5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 extern const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "charclass_invlists.h"
91 #include "inline_invlist.c"
92 #include "unicode_constants.h"
94 #define HAS_NONLATIN1_FOLD_CLOSURE(i) _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
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))
5507 && SvRV(msv) == SvRV(sv))
5512 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5514 orig_patlen = SvCUR(pat);
5515 sv_catsv_nomg(pat, msv);
5518 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
5521 /* extract any code blocks within any embedded qr//'s */
5522 if (rx && SvTYPE(rx) == SVt_REGEXP
5523 && RX_ENGINE((REGEXP*)rx)->op_comp)
5526 RXi_GET_DECL(((struct regexp*)SvANY(rx)), ri);
5527 if (ri->num_code_blocks) {
5529 /* the presence of an embedded qr// with code means
5530 * we should always recompile: the text of the
5531 * qr// may not have changed, but it may be a
5532 * different closure than last time */
5534 Renew(pRExC_state->code_blocks,
5535 pRExC_state->num_code_blocks + ri->num_code_blocks,
5536 struct reg_code_block);
5537 pRExC_state->num_code_blocks += ri->num_code_blocks;
5538 for (i=0; i < ri->num_code_blocks; i++) {
5539 struct reg_code_block *src, *dst;
5540 STRLEN offset = orig_patlen
5541 + ((struct regexp *)SvANY(rx))->pre_prefix;
5542 assert(n < pRExC_state->num_code_blocks);
5543 src = &ri->code_blocks[i];
5544 dst = &pRExC_state->code_blocks[n];
5545 dst->start = src->start + offset;
5546 dst->end = src->end + offset;
5547 dst->block = src->block;
5548 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
5562 while (SvAMAGIC(pat)
5563 && (sv = AMG_CALLunary(pat, string_amg))
5571 /* handle bare regex: foo =~ $re */
5576 if (SvTYPE(re) == SVt_REGEXP) {
5580 Safefree(pRExC_state->code_blocks);
5586 /* not a list of SVs, so must be a list of OPs */
5588 if (expr->op_type == OP_LIST) {
5593 pat = newSVpvn("", 0);
5598 /* given a list of CONSTs and DO blocks in expr, append all
5599 * the CONSTs to pat, and record the start and end of each
5600 * code block in code_blocks[] (each DO{} op is followed by an
5601 * OP_CONST containing the corresponding literal '(?{...})
5604 for (o = cLISTOPx(expr)->op_first; o; o = o->op_sibling) {
5605 if (o->op_type == OP_CONST) {
5606 sv_catsv(pat, cSVOPo_sv);
5608 pRExC_state->code_blocks[i].end = SvCUR(pat)-1;
5612 else if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL)) {
5613 assert(i+1 < pRExC_state->num_code_blocks);
5614 pRExC_state->code_blocks[++i].start = SvCUR(pat);
5615 pRExC_state->code_blocks[i].block = o;
5616 pRExC_state->code_blocks[i].src_regex = NULL;
5622 assert(expr->op_type == OP_CONST);
5623 pat = cSVOPx_sv(expr);
5627 exp = SvPV_nomg(pat, plen);
5629 if (!eng->op_comp) {
5630 if ((SvUTF8(pat) && IN_BYTES)
5631 || SvGMAGICAL(pat) || SvAMAGIC(pat))
5633 /* make a temporary copy; either to convert to bytes,
5634 * or to avoid repeating get-magic / overloaded stringify */
5635 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
5636 (IN_BYTES ? 0 : SvUTF8(pat)));
5638 Safefree(pRExC_state->code_blocks);
5639 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
5642 /* ignore the utf8ness if the pattern is 0 length */
5643 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
5644 RExC_uni_semantics = 0;
5645 RExC_contains_locale = 0;
5646 pRExC_state->runtime_code_qr = NULL;
5648 /****************** LONG JUMP TARGET HERE***********************/
5649 /* Longjmp back to here if have to switch in midstream to utf8 */
5650 if (! RExC_orig_utf8) {
5651 JMPENV_PUSH(jump_ret);
5652 used_setjump = TRUE;
5655 if (jump_ret == 0) { /* First time through */
5659 SV *dsv= sv_newmortal();
5660 RE_PV_QUOTED_DECL(s, RExC_utf8,
5661 dsv, exp, plen, 60);
5662 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
5663 PL_colors[4],PL_colors[5],s);
5666 else { /* longjumped back */
5669 STRLEN s = 0, d = 0;
5672 /* If the cause for the longjmp was other than changing to utf8, pop
5673 * our own setjmp, and longjmp to the correct handler */
5674 if (jump_ret != UTF8_LONGJMP) {
5676 JMPENV_JUMP(jump_ret);
5681 /* It's possible to write a regexp in ascii that represents Unicode
5682 codepoints outside of the byte range, such as via \x{100}. If we
5683 detect such a sequence we have to convert the entire pattern to utf8
5684 and then recompile, as our sizing calculation will have been based
5685 on 1 byte == 1 character, but we will need to use utf8 to encode
5686 at least some part of the pattern, and therefore must convert the whole
5689 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5690 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5692 /* upgrade pattern to UTF8, and if there are code blocks,
5693 * recalculate the indices.
5694 * This is essentially an unrolled Perl_bytes_to_utf8() */
5696 src = (U8*)SvPV_nomg(pat, plen);
5697 Newx(dst, plen * 2 + 1, U8);
5700 const UV uv = NATIVE_TO_ASCII(src[s]);
5701 if (UNI_IS_INVARIANT(uv))
5702 dst[d] = (U8)UTF_TO_NATIVE(uv);
5704 dst[d++] = (U8)UTF8_EIGHT_BIT_HI(uv);
5705 dst[d] = (U8)UTF8_EIGHT_BIT_LO(uv);
5707 if (n < pRExC_state->num_code_blocks) {
5708 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5709 pRExC_state->code_blocks[n].start = d;
5710 assert(dst[d] == '(');
5713 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5714 pRExC_state->code_blocks[n].end = d;
5715 assert(dst[d] == ')');
5728 RExC_orig_utf8 = RExC_utf8 = 1;
5731 /* return old regex if pattern hasn't changed */
5735 && !!RX_UTF8(old_re) == !!RExC_utf8
5736 && RX_PRECOMP(old_re)
5737 && RX_PRELEN(old_re) == plen
5738 && memEQ(RX_PRECOMP(old_re), exp, plen))
5740 /* with runtime code, always recompile */
5741 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5743 if (!runtime_code) {
5747 Safefree(pRExC_state->code_blocks);
5751 else if ((pm_flags & PMf_USE_RE_EVAL)
5752 /* this second condition covers the non-regex literal case,
5753 * i.e. $foo =~ '(?{})'. */
5754 || ( !PL_reg_state.re_reparsing && IN_PERL_COMPILETIME
5755 && (PL_hints & HINT_RE_EVAL))
5757 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, expr, pm_flags,
5760 #ifdef TRIE_STUDY_OPT
5764 rx_flags = orig_rx_flags;
5766 if (initial_charset == REGEX_LOCALE_CHARSET) {
5767 RExC_contains_locale = 1;
5769 else if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
5771 /* Set to use unicode semantics if the pattern is in utf8 and has the
5772 * 'depends' charset specified, as it means unicode when utf8 */
5773 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5777 RExC_flags = rx_flags;
5778 RExC_pm_flags = pm_flags;
5781 if (PL_tainting && PL_tainted)
5782 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
5784 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
5785 /* whoops, we have a non-utf8 pattern, whilst run-time code
5786 * got compiled as utf8. Try again with a utf8 pattern */
5787 JMPENV_JUMP(UTF8_LONGJMP);
5790 assert(!pRExC_state->runtime_code_qr);
5795 RExC_in_lookbehind = 0;
5796 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
5798 RExC_override_recoding = 0;
5800 /* First pass: determine size, legality. */
5808 RExC_emit = &PL_regdummy;
5809 RExC_whilem_seen = 0;
5810 RExC_open_parens = NULL;
5811 RExC_close_parens = NULL;
5813 RExC_paren_names = NULL;
5815 RExC_paren_name_list = NULL;
5817 RExC_recurse = NULL;
5818 RExC_recurse_count = 0;
5819 pRExC_state->code_index = 0;
5821 #if 0 /* REGC() is (currently) a NOP at the first pass.
5822 * Clever compilers notice this and complain. --jhi */
5823 REGC((U8)REG_MAGIC, (char*)RExC_emit);
5826 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
5828 RExC_lastparse=NULL;
5830 if (reg(pRExC_state, 0, &flags,1) == NULL) {
5831 RExC_precomp = NULL;
5832 Safefree(pRExC_state->code_blocks);
5836 /* Here, finished first pass. Get rid of any added setjmp */
5842 PerlIO_printf(Perl_debug_log,
5843 "Required size %"IVdf" nodes\n"
5844 "Starting second pass (creation)\n",
5847 RExC_lastparse=NULL;
5850 /* The first pass could have found things that force Unicode semantics */
5851 if ((RExC_utf8 || RExC_uni_semantics)
5852 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
5854 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
5857 /* Small enough for pointer-storage convention?
5858 If extralen==0, this means that we will not need long jumps. */
5859 if (RExC_size >= 0x10000L && RExC_extralen)
5860 RExC_size += RExC_extralen;
5863 if (RExC_whilem_seen > 15)
5864 RExC_whilem_seen = 15;
5866 /* Allocate space and zero-initialize. Note, the two step process
5867 of zeroing when in debug mode, thus anything assigned has to
5868 happen after that */
5869 rx = (REGEXP*) newSV_type(SVt_REGEXP);
5870 r = (struct regexp*)SvANY(rx);
5871 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
5872 char, regexp_internal);
5873 if ( r == NULL || ri == NULL )
5874 FAIL("Regexp out of space");
5876 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
5877 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode), char);
5879 /* bulk initialize base fields with 0. */
5880 Zero(ri, sizeof(regexp_internal), char);
5883 /* non-zero initialization begins here */
5886 r->extflags = rx_flags;
5887 if (pm_flags & PMf_IS_QR) {
5888 ri->code_blocks = pRExC_state->code_blocks;
5889 ri->num_code_blocks = pRExC_state->num_code_blocks;
5892 SAVEFREEPV(pRExC_state->code_blocks);
5895 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
5896 bool has_charset = (get_regex_charset(r->extflags) != REGEX_DEPENDS_CHARSET);
5898 /* The caret is output if there are any defaults: if not all the STD
5899 * flags are set, or if no character set specifier is needed */
5901 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
5903 bool has_runon = ((RExC_seen & REG_SEEN_RUN_ON_COMMENT)==REG_SEEN_RUN_ON_COMMENT);
5904 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
5905 >> RXf_PMf_STD_PMMOD_SHIFT);
5906 const char *fptr = STD_PAT_MODS; /*"msix"*/
5908 /* Allocate for the worst case, which is all the std flags are turned
5909 * on. If more precision is desired, we could do a population count of
5910 * the flags set. This could be done with a small lookup table, or by
5911 * shifting, masking and adding, or even, when available, assembly
5912 * language for a machine-language population count.
