_make_trie_failtable(pTHX_ RExC_state_t *pRExC_state, regnode *source, regnode *stclass, U32 depth)
2454 /* The Trie is constructed and compressed now so we can build a fail array if it's needed
2456 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and 3.32 in the
2457 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi, Ullman 1985/88
2460 We find the fail state for each state in the trie, this state is the longest proper
2461 suffix of the current state's 'word' that is also a proper prefix of another word in our
2462 trie. State 1 represents the word '' and is thus the default fail state. This allows
2463 the DFA not to have to restart after its tried and failed a word at a given point, it
2464 simply continues as though it had been matching the other word in the first place.
2466 'abcdgu'=~/abcdefg|cdgu/
2467 When we get to 'd' we are still matching the first word, we would encounter 'g' which would
2468 fail, which would bring us to the state representing 'd' in the second word where we would
2469 try 'g' and succeed, proceeding to match 'cdgu'.
2471 /* add a fail transition */
2472 const U32 trie_offset = ARG(source);
2473 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
2475 const U32 ucharcount = trie->uniquecharcount;
2476 const U32 numstates = trie->statecount;
2477 const U32 ubound = trie->lasttrans + ucharcount;
2481 U32 base = trie->states[ 1 ].trans.base;
2484 const U32 data_slot = add_data( pRExC_state, 1, "T" );
2485 GET_RE_DEBUG_FLAGS_DECL;
2487 PERL_ARGS_ASSERT_MAKE_TRIE_FAILTABLE;
2489 PERL_UNUSED_ARG(depth);
2493 ARG_SET( stclass, data_slot );
2494 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
2495 RExC_rxi->data->data[ data_slot ] = (void*)aho;
2496 aho->trie=trie_offset;
2497 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
2498 Copy( trie->states, aho->states, numstates, reg_trie_state );
2499 Newxz( q, numstates, U32);
2500 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
2503 /* initialize fail[0..1] to be 1 so that we always have
2504 a valid final fail state */
2505 fail[ 0 ] = fail[ 1 ] = 1;
2507 for ( charid = 0; charid < ucharcount ; charid++ ) {
2508 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
2510 q[ q_write ] = newstate;
2511 /* set to point at the root */
2512 fail[ q[ q_write++ ] ]=1;
2515 while ( q_read < q_write) {
2516 const U32 cur = q[ q_read++ % numstates ];
2517 base = trie->states[ cur ].trans.base;
2519 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
2520 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
2522 U32 fail_state = cur;
2525 fail_state = fail[ fail_state ];
2526 fail_base = aho->states[ fail_state ].trans.base;
2527 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
2529 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
2530 fail[ ch_state ] = fail_state;
2531 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
2533 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
2535 q[ q_write++ % numstates] = ch_state;
2539 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
2540 when we fail in state 1, this allows us to use the
2541 charclass scan to find a valid start char. This is based on the principle
2542 that theres a good chance the string being searched contains lots of stuff
2543 that cant be a start char.
2545 fail[ 0 ] = fail[ 1 ] = 0;
2546 DEBUG_TRIE_COMPILE_r({
2547 PerlIO_printf(Perl_debug_log,
2548 "%*sStclass Failtable (%"UVuf" states): 0",
2549 (int)(depth * 2), "", (UV)numstates
2551 for( q_read=1; q_read<numstates; q_read++ ) {
2552 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
2554 PerlIO_printf(Perl_debug_log, "\n");
2557 /*RExC_seen |= REG_SEEN_TRIEDFA;*/
2562 * There are strange code-generation bugs caused on sparc64 by gcc-2.95.2.
2563 * These need to be revisited when a newer toolchain becomes available.
2565 #if defined(__sparc64__) && defined(__GNUC__)
2566 # if __GNUC__ < 2 || (__GNUC__ == 2 && __GNUC_MINOR__ < 96)
2567 # undef SPARC64_GCC_WORKAROUND
2568 # define SPARC64_GCC_WORKAROUND 1
2572 #define DEBUG_PEEP(str,scan,depth) \
2573 DEBUG_OPTIMISE_r({if (scan){ \
2574 SV * const mysv=sv_newmortal(); \
2575 regnode *Next = regnext(scan); \
2576 regprop(RExC_rx, mysv, scan); \
2577 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)\n", \
2578 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(mysv),\
2579 Next ? (REG_NODE_NUM(Next)) : 0 ); \
2583 /* The below joins as many adjacent EXACTish nodes as possible into a single
2584 * one, and looks for problematic sequences of characters whose folds vs.
2585 * non-folds have sufficiently different lengths, that the optimizer would be
2586 * fooled into rejecting legitimate matches of them, and the trie construction
2587 * code needs to handle specially. The joining is only done if:
2588 * 1) there is room in the current conglomerated node to entirely contain the
2590 * 2) they are the exact same node type
2592 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
2593 * these get optimized out
2595 * If there are problematic code sequences, *min_subtract is set to the delta
2596 * that the minimum size of the node can be less than its actual size. And,
2597 * the node type of the result is changed to reflect that it contains these
2600 * And *has_exactf_sharp_s is set to indicate whether or not the node is EXACTF
2601 * and contains LATIN SMALL LETTER SHARP S
2603 * This is as good a place as any to discuss the design of handling these
2604 * problematic sequences. It's been wrong in Perl for a very long time. There
2605 * are three code points currently in Unicode whose folded lengths differ so
2606 * much from the un-folded lengths that it causes problems for the optimizer
2607 * and trie construction. Why only these are problematic, and not others where
2608 * lengths also differ is something I (khw) do not understand. New versions of
2609 * Unicode might add more such code points. Hopefully the logic in
2610 * fold_grind.t that figures out what to test (in part by verifying that each
2611 * size-combination gets tested) will catch any that do come along, so they can
2612 * be added to the special handling below. The chances of new ones are
2613 * actually rather small, as most, if not all, of the world's scripts that have
2614 * casefolding have already been encoded by Unicode. Also, a number of
2615 * Unicode's decisions were made to allow compatibility with pre-existing
2616 * standards, and almost all of those have already been dealt with. These
2617 * would otherwise be the most likely candidates for generating further tricky
2618 * sequences. In other words, Unicode by itself is unlikely to add new ones
2619 * unless it is for compatibility with pre-existing standards, and there aren't
2620 * many of those left.
2622 * The previous designs for dealing with these involved assigning a special
2623 * node for them. This approach doesn't work, as evidenced by this example:
2624 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
2625 * Both these fold to "sss", but if the pattern is parsed to create a node
2626 * that would match just the \xDF, it won't be able to handle the case where a
2627 * successful match would have to cross the node's boundary. The new approach
2628 * that hopefully generally solves the problem generates an EXACTFU_SS node
2631 * There are a number of components to the approach (a lot of work for just
2632 * three code points!):
2633 * 1) This routine examines each EXACTFish node that could contain the
2634 * problematic sequences. It returns in *min_subtract how much to
2635 * subtract from the the actual length of the string to get a real minimum
2636 * for one that could match it. This number is usually 0 except for the
2637 * problematic sequences. This delta is used by the caller to adjust the
2638 * min length of the match, and the delta between min and max, so that the
2639 * optimizer doesn't reject these possibilities based on size constraints.
2640 * 2) These sequences require special handling by the trie code, so this code
2641 * changes the joined node type to special ops: EXACTFU_TRICKYFOLD and
2643 * 3) This is sufficient for the two Greek sequences (described below), but
2644 * the one involving the Sharp s (\xDF) needs more. The node type
2645 * EXACTFU_SS is used for an EXACTFU node that contains at least one "ss"
2646 * sequence in it. For non-UTF-8 patterns and strings, this is the only
2647 * case where there is a possible fold length change. That means that a
2648 * regular EXACTFU node without UTF-8 involvement doesn't have to concern
2649 * itself with length changes, and so can be processed faster. regexec.c
2650 * takes advantage of this. Generally, an EXACTFish node that is in UTF-8
2651 * is pre-folded by regcomp.c. This saves effort in regex matching.
2652 * However, the pre-folding isn't done for non-UTF8 patterns because the
2653 * fold of the MICRO SIGN requires UTF-8. Also what EXACTF and EXACTFL
2654 * nodes fold to isn't known until runtime. The fold possibilities for
2655 * the non-UTF8 patterns are quite simple, except for the sharp s. All
2656 * the ones that don't involve a UTF-8 target string are members of a
2657 * fold-pair, and arrays are set up for all of them so that the other
2658 * member of the pair can be found quickly. Code elsewhere in this file
2659 * makes sure that in EXACTFU nodes, the sharp s gets folded to 'ss', even
2660 * if the pattern isn't UTF-8. This avoids the issues described in the
2662 * 4) A problem remains for the sharp s in EXACTF nodes. Whether it matches
2663 * 'ss' or not is not knowable at compile time. It will match iff the
2664 * target string is in UTF-8, unlike the EXACTFU nodes, where it always
2665 * matches; and the EXACTFL and EXACTFA nodes where it never does. Thus
2666 * it can't be folded to "ss" at compile time, unlike EXACTFU does (as
2667 * described in item 3). An assumption that the optimizer part of
2668 * regexec.c (probably unwittingly) makes is that a character in the
2669 * pattern corresponds to at most a single character in the target string.
2670 * (And I do mean character, and not byte here, unlike other parts of the
2671 * documentation that have never been updated to account for multibyte
2672 * Unicode.) This assumption is wrong only in this case, as all other
2673 * cases are either 1-1 folds when no UTF-8 is involved; or is true by
2674 * virtue of having this file pre-fold UTF-8 patterns. I'm
2675 * reluctant to try to change this assumption, so instead the code punts.
2676 * This routine examines EXACTF nodes for the sharp s, and returns a
2677 * boolean indicating whether or not the node is an EXACTF node that
2678 * contains a sharp s. When it is true, the caller sets a flag that later
2679 * causes the optimizer in this file to not set values for the floating
2680 * and fixed string lengths, and thus avoids the optimizer code in
2681 * regexec.c that makes the invalid assumption. Thus, there is no
2682 * optimization based on string lengths for EXACTF nodes that contain the
2683 * sharp s. This only happens for /id rules (which means the pattern
2687 #define JOIN_EXACT(scan,min_subtract,has_exactf_sharp_s, flags) \
2688 if (PL_regkind[OP(scan)] == EXACT) \
2689 join_exact(pRExC_state,(scan),(min_subtract),has_exactf_sharp_s, (flags),NULL,depth+1)
2692 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) {
2693 /* Merge several consecutive EXACTish nodes into one. */
2694 regnode *n = regnext(scan);
2696 regnode *next = scan + NODE_SZ_STR(scan);
2700 regnode *stop = scan;
2701 GET_RE_DEBUG_FLAGS_DECL;
2703 PERL_UNUSED_ARG(depth);
2706 PERL_ARGS_ASSERT_JOIN_EXACT;
2707 #ifndef EXPERIMENTAL_INPLACESCAN
2708 PERL_UNUSED_ARG(flags);
2709 PERL_UNUSED_ARG(val);
2711 DEBUG_PEEP("join",scan,depth);
2713 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
2714 * EXACT ones that are mergeable to the current one. */
2716 && (PL_regkind[OP(n)] == NOTHING
2717 || (stringok && OP(n) == OP(scan)))
2719 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
2722 if (OP(n) == TAIL || n > next)
2724 if (PL_regkind[OP(n)] == NOTHING) {
2725 DEBUG_PEEP("skip:",n,depth);
2726 NEXT_OFF(scan) += NEXT_OFF(n);
2727 next = n + NODE_STEP_REGNODE;
2734 else if (stringok) {
2735 const unsigned int oldl = STR_LEN(scan);
2736 regnode * const nnext = regnext(n);
2738 /* XXX I (khw) kind of doubt that this works on platforms where
2739 * U8_MAX is above 255 because of lots of other assumptions */
2740 if (oldl + STR_LEN(n) > U8_MAX)
2743 DEBUG_PEEP("merg",n,depth);
2746 NEXT_OFF(scan) += NEXT_OFF(n);
2747 STR_LEN(scan) += STR_LEN(n);
2748 next = n + NODE_SZ_STR(n);
2749 /* Now we can overwrite *n : */
2750 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
2758 #ifdef EXPERIMENTAL_INPLACESCAN
2759 if (flags && !NEXT_OFF(n)) {
2760 DEBUG_PEEP("atch", val, depth);
2761 if (reg_off_by_arg[OP(n)]) {
2762 ARG_SET(n, val - n);
2765 NEXT_OFF(n) = val - n;
2773 *has_exactf_sharp_s = FALSE;
2775 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
2776 * can now analyze for sequences of problematic code points. (Prior to
2777 * this final joining, sequences could have been split over boundaries, and
2778 * hence missed). The sequences only happen in folding, hence for any
2779 * non-EXACT EXACTish node */
2780 if (OP(scan) != EXACT) {
2782 U8 * s0 = (U8*) STRING(scan);
2783 U8 * const s_end = s0 + STR_LEN(scan);
2785 /* The below is perhaps overboard, but this allows us to save a test
2786 * each time through the loop at the expense of a mask. This is
2787 * because on both EBCDIC and ASCII machines, 'S' and 's' differ by a
2788 * single bit. On ASCII they are 32 apart; on EBCDIC, they are 64.
2789 * This uses an exclusive 'or' to find that bit and then inverts it to
2790 * form a mask, with just a single 0, in the bit position where 'S' and
2792 const U8 S_or_s_mask = (U8) ~ ('S' ^ 's');
2793 const U8 s_masked = 's' & S_or_s_mask;
2795 /* One pass is made over the node's string looking for all the
2796 * possibilities. to avoid some tests in the loop, there are two main
2797 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
2801 /* There are two problematic Greek code points in Unicode
2804 * U+0390 - GREEK SMALL LETTER IOTA WITH DIALYTIKA AND TONOS
2805 * U+03B0 - GREEK SMALL LETTER UPSILON WITH DIALYTIKA AND TONOS
2811 * U+03B9 U+0308 U+0301 0xCE 0xB9 0xCC 0x88 0xCC 0x81
2812 * U+03C5 U+0308 U+0301 0xCF 0x85 0xCC 0x88 0xCC 0x81
2814 * This means that in case-insensitive matching (or "loose
2815 * matching", as Unicode calls it), an EXACTF of length six (the
2816 * UTF-8 encoded byte length of the above casefolded versions) can
2817 * match a target string of length two (the byte length of UTF-8
2818 * encoded U+0390 or U+03B0). This would rather mess up the
2819 * minimum length computation. (there are other code points that
2820 * also fold to these two sequences, but the delta is smaller)
2822 * If these sequences are found, the minimum length is decreased by
2823 * four (six minus two).
2825 * Similarly, 'ss' may match the single char and byte LATIN SMALL
2826 * LETTER SHARP S. We decrease the min length by 1 for each
2827 * occurrence of 'ss' found */
2829 #define U390_FIRST_BYTE GREEK_SMALL_LETTER_IOTA_UTF8_FIRST_BYTE
2830 #define U3B0_FIRST_BYTE GREEK_SMALL_LETTER_UPSILON_UTF8_FIRST_BYTE
2831 const U8 U390_tail[] = GREEK_SMALL_LETTER_IOTA_UTF8_TAIL
2832 COMBINING_DIAERESIS_UTF8
2833 COMBINING_ACUTE_ACCENT_UTF8;
2834 const U8 U3B0_tail[] = GREEK_SMALL_LETTER_UPSILON_UTF8_TAIL
2835 COMBINING_DIAERESIS_UTF8
2836 COMBINING_ACUTE_ACCENT_UTF8;
2837 const U8 len = sizeof(U390_tail); /* (-1 for NUL; +1 for 1st byte;
2838 yields a net of 0 */
2839 /* Examine the string for one of the problematic sequences */
2841 s < s_end - 1; /* Can stop 1 before the end, as minimum length
2842 * sequence we are looking for is 2 */
2846 /* Look for the first byte in each problematic sequence */
2848 /* We don't have to worry about other things that fold to
2849 * 's' (such as the long s, U+017F), as all above-latin1
2850 * code points have been pre-folded */
2854 /* Current character is an 's' or 'S'. If next one is
2855 * as well, we have the dreaded sequence */
2856 if (((*(s+1) & S_or_s_mask) == s_masked)
2857 /* These two node types don't have special handling
2859 && OP(scan) != EXACTFL && OP(scan) != EXACTFA)
2862 OP(scan) = EXACTFU_SS;
2863 s++; /* No need to look at this character again */
2867 case U390_FIRST_BYTE:
2868 if (s_end - s >= len
2870 /* The 1's are because are skipping comparing the
2872 && memEQ(s + 1, U390_tail, len - 1))
2874 goto greek_sequence;
2878 case U3B0_FIRST_BYTE:
2879 if (! (s_end - s >= len
2880 && memEQ(s + 1, U3B0_tail, len - 1)))
2887 /* This requires special handling by trie's, so change
2888 * the node type to indicate this. If EXACTFA and
2889 * EXACTFL were ever to be handled by trie's, this
2890 * would have to be changed. If this node has already
2891 * been changed to EXACTFU_SS in this loop, leave it as
2892 * is. (I (khw) think it doesn't matter in regexec.c
2893 * for UTF patterns, but no need to change it */
2894 if (OP(scan) == EXACTFU) {
2895 OP(scan) = EXACTFU_TRICKYFOLD;
2897 s += 6; /* We already know what this sequence is. Skip
2903 else if (OP(scan) != EXACTFL && OP(scan) != EXACTFA) {
2905 /* Here, the pattern is not UTF-8. We need to look only for the
2906 * 'ss' sequence, and in the EXACTF case, the sharp s, which can be
2907 * in the final position. Otherwise we can stop looking 1 byte
2908 * earlier because have to find both the first and second 's' */
2909 const U8* upper = (OP(scan) == EXACTF) ? s_end : s_end -1;
2911 for (s = s0; s < upper; s++) {
2916 && ((*(s+1) & S_or_s_mask) == s_masked))
2920 /* EXACTF nodes need to know that the minimum
2921 * length changed so that a sharp s in the string
2922 * can match this ss in the pattern, but they
2923 * remain EXACTF nodes, as they won't match this
2924 * unless the target string is is UTF-8, which we
2925 * don't know until runtime */
2926 if (OP(scan) != EXACTF) {
2927 OP(scan) = EXACTFU_SS;
2932 case LATIN_SMALL_LETTER_SHARP_S:
2933 if (OP(scan) == EXACTF) {
2934 *has_exactf_sharp_s = TRUE;
2943 /* Allow dumping but overwriting the collection of skipped
2944 * ops and/or strings with fake optimized ops */
2945 n = scan + NODE_SZ_STR(scan);
2953 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
2957 /* REx optimizer. Converts nodes into quicker variants "in place".
2958 Finds fixed substrings. */
2960 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
2961 to the position after last scanned or to NULL. */
2963 #define INIT_AND_WITHP \
2964 assert(!and_withp); \
2965 Newx(and_withp,1,struct regnode_charclass_class); \
2966 SAVEFREEPV(and_withp)
2968 /* this is a chain of data about sub patterns we are processing that
2969 need to be handled separately/specially in study_chunk. Its so
2970 we can simulate recursion without losing state. */
2972 typedef struct scan_frame {
2973 regnode *last; /* last node to process in this frame */
2974 regnode *next; /* next node to process when last is reached */
2975 struct scan_frame *prev; /*previous frame*/
2976 I32 stop; /* what stopparen do we use */
2980 #define SCAN_COMMIT(s, data, m) scan_commit(s, data, m, is_inf)
2982 #define CASE_SYNST_FNC(nAmE) \
2984 if (flags & SCF_DO_STCLASS_AND) { \
2985 for (value = 0; value < 256; value++) \
2986 if (!is_ ## nAmE ## _cp(value)) \
2987 ANYOF_BITMAP_CLEAR(data->start_class, value); \
2990 for (value = 0; value < 256; value++) \
2991 if (is_ ## nAmE ## _cp(value)) \
2992 ANYOF_BITMAP_SET(data->start_class, value); \
2996 if (flags & SCF_DO_STCLASS_AND) { \
2997 for (value = 0; value < 256; value++) \
2998 if (is_ ## nAmE ## _cp(value)) \
2999 ANYOF_BITMAP_CLEAR(data->start_class, value); \
3002 for (value = 0; value < 256; value++) \
3003 if (!is_ ## nAmE ## _cp(value)) \
3004 ANYOF_BITMAP_SET(data->start_class, value); \
3011 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3012 I32 *minlenp, I32 *deltap,
3017 struct regnode_charclass_class *and_withp,
3018 U32 flags, U32 depth)
3019 /* scanp: Start here (read-write). */
3020 /* deltap: Write maxlen-minlen here. */
3021 /* last: Stop before this one. */
3022 /* data: string data about the pattern */
3023 /* stopparen: treat close N as END */
3024 /* recursed: which subroutines have we recursed into */
3025 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3028 I32 min = 0, pars = 0, code;
3029 regnode *scan = *scanp, *next;
3031 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3032 int is_inf_internal = 0; /* The studied chunk is infinite */
3033 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3034 scan_data_t data_fake;
3035 SV *re_trie_maxbuff = NULL;
3036 regnode *first_non_open = scan;
3037 I32 stopmin = I32_MAX;
3038 scan_frame *frame = NULL;
3039 GET_RE_DEBUG_FLAGS_DECL;
3041 PERL_ARGS_ASSERT_STUDY_CHUNK;
3044 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3048 while (first_non_open && OP(first_non_open) == OPEN)
3049 first_non_open=regnext(first_non_open);
3054 while ( scan && OP(scan) != END && scan < last ){
3055 UV min_subtract = 0; /* How much to subtract from the minimum node
3056 length to get a real minimum (because the
3057 folded version may be shorter) */
3058 bool has_exactf_sharp_s = FALSE;
3059 /* Peephole optimizer: */
3060 DEBUG_STUDYDATA("Peep:", data,depth);
3061 DEBUG_PEEP("Peep",scan,depth);
3063 /* Its not clear to khw or hv why this is done here, and not in the
3064 * clauses that deal with EXACT nodes. khw's guess is that it's
3065 * because of a previous design */
3066 JOIN_EXACT(scan,&min_subtract, &has_exactf_sharp_s, 0);
3068 /* Follow the next-chain of the current node and optimize
3069 away all the NOTHINGs from it. */
3070 if (OP(scan) != CURLYX) {
3071 const int max = (reg_off_by_arg[OP(scan)]
3073 /* I32 may be smaller than U16 on CRAYs! */
3074 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3075 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3079 /* Skip NOTHING and LONGJMP. */
3080 while ((n = regnext(n))
3081 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3082 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3083 && off + noff < max)
3085 if (reg_off_by_arg[OP(scan)])
3088 NEXT_OFF(scan) = off;
3093 /* The principal pseudo-switch. Cannot be a switch, since we
3094 look into several different things. */
3095 if (OP(scan) == BRANCH || OP(scan) == BRANCHJ
3096 || OP(scan) == IFTHEN) {
3097 next = regnext(scan);
3099 /* demq: the op(next)==code check is to see if we have "branch-branch" AFAICT */
3101 if (OP(next) == code || code == IFTHEN) {
3102 /* NOTE - There is similar code to this block below for handling
3103 TRIE nodes on a re-study. If you change stuff here check there
3105 I32 max1 = 0, min1 = I32_MAX, num = 0;
3106 struct regnode_charclass_class accum;
3107 regnode * const startbranch=scan;
3109 if (flags & SCF_DO_SUBSTR)
3110 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot merge strings after this. */
3111 if (flags & SCF_DO_STCLASS)
3112 cl_init_zero(pRExC_state, &accum);
3114 while (OP(scan) == code) {
3115 I32 deltanext, minnext, f = 0, fake;
3116 struct regnode_charclass_class this_class;
3119 data_fake.flags = 0;
3121 data_fake.whilem_c = data->whilem_c;
3122 data_fake.last_closep = data->last_closep;
3125 data_fake.last_closep = &fake;
3127 data_fake.pos_delta = delta;
3128 next = regnext(scan);
3129 scan = NEXTOPER(scan);
3131 scan = NEXTOPER(scan);
3132 if (flags & SCF_DO_STCLASS) {
3133 cl_init(pRExC_state, &this_class);
3134 data_fake.start_class = &this_class;
3135 f = SCF_DO_STCLASS_AND;
3137 if (flags & SCF_WHILEM_VISITED_POS)
3138 f |= SCF_WHILEM_VISITED_POS;
3140 /* we suppose the run is continuous, last=next...*/
3141 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3143 stopparen, recursed, NULL, f,depth+1);
3146 if (max1 < minnext + deltanext)
3147 max1 = minnext + deltanext;
3148 if (deltanext == I32_MAX)
3149 is_inf = is_inf_internal = 1;
3151 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3153 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3154 if ( stopmin > minnext)
3155 stopmin = min + min1;
3156 flags &= ~SCF_DO_SUBSTR;
3158 data->flags |= SCF_SEEN_ACCEPT;
3161 if (data_fake.flags & SF_HAS_EVAL)
3162 data->flags |= SF_HAS_EVAL;
3163 data->whilem_c = data_fake.whilem_c;
3165 if (flags & SCF_DO_STCLASS)
3166 cl_or(pRExC_state, &accum, &this_class);
3168 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3170 if (flags & SCF_DO_SUBSTR) {
3171 data->pos_min += min1;
3172 data->pos_delta += max1 - min1;
3173 if (max1 != min1 || is_inf)
3174 data->longest = &(data->longest_float);
3177 delta += max1 - min1;
3178 if (flags & SCF_DO_STCLASS_OR) {
3179 cl_or(pRExC_state, data->start_class, &accum);
3181 cl_and(data->start_class, and_withp);
3182 flags &= ~SCF_DO_STCLASS;
3185 else if (flags & SCF_DO_STCLASS_AND) {
3187 cl_and(data->start_class, &accum);
3188 flags &= ~SCF_DO_STCLASS;
3191 /* Switch to OR mode: cache the old value of
3192 * data->start_class */
3194 StructCopy(data->start_class, and_withp,
3195 struct regnode_charclass_class);
3196 flags &= ~SCF_DO_STCLASS_AND;
3197 StructCopy(&accum, data->start_class,
3198 struct regnode_charclass_class);
3199 flags |= SCF_DO_STCLASS_OR;
3200 data->start_class->flags |= ANYOF_EOS;
3204 if (PERL_ENABLE_TRIE_OPTIMISATION && OP( startbranch ) == BRANCH ) {
3207 Assuming this was/is a branch we are dealing with: 'scan' now
3208 points at the item that follows the branch sequence, whatever
3209 it is. We now start at the beginning of the sequence and look
3216 which would be constructed from a pattern like /A|LIST|OF|WORDS/
3218 If we can find such a subsequence we need to turn the first
3219 element into a trie and then add the subsequent branch exact
3220 strings to the trie.
3224 1. patterns where the whole set of branches can be converted.
3226 2. patterns where only a subset can be converted.
3228 In case 1 we can replace the whole set with a single regop
3229 for the trie. In case 2 we need to keep the start and end
3232 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
3233 becomes BRANCH TRIE; BRANCH X;
3235 There is an additional case, that being where there is a
3236 common prefix, which gets split out into an EXACT like node
3237 preceding the TRIE node.
3239 If x(1..n)==tail then we can do a simple trie, if not we make
3240 a "jump" trie, such that when we match the appropriate word
3241 we "jump" to the appropriate tail node. Essentially we turn
3242 a nested if into a case structure of sorts.
3247 if (!re_trie_maxbuff) {
3248 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
3249 if (!SvIOK(re_trie_maxbuff))
3250 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
3252 if ( SvIV(re_trie_maxbuff)>=0 ) {
3254 regnode *first = (regnode *)NULL;
3255 regnode *last = (regnode *)NULL;
3256 regnode *tail = scan;
3261 SV * const mysv = sv_newmortal(); /* for dumping */
3263 /* var tail is used because there may be a TAIL
3264 regop in the way. Ie, the exacts will point to the
3265 thing following the TAIL, but the last branch will
3266 point at the TAIL. So we advance tail. If we
3267 have nested (?:) we may have to move through several
3271 while ( OP( tail ) == TAIL ) {
3272 /* this is the TAIL generated by (?:) */
3273 tail = regnext( tail );
3277 DEBUG_TRIE_COMPILE_r({
3278 regprop(RExC_rx, mysv, tail );
3279 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
3280 (int)depth * 2 + 2, "",
3281 "Looking for TRIE'able sequences. Tail node is: ",
3282 SvPV_nolen_const( mysv )
3288 Step through the branches
3289 cur represents each branch,
3290 noper is the first thing to be matched as part of that branch
3291 noper_next is the regnext() of that node.
3293 We normally handle a case like this /FOO[xyz]|BAR[pqr]/
3294 via a "jump trie" but we also support building with NOJUMPTRIE,
3295 which restricts the trie logic to structures like /FOO|BAR/.
3297 If noper is a trieable nodetype then the branch is a possible optimization
3298 target. If we are building under NOJUMPTRIE then we require that noper_next
3299 is the same as scan (our current position in the regex program).
3301 Once we have two or more consecutive such branches we can create a
3302 trie of the EXACT's contents and stitch it in place into the program.
3304 If the sequence represents all of the branches in the alternation we
3305 replace the entire thing with a single TRIE node.
3307 Otherwise when it is a subsequence we need to stitch it in place and
3308 replace only the relevant branches. This means the first branch has
3309 to remain as it is used by the alternation logic, and its next pointer,
3310 and needs to be repointed at the item on the branch chain following
3311 the last branch we have optimized away.
3313 This could be either a BRANCH, in which case the subsequence is internal,
3314 or it could be the item following the branch sequence in which case the
3315 subsequence is at the end (which does not necessarily mean the first node
3316 is the start of the alternation).
3318 TRIE_TYPE(X) is a define which maps the optype to a trietype.
3321 ----------------+-----------
3325 EXACTFU_SS | EXACTFU
3326 EXACTFU_TRICKYFOLD | EXACTFU
3331 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) ? NOTHING : \
3332 ( EXACT == (X) ) ? EXACT : \
3333 ( EXACTFU == (X) || EXACTFU_SS == (X) || EXACTFU_TRICKYFOLD == (X) ) ? EXACTFU : \
3336 /* dont use tail as the end marker for this traverse */
3337 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
3338 regnode * const noper = NEXTOPER( cur );
3339 U8 noper_type = OP( noper );
3340 U8 noper_trietype = TRIE_TYPE( noper_type );
3341 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
3342 regnode * const noper_next = regnext( noper );
3343 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
3344 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
3347 DEBUG_TRIE_COMPILE_r({
3348 regprop(RExC_rx, mysv, cur);
3349 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
3350 (int)depth * 2 + 2,"", SvPV_nolen_const( mysv ), REG_NODE_NUM(cur) );
3352 regprop(RExC_rx, mysv, noper);
3353 PerlIO_printf( Perl_debug_log, " -> %s",
3354 SvPV_nolen_const(mysv));
3357 regprop(RExC_rx, mysv, noper_next );
3358 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
3359 SvPV_nolen_const(mysv));
3361 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
3362 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
3363 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
3367 /* Is noper a trieable nodetype that can be merged with the
3368 * current trie (if there is one)? */
3372 ( noper_trietype == NOTHING)
3373 || ( trietype == NOTHING )
3374 || ( trietype == noper_trietype )
3377 && noper_next == tail
3381 /* Handle mergable triable node
3382 * Either we are the first node in a new trieable sequence,
3383 * in which case we do some bookkeeping, otherwise we update
3384 * the end pointer. */
3387 if ( noper_trietype == NOTHING ) {
3388 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
3389 regnode * const noper_next = regnext( noper );
3390 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
3391 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
3394 if ( noper_next_trietype ) {
3395 trietype = noper_next_trietype;
3396 } else if (noper_next_type) {
3397 /* a NOTHING regop is 1 regop wide. We need at least two
3398 * for a trie so we can't merge this in */
3402 trietype = noper_trietype;
3405 if ( trietype == NOTHING )
3406 trietype = noper_trietype;
3411 } /* end handle mergable triable node */
3413 /* handle unmergable node -
3414 * noper may either be a triable node which can not be tried
3415 * together with the current trie, or a non triable node */
3417 /* If last is set and trietype is not NOTHING then we have found
3418 * at least two triable branch sequences in a row of a similar
3419 * trietype so we can turn them into a trie. If/when we
3420 * allow NOTHING to start a trie sequence this condition will be
3421 * required, and it isn't expensive so we leave it in for now. */
3422 if ( trietype != NOTHING )
3423 make_trie( pRExC_state,
3424 startbranch, first, cur, tail, count,
3425 trietype, depth+1 );
3426 last = NULL; /* note: we clear/update first, trietype etc below, so we dont do it here */
3430 && noper_next == tail
3433 /* noper is triable, so we can start a new trie sequence */
3436 trietype = noper_trietype;
3438 /* if we already saw a first but the current node is not triable then we have
3439 * to reset the first information. */
3444 } /* end handle unmergable node */
3445 } /* loop over branches */
3446 DEBUG_TRIE_COMPILE_r({
3447 regprop(RExC_rx, mysv, cur);
3448 PerlIO_printf( Perl_debug_log,
3449 "%*s- %s (%d) <SCAN FINISHED>\n", (int)depth * 2 + 2,
3450 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3454 if ( trietype != NOTHING ) {
3455 /* the last branch of the sequence was part of a trie,
3456 * so we have to construct it here outside of the loop
3458 made= make_trie( pRExC_state, startbranch, first, scan, tail, count, trietype, depth+1 );
3459 #ifdef TRIE_STUDY_OPT
3460 if ( ((made == MADE_EXACT_TRIE &&
3461 startbranch == first)
3462 || ( first_non_open == first )) &&
3464 flags |= SCF_TRIE_RESTUDY;
3465 if ( startbranch == first
3468 RExC_seen &=~REG_TOP_LEVEL_BRANCHES;
3473 /* at this point we know whatever we have is a NOTHING sequence/branch
3474 * AND if 'startbranch' is 'first' then we can turn the whole thing into a NOTHING
3476 if ( startbranch == first ) {
3478 /* the entire thing is a NOTHING sequence, something like this:
3479 * (?:|) So we can turn it into a plain NOTHING op. */
3480 DEBUG_TRIE_COMPILE_r({
3481 regprop(RExC_rx, mysv, cur);
3482 PerlIO_printf( Perl_debug_log,
3483 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
3484 "", SvPV_nolen_const( mysv ),REG_NODE_NUM(cur));
3487 OP(startbranch)= NOTHING;
3488 NEXT_OFF(startbranch)= tail - startbranch;
3489 for ( opt= startbranch + 1; opt < tail ; opt++ )
3493 } /* end if ( last) */
3494 } /* TRIE_MAXBUF is non zero */
3499 else if ( code == BRANCHJ ) { /* single branch is optimized. */
3500 scan = NEXTOPER(NEXTOPER(scan));
3501 } else /* single branch is optimized. */
3502 scan = NEXTOPER(scan);
3504 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
3505 scan_frame *newframe = NULL;
3510 if (OP(scan) != SUSPEND) {
3511 /* set the pointer */
3512 if (OP(scan) == GOSUB) {
3514 RExC_recurse[ARG2L(scan)] = scan;
3515 start = RExC_open_parens[paren-1];
3516 end = RExC_close_parens[paren-1];
3519 start = RExC_rxi->program + 1;
3523 Newxz(recursed, (((RExC_npar)>>3) +1), U8);
3524 SAVEFREEPV(recursed);
3526 if (!PAREN_TEST(recursed,paren+1)) {
3527 PAREN_SET(recursed,paren+1);
3528 Newx(newframe,1,scan_frame);
3530 if (flags & SCF_DO_SUBSTR) {
3531 SCAN_COMMIT(pRExC_state,data,minlenp);
3532 data->longest = &(data->longest_float);
3534 is_inf = is_inf_internal = 1;
3535 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
3536 cl_anything(pRExC_state, data->start_class);
3537 flags &= ~SCF_DO_STCLASS;
3540 Newx(newframe,1,scan_frame);
3543 end = regnext(scan);
3548 SAVEFREEPV(newframe);
3549 newframe->next = regnext(scan);
3550 newframe->last = last;
3551 newframe->stop = stopparen;
3552 newframe->prev = frame;
3562 else if (OP(scan) == EXACT) {
3563 I32 l = STR_LEN(scan);
3566 const U8 * const s = (U8*)STRING(scan);
3567 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3568 l = utf8_length(s, s + l);
3570 uc = *((U8*)STRING(scan));
3573 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
3574 /* The code below prefers earlier match for fixed
3575 offset, later match for variable offset. */
3576 if (data->last_end == -1) { /* Update the start info. */
3577 data->last_start_min = data->pos_min;
3578 data->last_start_max = is_inf
3579 ? I32_MAX : data->pos_min + data->pos_delta;
3581 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
3583 SvUTF8_on(data->last_found);
3585 SV * const sv = data->last_found;
3586 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
3587 mg_find(sv, PERL_MAGIC_utf8) : NULL;
3588 if (mg && mg->mg_len >= 0)
3589 mg->mg_len += utf8_length((U8*)STRING(scan),
3590 (U8*)STRING(scan)+STR_LEN(scan));
3592 data->last_end = data->pos_min + l;
3593 data->pos_min += l; /* As in the first entry. */
3594 data->flags &= ~SF_BEFORE_EOL;
3596 if (flags & SCF_DO_STCLASS_AND) {
3597 /* Check whether it is compatible with what we know already! */
3601 /* If compatible, we or it in below. It is compatible if is
3602 * in the bitmp and either 1) its bit or its fold is set, or 2)
3603 * it's for a locale. Even if there isn't unicode semantics
3604 * here, at runtime there may be because of matching against a
3605 * utf8 string, so accept a possible false positive for
3606 * latin1-range folds */
3608 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3609 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3610 && (!(data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD)
3611 || !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3616 ANYOF_CLASS_ZERO(data->start_class);
3617 ANYOF_BITMAP_ZERO(data->start_class);
3619 ANYOF_BITMAP_SET(data->start_class, uc);
3620 else if (uc >= 0x100) {
3623 /* Some Unicode code points fold to the Latin1 range; as
3624 * XXX temporary code, instead of figuring out if this is
3625 * one, just assume it is and set all the start class bits
3626 * that could be some such above 255 code point's fold
3627 * which will generate fals positives. As the code
3628 * elsewhere that does compute the fold settles down, it
3629 * can be extracted out and re-used here */
3630 for (i = 0; i < 256; i++){
3631 if (HAS_NONLATIN1_FOLD_CLOSURE(i)) {
3632 ANYOF_BITMAP_SET(data->start_class, i);
3636 data->start_class->flags &= ~ANYOF_EOS;
3638 data->start_class->flags &= ~ANYOF_UNICODE_ALL;
3640 else if (flags & SCF_DO_STCLASS_OR) {
3641 /* false positive possible if the class is case-folded */
3643 ANYOF_BITMAP_SET(data->start_class, uc);
3645 data->start_class->flags |= ANYOF_UNICODE_ALL;
3646 data->start_class->flags &= ~ANYOF_EOS;
3647 cl_and(data->start_class, and_withp);
3649 flags &= ~SCF_DO_STCLASS;
3651 else if (PL_regkind[OP(scan)] == EXACT) { /* But OP != EXACT! */
3652 I32 l = STR_LEN(scan);
3653 UV uc = *((U8*)STRING(scan));
3655 /* Search for fixed substrings supports EXACT only. */
3656 if (flags & SCF_DO_SUBSTR) {
3658 SCAN_COMMIT(pRExC_state, data, minlenp);
3661 const U8 * const s = (U8 *)STRING(scan);
3662 uc = utf8_to_uvchr_buf(s, s + l, NULL);
3663 l = utf8_length(s, s + l);
3665 if (has_exactf_sharp_s) {
3666 RExC_seen |= REG_SEEN_EXACTF_SHARP_S;
3668 min += l - min_subtract;
3672 delta += min_subtract;
3673 if (flags & SCF_DO_SUBSTR) {
3674 data->pos_min += l - min_subtract;
3675 if (data->pos_min < 0) {
3678 data->pos_delta += min_subtract;
3680 data->longest = &(data->longest_float);
3683 if (flags & SCF_DO_STCLASS_AND) {
3684 /* Check whether it is compatible with what we know already! */
3687 (!(data->start_class->flags & (ANYOF_CLASS | ANYOF_LOCALE))
3688 && !ANYOF_BITMAP_TEST(data->start_class, uc)
3689 && !ANYOF_BITMAP_TEST(data->start_class, PL_fold_latin1[uc])))
3693 ANYOF_CLASS_ZERO(data->start_class);
3694 ANYOF_BITMAP_ZERO(data->start_class);
3696 ANYOF_BITMAP_SET(data->start_class, uc);
3697 data->start_class->flags &= ~ANYOF_EOS;
3698 data->start_class->flags |= ANYOF_LOC_NONBITMAP_FOLD;
3699 if (OP(scan) == EXACTFL) {
3700 /* XXX This set is probably no longer necessary, and
3701 * probably wrong as LOCALE now is on in the initial
3703 data->start_class->flags |= ANYOF_LOCALE;
3707 /* Also set the other member of the fold pair. In case
3708 * that unicode semantics is called for at runtime, use
3709 * the full latin1 fold. (Can't do this for locale,
3710 * because not known until runtime) */
3711 ANYOF_BITMAP_SET(data->start_class, PL_fold_latin1[uc]);
3713 /* All other (EXACTFL handled above) folds except under
3714 * /iaa that include s, S, and sharp_s also may include
3716 if (OP(scan) != EXACTFA) {
3717 if (uc == 's' || uc == 'S') {
3718 ANYOF_BITMAP_SET(data->start_class,
3719 LATIN_SMALL_LETTER_SHARP_S);
3721 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3722 ANYOF_BITMAP_SET(data->start_class, 's');
3723 ANYOF_BITMAP_SET(data->start_class, 'S');
3728 else if (uc >= 0x100) {
3730 for (i = 0; i < 256; i++){
3731 if (_HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)) {
3732 ANYOF_BITMAP_SET(data->start_class, i);
3737 else if (flags & SCF_DO_STCLASS_OR) {
3738 if (data->start_class->flags & ANYOF_LOC_NONBITMAP_FOLD) {
3739 /* false positive possible if the class is case-folded.