5913 * We never output a minus, as all those are defaults, so are
5914 * covered by the caret */
5915 const STRLEN wraplen = plen + has_p + has_runon
5916 + has_default /* If needs a caret */
5918 /* If needs a character set specifier */
5919 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
5920 + (sizeof(STD_PAT_MODS) - 1)
5921 + (sizeof("(?:)") - 1);
5923 p = sv_grow(MUTABLE_SV(rx), wraplen + 1); /* +1 for the ending NUL */
5926 SvFLAGS(rx) |= SVf_UTF8;
5929 /* If a default, cover it using the caret */
5931 *p++= DEFAULT_PAT_MOD;
5935 const char* const name = get_regex_charset_name(r->extflags, &len);
5936 Copy(name, p, len, char);
5940 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
5943 while((ch = *fptr++)) {
5951 Copy(RExC_precomp, p, plen, char);
5952 assert ((RX_WRAPPED(rx) - p) < 16);
5953 r->pre_prefix = p - RX_WRAPPED(rx);
5959 SvCUR_set(rx, p - SvPVX_const(rx));
5963 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
5965 if (RExC_seen & REG_SEEN_RECURSE) {
5966 Newxz(RExC_open_parens, RExC_npar,regnode *);
5967 SAVEFREEPV(RExC_open_parens);
5968 Newxz(RExC_close_parens,RExC_npar,regnode *);
5969 SAVEFREEPV(RExC_close_parens);
5972 /* Useful during FAIL. */
5973 #ifdef RE_TRACK_PATTERN_OFFSETS
5974 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
5975 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
5976 "%s %"UVuf" bytes for offset annotations.\n",
5977 ri->u.offsets ? "Got" : "Couldn't get",
5978 (UV)((2*RExC_size+1) * sizeof(U32))));
5980 SetProgLen(ri,RExC_size);
5985 /* Second pass: emit code. */
5986 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
5987 RExC_pm_flags = pm_flags;
5992 RExC_emit_start = ri->program;
5993 RExC_emit = ri->program;
5994 RExC_emit_bound = ri->program + RExC_size + 1;
5995 pRExC_state->code_index = 0;
5997 REGC((U8)REG_MAGIC, (char*) RExC_emit++);
5998 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6002 /* XXXX To minimize changes to RE engine we always allocate
6003 3-units-long substrs field. */
6004 Newx(r->substrs, 1, struct reg_substr_data);
6005 if (RExC_recurse_count) {
6006 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6007 SAVEFREEPV(RExC_recurse);
6011 r->minlen = minlen = sawlookahead = sawplus = sawopen = 0;
6012 Zero(r->substrs, 1, struct reg_substr_data);
6014 #ifdef TRIE_STUDY_OPT
6016 StructCopy(&zero_scan_data, &data, scan_data_t);
6017 copyRExC_state = RExC_state;
6020 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6022 RExC_state = copyRExC_state;
6023 if (seen & REG_TOP_LEVEL_BRANCHES)
6024 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
6026 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES;
6027 if (data.last_found) {
6028 SvREFCNT_dec(data.longest_fixed);
6029 SvREFCNT_dec(data.longest_float);
6030 SvREFCNT_dec(data.last_found);
6032 StructCopy(&zero_scan_data, &data, scan_data_t);
6035 StructCopy(&zero_scan_data, &data, scan_data_t);
6038 /* Dig out information for optimizations. */
6039 r->extflags = RExC_flags; /* was pm_op */
6040 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6043 SvUTF8_on(rx); /* Unicode in it? */
6044 ri->regstclass = NULL;
6045 if (RExC_naughty >= 10) /* Probably an expensive pattern. */
6046 r->intflags |= PREGf_NAUGHTY;
6047 scan = ri->program + 1; /* First BRANCH. */
6049 /* testing for BRANCH here tells us whether there is "must appear"
6050 data in the pattern. If there is then we can use it for optimisations */
6051 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES)) { /* Only one top-level choice. */
6053 STRLEN longest_float_length, longest_fixed_length;
6054 struct regnode_charclass_class ch_class; /* pointed to by data */
6056 I32 last_close = 0; /* pointed to by data */
6057 regnode *first= scan;
6058 regnode *first_next= regnext(first);
6060 * Skip introductions and multiplicators >= 1
6061 * so that we can extract the 'meat' of the pattern that must
6062 * match in the large if() sequence following.
6063 * NOTE that EXACT is NOT covered here, as it is normally
6064 * picked up by the optimiser separately.
6066 * This is unfortunate as the optimiser isnt handling lookahead
6067 * properly currently.
6070 while ((OP(first) == OPEN && (sawopen = 1)) ||
6071 /* An OR of *one* alternative - should not happen now. */
6072 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6073 /* for now we can't handle lookbehind IFMATCH*/
6074 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6075 (OP(first) == PLUS) ||
6076 (OP(first) == MINMOD) ||
6077 /* An {n,m} with n>0 */
6078 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6079 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6082 * the only op that could be a regnode is PLUS, all the rest
6083 * will be regnode_1 or regnode_2.
6086 if (OP(first) == PLUS)
6089 first += regarglen[OP(first)];
6091 first = NEXTOPER(first);
6092 first_next= regnext(first);
6095 /* Starting-point info. */
6097 DEBUG_PEEP("first:",first,0);
6098 /* Ignore EXACT as we deal with it later. */
6099 if (PL_regkind[OP(first)] == EXACT) {
6100 if (OP(first) == EXACT)
6101 NOOP; /* Empty, get anchored substr later. */
6103 ri->regstclass = first;
6106 else if (PL_regkind[OP(first)] == TRIE &&
6107 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
6110 /* this can happen only on restudy */
6111 if ( OP(first) == TRIE ) {
6112 struct regnode_1 *trieop = (struct regnode_1 *)
6113 PerlMemShared_calloc(1, sizeof(struct regnode_1));
6114 StructCopy(first,trieop,struct regnode_1);
6115 trie_op=(regnode *)trieop;
6117 struct regnode_charclass *trieop = (struct regnode_charclass *)
6118 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
6119 StructCopy(first,trieop,struct regnode_charclass);
6120 trie_op=(regnode *)trieop;
6123 make_trie_failtable(pRExC_state, (regnode *)first, trie_op, 0);
6124 ri->regstclass = trie_op;
6127 else if (REGNODE_SIMPLE(OP(first)))
6128 ri->regstclass = first;
6129 else if (PL_regkind[OP(first)] == BOUND ||
6130 PL_regkind[OP(first)] == NBOUND)
6131 ri->regstclass = first;
6132 else if (PL_regkind[OP(first)] == BOL) {
6133 r->extflags |= (OP(first) == MBOL
6135 : (OP(first) == SBOL
6138 first = NEXTOPER(first);
6141 else if (OP(first) == GPOS) {
6142 r->extflags |= RXf_ANCH_GPOS;
6143 first = NEXTOPER(first);
6146 else if ((!sawopen || !RExC_sawback) &&
6147 (OP(first) == STAR &&
6148 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
6149 !(r->extflags & RXf_ANCH) && !pRExC_state->num_code_blocks)
6151 /* turn .* into ^.* with an implied $*=1 */
6153 (OP(NEXTOPER(first)) == REG_ANY)
6156 r->extflags |= type;
6157 r->intflags |= PREGf_IMPLICIT;
6158 first = NEXTOPER(first);
6161 if (sawplus && !sawlookahead && (!sawopen || !RExC_sawback)
6162 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
6163 /* x+ must match at the 1st pos of run of x's */
6164 r->intflags |= PREGf_SKIP;
6166 /* Scan is after the zeroth branch, first is atomic matcher. */
6167 #ifdef TRIE_STUDY_OPT
6170 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6171 (IV)(first - scan + 1))
6175 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
6176 (IV)(first - scan + 1))
6182 * If there's something expensive in the r.e., find the
6183 * longest literal string that must appear and make it the
6184 * regmust. Resolve ties in favor of later strings, since
6185 * the regstart check works with the beginning of the r.e.
6186 * and avoiding duplication strengthens checking. Not a
6187 * strong reason, but sufficient in the absence of others.
6188 * [Now we resolve ties in favor of the earlier string if
6189 * it happens that c_offset_min has been invalidated, since the
6190 * earlier string may buy us something the later one won't.]
6193 data.longest_fixed = newSVpvs("");
6194 data.longest_float = newSVpvs("");
6195 data.last_found = newSVpvs("");
6196 data.longest = &(data.longest_fixed);
6198 if (!ri->regstclass) {
6199 cl_init(pRExC_state, &ch_class);
6200 data.start_class = &ch_class;
6201 stclass_flag = SCF_DO_STCLASS_AND;
6202 } else /* XXXX Check for BOUND? */
6204 data.last_closep = &last_close;
6206 minlen = study_chunk(pRExC_state, &first, &minlen, &fake, scan + RExC_size, /* Up to end */
6207 &data, -1, NULL, NULL,
6208 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag,0);
6214 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
6215 && data.last_start_min == 0 && data.last_end > 0
6216 && !RExC_seen_zerolen
6217 && !(RExC_seen & REG_SEEN_VERBARG)
6218 && (!(RExC_seen & REG_SEEN_GPOS) || (r->extflags & RXf_ANCH_GPOS)))
6219 r->extflags |= RXf_CHECK_ALL;
6220 scan_commit(pRExC_state, &data,&minlen,0);
6221 SvREFCNT_dec(data.last_found);
6223 longest_float_length = CHR_SVLEN(data.longest_float);
6225 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
6226 && data.offset_fixed == data.offset_float_min
6227 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
6228 && S_setup_longest (aTHX_ pRExC_state,
6232 &(r->float_end_shift),
6233 data.lookbehind_float,
6234 data.offset_float_min,
6236 longest_float_length,
6237 data.flags & SF_FL_BEFORE_EOL,
6238 data.flags & SF_FL_BEFORE_MEOL))
6240 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
6241 r->float_max_offset = data.offset_float_max;
6242 if (data.offset_float_max < I32_MAX) /* Don't offset infinity */
6243 r->float_max_offset -= data.lookbehind_float;
6246 r->float_substr = r->float_utf8 = NULL;
6247 SvREFCNT_dec(data.longest_float);
6248 longest_float_length = 0;
6251 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
6253 if (S_setup_longest (aTHX_ pRExC_state,
6255 &(r->anchored_utf8),
6256 &(r->anchored_substr),
6257 &(r->anchored_end_shift),
6258 data.lookbehind_fixed,
6261 longest_fixed_length,
6262 data.flags & SF_FIX_BEFORE_EOL,
6263 data.flags & SF_FIX_BEFORE_MEOL))
6265 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
6268 r->anchored_substr = r->anchored_utf8 = NULL;
6269 SvREFCNT_dec(data.longest_fixed);
6270 longest_fixed_length = 0;
6274 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
6275 ri->regstclass = NULL;
6277 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
6279 && !(data.start_class->flags & ANYOF_EOS)
6280 && !cl_is_anything(data.start_class))
6282 const U32 n = add_data(pRExC_state, 1, "f");
6283 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
6285 Newx(RExC_rxi->data->data[n], 1,
6286 struct regnode_charclass_class);
6287 StructCopy(data.start_class,
6288 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6289 struct regnode_charclass_class);
6290 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6291 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6292 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
6293 regprop(r, sv, (regnode*)data.start_class);
6294 PerlIO_printf(Perl_debug_log,
6295 "synthetic stclass \"%s\".\n",
6296 SvPVX_const(sv));});
6299 /* A temporary algorithm prefers floated substr to fixed one to dig more info. */
6300 if (longest_fixed_length > longest_float_length) {
6301 r->check_end_shift = r->anchored_end_shift;
6302 r->check_substr = r->anchored_substr;
6303 r->check_utf8 = r->anchored_utf8;
6304 r->check_offset_min = r->check_offset_max = r->anchored_offset;
6305 if (r->extflags & RXf_ANCH_SINGLE)
6306 r->extflags |= RXf_NOSCAN;
6309 r->check_end_shift = r->float_end_shift;
6310 r->check_substr = r->float_substr;
6311 r->check_utf8 = r->float_utf8;
6312 r->check_offset_min = r->float_min_offset;
6313 r->check_offset_max = r->float_max_offset;
6315 /* XXXX Currently intuiting is not compatible with ANCH_GPOS.
6316 This should be changed ASAP! */
6317 if ((r->check_substr || r->check_utf8) && !(r->extflags & RXf_ANCH_GPOS)) {
6318 r->extflags |= RXf_USE_INTUIT;
6319 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
6320 r->extflags |= RXf_INTUIT_TAIL;
6322 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
6323 if ( (STRLEN)minlen < longest_float_length )
6324 minlen= longest_float_length;
6325 if ( (STRLEN)minlen < longest_fixed_length )
6326 minlen= longest_fixed_length;
6330 /* Several toplevels. Best we can is to set minlen. */
6332 struct regnode_charclass_class ch_class;
6335 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
6337 scan = ri->program + 1;
6338 cl_init(pRExC_state, &ch_class);
6339 data.start_class = &ch_class;
6340 data.last_closep = &last_close;
6343 minlen = study_chunk(pRExC_state, &scan, &minlen, &fake, scan + RExC_size,
6344 &data, -1, NULL, NULL, SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS,0);
6348 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
6349 = r->float_substr = r->float_utf8 = NULL;
6351 if (!(data.start_class->flags & ANYOF_EOS)
6352 && !cl_is_anything(data.start_class))
6354 const U32 n = add_data(pRExC_state, 1, "f");
6355 data.start_class->flags |= ANYOF_IS_SYNTHETIC;
6357 Newx(RExC_rxi->data->data[n], 1,
6358 struct regnode_charclass_class);
6359 StructCopy(data.start_class,
6360 (struct regnode_charclass_class*)RExC_rxi->data->data[n],
6361 struct regnode_charclass_class);
6362 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
6363 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
6364 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
6365 regprop(r, sv, (regnode*)data.start_class);
6366 PerlIO_printf(Perl_debug_log,
6367 "synthetic stclass \"%s\".\n",
6368 SvPVX_const(sv));});
6372 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
6373 the "real" pattern. */
6375 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf"\n",
6376 (IV)minlen, (IV)r->minlen);
6378 r->minlenret = minlen;
6379 if (r->minlen < minlen)
6382 if (RExC_seen & REG_SEEN_GPOS)
6383 r->extflags |= RXf_GPOS_SEEN;
6384 if (RExC_seen & REG_SEEN_LOOKBEHIND)
6385 r->extflags |= RXf_LOOKBEHIND_SEEN;
6386 if (pRExC_state->num_code_blocks)
6387 r->extflags |= RXf_EVAL_SEEN;
6388 if (RExC_seen & REG_SEEN_CANY)
6389 r->extflags |= RXf_CANY_SEEN;
6390 if (RExC_seen & REG_SEEN_VERBARG)
6391 r->intflags |= PREGf_VERBARG_SEEN;
6392 if (RExC_seen & REG_SEEN_CUTGROUP)
6393 r->intflags |= PREGf_CUTGROUP_SEEN;
6394 if (pm_flags & PMf_USE_RE_EVAL)
6395 r->intflags |= PREGf_USE_RE_EVAL;
6396 if (RExC_paren_names)
6397 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
6399 RXp_PAREN_NAMES(r) = NULL;
6401 #ifdef STUPID_PATTERN_CHECKS
6402 if (RX_PRELEN(rx) == 0)
6403 r->extflags |= RXf_NULL;
6404 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
6405 /* XXX: this should happen BEFORE we compile */
6406 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6407 else if (RX_PRELEN(rx) == 3 && memEQ("\\s+", RX_PRECOMP(rx), 3))
6408 r->extflags |= RXf_WHITE;
6409 else if (RX_PRELEN(rx) == 1 && RXp_PRECOMP(rx)[0] == '^')
6410 r->extflags |= RXf_START_ONLY;
6412 if (r->extflags & RXf_SPLIT && RX_PRELEN(rx) == 1 && RX_PRECOMP(rx)[0] == ' ')
6413 /* XXX: this should happen BEFORE we compile */
6414 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
6416 regnode *first = ri->program + 1;
6419 if (PL_regkind[fop] == NOTHING && OP(NEXTOPER(first)) == END)
6420 r->extflags |= RXf_NULL;
6421 else if (PL_regkind[fop] == BOL && OP(NEXTOPER(first)) == END)
6422 r->extflags |= RXf_START_ONLY;
6423 else if (fop == PLUS && OP(NEXTOPER(first)) == SPACE
6424 && OP(regnext(first)) == END)
6425 r->extflags |= RXf_WHITE;
6429 if (RExC_paren_names) {
6430 ri->name_list_idx = add_data( pRExC_state, 1, "a" );
6431 ri->data->data[ri->name_list_idx] = (void*)SvREFCNT_inc(RExC_paren_name_list);
6434 ri->name_list_idx = 0;
6436 if (RExC_recurse_count) {
6437 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
6438 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
6439 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
6442 Newxz(r->offs, RExC_npar, regexp_paren_pair);
6443 /* assume we don't need to swap parens around before we match */
6446 PerlIO_printf(Perl_debug_log,"Final program:\n");
6449 #ifdef RE_TRACK_PATTERN_OFFSETS
6450 DEBUG_OFFSETS_r(if (ri->u.offsets) {
6451 const U32 len = ri->u.offsets[0];
6453 GET_RE_DEBUG_FLAGS_DECL;
6454 PerlIO_printf(Perl_debug_log, "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
6455 for (i = 1; i <= len; i++) {
6456 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
6457 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
6458 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
6460 PerlIO_printf(Perl_debug_log, "\n");
6468 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
6471 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
6473 PERL_UNUSED_ARG(value);
6475 if (flags & RXapif_FETCH) {
6476 return reg_named_buff_fetch(rx, key, flags);
6477 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
6478 Perl_croak_no_modify(aTHX);
6480 } else if (flags & RXapif_EXISTS) {
6481 return reg_named_buff_exists(rx, key, flags)
6484 } else if (flags & RXapif_REGNAMES) {
6485 return reg_named_buff_all(rx, flags);
6486 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
6487 return reg_named_buff_scalar(rx, flags);
6489 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
6495 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
6498 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
6499 PERL_UNUSED_ARG(lastkey);
6501 if (flags & RXapif_FIRSTKEY)
6502 return reg_named_buff_firstkey(rx, flags);
6503 else if (flags & RXapif_NEXTKEY)
6504 return reg_named_buff_nextkey(rx, flags);
6506 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter", (int)flags);
6512 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
6515 AV *retarray = NULL;
6517 struct regexp *const rx = (struct regexp *)SvANY(r);
6519 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
6521 if (flags & RXapif_ALL)
6524 if (rx && RXp_PAREN_NAMES(rx)) {
6525 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
6528 SV* sv_dat=HeVAL(he_str);
6529 I32 *nums=(I32*)SvPVX(sv_dat);
6530 for ( i=0; i<SvIVX(sv_dat); i++ ) {
6531 if ((I32)(rx->nparens) >= nums[i]
6532 && rx->offs[nums[i]].start != -1
6533 && rx->offs[nums[i]].end != -1)
6536 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
6541 ret = newSVsv(&PL_sv_undef);
6544 av_push(retarray, ret);
6547 return newRV_noinc(MUTABLE_SV(retarray));
6554 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
6557 struct regexp *const rx = (struct regexp *)SvANY(r);
6559 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
6561 if (rx && RXp_PAREN_NAMES(rx)) {
6562 if (flags & RXapif_ALL) {
6563 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
6565 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
6579 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
6581 struct regexp *const rx = (struct regexp *)SvANY(r);
6583 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
6585 if ( rx && RXp_PAREN_NAMES(rx) ) {
6586 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
6588 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
6595 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
6597 struct regexp *const rx = (struct regexp *)SvANY(r);
6598 GET_RE_DEBUG_FLAGS_DECL;
6600 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
6602 if (rx && RXp_PAREN_NAMES(rx)) {
6603 HV *hv = RXp_PAREN_NAMES(rx);
6605 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6608 SV* sv_dat = HeVAL(temphe);
6609 I32 *nums = (I32*)SvPVX(sv_dat);
6610 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6611 if ((I32)(rx->lastparen) >= nums[i] &&
6612 rx->offs[nums[i]].start != -1 &&
6613 rx->offs[nums[i]].end != -1)
6619 if (parno || flags & RXapif_ALL) {
6620 return newSVhek(HeKEY_hek(temphe));
6628 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
6633 struct regexp *const rx = (struct regexp *)SvANY(r);
6635 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
6637 if (rx && RXp_PAREN_NAMES(rx)) {
6638 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
6639 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
6640 } else if (flags & RXapif_ONE) {
6641 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
6642 av = MUTABLE_AV(SvRV(ret));
6643 length = av_len(av);
6645 return newSViv(length + 1);
6647 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar", (int)flags);
6651 return &PL_sv_undef;
6655 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
6657 struct regexp *const rx = (struct regexp *)SvANY(r);
6660 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
6662 if (rx && RXp_PAREN_NAMES(rx)) {
6663 HV *hv= RXp_PAREN_NAMES(rx);
6665 (void)hv_iterinit(hv);
6666 while ( (temphe = hv_iternext_flags(hv,0)) ) {
6669 SV* sv_dat = HeVAL(temphe);
6670 I32 *nums = (I32*)SvPVX(sv_dat);
6671 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
6672 if ((I32)(rx->lastparen) >= nums[i] &&
6673 rx->offs[nums[i]].start != -1 &&
6674 rx->offs[nums[i]].end != -1)
6680 if (parno || flags & RXapif_ALL) {
6681 av_push(av, newSVhek(HeKEY_hek(temphe)));
6686 return newRV_noinc(MUTABLE_SV(av));
6690 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
6693 struct regexp *const rx = (struct regexp *)SvANY(r);
6699 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
6701 if ( ( n == RX_BUFF_IDX_CARET_PREMATCH
6702 || n == RX_BUFF_IDX_CARET_FULLMATCH
6703 || n == RX_BUFF_IDX_CARET_POSTMATCH
6705 && !(rx->extflags & RXf_PMf_KEEPCOPY)
6712 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
6713 /* no need to distinguish between them any more */
6714 n = RX_BUFF_IDX_FULLMATCH;
6716 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
6717 && rx->offs[0].start != -1)
6719 /* $`, ${^PREMATCH} */
6720 i = rx->offs[0].start;
6724 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
6725 && rx->offs[0].end != -1)
6727 /* $', ${^POSTMATCH} */
6728 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
6729 i = rx->sublen + rx->suboffset - rx->offs[0].end;
6732 if ( 0 <= n && n <= (I32)rx->nparens &&
6733 (s1 = rx->offs[n].start) != -1 &&
6734 (t1 = rx->offs[n].end) != -1)
6736 /* $&, ${^MATCH}, $1 ... */
6738 s = rx->subbeg + s1 - rx->suboffset;
6743 assert(s >= rx->subbeg);
6744 assert(rx->sublen >= (s - rx->subbeg) + i );
6746 const int oldtainted = PL_tainted;
6748 sv_setpvn(sv, s, i);
6749 PL_tainted = oldtainted;
6750 if ( (rx->extflags & RXf_CANY_SEEN)
6751 ? (RXp_MATCH_UTF8(rx)
6752 && (!i || is_utf8_string((U8*)s, i)))
6753 : (RXp_MATCH_UTF8(rx)) )
6760 if (RXp_MATCH_TAINTED(rx)) {
6761 if (SvTYPE(sv) >= SVt_PVMG) {
6762 MAGIC* const mg = SvMAGIC(sv);
6765 SvMAGIC_set(sv, mg->mg_moremagic);
6767 if ((mgt = SvMAGIC(sv))) {
6768 mg->mg_moremagic = mgt;
6769 SvMAGIC_set(sv, mg);
6780 sv_setsv(sv,&PL_sv_undef);
6786 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
6787 SV const * const value)
6789 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
6791 PERL_UNUSED_ARG(rx);
6792 PERL_UNUSED_ARG(paren);
6793 PERL_UNUSED_ARG(value);
6796 Perl_croak_no_modify(aTHX);
6800 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
6803 struct regexp *const rx = (struct regexp *)SvANY(r);
6807 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
6809 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
6811 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
6812 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6816 case RX_BUFF_IDX_PREMATCH: /* $` */
6817 if (rx->offs[0].start != -1) {
6818 i = rx->offs[0].start;
6827 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
6828 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6830 case RX_BUFF_IDX_POSTMATCH: /* $' */
6831 if (rx->offs[0].end != -1) {
6832 i = rx->sublen - rx->offs[0].end;
6834 s1 = rx->offs[0].end;
6841 case RX_BUFF_IDX_CARET_FULLMATCH: /* ${^MATCH} */
6842 if (!(rx->extflags & RXf_PMf_KEEPCOPY))
6846 /* $& / ${^MATCH}, $1, $2, ... */
6848 if (paren <= (I32)rx->nparens &&
6849 (s1 = rx->offs[paren].start) != -1 &&
6850 (t1 = rx->offs[paren].end) != -1)
6856 if (ckWARN(WARN_UNINITIALIZED))
6857 report_uninit((const SV *)sv);
6862 if (i > 0 && RXp_MATCH_UTF8(rx)) {
6863 const char * const s = rx->subbeg - rx->suboffset + s1;
6868 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
6875 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
6877 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
6878 PERL_UNUSED_ARG(rx);
6882 return newSVpvs("Regexp");
6885 /* Scans the name of a named buffer from the pattern.