3740 Assume that the locale settings are the same... */
3742 ANYOF_BITMAP_SET(data->start_class, uc);
3743 if (OP(scan) != EXACTFL) {
3745 /* And set the other member of the fold pair, but
3746 * can't do that in locale because not known until
3748 ANYOF_BITMAP_SET(data->start_class,
3749 PL_fold_latin1[uc]);
3751 /* All folds except under /iaa that include s, S,
3752 * and sharp_s also may include the others */
3753 if (OP(scan) != EXACTFA) {
3754 if (uc == 's' || uc == 'S') {
3755 ANYOF_BITMAP_SET(data->start_class,
3756 LATIN_SMALL_LETTER_SHARP_S);
3758 else if (uc == LATIN_SMALL_LETTER_SHARP_S) {
3759 ANYOF_BITMAP_SET(data->start_class, 's');
3760 ANYOF_BITMAP_SET(data->start_class, 'S');
3765 data->start_class->flags &= ~ANYOF_EOS;
3767 cl_and(data->start_class, and_withp);
3769 flags &= ~SCF_DO_STCLASS;
3771 else if (REGNODE_VARIES(OP(scan))) {
3772 I32 mincount, maxcount, minnext, deltanext, fl = 0;
3773 I32 f = flags, pos_before = 0;
3774 regnode * const oscan = scan;
3775 struct regnode_charclass_class this_class;
3776 struct regnode_charclass_class *oclass = NULL;
3777 I32 next_is_eval = 0;
3779 switch (PL_regkind[OP(scan)]) {
3780 case WHILEM: /* End of (?:...)* . */
3781 scan = NEXTOPER(scan);
3784 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
3785 next = NEXTOPER(scan);
3786 if (OP(next) == EXACT || (flags & SCF_DO_STCLASS)) {
3788 maxcount = REG_INFTY;
3789 next = regnext(scan);
3790 scan = NEXTOPER(scan);
3794 if (flags & SCF_DO_SUBSTR)
3799 if (flags & SCF_DO_STCLASS) {
3801 maxcount = REG_INFTY;
3802 next = regnext(scan);
3803 scan = NEXTOPER(scan);
3806 is_inf = is_inf_internal = 1;
3807 scan = regnext(scan);
3808 if (flags & SCF_DO_SUBSTR) {
3809 SCAN_COMMIT(pRExC_state, data, minlenp); /* Cannot extend fixed substrings */
3810 data->longest = &(data->longest_float);
3812 goto optimize_curly_tail;
3814 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
3815 && (scan->flags == stopparen))
3820 mincount = ARG1(scan);
3821 maxcount = ARG2(scan);
3823 next = regnext(scan);
3824 if (OP(scan) == CURLYX) {
3825 I32 lp = (data ? *(data->last_closep) : 0);
3826 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
3828 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
3829 next_is_eval = (OP(scan) == EVAL);
3831 if (flags & SCF_DO_SUBSTR) {
3832 if (mincount == 0) SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot extend fixed substrings */
3833 pos_before = data->pos_min;
3837 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
3839 data->flags |= SF_IS_INF;
3841 if (flags & SCF_DO_STCLASS) {
3842 cl_init(pRExC_state, &this_class);
3843 oclass = data->start_class;
3844 data->start_class = &this_class;
3845 f |= SCF_DO_STCLASS_AND;
3846 f &= ~SCF_DO_STCLASS_OR;
3848 /* Exclude from super-linear cache processing any {n,m}
3849 regops for which the combination of input pos and regex
3850 pos is not enough information to determine if a match
3853 For example, in the regex /foo(bar\s*){4,8}baz/ with the
3854 regex pos at the \s*, the prospects for a match depend not
3855 only on the input position but also on how many (bar\s*)
3856 repeats into the {4,8} we are. */
3857 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
3858 f &= ~SCF_WHILEM_VISITED_POS;
3860 /* This will finish on WHILEM, setting scan, or on NULL: */
3861 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
3862 last, data, stopparen, recursed, NULL,
3864 ? (f & ~SCF_DO_SUBSTR) : f),depth+1);
3866 if (flags & SCF_DO_STCLASS)
3867 data->start_class = oclass;
3868 if (mincount == 0 || minnext == 0) {
3869 if (flags & SCF_DO_STCLASS_OR) {
3870 cl_or(pRExC_state, data->start_class, &this_class);
3872 else if (flags & SCF_DO_STCLASS_AND) {
3873 /* Switch to OR mode: cache the old value of
3874 * data->start_class */
3876 StructCopy(data->start_class, and_withp,
3877 struct regnode_charclass_class);
3878 flags &= ~SCF_DO_STCLASS_AND;
3879 StructCopy(&this_class, data->start_class,
3880 struct regnode_charclass_class);
3881 flags |= SCF_DO_STCLASS_OR;
3882 data->start_class->flags |= ANYOF_EOS;
3884 } else { /* Non-zero len */
3885 if (flags & SCF_DO_STCLASS_OR) {
3886 cl_or(pRExC_state, data->start_class, &this_class);
3887 cl_and(data->start_class, and_withp);
3889 else if (flags & SCF_DO_STCLASS_AND)
3890 cl_and(data->start_class, &this_class);
3891 flags &= ~SCF_DO_STCLASS;
3893 if (!scan) /* It was not CURLYX, but CURLY. */
3895 if ( /* ? quantifier ok, except for (?{ ... }) */
3896 (next_is_eval || !(mincount == 0 && maxcount == 1))
3897 && (minnext == 0) && (deltanext == 0)
3898 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
3899 && maxcount <= REG_INFTY/3) /* Complement check for big count */
3901 ckWARNreg(RExC_parse,
3902 "Quantifier unexpected on zero-length expression");
3905 min += minnext * mincount;
3906 is_inf_internal |= ((maxcount == REG_INFTY
3907 && (minnext + deltanext) > 0)
3908 || deltanext == I32_MAX);
3909 is_inf |= is_inf_internal;
3910 delta += (minnext + deltanext) * maxcount - minnext * mincount;
3912 /* Try powerful optimization CURLYX => CURLYN. */
3913 if ( OP(oscan) == CURLYX && data
3914 && data->flags & SF_IN_PAR
3915 && !(data->flags & SF_HAS_EVAL)
3916 && !deltanext && minnext == 1 ) {
3917 /* Try to optimize to CURLYN. */
3918 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
3919 regnode * const nxt1 = nxt;
3926 if (!REGNODE_SIMPLE(OP(nxt))
3927 && !(PL_regkind[OP(nxt)] == EXACT
3928 && STR_LEN(nxt) == 1))
3934 if (OP(nxt) != CLOSE)
3936 if (RExC_open_parens) {
3937 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3938 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
3940 /* Now we know that nxt2 is the only contents: */
3941 oscan->flags = (U8)ARG(nxt);
3943 OP(nxt1) = NOTHING; /* was OPEN. */
3946 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3947 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
3948 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
3949 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3950 OP(nxt + 1) = OPTIMIZED; /* was count. */
3951 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
3956 /* Try optimization CURLYX => CURLYM. */
3957 if ( OP(oscan) == CURLYX && data
3958 && !(data->flags & SF_HAS_PAR)
3959 && !(data->flags & SF_HAS_EVAL)
3960 && !deltanext /* atom is fixed width */
3961 && minnext != 0 /* CURLYM can't handle zero width */
3962 && ! (RExC_seen & REG_SEEN_EXACTF_SHARP_S) /* Nor \xDF */
3964 /* XXXX How to optimize if data == 0? */
3965 /* Optimize to a simpler form. */
3966 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
3970 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
3971 && (OP(nxt2) != WHILEM))
3973 OP(nxt2) = SUCCEED; /* Whas WHILEM */
3974 /* Need to optimize away parenths. */
3975 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
3976 /* Set the parenth number. */
3977 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
3979 oscan->flags = (U8)ARG(nxt);
3980 if (RExC_open_parens) {
3981 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
3982 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
3984 OP(nxt1) = OPTIMIZED; /* was OPEN. */
3985 OP(nxt) = OPTIMIZED; /* was CLOSE. */
3988 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
3989 OP(nxt + 1) = OPTIMIZED; /* was count. */
3990 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
3991 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
3994 while ( nxt1 && (OP(nxt1) != WHILEM)) {
3995 regnode *nnxt = regnext(nxt1);
3997 if (reg_off_by_arg[OP(nxt1)])
3998 ARG_SET(nxt1, nxt2 - nxt1);
3999 else if (nxt2 - nxt1 < U16_MAX)
4000 NEXT_OFF(nxt1) = nxt2 - nxt1;
4002 OP(nxt) = NOTHING; /* Cannot beautify */
4007 /* Optimize again: */
4008 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4009 NULL, stopparen, recursed, NULL, 0,depth+1);
4014 else if ((OP(oscan) == CURLYX)
4015 && (flags & SCF_WHILEM_VISITED_POS)
4016 /* See the comment on a similar expression above.
4017 However, this time it's not a subexpression
4018 we care about, but the expression itself. */
4019 && (maxcount == REG_INFTY)
4020 && data && ++data->whilem_c < 16) {
4021 /* This stays as CURLYX, we can put the count/of pair. */
4022 /* Find WHILEM (as in regexec.c) */
4023 regnode *nxt = oscan + NEXT_OFF(oscan);
4025 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4027 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4028 | (RExC_whilem_seen << 4)); /* On WHILEM */
4030 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4032 if (flags & SCF_DO_SUBSTR) {
4033 SV *last_str = NULL;
4034 int counted = mincount != 0;
4036 if (data->last_end > 0 && mincount != 0) { /* Ends with a string. */
4037 #if defined(SPARC64_GCC_WORKAROUND)
4040 const char *s = NULL;
4043 if (pos_before >= data->last_start_min)
4046 b = data->last_start_min;
4049 s = SvPV_const(data->last_found, l);
4050 old = b - data->last_start_min;
4053 I32 b = pos_before >= data->last_start_min
4054 ? pos_before : data->last_start_min;
4056 const char * const s = SvPV_const(data->last_found, l);
4057 I32 old = b - data->last_start_min;
4061 old = utf8_hop((U8*)s, old) - (U8*)s;
4063 /* Get the added string: */
4064 last_str = newSVpvn_utf8(s + old, l, UTF);
4065 if (deltanext == 0 && pos_before == b) {
4066 /* What was added is a constant string */
4068 SvGROW(last_str, (mincount * l) + 1);
4069 repeatcpy(SvPVX(last_str) + l,
4070 SvPVX_const(last_str), l, mincount - 1);
4071 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4072 /* Add additional parts. */
4073 SvCUR_set(data->last_found,
4074 SvCUR(data->last_found) - l);
4075 sv_catsv(data->last_found, last_str);
4077 SV * sv = data->last_found;
4079 SvUTF8(sv) && SvMAGICAL(sv) ?
4080 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4081 if (mg && mg->mg_len >= 0)
4082 mg->mg_len += CHR_SVLEN(last_str) - l;
4084 data->last_end += l * (mincount - 1);
4087 /* start offset must point into the last copy */
4088 data->last_start_min += minnext * (mincount - 1);
4089 data->last_start_max += is_inf ? I32_MAX
4090 : (maxcount - 1) * (minnext + data->pos_delta);
4093 /* It is counted once already... */
4094 data->pos_min += minnext * (mincount - counted);
4095 data->pos_delta += - counted * deltanext +
4096 (minnext + deltanext) * maxcount - minnext * mincount;
4097 if (mincount != maxcount) {
4098 /* Cannot extend fixed substrings found inside
4100 SCAN_COMMIT(pRExC_state,data,minlenp);
4101 if (mincount && last_str) {
4102 SV * const sv = data->last_found;
4103 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4104 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4108 sv_setsv(sv, last_str);
4109 data->last_end = data->pos_min;
4110 data->last_start_min =
4111 data->pos_min - CHR_SVLEN(last_str);
4112 data->last_start_max = is_inf
4114 : data->pos_min + data->pos_delta
4115 - CHR_SVLEN(last_str);
4117 data->longest = &(data->longest_float);
4119 SvREFCNT_dec(last_str);
4121 if (data && (fl & SF_HAS_EVAL))
4122 data->flags |= SF_HAS_EVAL;
4123 optimize_curly_tail:
4124 if (OP(oscan) != CURLYX) {
4125 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4127 NEXT_OFF(oscan) += NEXT_OFF(next);
4130 default: /* REF, ANYOFV, and CLUMP only? */
4131 if (flags & SCF_DO_SUBSTR) {
4132 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4133 data->longest = &(data->longest_float);
4135 is_inf = is_inf_internal = 1;
4136 if (flags & SCF_DO_STCLASS_OR)
4137 cl_anything(pRExC_state, data->start_class);
4138 flags &= ~SCF_DO_STCLASS;
4142 else if (OP(scan) == LNBREAK) {
4143 if (flags & SCF_DO_STCLASS) {
4145 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
4146 if (flags & SCF_DO_STCLASS_AND) {
4147 for (value = 0; value < 256; value++)
4148 if (!is_VERTWS_cp(value))
4149 ANYOF_BITMAP_CLEAR(data->start_class, value);
4152 for (value = 0; value < 256; value++)
4153 if (is_VERTWS_cp(value))
4154 ANYOF_BITMAP_SET(data->start_class, value);
4156 if (flags & SCF_DO_STCLASS_OR)
4157 cl_and(data->start_class, and_withp);
4158 flags &= ~SCF_DO_STCLASS;
4162 if (flags & SCF_DO_SUBSTR) {
4163 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4165 data->pos_delta += 1;
4166 data->longest = &(data->longest_float);
4169 else if (REGNODE_SIMPLE(OP(scan))) {
4172 if (flags & SCF_DO_SUBSTR) {
4173 SCAN_COMMIT(pRExC_state,data,minlenp);
4177 if (flags & SCF_DO_STCLASS) {
4178 data->start_class->flags &= ~ANYOF_EOS; /* No match on empty */
4180 /* Some of the logic below assumes that switching
4181 locale on will only add false positives. */
4182 switch (PL_regkind[OP(scan)]) {
4186 /* Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d", OP(scan)); */
4187 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4188 cl_anything(pRExC_state, data->start_class);
4191 if (OP(scan) == SANY)
4193 if (flags & SCF_DO_STCLASS_OR) { /* Everything but \n */
4194 value = (ANYOF_BITMAP_TEST(data->start_class,'\n')
4195 || ANYOF_CLASS_TEST_ANY_SET(data->start_class));
4196 cl_anything(pRExC_state, data->start_class);
4198 if (flags & SCF_DO_STCLASS_AND || !value)
4199 ANYOF_BITMAP_CLEAR(data->start_class,'\n');
4202 if (flags & SCF_DO_STCLASS_AND)
4203 cl_and(data->start_class,
4204 (struct regnode_charclass_class*)scan);
4206 cl_or(pRExC_state, data->start_class,
4207 (struct regnode_charclass_class*)scan);
4210 if (flags & SCF_DO_STCLASS_AND) {
4211 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4212 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NALNUM);
4213 if (OP(scan) == ALNUMU) {
4214 for (value = 0; value < 256; value++) {
4215 if (!isWORDCHAR_L1(value)) {
4216 ANYOF_BITMAP_CLEAR(data->start_class, value);
4220 for (value = 0; value < 256; value++) {
4221 if (!isALNUM(value)) {
4222 ANYOF_BITMAP_CLEAR(data->start_class, value);
4229 if (data->start_class->flags & ANYOF_LOCALE)
4230 ANYOF_CLASS_SET(data->start_class,ANYOF_ALNUM);
4232 /* Even if under locale, set the bits for non-locale
4233 * in case it isn't a true locale-node. This will
4234 * create false positives if it truly is locale */
4235 if (OP(scan) == ALNUMU) {
4236 for (value = 0; value < 256; value++) {
4237 if (isWORDCHAR_L1(value)) {
4238 ANYOF_BITMAP_SET(data->start_class, value);
4242 for (value = 0; value < 256; value++) {
4243 if (isALNUM(value)) {
4244 ANYOF_BITMAP_SET(data->start_class, value);
4251 if (flags & SCF_DO_STCLASS_AND) {
4252 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4253 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_ALNUM);
4254 if (OP(scan) == NALNUMU) {
4255 for (value = 0; value < 256; value++) {
4256 if (isWORDCHAR_L1(value)) {
4257 ANYOF_BITMAP_CLEAR(data->start_class, value);
4261 for (value = 0; value < 256; value++) {
4262 if (isALNUM(value)) {
4263 ANYOF_BITMAP_CLEAR(data->start_class, value);
4270 if (data->start_class->flags & ANYOF_LOCALE)
4271 ANYOF_CLASS_SET(data->start_class,ANYOF_NALNUM);
4273 /* Even if under locale, set the bits for non-locale in
4274 * case it isn't a true locale-node. This will create
4275 * false positives if it truly is locale */
4276 if (OP(scan) == NALNUMU) {
4277 for (value = 0; value < 256; value++) {
4278 if (! isWORDCHAR_L1(value)) {
4279 ANYOF_BITMAP_SET(data->start_class, value);
4283 for (value = 0; value < 256; value++) {
4284 if (! isALNUM(value)) {
4285 ANYOF_BITMAP_SET(data->start_class, value);
4292 if (flags & SCF_DO_STCLASS_AND) {
4293 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4294 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NSPACE);
4295 if (OP(scan) == SPACEU) {
4296 for (value = 0; value < 256; value++) {
4297 if (!isSPACE_L1(value)) {
4298 ANYOF_BITMAP_CLEAR(data->start_class, value);
4302 for (value = 0; value < 256; value++) {
4303 if (!isSPACE(value)) {
4304 ANYOF_BITMAP_CLEAR(data->start_class, value);
4311 if (data->start_class->flags & ANYOF_LOCALE) {
4312 ANYOF_CLASS_SET(data->start_class,ANYOF_SPACE);
4314 if (OP(scan) == SPACEU) {
4315 for (value = 0; value < 256; value++) {
4316 if (isSPACE_L1(value)) {
4317 ANYOF_BITMAP_SET(data->start_class, value);
4321 for (value = 0; value < 256; value++) {
4322 if (isSPACE(value)) {
4323 ANYOF_BITMAP_SET(data->start_class, value);
4330 if (flags & SCF_DO_STCLASS_AND) {
4331 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4332 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_SPACE);
4333 if (OP(scan) == NSPACEU) {
4334 for (value = 0; value < 256; value++) {
4335 if (isSPACE_L1(value)) {
4336 ANYOF_BITMAP_CLEAR(data->start_class, value);
4340 for (value = 0; value < 256; value++) {
4341 if (isSPACE(value)) {
4342 ANYOF_BITMAP_CLEAR(data->start_class, value);
4349 if (data->start_class->flags & ANYOF_LOCALE)
4350 ANYOF_CLASS_SET(data->start_class,ANYOF_NSPACE);
4351 if (OP(scan) == NSPACEU) {
4352 for (value = 0; value < 256; value++) {
4353 if (!isSPACE_L1(value)) {
4354 ANYOF_BITMAP_SET(data->start_class, value);
4359 for (value = 0; value < 256; value++) {
4360 if (!isSPACE(value)) {
4361 ANYOF_BITMAP_SET(data->start_class, value);
4368 if (flags & SCF_DO_STCLASS_AND) {
4369 if (!(data->start_class->flags & ANYOF_LOCALE)) {
4370 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_NDIGIT);
4371 for (value = 0; value < 256; value++)
4372 if (!isDIGIT(value))
4373 ANYOF_BITMAP_CLEAR(data->start_class, value);
4377 if (data->start_class->flags & ANYOF_LOCALE)
4378 ANYOF_CLASS_SET(data->start_class,ANYOF_DIGIT);
4379 for (value = 0; value < 256; value++)
4381 ANYOF_BITMAP_SET(data->start_class, value);
4385 if (flags & SCF_DO_STCLASS_AND) {
4386 if (!(data->start_class->flags & ANYOF_LOCALE))
4387 ANYOF_CLASS_CLEAR(data->start_class,ANYOF_DIGIT);
4388 for (value = 0; value < 256; value++)
4390 ANYOF_BITMAP_CLEAR(data->start_class, value);
4393 if (data->start_class->flags & ANYOF_LOCALE)
4394 ANYOF_CLASS_SET(data->start_class,ANYOF_NDIGIT);
4395 for (value = 0; value < 256; value++)
4396 if (!isDIGIT(value))
4397 ANYOF_BITMAP_SET(data->start_class, value);
4400 CASE_SYNST_FNC(VERTWS);
4401 CASE_SYNST_FNC(HORIZWS);
4404 if (flags & SCF_DO_STCLASS_OR)
4405 cl_and(data->start_class, and_withp);
4406 flags &= ~SCF_DO_STCLASS;
4409 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
4410 data->flags |= (OP(scan) == MEOL
4413 SCAN_COMMIT(pRExC_state, data, minlenp);
4416 else if ( PL_regkind[OP(scan)] == BRANCHJ
4417 /* Lookbehind, or need to calculate parens/evals/stclass: */
4418 && (scan->flags || data || (flags & SCF_DO_STCLASS))
4419 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM)) {
4420 if ( OP(scan) == UNLESSM &&
4422 OP(NEXTOPER(NEXTOPER(scan))) == NOTHING &&
4423 OP(regnext(NEXTOPER(NEXTOPER(scan)))) == SUCCEED
4426 regnode *upto= regnext(scan);
4428 SV * const mysv_val=sv_newmortal();
4429 DEBUG_STUDYDATA("OPFAIL",data,depth);
4431 /*DEBUG_PARSE_MSG("opfail");*/
4432 regprop(RExC_rx, mysv_val, upto);
4433 PerlIO_printf(Perl_debug_log, "~ replace with OPFAIL pointed at %s (%"IVdf") offset %"IVdf"\n",
4434 SvPV_nolen_const(mysv_val),
4435 (IV)REG_NODE_NUM(upto),
4440 NEXT_OFF(scan) = upto - scan;
4441 for (opt= scan + 1; opt < upto ; opt++)
4442 OP(opt) = OPTIMIZED;
4446 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4447 || OP(scan) == UNLESSM )
4449 /* Negative Lookahead/lookbehind
4450 In this case we can't do fixed string optimisation.
4453 I32 deltanext, minnext, fake = 0;
4455 struct regnode_charclass_class intrnl;
4458 data_fake.flags = 0;
4460 data_fake.whilem_c = data->whilem_c;
4461 data_fake.last_closep = data->last_closep;
4464 data_fake.last_closep = &fake;
4465 data_fake.pos_delta = delta;
4466 if ( flags & SCF_DO_STCLASS && !scan->flags
4467 && OP(scan) == IFMATCH ) { /* Lookahead */
4468 cl_init(pRExC_state, &intrnl);
4469 data_fake.start_class = &intrnl;
4470 f |= SCF_DO_STCLASS_AND;
4472 if (flags & SCF_WHILEM_VISITED_POS)
4473 f |= SCF_WHILEM_VISITED_POS;
4474 next = regnext(scan);
4475 nscan = NEXTOPER(NEXTOPER(scan));
4476 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
4477 last, &data_fake, stopparen, recursed, NULL, f, depth+1);
4480 FAIL("Variable length lookbehind not implemented");
4482 else if (minnext > (I32)U8_MAX) {
4483 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4485 scan->flags = (U8)minnext;
4488 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4490 if (data_fake.flags & SF_HAS_EVAL)
4491 data->flags |= SF_HAS_EVAL;
4492 data->whilem_c = data_fake.whilem_c;
4494 if (f & SCF_DO_STCLASS_AND) {
4495 if (flags & SCF_DO_STCLASS_OR) {
4496 /* OR before, AND after: ideally we would recurse with
4497 * data_fake to get the AND applied by study of the
4498 * remainder of the pattern, and then derecurse;
4499 * *** HACK *** for now just treat as "no information".
4500 * See [perl #56690].
4502 cl_init(pRExC_state, data->start_class);
4504 /* AND before and after: combine and continue */
4505 const int was = (data->start_class->flags & ANYOF_EOS);
4507 cl_and(data->start_class, &intrnl);
4509 data->start_class->flags |= ANYOF_EOS;
4513 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
4515 /* Positive Lookahead/lookbehind
4516 In this case we can do fixed string optimisation,
4517 but we must be careful about it. Note in the case of
4518 lookbehind the positions will be offset by the minimum
4519 length of the pattern, something we won't know about
4520 until after the recurse.
4522 I32 deltanext, fake = 0;
4524 struct regnode_charclass_class intrnl;
4526 /* We use SAVEFREEPV so that when the full compile
4527 is finished perl will clean up the allocated
4528 minlens when it's all done. This way we don't
4529 have to worry about freeing them when we know
4530 they wont be used, which would be a pain.
4533 Newx( minnextp, 1, I32 );
4534 SAVEFREEPV(minnextp);
4537 StructCopy(data, &data_fake, scan_data_t);
4538 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
4541 SCAN_COMMIT(pRExC_state, &data_fake,minlenp);
4542 data_fake.last_found=newSVsv(data->last_found);
4546 data_fake.last_closep = &fake;
4547 data_fake.flags = 0;
4548 data_fake.pos_delta = delta;
4550 data_fake.flags |= SF_IS_INF;
4551 if ( flags & SCF_DO_STCLASS && !scan->flags
4552 && OP(scan) == IFMATCH ) { /* Lookahead */
4553 cl_init(pRExC_state, &intrnl);
4554 data_fake.start_class = &intrnl;
4555 f |= SCF_DO_STCLASS_AND;
4557 if (flags & SCF_WHILEM_VISITED_POS)
4558 f |= SCF_WHILEM_VISITED_POS;
4559 next = regnext(scan);
4560 nscan = NEXTOPER(NEXTOPER(scan));
4562 *minnextp = study_chunk(pRExC_state, &nscan, minnextp, &deltanext,
4563 last, &data_fake, stopparen, recursed, NULL, f,depth+1);
4566 FAIL("Variable length lookbehind not implemented");
4568 else if (*minnextp > (I32)U8_MAX) {
4569 FAIL2("Lookbehind longer than %"UVuf" not implemented", (UV)U8_MAX);
4571 scan->flags = (U8)*minnextp;
4576 if (f & SCF_DO_STCLASS_AND) {
4577 const int was = (data->start_class->flags & ANYOF_EOS);
4579 cl_and(data->start_class, &intrnl);
4581 data->start_class->flags |= ANYOF_EOS;
4584 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4586 if (data_fake.flags & SF_HAS_EVAL)
4587 data->flags |= SF_HAS_EVAL;
4588 data->whilem_c = data_fake.whilem_c;
4589 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
4590 if (RExC_rx->minlen<*minnextp)
4591 RExC_rx->minlen=*minnextp;
4592 SCAN_COMMIT(pRExC_state, &data_fake, minnextp);
4593 SvREFCNT_dec(data_fake.last_found);
4595 if ( data_fake.minlen_fixed != minlenp )
4597 data->offset_fixed= data_fake.offset_fixed;
4598 data->minlen_fixed= data_fake.minlen_fixed;
4599 data->lookbehind_fixed+= scan->flags;
4601 if ( data_fake.minlen_float != minlenp )
4603 data->minlen_float= data_fake.minlen_float;
4604 data->offset_float_min=data_fake.offset_float_min;
4605 data->offset_float_max=data_fake.offset_float_max;
4606 data->lookbehind_float+= scan->flags;
4613 else if (OP(scan) == OPEN) {
4614 if (stopparen != (I32)ARG(scan))
4617 else if (OP(scan) == CLOSE) {
4618 if (stopparen == (I32)ARG(scan)) {
4621 if ((I32)ARG(scan) == is_par) {
4622 next = regnext(scan);
4624 if ( next && (OP(next) != WHILEM) && next < last)
4625 is_par = 0; /* Disable optimization */
4628 *(data->last_closep) = ARG(scan);
4630 else if (OP(scan) == EVAL) {
4632 data->flags |= SF_HAS_EVAL;
4634 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
4635 if (flags & SCF_DO_SUBSTR) {
4636 SCAN_COMMIT(pRExC_state,data,minlenp);
4637 flags &= ~SCF_DO_SUBSTR;
4639 if (data && OP(scan)==ACCEPT) {
4640 data->flags |= SCF_SEEN_ACCEPT;
4645 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
4647 if (flags & SCF_DO_SUBSTR) {
4648 SCAN_COMMIT(pRExC_state,data,minlenp);
4649 data->longest = &(data->longest_float);
4651 is_inf = is_inf_internal = 1;
4652 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4653 cl_anything(pRExC_state, data->start_class);
4654 flags &= ~SCF_DO_STCLASS;
4656 else if (OP(scan) == GPOS) {
4657 if (!(RExC_rx->extflags & RXf_GPOS_FLOAT) &&
4658 !(delta || is_inf || (data && data->pos_delta)))
4660 if (!(RExC_rx->extflags & RXf_ANCH) && (flags & SCF_DO_SUBSTR))
4661 RExC_rx->extflags |= RXf_ANCH_GPOS;
4662 if (RExC_rx->gofs < (U32)min)
4663 RExC_rx->gofs = min;
4665 RExC_rx->extflags |= RXf_GPOS_FLOAT;
4669 #ifdef TRIE_STUDY_OPT
4670 #ifdef FULL_TRIE_STUDY
4671 else if (PL_regkind[OP(scan)] == TRIE) {
4672 /* NOTE - There is similar code to this block above for handling
4673 BRANCH nodes on the initial study. If you change stuff here
4675 regnode *trie_node= scan;
4676 regnode *tail= regnext(scan);
4677 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4678 I32 max1 = 0, min1 = I32_MAX;
4679 struct regnode_charclass_class accum;
4681 if (flags & SCF_DO_SUBSTR) /* XXXX Add !SUSPEND? */
4682 SCAN_COMMIT(pRExC_state, data,minlenp); /* Cannot merge strings after this. */
4683 if (flags & SCF_DO_STCLASS)
4684 cl_init_zero(pRExC_state, &accum);
4690 const regnode *nextbranch= NULL;
4693 for ( word=1 ; word <= trie->wordcount ; word++)
4695 I32 deltanext=0, minnext=0, f = 0, fake;
4696 struct regnode_charclass_class this_class;
4698 data_fake.flags = 0;
4700 data_fake.whilem_c = data->whilem_c;
4701 data_fake.last_closep = data->last_closep;
4704 data_fake.last_closep = &fake;
4705 data_fake.pos_delta = delta;
4706 if (flags & SCF_DO_STCLASS) {
4707 cl_init(pRExC_state, &this_class);
4708 data_fake.start_class = &this_class;
4709 f = SCF_DO_STCLASS_AND;
4711 if (flags & SCF_WHILEM_VISITED_POS)
4712 f |= SCF_WHILEM_VISITED_POS;
4714 if (trie->jump[word]) {
4716 nextbranch = trie_node + trie->jump[0];
4717 scan= trie_node + trie->jump[word];
4718 /* We go from the jump point to the branch that follows
4719 it. Note this means we need the vestigal unused branches
4720 even though they arent otherwise used.
4722 minnext = study_chunk(pRExC_state, &scan, minlenp,
4723 &deltanext, (regnode *)nextbranch, &data_fake,
4724 stopparen, recursed, NULL, f,depth+1);
4726 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
4727 nextbranch= regnext((regnode*)nextbranch);
4729 if (min1 > (I32)(minnext + trie->minlen))
4730 min1 = minnext + trie->minlen;
4731 if (max1 < (I32)(minnext + deltanext + trie->maxlen))
4732 max1 = minnext + deltanext + trie->maxlen;
4733 if (deltanext == I32_MAX)
4734 is_inf = is_inf_internal = 1;
4736 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
4738 if (data_fake.flags & SCF_SEEN_ACCEPT) {
4739 if ( stopmin > min + min1)
4740 stopmin = min + min1;
4741 flags &= ~SCF_DO_SUBSTR;
4743 data->flags |= SCF_SEEN_ACCEPT;
4746 if (data_fake.flags & SF_HAS_EVAL)
4747 data->flags |= SF_HAS_EVAL;
4748 data->whilem_c = data_fake.whilem_c;
4750 if (flags & SCF_DO_STCLASS)
4751 cl_or(pRExC_state, &accum, &this_class);
4754 if (flags & SCF_DO_SUBSTR) {
4755 data->pos_min += min1;
4756 data->pos_delta += max1 - min1;
4757 if (max1 != min1 || is_inf)
4758 data->longest = &(data->longest_float);
4761 delta += max1 - min1;
4762 if (flags & SCF_DO_STCLASS_OR) {
4763 cl_or(pRExC_state, data->start_class, &accum);
4765 cl_and(data->start_class, and_withp);
4766 flags &= ~SCF_DO_STCLASS;
4769 else if (flags & SCF_DO_STCLASS_AND) {
4771 cl_and(data->start_class, &accum);
4772 flags &= ~SCF_DO_STCLASS;
4775 /* Switch to OR mode: cache the old value of
4776 * data->start_class */
4778 StructCopy(data->start_class, and_withp,
4779 struct regnode_charclass_class);
4780 flags &= ~SCF_DO_STCLASS_AND;
4781 StructCopy(&accum, data->start_class,
4782 struct regnode_charclass_class);
4783 flags |= SCF_DO_STCLASS_OR;
4784 data->start_class->flags |= ANYOF_EOS;
4791 else if (PL_regkind[OP(scan)] == TRIE) {
4792 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
4795 min += trie->minlen;
4796 delta += (trie->maxlen - trie->minlen);
4797 flags &= ~SCF_DO_STCLASS; /* xxx */
4798 if (flags & SCF_DO_SUBSTR) {
4799 SCAN_COMMIT(pRExC_state,data,minlenp); /* Cannot expect anything... */
4800 data->pos_min += trie->minlen;
4801 data->pos_delta += (trie->maxlen - trie->minlen);
4802 if (trie->maxlen != trie->minlen)
4803 data->longest = &(data->longest_float);
4805 if (trie->jump) /* no more substrings -- for now /grr*/
4806 flags &= ~SCF_DO_SUBSTR;
4808 #endif /* old or new */
4809 #endif /* TRIE_STUDY_OPT */
4811 /* Else: zero-length, ignore. */
4812 scan = regnext(scan);
4817 stopparen = frame->stop;
4818 frame = frame->prev;
4819 goto fake_study_recurse;
4824 DEBUG_STUDYDATA("pre-fin:",data,depth);
4827 *deltap = is_inf_internal ? I32_MAX : delta;
4828 if (flags & SCF_DO_SUBSTR && is_inf)
4829 data->pos_delta = I32_MAX - data->pos_min;
4830 if (is_par > (I32)U8_MAX)
4832 if (is_par && pars==1 && data) {
4833 data->flags |= SF_IN_PAR;
4834 data->flags &= ~SF_HAS_PAR;
4836 else if (pars && data) {
4837 data->flags |= SF_HAS_PAR;
4838 data->flags &= ~SF_IN_PAR;
4840 if (flags & SCF_DO_STCLASS_OR)
4841 cl_and(data->start_class, and_withp);
4842 if (flags & SCF_TRIE_RESTUDY)
4843 data->flags |= SCF_TRIE_RESTUDY;
4845 DEBUG_STUDYDATA("post-fin:",data,depth);
4847 return min < stopmin ? min : stopmin;
4851 S_add_data(RExC_state_t *pRExC_state, U32 n, const char *s)
4853 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
4855 PERL_ARGS_ASSERT_ADD_DATA;
4857 Renewc(RExC_rxi->data,
4858 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
4859 char, struct reg_data);
4861 Renew(RExC_rxi->data->what, count + n, U8);
4863 Newx(RExC_rxi->data->what, n, U8);
4864 RExC_rxi->data->count = count + n;
4865 Copy(s, RExC_rxi->data->what + count, n, U8);
4869 /*XXX: todo make this not included in a non debugging perl */
4870 #ifndef PERL_IN_XSUB_RE
4872 Perl_reginitcolors(pTHX)
4875 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
4877 char *t = savepv(s);
4881 t = strchr(t, '\t');
4887 PL_colors[i] = t = (char *)"";
4892 PL_colors[i++] = (char *)"";
4899 #ifdef TRIE_STUDY_OPT
4900 #define CHECK_RESTUDY_GOTO \
4902 (data.flags & SCF_TRIE_RESTUDY) \
4906 #define CHECK_RESTUDY_GOTO
4910 * pregcomp - compile a regular expression into internal code
4912 * Decides which engine's compiler to call based on the hint currently in
4916 #ifndef PERL_IN_XSUB_RE
4918 /* return the currently in-scope regex engine (or the default if none) */
4920 regexp_engine const *
4921 Perl_current_re_engine(pTHX)
4925 if (IN_PERL_COMPILETIME) {
4926 HV * const table = GvHV(PL_hintgv);
4930 return &PL_core_reg_engine;
4931 ptr = hv_fetchs(table, "regcomp", FALSE);
4932 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
4933 return &PL_core_reg_engine;
4934 return INT2PTR(regexp_engine*,SvIV(*ptr));
4938 if (!PL_curcop->cop_hints_hash)
4939 return &PL_core_reg_engine;
4940 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
4941 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
4942 return &PL_core_reg_engine;
4943 return INT2PTR(regexp_engine*,SvIV(ptr));
4949 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
4952 regexp_engine const *eng = current_re_engine();
4953 GET_RE_DEBUG_FLAGS_DECL;
4955 PERL_ARGS_ASSERT_PREGCOMP;
4957 /* Dispatch a request to compile a regexp to correct regexp engine. */
4959 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
4962 return CALLREGCOMP_ENG(eng, pattern, flags);
4966 /* public(ish) entry point for the perl core's own regex compiling code.
4967 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
4968 * pattern rather than a list of OPs, and uses the internal engine rather
4969 * than the current one */
4972 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
4974 SV *pat = pattern; /* defeat constness! */
4975 PERL_ARGS_ASSERT_RE_COMPILE;
4976 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
4977 #ifdef PERL_IN_XSUB_RE
4980 &PL_core_reg_engine,
4982 NULL, NULL, rx_flags, 0);
4985 /* see if there are any run-time code blocks in the pattern.
4986 * False positives are allowed */
4989 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state, OP *expr,
4990 U32 pm_flags, char *pat, STRLEN plen)
4995 /* avoid infinitely recursing when we recompile the pattern parcelled up
4996 * as qr'...'. A single constant qr// string can't have have any
4997 * run-time component in it, and thus, no runtime code. (A non-qr
4998 * string, however, can, e.g. $x =~ '(?{})') */
4999 if ((pm_flags & PMf_IS_QR) && expr && expr->op_type == OP_CONST)
5002 for (s = 0; s < plen; s++) {
5003 if (n < pRExC_state->num_code_blocks
5004 && s == pRExC_state->code_blocks[n].start)
5006 s = pRExC_state->code_blocks[n].end;
5010 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
5012 if (pat[s] == '(' && pat[s+1] == '?' &&
5013 (pat[s+2] == '{' || (pat[s+2] == '?' && pat[s+3] == '{'))
5020 /* Handle run-time code blocks. We will already have compiled any direct
5021 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
5022 * copy of it, but with any literal code blocks blanked out and
5023 * appropriate chars escaped; then feed it into
5025 * eval "qr'modified_pattern'"
5029 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
5033 * qr'a\\bc def\'ghi\\\\jkl(?{"this is runtime"})mno'
5035 * After eval_sv()-ing that, grab any new code blocks from the returned qr
5036 * and merge them with any code blocks of the original regexp.
5038 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
5039 * instead, just save the qr and return FALSE; this tells our caller that
5040 * the original pattern needs upgrading to utf8.
5044 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
5045 char *pat, STRLEN plen)
5049 GET_RE_DEBUG_FLAGS_DECL;
5051 if (pRExC_state->runtime_code_qr) {
5052 /* this is the second time we've been called; this should
5053 * only happen if the main pattern got upgraded to utf8
5054 * during compilation; re-use the qr we compiled first time
5055 * round (which should be utf8 too)
5057 qr = pRExC_state->runtime_code_qr;
5058 pRExC_state->runtime_code_qr = NULL;
5059 assert(RExC_utf8 && SvUTF8(qr));
5065 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
5069 /* determine how many extra chars we need for ' and \ escaping */
5070 for (s = 0; s < plen; s++) {
5071 if (pat[s] == '\'' || pat[s] == '\\')
5075 Newx(newpat, newlen, char);
5077 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
5079 for (s = 0; s < plen; s++) {
5080 if (n < pRExC_state->num_code_blocks
5081 && s == pRExC_state->code_blocks[n].start)
5083 /* blank out literal code block */
5084 assert(pat[s] == '(');
5085 while (s <= pRExC_state->code_blocks[n].end) {
5093 if (pat[s] == '\'' || pat[s] == '\\')
5098 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
5102 PerlIO_printf(Perl_debug_log,
5103 "%sre-parsing pattern for runtime code:%s %s\n",
5104 PL_colors[4],PL_colors[5],newpat);
5107 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
5113 PUSHSTACKi(PERLSI_REQUIRE);
5114 /* this causes the toker to collapse \\ into \ when parsing
5115 * qr''; normally only q'' does this. It also alters hints
5117 PL_reg_state.re_reparsing = TRUE;
5118 eval_sv(sv, G_SCALAR);
5124 Perl_croak(aTHX_ "%s", SvPVx_nolen_const(ERRSV));
5125 assert(SvROK(qr_ref));
5127 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
5128 /* the leaving below frees the tmp qr_ref.
5129 * Give qr a life of its own */
5137 if (!RExC_utf8 && SvUTF8(qr)) {
5138 /* first time through; the pattern got upgraded; save the
5139 * qr for the next time through */
5140 assert(!pRExC_state->runtime_code_qr);
5141 pRExC_state->runtime_code_qr = qr;
5146 /* extract any code blocks within the returned qr// */
5149 /* merge the main (r1) and run-time (r2) code blocks into one */
5151 RXi_GET_DECL(((struct regexp*)SvANY(qr)), r2);
5152 struct reg_code_block *new_block, *dst;
5153 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
5156 if (!r2->num_code_blocks) /* we guessed wrong */
5160 r1->num_code_blocks + r2->num_code_blocks,
5161 struct reg_code_block);
5164 while ( i1 < r1->num_code_blocks
5165 || i2 < r2->num_code_blocks)
5167 struct reg_code_block *src;
5170 if (i1 == r1->num_code_blocks) {
5171 src = &r2->code_blocks[i2++];
5174 else if (i2 == r2->num_code_blocks)
5175 src = &r1->code_blocks[i1++];
5176 else if ( r1->code_blocks[i1].start
5177 < r2->code_blocks[i2].start)
5179 src = &r1->code_blocks[i1++];
5180 assert(src->end < r2->code_blocks[i2].start);
5183 assert( r1->code_blocks[i1].start
5184 > r2->code_blocks[i2].start);
5185 src = &r2->code_blocks[i2++];
5187 assert(src->end < r1->code_blocks[i1].start);
5190 assert(pat[src->start] == '(');
5191 assert(pat[src->end] == ')');
5192 dst->start = src->start;
5193 dst->end = src->end;
5194 dst->block = src->block;
5195 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
5199 r1->num_code_blocks += r2->num_code_blocks;
5200 Safefree(r1->code_blocks);
5201 r1->code_blocks = new_block;
5210 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)
5212 /* This is the common code for setting up the floating and fixed length
5213 * string data extracted from Perlre_op_compile() below. Returns a boolean
5214 * as to whether succeeded or not */
5218 if (! (longest_length
5219 || (eol /* Can't have SEOL and MULTI */
5220 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
5222 /* See comments for join_exact for why REG_SEEN_EXACTF_SHARP_S */
5223 || (RExC_seen & REG_SEEN_EXACTF_SHARP_S))
5228 /* copy the information about the longest from the reg_scan_data
5229 over to the program. */
5230 if (SvUTF8(sv_longest)) {
5231 *rx_utf8 = sv_longest;
5234 *rx_substr = sv_longest;
5237 /* end_shift is how many chars that must be matched that
5238 follow this item. We calculate it ahead of time as once the
5239 lookbehind offset is added in we lose the ability to correctly
5241 ml = minlen ? *(minlen) : (I32)longest_length;
5242 *rx_end_shift = ml - offset
5243 - longest_length + (SvTAIL(sv_longest) != 0)
5246 t = (eol/* Can't have SEOL and MULTI */
5247 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
5248 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
5254 * Perl_re_op_compile - the perl internal RE engine's function to compile a
5255 * regular expression into internal code.
5256 * The pattern may be passed either as:
5257 * a list of SVs (patternp plus pat_count)
5258 * a list of OPs (expr)
5259 * If both are passed, the SV list is used, but the OP list indicates
5260 * which SVs are actually pre-compiled code blocks
5262 * The SVs in the list have magic and qr overloading applied to them (and
5263 * the list may be modified in-place with replacement SVs in the latter
5266 * If the pattern hasn't changed from old_re, then old_re will be
5269 * eng is the current engine. If that engine has an op_comp method, then
5270 * handle directly (i.e. we assume that op_comp was us); otherwise, just
5271 * do the initial concatenation of arguments and pass on to the external
5274 * If is_bare_re is not null, set it to a boolean indicating whether the
5275 * arg list reduced (after overloading) to a single bare regex which has
5276 * been returned (i.e. /$qr/).
5278 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
5280 * pm_flags contains the PMf_* flags, typically based on those from the
5281 * pm_flags field of the related PMOP. Currently we're only interested in
5282 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
5284 * We can't allocate space until we know how big the compiled form will be,
5285 * but we can't compile it (and thus know how big it is) until we've got a
5286 * place to put the code. So we cheat: we compile it twice, once with code
5287 * generation turned off and size counting turned on, and once "for real".
5288 * This also means that we don't allocate space until we are sure that the
5289 * thing really will compile successfully, and we never have to move the
5290 * code and thus invalidate pointers into it. (Note that it has to be in
5291 * one piece because free() must be able to free it all.) [NB: not true in perl]
5293 * Beware that the optimization-preparation code in here knows about some
5294 * of the structure of the compiled regexp. [I'll say.]
5298 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
5299 OP *expr, const regexp_engine* eng, REGEXP *VOL old_re,
5300 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
5305 regexp_internal *ri;
5315 /* these are all flags - maybe they should be turned
5316 * into a single int with different bit masks */
5317 I32 sawlookahead = 0;
5320 bool used_setjump = FALSE;
5321 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
5322 bool code_is_utf8 = 0;
5323 bool VOL recompile = 0;
5324 bool runtime_code = 0;
5328 RExC_state_t RExC_state;
5329 RExC_state_t * const pRExC_state = &RExC_state;
5330 #ifdef TRIE_STUDY_OPT
5332 RExC_state_t copyRExC_state;
5334 GET_RE_DEBUG_FLAGS_DECL;
5336 PERL_ARGS_ASSERT_RE_OP_COMPILE;
5338 DEBUG_r(if (!PL_colorset) reginitcolors());
5340 #ifndef PERL_IN_XSUB_RE
5341 /* Initialize these here instead of as-needed, as is quick and avoids
5342 * having to test them each time otherwise */
5343 if (! PL_AboveLatin1) {
5344 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
5345 PL_ASCII = _new_invlist_C_array(ASCII_invlist);
5346 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
5348 PL_L1PosixAlnum = _new_invlist_C_array(L1PosixAlnum_invlist);
5349 PL_PosixAlnum = _new_invlist_C_array(PosixAlnum_invlist);
5351 PL_L1PosixAlpha = _new_invlist_C_array(L1PosixAlpha_invlist);
5352 PL_PosixAlpha = _new_invlist_C_array(PosixAlpha_invlist);
5354 PL_PosixBlank = _new_invlist_C_array(PosixBlank_invlist);
5355 PL_XPosixBlank = _new_invlist_C_array(XPosixBlank_invlist);
5357 PL_L1Cased = _new_invlist_C_array(L1Cased_invlist);
5359 PL_PosixCntrl = _new_invlist_C_array(PosixCntrl_invlist);
5360 PL_XPosixCntrl = _new_invlist_C_array(XPosixCntrl_invlist);
5362 PL_PosixDigit = _new_invlist_C_array(PosixDigit_invlist);
5364 PL_L1PosixGraph = _new_invlist_C_array(L1PosixGraph_invlist);
5365 PL_PosixGraph = _new_invlist_C_array(PosixGraph_invlist);
5367 PL_L1PosixLower = _new_invlist_C_array(L1PosixLower_invlist);
5368 PL_PosixLower = _new_invlist_C_array(PosixLower_invlist);
5370 PL_L1PosixPrint = _new_invlist_C_array(L1PosixPrint_invlist);
5371 PL_PosixPrint = _new_invlist_C_array(PosixPrint_invlist);
5373 PL_L1PosixPunct = _new_invlist_C_array(L1PosixPunct_invlist);
5374 PL_PosixPunct = _new_invlist_C_array(PosixPunct_invlist);
5376 PL_PerlSpace = _new_invlist_C_array(PerlSpace_invlist);
5377 PL_XPerlSpace = _new_invlist_C_array(XPerlSpace_invlist);
5379 PL_PosixSpace = _new_invlist_C_array(PosixSpace_invlist);
5380 PL_XPosixSpace = _new_invlist_C_array(XPosixSpace_invlist);
5382 PL_L1PosixUpper = _new_invlist_C_array(L1PosixUpper_invlist);
5383 PL_PosixUpper = _new_invlist_C_array(PosixUpper_invlist);
5385 PL_VertSpace = _new_invlist_C_array(VertSpace_invlist);
5387 PL_PosixWord = _new_invlist_C_array(PosixWord_invlist);
5388 PL_L1PosixWord = _new_invlist_C_array(L1PosixWord_invlist);
5390 PL_PosixXDigit = _new_invlist_C_array(PosixXDigit_invlist);
5391 PL_XPosixXDigit = _new_invlist_C_array(XPosixXDigit_invlist);
5395 pRExC_state->code_blocks = NULL;
5396 pRExC_state->num_code_blocks = 0;
5399 *is_bare_re = FALSE;
5401 if (expr && (expr->op_type == OP_LIST ||
5402 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
5404 /* is the source UTF8, and how many code blocks are there? */
5408 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5409 if (o->op_type == OP_CONST && SvUTF8(cSVOPo_sv))
5411 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
5412 /* count of DO blocks */
5416 pRExC_state->num_code_blocks = ncode;
5417 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
5422 /* handle a list of SVs */
5426 /* apply magic and RE overloading to each arg */
5427 for (svp = patternp; svp < patternp + pat_count; svp++) {
5430 if (SvROK(rx) && SvAMAGIC(rx)) {
5431 SV *sv = AMG_CALLunary(rx, regexp_amg);
5435 if (SvTYPE(sv) != SVt_REGEXP)
5436 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5442 if (pat_count > 1) {
5443 /* concat multiple args and find any code block indexes */
5448 STRLEN orig_patlen = 0;
5450 if (pRExC_state->num_code_blocks) {
5451 o = cLISTOPx(expr)->op_first;
5452 assert(o->op_type == OP_PUSHMARK);
5456 pat = newSVpvn("", 0);
5459 /* determine if the pattern is going to be utf8 (needed
5460 * in advance to align code block indices correctly).
5461 * XXX This could fail to be detected for an arg with
5462 * overloading but not concat overloading; but the main effect
5463 * in this obscure case is to need a 'use re eval' for a
5464 * literal code block */
5465 for (svp = patternp; svp < patternp + pat_count; svp++) {
5472 for (svp = patternp; svp < patternp + pat_count; svp++) {
5473 SV *sv, *msv = *svp;
5477 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5478 assert(n < pRExC_state->num_code_blocks);
5479 pRExC_state->code_blocks[n].start = SvCUR(pat);
5480 pRExC_state->code_blocks[n].block = o;
5481 pRExC_state->code_blocks[n].src_regex = NULL;
5484 o = o->op_sibling; /* skip CONST */
5490 if ((SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5491 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5494 /* overloading involved: all bets are off over literal
5495 * code. Pretend we haven't seen it */
5496 pRExC_state->num_code_blocks -= n;
5502 while (SvAMAGIC(msv)
5503 && (sv = AMG_CALLunary(msv, string_amg))
5509 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5511 orig_patlen = SvCUR(pat);
5512 sv_catsv_nomg(pat, msv);
5515 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5518 /* extract any code blocks within any embedded qr//'s */
5519 if (rx && SvTYPE(rx) == SVt_REGEXP
5520 && RX_ENGINE((REGEXP*)rx)->op_comp)
5523 RXi_GET_DECL(((struct regexp*)SvANY(rx)), ri);
5524 if (ri->num_code_blocks) {
5526 /* the presence of an embedded qr// with code means
5527 * we should always recompile: the text of the
5528 * qr// may not have changed, but it may be a
5529 * different closure than last time */
5531 Renew(pRExC_state->code_blocks,
5532 pRExC_state->num_code_blocks + ri->num_code_blocks,
5533 struct reg_code_block);
5534 pRExC_state->num_code_blocks += ri->num_code_blocks;
5535 for (i=0; i < ri->num_code_blocks; i++) {
5536 struct reg_code_block *src, *dst;
5537 STRLEN offset = orig_patlen
5538 + ((struct regexp *)SvANY(rx))->pre_prefix;
5539 assert(n < pRExC_state->num_code_blocks);
5540 src = &ri->code_blocks[i];
5541 dst = &pRExC_state->code_blocks[n];
5542 dst->start = src->start + offset;
5543 dst->end = src->end + offset;
5544 dst->block = src->block;
5545 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5559 while (SvAMAGIC(pat)
5560 && (sv = AMG_CALLunary(pat, string_amg))
5568 /* handle bare regex: foo =~ $re */
5573 if (SvTYPE(re) == SVt_REGEXP) {
5577 Safefree(pRExC_state->code_blocks);
5583 /* not a list of SVs, so must be a list of OPs */
5585 if (expr->op_type == OP_LIST) {
5590 pat = newSVpvn("", 0);
5595 /* given a list of CONSTs and DO blocks in expr, append all
5596 * the CONSTs to pat, and record the start and end of each
5597 * code block in code_blocks[] (each DO{} op is followed by an
5598 * OP_CONST containing the corresponding literal '(?{...})
5601 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5602 if (o->op_type == OP_CONST) {
5603 sv_catsv(pat, cSVOPo_sv);
5605 pRExC_state->code_blocks[i].end = SvCUR(pat)-1;
5609 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5610 assert(i+1 < pRExC_state->num_code_blocks);
5611 pRExC_state->code_blocks[++i].start = SvCUR(pat);
5612 pRExC_state->code_blocks[i].block = o;
5613 pRExC_state->code_blocks[i].src_regex = NULL;
5619 assert(expr->op_type == OP_CONST);
5620 pat = cSVOPx_sv(expr);
5624 exp = SvPV_nomg(pat, plen);
5626 if (!eng->op_comp) {
5627 if ((SvUTF8(pat) && IN_BYTES)
5628 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5630 /* make a temporary copy; either to convert to bytes,
5631 * or to avoid repeating get-magic / overloaded stringify */
5632 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5633 (IN_BYTES ? 0 : SvUTF8(pat)));
5635 Safefree(pRExC_state->code_blocks);
5636 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5639 /* ignore the utf8ness if the pattern is 0 length */
5640 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5641 RExC_uni_semantics = 0;
5642 RExC_contains_locale = 0;
5643 pRExC_state->runtime_code_qr = NULL;
5645 /****************** LONG JUMP TARGET HERE***********************/
5646 /* Longjmp back to here if have to switch in midstream to utf8 */
5647 if (! RExC_orig_utf8) {
5648 JMPENV_PUSH(jump_ret);
5649 used_setjump = TRUE;
5652 if (jump_ret == 0) { /* First time through */
5656 SV *dsv= sv_newmortal();
5657 RE_PV_QUOTED_DECL(s, RExC_utf8,
5658 dsv, exp, plen, 60);
5659 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5660 PL_colors[4],PL_colors[5],s);
5663 else { /* longjumped back */
5666 STRLEN s = 0, d = 0;
5669 /* If the cause for the longjmp was other than changing to utf8, pop
5670 * our own setjmp, and longjmp to the correct handler */
5671 if (jump_ret != UTF8_LONGJMP) {
5673 JMPENV_JUMP(jump_ret);
5678 /* It's possible to write a regexp in ascii that represents Unicode
5679 codepoints outside of the byte range, such as via \x{100}. If we
5680 detect such a sequence we have to convert the entire pattern to utf8
5681 and then recompile, as our sizing calculation will have been based
5682 on 1 byte == 1 character, but we will need to use utf8 to encode
5683 at least some part of the pattern, and therefore must convert the whole
5686 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5687 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5689 /* upgrade pattern to UTF8, and if there are code blocks,
5690 * recalculate the indices.
5691 * This is essentially an unrolled Perl_bytes_to_utf8() */
5693 src = (U8*)SvPV_nomg(pat, plen);
5694 Newx(dst, plen * 2 + 1, U8);
5697 const UV uv = NATIVE_TO_ASCII(src[s]);
5698 if (UNI_IS_INVARIANT(uv))
5699 dst[d] = (U8)UTF_TO_NATIVE(uv);
5701 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
5702 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
5704 if (n < pRExC_state->num_code_blocks) {
5705 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5706 pRExC_state->code_blocks[n].start = d;
5707 assert(dst[d] == '(');
5710 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5711 pRExC_state->code_blocks[n].end = d;
5712 assert(dst[d] == ')');
5725 RExC_orig_utf8 = RExC_utf8 = 1;
5728 /* return old regex if pattern hasn't changed */
5732 && !!RX_UTF8(old_re) == !!RExC_utf8
5733 && RX_PRECOMP(old_re)
5734 && RX_PRELEN(old_re) == plen
5735 && memEQ(RX_PRECOMP(old_re), exp, plen))
5737 /* with runtime code, always recompile */
5738 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5740 if (!runtime_code) {
5744 Safefree(pRExC_state->code_blocks);
5748 else if ((pm_flags & PMf_USE_RE_EVAL)
5749 /* this second condition covers the non-regex literal case,
5750 * i.e. $foo =~ '(?{})'. */
5751 || ( !PL_reg_state.re_reparsing && IN_PERL_COMPILETIME
5752 && (PL_hints & HINT_RE_EVAL))
5754 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5757 #ifdef TRIE_STUDY_OPT
5761 rx_flags = orig_rx_flags;
5763 if (initial_charset == REGEX_LOCALE_CHARSET) {
5764 RExC_contains_locale = 1;
5766 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5768 /* Set to use unicode semantics if the pattern is in utf8 and has the
5769 * 'depends' charset specified, as it means unicode when utf8 */
5770 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5774 RExC_flags = rx_flags;
5775 RExC_pm_flags = pm_flags;
5778 if (PL_tainting && PL_tainted)
5779 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5781 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5782 /* whoops, we have a non-utf8 pattern, whilst run-time code
5783 * got compiled as utf8. Try again with a utf8 pattern */
5784 JMPENV_JUMP(UTF8_LONGJMP);
5787 assert(!pRExC_state->runtime_code_qr);
5792 RExC_in_lookbehind = 0;
5793 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5795 RExC_override_recoding = 0;
5797 /* First pass: determine size, legality. */
5805 RExC_emit = &PL_regdummy;
5806 RExC_whilem_seen = 0;
5807 RExC_open_parens = NULL;
5808 RExC_close_parens = NULL;
5810 RExC_paren_names = NULL;
5812 RExC_paren_name_list = NULL;
5814 RExC_recurse = NULL;
5815 RExC_recurse_count = 0;
5816 pRExC_state->code_index = 0;
5818 #if 0 /* REGC() is (currently) a NOP at the first pass.
5819 * Clever compilers notice this and complain. --jhi */
5820 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5823 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5825 RExC_lastparse=NULL;
5827 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5828 RExC_precomp = NULL;
5829 Safefree(pRExC_state->code_blocks);
5833 /* Here, finished first pass. Get rid of any added setjmp */
5839 PerlIO_printf(Perl_debug_log,
5840 "Required size %"IVdf" nodes\n"
5841 "Starting second pass (creation)\n",
5844 RExC_lastparse=NULL;
5847 /* The first pass could have found things that force Unicode semantics */
5848 if ((RExC_utf8 || RExC_uni_semantics)
5849 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5851 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5854 /* Small enough for pointer-storage convention?
5855 If extralen==0, this means that we will not need long jumps. */
5856 if (RExC_size >= 0x10000L && RExC_extralen)
5857 RExC_size += RExC_extralen;
5860 if (RExC_whilem_seen > 15)
5861 RExC_whilem_seen = 15;
5863 /* Allocate space and zero-initialize. Note, the two step process
5864 of zeroing when in debug mode, thus anything assigned has to
5865 happen after that */
5866 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5867 r = (struct regexp*)SvANY(rx);
5868 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5869 char, regexp_internal);
5870 if ( r == NULL || ri == NULL )
5871 FAIL("Regexp out of space");
5873 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5874 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5876 /* bulk initialize base fields with 0. */
5877 Zero(ri, sizeof(regexp_internal), char);
5880 /* non-zero initialization begins here */
5883 r->extflags = rx_flags;
5884 if (pm_flags & PMf_IS_QR) {
5885 ri->code_blocks = pRExC_state->code_blocks;
5886 ri->num_code_blocks = pRExC_state->num_code_blocks;
5889 SAVEFREEPV(pRExC_state->code_blocks);
5892 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5893 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5895 /* The caret is output if there are any defaults: if not all the STD
5896 * flags are set, or if no character set specifier is needed */
5898 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5900 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5901 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5902 >> RXf_PMf_STD_PMMOD_SHIFT);
5903 const char *fptr = STD_PAT_MODS; /*"msix"*/
5905 /* Allocate for the worst case, which is all the std flags are turned
5906 * on. If more precision is desired, we could do a population count of
5907 * the flags set. This could be done with a small lookup table, or by
5908 * shifting, masking and adding, or even, when available, assembly
5909 * language for a machine-language population count.
5910 * We never output a minus, as all those are defaults, so are
5911 * covered by the caret */
5912 const STRLEN wraplen = plen + has_p + has_runon
5913 + has_default /* If needs a caret */
5915 /* If needs a character set specifier */
5916 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5917 + (sizeof(STD_PAT_MODS) - 1)
5918 + (sizeof("(?:)") - 1);
5920 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
5923 SvFLAGS(rx) |= SVf_UTF8;
5926 /* If a default, cover it using the caret */
5928 *p++= DEFAULT_PAT_MOD;
5932 const char* const name = get_regex_charset_name(r->extflags, &len);
5933 Copy(name, p, len, char);
5937 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5940 while((ch = *fptr++)) {
5948 Copy(RExC_precomp, p, plen, char);
5949 assert ((RX_WRAPPED(rx) - p) < 16);
5950 r->pre_prefix = p - RX_WRAPPED(rx);
5956 SvCUR_set(rx, p - SvPVX_const(rx));
5960 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5962 if (RExC_seen & REG_SEEN_RECURSE) {
5963 Newxz(RExC_open_parens, RExC_npar,regnode *);
5964 SAVEFREEPV(RExC_open_parens);
5965 Newxz(RExC_close_parens,RExC_npar,regnode *);
5966 SAVEFREEPV(RExC_close_parens);
5969 /* Useful during FAIL. */
5970 #ifdef RE_TRACK_PATTERN_OFFSETS
5971 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5972 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5973 "%s %"UVuf" bytes for offset annotations.\n",
5974 ri->u.offsets ? "Got" : "Couldn't get",
5975 (UV)((2*RExC_size+1) * sizeof(U32))));
5977 SetProgLen(ri,RExC_size);
5982 /* Second pass: emit code. */
5983 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5984 RExC_pm_flags = pm_flags;
5989 RExC_emit_start = ri->program;
5990 RExC_emit = ri->program;
5991 RExC_emit_bound = ri->program + RExC_size + 1;
5992 pRExC_state->code_index = 0;
5994 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5995 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5999 /* XXXX To minimize changes to RE engine we always allocate
6000 3-units-long substrs field. */
6001 Newx(r->substrs, 1, struct reg_substr_data);
6002 if (RExC_recurse_count) {
6003 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6004 SAVEFREEPV(RExC_recurse);
6008 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
6009 Zero(r->substrs, 1, struct reg_substr_data);
6011 #ifdef TRIE_STUDY_OPT
6013 StructCopy(&zero_scan_data, &data, scan_data_t);
6014 copyRExC_state = RExC_state;
6017 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6019 RExC_state = copyRExC_state;
6020 if (seen & REG_TOP_LEVEL_BRANCHES)
6021 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6023 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6024 if (data.last_found) {
6025 SvREFCNT_dec(data.longest_fixed);
6026 SvREFCNT_dec(data.longest_float);
6027 SvREFCNT_dec(data.last_found);
6029 StructCopy(&zero_scan_data, &data, scan_data_t);
6032 StructCopy(&zero_scan_data, &data, scan_data_t);
6035 /* Dig out information for optimizations. */
6036 r->extflags = RExC_flags; /* was pm_op */
6037 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6040 SvUTF8_on(rx); /* Unicode in it? */
6041 ri->regstclass = NULL;
6042 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6043 r->intflags |= PREGf_NAUGHTY;
6044 scan = ri->program + 1; /* First BRANCH. */
6046 /* testing for BRANCH here tells us whether there is "must appear"
6047 data in the pattern. If there is then we can use it for optimisations */
6048 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6050 STRLEN longest_float_length, longest_fixed_length;
6051 struct regnode_charclass_class ch_class; /* pointed to by data */
6053 I32 last_close = 0; /* pointed to by data */
6054 regnode *first= scan;
6055 regnode *first_next= regnext(first);
6057 * Skip introductions and multiplicators >= 1
6058 * so that we can extract the 'meat' of the pattern that must
6059 * match in the large if() sequence following.
6060 * NOTE that EXACT is NOT covered here, as it is normally
6061 * picked up by the optimiser separately.
6063 * This is unfortunate as the optimiser isnt handling lookahead
6064 * properly currently.
6067 while ((OP(first) == OPEN && (sawopen = 1)) ||
6068 /* An OR of *one* alternative - should not happen now. */
6069 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6070 /* for now we can't handle lookbehind IFMATCH*/
6071 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6072 (OP(first) == PLUS) ||
6073 (OP(first) == MINMOD) ||
6074 /* An {n,m} with n>0 */
6075 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6076 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6079 * the only op that could be a regnode is PLUS, all the rest
6080 * will be regnode_1 or regnode_2.
6083 if (OP(first) == PLUS)
6086 first += regarglen[OP(first)];
6088 first = NEXTOPER(first);
6089 first_next= regnext(first);
6092 /* Starting-point info. */
6094 DEBUG_PEEP("first:",first,0);
6095 /* Ignore EXACT as we deal with it later. */
6096 if (PL_regkind[OP(first)] == EXACT) {
6097 if (OP(first) == EXACT)
6098 NOOP; /* Empty, get anchored substr later. */
6100 ri->regstclass = first;
6103 else if (PL_regkind[OP(first)] == TRIE &&
6104 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6107 /* this can happen only on restudy */
6108 if ( OP(first) == TRIE ) {
6109 struct regnode_1 *trieop = (struct regnode_1 *)
6110 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6111 StructCopy(first,trieop,struct regnode_1);
6112 trie_op=(regnode *)trieop;
6114 struct regnode_charclass *trieop = (struct regnode_charclass *)
6115 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6116 StructCopy(first,trieop,struct regnode_charclass);
6117 trie_op=(regnode *)trieop;
6120 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6121 ri->regstclass = trie_op;
6124 else if (REGNODE_SIMPLE(OP(first)))
6125 ri->regstclass = first;
6126 else if (PL_regkind[OP(first)] == BOUND ||
6127 PL_regkind[OP(first)] == NBOUND)
6128 ri->regstclass = first;
6129 else if (PL_regkind[OP(first)] == BOL) {
6130 r->extflags |= (OP(first) == MBOL
6132 : (OP(first) == SBOL
6135 first = NEXTOPER(first);
6138 else if (OP(first) == GPOS) {
6139 r->extflags |= RXf_ANCH_GPOS;
6140 first = NEXTOPER(first);
6143 else if ((!sawopen || !RExC_sawback) &&
6144 (OP(first) == STAR &&
6145 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6146 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6148 /* turn .* into ^.* with an implied $*=1 */
6150 (OP(NEXTOPER(first)) == REG_ANY)
6153 r->extflags |= type;
6154 r->intflags |= PREGf_IMPLICIT;
6155 first = NEXTOPER(first);
6158 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6159 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6160 /* x+ must match at the 1st pos of run of x's */
6161 r->intflags |= PREGf_SKIP;
6163 /* Scan is after the zeroth branch, first is atomic matcher. */
6164 #ifdef TRIE_STUDY_OPT
6167 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6168 (IV)(first - scan + 1))
6172 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6173 (IV)(first - scan + 1))
6179 * If there's something expensive in the r.e., find the
6180 * longest literal string that must appear and make it the
6181 * regmust. Resolve ties in favor of later strings, since
6182 * the regstart check works with the beginning of the r.e.
6183 * and avoiding duplication strengthens checking. Not a
6184 * strong reason, but sufficient in the absence of others.
6185 * [Now we resolve ties in favor of the earlier string if
6186 * it happens that c_offset_min has been invalidated, since the
6187 * earlier string may buy us something the later one won't.]
6190 data.longest_fixed = newSVpvs("");
6191 data.longest_float = newSVpvs("");
6192 data.last_found = newSVpvs("");
6193 data.longest = &(data.longest_fixed);
6195 if (!ri->regstclass) {
6196 cl_init(pRExC_state, &ch_class);
6197 data.start_class = &ch_class;
6198 stclass_flag = SCF_DO_STCLASS_AND;
6199 } else /* XXXX Check for BOUND? */
6201 data.last_closep = &last_close;
6203 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6204 &data, -1, NULL, NULL,
6205 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6211 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6212 && data.last_start_min == 0 && data.last_end > 0
6213 && !RExC_seen_zerolen
6214 && !(RExC_seen & REG_SEEN_VERBARG)
6215 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6216 r->extflags |= RXf_CHECK_ALL;
6217 scan_commit(pRExC_state, &data,&minlen,0);
6218 SvREFCNT_dec(data.last_found);
6220 longest_float_length = CHR_SVLEN(data.longest_float);
6222 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6223 && data.offset_fixed == data.offset_float_min
6224 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6225 && S_setup_longest (aTHX_ pRExC_state,
6229 &(r->float_end_shift),
6230 data.lookbehind_float,
6231 data.offset_float_min,
6233 longest_float_length,
6234 data.flags & SF_FL_BEFORE_EOL,
6235 data.flags & SF_FL_BEFORE_MEOL))
6237 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6238 r->float_max_offset = data.offset_float_max;
6239 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6240 r->float_max_offset -= data.lookbehind_float;
6243 r->float_substr = r->float_utf8 = NULL;
6244 SvREFCNT_dec(data.longest_float);
6245 longest_float_length = 0;
6248 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6250 if (S_setup_longest (aTHX_ pRExC_state,
6252 &(r->anchored_utf8),
6253 &(r->anchored_substr),
6254 &(r->anchored_end_shift),
6255 data.lookbehind_fixed,
6258 longest_fixed_length,
6259 data.flags & SF_FIX_BEFORE_EOL,
6260 data.flags & SF_FIX_BEFORE_MEOL))
6262 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6265 r->anchored_substr = r->anchored_utf8 = NULL;
6266 SvREFCNT_dec(data.longest_fixed);
6267 longest_fixed_length = 0;
6271 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6272 ri->regstclass = NULL;
6274 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6276 && !(data.start_class->flags & ANYOF_EOS)
6277 && !cl_is_anything(data.start_class))
6279 const U32 n = add_data(pRExC_state, 1, "f");
6280 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
6282 Newx(RExC_rxi->data->data[n], 1,
6283 struct regnode_charclass_class);
6284 StructCopy(data.start_class,
6285 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6286 struct regnode_charclass_class);
6287 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6288 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6289 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6290 regprop(r, sv, (regnode*)data.start_class);
6291 PerlIO_printf(Perl_debug_log,
6292 "synthetic stclass \"%s\".\n",
6293 SvPVX_const(sv));});
6296 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6297 if (longest_fixed_length > longest_float_length) {
6298 r->check_end_shift = r->anchored_end_shift;
6299 r->check_substr = r->anchored_substr;
6300 r->check_utf8 = r->anchored_utf8;
6301 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6302 if (r->extflags & RXf_ANCH_SINGLE)
6303 r->extflags |= RXf_NOSCAN;
6306 r->check_end_shift = r->float_end_shift;
6307 r->check_substr = r->float_substr;
6308 r->check_utf8 = r->float_utf8;
6309 r->check_offset_min = r->float_min_offset;
6310 r->check_offset_max = r->float_max_offset;
6312 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6313 This should be changed ASAP! */
6314 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6315 r->extflags |= RXf_USE_INTUIT;
6316 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6317 r->extflags |= RXf_INTUIT_TAIL;
6319 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6320 if ( (STRLEN)minlen < longest_float_length )
6321 minlen= longest_float_length;
6322 if ( (STRLEN)minlen < longest_fixed_length )
6323 minlen= longest_fixed_length;
6327 /* Several toplevels. Best we can is to set minlen. */
6329 struct regnode_charclass_class ch_class;
6332 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6334 scan = ri->program + 1;
6335 cl_init(pRExC_state, &ch_class);
6336 data.start_class = &ch_class;
6337 data.last_closep = &last_close;
6340 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6341 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6345 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6346 = r->float_substr = r->float_utf8 = NULL;
6348 if (!(data.start_class->flags & ANYOF_EOS)
6349 && !cl_is_anything(data.start_class))
6351 const U32 n = add_data(pRExC_state, 1, "f");
6352 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
6354 Newx(RExC_rxi->data->data[n], 1,
6355 struct regnode_charclass_class);
6356 StructCopy(data.start_class,
6357 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6358 struct regnode_charclass_class);
6359 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6360 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6361 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6362 regprop(r, sv, (regnode*)data.start_class);
6363 PerlIO_printf(Perl_debug_log,
6364 "synthetic stclass \"%s\".\n",
6365 SvPVX_const(sv));});
6369 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6370 the "real" pattern. */
6372 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6373 (IV)minlen, (IV)r->minlen);
6375 r->minlenret = minlen;
6376 if (r->minlen < minlen)
6379 if (RExC_seen & REG_SEEN_GPOS)
6380 r->extflags |= RXf_GPOS_SEEN;
6381 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6382 r->extflags |= RXf_LOOKBEHIND_SEEN;
6383 if (pRExC_state->num_code_blocks)
6384 r->extflags |= RXf_EVAL_SEEN;
6385 if (RExC_seen & REG_SEEN_CANY)
6386 r->extflags |= RXf_CANY_SEEN;
6387 if (RExC_seen & REG_SEEN_VERBARG)
6388 r->intflags |= PREGf_VERBARG_SEEN;
6389 if (RExC_seen & REG_SEEN_CUTGROUP)
6390 r->intflags |= PREGf_CUTGROUP_SEEN;
6391 if (pm_flags & PMf_USE_RE_EVAL)
6392 r->intflags |= PREGf_USE_RE_EVAL;
6393 if (RExC_paren_names)
6394 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6396 RXp_PAREN_NAMES(r) = NULL;
6398 #ifdef STUPID_PATTERN_CHECKS
6399 if (RX_PRELEN(rx) == 0)
6400 r->extflags |= RXf_NULL;
6401 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
6402 /* XXX: this should happen BEFORE we compile */
6403 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6404 else if (RX_PRELEN(rx) == 3 && memEQ("\\s+", RX_PRECOMP(rx), 3))
6405 r->extflags |= RXf_WHITE;
6406 else if (RX_PRELEN(rx) == 1 && RXp_PRECOMP(rx)[0] == '^')
6407 r->extflags |= RXf_START_ONLY;
6409 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
6410 /* XXX: this should happen BEFORE we compile */
6411 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6413 regnode *first = ri->program + 1;
6416 if (PL_regkind[fop] == NOTHING && OP(NEXTOPER(first)) == END)
6417 r->extflags |= RXf_NULL;
6418 else if (PL_regkind[fop] == BOL && OP(NEXTOPER(first)) == END)
6419 r->extflags |= RXf_START_ONLY;
6420 else if (fop == PLUS && OP(NEXTOPER(first)) == SPACE
6421 && OP(regnext(first)) == END)
6422 r->extflags |= RXf_WHITE;
6426 if (RExC_paren_names) {
6427 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6428 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6431 ri->name_list_idx = 0;
6433 if (RExC_recurse_count) {
6434 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6435 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6436 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6439 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6440 /* assume we don't need to swap parens around before we match */
6443 PerlIO_printf(Perl_debug_log,"Final program:\n");
6446 #ifdef RE_TRACK_PATTERN_OFFSETS
6447 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6448 const U32 len = ri->u.offsets[0];
6450 GET_RE_DEBUG_FLAGS_DECL;
6451 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6452 for (i = 1; i <= len; i++) {
6453 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6454 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6455 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6457 PerlIO_printf(Perl_debug_log, "\n");
6465 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6468 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6470 PERL_UNUSED_ARG(value);
6472 if (flags & RXapif_FETCH) {
6473 return reg_named_buff_fetch(rx, key, flags);
6474 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6475 Perl_croak_no_modify(aTHX);
6477 } else if (flags & RXapif_EXISTS) {
6478 return reg_named_buff_exists(rx, key, flags)
6481 } else if (flags & RXapif_REGNAMES) {
6482 return reg_named_buff_all(rx, flags);
6483 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6484 return reg_named_buff_scalar(rx, flags);
6486 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6492 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6495 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6496 PERL_UNUSED_ARG(lastkey);
6498 if (flags & RXapif_FIRSTKEY)
6499 return reg_named_buff_firstkey(rx, flags);
6500 else if (flags & RXapif_NEXTKEY)
6501 return reg_named_buff_nextkey(rx, flags);
6503 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6509 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6512 AV *retarray = NULL;
6514 struct regexp *const rx = (struct regexp *)SvANY(r);
6516 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6518 if (flags & RXapif_ALL)
6521 if (rx && RXp_PAREN_NAMES(rx)) {
6522 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6525 SV* sv_dat=HeVAL(he_str);
6526 I32 *nums=(I32*)SvPVX(sv_dat);
6527 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6528 if ((I32)(rx->nparens) >= nums[i]
6529 && rx->offs[nums[i]].start != -1
6530 && rx->offs[nums[i]].end != -1)
6533 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6538 ret = newSVsv(&PL_sv_undef);
6541 av_push(retarray, ret);
6544 return newRV_noinc(MUTABLE_SV(retarray));
6551 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6554 struct regexp *const rx = (struct regexp *)SvANY(r);
6556 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6558 if (rx && RXp_PAREN_NAMES(rx)) {
6559 if (flags & RXapif_ALL) {
6560 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6562 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6576 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6578 struct regexp *const rx = (struct regexp *)SvANY(r);
6580 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6582 if ( rx && RXp_PAREN_NAMES(rx) ) {
6583 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6585 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6592 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6594 struct regexp *const rx = (struct regexp *)SvANY(r);
6595 GET_RE_DEBUG_FLAGS_DECL;
6597 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6599 if (rx && RXp_PAREN_NAMES(rx)) {
6600 HV *hv = RXp_PAREN_NAMES(rx);
6602 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6605 SV* sv_dat = HeVAL(temphe);
6606 I32 *nums = (I32*)SvPVX(sv_dat);
6607 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6608 if ((I32)(rx->lastparen) >= nums[i] &&
6609 rx->offs[nums[i]].start != -1 &&
6610 rx->offs[nums[i]].end != -1)
6616 if (parno || flags & RXapif_ALL) {
6617 return newSVhek(HeKEY_hek(temphe));
6625 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6630 struct regexp *const rx = (struct regexp *)SvANY(r);
6632 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6634 if (rx && RXp_PAREN_NAMES(rx)) {
6635 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6636 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6637 } else if (flags & RXapif_ONE) {
6638 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6639 av = MUTABLE_AV(SvRV(ret));
6640 length = av_len(av);
6642 return newSViv(length + 1);
6644 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6648 return &PL_sv_undef;
6652 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6654 struct regexp *const rx = (struct regexp *)SvANY(r);
6657 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6659 if (rx && RXp_PAREN_NAMES(rx)) {
6660 HV *hv= RXp_PAREN_NAMES(rx);
6662 (void)hv_iterinit(hv);
6663 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6666 SV* sv_dat = HeVAL(temphe);
6667 I32 *nums = (I32*)SvPVX(sv_dat);
6668 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6669 if ((I32)(rx->lastparen) >= nums[i] &&
6670 rx->offs[nums[i]].start != -1 &&
6671 rx->offs[nums[i]].end != -1)
6677 if (parno || flags & RXapif_ALL) {
6678 av_push(av, newSVhek(HeKEY_hek(temphe)));
6683 return newRV_noinc(MUTABLE_SV(av));
6687 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6690 struct regexp *const rx = (struct regexp *)SvANY(r);
6695 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6698 sv_setsv(sv,&PL_sv_undef);
6702 if (paren == RX_BUFF_IDX_PREMATCH && rx->offs[0].start != -1) {
6704 i = rx->offs[0].start;
6708 if (paren == RX_BUFF_IDX_POSTMATCH && rx->offs[0].end != -1) {
6710 s = rx->subbeg + rx->offs[0].end;
6711 i = rx->sublen - rx->offs[0].end;
6714 if ( 0 <= paren && paren <= (I32)rx->nparens &&
6715 (s1 = rx->offs[paren].start) != -1 &&
6716 (t1 = rx->offs[paren].end) != -1)
6720 s = rx->subbeg + s1;
6722 sv_setsv(sv,&PL_sv_undef);
6725 assert(rx->sublen >= (s - rx->subbeg) + i );
6727 const int oldtainted = PL_tainted;
6729 sv_setpvn(sv, s, i);
6730 PL_tainted = oldtainted;
6731 if ( (rx->extflags & RXf_CANY_SEEN)
6732 ? (RXp_MATCH_UTF8(rx)
6733 && (!i || is_utf8_string((U8*)s, i)))
6734 : (RXp_MATCH_UTF8(rx)) )
6741 if (RXp_MATCH_TAINTED(rx)) {
6742 if (SvTYPE(sv) >= SVt_PVMG) {
6743 MAGIC* const mg = SvMAGIC(sv);
6746 SvMAGIC_set(sv, mg->mg_moremagic);
6748 if ((mgt = SvMAGIC(sv))) {
6749 mg->mg_moremagic = mgt;
6750 SvMAGIC_set(sv, mg);
6760 sv_setsv(sv,&PL_sv_undef);
6766 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6767 SV const * const value)
6769 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6771 PERL_UNUSED_ARG(rx);
6772 PERL_UNUSED_ARG(paren);
6773 PERL_UNUSED_ARG(value);
6776 Perl_croak_no_modify(aTHX);
6780 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6783 struct regexp *const rx = (struct regexp *)SvANY(r);
6787 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6789 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6791 /* $` / ${^PREMATCH} */
6792 case RX_BUFF_IDX_PREMATCH:
6793 if (rx->offs[0].start != -1) {
6794 i = rx->offs[0].start;
6802 /* $' / ${^POSTMATCH} */
6803 case RX_BUFF_IDX_POSTMATCH:
6804 if (rx->offs[0].end != -1) {
6805 i = rx->sublen - rx->offs[0].end;
6807 s1 = rx->offs[0].end;
6813 /* $& / ${^MATCH}, $1, $2, ... */
6815 if (paren <= (I32)rx->nparens &&
6816 (s1 = rx->offs[paren].start) != -1 &&
6817 (t1 = rx->offs[paren].end) != -1)
6822 if (ckWARN(WARN_UNINITIALIZED))
6823 report_uninit((const SV *)sv);
6828 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6829 const char * const s = rx->subbeg + s1;
6834 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6841 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6843 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6844 PERL_UNUSED_ARG(rx);
6848 return newSVpvs("Regexp");
6851 /* Scans the name of a named buffer from the pattern.