6886 * If flags is REG_RSN_RETURN_NULL returns null.
6887 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
6888 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
6889 * to the parsed name as looked up in the RExC_paren_names hash.
6890 * If there is an error throws a vFAIL().. type exception.
6893 #define REG_RSN_RETURN_NULL 0
6894 #define REG_RSN_RETURN_NAME 1
6895 #define REG_RSN_RETURN_DATA 2
6898 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
6900 char *name_start = RExC_parse;
6902 PERL_ARGS_ASSERT_REG_SCAN_NAME;
6904 if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
6905 /* skip IDFIRST by using do...while */
6908 RExC_parse += UTF8SKIP(RExC_parse);
6909 } while (isALNUM_utf8((U8*)RExC_parse));
6913 } while (isALNUM(*RExC_parse));
6915 RExC_parse++; /* so the <- from the vFAIL is after the offending character */
6916 vFAIL("Group name must start with a non-digit word character");
6920 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
6921 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
6922 if ( flags == REG_RSN_RETURN_NAME)
6924 else if (flags==REG_RSN_RETURN_DATA) {
6927 if ( ! sv_name ) /* should not happen*/
6928 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
6929 if (RExC_paren_names)
6930 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
6932 sv_dat = HeVAL(he_str);
6934 vFAIL("Reference to nonexistent named group");
6938 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
6939 (unsigned long) flags);
6941 assert(0); /* NOT REACHED */
6946 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
6947 int rem=(int)(RExC_end - RExC_parse); \
6956 if (RExC_lastparse!=RExC_parse) \
6957 PerlIO_printf(Perl_debug_log," >%.*s%-*s", \
6960 iscut ? "..." : "<" \
6963 PerlIO_printf(Perl_debug_log,"%16s",""); \
6966 num = RExC_size + 1; \
6968 num=REG_NODE_NUM(RExC_emit); \
6969 if (RExC_lastnum!=num) \
6970 PerlIO_printf(Perl_debug_log,"|%4d",num); \
6972 PerlIO_printf(Perl_debug_log,"|%4s",""); \
6973 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
6974 (int)((depth*2)), "", \
6978 RExC_lastparse=RExC_parse; \
6983 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
6984 DEBUG_PARSE_MSG((funcname)); \
6985 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
6987 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
6988 DEBUG_PARSE_MSG((funcname)); \
6989 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
6992 /* This section of code defines the inversion list object and its methods. The
6993 * interfaces are highly subject to change, so as much as possible is static to
6994 * this file. An inversion list is here implemented as a malloc'd C UV array
6995 * with some added info that is placed as UVs at the beginning in a header
6996 * portion. An inversion list for Unicode is an array of code points, sorted
6997 * by ordinal number. The zeroth element is the first code point in the list.
6998 * The 1th element is the first element beyond that not in the list. In other
6999 * words, the first range is
7000 * invlist[0]..(invlist[1]-1)
7001 * The other ranges follow. Thus every element whose index is divisible by two
7002 * marks the beginning of a range that is in the list, and every element not
7003 * divisible by two marks the beginning of a range not in the list. A single
7004 * element inversion list that contains the single code point N generally
7005 * consists of two elements
7008 * (The exception is when N is the highest representable value on the
7009 * machine, in which case the list containing just it would be a single
7010 * element, itself. By extension, if the last range in the list extends to
7011 * infinity, then the first element of that range will be in the inversion list
7012 * at a position that is divisible by two, and is the final element in the
7014 * Taking the complement (inverting) an inversion list is quite simple, if the
7015 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7016 * This implementation reserves an element at the beginning of each inversion
7017 * list to contain 0 when the list contains 0, and contains 1 otherwise. The
7018 * actual beginning of the list is either that element if 0, or the next one if
7021 * More about inversion lists can be found in "Unicode Demystified"
7022 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7023 * More will be coming when functionality is added later.
7025 * The inversion list data structure is currently implemented as an SV pointing
7026 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7027 * array of UV whose memory management is automatically handled by the existing
7028 * facilities for SV's.
7030 * Some of the methods should always be private to the implementation, and some
7031 * should eventually be made public */
7033 /* The header definitions are in F<inline_invlist.c> */
7035 #define TO_INTERNAL_SIZE(x) ((x + HEADER_LENGTH) * sizeof(UV))
7036 #define FROM_INTERNAL_SIZE(x) ((x / sizeof(UV)) - HEADER_LENGTH)
7038 #define INVLIST_INITIAL_LEN 10
7040 PERL_STATIC_INLINE UV*
7041 S__invlist_array_init(pTHX_ SV* const invlist, const bool will_have_0)
7043 /* Returns a pointer to the first element in the inversion list's array.
7044 * This is called upon initialization of an inversion list. Where the
7045 * array begins depends on whether the list has the code point U+0000
7046 * in it or not. The other parameter tells it whether the code that
7047 * follows this call is about to put a 0 in the inversion list or not.
7048 * The first element is either the element with 0, if 0, or the next one,
7051 UV* zero = get_invlist_zero_addr(invlist);
7053 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
7056 assert(! *_get_invlist_len_addr(invlist));
7058 /* 1^1 = 0; 1^0 = 1 */
7059 *zero = 1 ^ will_have_0;
7060 return zero + *zero;
7063 PERL_STATIC_INLINE UV*
7064 S_invlist_array(pTHX_ SV* const invlist)
7066 /* Returns the pointer to the inversion list's array. Every time the
7067 * length changes, this needs to be called in case malloc or realloc moved
7070 PERL_ARGS_ASSERT_INVLIST_ARRAY;
7072 /* Must not be empty. If these fail, you probably didn't check for <len>
7073 * being non-zero before trying to get the array */
7074 assert(*_get_invlist_len_addr(invlist));
7075 assert(*get_invlist_zero_addr(invlist) == 0
7076 || *get_invlist_zero_addr(invlist) == 1);
7078 /* The array begins either at the element reserved for zero if the
7079 * list contains 0 (that element will be set to 0), or otherwise the next
7080 * element (in which case the reserved element will be set to 1). */
7081 return (UV *) (get_invlist_zero_addr(invlist)
7082 + *get_invlist_zero_addr(invlist));
7085 PERL_STATIC_INLINE void
7086 S_invlist_set_len(pTHX_ SV* const invlist, const UV len)
7088 /* Sets the current number of elements stored in the inversion list */
7090 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
7092 *_get_invlist_len_addr(invlist) = len;
7094 assert(len <= SvLEN(invlist));
7096 SvCUR_set(invlist, TO_INTERNAL_SIZE(len));
7097 /* If the list contains U+0000, that element is part of the header,
7098 * and should not be counted as part of the array. It will contain
7099 * 0 in that case, and 1 otherwise. So we could flop 0=>1, 1=>0 and
7101 * SvCUR_set(invlist,
7102 * TO_INTERNAL_SIZE(len
7103 * - (*get_invlist_zero_addr(inv_list) ^ 1)));
7104 * But, this is only valid if len is not 0. The consequences of not doing
7105 * this is that the memory allocation code may think that 1 more UV is
7106 * being used than actually is, and so might do an unnecessary grow. That
7107 * seems worth not bothering to make this the precise amount.
7109 * Note that when inverting, SvCUR shouldn't change */
7112 PERL_STATIC_INLINE IV*
7113 S_get_invlist_previous_index_addr(pTHX_ SV* invlist)
7115 /* Return the address of the UV that is reserved to hold the cached index
7118 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
7120 return (IV *) (SvPVX(invlist) + (INVLIST_PREVIOUS_INDEX_OFFSET * sizeof (UV)));
7123 PERL_STATIC_INLINE IV
7124 S_invlist_previous_index(pTHX_ SV* const invlist)
7126 /* Returns cached index of previous search */
7128 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
7130 return *get_invlist_previous_index_addr(invlist);
7133 PERL_STATIC_INLINE void
7134 S_invlist_set_previous_index(pTHX_ SV* const invlist, const IV index)
7136 /* Caches <index> for later retrieval */
7138 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
7140 assert(index == 0 || index < (int) _invlist_len(invlist));
7142 *get_invlist_previous_index_addr(invlist) = index;
7145 PERL_STATIC_INLINE UV
7146 S_invlist_max(pTHX_ SV* const invlist)
7148 /* Returns the maximum number of elements storable in the inversion list's
7149 * array, without having to realloc() */
7151 PERL_ARGS_ASSERT_INVLIST_MAX;
7153 return FROM_INTERNAL_SIZE(SvLEN(invlist));
7156 PERL_STATIC_INLINE UV*
7157 S_get_invlist_zero_addr(pTHX_ SV* invlist)
7159 /* Return the address of the UV that is reserved to hold 0 if the inversion
7160 * list contains 0. This has to be the last element of the heading, as the
7161 * list proper starts with either it if 0, or the next element if not.
7162 * (But we force it to contain either 0 or 1) */
7164 PERL_ARGS_ASSERT_GET_INVLIST_ZERO_ADDR;
7166 return (UV *) (SvPVX(invlist) + (INVLIST_ZERO_OFFSET * sizeof (UV)));
7169 #ifndef PERL_IN_XSUB_RE
7171 Perl__new_invlist(pTHX_ IV initial_size)
7174 /* Return a pointer to a newly constructed inversion list, with enough
7175 * space to store 'initial_size' elements. If that number is negative, a
7176 * system default is used instead */
7180 if (initial_size < 0) {
7181 initial_size = INVLIST_INITIAL_LEN;
7184 /* Allocate the initial space */
7185 new_list = newSV(TO_INTERNAL_SIZE(initial_size));
7186 invlist_set_len(new_list, 0);
7188 /* Force iterinit() to be used to get iteration to work */
7189 *get_invlist_iter_addr(new_list) = UV_MAX;
7191 /* This should force a segfault if a method doesn't initialize this
7193 *get_invlist_zero_addr(new_list) = UV_MAX;
7195 *get_invlist_previous_index_addr(new_list) = 0;
7196 *get_invlist_version_id_addr(new_list) = INVLIST_VERSION_ID;
7197 #if HEADER_LENGTH != 5
7198 # 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
7206 S__new_invlist_C_array(pTHX_ UV* list)
7208 /* Return a pointer to a newly constructed inversion list, initialized to
7209 * point to <list>, which has to be in the exact correct inversion list
7210 * form, including internal fields. Thus this is a dangerous routine that
7211 * should not be used in the wrong hands */
7213 SV* invlist = newSV_type(SVt_PV);
7215 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
7217 SvPV_set(invlist, (char *) list);
7218 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
7219 shouldn't touch it */
7220 SvCUR_set(invlist, TO_INTERNAL_SIZE(_invlist_len(invlist)));
7222 if (*get_invlist_version_id_addr(invlist) != INVLIST_VERSION_ID) {
7223 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
7230 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
7232 /* Grow the maximum size of an inversion list */
7234 PERL_ARGS_ASSERT_INVLIST_EXTEND;
7236 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max));
7239 PERL_STATIC_INLINE void
7240 S_invlist_trim(pTHX_ SV* const invlist)
7242 PERL_ARGS_ASSERT_INVLIST_TRIM;
7244 /* Change the length of the inversion list to how many entries it currently
7247 SvPV_shrink_to_cur((SV *) invlist);
7250 #define _invlist_union_complement_2nd(a, b, output) _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
7253 S__append_range_to_invlist(pTHX_ SV* const invlist, const UV start, const UV end)
7255 /* Subject to change or removal. Append the range from 'start' to 'end' at
7256 * the end of the inversion list. The range must be above any existing
7260 UV max = invlist_max(invlist);
7261 UV len = _invlist_len(invlist);
7263 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
7265 if (len == 0) { /* Empty lists must be initialized */
7266 array = _invlist_array_init(invlist, start == 0);
7269 /* Here, the existing list is non-empty. The current max entry in the
7270 * list is generally the first value not in the set, except when the
7271 * set extends to the end of permissible values, in which case it is
7272 * the first entry in that final set, and so this call is an attempt to
7273 * append out-of-order */
7275 UV final_element = len - 1;
7276 array = invlist_array(invlist);
7277 if (array[final_element] > start
7278 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
7280 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",
7281 array[final_element], start,
7282 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
7285 /* Here, it is a legal append. If the new range begins with the first
7286 * value not in the set, it is extending the set, so the new first
7287 * value not in the set is one greater than the newly extended range.