6852 * If flags is REG_RSN_RETURN_NULL returns null.
6853 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6854 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6855 * to the parsed name as looked up in the RExC_paren_names hash.
6856 * If there is an error throws a vFAIL().. type exception.
6859 #define REG_RSN_RETURN_NULL 0
6860 #define REG_RSN_RETURN_NAME 1
6861 #define REG_RSN_RETURN_DATA 2
6864 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6866 char *name_start = RExC_parse;
6868 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6870 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6871 /* skip IDFIRST by using do...while */
6874 RExC_parse += UTF8SKIP(RExC_parse);
6875 } while (isALNUM_utf8((U8*)RExC_parse));
6879 } while (isALNUM(*RExC_parse));
6881 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6882 vFAIL("Group name must start with a non-digit word character");
6886 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6887 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6888 if ( flags == REG_RSN_RETURN_NAME)
6890 else if (flags==REG_RSN_RETURN_DATA) {
6893 if ( ! sv_name ) /* should not happen*/
6894 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6895 if (RExC_paren_names)
6896 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6898 sv_dat = HeVAL(he_str);
6900 vFAIL("Reference to nonexistent named group");
6904 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6905 (unsigned long) flags);
6907 assert(0); /* NOT REACHED */
6912 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6913 int rem=(int)(RExC_end - RExC_parse); \
6922 if (RExC_lastparse!=RExC_parse) \
6923 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6926 iscut ? "..." : "<" \
6929 PerlIO_printf(Perl_debug_log,"%16s",""); \
6932 num = RExC_size + 1; \
6934 num=REG_NODE_NUM(RExC_emit); \
6935 if (RExC_lastnum!=num) \
6936 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6938 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6939 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6940 (int)((depth*2)), "", \
6944 RExC_lastparse=RExC_parse; \
6949 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6950 DEBUG_PARSE_MSG((funcname)); \
6951 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6953 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6954 DEBUG_PARSE_MSG((funcname)); \
6955 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6958 /* This section of code defines the inversion list object and its methods. The
6959 * interfaces are highly subject to change, so as much as possible is static to
6960 * this file. An inversion list is here implemented as a malloc'd C UV array
6961 * with some added info that is placed as UVs at the beginning in a header
6962 * portion. An inversion list for Unicode is an array of code points, sorted
6963 * by ordinal number. The zeroth element is the first code point in the list.
6964 * The 1th element is the first element beyond that not in the list. In other
6965 * words, the first range is
6966 * invlist[0]..(invlist[1]-1)
6967 * The other ranges follow. Thus every element whose index is divisible by two
6968 * marks the beginning of a range that is in the list, and every element not
6969 * divisible by two marks the beginning of a range not in the list. A single
6970 * element inversion list that contains the single code point N generally
6971 * consists of two elements
6974 * (The exception is when N is the highest representable value on the
6975 * machine, in which case the list containing just it would be a single
6976 * element, itself. By extension, if the last range in the list extends to
6977 * infinity, then the first element of that range will be in the inversion list
6978 * at a position that is divisible by two, and is the final element in the
6980 * Taking the complement (inverting) an inversion list is quite simple, if the
6981 * first element is 0, remove it; otherwise add a 0 element at the beginning.
6982 * This implementation reserves an element at the beginning of each inversion list
6983 * to contain 0 when the list contains 0, and contains 1 otherwise. The actual
6984 * beginning of the list is either that element if 0, or the next one if 1.
6986 * More about inversion lists can be found in "Unicode Demystified"
6987 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
6988 * More will be coming when functionality is added later.
6990 * The inversion list data structure is currently implemented as an SV pointing
6991 * to an array of UVs that the SV thinks are bytes. This allows us to have an
6992 * array of UV whose memory management is automatically handled by the existing
6993 * facilities for SV's.
6995 * Some of the methods should always be private to the implementation, and some
6996 * should eventually be made public */
6998 /* The header definitions are in F<inline_invlist.c> */
7000 #define TO_INTERNAL_SIZE(x) ((x + HEADER_LENGTH) * sizeof(UV))
7001 #define FROM_INTERNAL_SIZE(x) ((x / sizeof(UV)) - HEADER_LENGTH)
7003 #define INVLIST_INITIAL_LEN 10
7005 PERL_STATIC_INLINE UV*
7006 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7008 /* Returns a pointer to the first element in the inversion list's array.
7009 * This is called upon initialization of an inversion list. Where the
7010 * array begins depends on whether the list has the code point U+0000
7011 * in it or not. The other parameter tells it whether the code that
7012 * follows this call is about to put a 0 in the inversion list or not.
7013 * The first element is either the element with 0, if 0, or the next one,
7016 UV* zero = get_invlist_zero_addr(invlist);
7018 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7021 assert(! *_get_invlist_len_addr(invlist));
7023 /* 1^1 = 0; 1^0 = 1 */
7024 *zero = 1 ^ will_have_0;
7025 return zero + *zero;
7028 PERL_STATIC_INLINE UV*
7029 S_invlist_array(pTHX_ SV* const invlist)
7031 /* Returns the pointer to the inversion list's array. Every time the
7032 * length changes, this needs to be called in case malloc or realloc moved
7035 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7037 /* Must not be empty. If these fail, you probably didn't check for <len>
7038 * being non-zero before trying to get the array */
7039 assert(*_get_invlist_len_addr(invlist));
7040 assert(*get_invlist_zero_addr(invlist) == 0
7041 || *get_invlist_zero_addr(invlist) == 1);
7043 /* The array begins either at the element reserved for zero if the
7044 * list contains 0 (that element will be set to 0), or otherwise the next
7045 * element (in which case the reserved element will be set to 1). */
7046 return (UV *) (get_invlist_zero_addr(invlist)
7047 + *get_invlist_zero_addr(invlist));
7050 PERL_STATIC_INLINE void
7051 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7053 /* Sets the current number of elements stored in the inversion list */
7055 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7057 *_get_invlist_len_addr(invlist) = len;
7059 assert(len <= SvLEN(invlist));
7061 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7062 /* If the list contains U+0000, that element is part of the header,
7063 * and should not be counted as part of the array. It will contain
7064 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7066 * SvCUR_set(invlist,
7067 * TO_INTERNAL_SIZE(len
7068 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7069 * But, this is only valid if len is not 0. The consequences of not doing
7070 * this is that the memory allocation code may think that 1 more UV is
7071 * being used than actually is, and so might do an unnecessary grow. That
7072 * seems worth not bothering to make this the precise amount.
7074 * Note that when inverting, SvCUR shouldn't change */
7077 PERL_STATIC_INLINE IV*
7078 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7080 /* Return the address of the UV that is reserved to hold the cached index
7083 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7085 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7088 PERL_STATIC_INLINE IV
7089 S_invlist_previous_index(pTHX_ SV* const invlist)
7091 /* Returns cached index of previous search */
7093 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7095 return *get_invlist_previous_index_addr(invlist);
7098 PERL_STATIC_INLINE void
7099 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7101 /* Caches <index> for later retrieval */
7103 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7105 assert(index == 0 || index < (int) _invlist_len(invlist));
7107 *get_invlist_previous_index_addr(invlist) = index;
7110 PERL_STATIC_INLINE UV
7111 S_invlist_max(pTHX_ SV* const invlist)
7113 /* Returns the maximum number of elements storable in the inversion list's
7114 * array, without having to realloc() */
7116 PERL_ARGS_ASSERT_INVLIST_MAX;
7118 return FROM_INTERNAL_SIZE(SvLEN(invlist));
7121 PERL_STATIC_INLINE UV*
7122 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7124 /* Return the address of the UV that is reserved to hold 0 if the inversion
7125 * list contains 0. This has to be the last element of the heading, as the
7126 * list proper starts with either it if 0, or the next element if not.
7127 * (But we force it to contain either 0 or 1) */
7129 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7131 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7134 #ifndef PERL_IN_XSUB_RE
7136 Perl__new_invlist(pTHX_ IV initial_size)
7139 /* Return a pointer to a newly constructed inversion list, with enough
7140 * space to store 'initial_size' elements. If that number is negative, a
7141 * system default is used instead */
7145 if (initial_size < 0) {
7146 initial_size = INVLIST_INITIAL_LEN;
7149 /* Allocate the initial space */
7150 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7151 invlist_set_len(new_list, 0);
7153 /* Force iterinit() to be used to get iteration to work */
7154 *get_invlist_iter_addr(new_list) = UV_MAX;
7156 /* This should force a segfault if a method doesn't initialize this
7158 *get_invlist_zero_addr(new_list) = UV_MAX;
7160 *get_invlist_previous_index_addr(new_list) = 0;
7161 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7162 #if HEADER_LENGTH != 5
7163 # error Need to regenerate VERSION_ID by running perl -E 'say int(rand 2**31-1)', and then changing the #if to the new length
7171 S__new_invlist_C_array(pTHX_ UV* list)
7173 /* Return a pointer to a newly constructed inversion list, initialized to
7174 * point to <list>, which has to be in the exact correct inversion list
7175 * form, including internal fields. Thus this is a dangerous routine that
7176 * should not be used in the wrong hands */
7178 SV* invlist = newSV_type(SVt_PV);
7180 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7182 SvPV_set(invlist, (char *) list);
7183 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7184 shouldn't touch it */
7185 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7187 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7188 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7195 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7197 /* Grow the maximum size of an inversion list */
7199 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7201 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7204 PERL_STATIC_INLINE void
7205 S_invlist_trim(pTHX_ SV* const invlist)
7207 PERL_ARGS_ASSERT_INVLIST_TRIM;
7209 /* Change the length of the inversion list to how many entries it currently
7212 SvPV_shrink_to_cur((SV *) invlist);
7215 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7218 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7220 /* Subject to change or removal. Append the range from 'start' to 'end' at
7221 * the end of the inversion list. The range must be above any existing
7225 UV max = invlist_max(invlist);
7226 UV len = _invlist_len(invlist);
7228 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7230 if (len == 0) { /* Empty lists must be initialized */
7231 array = _invlist_array_init(invlist, start == 0);
7234 /* Here, the existing list is non-empty. The current max entry in the
7235 * list is generally the first value not in the set, except when the
7236 * set extends to the end of permissible values, in which case it is
7237 * the first entry in that final set, and so this call is an attempt to
7238 * append out-of-order */
7240 UV final_element = len - 1;
7241 array = invlist_array(invlist);
7242 if (array[final_element] > start
7243 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7245 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",
7246 array[final_element], start,
7247 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7250 /* Here, it is a legal append. If the new range begins with the first
7251 * value not in the set, it is extending the set, so the new first
7252 * value not in the set is one greater than the newly extended range.
7254 if (array[final_element] == start) {
7255 if (end != UV_MAX) {
7256 array[final_element] = end + 1;
7259 /* But if the end is the maximum representable on the machine,
7260 * just let the range that this would extend to have no end */
7261 invlist_set_len(invlist, len - 1);
7267 /* Here the new range doesn't extend any existing set. Add it */
7269 len += 2; /* Includes an element each for the start and end of range */
7271 /* If overflows the existing space, extend, which may cause the array to be
7274 invlist_extend(invlist, len);
7275 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7276 failure in invlist_array() */
7277 array = invlist_array(invlist);
7280 invlist_set_len(invlist, len);
7283 /* The next item on the list starts the range, the one after that is
7284 * one past the new range. */
7285 array[len - 2] = start;
7286 if (end != UV_MAX) {
7287 array[len - 1] = end + 1;
7290 /* But if the end is the maximum representable on the machine, just let
7291 * the range have no end */
7292 invlist_set_len(invlist, len - 1);
7296 #ifndef PERL_IN_XSUB_RE
7299 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7301 /* Searches the inversion list for the entry that contains the input code
7302 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7303 * return value is the index into the list's array of the range that
7308 IV high = _invlist_len(invlist);
7309 const IV highest_element = high - 1;
7312 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7314 /* If list is empty, return failure. */
7319 /* If the code point is before the first element, return failure. (We
7320 * can't combine this with the test above, because we can't get the array
7321 * unless we know the list is non-empty) */
7322 array = invlist_array(invlist);
7324 mid = invlist_previous_index(invlist);
7325 assert(mid >=0 && mid <= highest_element);
7327 /* <mid> contains the cache of the result of the previous call to this
7328 * function (0 the first time). See if this call is for the same result,
7329 * or if it is for mid-1. This is under the theory that calls to this
7330 * function will often be for related code points that are near each other.
7331 * And benchmarks show that caching gives better results. We also test
7332 * here if the code point is within the bounds of the list. These tests
7333 * replace others that would have had to be made anyway to make sure that
7334 * the array bounds were not exceeded, and give us extra information at the
7336 if (cp >= array[mid]) {
7337 if (cp >= array[highest_element]) {
7338 return highest_element;
7341 /* Here, array[mid] <= cp < array[highest_element]. This means that
7342 * the final element is not the answer, so can exclude it; it also
7343 * means that <mid> is not the final element, so can refer to 'mid + 1'
7345 if (cp < array[mid + 1]) {
7351 else { /* cp < aray[mid] */
7352 if (cp < array[0]) { /* Fail if outside the array */
7356 if (cp >= array[mid - 1]) {
7361 /* Binary search. What we are looking for is <i> such that
7362 * array[i] <= cp < array[i+1]
7363 * The loop below converges on the i+1. Note that there may not be an
7364 * (i+1)th element in the array, and things work nonetheless */
7365 while (low < high) {
7366 mid = (low + high) / 2;
7367 assert(mid <= highest_element);
7368 if (array[mid] <= cp) { /* cp >= array[mid] */
7371 /* We could do this extra test to exit the loop early.
7372 if (cp < array[low]) {
7377 else { /* cp < array[mid] */
7384 invlist_set_previous_index(invlist, high);
7389 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7391 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7392 * but is used when the swash has an inversion list. This makes this much
7393 * faster, as it uses a binary search instead of a linear one. This is
7394 * intimately tied to that function, and perhaps should be in utf8.c,
7395 * except it is intimately tied to inversion lists as well. It assumes
7396 * that <swatch> is all 0's on input */
7399 const IV len = _invlist_len(invlist);
7403 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7405 if (len == 0) { /* Empty inversion list */
7409 array = invlist_array(invlist);
7411 /* Find which element it is */
7412 i = _invlist_search(invlist, start);
7414 /* We populate from <start> to <end> */
7415 while (current < end) {
7418 /* The inversion list gives the results for every possible code point
7419 * after the first one in the list. Only those ranges whose index is
7420 * even are ones that the inversion list matches. For the odd ones,
7421 * and if the initial code point is not in the list, we have to skip
7422 * forward to the next element */
7423 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7425 if (i >= len) { /* Finished if beyond the end of the array */
7429 if (current >= end) { /* Finished if beyond the end of what we
7431 if (LIKELY(end < UV_MAX)) {
7435 /* We get here when the upper bound is the maximum
7436 * representable on the machine, and we are looking for just
7437 * that code point. Have to special case it */
7439 goto join_end_of_list;
7442 assert(current >= start);
7444 /* The current range ends one below the next one, except don't go past
7447 upper = (i < len && array[i] < end) ? array[i] : end;
7449 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7450 * for each code point in it */
7451 for (; current < upper; current++) {
7452 const STRLEN offset = (STRLEN)(current - start);
7453 swatch[offset >> 3] |= 1 << (offset & 7);
7458 /* Quit if at the end of the list */
7461 /* But first, have to deal with the highest possible code point on
7462 * the platform. The previous code assumes that <end> is one
7463 * beyond where we want to populate, but that is impossible at the
7464 * platform's infinity, so have to handle it specially */
7465 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7467 const STRLEN offset = (STRLEN)(end - start);
7468 swatch[offset >> 3] |= 1 << (offset & 7);
7473 /* Advance to the next range, which will be for code points not in the
7482 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7484 /* Take the union of two inversion lists and point <output> to it. *output
7485 * should be defined upon input, and if it points to one of the two lists,
7486 * the reference count to that list will be decremented. The first list,
7487 * <a>, may be NULL, in which case a copy of the second list is returned.
7488 * If <complement_b> is TRUE, the union is taken of the complement
7489 * (inversion) of <b> instead of b itself.
7491 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7492 * Richard Gillam, published by Addison-Wesley, and explained at some
7493 * length there. The preface says to incorporate its examples into your
7494 * code at your own risk.
7496 * The algorithm is like a merge sort.
7498 * XXX A potential performance improvement is to keep track as we go along
7499 * if only one of the inputs contributes to the result, meaning the other
7500 * is a subset of that one. In that case, we can skip the final copy and
7501 * return the larger of the input lists, but then outside code might need
7502 * to keep track of whether to free the input list or not */
7504 UV* array_a; /* a's array */
7506 UV len_a; /* length of a's array */
7509 SV* u; /* the resulting union */
7513 UV i_a = 0; /* current index into a's array */
7517 /* running count, as explained in the algorithm source book; items are
7518 * stopped accumulating and are output when the count changes to/from 0.
7519 * The count is incremented when we start a range that's in the set, and
7520 * decremented when we start a range that's not in the set. So its range
7521 * is 0 to 2. Only when the count is zero is something not in the set.
7525 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7528 /* If either one is empty, the union is the other one */
7529 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7536 *output = invlist_clone(b);
7538 _invlist_invert(*output);
7540 } /* else *output already = b; */
7543 else if ((len_b = _invlist_len(b)) == 0) {
7548 /* The complement of an empty list is a list that has everything in it,
7549 * so the union with <a> includes everything too */
7554 *output = _new_invlist(1);
7555 _append_range_to_invlist(*output, 0, UV_MAX);
7557 else if (*output != a) {
7558 *output = invlist_clone(a);
7560 /* else *output already = a; */
7564 /* Here both lists exist and are non-empty */
7565 array_a = invlist_array(a);
7566 array_b = invlist_array(b);
7568 /* If are to take the union of 'a' with the complement of b, set it
7569 * up so are looking at b's complement. */
7572 /* To complement, we invert: if the first element is 0, remove it. To
7573 * do this, we just pretend the array starts one later, and clear the
7574 * flag as we don't have to do anything else later */
7575 if (array_b[0] == 0) {
7578 complement_b = FALSE;
7582 /* But if the first element is not zero, we unshift a 0 before the
7583 * array. The data structure reserves a space for that 0 (which
7584 * should be a '1' right now), so physical shifting is unneeded,
7585 * but temporarily change that element to 0. Before exiting the
7586 * routine, we must restore the element to '1' */
7593 /* Size the union for the worst case: that the sets are completely
7595 u = _new_invlist(len_a + len_b);
7597 /* Will contain U+0000 if either component does */
7598 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7599 || (len_b > 0 && array_b[0] == 0));
7601 /* Go through each list item by item, stopping when exhausted one of
7603 while (i_a < len_a && i_b < len_b) {
7604 UV cp; /* The element to potentially add to the union's array */
7605 bool cp_in_set; /* is it in the the input list's set or not */
7607 /* We need to take one or the other of the two inputs for the union.
7608 * Since we are merging two sorted lists, we take the smaller of the
7609 * next items. In case of a tie, we take the one that is in its set
7610 * first. If we took one not in the set first, it would decrement the
7611 * count, possibly to 0 which would cause it to be output as ending the
7612 * range, and the next time through we would take the same number, and
7613 * output it again as beginning the next range. By doing it the
7614 * opposite way, there is no possibility that the count will be
7615 * momentarily decremented to 0, and thus the two adjoining ranges will
7616 * be seamlessly merged. (In a tie and both are in the set or both not
7617 * in the set, it doesn't matter which we take first.) */
7618 if (array_a[i_a] < array_b[i_b]
7619 || (array_a[i_a] == array_b[i_b]
7620 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7622 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7626 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7630 /* Here, have chosen which of the two inputs to look at. Only output
7631 * if the running count changes to/from 0, which marks the
7632 * beginning/end of a range in that's in the set */
7635 array_u[i_u++] = cp;
7642 array_u[i_u++] = cp;
7647 /* Here, we are finished going through at least one of the lists, which
7648 * means there is something remaining in at most one. We check if the list
7649 * that hasn't been exhausted is positioned such that we are in the middle
7650 * of a range in its set or not. (i_a and i_b point to the element beyond
7651 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7652 * is potentially more to output.
7653 * There are four cases:
7654 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7655 * in the union is entirely from the non-exhausted set.
7656 * 2) Both were in their sets, count is 2. Nothing further should
7657 * be output, as everything that remains will be in the exhausted
7658 * list's set, hence in the union; decrementing to 1 but not 0 insures
7660 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7661 * Nothing further should be output because the union includes
7662 * everything from the exhausted set. Not decrementing ensures that.
7663 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7664 * decrementing to 0 insures that we look at the remainder of the
7665 * non-exhausted set */
7666 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7667 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7672 /* The final length is what we've output so far, plus what else is about to
7673 * be output. (If 'count' is non-zero, then the input list we exhausted
7674 * has everything remaining up to the machine's limit in its set, and hence
7675 * in the union, so there will be no further output. */
7678 /* At most one of the subexpressions will be non-zero */
7679 len_u += (len_a - i_a) + (len_b - i_b);
7682 /* Set result to final length, which can change the pointer to array_u, so
7684 if (len_u != _invlist_len(u)) {
7685 invlist_set_len(u, len_u);
7687 array_u = invlist_array(u);
7690 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7691 * the other) ended with everything above it not in its set. That means
7692 * that the remaining part of the union is precisely the same as the
7693 * non-exhausted list, so can just copy it unchanged. (If both list were
7694 * exhausted at the same time, then the operations below will be both 0.)
7697 IV copy_count; /* At most one will have a non-zero copy count */
7698 if ((copy_count = len_a - i_a) > 0) {
7699 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7701 else if ((copy_count = len_b - i_b) > 0) {
7702 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7706 /* We may be removing a reference to one of the inputs */
7707 if (a == *output || b == *output) {
7708 SvREFCNT_dec(*output);
7711 /* If we've changed b, restore it */
7721 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7723 /* Take the intersection of two inversion lists and point <i> to it. *i
7724 * should be defined upon input, and if it points to one of the two lists,
7725 * the reference count to that list will be decremented.
7726 * If <complement_b> is TRUE, the result will be the intersection of <a>
7727 * and the complement (or inversion) of <b> instead of <b> directly.
7729 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7730 * Richard Gillam, published by Addison-Wesley, and explained at some
7731 * length there. The preface says to incorporate its examples into your
7732 * code at your own risk. In fact, it had bugs
7734 * The algorithm is like a merge sort, and is essentially the same as the
7738 UV* array_a; /* a's array */
7740 UV len_a; /* length of a's array */
7743 SV* r; /* the resulting intersection */
7747 UV i_a = 0; /* current index into a's array */
7751 /* running count, as explained in the algorithm source book; items are
7752 * stopped accumulating and are output when the count changes to/from 2.
7753 * The count is incremented when we start a range that's in the set, and
7754 * decremented when we start a range that's not in the set. So its range
7755 * is 0 to 2. Only when the count is 2 is something in the intersection.
7759 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7762 /* Special case if either one is empty */
7763 len_a = _invlist_len(a);
7764 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7766 if (len_a != 0 && complement_b) {
7768 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7769 * be empty. Here, also we are using 'b's complement, which hence
7770 * must be every possible code point. Thus the intersection is
7773 *i = invlist_clone(a);
7779 /* else *i is already 'a' */
7783 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7784 * intersection must be empty */
7791 *i = _new_invlist(0);
7795 /* Here both lists exist and are non-empty */
7796 array_a = invlist_array(a);
7797 array_b = invlist_array(b);
7799 /* If are to take the intersection of 'a' with the complement of b, set it
7800 * up so are looking at b's complement. */
7803 /* To complement, we invert: if the first element is 0, remove it. To
7804 * do this, we just pretend the array starts one later, and clear the
7805 * flag as we don't have to do anything else later */
7806 if (array_b[0] == 0) {
7809 complement_b = FALSE;
7813 /* But if the first element is not zero, we unshift a 0 before the
7814 * array. The data structure reserves a space for that 0 (which
7815 * should be a '1' right now), so physical shifting is unneeded,
7816 * but temporarily change that element to 0. Before exiting the
7817 * routine, we must restore the element to '1' */
7824 /* Size the intersection for the worst case: that the intersection ends up
7825 * fragmenting everything to be completely disjoint */
7826 r= _new_invlist(len_a + len_b);
7828 /* Will contain U+0000 iff both components do */
7829 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7830 && len_b > 0 && array_b[0] == 0);
7832 /* Go through each list item by item, stopping when exhausted one of
7834 while (i_a < len_a && i_b < len_b) {
7835 UV cp; /* The element to potentially add to the intersection's
7837 bool cp_in_set; /* Is it in the input list's set or not */
7839 /* We need to take one or the other of the two inputs for the
7840 * intersection. Since we are merging two sorted lists, we take the
7841 * smaller of the next items. In case of a tie, we take the one that
7842 * is not in its set first (a difference from the union algorithm). If
7843 * we took one in the set first, it would increment the count, possibly
7844 * to 2 which would cause it to be output as starting a range in the
7845 * intersection, and the next time through we would take that same
7846 * number, and output it again as ending the set. By doing it the
7847 * opposite of this, there is no possibility that the count will be
7848 * momentarily incremented to 2. (In a tie and both are in the set or
7849 * both not in the set, it doesn't matter which we take first.) */
7850 if (array_a[i_a] < array_b[i_b]
7851 || (array_a[i_a] == array_b[i_b]
7852 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7854 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7858 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7862 /* Here, have chosen which of the two inputs to look at. Only output
7863 * if the running count changes to/from 2, which marks the
7864 * beginning/end of a range that's in the intersection */
7868 array_r[i_r++] = cp;
7873 array_r[i_r++] = cp;
7879 /* Here, we are finished going through at least one of the lists, which
7880 * means there is something remaining in at most one. We check if the list
7881 * that has been exhausted is positioned such that we are in the middle
7882 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7883 * the ones we care about.) There are four cases:
7884 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7885 * nothing left in the intersection.
7886 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7887 * above 2. What should be output is exactly that which is in the
7888 * non-exhausted set, as everything it has is also in the intersection
7889 * set, and everything it doesn't have can't be in the intersection
7890 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7891 * gets incremented to 2. Like the previous case, the intersection is
7892 * everything that remains in the non-exhausted set.
7893 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7894 * remains 1. And the intersection has nothing more. */
7895 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7896 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7901 /* The final length is what we've output so far plus what else is in the
7902 * intersection. At most one of the subexpressions below will be non-zero */
7905 len_r += (len_a - i_a) + (len_b - i_b);
7908 /* Set result to final length, which can change the pointer to array_r, so
7910 if (len_r != _invlist_len(r)) {
7911 invlist_set_len(r, len_r);
7913 array_r = invlist_array(r);
7916 /* Finish outputting any remaining */
7917 if (count >= 2) { /* At most one will have a non-zero copy count */
7919 if ((copy_count = len_a - i_a) > 0) {
7920 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7922 else if ((copy_count = len_b - i_b) > 0) {
7923 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7927 /* We may be removing a reference to one of the inputs */
7928 if (a == *i || b == *i) {
7932 /* If we've changed b, restore it */
7942 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7944 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7945 * set. A pointer to the inversion list is returned. This may actually be
7946 * a new list, in which case the passed in one has been destroyed. The
7947 * passed in inversion list can be NULL, in which case a new one is created
7948 * with just the one range in it */
7953 if (invlist == NULL) {
7954 invlist = _new_invlist(2);
7958 len = _invlist_len(invlist);
7961 /* If comes after the final entry, can just append it to the end */
7963 || start >= invlist_array(invlist)
7964 [_invlist_len(invlist) - 1])
7966 _append_range_to_invlist(invlist, start, end);
7970 /* Here, can't just append things, create and return a new inversion list
7971 * which is the union of this range and the existing inversion list */
7972 range_invlist = _new_invlist(2);
7973 _append_range_to_invlist(range_invlist, start, end);
7975 _invlist_union(invlist, range_invlist, &invlist);
7977 /* The temporary can be freed */
7978 SvREFCNT_dec(range_invlist);
7985 PERL_STATIC_INLINE SV*
7986 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
7987 return _add_range_to_invlist(invlist, cp, cp);
7990 #ifndef PERL_IN_XSUB_RE
7992 Perl__invlist_invert(pTHX_ SV* const invlist)
7994 /* Complement the input inversion list. This adds a 0 if the list didn't
7995 * have a zero; removes it otherwise. As described above, the data
7996 * structure is set up so that this is very efficient */
7998 UV* len_pos = _get_invlist_len_addr(invlist);
8000 PERL_ARGS_ASSERT__INVLIST_INVERT;
8002 /* The inverse of matching nothing is matching everything */
8003 if (*len_pos == 0) {
8004 _append_range_to_invlist(invlist, 0, UV_MAX);
8008 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8009 * zero element was a 0, so it is being removed, so the length decrements
8010 * by 1; and vice-versa. SvCUR is unaffected */
8011 if (*get_invlist_zero_addr(invlist) ^= 1) {
8020 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8022 /* Complement the input inversion list (which must be a Unicode property,
8023 * all of which don't match above the Unicode maximum code point.) And
8024 * Perl has chosen to not have the inversion match above that either. This
8025 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8031 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8033 _invlist_invert(invlist);
8035 len = _invlist_len(invlist);
8037 if (len != 0) { /* If empty do nothing */
8038 array = invlist_array(invlist);
8039 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8040 /* Add 0x110000. First, grow if necessary */
8042 if (invlist_max(invlist) < len) {
8043 invlist_extend(invlist, len);
8044 array = invlist_array(invlist);
8046 invlist_set_len(invlist, len);
8047 array[len - 1] = PERL_UNICODE_MAX + 1;
8049 else { /* Remove the 0x110000 */
8050 invlist_set_len(invlist, len - 1);
8058 PERL_STATIC_INLINE SV*
8059 S_invlist_clone(pTHX_ SV* const invlist)
8062 /* Return a new inversion list that is a copy of the input one, which is
8065 /* Need to allocate extra space to accommodate Perl's addition of a
8066 * trailing NUL to SvPV's, since it thinks they are always strings */
8067 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8068 STRLEN length = SvCUR(invlist);
8070 PERL_ARGS_ASSERT_INVLIST_CLONE;
8072 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8073 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8078 PERL_STATIC_INLINE UV*
8079 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8081 /* Return the address of the UV that contains the current iteration
8084 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8086 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8089 PERL_STATIC_INLINE UV*
8090 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8092 /* Return the address of the UV that contains the version id. */
8094 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8096 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8099 PERL_STATIC_INLINE void
8100 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8102 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8104 *get_invlist_iter_addr(invlist) = 0;
8108 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8110 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8111 * This call sets in <*start> and <*end>, the next range in <invlist>.
8112 * Returns <TRUE> if successful and the next call will return the next
8113 * range; <FALSE> if was already at the end of the list. If the latter,
8114 * <*start> and <*end> are unchanged, and the next call to this function
8115 * will start over at the beginning of the list */
8117 UV* pos = get_invlist_iter_addr(invlist);
8118 UV len = _invlist_len(invlist);
8121 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8124 *pos = UV_MAX; /* Force iternit() to be required next time */
8128 array = invlist_array(invlist);
8130 *start = array[(*pos)++];
8136 *end = array[(*pos)++] - 1;
8142 PERL_STATIC_INLINE UV
8143 S_invlist_highest(pTHX_ SV* const invlist)
8145 /* Returns the highest code point that matches an inversion list. This API
8146 * has an ambiguity, as it returns 0 under either the highest is actually
8147 * 0, or if the list is empty. If this distinction matters to you, check
8148 * for emptiness before calling this function */
8150 UV len = _invlist_len(invlist);
8153 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8159 array = invlist_array(invlist);
8161 /* The last element in the array in the inversion list always starts a
8162 * range that goes to infinity. That range may be for code points that are
8163 * matched in the inversion list, or it may be for ones that aren't
8164 * matched. In the latter case, the highest code point in the set is one
8165 * less than the beginning of this range; otherwise it is the final element
8166 * of this range: infinity */
8167 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8169 : array[len - 1] - 1;
8172 #ifndef PERL_IN_XSUB_RE
8174 Perl__invlist_contents(pTHX_ SV* const invlist)
8176 /* Get the contents of an inversion list into a string SV so that they can
8177 * be printed out. It uses the format traditionally done for debug tracing
8181 SV* output = newSVpvs("\n");
8183 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8185 invlist_iterinit(invlist);
8186 while (invlist_iternext(invlist, &start, &end)) {
8187 if (end == UV_MAX) {
8188 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8190 else if (end != start) {
8191 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8195 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8205 S_invlist_dump(pTHX_ SV* const invlist, const char * const header)
8207 /* Dumps out the ranges in an inversion list. The string 'header'
8208 * if present is output on a line before the first range */
8212 if (header && strlen(header)) {
8213 PerlIO_printf(Perl_debug_log, "%s\n", header);
8215 invlist_iterinit(invlist);
8216 while (invlist_iternext(invlist, &start, &end)) {
8217 if (end == UV_MAX) {
8218 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8221 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n", start, end);
8229 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8231 /* Return a boolean as to if the two passed in inversion lists are
8232 * identical. The final argument, if TRUE, says to take the complement of
8233 * the second inversion list before doing the comparison */
8235 UV* array_a = invlist_array(a);
8236 UV* array_b = invlist_array(b);
8237 UV len_a = _invlist_len(a);
8238 UV len_b = _invlist_len(b);
8240 UV i = 0; /* current index into the arrays */
8241 bool retval = TRUE; /* Assume are identical until proven otherwise */
8243 PERL_ARGS_ASSERT__INVLISTEQ;
8245 /* If are to compare 'a' with the complement of b, set it
8246 * up so are looking at b's complement. */
8249 /* The complement of nothing is everything, so <a> would have to have
8250 * just one element, starting at zero (ending at infinity) */
8252 return (len_a == 1 && array_a[0] == 0);
8254 else if (array_b[0] == 0) {
8256 /* Otherwise, to complement, we invert. Here, the first element is
8257 * 0, just remove it. To do this, we just pretend the array starts
8258 * one later, and clear the flag as we don't have to do anything
8263 complement_b = FALSE;
8267 /* But if the first element is not zero, we unshift a 0 before the
8268 * array. The data structure reserves a space for that 0 (which
8269 * should be a '1' right now), so physical shifting is unneeded,
8270 * but temporarily change that element to 0. Before exiting the
8271 * routine, we must restore the element to '1' */
8278 /* Make sure that the lengths are the same, as well as the final element
8279 * before looping through the remainder. (Thus we test the length, final,
8280 * and first elements right off the bat) */
8281 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8284 else for (i = 0; i < len_a - 1; i++) {
8285 if (array_a[i] != array_b[i]) {
8298 #undef HEADER_LENGTH
8299 #undef INVLIST_INITIAL_LENGTH
8300 #undef TO_INTERNAL_SIZE
8301 #undef FROM_INTERNAL_SIZE
8302 #undef INVLIST_LEN_OFFSET
8303 #undef INVLIST_ZERO_OFFSET
8304 #undef INVLIST_ITER_OFFSET
8305 #undef INVLIST_VERSION_ID
8307 /* End of inversion list object */
8310 - reg - regular expression, i.e. main body or parenthesized thing
8312 * Caller must absorb opening parenthesis.
8314 * Combining parenthesis handling with the base level of regular expression
8315 * is a trifle forced, but the need to tie the tails of the branches to what
8316 * follows makes it hard to avoid.
8318 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8320 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8322 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8326 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8327 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8330 regnode *ret; /* Will be the head of the group. */
8333 regnode *ender = NULL;
8336 U32 oregflags = RExC_flags;
8337 bool have_branch = 0;
8339 I32 freeze_paren = 0;
8340 I32 after_freeze = 0;
8342 /* for (?g), (?gc), and (?o) warnings; warning
8343 about (?c) will warn about (?g) -- japhy */
8345 #define WASTED_O 0x01
8346 #define WASTED_G 0x02
8347 #define WASTED_C 0x04
8348 #define WASTED_GC (0x02|0x04)
8349 I32 wastedflags = 0x00;
8351 char * parse_start = RExC_parse; /* MJD */
8352 char * const oregcomp_parse = RExC_parse;
8354 GET_RE_DEBUG_FLAGS_DECL;
8356 PERL_ARGS_ASSERT_REG;
8357 DEBUG_PARSE("reg ");
8359 *flagp = 0; /* Tentatively. */
8362 /* Make an OPEN node, if parenthesized. */
8364 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8365 char *start_verb = RExC_parse;
8366 STRLEN verb_len = 0;
8367 char *start_arg = NULL;
8368 unsigned char op = 0;
8370 int internal_argval = 0; /* internal_argval is only useful if !argok */
8371 while ( *RExC_parse && *RExC_parse != ')' ) {
8372 if ( *RExC_parse == ':' ) {
8373 start_arg = RExC_parse + 1;
8379 verb_len = RExC_parse - start_verb;
8382 while ( *RExC_parse && *RExC_parse != ')' )
8384 if ( *RExC_parse != ')' )
8385 vFAIL("Unterminated verb pattern argument");
8386 if ( RExC_parse == start_arg )
8389 if ( *RExC_parse != ')' )
8390 vFAIL("Unterminated verb pattern");
8393 switch ( *start_verb ) {
8394 case 'A': /* (*ACCEPT) */
8395 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8397 internal_argval = RExC_nestroot;
8400 case 'C': /* (*COMMIT) */
8401 if ( memEQs(start_verb,verb_len,"COMMIT") )
8404 case 'F': /* (*FAIL) */
8405 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8410 case ':': /* (*:NAME) */
8411 case 'M': /* (*MARK:NAME) */
8412 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8417 case 'P': /* (*PRUNE) */
8418 if ( memEQs(start_verb,verb_len,"PRUNE") )
8421 case 'S': /* (*SKIP) */
8422 if ( memEQs(start_verb,verb_len,"SKIP") )
8425 case 'T': /* (*THEN) */
8426 /* [19:06] <TimToady> :: is then */
8427 if ( memEQs(start_verb,verb_len,"THEN") ) {
8429 RExC_seen |= REG_SEEN_CUTGROUP;
8435 vFAIL3("Unknown verb pattern '%.*s'",
8436 verb_len, start_verb);
8439 if ( start_arg && internal_argval ) {
8440 vFAIL3("Verb pattern '%.*s' may not have an argument",
8441 verb_len, start_verb);
8442 } else if ( argok < 0 && !start_arg ) {
8443 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8444 verb_len, start_verb);
8446 ret = reganode(pRExC_state, op, internal_argval);
8447 if ( ! internal_argval && ! SIZE_ONLY ) {
8449 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8450 ARG(ret) = add_data( pRExC_state, 1, "S" );
8451 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8458 if (!internal_argval)
8459 RExC_seen |= REG_SEEN_VERBARG;
8460 } else if ( start_arg ) {
8461 vFAIL3("Verb pattern '%.*s' may not have an argument",
8462 verb_len, start_verb);
8464 ret = reg_node(pRExC_state, op);
8466 nextchar(pRExC_state);
8469 if (*RExC_parse == '?') { /* (?...) */
8470 bool is_logical = 0;
8471 const char * const seqstart = RExC_parse;
8472 bool has_use_defaults = FALSE;
8475 paren = *RExC_parse++;
8476 ret = NULL; /* For look-ahead/behind. */
8479 case 'P': /* (?P...) variants for those used to PCRE/Python */
8480 paren = *RExC_parse++;
8481 if ( paren == '<') /* (?P<...>) named capture */
8483 else if (paren == '>') { /* (?P>name) named recursion */
8484 goto named_recursion;
8486 else if (paren == '=') { /* (?P=...) named backref */
8487 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8488 you change this make sure you change that */
8489 char* name_start = RExC_parse;
8491 SV *sv_dat = reg_scan_name(pRExC_state,
8492 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8493 if (RExC_parse == name_start || *RExC_parse != ')')
8494 vFAIL2("Sequence %.3s... not terminated",parse_start);
8497 num = add_data( pRExC_state, 1, "S" );
8498 RExC_rxi->data->data[num]=(void*)sv_dat;
8499 SvREFCNT_inc_simple_void(sv_dat);
8502 ret = reganode(pRExC_state,
8505 : (ASCII_FOLD_RESTRICTED)
8507 : (AT_LEAST_UNI_SEMANTICS)
8515 Set_Node_Offset(ret, parse_start+1);
8516 Set_Node_Cur_Length(ret); /* MJD */
8518 nextchar(pRExC_state);
8522 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8524 case '<': /* (?<...) */
8525 if (*RExC_parse == '!')
8527 else if (*RExC_parse != '=')
8533 case '\'': /* (?'...') */
8534 name_start= RExC_parse;
8535 svname = reg_scan_name(pRExC_state,
8536 SIZE_ONLY ? /* reverse test from the others */
8537 REG_RSN_RETURN_NAME :
8538 REG_RSN_RETURN_NULL);
8539 if (RExC_parse == name_start) {
8541 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8544 if (*RExC_parse != paren)
8545 vFAIL2("Sequence (?%c... not terminated",
8546 paren=='>' ? '<' : paren);
8550 if (!svname) /* shouldn't happen */
8552 "panic: reg_scan_name returned NULL");
8553 if (!RExC_paren_names) {
8554 RExC_paren_names= newHV();
8555 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8557 RExC_paren_name_list= newAV();
8558 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8561 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8563 sv_dat = HeVAL(he_str);
8565 /* croak baby croak */
8567 "panic: paren_name hash element allocation failed");
8568 } else if ( SvPOK(sv_dat) ) {
8569 /* (?|...) can mean we have dupes so scan to check
8570 its already been stored. Maybe a flag indicating
8571 we are inside such a construct would be useful,
8572 but the arrays are likely to be quite small, so
8573 for now we punt -- dmq */
8574 IV count = SvIV(sv_dat);
8575 I32 *pv = (I32*)SvPVX(sv_dat);
8577 for ( i = 0 ; i < count ; i++ ) {
8578 if ( pv[i] == RExC_npar ) {
8584 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8585 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8586 pv[count] = RExC_npar;
8587 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8590 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8591 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8593 SvIV_set(sv_dat, 1);
8596 /* Yes this does cause a memory leak in debugging Perls */
8597 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8598 SvREFCNT_dec(svname);
8601 /*sv_dump(sv_dat);*/
8603 nextchar(pRExC_state);
8605 goto capturing_parens;
8607 RExC_seen |= REG_SEEN_LOOKBEHIND;
8608 RExC_in_lookbehind++;
8610 case '=': /* (?=...) */
8611 RExC_seen_zerolen++;
8613 case '!': /* (?!...) */
8614 RExC_seen_zerolen++;
8615 if (*RExC_parse == ')') {
8616 ret=reg_node(pRExC_state, OPFAIL);
8617 nextchar(pRExC_state);
8621 case '|': /* (?|...) */
8622 /* branch reset, behave like a (?:...) except that
8623 buffers in alternations share the same numbers */
8625 after_freeze = freeze_paren = RExC_npar;
8627 case ':': /* (?:...) */
8628 case '>': /* (?>...) */
8630 case '$': /* (?$...) */
8631 case '@': /* (?@...) */
8632 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8634 case '#': /* (?#...) */
8635 while (*RExC_parse && *RExC_parse != ')')
8637 if (*RExC_parse != ')')
8638 FAIL("Sequence (?#... not terminated");
8639 nextchar(pRExC_state);
8642 case '0' : /* (?0) */
8643 case 'R' : /* (?R) */
8644 if (*RExC_parse != ')')
8645 FAIL("Sequence (?R) not terminated");
8646 ret = reg_node(pRExC_state, GOSTART);
8647 *flagp |= POSTPONED;
8648 nextchar(pRExC_state);
8651 { /* named and numeric backreferences */
8653 case '&': /* (?&NAME) */
8654 parse_start = RExC_parse - 1;
8657 SV *sv_dat = reg_scan_name(pRExC_state,
8658 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8659 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8661 goto gen_recurse_regop;
8662 assert(0); /* NOT REACHED */
8664 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8666 vFAIL("Illegal pattern");
8668 goto parse_recursion;
8670 case '-': /* (?-1) */
8671 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8672 RExC_parse--; /* rewind to let it be handled later */
8676 case '1': case '2': case '3': case '4': /* (?1) */
8677 case '5': case '6': case '7': case '8': case '9':
8680 num = atoi(RExC_parse);
8681 parse_start = RExC_parse - 1; /* MJD */
8682 if (*RExC_parse == '-')
8684 while (isDIGIT(*RExC_parse))
8686 if (*RExC_parse!=')')
8687 vFAIL("Expecting close bracket");
8690 if ( paren == '-' ) {
8692 Diagram of capture buffer numbering.
8693 Top line is the normal capture buffer numbers
8694 Bottom line is the negative indexing as from
8698 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8702 num = RExC_npar + num;
8705 vFAIL("Reference to nonexistent group");
8707 } else if ( paren == '+' ) {
8708 num = RExC_npar + num - 1;
8711 ret = reganode(pRExC_state, GOSUB, num);
8713 if (num > (I32)RExC_rx->nparens) {
8715 vFAIL("Reference to nonexistent group");
8717 ARG2L_SET( ret, RExC_recurse_count++);
8719 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8720 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8724 RExC_seen |= REG_SEEN_RECURSE;
8725 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8726 Set_Node_Offset(ret, parse_start); /* MJD */
8728 *flagp |= POSTPONED;
8729 nextchar(pRExC_state);
8731 } /* named and numeric backreferences */
8732 assert(0); /* NOT REACHED */
8734 case '?': /* (??...) */
8736 if (*RExC_parse != '{') {
8738 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8741 *flagp |= POSTPONED;
8742 paren = *RExC_parse++;
8744 case '{': /* (?{...}) */
8747 struct reg_code_block *cb;
8749 RExC_seen_zerolen++;
8751 if ( !pRExC_state->num_code_blocks
8752 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8753 || pRExC_state->code_blocks[pRExC_state->code_index].start
8754 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8757 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8758 FAIL("panic: Sequence (?{...}): no code block found\n");
8759 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8761 /* this is a pre-compiled code block (?{...}) */
8762 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8763 RExC_parse = RExC_start + cb->end;
8766 if (cb->src_regex) {
8767 n = add_data(pRExC_state, 2, "rl");
8768 RExC_rxi->data->data[n] =
8769 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8770 RExC_rxi->data->data[n+1] = (void*)o;
8773 n = add_data(pRExC_state, 1,
8774 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8775 RExC_rxi->data->data[n] = (void*)o;
8778 pRExC_state->code_index++;
8779 nextchar(pRExC_state);
8783 ret = reg_node(pRExC_state, LOGICAL);
8784 eval = reganode(pRExC_state, EVAL, n);
8787 /* for later propagation into (??{}) return value */
8788 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8790 REGTAIL(pRExC_state, ret, eval);
8791 /* deal with the length of this later - MJD */
8794 ret = reganode(pRExC_state, EVAL, n);
8795 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8796 Set_Node_Offset(ret, parse_start);
8799 case '(': /* (?(?{...})...) and (?(?=...)...) */
8802 if (RExC_parse[0] == '?') { /* (?(?...)) */
8803 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8804 || RExC_parse[1] == '<'
8805 || RExC_parse[1] == '{') { /* Lookahead or eval. */
8808 ret = reg_node(pRExC_state, LOGICAL);
8811 REGTAIL(pRExC_state, ret, reg(pRExC_state, 1, &flag,depth+1));
8815 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
8816 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
8818 char ch = RExC_parse[0] == '<' ? '>' : '\'';
8819 char *name_start= RExC_parse++;
8821 SV *sv_dat=reg_scan_name(pRExC_state,
8822 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8823 if (RExC_parse == name_start || *RExC_parse != ch)
8824 vFAIL2("Sequence (?(%c... not terminated",
8825 (ch == '>' ? '<' : ch));
8828 num = add_data( pRExC_state, 1, "S" );
8829 RExC_rxi->data->data[num]=(void*)sv_dat;
8830 SvREFCNT_inc_simple_void(sv_dat);
8832 ret = reganode(pRExC_state,NGROUPP,num);
8833 goto insert_if_check_paren;
8835 else if (RExC_parse[0] == 'D' &&
8836 RExC_parse[1] == 'E' &&
8837 RExC_parse[2] == 'F' &&
8838 RExC_parse[3] == 'I' &&
8839 RExC_parse[4] == 'N' &&
8840 RExC_parse[5] == 'E')
8842 ret = reganode(pRExC_state,DEFINEP,0);
8845 goto insert_if_check_paren;
8847 else if (RExC_parse[0] == 'R') {
8850 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
8851 parno = atoi(RExC_parse++);
8852 while (isDIGIT(*RExC_parse))
8854 } else if (RExC_parse[0] == '&') {
8857 sv_dat = reg_scan_name(pRExC_state,
8858 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8859 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8861 ret = reganode(pRExC_state,INSUBP,parno);
8862 goto insert_if_check_paren;
8864 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
8867 parno = atoi(RExC_parse++);
8869 while (isDIGIT(*RExC_parse))
8871 ret = reganode(pRExC_state, GROUPP, parno);
8873 insert_if_check_paren:
8874 if ((c = *nextchar(pRExC_state)) != ')')
8875 vFAIL("Switch condition not recognized");
8877 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
8878 br = regbranch(pRExC_state, &flags, 1,depth+1);
8880 br = reganode(pRExC_state, LONGJMP, 0);
8882 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
8883 c = *nextchar(pRExC_state);
8888 vFAIL("(?(DEFINE)....) does not allow branches");
8889 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
8890 regbranch(pRExC_state, &flags, 1,depth+1);
8891 REGTAIL(pRExC_state, ret, lastbr);
8894 c = *nextchar(pRExC_state);
8899 vFAIL("Switch (?(condition)... contains too many branches");
8900 ender = reg_node(pRExC_state, TAIL);
8901 REGTAIL(pRExC_state, br, ender);
8903 REGTAIL(pRExC_state, lastbr, ender);
8904 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
8907 REGTAIL(pRExC_state, ret, ender);
8908 RExC_size++; /* XXX WHY do we need this?!!
8909 For large programs it seems to be required
8910 but I can't figure out why. -- dmq*/
8914 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
8918 RExC_parse--; /* for vFAIL to print correctly */
8919 vFAIL("Sequence (? incomplete");
8921 case DEFAULT_PAT_MOD: /* Use default flags with the exceptions
8923 has_use_defaults = TRUE;
8924 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8925 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8926 ? REGEX_UNICODE_CHARSET
8927 : REGEX_DEPENDS_CHARSET);
8931 parse_flags: /* (?i) */
8933 U32 posflags = 0, negflags = 0;
8934 U32 *flagsp = &posflags;
8935 char has_charset_modifier = '\0';
8936 regex_charset cs = get_regex_charset(RExC_flags);
8937 if (cs == REGEX_DEPENDS_CHARSET
8938 && (RExC_utf8 || RExC_uni_semantics))
8940 cs = REGEX_UNICODE_CHARSET;
8943 while (*RExC_parse) {
8944 /* && strchr("iogcmsx", *RExC_parse) */
8945 /* (?g), (?gc) and (?o) are useless here
8946 and must be globally applied -- japhy */
8947 switch (*RExC_parse) {
8948 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8949 case LOCALE_PAT_MOD:
8950 if (has_charset_modifier) {
8951 goto excess_modifier;
8953 else if (flagsp == &negflags) {
8956 cs = REGEX_LOCALE_CHARSET;
8957 has_charset_modifier = LOCALE_PAT_MOD;
8958 RExC_contains_locale = 1;
8960 case UNICODE_PAT_MOD:
8961 if (has_charset_modifier) {
8962 goto excess_modifier;
8964 else if (flagsp == &negflags) {
8967 cs = REGEX_UNICODE_CHARSET;
8968 has_charset_modifier = UNICODE_PAT_MOD;
8970 case ASCII_RESTRICT_PAT_MOD:
8971 if (flagsp == &negflags) {
8974 if (has_charset_modifier) {
8975 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
8976 goto excess_modifier;
8978 /* Doubled modifier implies more restricted */
8979 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
8982 cs = REGEX_ASCII_RESTRICTED_CHARSET;
8984 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
8986 case DEPENDS_PAT_MOD:
8987 if (has_use_defaults) {
8988 goto fail_modifiers;
8990 else if (flagsp == &negflags) {
8993 else if (has_charset_modifier) {
8994 goto excess_modifier;
8997 /* The dual charset means unicode semantics if the
8998 * pattern (or target, not known until runtime) are
8999 * utf8, or something in the pattern indicates unicode
9001 cs = (RExC_utf8 || RExC_uni_semantics)
9002 ? REGEX_UNICODE_CHARSET
9003 : REGEX_DEPENDS_CHARSET;
9004 has_charset_modifier = DEPENDS_PAT_MOD;
9008 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9009 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9011 else if (has_charset_modifier == *(RExC_parse - 1)) {
9012 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
9015 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9020 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
9022 case ONCE_PAT_MOD: /* 'o' */
9023 case GLOBAL_PAT_MOD: /* 'g' */
9024 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9025 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
9026 if (! (wastedflags & wflagbit) ) {
9027 wastedflags |= wflagbit;
9030 "Useless (%s%c) - %suse /%c modifier",
9031 flagsp == &negflags ? "?-" : "?",
9033 flagsp == &negflags ? "don't " : "",
9040 case CONTINUE_PAT_MOD: /* 'c' */
9041 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9042 if (! (wastedflags & WASTED_C) ) {
9043 wastedflags |= WASTED_GC;
9046 "Useless (%sc) - %suse /gc modifier",
9047 flagsp == &negflags ? "?-" : "?",
9048 flagsp == &negflags ? "don't " : ""
9053 case KEEPCOPY_PAT_MOD: /* 'p' */
9054 if (flagsp == &negflags) {
9056 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9058 *flagsp |= RXf_PMf_KEEPCOPY;
9062 /* A flag is a default iff it is following a minus, so
9063 * if there is a minus, it means will be trying to
9064 * re-specify a default which is an error */
9065 if (has_use_defaults || flagsp == &negflags) {
9068 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9072 wastedflags = 0; /* reset so (?g-c) warns twice */
9078 RExC_flags |= posflags;
9079 RExC_flags &= ~negflags;
9080 set_regex_charset(&RExC_flags, cs);
9082 oregflags |= posflags;
9083 oregflags &= ~negflags;
9084 set_regex_charset(&oregflags, cs);
9086 nextchar(pRExC_state);
9097 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9102 }} /* one for the default block, one for the switch */
9109 ret = reganode(pRExC_state, OPEN, parno);
9112 RExC_nestroot = parno;
9113 if (RExC_seen & REG_SEEN_RECURSE
9114 && !RExC_open_parens[parno-1])
9116 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9117 "Setting open paren #%"IVdf" to %d\n",
9118 (IV)parno, REG_NODE_NUM(ret)));
9119 RExC_open_parens[parno-1]= ret;
9122 Set_Node_Length(ret, 1); /* MJD */
9123 Set_Node_Offset(ret, RExC_parse); /* MJD */
9131 /* Pick up the branches, linking them together. */
9132 parse_start = RExC_parse; /* MJD */
9133 br = regbranch(pRExC_state, &flags, 1,depth+1);
9135 /* branch_len = (paren != 0); */
9139 if (*RExC_parse == '|') {
9140 if (!SIZE_ONLY && RExC_extralen) {
9141 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9144 reginsert(pRExC_state, BRANCH, br, depth+1);
9145 Set_Node_Length(br, paren != 0);
9146 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9150 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9152 else if (paren == ':') {
9153 *flagp |= flags&SIMPLE;
9155 if (is_open) { /* Starts with OPEN. */
9156 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9158 else if (paren != '?') /* Not Conditional */
9160 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9162 while (*RExC_parse == '|') {
9163 if (!SIZE_ONLY && RExC_extralen) {
9164 ender = reganode(pRExC_state, LONGJMP,0);
9165 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9168 RExC_extralen += 2; /* Account for LONGJMP. */
9169 nextchar(pRExC_state);
9171 if (RExC_npar > after_freeze)
9172 after_freeze = RExC_npar;
9173 RExC_npar = freeze_paren;
9175 br = regbranch(pRExC_state, &flags, 0, depth+1);
9179 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9181 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9184 if (have_branch || paren != ':') {
9185 /* Make a closing node, and hook it on the end. */
9188 ender = reg_node(pRExC_state, TAIL);
9191 ender = reganode(pRExC_state, CLOSE, parno);
9192 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9193 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9194 "Setting close paren #%"IVdf" to %d\n",
9195 (IV)parno, REG_NODE_NUM(ender)));
9196 RExC_close_parens[parno-1]= ender;
9197 if (RExC_nestroot == parno)
9200 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9201 Set_Node_Length(ender,1); /* MJD */
9207 *flagp &= ~HASWIDTH;
9210 ender = reg_node(pRExC_state, SUCCEED);
9213 ender = reg_node(pRExC_state, END);
9215 assert(!RExC_opend); /* there can only be one! */
9220 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9221 SV * const mysv_val1=sv_newmortal();
9222 SV * const mysv_val2=sv_newmortal();
9223 DEBUG_PARSE_MSG("lsbr");
9224 regprop(RExC_rx, mysv_val1, lastbr);
9225 regprop(RExC_rx, mysv_val2, ender);
9226 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9227 SvPV_nolen_const(mysv_val1),
9228 (IV)REG_NODE_NUM(lastbr),
9229 SvPV_nolen_const(mysv_val2),
9230 (IV)REG_NODE_NUM(ender),
9231 (IV)(ender - lastbr)
9234 REGTAIL(pRExC_state, lastbr, ender);
9236 if (have_branch && !SIZE_ONLY) {
9239 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9241 /* Hook the tails of the branches to the closing node. */
9242 for (br = ret; br; br = regnext(br)) {
9243 const U8 op = PL_regkind[OP(br)];
9245 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9246 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9249 else if (op == BRANCHJ) {
9250 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9251 /* for now we always disable this optimisation * /
9252 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9258 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9259 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9260 SV * const mysv_val1=sv_newmortal();
9261 SV * const mysv_val2=sv_newmortal();
9262 DEBUG_PARSE_MSG("NADA");
9263 regprop(RExC_rx, mysv_val1, ret);
9264 regprop(RExC_rx, mysv_val2, ender);
9265 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9266 SvPV_nolen_const(mysv_val1),
9267 (IV)REG_NODE_NUM(ret),
9268 SvPV_nolen_const(mysv_val2),
9269 (IV)REG_NODE_NUM(ender),
9274 if (OP(ender) == TAIL) {
9279 for ( opt= br + 1; opt < ender ; opt++ )
9281 NEXT_OFF(br)= ender - br;
9289 static const char parens[] = "=!<,>";
9291 if (paren && (p = strchr(parens, paren))) {
9292 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9293 int flag = (p - parens) > 1;
9296 node = SUSPEND, flag = 0;
9297 reginsert(pRExC_state, node,ret, depth+1);
9298 Set_Node_Cur_Length(ret);
9299 Set_Node_Offset(ret, parse_start + 1);
9301 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9305 /* Check for proper termination. */
9307 RExC_flags = oregflags;
9308 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9309 RExC_parse = oregcomp_parse;
9310 vFAIL("Unmatched (");
9313 else if (!paren && RExC_parse < RExC_end) {
9314 if (*RExC_parse == ')') {
9316 vFAIL("Unmatched )");
9319 FAIL("Junk on end of regexp"); /* "Can't happen". */
9320 assert(0); /* NOTREACHED */
9323 if (RExC_in_lookbehind) {
9324 RExC_in_lookbehind--;
9326 if (after_freeze > RExC_npar)
9327 RExC_npar = after_freeze;
9332 - regbranch - one alternative of an | operator
9334 * Implements the concatenation operator.
9337 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9341 regnode *chain = NULL;
9343 I32 flags = 0, c = 0;
9344 GET_RE_DEBUG_FLAGS_DECL;
9346 PERL_ARGS_ASSERT_REGBRANCH;
9348 DEBUG_PARSE("brnc");
9353 if (!SIZE_ONLY && RExC_extralen)
9354 ret = reganode(pRExC_state, BRANCHJ,0);
9356 ret = reg_node(pRExC_state, BRANCH);
9357 Set_Node_Length(ret, 1);
9361 if (!first && SIZE_ONLY)
9362 RExC_extralen += 1; /* BRANCHJ */
9364 *flagp = WORST; /* Tentatively. */
9367 nextchar(pRExC_state);
9368 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9370 latest = regpiece(pRExC_state, &flags,depth+1);
9371 if (latest == NULL) {
9372 if (flags & TRYAGAIN)
9376 else if (ret == NULL)
9378 *flagp |= flags&(HASWIDTH|POSTPONED);
9379 if (chain == NULL) /* First piece. */
9380 *flagp |= flags&SPSTART;
9383 REGTAIL(pRExC_state, chain, latest);
9388 if (chain == NULL) { /* Loop ran zero times. */
9389 chain = reg_node(pRExC_state, NOTHING);
9394 *flagp |= flags&SIMPLE;
9401 - regpiece - something followed by possible [*+?]
9403 * Note that the branching code sequences used for ? and the general cases
9404 * of * and + are somewhat optimized: they use the same NOTHING node as
9405 * both the endmarker for their branch list and the body of the last branch.
9406 * It might seem that this node could be dispensed with entirely, but the
9407 * endmarker role is not redundant.
9410 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9417 const char * const origparse = RExC_parse;
9419 I32 max = REG_INFTY;
9420 #ifdef RE_TRACK_PATTERN_OFFSETS
9423 const char *maxpos = NULL;
9424 GET_RE_DEBUG_FLAGS_DECL;
9426 PERL_ARGS_ASSERT_REGPIECE;
9428 DEBUG_PARSE("piec");
9430 ret = regatom(pRExC_state, &flags,depth+1);
9432 if (flags & TRYAGAIN)
9439 if (op == '{' && regcurly(RExC_parse)) {
9441 #ifdef RE_TRACK_PATTERN_OFFSETS
9442 parse_start = RExC_parse; /* MJD */
9444 next = RExC_parse + 1;
9445 while (isDIGIT(*next) || *next == ',') {
9454 if (*next == '}') { /* got one */
9458 min = atoi(RExC_parse);
9462 maxpos = RExC_parse;
9464 if (!max && *maxpos != '0')
9465 max = REG_INFTY; /* meaning "infinity" */
9466 else if (max >= REG_INFTY)
9467 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9469 nextchar(pRExC_state);
9472 if ((flags&SIMPLE)) {
9473 RExC_naughty += 2 + RExC_naughty / 2;
9474 reginsert(pRExC_state, CURLY, ret, depth+1);
9475 Set_Node_Offset(ret, parse_start+1); /* MJD */
9476 Set_Node_Cur_Length(ret);
9479 regnode * const w = reg_node(pRExC_state, WHILEM);
9482 REGTAIL(pRExC_state, ret, w);
9483 if (!SIZE_ONLY && RExC_extralen) {
9484 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9485 reginsert(pRExC_state, NOTHING,ret, depth+1);
9486 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9488 reginsert(pRExC_state, CURLYX,ret, depth+1);
9490 Set_Node_Offset(ret, parse_start+1);
9491 Set_Node_Length(ret,
9492 op == '{' ? (RExC_parse - parse_start) : 1);
9494 if (!SIZE_ONLY && RExC_extralen)
9495 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9496 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9498 RExC_whilem_seen++, RExC_extralen += 3;
9499 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9508 vFAIL("Can't do {n,m} with n > m");
9510 ARG1_SET(ret, (U16)min);
9511 ARG2_SET(ret, (U16)max);
9523 #if 0 /* Now runtime fix should be reliable. */
9525 /* if this is reinstated, don't forget to put this back into perldiag:
9527 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9529 (F) The part of the regexp subject to either the * or + quantifier
9530 could match an empty string. The {#} shows in the regular
9531 expression about where the problem was discovered.
9535 if (!(flags&HASWIDTH) && op != '?')
9536 vFAIL("Regexp *+ operand could be empty");
9539 #ifdef RE_TRACK_PATTERN_OFFSETS
9540 parse_start = RExC_parse;
9542 nextchar(pRExC_state);
9544 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9546 if (op == '*' && (flags&SIMPLE)) {
9547 reginsert(pRExC_state, STAR, ret, depth+1);
9551 else if (op == '*') {
9555 else if (op == '+' && (flags&SIMPLE)) {
9556 reginsert(pRExC_state, PLUS, ret, depth+1);
9560 else if (op == '+') {
9564 else if (op == '?') {
9569 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9570 ckWARN3reg(RExC_parse,
9571 "%.*s matches null string many times",
9572 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9576 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9577 nextchar(pRExC_state);
9578 reginsert(pRExC_state, MINMOD, ret, depth+1);
9579 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9581 #ifndef REG_ALLOW_MINMOD_SUSPEND
9584 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9586 nextchar(pRExC_state);
9587 ender = reg_node(pRExC_state, SUCCEED);
9588 REGTAIL(pRExC_state, ret, ender);
9589 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9591 ender = reg_node(pRExC_state, TAIL);
9592 REGTAIL(pRExC_state, ret, ender);
9596 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9598 vFAIL("Nested quantifiers");
9605 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class)
9608 /* This is expected to be called by a parser routine that has recognized '\N'
9609 and needs to handle the rest. RExC_parse is expected to point at the first
9610 char following the N at the time of the call. On successful return,
9611 RExC_parse has been updated to point to just after the sequence identified
9612 by this routine, and <*flagp> has been updated.
9614 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9617 \N may begin either a named sequence, or if outside a character class, mean
9618 to match a non-newline. For non single-quoted regexes, the tokenizer has
9619 attempted to decide which, and in the case of a named sequence, converted it
9620 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9621 where c1... are the characters in the sequence. For single-quoted regexes,
9622 the tokenizer passes the \N sequence through unchanged; this code will not
9623 attempt to determine this nor expand those, instead raising a syntax error.
9624 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9625 or there is no '}', it signals that this \N occurrence means to match a
9628 Only the \N{U+...} form should occur in a character class, for the same
9629 reason that '.' inside a character class means to just match a period: it
9630 just doesn't make sense.
9632 The function raises an error (via vFAIL), and doesn't return for various
9633 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9634 success; it returns FALSE otherwise.
9636 If <valuep> is non-null, it means the caller can accept an input sequence
9637 consisting of a just a single code point; <*valuep> is set to that value
9638 if the input is such.
9640 If <node_p> is non-null it signifies that the caller can accept any other
9641 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9643 1) \N means not-a-NL: points to a newly created REG_ANY node;
9644 2) \N{}: points to a new NOTHING node;
9645 3) otherwise: points to a new EXACT node containing the resolved
9647 Note that FALSE is returned for single code point sequences if <valuep> is
9651 char * endbrace; /* '}' following the name */
9653 char *endchar; /* Points to '.' or '}' ending cur char in the input
9655 bool has_multiple_chars; /* true if the input stream contains a sequence of
9656 more than one character */
9658 GET_RE_DEBUG_FLAGS_DECL;
9660 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9664 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9666 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9667 * modifier. The other meaning does not */
9668 p = (RExC_flags & RXf_PMf_EXTENDED)
9669 ? regwhite( pRExC_state, RExC_parse )
9672 /* Disambiguate between \N meaning a named character versus \N meaning
9673 * [^\n]. The former is assumed when it can't be the latter. */
9674 if (*p != '{' || regcurly(p)) {
9677 /* no bare \N in a charclass */
9678 if (in_char_class) {
9679 vFAIL("\\N in a character class must be a named character: \\N{...}");
9683 nextchar(pRExC_state);
9684 *node_p = reg_node(pRExC_state, REG_ANY);
9685 *flagp |= HASWIDTH|SIMPLE;
9688 Set_Node_Length(*node_p, 1); /* MJD */
9692 /* Here, we have decided it should be a named character or sequence */
9694 /* The test above made sure that the next real character is a '{', but
9695 * under the /x modifier, it could be separated by space (or a comment and
9696 * \n) and this is not allowed (for consistency with \x{...} and the
9697 * tokenizer handling of \N{NAME}). */
9698 if (*RExC_parse != '{') {
9699 vFAIL("Missing braces on \\N{}");
9702 RExC_parse++; /* Skip past the '{' */
9704 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9705 || ! (endbrace == RExC_parse /* nothing between the {} */
9706 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9707 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9709 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9710 vFAIL("\\N{NAME} must be resolved by the lexer");
9713 if (endbrace == RExC_parse) { /* empty: \N{} */
9716 *node_p = reg_node(pRExC_state,NOTHING);
9718 else if (in_char_class) {
9719 if (SIZE_ONLY && in_char_class) {
9720 ckWARNreg(RExC_parse,
9721 "Ignoring zero length \\N{} in character class"
9729 nextchar(pRExC_state);
9733 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9734 RExC_parse += 2; /* Skip past the 'U+' */
9736 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9738 /* Code points are separated by dots. If none, there is only one code
9739 * point, and is terminated by the brace */
9740 has_multiple_chars = (endchar < endbrace);
9742 if (valuep && (! has_multiple_chars || in_char_class)) {
9743 /* We only pay attention to the first char of
9744 multichar strings being returned in char classes. I kinda wonder
9745 if this makes sense as it does change the behaviour
9746 from earlier versions, OTOH that behaviour was broken
9747 as well. XXX Solution is to recharacterize as
9748 [rest-of-class]|multi1|multi2... */
9750 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9751 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9752 | PERL_SCAN_DISALLOW_PREFIX
9753 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9755 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9757 /* The tokenizer should have guaranteed validity, but it's possible to
9758 * bypass it by using single quoting, so check */
9759 if (length_of_hex == 0
9760 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9762 RExC_parse += length_of_hex; /* Includes all the valid */
9763 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9764 ? UTF8SKIP(RExC_parse)
9766 /* Guard against malformed utf8 */
9767 if (RExC_parse >= endchar) {
9768 RExC_parse = endchar;
9770 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9773 if (in_char_class && has_multiple_chars) {
9774 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9776 RExC_parse = endbrace + 1;
9778 else if (! node_p || ! has_multiple_chars) {
9780 /* Here, the input is legal, but not according to the caller's
9781 * options. We fail without advancing the parse, so that the
9782 * caller can try again */
9788 /* What is done here is to convert this to a sub-pattern of the form
9789 * (?:\x{char1}\x{char2}...)
9790 * and then call reg recursively. That way, it retains its atomicness,
9791 * while not having to worry about special handling that some code
9792 * points may have. toke.c has converted the original Unicode values
9793 * to native, so that we can just pass on the hex values unchanged. We
9794 * do have to set a flag to keep recoding from happening in the
9797 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9799 char *orig_end = RExC_end;
9802 while (RExC_parse < endbrace) {
9804 /* Convert to notation the rest of the code understands */
9805 sv_catpv(substitute_parse, "\\x{");
9806 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9807 sv_catpv(substitute_parse, "}");
9809 /* Point to the beginning of the next character in the sequence. */
9810 RExC_parse = endchar + 1;
9811 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9813 sv_catpv(substitute_parse, ")");
9815 RExC_parse = SvPV(substitute_parse, len);
9817 /* Don't allow empty number */
9819 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9821 RExC_end = RExC_parse + len;
9823 /* The values are Unicode, and therefore not subject to recoding */
9824 RExC_override_recoding = 1;
9826 *node_p = reg(pRExC_state, 1, &flags, depth+1);
9827 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9829 RExC_parse = endbrace;
9830 RExC_end = orig_end;
9831 RExC_override_recoding = 0;
9833 nextchar(pRExC_state);
9843 * It returns the code point in utf8 for the value in *encp.
9844 * value: a code value in the source encoding
9845 * encp: a pointer to an Encode object
9847 * If the result from Encode is not a single character,
9848 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9851 S_reg_recode(pTHX_ const char value, SV **encp)
9854 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9855 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9856 const STRLEN newlen = SvCUR(sv);
9857 UV uv = UNICODE_REPLACEMENT;
9859 PERL_ARGS_ASSERT_REG_RECODE;
9863 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9866 if (!newlen || numlen != newlen) {
9867 uv = UNICODE_REPLACEMENT;
9873 PERL_STATIC_INLINE U8
9874 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
9878 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
9884 op = get_regex_charset(RExC_flags);
9885 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
9886 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
9887 been, so there is no hole */
9893 PERL_STATIC_INLINE void
9894 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
9896 /* This knows the details about sizing an EXACTish node, setting flags for
9897 * it (by setting <*flagp>, and potentially populating it with a single
9900 * If <len> is non-zero, this function assumes that the node has already
9901 * been populated, and just does the sizing. In this case <code_point>
9902 * should be the final code point that has already been placed into the
9903 * node. This value will be ignored except that under some circumstances
9904 * <*flagp> is set based on it.