7289 if (array[final_element] == start) {
7290 if (end != UV_MAX) {
7291 array[final_element] = end + 1;
7294 /* But if the end is the maximum representable on the machine,
7295 * just let the range that this would extend to have no end */
7296 invlist_set_len(invlist, len - 1);
7302 /* Here the new range doesn't extend any existing set. Add it */
7304 len += 2; /* Includes an element each for the start and end of range */
7306 /* If overflows the existing space, extend, which may cause the array to be
7309 invlist_extend(invlist, len);
7310 invlist_set_len(invlist, len); /* Have to set len here to avoid assert
7311 failure in invlist_array() */
7312 array = invlist_array(invlist);
7315 invlist_set_len(invlist, len);
7318 /* The next item on the list starts the range, the one after that is
7319 * one past the new range. */
7320 array[len - 2] = start;
7321 if (end != UV_MAX) {
7322 array[len - 1] = end + 1;
7325 /* But if the end is the maximum representable on the machine, just let
7326 * the range have no end */
7327 invlist_set_len(invlist, len - 1);
7331 #ifndef PERL_IN_XSUB_RE
7334 Perl__invlist_search(pTHX_ SV* const invlist, const UV cp)
7336 /* Searches the inversion list for the entry that contains the input code
7337 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
7338 * return value is the index into the list's array of the range that
7343 IV high = _invlist_len(invlist);
7344 const IV highest_element = high - 1;
7347 PERL_ARGS_ASSERT__INVLIST_SEARCH;
7349 /* If list is empty, return failure. */
7354 /* If the code point is before the first element, return failure. (We
7355 * can't combine this with the test above, because we can't get the array
7356 * unless we know the list is non-empty) */
7357 array = invlist_array(invlist);
7359 mid = invlist_previous_index(invlist);
7360 assert(mid >=0 && mid <= highest_element);
7362 /* <mid> contains the cache of the result of the previous call to this
7363 * function (0 the first time). See if this call is for the same result,
7364 * or if it is for mid-1. This is under the theory that calls to this
7365 * function will often be for related code points that are near each other.
7366 * And benchmarks show that caching gives better results. We also test
7367 * here if the code point is within the bounds of the list. These tests
7368 * replace others that would have had to be made anyway to make sure that
7369 * the array bounds were not exceeded, and give us extra information at the
7371 if (cp >= array[mid]) {
7372 if (cp >= array[highest_element]) {
7373 return highest_element;
7376 /* Here, array[mid] <= cp < array[highest_element]. This means that
7377 * the final element is not the answer, so can exclude it; it also
7378 * means that <mid> is not the final element, so can refer to 'mid + 1'
7380 if (cp < array[mid + 1]) {
7386 else { /* cp < aray[mid] */
7387 if (cp < array[0]) { /* Fail if outside the array */
7391 if (cp >= array[mid - 1]) {
7396 /* Binary search. What we are looking for is <i> such that
7397 * array[i] <= cp < array[i+1]
7398 * The loop below converges on the i+1. Note that there may not be an
7399 * (i+1)th element in the array, and things work nonetheless */
7400 while (low < high) {
7401 mid = (low + high) / 2;
7402 assert(mid <= highest_element);
7403 if (array[mid] <= cp) { /* cp >= array[mid] */
7406 /* We could do this extra test to exit the loop early.
7407 if (cp < array[low]) {
7412 else { /* cp < array[mid] */
7419 invlist_set_previous_index(invlist, high);
7424 Perl__invlist_populate_swatch(pTHX_ SV* const invlist, const UV start, const UV end, U8* swatch)
7426 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
7427 * but is used when the swash has an inversion list. This makes this much
7428 * faster, as it uses a binary search instead of a linear one. This is
7429 * intimately tied to that function, and perhaps should be in utf8.c,
7430 * except it is intimately tied to inversion lists as well. It assumes
7431 * that <swatch> is all 0's on input */
7434 const IV len = _invlist_len(invlist);
7438 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
7440 if (len == 0) { /* Empty inversion list */
7444 array = invlist_array(invlist);
7446 /* Find which element it is */
7447 i = _invlist_search(invlist, start);
7449 /* We populate from <start> to <end> */
7450 while (current < end) {
7453 /* The inversion list gives the results for every possible code point
7454 * after the first one in the list. Only those ranges whose index is
7455 * even are ones that the inversion list matches. For the odd ones,
7456 * and if the initial code point is not in the list, we have to skip
7457 * forward to the next element */
7458 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
7460 if (i >= len) { /* Finished if beyond the end of the array */
7464 if (current >= end) { /* Finished if beyond the end of what we
7466 if (LIKELY(end < UV_MAX)) {
7470 /* We get here when the upper bound is the maximum
7471 * representable on the machine, and we are looking for just
7472 * that code point. Have to special case it */
7474 goto join_end_of_list;
7477 assert(current >= start);
7479 /* The current range ends one below the next one, except don't go past
7482 upper = (i < len && array[i] < end) ? array[i] : end;
7484 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
7485 * for each code point in it */
7486 for (; current < upper; current++) {
7487 const STRLEN offset = (STRLEN)(current - start);
7488 swatch[offset >> 3] |= 1 << (offset & 7);
7493 /* Quit if at the end of the list */
7496 /* But first, have to deal with the highest possible code point on
7497 * the platform. The previous code assumes that <end> is one
7498 * beyond where we want to populate, but that is impossible at the
7499 * platform's infinity, so have to handle it specially */
7500 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
7502 const STRLEN offset = (STRLEN)(end - start);
7503 swatch[offset >> 3] |= 1 << (offset & 7);
7508 /* Advance to the next range, which will be for code points not in the
7517 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** output)
7519 /* Take the union of two inversion lists and point <output> to it. *output
7520 * should be defined upon input, and if it points to one of the two lists,
7521 * the reference count to that list will be decremented. The first list,
7522 * <a>, may be NULL, in which case a copy of the second list is returned.
7523 * If <complement_b> is TRUE, the union is taken of the complement
7524 * (inversion) of <b> instead of b itself.
7526 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7527 * Richard Gillam, published by Addison-Wesley, and explained at some
7528 * length there. The preface says to incorporate its examples into your
7529 * code at your own risk.
7531 * The algorithm is like a merge sort.
7533 * XXX A potential performance improvement is to keep track as we go along
7534 * if only one of the inputs contributes to the result, meaning the other
7535 * is a subset of that one. In that case, we can skip the final copy and
7536 * return the larger of the input lists, but then outside code might need
7537 * to keep track of whether to free the input list or not */
7539 UV* array_a; /* a's array */
7541 UV len_a; /* length of a's array */
7544 SV* u; /* the resulting union */
7548 UV i_a = 0; /* current index into a's array */
7552 /* running count, as explained in the algorithm source book; items are
7553 * stopped accumulating and are output when the count changes to/from 0.
7554 * The count is incremented when we start a range that's in the set, and
7555 * decremented when we start a range that's not in the set. So its range
7556 * is 0 to 2. Only when the count is zero is something not in the set.
7560 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
7563 /* If either one is empty, the union is the other one */
7564 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
7571 *output = invlist_clone(b);
7573 _invlist_invert(*output);
7575 } /* else *output already = b; */
7578 else if ((len_b = _invlist_len(b)) == 0) {
7583 /* The complement of an empty list is a list that has everything in it,
7584 * so the union with <a> includes everything too */
7589 *output = _new_invlist(1);
7590 _append_range_to_invlist(*output, 0, UV_MAX);
7592 else if (*output != a) {
7593 *output = invlist_clone(a);
7595 /* else *output already = a; */
7599 /* Here both lists exist and are non-empty */
7600 array_a = invlist_array(a);
7601 array_b = invlist_array(b);
7603 /* If are to take the union of 'a' with the complement of b, set it
7604 * up so are looking at b's complement. */
7607 /* To complement, we invert: if the first element is 0, remove it. To
7608 * do this, we just pretend the array starts one later, and clear the
7609 * flag as we don't have to do anything else later */
7610 if (array_b[0] == 0) {
7613 complement_b = FALSE;
7617 /* But if the first element is not zero, we unshift a 0 before the
7618 * array. The data structure reserves a space for that 0 (which
7619 * should be a '1' right now), so physical shifting is unneeded,
7620 * but temporarily change that element to 0. Before exiting the
7621 * routine, we must restore the element to '1' */
7628 /* Size the union for the worst case: that the sets are completely
7630 u = _new_invlist(len_a + len_b);
7632 /* Will contain U+0000 if either component does */
7633 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
7634 || (len_b > 0 && array_b[0] == 0));
7636 /* Go through each list item by item, stopping when exhausted one of
7638 while (i_a < len_a && i_b < len_b) {
7639 UV cp; /* The element to potentially add to the union's array */
7640 bool cp_in_set; /* is it in the the input list's set or not */
7642 /* We need to take one or the other of the two inputs for the union.
7643 * Since we are merging two sorted lists, we take the smaller of the
7644 * next items. In case of a tie, we take the one that is in its set
7645 * first. If we took one not in the set first, it would decrement the
7646 * count, possibly to 0 which would cause it to be output as ending the
7647 * range, and the next time through we would take the same number, and
7648 * output it again as beginning the next range. By doing it the
7649 * opposite way, there is no possibility that the count will be
7650 * momentarily decremented to 0, and thus the two adjoining ranges will
7651 * be seamlessly merged. (In a tie and both are in the set or both not
7652 * in the set, it doesn't matter which we take first.) */
7653 if (array_a[i_a] < array_b[i_b]
7654 || (array_a[i_a] == array_b[i_b]
7655 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7657 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7661 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7665 /* Here, have chosen which of the two inputs to look at. Only output
7666 * if the running count changes to/from 0, which marks the
7667 * beginning/end of a range in that's in the set */
7670 array_u[i_u++] = cp;
7677 array_u[i_u++] = cp;
7682 /* Here, we are finished going through at least one of the lists, which
7683 * means there is something remaining in at most one. We check if the list
7684 * that hasn't been exhausted is positioned such that we are in the middle
7685 * of a range in its set or not. (i_a and i_b point to the element beyond
7686 * the one we care about.) If in the set, we decrement 'count'; if 0, there
7687 * is potentially more to output.
7688 * There are four cases:
7689 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
7690 * in the union is entirely from the non-exhausted set.
7691 * 2) Both were in their sets, count is 2. Nothing further should
7692 * be output, as everything that remains will be in the exhausted
7693 * list's set, hence in the union; decrementing to 1 but not 0 insures
7695 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
7696 * Nothing further should be output because the union includes
7697 * everything from the exhausted set. Not decrementing ensures that.