9906 * If <len is zero, the function assumes that the node is to contain only
9907 * the single character given by <code_point> and calculates what <len>
9908 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
9909 * additionally will populate the node's STRING with <code_point>, if <len>
9910 * is 0. In both cases <*flagp> is appropriately set
9912 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
9913 * folded (the latter only when the rules indicate it can match 'ss') */
9915 bool len_passed_in = cBOOL(len != 0);
9916 U8 character[UTF8_MAXBYTES_CASE+1];
9918 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
9920 if (! len_passed_in) {
9923 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
9926 uvchr_to_utf8( character, code_point);
9927 len = UTF8SKIP(character);
9931 || code_point != LATIN_SMALL_LETTER_SHARP_S
9932 || ASCII_FOLD_RESTRICTED
9933 || ! AT_LEAST_UNI_SEMANTICS)
9935 *character = (U8) code_point;
9940 *(character + 1) = 's';
9946 RExC_size += STR_SZ(len);
9949 RExC_emit += STR_SZ(len);
9950 STR_LEN(node) = len;
9951 if (! len_passed_in) {
9952 Copy((char *) character, STRING(node), len, char);
9957 if (len == 1 && UNI_IS_INVARIANT(code_point))
9962 - regatom - the lowest level
9964 Try to identify anything special at the start of the pattern. If there
9965 is, then handle it as required. This may involve generating a single regop,
9966 such as for an assertion; or it may involve recursing, such as to
9967 handle a () structure.
9969 If the string doesn't start with something special then we gobble up
9970 as much literal text as we can.
9972 Once we have been able to handle whatever type of thing started the
9973 sequence, we return.
9975 Note: we have to be careful with escapes, as they can be both literal
9976 and special, and in the case of \10 and friends, context determines which.
9978 A summary of the code structure is:
9980 switch (first_byte) {
9981 cases for each special:
9982 handle this special;
9986 cases for each unambiguous special:
9987 handle this special;
9989 cases for each ambigous special/literal:
9991 if (special) handle here
9993 default: // unambiguously literal:
9996 default: // is a literal char
9999 create EXACTish node for literal;
10000 while (more input and node isn't full) {
10001 switch (input_byte) {
10002 cases for each special;
10003 make sure parse pointer is set so that the next call to
10004 regatom will see this special first
10005 goto loopdone; // EXACTish node terminated by prev. char
10007 append char to EXACTISH node;
10009 get next input byte;
10013 return the generated node;
10015 Specifically there are two separate switches for handling
10016 escape sequences, with the one for handling literal escapes requiring
10017 a dummy entry for all of the special escapes that are actually handled
10022 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10025 regnode *ret = NULL;
10027 char *parse_start = RExC_parse;
10029 GET_RE_DEBUG_FLAGS_DECL;
10030 DEBUG_PARSE("atom");
10031 *flagp = WORST; /* Tentatively. */
10033 PERL_ARGS_ASSERT_REGATOM;
10036 switch ((U8)*RExC_parse) {
10038 RExC_seen_zerolen++;
10039 nextchar(pRExC_state);
10040 if (RExC_flags & RXf_PMf_MULTILINE)
10041 ret = reg_node(pRExC_state, MBOL);
10042 else if (RExC_flags & RXf_PMf_SINGLELINE)
10043 ret = reg_node(pRExC_state, SBOL);
10045 ret = reg_node(pRExC_state, BOL);
10046 Set_Node_Length(ret, 1); /* MJD */
10049 nextchar(pRExC_state);
10051 RExC_seen_zerolen++;
10052 if (RExC_flags & RXf_PMf_MULTILINE)
10053 ret = reg_node(pRExC_state, MEOL);
10054 else if (RExC_flags & RXf_PMf_SINGLELINE)
10055 ret = reg_node(pRExC_state, SEOL);
10057 ret = reg_node(pRExC_state, EOL);
10058 Set_Node_Length(ret, 1); /* MJD */
10061 nextchar(pRExC_state);
10062 if (RExC_flags & RXf_PMf_SINGLELINE)
10063 ret = reg_node(pRExC_state, SANY);
10065 ret = reg_node(pRExC_state, REG_ANY);
10066 *flagp |= HASWIDTH|SIMPLE;
10068 Set_Node_Length(ret, 1); /* MJD */
10072 char * const oregcomp_parse = ++RExC_parse;
10073 ret = regclass(pRExC_state, flagp,depth+1);
10074 if (*RExC_parse != ']') {
10075 RExC_parse = oregcomp_parse;
10076 vFAIL("Unmatched [");
10078 nextchar(pRExC_state);
10079 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10083 nextchar(pRExC_state);
10084 ret = reg(pRExC_state, 1, &flags,depth+1);
10086 if (flags & TRYAGAIN) {
10087 if (RExC_parse == RExC_end) {
10088 /* Make parent create an empty node if needed. */
10089 *flagp |= TRYAGAIN;
10096 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10100 if (flags & TRYAGAIN) {
10101 *flagp |= TRYAGAIN;
10104 vFAIL("Internal urp");
10105 /* Supposed to be caught earlier. */
10111 vFAIL("Quantifier follows nothing");
10116 This switch handles escape sequences that resolve to some kind
10117 of special regop and not to literal text. Escape sequnces that
10118 resolve to literal text are handled below in the switch marked
10121 Every entry in this switch *must* have a corresponding entry
10122 in the literal escape switch. However, the opposite is not
10123 required, as the default for this switch is to jump to the
10124 literal text handling code.
10126 switch ((U8)*++RExC_parse) {
10127 /* Special Escapes */
10129 RExC_seen_zerolen++;
10130 ret = reg_node(pRExC_state, SBOL);
10132 goto finish_meta_pat;
10134 ret = reg_node(pRExC_state, GPOS);
10135 RExC_seen |= REG_SEEN_GPOS;
10137 goto finish_meta_pat;
10139 RExC_seen_zerolen++;
10140 ret = reg_node(pRExC_state, KEEPS);
10142 /* XXX:dmq : disabling in-place substitution seems to
10143 * be necessary here to avoid cases of memory corruption, as
10144 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10146 RExC_seen |= REG_SEEN_LOOKBEHIND;
10147 goto finish_meta_pat;
10149 ret = reg_node(pRExC_state, SEOL);
10151 RExC_seen_zerolen++; /* Do not optimize RE away */
10152 goto finish_meta_pat;
10154 ret = reg_node(pRExC_state, EOS);
10156 RExC_seen_zerolen++; /* Do not optimize RE away */
10157 goto finish_meta_pat;
10159 ret = reg_node(pRExC_state, CANY);
10160 RExC_seen |= REG_SEEN_CANY;
10161 *flagp |= HASWIDTH|SIMPLE;
10162 goto finish_meta_pat;
10164 ret = reg_node(pRExC_state, CLUMP);
10165 *flagp |= HASWIDTH;
10166 goto finish_meta_pat;
10168 op = ALNUM + get_regex_charset(RExC_flags);
10169 if (op > ALNUMA) { /* /aa is same as /a */
10172 ret = reg_node(pRExC_state, op);
10173 *flagp |= HASWIDTH|SIMPLE;
10174 goto finish_meta_pat;
10176 op = NALNUM + get_regex_charset(RExC_flags);
10177 if (op > NALNUMA) { /* /aa is same as /a */
10180 ret = reg_node(pRExC_state, op);
10181 *flagp |= HASWIDTH|SIMPLE;
10182 goto finish_meta_pat;
10184 RExC_seen_zerolen++;
10185 RExC_seen |= REG_SEEN_LOOKBEHIND;
10186 op = BOUND + get_regex_charset(RExC_flags);
10187 if (op > BOUNDA) { /* /aa is same as /a */
10190 ret = reg_node(pRExC_state, op);
10191 FLAGS(ret) = get_regex_charset(RExC_flags);
10193 goto finish_meta_pat;
10195 RExC_seen_zerolen++;
10196 RExC_seen |= REG_SEEN_LOOKBEHIND;
10197 op = NBOUND + get_regex_charset(RExC_flags);
10198 if (op > NBOUNDA) { /* /aa is same as /a */
10201 ret = reg_node(pRExC_state, op);
10202 FLAGS(ret) = get_regex_charset(RExC_flags);
10204 goto finish_meta_pat;
10206 op = SPACE + get_regex_charset(RExC_flags);
10207 if (op > SPACEA) { /* /aa is same as /a */
10210 ret = reg_node(pRExC_state, op);
10211 *flagp |= HASWIDTH|SIMPLE;
10212 goto finish_meta_pat;
10214 op = NSPACE + get_regex_charset(RExC_flags);
10215 if (op > NSPACEA) { /* /aa is same as /a */
10218 ret = reg_node(pRExC_state, op);
10219 *flagp |= HASWIDTH|SIMPLE;
10220 goto finish_meta_pat;
10228 U8 offset = get_regex_charset(RExC_flags);
10229 if (offset == REGEX_UNICODE_CHARSET) {
10230 offset = REGEX_DEPENDS_CHARSET;
10232 else if (offset == REGEX_ASCII_MORE_RESTRICTED_CHARSET) {
10233 offset = REGEX_ASCII_RESTRICTED_CHARSET;
10237 ret = reg_node(pRExC_state, op);
10238 *flagp |= HASWIDTH|SIMPLE;
10239 goto finish_meta_pat;
10241 ret = reg_node(pRExC_state, LNBREAK);
10242 *flagp |= HASWIDTH|SIMPLE;
10243 goto finish_meta_pat;
10245 ret = reg_node(pRExC_state, HORIZWS);
10246 *flagp |= HASWIDTH|SIMPLE;
10247 goto finish_meta_pat;
10249 ret = reg_node(pRExC_state, NHORIZWS);
10250 *flagp |= HASWIDTH|SIMPLE;
10251 goto finish_meta_pat;
10253 ret = reg_node(pRExC_state, VERTWS);
10254 *flagp |= HASWIDTH|SIMPLE;
10255 goto finish_meta_pat;
10257 ret = reg_node(pRExC_state, NVERTWS);
10258 *flagp |= HASWIDTH|SIMPLE;
10260 nextchar(pRExC_state);
10261 Set_Node_Length(ret, 2); /* MJD */
10266 char* const oldregxend = RExC_end;
10268 char* parse_start = RExC_parse - 2;
10271 if (RExC_parse[1] == '{') {
10272 /* a lovely hack--pretend we saw [\pX] instead */
10273 RExC_end = strchr(RExC_parse, '}');
10275 const U8 c = (U8)*RExC_parse;
10277 RExC_end = oldregxend;
10278 vFAIL2("Missing right brace on \\%c{}", c);
10283 RExC_end = RExC_parse + 2;
10284 if (RExC_end > oldregxend)
10285 RExC_end = oldregxend;
10289 ret = regclass(pRExC_state, flagp,depth+1);
10291 RExC_end = oldregxend;
10294 Set_Node_Offset(ret, parse_start + 2);
10295 Set_Node_Cur_Length(ret);
10296 nextchar(pRExC_state);
10300 /* Handle \N and \N{NAME} with multiple code points here and not
10301 * below because it can be multicharacter. join_exact() will join
10302 * them up later on. Also this makes sure that things like
10303 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10304 * The options to the grok function call causes it to fail if the
10305 * sequence is just a single code point. We then go treat it as
10306 * just another character in the current EXACT node, and hence it
10307 * gets uniform treatment with all the other characters. The
10308 * special treatment for quantifiers is not needed for such single
10309 * character sequences */
10311 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE)) {
10316 case 'k': /* Handle \k<NAME> and \k'NAME' */
10319 char ch= RExC_parse[1];
10320 if (ch != '<' && ch != '\'' && ch != '{') {
10322 vFAIL2("Sequence %.2s... not terminated",parse_start);
10324 /* this pretty much dupes the code for (?P=...) in reg(), if
10325 you change this make sure you change that */
10326 char* name_start = (RExC_parse += 2);
10328 SV *sv_dat = reg_scan_name(pRExC_state,
10329 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10330 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10331 if (RExC_parse == name_start || *RExC_parse != ch)
10332 vFAIL2("Sequence %.3s... not terminated",parse_start);
10335 num = add_data( pRExC_state, 1, "S" );
10336 RExC_rxi->data->data[num]=(void*)sv_dat;
10337 SvREFCNT_inc_simple_void(sv_dat);
10341 ret = reganode(pRExC_state,
10344 : (ASCII_FOLD_RESTRICTED)
10346 : (AT_LEAST_UNI_SEMANTICS)
10352 *flagp |= HASWIDTH;
10354 /* override incorrect value set in reganode MJD */
10355 Set_Node_Offset(ret, parse_start+1);
10356 Set_Node_Cur_Length(ret); /* MJD */
10357 nextchar(pRExC_state);
10363 case '1': case '2': case '3': case '4':
10364 case '5': case '6': case '7': case '8': case '9':
10367 bool isg = *RExC_parse == 'g';
10372 if (*RExC_parse == '{') {
10376 if (*RExC_parse == '-') {
10380 if (hasbrace && !isDIGIT(*RExC_parse)) {
10381 if (isrel) RExC_parse--;
10383 goto parse_named_seq;
10385 num = atoi(RExC_parse);
10386 if (isg && num == 0)
10387 vFAIL("Reference to invalid group 0");
10389 num = RExC_npar - num;
10391 vFAIL("Reference to nonexistent or unclosed group");
10393 if (!isg && num > 9 && num >= RExC_npar)
10394 /* Probably a character specified in octal, e.g. \35 */
10397 char * const parse_start = RExC_parse - 1; /* MJD */
10398 while (isDIGIT(*RExC_parse))
10400 if (parse_start == RExC_parse - 1)
10401 vFAIL("Unterminated \\g... pattern");
10403 if (*RExC_parse != '}')
10404 vFAIL("Unterminated \\g{...} pattern");
10408 if (num > (I32)RExC_rx->nparens)
10409 vFAIL("Reference to nonexistent group");
10412 ret = reganode(pRExC_state,
10415 : (ASCII_FOLD_RESTRICTED)
10417 : (AT_LEAST_UNI_SEMANTICS)
10423 *flagp |= HASWIDTH;
10425 /* override incorrect value set in reganode MJD */
10426 Set_Node_Offset(ret, parse_start+1);
10427 Set_Node_Cur_Length(ret); /* MJD */
10429 nextchar(pRExC_state);
10434 if (RExC_parse >= RExC_end)
10435 FAIL("Trailing \\");
10438 /* Do not generate "unrecognized" warnings here, we fall
10439 back into the quick-grab loop below */
10446 if (RExC_flags & RXf_PMf_EXTENDED) {
10447 if ( reg_skipcomment( pRExC_state ) )
10454 parse_start = RExC_parse - 1;
10463 #define MAX_NODE_STRING_SIZE 127
10464 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10466 U8 upper_parse = MAX_NODE_STRING_SIZE;
10469 bool next_is_quantifier;
10470 char * oldp = NULL;
10473 node_type = compute_EXACTish(pRExC_state);
10474 ret = reg_node(pRExC_state, node_type);
10476 /* In pass1, folded, we use a temporary buffer instead of the
10477 * actual node, as the node doesn't exist yet */
10478 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10484 /* XXX The node can hold up to 255 bytes, yet this only goes to
10485 * 127. I (khw) do not know why. Keeping it somewhat less than
10486 * 255 allows us to not have to worry about overflow due to
10487 * converting to utf8 and fold expansion, but that value is
10488 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10489 * split up by this limit into a single one using the real max of
10490 * 255. Even at 127, this breaks under rare circumstances. If
10491 * folding, we do not want to split a node at a character that is a
10492 * non-final in a multi-char fold, as an input string could just
10493 * happen to want to match across the node boundary. The join
10494 * would solve that problem if the join actually happens. But a
10495 * series of more than two nodes in a row each of 127 would cause
10496 * the first join to succeed to get to 254, but then there wouldn't
10497 * be room for the next one, which could at be one of those split
10498 * multi-char folds. I don't know of any fool-proof solution. One
10499 * could back off to end with only a code point that isn't such a
10500 * non-final, but it is possible for there not to be any in the
10502 for (p = RExC_parse - 1;
10503 len < upper_parse && p < RExC_end;
10508 if (RExC_flags & RXf_PMf_EXTENDED)
10509 p = regwhite( pRExC_state, p );
10520 /* Literal Escapes Switch
10522 This switch is meant to handle escape sequences that
10523 resolve to a literal character.
10525 Every escape sequence that represents something
10526 else, like an assertion or a char class, is handled
10527 in the switch marked 'Special Escapes' above in this
10528 routine, but also has an entry here as anything that
10529 isn't explicitly mentioned here will be treated as
10530 an unescaped equivalent literal.
10533 switch ((U8)*++p) {
10534 /* These are all the special escapes. */
10535 case 'A': /* Start assertion */
10536 case 'b': case 'B': /* Word-boundary assertion*/
10537 case 'C': /* Single char !DANGEROUS! */
10538 case 'd': case 'D': /* digit class */
10539 case 'g': case 'G': /* generic-backref, pos assertion */
10540 case 'h': case 'H': /* HORIZWS */
10541 case 'k': case 'K': /* named backref, keep marker */
10542 case 'p': case 'P': /* Unicode property */
10543 case 'R': /* LNBREAK */
10544 case 's': case 'S': /* space class */
10545 case 'v': case 'V': /* VERTWS */
10546 case 'w': case 'W': /* word class */
10547 case 'X': /* eXtended Unicode "combining character sequence" */
10548 case 'z': case 'Z': /* End of line/string assertion */
10552 /* Anything after here is an escape that resolves to a
10553 literal. (Except digits, which may or may not)
10559 case 'N': /* Handle a single-code point named character. */
10560 /* The options cause it to fail if a multiple code
10561 * point sequence. Handle those in the switch() above
10563 RExC_parse = p + 1;
10564 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10565 flagp, depth, FALSE))
10567 RExC_parse = p = oldp;
10571 if (ender > 0xff) {
10588 ender = ASCII_TO_NATIVE('\033');
10592 ender = ASCII_TO_NATIVE('\007');
10597 STRLEN brace_len = len;
10599 const char* error_msg;
10601 bool valid = grok_bslash_o(p,
10608 RExC_parse = p; /* going to die anyway; point
10609 to exact spot of failure */
10616 if (PL_encoding && ender < 0x100) {
10617 goto recode_encoding;
10619 if (ender > 0xff) {
10626 STRLEN brace_len = len;
10628 const char* error_msg;
10630 bool valid = grok_bslash_x(p,
10637 RExC_parse = p; /* going to die anyway; point
10638 to exact spot of failure */
10644 if (PL_encoding && ender < 0x100) {
10645 goto recode_encoding;
10647 if (ender > 0xff) {
10654 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10656 case '0': case '1': case '2': case '3':case '4':
10657 case '5': case '6': case '7':
10659 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10661 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10663 ender = grok_oct(p, &numlen, &flags, NULL);
10664 if (ender > 0xff) {
10673 if (PL_encoding && ender < 0x100)
10674 goto recode_encoding;
10677 if (! RExC_override_recoding) {
10678 SV* enc = PL_encoding;
10679 ender = reg_recode((const char)(U8)ender, &enc);
10680 if (!enc && SIZE_ONLY)
10681 ckWARNreg(p, "Invalid escape in the specified encoding");
10687 FAIL("Trailing \\");
10690 if (!SIZE_ONLY&& isALNUMC(*p)) {
10691 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s passed through", p);
10693 goto normal_default;
10697 /* Currently we don't warn when the lbrace is at the start
10698 * of a construct. This catches it in the middle of a
10699 * literal string, or when its the first thing after
10700 * something like "\b" */
10702 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
10704 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
10709 if (UTF8_IS_START(*p) && UTF) {
10711 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10712 &numlen, UTF8_ALLOW_DEFAULT);
10718 } /* End of switch on the literal */
10720 /* Here, have looked at the literal character and <ender>
10721 * contains its ordinal, <p> points to the character after it
10724 if ( RExC_flags & RXf_PMf_EXTENDED)
10725 p = regwhite( pRExC_state, p );
10727 /* If the next thing is a quantifier, it applies to this
10728 * character only, which means that this character has to be in
10729 * its own node and can't just be appended to the string in an
10730 * existing node, so if there are already other characters in
10731 * the node, close the node with just them, and set up to do
10732 * this character again next time through, when it will be the
10733 * only thing in its new node */
10734 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10742 /* See comments for join_exact() as to why we fold
10743 * this non-UTF at compile time */
10744 || (node_type == EXACTFU
10745 && ender == LATIN_SMALL_LETTER_SHARP_S))
10749 /* Prime the casefolded buffer. Locale rules, which
10750 * apply only to code points < 256, aren't known until
10751 * execution, so for them, just output the original
10752 * character using utf8. If we start to fold non-UTF
10753 * patterns, be sure to update join_exact() */
10754 if (LOC && ender < 256) {
10755 if (UNI_IS_INVARIANT(ender)) {
10759 *s = UTF8_TWO_BYTE_HI(ender);
10760 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10765 ender = _to_uni_fold_flags(ender, (U8 *) s, &foldlen,
10767 | ((LOC) ? FOLD_FLAGS_LOCALE
10768 : (ASCII_FOLD_RESTRICTED)
10769 ? FOLD_FLAGS_NOMIX_ASCII
10775 /* The loop increments <len> each time, as all but this
10776 * path (and the one just below for UTF) through it add
10777 * a single byte to the EXACTish node. But this one
10778 * has changed len to be the correct final value, so
10779 * subtract one to cancel out the increment that
10781 len += foldlen - 1;
10788 const STRLEN unilen = reguni(pRExC_state, ender, s);
10794 /* See comment just above for - 1 */
10798 REGC((char)ender, s++);
10801 if (next_is_quantifier) {
10803 /* Here, the next input is a quantifier, and to get here,
10804 * the current character is the only one in the node.
10805 * Also, here <len> doesn't include the final byte for this
10811 } /* End of loop through literal characters */
10813 /* Here we have either exhausted the input or ran out of room in
10814 * the node. (If we encountered a character that can't be in the
10815 * node, transfer is made directly to <loopdone>, and so we
10816 * wouldn't have fallen off the end of the loop.) In the latter
10817 * case, we artificially have to split the node into two, because
10818 * we just don't have enough space to hold everything. This
10819 * creates a problem if the final character participates in a
10820 * multi-character fold in the non-final position, as a match that
10821 * should have occurred won't, due to the way nodes are matched,
10822 * and our artificial boundary. So back off until we find a non-
10823 * problematic character -- one that isn't at the beginning or
10824 * middle of such a fold. (Either it doesn't participate in any
10825 * folds, or appears only in the final position of all the folds it
10826 * does participate in.) A better solution with far fewer false
10827 * positives, and that would fill the nodes more completely, would
10828 * be to actually have available all the multi-character folds to
10829 * test against, and to back-off only far enough to be sure that
10830 * this node isn't ending with a partial one. <upper_parse> is set
10831 * further below (if we need to reparse the node) to include just
10832 * up through that final non-problematic character that this code
10833 * identifies, so when it is set to less than the full node, we can
10834 * skip the rest of this */
10835 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
10837 const STRLEN full_len = len;
10839 assert(len >= MAX_NODE_STRING_SIZE);
10841 /* Here, <s> points to the final byte of the final character.
10842 * Look backwards through the string until find a non-
10843 * problematic character */
10847 /* These two have no multi-char folds to non-UTF characters
10849 if (ASCII_FOLD_RESTRICTED || LOC) {
10853 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
10857 if (! PL_NonL1NonFinalFold) {
10858 PL_NonL1NonFinalFold = _new_invlist_C_array(
10859 NonL1_Perl_Non_Final_Folds_invlist);
10862 /* Point to the first byte of the final character */
10863 s = (char *) utf8_hop((U8 *) s, -1);
10865 while (s >= s0) { /* Search backwards until find
10866 non-problematic char */
10867 if (UTF8_IS_INVARIANT(*s)) {
10869 /* There are no ascii characters that participate
10870 * in multi-char folds under /aa. In EBCDIC, the
10871 * non-ascii invariants are all control characters,
10872 * so don't ever participate in any folds. */
10873 if (ASCII_FOLD_RESTRICTED
10874 || ! IS_NON_FINAL_FOLD(*s))
10879 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
10881 /* No Latin1 characters participate in multi-char
10882 * folds under /l */
10884 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
10890 else if (! _invlist_contains_cp(
10891 PL_NonL1NonFinalFold,
10892 valid_utf8_to_uvchr((U8 *) s, NULL)))
10897 /* Here, the current character is problematic in that
10898 * it does occur in the non-final position of some
10899 * fold, so try the character before it, but have to
10900 * special case the very first byte in the string, so
10901 * we don't read outside the string */
10902 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
10903 } /* End of loop backwards through the string */
10905 /* If there were only problematic characters in the string,
10906 * <s> will point to before s0, in which case the length
10907 * should be 0, otherwise include the length of the
10908 * non-problematic character just found */
10909 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
10912 /* Here, have found the final character, if any, that is
10913 * non-problematic as far as ending the node without splitting
10914 * it across a potential multi-char fold. <len> contains the
10915 * number of bytes in the node up-to and including that
10916 * character, or is 0 if there is no such character, meaning
10917 * the whole node contains only problematic characters. In
10918 * this case, give up and just take the node as-is. We can't
10924 /* Here, the node does contain some characters that aren't
10925 * problematic. If one such is the final character in the
10926 * node, we are done */
10927 if (len == full_len) {
10930 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
10932 /* If the final character is problematic, but the
10933 * penultimate is not, back-off that last character to
10934 * later start a new node with it */
10939 /* Here, the final non-problematic character is earlier
10940 * in the input than the penultimate character. What we do
10941 * is reparse from the beginning, going up only as far as
10942 * this final ok one, thus guaranteeing that the node ends
10943 * in an acceptable character. The reason we reparse is
10944 * that we know how far in the character is, but we don't
10945 * know how to correlate its position with the input parse.
10946 * An alternate implementation would be to build that
10947 * correlation as we go along during the original parse,
10948 * but that would entail extra work for every node, whereas
10949 * this code gets executed only when the string is too
10950 * large for the node, and the final two characters are
10951 * problematic, an infrequent occurrence. Yet another
10952 * possible strategy would be to save the tail of the
10953 * string, and the next time regatom is called, initialize
10954 * with that. The problem with this is that unless you
10955 * back off one more character, you won't be guaranteed
10956 * regatom will get called again, unless regbranch,
10957 * regpiece ... are also changed. If you do back off that
10958 * extra character, so that there is input guaranteed to
10959 * force calling regatom, you can't handle the case where
10960 * just the first character in the node is acceptable. I
10961 * (khw) decided to try this method which doesn't have that
10962 * pitfall; if performance issues are found, we can do a
10963 * combination of the current approach plus that one */
10969 } /* End of verifying node ends with an appropriate char */
10971 loopdone: /* Jumped to when encounters something that shouldn't be in
10974 /* I (khw) don't know if you can get here with zero length, but the
10975 * old code handled this situation by creating a zero-length EXACT
10976 * node. Might as well be NOTHING instead */
10981 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
10984 RExC_parse = p - 1;
10985 Set_Node_Cur_Length(ret); /* MJD */
10986 nextchar(pRExC_state);
10988 /* len is STRLEN which is unsigned, need to copy to signed */
10991 vFAIL("Internal disaster");
10994 } /* End of label 'defchar:' */
10996 } /* End of giant switch on input character */
11002 S_regwhite( RExC_state_t *pRExC_state, char *p )
11004 const char *e = RExC_end;
11006 PERL_ARGS_ASSERT_REGWHITE;
11011 else if (*p == '#') {
11014 if (*p++ == '\n') {
11020 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11028 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11029 Character classes ([:foo:]) can also be negated ([:^foo:]).
11030 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11031 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11032 but trigger failures because they are currently unimplemented. */
11034 #define POSIXCC_DONE(c) ((c) == ':')
11035 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11036 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11039 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value)
11042 I32 namedclass = OOB_NAMEDCLASS;
11044 PERL_ARGS_ASSERT_REGPPOSIXCC;
11046 if (value == '[' && RExC_parse + 1 < RExC_end &&
11047 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11048 POSIXCC(UCHARAT(RExC_parse))) {
11049 const char c = UCHARAT(RExC_parse);
11050 char* const s = RExC_parse++;
11052 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11054 if (RExC_parse == RExC_end)
11055 /* Grandfather lone [:, [=, [. */
11058 const char* const t = RExC_parse++; /* skip over the c */
11061 if (UCHARAT(RExC_parse) == ']') {
11062 const char *posixcc = s + 1;
11063 RExC_parse++; /* skip over the ending ] */
11066 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11067 const I32 skip = t - posixcc;
11069 /* Initially switch on the length of the name. */
11072 if (memEQ(posixcc, "word", 4)) /* this is not POSIX, this is the Perl \w */
11073 namedclass = ANYOF_ALNUM;
11076 /* Names all of length 5. */
11077 /* alnum alpha ascii blank cntrl digit graph lower
11078 print punct space upper */
11079 /* Offset 4 gives the best switch position. */
11080 switch (posixcc[4]) {
11082 if (memEQ(posixcc, "alph", 4)) /* alpha */
11083 namedclass = ANYOF_ALPHA;
11086 if (memEQ(posixcc, "spac", 4)) /* space */
11087 namedclass = ANYOF_PSXSPC;
11090 if (memEQ(posixcc, "grap", 4)) /* graph */
11091 namedclass = ANYOF_GRAPH;
11094 if (memEQ(posixcc, "asci", 4)) /* ascii */
11095 namedclass = ANYOF_ASCII;
11098 if (memEQ(posixcc, "blan", 4)) /* blank */
11099 namedclass = ANYOF_BLANK;
11102 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11103 namedclass = ANYOF_CNTRL;
11106 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11107 namedclass = ANYOF_ALNUMC;
11110 if (memEQ(posixcc, "lowe", 4)) /* lower */
11111 namedclass = ANYOF_LOWER;
11112 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11113 namedclass = ANYOF_UPPER;
11116 if (memEQ(posixcc, "digi", 4)) /* digit */
11117 namedclass = ANYOF_DIGIT;
11118 else if (memEQ(posixcc, "prin", 4)) /* print */
11119 namedclass = ANYOF_PRINT;
11120 else if (memEQ(posixcc, "punc", 4)) /* punct */
11121 namedclass = ANYOF_PUNCT;
11126 if (memEQ(posixcc, "xdigit", 6))
11127 namedclass = ANYOF_XDIGIT;
11131 if (namedclass == OOB_NAMEDCLASS)
11132 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11135 /* The #defines are structured so each complement is +1 to
11136 * the normal one */
11140 assert (posixcc[skip] == ':');
11141 assert (posixcc[skip+1] == ']');
11142 } else if (!SIZE_ONLY) {
11143 /* [[=foo=]] and [[.foo.]] are still future. */
11145 /* adjust RExC_parse so the warning shows after
11146 the class closes */
11147 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11149 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11152 /* Maternal grandfather:
11153 * "[:" ending in ":" but not in ":]" */
11163 S_checkposixcc(pTHX_ RExC_state_t *pRExC_state)
11167 PERL_ARGS_ASSERT_CHECKPOSIXCC;
11169 if (POSIXCC(UCHARAT(RExC_parse))) {
11170 const char *s = RExC_parse;
11171 const char c = *s++;
11173 while (isALNUM(*s))
11175 if (*s && c == *s && s[1] == ']') {
11177 "POSIX syntax [%c %c] belongs inside character classes",
11180 /* [[=foo=]] and [[.foo.]] are still future. */
11181 if (POSIXCC_NOTYET(c)) {
11182 /* adjust RExC_parse so the error shows after
11183 the class closes */
11184 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse++) != ']')
11186 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11192 /* Generate the code to add a full posix character <class> to the bracketed
11193 * character class given by <node>. (<node> is needed only under locale rules)
11194 * destlist is the inversion list for non-locale rules that this class is
11196 * sourcelist is the ASCII-range inversion list to add under /a rules
11197 * Xsourcelist is the full Unicode range list to use otherwise. */
11198 #define DO_POSIX(node, class, destlist, sourcelist, Xsourcelist) \
11200 SV* scratch_list = NULL; \
11202 /* Set this class in the node for runtime matching */ \
11203 ANYOF_CLASS_SET(node, class); \
11205 /* For above Latin1 code points, we use the full Unicode range */ \
11206 _invlist_intersection(PL_AboveLatin1, \
11209 /* And set the output to it, adding instead if there already is an \
11210 * output. Checking if <destlist> is NULL first saves an extra \
11211 * clone. Its reference count will be decremented at the next \
11212 * union, etc, or if this is the only instance, at the end of the \
11214 if (! destlist) { \
11215 destlist = scratch_list; \
11218 _invlist_union(destlist, scratch_list, &destlist); \
11219 SvREFCNT_dec(scratch_list); \
11223 /* For non-locale, just add it to any existing list */ \
11224 _invlist_union(destlist, \
11225 (AT_LEAST_ASCII_RESTRICTED) \
11231 /* Like DO_POSIX, but matches the complement of <sourcelist> and <Xsourcelist>.
11233 #define DO_N_POSIX(node, class, destlist, sourcelist, Xsourcelist) \
11235 SV* scratch_list = NULL; \
11236 ANYOF_CLASS_SET(node, class); \
11237 _invlist_subtract(PL_AboveLatin1, Xsourcelist, &scratch_list); \
11238 if (! destlist) { \
11239 destlist = scratch_list; \
11242 _invlist_union(destlist, scratch_list, &destlist); \
11243 SvREFCNT_dec(scratch_list); \
11247 _invlist_union_complement_2nd(destlist, \
11248 (AT_LEAST_ASCII_RESTRICTED) \
11252 /* Under /d, everything in the upper half of the Latin1 range \
11253 * matches this complement */ \
11254 if (DEPENDS_SEMANTICS) { \
11255 ANYOF_FLAGS(node) |= ANYOF_NON_UTF8_LATIN1_ALL; \
11259 /* Generate the code to add a posix character <class> to the bracketed
11260 * character class given by <node>. (<node> is needed only under locale rules)
11261 * destlist is the inversion list for non-locale rules that this class is
11263 * sourcelist is the ASCII-range inversion list to add under /a rules
11264 * l1_sourcelist is the Latin1 range list to use otherwise.
11265 * Xpropertyname is the name to add to <run_time_list> of the property to
11266 * specify the code points above Latin1 that will have to be
11267 * determined at run-time
11268 * run_time_list is a SV* that contains text names of properties that are to
11269 * be computed at run time. This concatenates <Xpropertyname>
11270 * to it, appropriately
11271 * This is essentially DO_POSIX, but we know only the Latin1 values at compile
11273 #define DO_POSIX_LATIN1_ONLY_KNOWN(node, class, destlist, sourcelist, \
11274 l1_sourcelist, Xpropertyname, run_time_list) \
11275 /* First, resolve whether to use the ASCII-only list or the L1 \
11277 DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(node, class, destlist, \
11278 ((AT_LEAST_ASCII_RESTRICTED) ? sourcelist : l1_sourcelist),\
11279 Xpropertyname, run_time_list)
11281 #define DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(node, class, destlist, sourcelist, \
11282 Xpropertyname, run_time_list) \
11283 /* If not /a matching, there are going to be code points we will have \
11284 * to defer to runtime to look-up */ \
11285 if (! AT_LEAST_ASCII_RESTRICTED) { \
11286 Perl_sv_catpvf(aTHX_ run_time_list, "+utf8::%s\n", Xpropertyname); \
11289 ANYOF_CLASS_SET(node, class); \
11292 _invlist_union(destlist, sourcelist, &destlist); \
11295 /* Like DO_POSIX_LATIN1_ONLY_KNOWN, but for the complement. A combination of
11296 * this and DO_N_POSIX. Sets <matches_above_unicode> only if it can; unchanged
11298 #define DO_N_POSIX_LATIN1_ONLY_KNOWN(node, class, destlist, sourcelist, \
11299 l1_sourcelist, Xpropertyname, run_time_list, matches_above_unicode) \
11300 if (AT_LEAST_ASCII_RESTRICTED) { \
11301 _invlist_union_complement_2nd(destlist, sourcelist, &destlist); \
11304 Perl_sv_catpvf(aTHX_ run_time_list, "!utf8::%s\n", Xpropertyname); \
11305 matches_above_unicode = TRUE; \
11307 ANYOF_CLASS_SET(node, namedclass); \
11310 SV* scratch_list = NULL; \
11311 _invlist_subtract(PL_Latin1, l1_sourcelist, &scratch_list); \
11312 if (! destlist) { \
11313 destlist = scratch_list; \
11316 _invlist_union(destlist, scratch_list, &destlist); \
11317 SvREFCNT_dec(scratch_list); \
11319 if (DEPENDS_SEMANTICS) { \
11320 ANYOF_FLAGS(node) |= ANYOF_NON_UTF8_LATIN1_ALL; \
11326 S_add_alternate(pTHX_ AV** alternate_ptr, U8* string, STRLEN len)
11328 /* Adds input 'string' with length 'len' to the ANYOF node's unicode
11329 * alternate list, pointed to by 'alternate_ptr'. This is an array of
11330 * the multi-character folds of characters in the node */
11333 PERL_ARGS_ASSERT_ADD_ALTERNATE;
11335 if (! *alternate_ptr) {
11336 *alternate_ptr = newAV();
11338 sv = newSVpvn_utf8((char*)string, len, TRUE);
11339 av_push(*alternate_ptr, sv);
11343 /* The names of properties whose definitions are not known at compile time are
11344 * stored in this SV, after a constant heading. So if the length has been
11345 * changed since initialization, then there is a run-time definition. */
11346 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11348 /* This converts the named class defined in regcomp.h to its equivalent class
11349 * number defined in handy.h. */
11350 #define namedclass_to_classnum(class) ((class) / 2)
11353 parse a class specification and produce either an ANYOF node that
11354 matches the pattern or perhaps will be optimized into an EXACTish node
11355 instead. The node contains a bit map for the first 256 characters, with the
11356 corresponding bit set if that character is in the list. For characters
11357 above 255, a range list is used */
11360 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11364 UV prevvalue = OOB_UNICODE;
11369 IV namedclass = OOB_NAMEDCLASS;
11370 char *rangebegin = NULL;
11371 bool need_class = 0;
11372 bool allow_full_fold = TRUE; /* Assume wants multi-char folding */
11374 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
11375 than just initialized. */
11376 SV* properties = NULL; /* Code points that match \p{} \P{} */
11377 SV* posixes = NULL; /* Code points that match classes like, [:word:],
11378 extended beyond the Latin1 range */
11379 UV element_count = 0; /* Number of distinct elements in the class.
11380 Optimizations may be possible if this is tiny */
11383 /* Unicode properties are stored in a swash; this holds the current one
11384 * being parsed. If this swash is the only above-latin1 component of the
11385 * character class, an optimization is to pass it directly on to the
11386 * execution engine. Otherwise, it is set to NULL to indicate that there
11387 * are other things in the class that have to be dealt with at execution
11389 SV* swash = NULL; /* Code points that match \p{} \P{} */
11391 /* Set if a component of this character class is user-defined; just passed
11392 * on to the engine */
11393 bool has_user_defined_property = FALSE;
11395 /* inversion list of code points this node matches only when the target
11396 * string is in UTF-8. (Because is under /d) */
11397 SV* depends_list = NULL;
11399 /* inversion list of code points this node matches. For much of the
11400 * function, it includes only those that match regardless of the utf8ness
11401 * of the target string */
11402 SV* cp_list = NULL;
11404 /* List of multi-character folds that are matched by this node */
11405 AV* unicode_alternate = NULL;
11407 /* In a range, counts how many 0-2 of the ends of it came from literals,
11408 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
11409 UV literal_endpoint = 0;
11411 bool invert = FALSE; /* Is this class to be complemented */
11413 /* Is there any thing like \W or [:^digit:] that matches above the legal
11414 * Unicode range? */
11415 bool runtime_posix_matches_above_Unicode = FALSE;
11417 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
11418 case we need to change the emitted regop to an EXACT. */
11419 const char * orig_parse = RExC_parse;
11420 const I32 orig_size = RExC_size;
11421 GET_RE_DEBUG_FLAGS_DECL;
11423 PERL_ARGS_ASSERT_REGCLASS;
11425 PERL_UNUSED_ARG(depth);
11428 DEBUG_PARSE("clas");
11430 /* Assume we are going to generate an ANYOF node. */
11431 ret = reganode(pRExC_state, ANYOF, 0);
11435 ANYOF_FLAGS(ret) = 0;
11438 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
11443 /* We have decided to not allow multi-char folds in inverted character
11444 * classes, due to the confusion that can happen, especially with
11445 * classes that are designed for a non-Unicode world: You have the
11446 * peculiar case that:
11447 "s s" =~ /^[^\xDF]+$/i => Y
11448 "ss" =~ /^[^\xDF]+$/i => N
11450 * See [perl #89750] */
11451 allow_full_fold = FALSE;
11455 RExC_size += ANYOF_SKIP;
11456 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
11459 RExC_emit += ANYOF_SKIP;
11461 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
11463 listsv = newSVpvs("# comment\n");
11464 initial_listsv_len = SvCUR(listsv);
11467 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
11469 if (!SIZE_ONLY && POSIXCC(nextvalue))
11470 checkposixcc(pRExC_state);
11472 /* allow 1st char to be ] (allowing it to be - is dealt with later) */
11473 if (UCHARAT(RExC_parse) == ']')
11474 goto charclassloop;
11477 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11481 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
11484 rangebegin = RExC_parse;
11488 value = utf8n_to_uvchr((U8*)RExC_parse,
11489 RExC_end - RExC_parse,
11490 &numlen, UTF8_ALLOW_DEFAULT);
11491 RExC_parse += numlen;
11494 value = UCHARAT(RExC_parse++);
11496 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
11497 if (value == '[' && POSIXCC(nextvalue))
11498 namedclass = regpposixcc(pRExC_state, value);
11499 else if (value == '\\') {
11501 value = utf8n_to_uvchr((U8*)RExC_parse,
11502 RExC_end - RExC_parse,
11503 &numlen, UTF8_ALLOW_DEFAULT);
11504 RExC_parse += numlen;
11507 value = UCHARAT(RExC_parse++);
11508 /* Some compilers cannot handle switching on 64-bit integer
11509 * values, therefore value cannot be an UV. Yes, this will
11510 * be a problem later if we want switch on Unicode.