7698 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
7699 * decrementing to 0 insures that we look at the remainder of the
7700 * non-exhausted set */
7701 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7702 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7707 /* The final length is what we've output so far, plus what else is about to
7708 * be output. (If 'count' is non-zero, then the input list we exhausted
7709 * has everything remaining up to the machine's limit in its set, and hence
7710 * in the union, so there will be no further output. */
7713 /* At most one of the subexpressions will be non-zero */
7714 len_u += (len_a - i_a) + (len_b - i_b);
7717 /* Set result to final length, which can change the pointer to array_u, so
7719 if (len_u != _invlist_len(u)) {
7720 invlist_set_len(u, len_u);
7722 array_u = invlist_array(u);
7725 /* When 'count' is 0, the list that was exhausted (if one was shorter than
7726 * the other) ended with everything above it not in its set. That means
7727 * that the remaining part of the union is precisely the same as the
7728 * non-exhausted list, so can just copy it unchanged. (If both list were
7729 * exhausted at the same time, then the operations below will be both 0.)
7732 IV copy_count; /* At most one will have a non-zero copy count */
7733 if ((copy_count = len_a - i_a) > 0) {
7734 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
7736 else if ((copy_count = len_b - i_b) > 0) {
7737 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
7741 /* We may be removing a reference to one of the inputs */
7742 if (a == *output || b == *output) {
7743 SvREFCNT_dec(*output);
7746 /* If we've changed b, restore it */
7756 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b, bool complement_b, SV** i)
7758 /* Take the intersection of two inversion lists and point <i> to it. *i
7759 * should be defined upon input, and if it points to one of the two lists,
7760 * the reference count to that list will be decremented.
7761 * If <complement_b> is TRUE, the result will be the intersection of <a>
7762 * and the complement (or inversion) of <b> instead of <b> directly.
7764 * The basis for this comes from "Unicode Demystified" Chapter 13 by
7765 * Richard Gillam, published by Addison-Wesley, and explained at some
7766 * length there. The preface says to incorporate its examples into your
7767 * code at your own risk. In fact, it had bugs
7769 * The algorithm is like a merge sort, and is essentially the same as the
7773 UV* array_a; /* a's array */
7775 UV len_a; /* length of a's array */
7778 SV* r; /* the resulting intersection */
7782 UV i_a = 0; /* current index into a's array */
7786 /* running count, as explained in the algorithm source book; items are
7787 * stopped accumulating and are output when the count changes to/from 2.
7788 * The count is incremented when we start a range that's in the set, and
7789 * decremented when we start a range that's not in the set. So its range
7790 * is 0 to 2. Only when the count is 2 is something in the intersection.
7794 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
7797 /* Special case if either one is empty */
7798 len_a = _invlist_len(a);
7799 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
7801 if (len_a != 0 && complement_b) {
7803 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
7804 * be empty. Here, also we are using 'b's complement, which hence
7805 * must be every possible code point. Thus the intersection is
7808 *i = invlist_clone(a);
7814 /* else *i is already 'a' */
7818 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
7819 * intersection must be empty */
7826 *i = _new_invlist(0);
7830 /* Here both lists exist and are non-empty */
7831 array_a = invlist_array(a);
7832 array_b = invlist_array(b);
7834 /* If are to take the intersection of 'a' with the complement of b, set it
7835 * up so are looking at b's complement. */
7838 /* To complement, we invert: if the first element is 0, remove it. To
7839 * do this, we just pretend the array starts one later, and clear the
7840 * flag as we don't have to do anything else later */
7841 if (array_b[0] == 0) {
7844 complement_b = FALSE;
7848 /* But if the first element is not zero, we unshift a 0 before the
7849 * array. The data structure reserves a space for that 0 (which
7850 * should be a '1' right now), so physical shifting is unneeded,
7851 * but temporarily change that element to 0. Before exiting the
7852 * routine, we must restore the element to '1' */
7859 /* Size the intersection for the worst case: that the intersection ends up
7860 * fragmenting everything to be completely disjoint */
7861 r= _new_invlist(len_a + len_b);
7863 /* Will contain U+0000 iff both components do */
7864 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
7865 && len_b > 0 && array_b[0] == 0);
7867 /* Go through each list item by item, stopping when exhausted one of
7869 while (i_a < len_a && i_b < len_b) {
7870 UV cp; /* The element to potentially add to the intersection's
7872 bool cp_in_set; /* Is it in the input list's set or not */
7874 /* We need to take one or the other of the two inputs for the
7875 * intersection. Since we are merging two sorted lists, we take the
7876 * smaller of the next items. In case of a tie, we take the one that
7877 * is not in its set first (a difference from the union algorithm). If
7878 * we took one in the set first, it would increment the count, possibly
7879 * to 2 which would cause it to be output as starting a range in the
7880 * intersection, and the next time through we would take that same
7881 * number, and output it again as ending the set. By doing it the
7882 * opposite of this, there is no possibility that the count will be
7883 * momentarily incremented to 2. (In a tie and both are in the set or
7884 * both not in the set, it doesn't matter which we take first.) */
7885 if (array_a[i_a] < array_b[i_b]
7886 || (array_a[i_a] == array_b[i_b]
7887 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
7889 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
7893 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
7897 /* Here, have chosen which of the two inputs to look at. Only output
7898 * if the running count changes to/from 2, which marks the
7899 * beginning/end of a range that's in the intersection */
7903 array_r[i_r++] = cp;
7908 array_r[i_r++] = cp;
7914 /* Here, we are finished going through at least one of the lists, which
7915 * means there is something remaining in at most one. We check if the list
7916 * that has been exhausted is positioned such that we are in the middle
7917 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
7918 * the ones we care about.) There are four cases:
7919 * 1) Both weren't in their sets, count is 0, and remains 0. There's
7920 * nothing left in the intersection.
7921 * 2) Both were in their sets, count is 2 and perhaps is incremented to
7922 * above 2. What should be output is exactly that which is in the
7923 * non-exhausted set, as everything it has is also in the intersection
7924 * set, and everything it doesn't have can't be in the intersection
7925 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
7926 * gets incremented to 2. Like the previous case, the intersection is
7927 * everything that remains in the non-exhausted set.
7928 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
7929 * remains 1. And the intersection has nothing more. */
7930 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
7931 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
7936 /* The final length is what we've output so far plus what else is in the
7937 * intersection. At most one of the subexpressions below will be non-zero */
7940 len_r += (len_a - i_a) + (len_b - i_b);
7943 /* Set result to final length, which can change the pointer to array_r, so
7945 if (len_r != _invlist_len(r)) {
7946 invlist_set_len(r, len_r);
7948 array_r = invlist_array(r);
7951 /* Finish outputting any remaining */
7952 if (count >= 2) { /* At most one will have a non-zero copy count */
7954 if ((copy_count = len_a - i_a) > 0) {
7955 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
7957 else if ((copy_count = len_b - i_b) > 0) {
7958 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
7962 /* We may be removing a reference to one of the inputs */
7963 if (a == *i || b == *i) {
7967 /* If we've changed b, restore it */
7977 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
7979 /* Add the range from 'start' to 'end' inclusive to the inversion list's
7980 * set. A pointer to the inversion list is returned. This may actually be
7981 * a new list, in which case the passed in one has been destroyed. The
7982 * passed in inversion list can be NULL, in which case a new one is created
7983 * with just the one range in it */
7988 if (invlist == NULL) {
7989 invlist = _new_invlist(2);
7993 len = _invlist_len(invlist);
7996 /* If comes after the final entry, can just append it to the end */
7998 || start >= invlist_array(invlist)
7999 [_invlist_len(invlist) - 1])
8001 _append_range_to_invlist(invlist, start, end);
8005 /* Here, can't just append things, create and return a new inversion list
8006 * which is the union of this range and the existing inversion list */
8007 range_invlist = _new_invlist(2);
8008 _append_range_to_invlist(range_invlist, start, end);
8010 _invlist_union(invlist, range_invlist, &invlist);
8012 /* The temporary can be freed */
8013 SvREFCNT_dec(range_invlist);
8020 PERL_STATIC_INLINE SV*
8021 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
8022 return _add_range_to_invlist(invlist, cp, cp);
8025 #ifndef PERL_IN_XSUB_RE
8027 Perl__invlist_invert(pTHX_ SV* const invlist)
8029 /* Complement the input inversion list. This adds a 0 if the list didn't
8030 * have a zero; removes it otherwise. As described above, the data
8031 * structure is set up so that this is very efficient */
8033 UV* len_pos = _get_invlist_len_addr(invlist);
8035 PERL_ARGS_ASSERT__INVLIST_INVERT;
8037 /* The inverse of matching nothing is matching everything */
8038 if (*len_pos == 0) {
8039 _append_range_to_invlist(invlist, 0, UV_MAX);
8043 /* The exclusive or complents 0 to 1; and 1 to 0. If the result is 1, the
8044 * zero element was a 0, so it is being removed, so the length decrements
8045 * by 1; and vice-versa. SvCUR is unaffected */
8046 if (*get_invlist_zero_addr(invlist) ^= 1) {
8055 Perl__invlist_invert_prop(pTHX_ SV* const invlist)
8057 /* Complement the input inversion list (which must be a Unicode property,
8058 * all of which don't match above the Unicode maximum code point.) And
8059 * Perl has chosen to not have the inversion match above that either. This
8060 * adds a 0x110000 if the list didn't end with it, and removes it if it did
8066 PERL_ARGS_ASSERT__INVLIST_INVERT_PROP;
8068 _invlist_invert(invlist);
8070 len = _invlist_len(invlist);
8072 if (len != 0) { /* If empty do nothing */
8073 array = invlist_array(invlist);
8074 if (array[len - 1] != PERL_UNICODE_MAX + 1) {
8075 /* Add 0x110000. First, grow if necessary */
8077 if (invlist_max(invlist) < len) {
8078 invlist_extend(invlist, len);
8079 array = invlist_array(invlist);
8081 invlist_set_len(invlist, len);
8082 array[len - 1] = PERL_UNICODE_MAX + 1;
8084 else { /* Remove the 0x110000 */
8085 invlist_set_len(invlist, len - 1);
8093 PERL_STATIC_INLINE SV*
8094 S_invlist_clone(pTHX_ SV* const invlist)
8097 /* Return a new inversion list that is a copy of the input one, which is
8100 /* Need to allocate extra space to accommodate Perl's addition of a
8101 * trailing NUL to SvPV's, since it thinks they are always strings */
8102 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
8103 STRLEN length = SvCUR(invlist);
8105 PERL_ARGS_ASSERT_INVLIST_CLONE;
8107 SvCUR_set(new_invlist, length); /* This isn't done automatically */
8108 Copy(SvPVX(invlist), SvPVX(new_invlist), length, char);
8113 PERL_STATIC_INLINE UV*
8114 S_get_invlist_iter_addr(pTHX_ SV* invlist)
8116 /* Return the address of the UV that contains the current iteration
8119 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
8121 return (UV *) (SvPVX(invlist) + (INVLIST_ITER_OFFSET * sizeof (UV)));
8124 PERL_STATIC_INLINE UV*
8125 S_get_invlist_version_id_addr(pTHX_ SV* invlist)
8127 /* Return the address of the UV that contains the version id. */
8129 PERL_ARGS_ASSERT_GET_INVLIST_VERSION_ID_ADDR;
8131 return (UV *) (SvPVX(invlist) + (INVLIST_VERSION_ID_OFFSET * sizeof (UV)));
8134 PERL_STATIC_INLINE void
8135 S_invlist_iterinit(pTHX_ SV* invlist) /* Initialize iterator for invlist */
8137 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
8139 *get_invlist_iter_addr(invlist) = 0;
8143 S_invlist_iternext(pTHX_ SV* invlist, UV* start, UV* end)
8145 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
8146 * This call sets in <*start> and <*end>, the next range in <invlist>.
8147 * Returns <TRUE> if successful and the next call will return the next
8148 * range; <FALSE> if was already at the end of the list. If the latter,
8149 * <*start> and <*end> are unchanged, and the next call to this function
8150 * will start over at the beginning of the list */
8152 UV* pos = get_invlist_iter_addr(invlist);
8153 UV len = _invlist_len(invlist);
8156 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
8159 *pos = UV_MAX; /* Force iternit() to be required next time */
8163 array = invlist_array(invlist);
8165 *start = array[(*pos)++];
8171 *end = array[(*pos)++] - 1;
8177 PERL_STATIC_INLINE UV
8178 S_invlist_highest(pTHX_ SV* const invlist)
8180 /* Returns the highest code point that matches an inversion list. This API
8181 * has an ambiguity, as it returns 0 under either the highest is actually
8182 * 0, or if the list is empty. If this distinction matters to you, check
8183 * for emptiness before calling this function */
8185 UV len = _invlist_len(invlist);
8188 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
8194 array = invlist_array(invlist);
8196 /* The last element in the array in the inversion list always starts a
8197 * range that goes to infinity. That range may be for code points that are
8198 * matched in the inversion list, or it may be for ones that aren't
8199 * matched. In the latter case, the highest code point in the set is one
8200 * less than the beginning of this range; otherwise it is the final element
8201 * of this range: infinity */
8202 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
8204 : array[len - 1] - 1;
8207 #ifndef PERL_IN_XSUB_RE
8209 Perl__invlist_contents(pTHX_ SV* const invlist)
8211 /* Get the contents of an inversion list into a string SV so that they can
8212 * be printed out. It uses the format traditionally done for debug tracing
8216 SV* output = newSVpvs("\n");
8218 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
8220 invlist_iterinit(invlist);
8221 while (invlist_iternext(invlist, &start, &end)) {
8222 if (end == UV_MAX) {
8223 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
8225 else if (end != start) {
8226 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
8230 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
8240 S_invlist_dump(pTHX_ SV* const invlist, const char * const header)
8242 /* Dumps out the ranges in an inversion list. The string 'header'
8243 * if present is output on a line before the first range */
8247 if (header && strlen(header)) {
8248 PerlIO_printf(Perl_debug_log, "%s\n", header);
8250 invlist_iterinit(invlist);
8251 while (invlist_iternext(invlist, &start, &end)) {
8252 if (end == UV_MAX) {
8253 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. INFINITY\n", start);
8256 PerlIO_printf(Perl_debug_log, "0x%04"UVXf" .. 0x%04"UVXf"\n", start, end);
8264 S__invlistEQ(pTHX_ SV* const a, SV* const b, bool complement_b)
8266 /* Return a boolean as to if the two passed in inversion lists are
8267 * identical. The final argument, if TRUE, says to take the complement of
8268 * the second inversion list before doing the comparison */
8270 UV* array_a = invlist_array(a);
8271 UV* array_b = invlist_array(b);
8272 UV len_a = _invlist_len(a);
8273 UV len_b = _invlist_len(b);
8275 UV i = 0; /* current index into the arrays */
8276 bool retval = TRUE; /* Assume are identical until proven otherwise */
8278 PERL_ARGS_ASSERT__INVLISTEQ;
8280 /* If are to compare 'a' with the complement of b, set it
8281 * up so are looking at b's complement. */
8284 /* The complement of nothing is everything, so <a> would have to have
8285 * just one element, starting at zero (ending at infinity) */
8287 return (len_a == 1 && array_a[0] == 0);
8289 else if (array_b[0] == 0) {
8291 /* Otherwise, to complement, we invert. Here, the first element is
8292 * 0, just remove it. To do this, we just pretend the array starts
8293 * one later, and clear the flag as we don't have to do anything
8298 complement_b = FALSE;
8302 /* But if the first element is not zero, we unshift a 0 before the
8303 * array. The data structure reserves a space for that 0 (which
8304 * should be a '1' right now), so physical shifting is unneeded,
8305 * but temporarily change that element to 0. Before exiting the
8306 * routine, we must restore the element to '1' */
8313 /* Make sure that the lengths are the same, as well as the final element
8314 * before looping through the remainder. (Thus we test the length, final,
8315 * and first elements right off the bat) */
8316 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
8319 else for (i = 0; i < len_a - 1; i++) {
8320 if (array_a[i] != array_b[i]) {
8333 #undef HEADER_LENGTH
8334 #undef INVLIST_INITIAL_LENGTH
8335 #undef TO_INTERNAL_SIZE
8336 #undef FROM_INTERNAL_SIZE
8337 #undef INVLIST_LEN_OFFSET
8338 #undef INVLIST_ZERO_OFFSET
8339 #undef INVLIST_ITER_OFFSET
8340 #undef INVLIST_VERSION_ID
8342 /* End of inversion list object */
8345 - reg - regular expression, i.e. main body or parenthesized thing
8347 * Caller must absorb opening parenthesis.
8349 * Combining parenthesis handling with the base level of regular expression
8350 * is a trifle forced, but the need to tie the tails of the branches to what
8351 * follows makes it hard to avoid.
8353 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
8355 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
8357 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
8361 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
8362 /* paren: Parenthesized? 0=top, 1=(, inside: changed to letter. */
8365 regnode *ret; /* Will be the head of the group. */
8368 regnode *ender = NULL;
8371 U32 oregflags = RExC_flags;
8372 bool have_branch = 0;
8374 I32 freeze_paren = 0;
8375 I32 after_freeze = 0;
8377 /* for (?g), (?gc), and (?o) warnings; warning
8378 about (?c) will warn about (?g) -- japhy */
8380 #define WASTED_O 0x01
8381 #define WASTED_G 0x02
8382 #define WASTED_C 0x04
8383 #define WASTED_GC (0x02|0x04)
8384 I32 wastedflags = 0x00;
8386 char * parse_start = RExC_parse; /* MJD */
8387 char * const oregcomp_parse = RExC_parse;
8389 GET_RE_DEBUG_FLAGS_DECL;
8391 PERL_ARGS_ASSERT_REG;
8392 DEBUG_PARSE("reg ");
8394 *flagp = 0; /* Tentatively. */
8397 /* Make an OPEN node, if parenthesized. */
8399 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
8400 char *start_verb = RExC_parse;
8401 STRLEN verb_len = 0;
8402 char *start_arg = NULL;
8403 unsigned char op = 0;
8405 int internal_argval = 0; /* internal_argval is only useful if !argok */
8406 while ( *RExC_parse && *RExC_parse != ')' ) {
8407 if ( *RExC_parse == ':' ) {
8408 start_arg = RExC_parse + 1;
8414 verb_len = RExC_parse - start_verb;
8417 while ( *RExC_parse && *RExC_parse != ')' )
8419 if ( *RExC_parse != ')' )
8420 vFAIL("Unterminated verb pattern argument");
8421 if ( RExC_parse == start_arg )
8424 if ( *RExC_parse != ')' )
8425 vFAIL("Unterminated verb pattern");
8428 switch ( *start_verb ) {
8429 case 'A': /* (*ACCEPT) */
8430 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
8432 internal_argval = RExC_nestroot;
8435 case 'C': /* (*COMMIT) */
8436 if ( memEQs(start_verb,verb_len,"COMMIT") )
8439 case 'F': /* (*FAIL) */
8440 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
8445 case ':': /* (*:NAME) */
8446 case 'M': /* (*MARK:NAME) */
8447 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
8452 case 'P': /* (*PRUNE) */
8453 if ( memEQs(start_verb,verb_len,"PRUNE") )
8456 case 'S': /* (*SKIP) */
8457 if ( memEQs(start_verb,verb_len,"SKIP") )
8460 case 'T': /* (*THEN) */
8461 /* [19:06] <TimToady> :: is then */
8462 if ( memEQs(start_verb,verb_len,"THEN") ) {
8464 RExC_seen |= REG_SEEN_CUTGROUP;
8470 vFAIL3("Unknown verb pattern '%.*s'",
8471 verb_len, start_verb);
8474 if ( start_arg && internal_argval ) {
8475 vFAIL3("Verb pattern '%.*s' may not have an argument",
8476 verb_len, start_verb);
8477 } else if ( argok < 0 && !start_arg ) {
8478 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
8479 verb_len, start_verb);
8481 ret = reganode(pRExC_state, op, internal_argval);
8482 if ( ! internal_argval && ! SIZE_ONLY ) {
8484 SV *sv = newSVpvn( start_arg, RExC_parse - start_arg);
8485 ARG(ret) = add_data( pRExC_state, 1, "S" );
8486 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
8493 if (!internal_argval)
8494 RExC_seen |= REG_SEEN_VERBARG;
8495 } else if ( start_arg ) {
8496 vFAIL3("Verb pattern '%.*s' may not have an argument",
8497 verb_len, start_verb);
8499 ret = reg_node(pRExC_state, op);
8501 nextchar(pRExC_state);
8504 if (*RExC_parse == '?') { /* (?...) */
8505 bool is_logical = 0;
8506 const char * const seqstart = RExC_parse;
8507 bool has_use_defaults = FALSE;
8510 paren = *RExC_parse++;
8511 ret = NULL; /* For look-ahead/behind. */
8514 case 'P': /* (?P...) variants for those used to PCRE/Python */
8515 paren = *RExC_parse++;
8516 if ( paren == '<') /* (?P<...>) named capture */
8518 else if (paren == '>') { /* (?P>name) named recursion */
8519 goto named_recursion;
8521 else if (paren == '=') { /* (?P=...) named backref */
8522 /* this pretty much dupes the code for \k<NAME> in regatom(), if
8523 you change this make sure you change that */
8524 char* name_start = RExC_parse;
8526 SV *sv_dat = reg_scan_name(pRExC_state,
8527 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8528 if (RExC_parse == name_start || *RExC_parse != ')')
8529 vFAIL2("Sequence %.3s... not terminated",parse_start);
8532 num = add_data( pRExC_state, 1, "S" );
8533 RExC_rxi->data->data[num]=(void*)sv_dat;
8534 SvREFCNT_inc_simple_void(sv_dat);
8537 ret = reganode(pRExC_state,
8540 : (ASCII_FOLD_RESTRICTED)
8542 : (AT_LEAST_UNI_SEMANTICS)
8550 Set_Node_Offset(ret, parse_start+1);
8551 Set_Node_Cur_Length(ret); /* MJD */
8553 nextchar(pRExC_state);
8557 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8559 case '<': /* (?<...) */
8560 if (*RExC_parse == '!')