11511 * A similar issue a little bit later when switching on
11512 * namedclass. --jhi */
11513 switch ((I32)value) {
11514 case 'w': namedclass = ANYOF_ALNUM; break;
11515 case 'W': namedclass = ANYOF_NALNUM; break;
11516 case 's': namedclass = ANYOF_SPACE; break;
11517 case 'S': namedclass = ANYOF_NSPACE; break;
11518 case 'd': namedclass = ANYOF_DIGIT; break;
11519 case 'D': namedclass = ANYOF_NDIGIT; break;
11520 case 'v': namedclass = ANYOF_VERTWS; break;
11521 case 'V': namedclass = ANYOF_NVERTWS; break;
11522 case 'h': namedclass = ANYOF_HORIZWS; break;
11523 case 'H': namedclass = ANYOF_NHORIZWS; break;
11524 case 'N': /* Handle \N{NAME} in class */
11526 /* We only pay attention to the first char of
11527 multichar strings being returned. I kinda wonder
11528 if this makes sense as it does change the behaviour
11529 from earlier versions, OTOH that behaviour was broken
11531 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
11532 TRUE /* => charclass */))
11543 /* This routine will handle any undefined properties */
11544 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
11546 if (RExC_parse >= RExC_end)
11547 vFAIL2("Empty \\%c{}", (U8)value);
11548 if (*RExC_parse == '{') {
11549 const U8 c = (U8)value;
11550 e = strchr(RExC_parse++, '}');
11552 vFAIL2("Missing right brace on \\%c{}", c);
11553 while (isSPACE(UCHARAT(RExC_parse)))
11555 if (e == RExC_parse)
11556 vFAIL2("Empty \\%c{}", c);
11557 n = e - RExC_parse;
11558 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
11569 if (UCHARAT(RExC_parse) == '^') {
11572 value = value == 'p' ? 'P' : 'p'; /* toggle */
11573 while (isSPACE(UCHARAT(RExC_parse))) {
11578 /* Try to get the definition of the property into
11579 * <invlist>. If /i is in effect, the effective property
11580 * will have its name be <__NAME_i>. The design is
11581 * discussed in commit
11582 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
11583 Newx(name, n + sizeof("_i__\n"), char);
11585 sprintf(name, "%s%.*s%s\n",
11586 (FOLD) ? "__" : "",
11592 /* Look up the property name, and get its swash and
11593 * inversion list, if the property is found */
11595 SvREFCNT_dec(swash);
11597 swash = _core_swash_init("utf8", name, &PL_sv_undef,
11600 NULL, /* No inversion list */
11603 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
11605 SvREFCNT_dec(swash);
11609 /* Here didn't find it. It could be a user-defined
11610 * property that will be available at run-time. Add it
11611 * to the list to look up then */
11612 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
11613 (value == 'p' ? '+' : '!'),
11615 has_user_defined_property = TRUE;
11617 /* We don't know yet, so have to assume that the
11618 * property could match something in the Latin1 range,
11619 * hence something that isn't utf8. Note that this
11620 * would cause things in <depends_list> to match
11621 * inappropriately, except that any \p{}, including
11622 * this one forces Unicode semantics, which means there
11623 * is <no depends_list> */
11624 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
11628 /* Here, did get the swash and its inversion list. If
11629 * the swash is from a user-defined property, then this
11630 * whole character class should be regarded as such */
11631 has_user_defined_property =
11633 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
11635 /* Invert if asking for the complement */
11636 if (value == 'P') {
11637 _invlist_union_complement_2nd(properties,
11641 /* The swash can't be used as-is, because we've
11642 * inverted things; delay removing it to here after
11643 * have copied its invlist above */
11644 SvREFCNT_dec(swash);
11648 _invlist_union(properties, invlist, &properties);
11653 RExC_parse = e + 1;
11654 namedclass = ANYOF_MAX; /* no official name, but it's named */
11656 /* \p means they want Unicode semantics */
11657 RExC_uni_semantics = 1;
11660 case 'n': value = '\n'; break;
11661 case 'r': value = '\r'; break;
11662 case 't': value = '\t'; break;
11663 case 'f': value = '\f'; break;
11664 case 'b': value = '\b'; break;
11665 case 'e': value = ASCII_TO_NATIVE('\033');break;
11666 case 'a': value = ASCII_TO_NATIVE('\007');break;
11668 RExC_parse--; /* function expects to be pointed at the 'o' */
11670 const char* error_msg;
11671 bool valid = grok_bslash_o(RExC_parse,
11676 RExC_parse += numlen;
11681 if (PL_encoding && value < 0x100) {
11682 goto recode_encoding;
11686 RExC_parse--; /* function expects to be pointed at the 'x' */
11688 const char* error_msg;
11689 bool valid = grok_bslash_x(RExC_parse,
11694 RExC_parse += numlen;
11699 if (PL_encoding && value < 0x100)
11700 goto recode_encoding;
11703 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
11705 case '0': case '1': case '2': case '3': case '4':
11706 case '5': case '6': case '7':
11708 /* Take 1-3 octal digits */
11709 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11711 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
11712 RExC_parse += numlen;
11713 if (PL_encoding && value < 0x100)
11714 goto recode_encoding;
11718 if (! RExC_override_recoding) {
11719 SV* enc = PL_encoding;
11720 value = reg_recode((const char)(U8)value, &enc);
11721 if (!enc && SIZE_ONLY)
11722 ckWARNreg(RExC_parse,
11723 "Invalid escape in the specified encoding");
11727 /* Allow \_ to not give an error */
11728 if (!SIZE_ONLY && isALNUM(value) && value != '_') {
11729 ckWARN2reg(RExC_parse,
11730 "Unrecognized escape \\%c in character class passed through",
11735 } /* end of \blah */
11738 literal_endpoint++;
11741 /* What matches in a locale is not known until runtime. This
11742 * includes what the Posix classes (like \w, [:space:]) match.
11743 * Room must be reserved (one time per class) to store such
11744 * classes, either if Perl is compiled so that locale nodes always
11745 * should have this space, or if there is such class info to be
11746 * stored. The space will contain a bit for each named class that
11747 * is to be matched against. This isn't needed for \p{} and
11748 * pseudo-classes, as they are not affected by locale, and hence
11749 * are dealt with separately */
11752 && (ANYOF_LOCALE == ANYOF_CLASS
11753 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
11757 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
11760 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
11761 ANYOF_CLASS_ZERO(ret);
11763 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
11766 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
11768 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
11769 * literal, as is the character that began the false range, i.e.
11770 * the 'a' in the examples */
11774 RExC_parse >= rangebegin ?
11775 RExC_parse - rangebegin : 0;
11776 ckWARN4reg(RExC_parse,
11777 "False [] range \"%*.*s\"",
11779 cp_list = add_cp_to_invlist(cp_list, '-');
11780 cp_list = add_cp_to_invlist(cp_list, prevvalue);
11783 range = 0; /* this was not a true range */
11784 element_count += 2; /* So counts for three values */
11788 switch ((I32)namedclass) {
11790 case ANYOF_ALNUMC: /* C's alnum, in contrast to \w */
11791 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11792 PL_PosixAlnum, PL_L1PosixAlnum, "XPosixAlnum", listsv);
11794 case ANYOF_NALNUMC:
11795 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11796 PL_PosixAlnum, PL_L1PosixAlnum, "XPosixAlnum", listsv,
11797 runtime_posix_matches_above_Unicode);
11800 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11801 PL_PosixAlpha, PL_L1PosixAlpha, "XPosixAlpha", listsv);
11804 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11805 PL_PosixAlpha, PL_L1PosixAlpha, "XPosixAlpha", listsv,
11806 runtime_posix_matches_above_Unicode);
11810 ANYOF_CLASS_SET(ret, namedclass);
11813 _invlist_union(posixes, PL_ASCII, &posixes);
11818 ANYOF_CLASS_SET(ret, namedclass);
11821 _invlist_union_complement_2nd(posixes,
11822 PL_ASCII, &posixes);
11823 if (DEPENDS_SEMANTICS) {
11824 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
11829 DO_POSIX(ret, namedclass, posixes,
11830 PL_PosixBlank, PL_XPosixBlank);
11833 DO_N_POSIX(ret, namedclass, posixes,
11834 PL_PosixBlank, PL_XPosixBlank);
11837 DO_POSIX(ret, namedclass, posixes,
11838 PL_PosixCntrl, PL_XPosixCntrl);
11841 DO_N_POSIX(ret, namedclass, posixes,
11842 PL_PosixCntrl, PL_XPosixCntrl);
11845 /* There are no digits in the Latin1 range outside of
11846 * ASCII, so call the macro that doesn't have to resolve
11848 DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(ret, namedclass, posixes,
11849 PL_PosixDigit, "XPosixDigit", listsv);
11852 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11853 PL_PosixDigit, PL_PosixDigit, "XPosixDigit", listsv,
11854 runtime_posix_matches_above_Unicode);
11857 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11858 PL_PosixGraph, PL_L1PosixGraph, "XPosixGraph", listsv);
11861 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11862 PL_PosixGraph, PL_L1PosixGraph, "XPosixGraph", listsv,
11863 runtime_posix_matches_above_Unicode);
11865 case ANYOF_HORIZWS:
11866 /* For these, we use the cp_list, as /d doesn't make a
11867 * difference in what these match. There would be problems
11868 * if these characters had folds other than themselves, as
11869 * cp_list is subject to folding. It turns out that \h
11870 * is just a synonym for XPosixBlank */
11871 _invlist_union(cp_list, PL_XPosixBlank, &cp_list);
11873 case ANYOF_NHORIZWS:
11874 _invlist_union_complement_2nd(cp_list,
11875 PL_XPosixBlank, &cp_list);
11879 { /* These require special handling, as they differ under
11880 folding, matching Cased there (which in the ASCII range
11881 is the same as Alpha */
11887 if (FOLD && ! LOC) {
11888 ascii_source = PL_PosixAlpha;
11889 l1_source = PL_L1Cased;
11893 ascii_source = PL_PosixLower;
11894 l1_source = PL_L1PosixLower;
11895 Xname = "XPosixLower";
11897 if (namedclass == ANYOF_LOWER) {
11898 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11899 ascii_source, l1_source, Xname, listsv);
11902 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass,
11903 posixes, ascii_source, l1_source, Xname, listsv,
11904 runtime_posix_matches_above_Unicode);
11909 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11910 PL_PosixPrint, PL_L1PosixPrint, "XPosixPrint", listsv);
11913 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11914 PL_PosixPrint, PL_L1PosixPrint, "XPosixPrint", listsv,
11915 runtime_posix_matches_above_Unicode);
11918 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11919 PL_PosixPunct, PL_L1PosixPunct, "XPosixPunct", listsv);
11922 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11923 PL_PosixPunct, PL_L1PosixPunct, "XPosixPunct", listsv,
11924 runtime_posix_matches_above_Unicode);
11927 DO_POSIX(ret, namedclass, posixes,
11928 PL_PosixSpace, PL_XPosixSpace);
11930 case ANYOF_NPSXSPC:
11931 DO_N_POSIX(ret, namedclass, posixes,
11932 PL_PosixSpace, PL_XPosixSpace);
11935 DO_POSIX(ret, namedclass, posixes,
11936 PL_PerlSpace, PL_XPerlSpace);
11939 DO_N_POSIX(ret, namedclass, posixes,
11940 PL_PerlSpace, PL_XPerlSpace);
11942 case ANYOF_UPPER: /* Same as LOWER, above */
11949 if (FOLD && ! LOC) {
11950 ascii_source = PL_PosixAlpha;
11951 l1_source = PL_L1Cased;
11955 ascii_source = PL_PosixUpper;
11956 l1_source = PL_L1PosixUpper;
11957 Xname = "XPosixUpper";
11959 if (namedclass == ANYOF_UPPER) {
11960 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11961 ascii_source, l1_source, Xname, listsv);
11964 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass,
11965 posixes, ascii_source, l1_source, Xname, listsv,
11966 runtime_posix_matches_above_Unicode);
11970 case ANYOF_ALNUM: /* Really is 'Word' */
11971 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11972 PL_PosixWord, PL_L1PosixWord, "XPosixWord", listsv);
11975 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11976 PL_PosixWord, PL_L1PosixWord, "XPosixWord", listsv,
11977 runtime_posix_matches_above_Unicode);
11980 /* For these, we use the cp_list, as /d doesn't make a
11981 * difference in what these match. There would be problems
11982 * if these characters had folds other than themselves, as
11983 * cp_list is subject to folding */
11984 _invlist_union(cp_list, PL_VertSpace, &cp_list);
11986 case ANYOF_NVERTWS:
11987 _invlist_union_complement_2nd(cp_list,
11988 PL_VertSpace, &cp_list);
11991 DO_POSIX(ret, namedclass, posixes,
11992 PL_PosixXDigit, PL_XPosixXDigit);
11994 case ANYOF_NXDIGIT:
11995 DO_N_POSIX(ret, namedclass, posixes,
11996 PL_PosixXDigit, PL_XPosixXDigit);
11999 /* this is to handle \p and \P */
12002 vFAIL("Invalid [::] class");
12006 continue; /* Go get next character */
12008 } /* end of namedclass \blah */
12011 if (prevvalue > value) /* b-a */ {
12012 const int w = RExC_parse - rangebegin;
12013 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12014 range = 0; /* not a valid range */
12018 prevvalue = value; /* save the beginning of the potential range */
12019 if (RExC_parse+1 < RExC_end
12020 && *RExC_parse == '-'
12021 && RExC_parse[1] != ']')
12025 /* a bad range like \w-, [:word:]- ? */
12026 if (namedclass > OOB_NAMEDCLASS) {
12027 if (ckWARN(WARN_REGEXP)) {
12029 RExC_parse >= rangebegin ?
12030 RExC_parse - rangebegin : 0;
12032 "False [] range \"%*.*s\"",
12036 cp_list = add_cp_to_invlist(cp_list, '-');
12040 range = 1; /* yeah, it's a range! */
12041 continue; /* but do it the next time */
12045 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12048 /* non-Latin1 code point implies unicode semantics. Must be set in
12049 * pass1 so is there for the whole of pass 2 */
12051 RExC_uni_semantics = 1;
12054 /* Ready to process either the single value, or the completed range */
12057 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12059 UV* this_range = _new_invlist(1);
12060 _append_range_to_invlist(this_range, prevvalue, value);
12062 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12063 * If this range was specified using something like 'i-j', we want
12064 * to include only the 'i' and the 'j', and not anything in
12065 * between, so exclude non-ASCII, non-alphabetics from it.
12066 * However, if the range was specified with something like
12067 * [\x89-\x91] or [\x89-j], all code points within it should be
12068 * included. literal_endpoint==2 means both ends of the range used
12069 * a literal character, not \x{foo} */
12070 if (literal_endpoint == 2
12071 && (prevvalue >= 'a' && value <= 'z')
12072 || (prevvalue >= 'A' && value <= 'Z'))
12074 _invlist_intersection(this_range, PL_ASCII, &this_range, );
12075 _invlist_intersection(this_range, PL_Alpha, &this_range, );
12077 _invlist_union(cp_list, this_range, &cp_list);
12078 literal_endpoint = 0;
12082 range = 0; /* this range (if it was one) is done now */
12083 } /* End of loop through all the text within the brackets */
12085 /* If the character class contains only a single element, it may be
12086 * optimizable into another node type which is smaller and runs faster.
12087 * Check if this is the case for this class */
12088 if (element_count == 1) {
12092 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
12093 [:digit:] or \p{foo} */
12095 /* Certain named classes have equivalents that can appear outside a
12096 * character class, e.g. \w, \H. We use these instead of a
12097 * character class. */
12098 switch ((I32)namedclass) {
12101 /* The first group is for node types that depend on the charset
12102 * modifier to the regex. We first calculate the base node
12103 * type, and if it should be inverted */
12110 goto join_charset_classes;
12117 goto join_charset_classes;
12125 join_charset_classes:
12127 /* Now that we have the base node type, we take advantage
12128 * of the enum ordering of the charset modifiers to get the
12129 * exact node type, For example the base SPACE also has
12130 * SPACEL, SPACEU, and SPACEA */
12132 offset = get_regex_charset(RExC_flags);
12134 /* /aa is the same as /a for these */
12135 if (offset == REGEX_ASCII_MORE_RESTRICTED_CHARSET) {
12136 offset = REGEX_ASCII_RESTRICTED_CHARSET;
12138 else if (op == DIGIT && offset == REGEX_UNICODE_CHARSET) {
12139 offset = REGEX_DEPENDS_CHARSET; /* There is no DIGITU */
12144 /* The number of varieties of each of these is the same,
12145 * hence, so is the delta between the normal and
12146 * complemented nodes */
12148 op += NALNUM - ALNUM;
12150 *flagp |= HASWIDTH|SIMPLE;
12153 /* The second group doesn't depend of the charset modifiers.
12154 * We just have normal and complemented */
12155 case ANYOF_NHORIZWS:
12158 case ANYOF_HORIZWS:
12160 op = (invert) ? NHORIZWS : HORIZWS;
12161 *flagp |= HASWIDTH|SIMPLE;
12164 case ANYOF_NVERTWS:
12168 op = (invert) ? NVERTWS : VERTWS;
12169 *flagp |= HASWIDTH|SIMPLE;
12179 if (AT_LEAST_UNI_SEMANTICS && ! AT_LEAST_ASCII_RESTRICTED) {
12184 /* A generic posix class. All the /a ones can be handled
12185 * by the POSIXA opcode. And all are closed under folding
12186 * in the ASCII range, so FOLD doesn't matter */
12187 if (AT_LEAST_ASCII_RESTRICTED
12188 || (! LOC && namedclass == ANYOF_ASCII))
12190 /* The odd numbered ones are the complements of the
12191 * next-lower even number one */
12192 if (namedclass % 2 == 1) {
12196 arg = namedclass_to_classnum(namedclass);
12197 op = (invert) ? NPOSIXA : POSIXA;
12202 else if (value == prevvalue) {
12204 /* Here, the class consists of just a single code point */
12207 if (! LOC && value == '\n') {
12208 op = REG_ANY; /* Optimize [^\n] */
12209 *flagp |= HASWIDTH|SIMPLE;
12213 else if (value < 256 || UTF) {
12215 /* Optimize a single value into an EXACTish node, but not if it
12216 * would require converting the pattern to UTF-8. */
12217 op = compute_EXACTish(pRExC_state);
12219 } /* Otherwise is a range */
12220 else if (! LOC) { /* locale could vary these */
12221 if (prevvalue == '0') {
12222 if (value == '9') {
12223 op = (invert) ? NDIGITA : DIGITA;
12224 *flagp |= HASWIDTH|SIMPLE;
12229 /* Here, we have changed <op> away from its initial value iff we found
12230 * an optimization */
12233 /* Throw away this ANYOF regnode, and emit the calculated one,
12234 * which should correspond to the beginning, not current, state of
12236 const char * cur_parse = RExC_parse;
12237 RExC_parse = (char *)orig_parse;
12241 /* To get locale nodes to not use the full ANYOF size would
12242 * require moving the code above that writes the portions
12243 * of it that aren't in other nodes to after this point.
12244 * e.g. ANYOF_CLASS_SET */
12245 RExC_size = orig_size;
12249 RExC_emit = (regnode *)orig_emit;
12252 ret = reg_node(pRExC_state, op);
12254 if (PL_regkind[op] == POSIXD) {
12258 *flagp |= HASWIDTH|SIMPLE;
12260 else if (PL_regkind[op] == EXACT) {
12261 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
12264 RExC_parse = (char *) cur_parse;
12266 SvREFCNT_dec(listsv);
12273 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
12275 /* If folding, we calculate all characters that could fold to or from the
12276 * ones already on the list */
12277 if (FOLD && cp_list) {
12278 UV start, end; /* End points of code point ranges */
12280 SV* fold_intersection = NULL;
12282 /* In the Latin1 range, the characters that can be folded-to or -from
12283 * are precisely the alphabetic characters. If the highest code point
12284 * is within Latin1, we can use the compiled-in list, and not have to
12285 * go out to disk. */
12286 if (invlist_highest(cp_list) < 256) {
12287 _invlist_intersection(PL_L1PosixAlpha, cp_list, &fold_intersection);
12291 /* Here, there are non-Latin1 code points, so we will have to go
12292 * fetch the list of all the characters that participate in folds
12294 if (! PL_utf8_foldable) {
12295 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
12296 &PL_sv_undef, 1, 0);
12297 PL_utf8_foldable = _get_swash_invlist(swash);
12298 SvREFCNT_dec(swash);
12301 /* This is a hash that for a particular fold gives all characters
12302 * that are involved in it */
12303 if (! PL_utf8_foldclosures) {
12305 /* If we were unable to find any folds, then we likely won't be
12306 * able to find the closures. So just create an empty list.
12307 * Folding will effectively be restricted to the non-Unicode
12308 * rules hard-coded into Perl. (This case happens legitimately
12309 * during compilation of Perl itself before the Unicode tables
12310 * are generated) */
12311 if (_invlist_len(PL_utf8_foldable) == 0) {
12312 PL_utf8_foldclosures = newHV();
12315 /* If the folds haven't been read in, call a fold function
12317 if (! PL_utf8_tofold) {
12318 U8 dummy[UTF8_MAXBYTES+1];
12321 /* This string is just a short named one above \xff */
12322 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, &dummy_len);
12323 assert(PL_utf8_tofold); /* Verify that worked */
12325 PL_utf8_foldclosures =
12326 _swash_inversion_hash(PL_utf8_tofold);
12330 /* Only the characters in this class that participate in folds need
12331 * be checked. Get the intersection of this class and all the
12332 * possible characters that are foldable. This can quickly narrow
12333 * down a large class */
12334 _invlist_intersection(PL_utf8_foldable, cp_list,
12335 &fold_intersection);
12338 /* Now look at the foldable characters in this class individually */
12339 invlist_iterinit(fold_intersection);
12340 while (invlist_iternext(fold_intersection, &start, &end)) {
12343 /* Locale folding for Latin1 characters is deferred until runtime */
12344 if (LOC && start < 256) {
12348 /* Look at every character in the range */
12349 for (j = start; j <= end; j++) {
12351 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
12357 /* We have the latin1 folding rules hard-coded here so that
12358 * an innocent-looking character class, like /[ks]/i won't
12359 * have to go out to disk to find the possible matches.
12360 * XXX It would be better to generate these via regen, in
12361 * case a new version of the Unicode standard adds new
12362 * mappings, though that is not really likely, and may be
12363 * caught by the default: case of the switch below. */
12365 if (PL_fold_latin1[j] != j) {
12367 /* ASCII is always matched; non-ASCII is matched only
12368 * under Unicode rules */
12369 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
12371 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
12375 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
12379 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
12380 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
12382 /* Certain Latin1 characters have matches outside
12383 * Latin1, or are multi-character. To get here, 'j' is
12384 * one of those characters. None of these matches is
12385 * valid for ASCII characters under /aa, which is why
12386 * the 'if' just above excludes those. The matches
12387 * fall into three categories:
12388 * 1) They are singly folded-to or -from an above 255
12389 * character, e.g., LATIN SMALL LETTER Y WITH
12390 * DIAERESIS and LATIN CAPITAL LETTER Y WITH
12392 * 2) They are part of a multi-char fold with another
12393 * latin1 character; only LATIN SMALL LETTER
12394 * SHARP S => "ss" fits this;
12395 * 3) They are part of a multi-char fold with a
12396 * character outside of Latin1, such as various
12398 * We aren't dealing fully with multi-char folds, except
12399 * we do deal with the pattern containing a character
12400 * that has a multi-char fold (not so much the inverse).
12401 * For types 1) and 3), the matches only happen when the
12402 * target string is utf8; that's not true for 2), and we
12403 * set a flag for it.
12405 * The code below adds the single fold closures for 'j'
12406 * to the inversion list. */
12411 add_cp_to_invlist(cp_list, KELVIN_SIGN);
12415 cp_list = add_cp_to_invlist(cp_list,
12416 LATIN_SMALL_LETTER_LONG_S);
12419 cp_list = add_cp_to_invlist(cp_list,
12420 GREEK_CAPITAL_LETTER_MU);
12421 cp_list = add_cp_to_invlist(cp_list,
12422 GREEK_SMALL_LETTER_MU);
12424 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
12425 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
12427 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
12429 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
12430 cp_list = add_cp_to_invlist(cp_list,
12431 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
12433 case LATIN_SMALL_LETTER_SHARP_S:
12434 cp_list = add_cp_to_invlist(cp_list,
12435 LATIN_CAPITAL_LETTER_SHARP_S);
12437 /* Under /a, /d, and /u, this can match the two
12439 if (! ASCII_FOLD_RESTRICTED) {
12440 add_alternate(&unicode_alternate,
12443 /* And under /u or /a, it can match even if
12444 * the target is not utf8 */
12445 if (AT_LEAST_UNI_SEMANTICS) {
12446 ANYOF_FLAGS(ret) |=
12447 ANYOF_NONBITMAP_NON_UTF8;
12451 case 'F': case 'f':
12452 case 'I': case 'i':
12453 case 'L': case 'l':
12454 case 'T': case 't':
12455 case 'A': case 'a':
12456 case 'H': case 'h':
12457 case 'J': case 'j':
12458 case 'N': case 'n':
12459 case 'W': case 'w':
12460 case 'Y': case 'y':
12461 /* These all are targets of multi-character
12462 * folds from code points that require UTF8 to
12463 * express, so they can't match unless the
12464 * target string is in UTF-8, so no action here
12465 * is necessary, as regexec.c properly handles
12466 * the general case for UTF-8 matching */
12469 /* Use deprecated warning to increase the
12470 * chances of this being output */
12471 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
12478 /* Here is an above Latin1 character. We don't have the rules
12479 * hard-coded for it. First, get its fold */
12480 f = _to_uni_fold_flags(j, foldbuf, &foldlen,
12481 ((allow_full_fold) ? FOLD_FLAGS_FULL : 0)
12483 ? FOLD_FLAGS_LOCALE
12484 : (ASCII_FOLD_RESTRICTED)
12485 ? FOLD_FLAGS_NOMIX_ASCII
12488 if (foldlen > (STRLEN)UNISKIP(f)) {
12490 /* Any multicharacter foldings (disallowed in lookbehind
12491 * patterns) require the following transform: [ABCDEF] ->
12492 * (?:[ABCabcDEFd]|pq|rst) where E folds into "pq" and F
12493 * folds into "rst", all other characters fold to single
12494 * characters. We save away these multicharacter foldings,
12495 * to be later saved as part of the additional "s" data. */
12496 if (! RExC_in_lookbehind) {
12498 U8* e = foldbuf + foldlen;
12500 /* If any of the folded characters of this are in the
12501 * Latin1 range, tell the regex engine that this can
12502 * match a non-utf8 target string. */
12504 if (UTF8_IS_INVARIANT(*loc)
12505 || UTF8_IS_DOWNGRADEABLE_START(*loc))
12508 |= ANYOF_NONBITMAP_NON_UTF8;
12511 loc += UTF8SKIP(loc);
12514 add_alternate(&unicode_alternate, foldbuf, foldlen);
12518 /* Single character fold of above Latin1. Add everything
12519 * in its fold closure to the list that this node should
12523 /* The fold closures data structure is a hash with the keys
12524 * being every character that is folded to, like 'k', and
12525 * the values each an array of everything that folds to its
12526 * key. e.g. [ 'k', 'K', KELVIN_SIGN ] */
12527 if ((listp = hv_fetch(PL_utf8_foldclosures,
12528 (char *) foldbuf, foldlen, FALSE)))
12530 AV* list = (AV*) *listp;
12532 for (k = 0; k <= av_len(list); k++) {
12533 SV** c_p = av_fetch(list, k, FALSE);
12536 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
12540 /* /aa doesn't allow folds between ASCII and non-;
12541 * /l doesn't allow them between above and below
12543 if ((ASCII_FOLD_RESTRICTED
12544 && (isASCII(c) != isASCII(j)))
12545 || (LOC && ((c < 256) != (j < 256))))
12550 /* Folds involving non-ascii Latin1 characters
12551 * under /d are added to a separate list */
12552 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
12554 cp_list = add_cp_to_invlist(cp_list, c);
12557 depends_list = add_cp_to_invlist(depends_list, c);
12564 SvREFCNT_dec(fold_intersection);
12567 /* And combine the result (if any) with any inversion list from posix
12568 * classes. The lists are kept separate up to now because we don't want to
12569 * fold the classes (folding of those is automatically handled by the swash
12570 * fetching code) */
12572 if (! DEPENDS_SEMANTICS) {
12574 _invlist_union(cp_list, posixes, &cp_list);
12575 SvREFCNT_dec(posixes);
12582 /* Under /d, we put into a separate list the Latin1 things that
12583 * match only when the target string is utf8 */
12584 SV* nonascii_but_latin1_properties = NULL;
12585 _invlist_intersection(posixes, PL_Latin1,
12586 &nonascii_but_latin1_properties);
12587 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
12588 &nonascii_but_latin1_properties);
12589 _invlist_subtract(posixes, nonascii_but_latin1_properties,
12592 _invlist_union(cp_list, posixes, &cp_list);
12593 SvREFCNT_dec(posixes);
12599 if (depends_list) {
12600 _invlist_union(depends_list, nonascii_but_latin1_properties,
12602 SvREFCNT_dec(nonascii_but_latin1_properties);
12605 depends_list = nonascii_but_latin1_properties;
12610 /* And combine the result (if any) with any inversion list from properties.
12611 * The lists are kept separate up to now so that we can distinguish the two
12612 * in regards to matching above-Unicode. A run-time warning is generated
12613 * if a Unicode property is matched against a non-Unicode code point. But,
12614 * we allow user-defined properties to match anything, without any warning,
12615 * and we also suppress the warning if there is a portion of the character
12616 * class that isn't a Unicode property, and which matches above Unicode, \W
12617 * or [\x{110000}] for example.
12618 * (Note that in this case, unlike the Posix one above, there is no
12619 * <depends_list>, because having a Unicode property forces Unicode
12622 bool warn_super = ! has_user_defined_property;
12625 /* If it matters to the final outcome, see if a non-property
12626 * component of the class matches above Unicode. If so, the
12627 * warning gets suppressed. This is true even if just a single
12628 * such code point is specified, as though not strictly correct if
12629 * another such code point is matched against, the fact that they
12630 * are using above-Unicode code points indicates they should know
12631 * the issues involved */
12633 bool non_prop_matches_above_Unicode =
12634 runtime_posix_matches_above_Unicode
12635 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
12637 non_prop_matches_above_Unicode =
12638 ! non_prop_matches_above_Unicode;
12640 warn_super = ! non_prop_matches_above_Unicode;
12643 _invlist_union(properties, cp_list, &cp_list);
12644 SvREFCNT_dec(properties);
12647 cp_list = properties;
12651 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
12655 /* Here, we have calculated what code points should be in the character
12658 * Now we can see about various optimizations. Fold calculation (which we
12659 * did above) needs to take place before inversion. Otherwise /[^k]/i
12660 * would invert to include K, which under /i would match k, which it
12661 * shouldn't. Therefore we can't invert folded locale now, as it won't be
12662 * folded until runtime */
12664 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
12665 * at compile time. Besides not inverting folded locale now, we can't invert
12666 * if there are things such as \w, which aren't known until runtime */
12668 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
12670 && ! unicode_alternate
12671 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12673 _invlist_invert(cp_list);
12675 /* Any swash can't be used as-is, because we've inverted things */
12677 SvREFCNT_dec(swash);
12681 /* Clear the invert flag since have just done it here */
12685 /* If we didn't do folding, it's because some information isn't available
12686 * until runtime; set the run-time fold flag for these. (We don't have to
12687 * worry about properties folding, as that is taken care of by the swash
12689 if (FOLD && (LOC || unicode_alternate))
12691 ANYOF_FLAGS(ret) |= ANYOF_LOC_NONBITMAP_FOLD;
12694 /* Some character classes are equivalent to other nodes. Such nodes take
12695 * up less room and generally fewer operations to execute than ANYOF nodes.
12696 * Above, we checked for and optimized into some such equivalents for
12697 * certain common classes that are easy to test. Getting to this point in
12698 * the code means that the class didn't get optimized there. Since this
12699 * code is only executed in Pass 2, it is too late to save space--it has
12700 * been allocated in Pass 1, and currently isn't given back. But turning
12701 * things into an EXACTish node can allow the optimizer to join it to any
12702 * adjacent such nodes. And if the class is equivalent to things like /./,
12703 * expensive run-time swashes can be avoided. Now that we have more
12704 * complete information, we can find things necessarily missed by the
12705 * earlier code. I (khw) am not sure how much to look for here. It would
12706 * be easy, but perhaps too slow, to check any candidates against all the
12707 * node types they could possibly match using _invlistEQ(). */
12710 && ! unicode_alternate
12713 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
12714 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12717 U8 op = END; /* The optimzation node-type */
12718 const char * cur_parse= RExC_parse;
12720 invlist_iterinit(cp_list);
12721 if (! invlist_iternext(cp_list, &start, &end)) {
12723 /* Here, the list is empty. This happens, for example, when a
12724 * Unicode property is the only thing in the character class, and
12725 * it doesn't match anything. (perluniprops.pod notes such
12728 *flagp |= HASWIDTH|SIMPLE;
12730 else if (start == end) { /* The range is a single code point */
12731 if (! invlist_iternext(cp_list, &start, &end)
12733 /* Don't do this optimization if it would require changing
12734 * the pattern to UTF-8 */
12735 && (start < 256 || UTF))
12737 /* Here, the list contains a single code point. Can optimize
12738 * into an EXACT node */
12747 /* A locale node under folding with one code point can be
12748 * an EXACTFL, as its fold won't be calculated until
12754 /* Here, we are generally folding, but there is only one
12755 * code point to match. If we have to, we use an EXACT
12756 * node, but it would be better for joining with adjacent
12757 * nodes in the optimization pass if we used the same
12758 * EXACTFish node that any such are likely to be. We can
12759 * do this iff the code point doesn't participate in any
12760 * folds. For example, an EXACTF of a colon is the same as
12761 * an EXACT one, since nothing folds to or from a colon.
12762 * In the Latin1 range, being an alpha means that the
12763 * character participates in a fold (except for the
12764 * feminine and masculine ordinals, which I (khw) don't
12765 * think are worrying about optimizing for). */
12767 if (isALPHA_L1(value)) {
12772 if (! PL_utf8_foldable) {
12773 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
12774 &PL_sv_undef, 1, 0);
12775 PL_utf8_foldable = _get_swash_invlist(swash);
12776 SvREFCNT_dec(swash);
12778 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
12783 /* If we haven't found the node type, above, it means we
12784 * can use the prevailing one */
12786 op = compute_EXACTish(pRExC_state);
12791 else if (start == 0) {
12792 if (end == UV_MAX) {
12794 *flagp |= HASWIDTH|SIMPLE;
12797 else if (end == '\n' - 1
12798 && invlist_iternext(cp_list, &start, &end)
12799 && start == '\n' + 1 && end == UV_MAX)
12802 *flagp |= HASWIDTH|SIMPLE;
12808 RExC_parse = (char *)orig_parse;
12809 RExC_emit = (regnode *)orig_emit;
12811 ret = reg_node(pRExC_state, op);
12813 RExC_parse = (char *)cur_parse;
12815 if (PL_regkind[op] == EXACT) {
12816 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
12819 SvREFCNT_dec(listsv);
12824 /* Here, <cp_list> contains all the code points we can determine at
12825 * compile time that match under all conditions. Go through it, and
12826 * for things that belong in the bitmap, put them there, and delete from
12827 * <cp_list>. While we are at it, see if everything above 255 is in the
12828 * list, and if so, set a flag to speed up execution */
12829 ANYOF_BITMAP_ZERO(ret);
12832 /* This gets set if we actually need to modify things */
12833 bool change_invlist = FALSE;
12837 /* Start looking through <cp_list> */
12838 invlist_iterinit(cp_list);
12839 while (invlist_iternext(cp_list, &start, &end)) {
12843 if (end == UV_MAX && start <= 256) {
12844 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
12847 /* Quit if are above what we should change */
12852 change_invlist = TRUE;
12854 /* Set all the bits in the range, up to the max that we are doing */
12855 high = (end < 255) ? end : 255;
12856 for (i = start; i <= (int) high; i++) {
12857 if (! ANYOF_BITMAP_TEST(ret, i)) {
12858 ANYOF_BITMAP_SET(ret, i);
12865 /* Done with loop; remove any code points that are in the bitmap from
12867 if (change_invlist) {
12868 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
12871 /* If have completely emptied it, remove it completely */
12872 if (_invlist_len(cp_list) == 0) {
12873 SvREFCNT_dec(cp_list);
12879 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
12882 /* Here, the bitmap has been populated with all the Latin1 code points that
12883 * always match. Can now add to the overall list those that match only
12884 * when the target string is UTF-8 (<depends_list>). */
12885 if (depends_list) {
12887 _invlist_union(cp_list, depends_list, &cp_list);
12888 SvREFCNT_dec(depends_list);
12891 cp_list = depends_list;
12895 /* If there is a swash and more than one element, we can't use the swash in
12896 * the optimization below. */
12897 if (swash && element_count > 1) {
12898 SvREFCNT_dec(swash);
12903 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
12904 && ! unicode_alternate)
12906 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
12907 SvREFCNT_dec(listsv);
12908 SvREFCNT_dec(unicode_alternate);
12911 /* av[0] stores the character class description in its textual form:
12912 * used later (regexec.c:Perl_regclass_swash()) to initialize the
12913 * appropriate swash, and is also useful for dumping the regnode.
12914 * av[1] if NULL, is a placeholder to later contain the swash computed
12915 * from av[0]. But if no further computation need be done, the
12916 * swash is stored there now.
12917 * av[2] stores the multicharacter foldings, used later in
12918 * regexec.c:S_reginclass().
12919 * av[3] stores the cp_list inversion list for use in addition or
12920 * instead of av[0]; used only if av[1] is NULL
12921 * av[4] is set if any component of the class is from a user-defined
12922 * property; used only if av[1] is NULL */
12923 AV * const av = newAV();
12926 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12930 av_store(av, 1, swash);
12931 SvREFCNT_dec(cp_list);
12934 av_store(av, 1, NULL);
12936 av_store(av, 3, cp_list);
12937 av_store(av, 4, newSVuv(has_user_defined_property));
12941 /* Store any computed multi-char folds only if we are allowing
12943 if (allow_full_fold) {
12944 av_store(av, 2, MUTABLE_SV(unicode_alternate));
12945 if (unicode_alternate) { /* This node is variable length */
12950 av_store(av, 2, NULL);
12952 rv = newRV_noinc(MUTABLE_SV(av));
12953 n = add_data(pRExC_state, 1, "s");
12954 RExC_rxi->data->data[n] = (void*)rv;
12958 *flagp |= HASWIDTH|SIMPLE;
12961 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
12964 /* reg_skipcomment()
12966 Absorbs an /x style # comments from the input stream.
12967 Returns true if there is more text remaining in the stream.
12968 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
12969 terminates the pattern without including a newline.
12971 Note its the callers responsibility to ensure that we are
12972 actually in /x mode
12977 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
12981 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
12983 while (RExC_parse < RExC_end)
12984 if (*RExC_parse++ == '\n') {
12989 /* we ran off the end of the pattern without ending
12990 the comment, so we have to add an \n when wrapping */
12991 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
12999 Advances the parse position, and optionally absorbs
13000 "whitespace" from the inputstream.
13002 Without /x "whitespace" means (?#...) style comments only,
13003 with /x this means (?#...) and # comments and whitespace proper.
13005 Returns the RExC_parse point from BEFORE the scan occurs.
13007 This is the /x friendly way of saying RExC_parse++.
13011 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13013 char* const retval = RExC_parse++;
13015 PERL_ARGS_ASSERT_NEXTCHAR;
13018 if (RExC_end - RExC_parse >= 3
13019 && *RExC_parse == '('
13020 && RExC_parse[1] == '?'
13021 && RExC_parse[2] == '#')
13023 while (*RExC_parse != ')') {
13024 if (RExC_parse == RExC_end)
13025 FAIL("Sequence (?#... not terminated");
13031 if (RExC_flags & RXf_PMf_EXTENDED) {
13032 if (isSPACE(*RExC_parse)) {
13036 else if (*RExC_parse == '#') {
13037 if ( reg_skipcomment( pRExC_state ) )
13046 - reg_node - emit a node
13048 STATIC regnode * /* Location. */
13049 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13053 regnode * const ret = RExC_emit;
13054 GET_RE_DEBUG_FLAGS_DECL;
13056 PERL_ARGS_ASSERT_REG_NODE;
13059 SIZE_ALIGN(RExC_size);
13063 if (RExC_emit >= RExC_emit_bound)
13064 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13065 op, RExC_emit, RExC_emit_bound);
13067 NODE_ALIGN_FILL(ret);
13069 FILL_ADVANCE_NODE(ptr, op);
13070 #ifdef RE_TRACK_PATTERN_OFFSETS
13071 if (RExC_offsets) { /* MJD */
13072 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13073 "reg_node", __LINE__,
13075 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13076 ? "Overwriting end of array!\n" : "OK",
13077 (UV)(RExC_emit - RExC_emit_start),
13078 (UV)(RExC_parse - RExC_start),
13079 (UV)RExC_offsets[0]));
13080 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13088 - reganode - emit a node with an argument
13090 STATIC regnode * /* Location. */
13091 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13095 regnode * const ret = RExC_emit;
13096 GET_RE_DEBUG_FLAGS_DECL;
13098 PERL_ARGS_ASSERT_REGANODE;
13101 SIZE_ALIGN(RExC_size);
13106 assert(2==regarglen[op]+1);
13108 Anything larger than this has to allocate the extra amount.
13109 If we changed this to be:
13111 RExC_size += (1 + regarglen[op]);
13113 then it wouldn't matter. Its not clear what side effect
13114 might come from that so its not done so far.
13119 if (RExC_emit >= RExC_emit_bound)
13120 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13121 op, RExC_emit, RExC_emit_bound);
13123 NODE_ALIGN_FILL(ret);
13125 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
13126 #ifdef RE_TRACK_PATTERN_OFFSETS
13127 if (RExC_offsets) { /* MJD */
13128 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
13132 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
13133 "Overwriting end of array!\n" : "OK",
13134 (UV)(RExC_emit - RExC_emit_start),
13135 (UV)(RExC_parse - RExC_start),
13136 (UV)RExC_offsets[0]));
13137 Set_Cur_Node_Offset;
13145 - reguni - emit (if appropriate) a Unicode character
13148 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
13152 PERL_ARGS_ASSERT_REGUNI;
13154 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
13158 - reginsert - insert an operator in front of already-emitted operand
13160 * Means relocating the operand.
13163 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
13169 const int offset = regarglen[(U8)op];
13170 const int size = NODE_STEP_REGNODE + offset;
13171 GET_RE_DEBUG_FLAGS_DECL;
13173 PERL_ARGS_ASSERT_REGINSERT;
13174 PERL_UNUSED_ARG(depth);
13175 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
13176 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
13185 if (RExC_open_parens) {
13187 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
13188 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
13189 if ( RExC_open_parens[paren] >= opnd ) {
13190 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
13191 RExC_open_parens[paren] += size;
13193 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
13195 if ( RExC_close_parens[paren] >= opnd ) {
13196 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
13197 RExC_close_parens[paren] += size;
13199 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
13204 while (src > opnd) {
13205 StructCopy(--src, --dst, regnode);
13206 #ifdef RE_TRACK_PATTERN_OFFSETS
13207 if (RExC_offsets) { /* MJD 20010112 */
13208 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
13212 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
13213 ? "Overwriting end of array!\n" : "OK",
13214 (UV)(src - RExC_emit_start),
13215 (UV)(dst - RExC_emit_start),
13216 (UV)RExC_offsets[0]));
13217 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
13218 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
13224 place = opnd; /* Op node, where operand used to be. */
13225 #ifdef RE_TRACK_PATTERN_OFFSETS
13226 if (RExC_offsets) { /* MJD */
13227 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
13231 (UV)(place - RExC_emit_start) > RExC_offsets[0]
13232 ? "Overwriting end of array!\n" : "OK",
13233 (UV)(place - RExC_emit_start),
13234 (UV)(RExC_parse - RExC_start),
13235 (UV)RExC_offsets[0]));
13236 Set_Node_Offset(place, RExC_parse);
13237 Set_Node_Length(place, 1);
13240 src = NEXTOPER(place);
13241 FILL_ADVANCE_NODE(place, op);
13242 Zero(src, offset, regnode);
13246 - regtail - set the next-pointer at the end of a node chain of p to val.
13247 - SEE ALSO: regtail_study
13249 /* TODO: All three parms should be const */
13251 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
13255 GET_RE_DEBUG_FLAGS_DECL;
13257 PERL_ARGS_ASSERT_REGTAIL;
13259 PERL_UNUSED_ARG(depth);
13265 /* Find last node. */
13268 regnode * const temp = regnext(scan);
13270 SV * const mysv=sv_newmortal();
13271 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
13272 regprop(RExC_rx, mysv, scan);
13273 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
13274 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
13275 (temp == NULL ? "->" : ""),
13276 (temp == NULL ? PL_reg_name[OP(val)] : "")
13284 if (reg_off_by_arg[OP(scan)]) {
13285 ARG_SET(scan, val - scan);
13288 NEXT_OFF(scan) = val - scan;
13294 - regtail_study - set the next-pointer at the end of a node chain of p to val.
13295 - Look for optimizable sequences at the same time.
13296 - currently only looks for EXACT chains.
13298 This is experimental code. The idea is to use this routine to perform
13299 in place optimizations on branches and groups as they are constructed,
13300 with the long term intention of removing optimization from study_chunk so
13301 that it is purely analytical.
13303 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
13304 to control which is which.
13307 /* TODO: All four parms should be const */
13310 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
13315 #ifdef EXPERIMENTAL_INPLACESCAN
13318 GET_RE_DEBUG_FLAGS_DECL;
13320 PERL_ARGS_ASSERT_REGTAIL_STUDY;
13326 /* Find last node. */
13330 regnode * const temp = regnext(scan);
13331 #ifdef EXPERIMENTAL_INPLACESCAN
13332 if (PL_regkind[OP(scan)] == EXACT) {
13333 bool has_exactf_sharp_s; /* Unexamined in this routine */
13334 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
13339 switch (OP(scan)) {
13345 case EXACTFU_TRICKYFOLD:
13347 if( exact == PSEUDO )
13349 else if ( exact != OP(scan) )
13358 SV * const mysv=sv_newmortal();
13359 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
13360 regprop(RExC_rx, mysv, scan);
13361 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
13362 SvPV_nolen_const(mysv),
13363 REG_NODE_NUM(scan),
13364 PL_reg_name[exact]);
13371 SV * const mysv_val=sv_newmortal();
13372 DEBUG_PARSE_MSG("");
13373 regprop(RExC_rx, mysv_val, val);
13374 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
13375 SvPV_nolen_const(mysv_val),
13376 (IV)REG_NODE_NUM(val),
13380 if (reg_off_by_arg[OP(scan)]) {
13381 ARG_SET(scan, val - scan);
13384 NEXT_OFF(scan) = val - scan;
13392 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
13396 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
13402 for (bit=0; bit<32; bit++) {
13403 if (flags & (1<<bit)) {
13404 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
13407 if (!set++ && lead)
13408 PerlIO_printf(Perl_debug_log, "%s",lead);
13409 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
13412 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
13413 if (!set++ && lead) {
13414 PerlIO_printf(Perl_debug_log, "%s",lead);
13417 case REGEX_UNICODE_CHARSET:
13418 PerlIO_printf(Perl_debug_log, "UNICODE");
13420 case REGEX_LOCALE_CHARSET:
13421 PerlIO_printf(Perl_debug_log, "LOCALE");
13423 case REGEX_ASCII_RESTRICTED_CHARSET:
13424 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
13426 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
13427 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
13430 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
13436 PerlIO_printf(Perl_debug_log, "\n");
13438 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
13444 Perl_regdump(pTHX_ const regexp *r)
13448 SV * const sv = sv_newmortal();
13449 SV *dsv= sv_newmortal();
13450 RXi_GET_DECL(r,ri);
13451 GET_RE_DEBUG_FLAGS_DECL;
13453 PERL_ARGS_ASSERT_REGDUMP;
13455 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
13457 /* Header fields of interest. */
13458 if (r->anchored_substr) {
13459 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
13460 RE_SV_DUMPLEN(r->anchored_substr), 30);
13461 PerlIO_printf(Perl_debug_log,
13462 "anchored %s%s at %"IVdf" ",
13463 s, RE_SV_TAIL(r->anchored_substr),
13464 (IV)r->anchored_offset);
13465 } else if (r->anchored_utf8) {
13466 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
13467 RE_SV_DUMPLEN(r->anchored_utf8), 30);
13468 PerlIO_printf(Perl_debug_log,
13469 "anchored utf8 %s%s at %"IVdf" ",
13470 s, RE_SV_TAIL(r->anchored_utf8),
13471 (IV)r->anchored_offset);
13473 if (r->float_substr) {
13474 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
13475 RE_SV_DUMPLEN(r->float_substr), 30);
13476 PerlIO_printf(Perl_debug_log,
13477 "floating %s%s at %"IVdf"..%"UVuf" ",
13478 s, RE_SV_TAIL(r->float_substr),
13479 (IV)r->float_min_offset, (UV)r->float_max_offset);
13480 } else if (r->float_utf8) {
13481 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
13482 RE_SV_DUMPLEN(r->float_utf8), 30);
13483 PerlIO_printf(Perl_debug_log,
13484 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
13485 s, RE_SV_TAIL(r->float_utf8),
13486 (IV)r->float_min_offset, (UV)r->float_max_offset);
13488 if (r->check_substr || r->check_utf8)
13489 PerlIO_printf(Perl_debug_log,
13491 (r->check_substr == r->float_substr
13492 && r->check_utf8 == r->float_utf8
13493 ? "(checking floating" : "(checking anchored"));
13494 if (r->extflags & RXf_NOSCAN)
13495 PerlIO_printf(Perl_debug_log, " noscan");
13496 if (r->extflags & RXf_CHECK_ALL)
13497 PerlIO_printf(Perl_debug_log, " isall");
13498 if (r->check_substr || r->check_utf8)
13499 PerlIO_printf(Perl_debug_log, ") ");
13501 if (ri->regstclass) {
13502 regprop(r, sv, ri->regstclass);
13503 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
13505 if (r->extflags & RXf_ANCH) {
13506 PerlIO_printf(Perl_debug_log, "anchored");
13507 if (r->extflags & RXf_ANCH_BOL)
13508 PerlIO_printf(Perl_debug_log, "(BOL)");
13509 if (r->extflags & RXf_ANCH_MBOL)
13510 PerlIO_printf(Perl_debug_log, "(MBOL)");
13511 if (r->extflags & RXf_ANCH_SBOL)
13512 PerlIO_printf(Perl_debug_log, "(SBOL)");
13513 if (r->extflags & RXf_ANCH_GPOS)
13514 PerlIO_printf(Perl_debug_log, "(GPOS)");
13515 PerlIO_putc(Perl_debug_log, ' ');
13517 if (r->extflags & RXf_GPOS_SEEN)
13518 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
13519 if (r->intflags & PREGf_SKIP)
13520 PerlIO_printf(Perl_debug_log, "plus ");
13521 if (r->intflags & PREGf_IMPLICIT)
13522 PerlIO_printf(Perl_debug_log, "implicit ");
13523 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
13524 if (r->extflags & RXf_EVAL_SEEN)
13525 PerlIO_printf(Perl_debug_log, "with eval ");
13526 PerlIO_printf(Perl_debug_log, "\n");
13527 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
13529 PERL_ARGS_ASSERT_REGDUMP;
13530 PERL_UNUSED_CONTEXT;
13531 PERL_UNUSED_ARG(r);
13532 #endif /* DEBUGGING */
13536 - regprop - printable representation of opcode
13538 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
13541 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
13542 if (flags & ANYOF_INVERT) \
13543 /*make sure the invert info is in each */ \
13544 sv_catpvs(sv, "^"); \
13550 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
13556 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
13557 static const char * const anyofs[] = {
13589 RXi_GET_DECL(prog,progi);
13590 GET_RE_DEBUG_FLAGS_DECL;
13592 PERL_ARGS_ASSERT_REGPROP;
13596 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
13597 /* It would be nice to FAIL() here, but this may be called from
13598 regexec.c, and it would be hard to supply pRExC_state. */
13599 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
13600 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
13602 k = PL_regkind[OP(o)];
13605 sv_catpvs(sv, " ");
13606 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
13607 * is a crude hack but it may be the best for now since
13608 * we have no flag "this EXACTish node was UTF-8"
13610 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
13611 PERL_PV_ESCAPE_UNI_DETECT |
13612 PERL_PV_ESCAPE_NONASCII |
13613 PERL_PV_PRETTY_ELLIPSES |
13614 PERL_PV_PRETTY_LTGT |
13615 PERL_PV_PRETTY_NOCLEAR
13617 } else if (k == TRIE) {
13618 /* print the details of the trie in dumpuntil instead, as
13619 * progi->data isn't available here */
13620 const char op = OP(o);
13621 const U32 n = ARG(o);
13622 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
13623 (reg_ac_data *)progi->data->data[n] :
13625 const reg_trie_data * const trie
13626 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
13628 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
13629 DEBUG_TRIE_COMPILE_r(
13630 Perl_sv_catpvf(aTHX_ sv,
13631 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
13632 (UV)trie->startstate,
13633 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
13634 (UV)trie->wordcount,
13637 (UV)TRIE_CHARCOUNT(trie),
13638 (UV)trie->uniquecharcount
13641 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
13643 int rangestart = -1;
13644 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
13645 sv_catpvs(sv, "[");
13646 for (i = 0; i <= 256; i++) {
13647 if (i < 256 && BITMAP_TEST(bitmap,i)) {
13648 if (rangestart == -1)
13650 } else if (rangestart != -1) {
13651 if (i <= rangestart + 3)
13652 for (; rangestart < i; rangestart++)
13653 put_byte(sv, rangestart);
13655 put_byte(sv, rangestart);
13656 sv_catpvs(sv, "-");
13657 put_byte(sv, i - 1);
13662 sv_catpvs(sv, "]");
13665 } else if (k == CURLY) {
13666 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
13667 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
13668 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
13670 else if (k == WHILEM && o->flags) /* Ordinal/of */
13671 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
13672 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
13673 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
13674 if ( RXp_PAREN_NAMES(prog) ) {
13675 if ( k != REF || (OP(o) < NREF)) {
13676 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
13677 SV **name= av_fetch(list, ARG(o), 0 );
13679 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
13682 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
13683 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
13684 I32 *nums=(I32*)SvPVX(sv_dat);
13685 SV **name= av_fetch(list, nums[0], 0 );
13688 for ( n=0; n<SvIVX(sv_dat); n++ ) {
13689 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
13690 (n ? "," : ""), (IV)nums[n]);
13692 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
13696 } else if (k == GOSUB)
13697 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
13698 else if (k == VERB) {
13700 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
13701 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
13702 } else if (k == LOGICAL)
13703 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
13704 else if (k == ANYOF) {
13705 int i, rangestart = -1;
13706 const U8 flags = ANYOF_FLAGS(o);
13710 if (flags & ANYOF_LOCALE)
13711 sv_catpvs(sv, "{loc}");
13712 if (flags & ANYOF_LOC_NONBITMAP_FOLD)
13713 sv_catpvs(sv, "{i}");
13714 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
13715 if (flags & ANYOF_INVERT)
13716 sv_catpvs(sv, "^");
13718 /* output what the standard cp 0-255 bitmap matches */
13719 for (i = 0; i <= 256; i++) {
13720 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
13721 if (rangestart == -1)
13723 } else if (rangestart != -1) {
13724 if (i <= rangestart + 3)
13725 for (; rangestart < i; rangestart++)
13726 put_byte(sv, rangestart);
13728 put_byte(sv, rangestart);
13729 sv_catpvs(sv, "-");
13730 put_byte(sv, i - 1);
13737 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
13738 /* output any special charclass tests (used entirely under use locale) */
13739 if (ANYOF_CLASS_TEST_ANY_SET(o))
13740 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
13741 if (ANYOF_CLASS_TEST(o,i)) {
13742 sv_catpv(sv, anyofs[i]);
13746 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
13748 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
13749 sv_catpvs(sv, "{non-utf8-latin1-all}");
13752 /* output information about the unicode matching */
13753 if (flags & ANYOF_UNICODE_ALL)
13754 sv_catpvs(sv, "{unicode_all}");
13755 else if (ANYOF_NONBITMAP(o))
13756 sv_catpvs(sv, "{unicode}");
13757 if (flags & ANYOF_NONBITMAP_NON_UTF8)
13758 sv_catpvs(sv, "{outside bitmap}");
13760 if (ANYOF_NONBITMAP(o)) {
13761 SV *lv; /* Set if there is something outside the bit map */
13762 SV * const sw = regclass_swash(prog, o, FALSE, &lv, 0);
13763 bool byte_output = FALSE; /* If something in the bitmap has been
13766 if (lv && lv != &PL_sv_undef) {
13768 U8 s[UTF8_MAXBYTES_CASE+1];
13770 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
13771 uvchr_to_utf8(s, i);
13774 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
13778 && swash_fetch(sw, s, TRUE))
13780 if (rangestart == -1)
13782 } else if (rangestart != -1) {
13783 byte_output = TRUE;
13784 if (i <= rangestart + 3)
13785 for (; rangestart < i; rangestart++) {
13786 put_byte(sv, rangestart);
13789 put_byte(sv, rangestart);
13790 sv_catpvs(sv, "-");
13799 char *s = savesvpv(lv);
13800 char * const origs = s;
13802 while (*s && *s != '\n')
13806 const char * const t = ++s;
13809 sv_catpvs(sv, " ");
13815 /* Truncate very long output */
13816 if (s - origs > 256) {
13817 Perl_sv_catpvf(aTHX_ sv,
13819 (int) (s - origs - 1),
13825 else if (*s == '\t') {
13844 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
13846 else if (k == POSIXD) {
13847 U8 index = FLAGS(o) * 2;
13848 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
13849 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
13852 sv_catpv(sv, anyofs[index]);
13855 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
13856 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
13858 PERL_UNUSED_CONTEXT;
13859 PERL_UNUSED_ARG(sv);
13860 PERL_UNUSED_ARG(o);
13861 PERL_UNUSED_ARG(prog);
13862 #endif /* DEBUGGING */
13866 Perl_re_intuit_string(pTHX_ REGEXP * const r)
13867 { /* Assume that RE_INTUIT is set */
13869 struct regexp *const prog = (struct regexp *)SvANY(r);
13870 GET_RE_DEBUG_FLAGS_DECL;
13872 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
13873 PERL_UNUSED_CONTEXT;
13877 const char * const s = SvPV_nolen_const(prog->check_substr
13878 ? prog->check_substr : prog->check_utf8);
13880 if (!PL_colorset) reginitcolors();
13881 PerlIO_printf(Perl_debug_log,
13882 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
13884 prog->check_substr ? "" : "utf8 ",
13885 PL_colors[5],PL_colors[0],
13888 (strlen(s) > 60 ? "..." : ""));
13891 return prog->check_substr ? prog->check_substr : prog->check_utf8;
13897 handles refcounting and freeing the perl core regexp structure. When
13898 it is necessary to actually free the structure the first thing it
13899 does is call the 'free' method of the regexp_engine associated to
13900 the regexp, allowing the handling of the void *pprivate; member
13901 first. (This routine is not overridable by extensions, which is why
13902 the extensions free is called first.)
13904 See regdupe and regdupe_internal if you change anything here.
13906 #ifndef PERL_IN_XSUB_RE
13908 Perl_pregfree(pTHX_ REGEXP *r)
13914 Perl_pregfree2(pTHX_ REGEXP *rx)
13917 struct regexp *const r = (struct regexp *)SvANY(rx);
13918 GET_RE_DEBUG_FLAGS_DECL;
13920 PERL_ARGS_ASSERT_PREGFREE2;
13922 if (r->mother_re) {
13923 ReREFCNT_dec(r->mother_re);
13925 CALLREGFREE_PVT(rx); /* free the private data */
13926 SvREFCNT_dec(RXp_PAREN_NAMES(r));
13929 SvREFCNT_dec(r->anchored_substr);
13930 SvREFCNT_dec(r->anchored_utf8);
13931 SvREFCNT_dec(r->float_substr);
13932 SvREFCNT_dec(r->float_utf8);
13933 Safefree(r->substrs);
13935 RX_MATCH_COPY_FREE(rx);
13936 #ifdef PERL_OLD_COPY_ON_WRITE
13937 SvREFCNT_dec(r->saved_copy);
13940 SvREFCNT_dec(r->qr_anoncv);
13945 This is a hacky workaround to the structural issue of match results
13946 being stored in the regexp structure which is in turn stored in
13947 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
13948 could be PL_curpm in multiple contexts, and could require multiple
13949 result sets being associated with the pattern simultaneously, such
13950 as when doing a recursive match with (??{$qr})
13952 The solution is to make a lightweight copy of the regexp structure
13953 when a qr// is returned from the code executed by (??{$qr}) this
13954 lightweight copy doesn't actually own any of its data except for
13955 the starp/end and the actual regexp structure itself.
13961 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
13963 struct regexp *ret;
13964 struct regexp *const r = (struct regexp *)SvANY(rx);
13966 PERL_ARGS_ASSERT_REG_TEMP_COPY;
13969 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
13970 ret = (struct regexp *)SvANY(ret_x);
13972 (void)ReREFCNT_inc(rx);
13973 /* We can take advantage of the existing "copied buffer" mechanism in SVs
13974 by pointing directly at the buffer, but flagging that the allocated
13975 space in the copy is zero. As we've just done a struct copy, it's now
13976 a case of zero-ing that, rather than copying the current length. */
13977 SvPV_set(ret_x, RX_WRAPPED(rx));
13978 SvFLAGS(ret_x) |= SvFLAGS(rx) & (SVf_POK|SVp_POK|SVf_UTF8);
13979 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
13980 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
13981 SvLEN_set(ret_x, 0);
13982 SvSTASH_set(ret_x, NULL);
13983 SvMAGIC_set(ret_x, NULL);
13985 const I32 npar = r->nparens+1;
13986 Newx(ret->offs, npar, regexp_paren_pair);
13987 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
13990 Newx(ret->substrs, 1, struct reg_substr_data);
13991 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
13993 SvREFCNT_inc_void(ret->anchored_substr);
13994 SvREFCNT_inc_void(ret->anchored_utf8);
13995 SvREFCNT_inc_void(ret->float_substr);
13996 SvREFCNT_inc_void(ret->float_utf8);
13998 /* check_substr and check_utf8, if non-NULL, point to either their
13999 anchored or float namesakes, and don't hold a second reference. */
14001 RX_MATCH_COPIED_off(ret_x);
14002 #ifdef PERL_OLD_COPY_ON_WRITE
14003 ret->saved_copy = NULL;
14005 ret->mother_re = rx;
14006 SvREFCNT_inc_void(ret->qr_anoncv);
14012 /* regfree_internal()
14014 Free the private data in a regexp. This is overloadable by
14015 extensions. Perl takes care of the regexp structure in pregfree(),
14016 this covers the *pprivate pointer which technically perl doesn't
14017 know about, however of course we have to handle the
14018 regexp_internal structure when no extension is in use.
14020 Note this is called before freeing anything in the regexp
14025 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14028 struct regexp *const r = (struct regexp *)SvANY(rx);
14029 RXi_GET_DECL(r,ri);
14030 GET_RE_DEBUG_FLAGS_DECL;
14032 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14038 SV *dsv= sv_newmortal();
14039 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14040 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14041 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14042 PL_colors[4],PL_colors[5],s);
14045 #ifdef RE_TRACK_PATTERN_OFFSETS
14047 Safefree(ri->u.offsets); /* 20010421 MJD */
14049 if (ri->code_blocks) {
14051 for (n = 0; n < ri->num_code_blocks; n++)
14052 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14053 Safefree(ri->code_blocks);
14057 int n = ri->data->count;
14060 /* If you add a ->what type here, update the comment in regcomp.h */
14061 switch (ri->data->what[n]) {
14067 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14070 Safefree(ri->data->data[n]);
14076 { /* Aho Corasick add-on structure for a trie node.
14077 Used in stclass optimization only */
14079 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
14081 refcount = --aho->refcount;
14084 PerlMemShared_free(aho->states);
14085 PerlMemShared_free(aho->fail);
14086 /* do this last!!!! */
14087 PerlMemShared_free(ri->data->data[n]);
14088 PerlMemShared_free(ri->regstclass);
14094 /* trie structure. */
14096 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
14098 refcount = --trie->refcount;
14101 PerlMemShared_free(trie->charmap);
14102 PerlMemShared_free(trie->states);
14103 PerlMemShared_free(trie->trans);
14105 PerlMemShared_free(trie->bitmap);
14107 PerlMemShared_free(trie->jump);
14108 PerlMemShared_free(trie->wordinfo);
14109 /* do this last!!!! */
14110 PerlMemShared_free(ri->data->data[n]);
14115 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
14118 Safefree(ri->data->what);
14119 Safefree(ri->data);
14125 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
14126 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
14127 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
14130 re_dup - duplicate a regexp.
14132 This routine is expected to clone a given regexp structure. It is only
14133 compiled under USE_ITHREADS.
14135 After all of the core data stored in struct regexp is duplicated
14136 the regexp_engine.dupe method is used to copy any private data
14137 stored in the *pprivate pointer. This allows extensions to handle
14138 any duplication it needs to do.
14140 See pregfree() and regfree_internal() if you change anything here.
14142 #if defined(USE_ITHREADS)
14143 #ifndef PERL_IN_XSUB_RE
14145 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
14149 const struct regexp *r = (const struct regexp *)SvANY(sstr);
14150 struct regexp *ret = (struct regexp *)SvANY(dstr);
14152 PERL_ARGS_ASSERT_RE_DUP_GUTS;
14154 npar = r->nparens+1;
14155 Newx(ret->offs, npar, regexp_paren_pair);
14156 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14158 /* no need to copy these */
14159 Newx(ret->swap, npar, regexp_paren_pair);
14162 if (ret->substrs) {
14163 /* Do it this way to avoid reading from *r after the StructCopy().
14164 That way, if any of the sv_dup_inc()s dislodge *r from the L1
14165 cache, it doesn't matter. */
14166 const bool anchored = r->check_substr
14167 ? r->check_substr == r->anchored_substr
14168 : r->check_utf8 == r->anchored_utf8;
14169 Newx(ret->substrs, 1, struct reg_substr_data);
14170 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14172 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
14173 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
14174 ret->float_substr = sv_dup_inc(ret->float_substr, param);
14175 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
14177 /* check_substr and check_utf8, if non-NULL, point to either their
14178 anchored or float namesakes, and don't hold a second reference. */
14180 if (ret->check_substr) {
14182 assert(r->check_utf8 == r->anchored_utf8);
14183 ret->check_substr = ret->anchored_substr;
14184 ret->check_utf8 = ret->anchored_utf8;
14186 assert(r->check_substr == r->float_substr);
14187 assert(r->check_utf8 == r->float_utf8);
14188 ret->check_substr = ret->float_substr;
14189 ret->check_utf8 = ret->float_utf8;
14191 } else if (ret->check_utf8) {
14193 ret->check_utf8 = ret->anchored_utf8;
14195 ret->check_utf8 = ret->float_utf8;
14200 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
14201 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
14204 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
14206 if (RX_MATCH_COPIED(dstr))
14207 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
14209 ret->subbeg = NULL;
14210 #ifdef PERL_OLD_COPY_ON_WRITE
14211 ret->saved_copy = NULL;
14214 if (ret->mother_re) {
14215 if (SvPVX_const(dstr) == SvPVX_const(ret->mother_re)) {
14216 /* Our storage points directly to our mother regexp, but that's
14217 1: a buffer in a different thread
14218 2: something we no longer hold a reference on
14219 so we need to copy it locally. */
14220 /* Note we need to use SvCUR(), rather than
14221 SvLEN(), on our mother_re, because it, in
14222 turn, may well be pointing to its own mother_re. */
14223 SvPV_set(dstr, SAVEPVN(SvPVX_const(ret->mother_re),
14224 SvCUR(ret->mother_re)+1));
14225 SvLEN_set(dstr, SvCUR(ret->mother_re)+1);
14227 ret->mother_re = NULL;
14231 #endif /* PERL_IN_XSUB_RE */
14236 This is the internal complement to regdupe() which is used to copy
14237 the structure pointed to by the *pprivate pointer in the regexp.
14238 This is the core version of the extension overridable cloning hook.
14239 The regexp structure being duplicated will be copied by perl prior
14240 to this and will be provided as the regexp *r argument, however
14241 with the /old/ structures pprivate pointer value. Thus this routine
14242 may override any copying normally done by perl.
14244 It returns a pointer to the new regexp_internal structure.
14248 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
14251 struct regexp *const r = (struct regexp *)SvANY(rx);
14252 regexp_internal *reti;
14254 RXi_GET_DECL(r,ri);
14256 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
14260 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
14261 Copy(ri->program, reti->program, len+1, regnode);
14263 reti->num_code_blocks = ri->num_code_blocks;
14264 if (ri->code_blocks) {
14266 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
14267 struct reg_code_block);
14268 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
14269 struct reg_code_block);
14270 for (n = 0; n < ri->num_code_blocks; n++)
14271 reti->code_blocks[n].src_regex = (REGEXP*)
14272 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
14275 reti->code_blocks = NULL;
14277 reti->regstclass = NULL;
14280 struct reg_data *d;
14281 const int count = ri->data->count;
14284 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
14285 char, struct reg_data);
14286 Newx(d->what, count, U8);
14289 for (i = 0; i < count; i++) {
14290 d->what[i] = ri->data->what[i];
14291 switch (d->what[i]) {
14292 /* see also regcomp.h and regfree_internal() */
14293 case 'a': /* actually an AV, but the dup function is identical. */
14297 case 'u': /* actually an HV, but the dup function is identical. */
14298 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
14301 /* This is cheating. */
14302 Newx(d->data[i], 1, struct regnode_charclass_class);
14303 StructCopy(ri->data->data[i], d->data[i],
14304 struct regnode_charclass_class);
14305 reti->regstclass = (regnode*)d->data[i];
14308 /* Trie stclasses are readonly and can thus be shared
14309 * without duplication. We free the stclass in pregfree
14310 * when the corresponding reg_ac_data struct is freed.
14312 reti->regstclass= ri->regstclass;
14316 ((reg_trie_data*)ri->data->data[i])->refcount++;
14321 d->data[i] = ri->data->data[i];
14324 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
14333 reti->name_list_idx = ri->name_list_idx;
14335 #ifdef RE_TRACK_PATTERN_OFFSETS
14336 if (ri->u.offsets) {
14337 Newx(reti->u.offsets, 2*len+1, U32);
14338 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
14341 SetProgLen(reti,len);
14344 return (void*)reti;
14347 #endif /* USE_ITHREADS */
14349 #ifndef PERL_IN_XSUB_RE
14352 - regnext - dig the "next" pointer out of a node
14355 Perl_regnext(pTHX_ register regnode *p)
14363 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
14364 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
14367 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
14376 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
14379 STRLEN l1 = strlen(pat1);
14380 STRLEN l2 = strlen(pat2);
14383 const char *message;
14385 PERL_ARGS_ASSERT_RE_CROAK2;
14391 Copy(pat1, buf, l1 , char);
14392 Copy(pat2, buf + l1, l2 , char);
14393 buf[l1 + l2] = '\n';
14394 buf[l1 + l2 + 1] = '\0';
14396 /* ANSI variant takes additional second argument */
14397 va_start(args, pat2);
14401 msv = vmess(buf, &args);
14403 message = SvPV_const(msv,l1);
14406 Copy(message, buf, l1 , char);
14407 buf[l1-1] = '\0'; /* Overwrite \n */
14408 Perl_croak(aTHX_ "%s", buf);
14411 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
14413 #ifndef PERL_IN_XSUB_RE
14415 Perl_save_re_context(pTHX)
14419 struct re_save_state *state;
14421 SAVEVPTR(PL_curcop);
14422 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
14424 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
14425 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
14426 SSPUSHUV(SAVEt_RE_STATE);
14428 Copy(&PL_reg_state, state, 1, struct re_save_state);
14430 PL_reg_oldsaved = NULL;
14431 PL_reg_oldsavedlen = 0;
14432 PL_reg_maxiter = 0;
14433 PL_reg_leftiter = 0;
14434 PL_reg_poscache = NULL;
14435 PL_reg_poscache_size = 0;
14436 #ifdef PERL_OLD_COPY_ON_WRITE
14440 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
14442 const REGEXP * const rx = PM_GETRE(PL_curpm);
14445 for (i = 1; i <= RX_NPARENS(rx); i++) {
14446 char digits[TYPE_CHARS(long)];
14447 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
14448 GV *const *const gvp
14449 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
14452 GV * const gv = *gvp;
14453 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
14463 clear_re(pTHX_ void *r)
14466 ReREFCNT_dec((REGEXP *)r);
14472 S_put_byte(pTHX_ SV *sv, int c)
14474 PERL_ARGS_ASSERT_PUT_BYTE;
14476 /* Our definition of isPRINT() ignores locales, so only bytes that are
14477 not part of UTF-8 are considered printable. I assume that the same
14478 holds for UTF-EBCDIC.
14479 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
14480 which Wikipedia says:
14482 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
14483 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
14484 identical, to the ASCII delete (DEL) or rubout control character.
14485 ) So the old condition can be simplified to !isPRINT(c) */
14488 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
14491 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
14495 const char string = c;
14496 if (c == '-' || c == ']' || c == '\\' || c == '^')
14497 sv_catpvs(sv, "\\");
14498 sv_catpvn(sv, &string, 1);
14503 #define CLEAR_OPTSTART \
14504 if (optstart) STMT_START { \
14505 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
14509 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
14511 STATIC const regnode *
14512 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
14513 const regnode *last, const regnode *plast,
14514 SV* sv, I32 indent, U32 depth)
14517 U8 op = PSEUDO; /* Arbitrary non-END op. */
14518 const regnode *next;
14519 const regnode *optstart= NULL;
14521 RXi_GET_DECL(r,ri);
14522 GET_RE_DEBUG_FLAGS_DECL;
14524 PERL_ARGS_ASSERT_DUMPUNTIL;
14526 #ifdef DEBUG_DUMPUNTIL
14527 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
14528 last ? last-start : 0,plast ? plast-start : 0);
14531 if (plast && plast < last)
14534 while (PL_regkind[op] != END && (!last || node < last)) {
14535 /* While that wasn't END last time... */
14538 if (op == CLOSE || op == WHILEM)
14540 next = regnext((regnode *)node);
14543 if (OP(node) == OPTIMIZED) {
14544 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
14551 regprop(r, sv, node);
14552 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
14553 (int)(2*indent + 1), "", SvPVX_const(sv));
14555 if (OP(node) != OPTIMIZED) {
14556 if (next == NULL) /* Next ptr. */
14557 PerlIO_printf(Perl_debug_log, " (0)");
14558 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
14559 PerlIO_printf(Perl_debug_log, " (FAIL)");
14561 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
14562 (void)PerlIO_putc(Perl_debug_log, '\n');
14566 if (PL_regkind[(U8)op] == BRANCHJ) {
14569 const regnode *nnode = (OP(next) == LONGJMP
14570 ? regnext((regnode *)next)
14572 if (last && nnode > last)
14574 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
14577 else if (PL_regkind[(U8)op] == BRANCH) {
14579 DUMPUNTIL(NEXTOPER(node), next);
14581 else if ( PL_regkind[(U8)op] == TRIE ) {
14582 const regnode *this_trie = node;
14583 const char op = OP(node);
14584 const U32 n = ARG(node);
14585 const reg_ac_data * const ac = op>=AHOCORASICK ?
14586 (reg_ac_data *)ri->data->data[n] :
14588 const reg_trie_data * const trie =
14589 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
14591 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
14593 const regnode *nextbranch= NULL;
14596 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
14597 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
14599 PerlIO_printf(Perl_debug_log, "%*s%s ",
14600 (int)(2*(indent+3)), "",
14601 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
14602 PL_colors[0], PL_colors[1],
14603 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
14604 PERL_PV_PRETTY_ELLIPSES |
14605 PERL_PV_PRETTY_LTGT
14610 U16 dist= trie->jump[word_idx+1];
14611 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
14612 (UV)((dist ? this_trie + dist : next) - start));
14615 nextbranch= this_trie + trie->jump[0];
14616 DUMPUNTIL(this_trie + dist, nextbranch);
14618 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
14619 nextbranch= regnext((regnode *)nextbranch);
14621 PerlIO_printf(Perl_debug_log, "\n");
14624 if (last && next > last)
14629 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
14630 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
14631 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
14633 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
14635 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
14637 else if ( op == PLUS || op == STAR) {
14638 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
14640 else if (PL_regkind[(U8)op] == ANYOF) {
14641 /* arglen 1 + class block */
14642 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
14643 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
14644 node = NEXTOPER(node);
14646 else if (PL_regkind[(U8)op] == EXACT) {
14647 /* Literal string, where present. */
14648 node += NODE_SZ_STR(node) - 1;
14649 node = NEXTOPER(node);
14652 node = NEXTOPER(node);
14653 node += regarglen[(U8)op];
14655 if (op == CURLYX || op == OPEN)
14659 #ifdef DEBUG_DUMPUNTIL
14660 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
14665 #endif /* DEBUGGING */
14669 * c-indentation-style: bsd
14670 * c-basic-offset: 4
14671 * indent-tabs-mode: nil
14674 * ex: set ts=8 sts=4 sw=4 et:
https://perl5.git.perl.org / perl5.git / blob