8562 else if (*RExC_parse != '=')
8568 case '\'': /* (?'...') */
8569 name_start= RExC_parse;
8570 svname = reg_scan_name(pRExC_state,
8571 SIZE_ONLY ? /* reverse test from the others */
8572 REG_RSN_RETURN_NAME :
8573 REG_RSN_RETURN_NULL);
8574 if (RExC_parse == name_start) {
8576 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8579 if (*RExC_parse != paren)
8580 vFAIL2("Sequence (?%c... not terminated",
8581 paren=='>' ? '<' : paren);
8585 if (!svname) /* shouldn't happen */
8587 "panic: reg_scan_name returned NULL");
8588 if (!RExC_paren_names) {
8589 RExC_paren_names= newHV();
8590 sv_2mortal(MUTABLE_SV(RExC_paren_names));
8592 RExC_paren_name_list= newAV();
8593 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
8596 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
8598 sv_dat = HeVAL(he_str);
8600 /* croak baby croak */
8602 "panic: paren_name hash element allocation failed");
8603 } else if ( SvPOK(sv_dat) ) {
8604 /* (?|...) can mean we have dupes so scan to check
8605 its already been stored. Maybe a flag indicating
8606 we are inside such a construct would be useful,
8607 but the arrays are likely to be quite small, so
8608 for now we punt -- dmq */
8609 IV count = SvIV(sv_dat);
8610 I32 *pv = (I32*)SvPVX(sv_dat);
8612 for ( i = 0 ; i < count ; i++ ) {
8613 if ( pv[i] == RExC_npar ) {
8619 pv = (I32*)SvGROW(sv_dat, SvCUR(sv_dat) + sizeof(I32)+1);
8620 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
8621 pv[count] = RExC_npar;
8622 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
8625 (void)SvUPGRADE(sv_dat,SVt_PVNV);
8626 sv_setpvn(sv_dat, (char *)&(RExC_npar), sizeof(I32));
8628 SvIV_set(sv_dat, 1);
8631 /* Yes this does cause a memory leak in debugging Perls */
8632 if (!av_store(RExC_paren_name_list, RExC_npar, SvREFCNT_inc(svname)))
8633 SvREFCNT_dec(svname);
8636 /*sv_dump(sv_dat);*/
8638 nextchar(pRExC_state);
8640 goto capturing_parens;
8642 RExC_seen |= REG_SEEN_LOOKBEHIND;
8643 RExC_in_lookbehind++;
8645 case '=': /* (?=...) */
8646 RExC_seen_zerolen++;
8648 case '!': /* (?!...) */
8649 RExC_seen_zerolen++;
8650 if (*RExC_parse == ')') {
8651 ret=reg_node(pRExC_state, OPFAIL);
8652 nextchar(pRExC_state);
8656 case '|': /* (?|...) */
8657 /* branch reset, behave like a (?:...) except that
8658 buffers in alternations share the same numbers */
8660 after_freeze = freeze_paren = RExC_npar;
8662 case ':': /* (?:...) */
8663 case '>': /* (?>...) */
8665 case '$': /* (?$...) */
8666 case '@': /* (?@...) */
8667 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
8669 case '#': /* (?#...) */
8670 while (*RExC_parse && *RExC_parse != ')')
8672 if (*RExC_parse != ')')
8673 FAIL("Sequence (?#... not terminated");
8674 nextchar(pRExC_state);
8677 case '0' : /* (?0) */
8678 case 'R' : /* (?R) */
8679 if (*RExC_parse != ')')
8680 FAIL("Sequence (?R) not terminated");
8681 ret = reg_node(pRExC_state, GOSTART);
8682 *flagp |= POSTPONED;
8683 nextchar(pRExC_state);
8686 { /* named and numeric backreferences */
8688 case '&': /* (?&NAME) */
8689 parse_start = RExC_parse - 1;
8692 SV *sv_dat = reg_scan_name(pRExC_state,
8693 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8694 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8696 goto gen_recurse_regop;
8697 assert(0); /* NOT REACHED */
8699 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8701 vFAIL("Illegal pattern");
8703 goto parse_recursion;
8705 case '-': /* (?-1) */
8706 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
8707 RExC_parse--; /* rewind to let it be handled later */
8711 case '1': case '2': case '3': case '4': /* (?1) */
8712 case '5': case '6': case '7': case '8': case '9':
8715 num = atoi(RExC_parse);
8716 parse_start = RExC_parse - 1; /* MJD */
8717 if (*RExC_parse == '-')
8719 while (isDIGIT(*RExC_parse))
8721 if (*RExC_parse!=')')
8722 vFAIL("Expecting close bracket");
8725 if ( paren == '-' ) {
8727 Diagram of capture buffer numbering.
8728 Top line is the normal capture buffer numbers
8729 Bottom line is the negative indexing as from
8733 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
8737 num = RExC_npar + num;
8740 vFAIL("Reference to nonexistent group");
8742 } else if ( paren == '+' ) {
8743 num = RExC_npar + num - 1;
8746 ret = reganode(pRExC_state, GOSUB, num);
8748 if (num > (I32)RExC_rx->nparens) {
8750 vFAIL("Reference to nonexistent group");
8752 ARG2L_SET( ret, RExC_recurse_count++);
8754 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
8755 "Recurse #%"UVuf" to %"IVdf"\n", (UV)ARG(ret), (IV)ARG2L(ret)));
8759 RExC_seen |= REG_SEEN_RECURSE;
8760 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
8761 Set_Node_Offset(ret, parse_start); /* MJD */
8763 *flagp |= POSTPONED;
8764 nextchar(pRExC_state);
8766 } /* named and numeric backreferences */
8767 assert(0); /* NOT REACHED */
8769 case '?': /* (??...) */
8771 if (*RExC_parse != '{') {
8773 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
8776 *flagp |= POSTPONED;
8777 paren = *RExC_parse++;
8779 case '{': /* (?{...}) */
8782 struct reg_code_block *cb;
8784 RExC_seen_zerolen++;
8786 if ( !pRExC_state->num_code_blocks
8787 || pRExC_state->code_index >= pRExC_state->num_code_blocks
8788 || pRExC_state->code_blocks[pRExC_state->code_index].start
8789 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
8792 if (RExC_pm_flags & PMf_USE_RE_EVAL)
8793 FAIL("panic: Sequence (?{...}): no code block found\n");
8794 FAIL("Eval-group not allowed at runtime, use re 'eval'");
8796 /* this is a pre-compiled code block (?{...}) */
8797 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
8798 RExC_parse = RExC_start + cb->end;
8801 if (cb->src_regex) {
8802 n = add_data(pRExC_state, 2, "rl");
8803 RExC_rxi->data->data[n] =
8804 (void*)SvREFCNT_inc((SV*)cb->src_regex);
8805 RExC_rxi->data->data[n+1] = (void*)o;
8808 n = add_data(pRExC_state, 1,
8809 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l");
8810 RExC_rxi->data->data[n] = (void*)o;
8813 pRExC_state->code_index++;
8814 nextchar(pRExC_state);
8818 ret = reg_node(pRExC_state, LOGICAL);
8819 eval = reganode(pRExC_state, EVAL, n);
8822 /* for later propagation into (??{}) return value */
8823 eval->flags = (U8) (RExC_flags & RXf_PMf_COMPILETIME);
8825 REGTAIL(pRExC_state, ret, eval);
8826 /* deal with the length of this later - MJD */
8829 ret = reganode(pRExC_state, EVAL, n);
8830 Set_Node_Length(ret, RExC_parse - parse_start + 1);
8831 Set_Node_Offset(ret, parse_start);
8834 case '(': /* (?(?{...})...) and (?(?=...)...) */
8837 if (RExC_parse[0] == '?') { /* (?(?...)) */
8838 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
8839 || RExC_parse[1] == '<'
8840 || RExC_parse[1] == '{') { /* Lookahead or eval. */
8843 ret = reg_node(pRExC_state, LOGICAL);
8846 REGTAIL(pRExC_state, ret, reg(pRExC_state, 1, &flag,depth+1));
8850 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
8851 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
8853 char ch = RExC_parse[0] == '<' ? '>' : '\'';
8854 char *name_start= RExC_parse++;
8856 SV *sv_dat=reg_scan_name(pRExC_state,
8857 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8858 if (RExC_parse == name_start || *RExC_parse != ch)
8859 vFAIL2("Sequence (?(%c... not terminated",
8860 (ch == '>' ? '<' : ch));
8863 num = add_data( pRExC_state, 1, "S" );
8864 RExC_rxi->data->data[num]=(void*)sv_dat;
8865 SvREFCNT_inc_simple_void(sv_dat);
8867 ret = reganode(pRExC_state,NGROUPP,num);
8868 goto insert_if_check_paren;
8870 else if (RExC_parse[0] == 'D' &&
8871 RExC_parse[1] == 'E' &&
8872 RExC_parse[2] == 'F' &&
8873 RExC_parse[3] == 'I' &&
8874 RExC_parse[4] == 'N' &&
8875 RExC_parse[5] == 'E')
8877 ret = reganode(pRExC_state,DEFINEP,0);
8880 goto insert_if_check_paren;
8882 else if (RExC_parse[0] == 'R') {
8885 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
8886 parno = atoi(RExC_parse++);
8887 while (isDIGIT(*RExC_parse))
8889 } else if (RExC_parse[0] == '&') {
8892 sv_dat = reg_scan_name(pRExC_state,
8893 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
8894 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
8896 ret = reganode(pRExC_state,INSUBP,parno);
8897 goto insert_if_check_paren;
8899 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
8902 parno = atoi(RExC_parse++);
8904 while (isDIGIT(*RExC_parse))
8906 ret = reganode(pRExC_state, GROUPP, parno);
8908 insert_if_check_paren:
8909 if ((c = *nextchar(pRExC_state)) != ')')
8910 vFAIL("Switch condition not recognized");
8912 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
8913 br = regbranch(pRExC_state, &flags, 1,depth+1);
8915 br = reganode(pRExC_state, LONGJMP, 0);
8917 REGTAIL(pRExC_state, br, reganode(pRExC_state, LONGJMP, 0));
8918 c = *nextchar(pRExC_state);
8923 vFAIL("(?(DEFINE)....) does not allow branches");
8924 lastbr = reganode(pRExC_state, IFTHEN, 0); /* Fake one for optimizer. */
8925 regbranch(pRExC_state, &flags, 1,depth+1);
8926 REGTAIL(pRExC_state, ret, lastbr);
8929 c = *nextchar(pRExC_state);
8934 vFAIL("Switch (?(condition)... contains too many branches");
8935 ender = reg_node(pRExC_state, TAIL);
8936 REGTAIL(pRExC_state, br, ender);
8938 REGTAIL(pRExC_state, lastbr, ender);
8939 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
8942 REGTAIL(pRExC_state, ret, ender);
8943 RExC_size++; /* XXX WHY do we need this?!!
8944 For large programs it seems to be required
8945 but I can't figure out why. -- dmq*/
8949 vFAIL2("Unknown switch condition (?(%.2s", RExC_parse);
8953 RExC_parse--; /* for vFAIL to print correctly */
8954 vFAIL("Sequence (? incomplete");
8956 case DEFAULT_PAT_MOD: /* Use default flags with the exceptions
8958 has_use_defaults = TRUE;
8959 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
8960 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
8961 ? REGEX_UNICODE_CHARSET
8962 : REGEX_DEPENDS_CHARSET);
8966 parse_flags: /* (?i) */
8968 U32 posflags = 0, negflags = 0;
8969 U32 *flagsp = &posflags;
8970 char has_charset_modifier = '\0';
8971 regex_charset cs = get_regex_charset(RExC_flags);
8972 if (cs == REGEX_DEPENDS_CHARSET
8973 && (RExC_utf8 || RExC_uni_semantics))
8975 cs = REGEX_UNICODE_CHARSET;
8978 while (*RExC_parse) {
8979 /* && strchr("iogcmsx", *RExC_parse) */
8980 /* (?g), (?gc) and (?o) are useless here
8981 and must be globally applied -- japhy */
8982 switch (*RExC_parse) {
8983 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp);
8984 case LOCALE_PAT_MOD:
8985 if (has_charset_modifier) {
8986 goto excess_modifier;
8988 else if (flagsp == &negflags) {
8991 cs = REGEX_LOCALE_CHARSET;
8992 has_charset_modifier = LOCALE_PAT_MOD;
8993 RExC_contains_locale = 1;
8995 case UNICODE_PAT_MOD:
8996 if (has_charset_modifier) {
8997 goto excess_modifier;
8999 else if (flagsp == &negflags) {
9002 cs = REGEX_UNICODE_CHARSET;
9003 has_charset_modifier = UNICODE_PAT_MOD;
9005 case ASCII_RESTRICT_PAT_MOD:
9006 if (flagsp == &negflags) {
9009 if (has_charset_modifier) {
9010 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9011 goto excess_modifier;
9013 /* Doubled modifier implies more restricted */
9014 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9017 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9019 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9021 case DEPENDS_PAT_MOD:
9022 if (has_use_defaults) {
9023 goto fail_modifiers;
9025 else if (flagsp == &negflags) {
9028 else if (has_charset_modifier) {
9029 goto excess_modifier;
9032 /* The dual charset means unicode semantics if the
9033 * pattern (or target, not known until runtime) are
9034 * utf8, or something in the pattern indicates unicode
9036 cs = (RExC_utf8 || RExC_uni_semantics)
9037 ? REGEX_UNICODE_CHARSET
9038 : REGEX_DEPENDS_CHARSET;
9039 has_charset_modifier = DEPENDS_PAT_MOD;
9043 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9044 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9046 else if (has_charset_modifier == *(RExC_parse - 1)) {
9047 vFAIL2("Regexp modifier \"%c\" may not appear twice", *(RExC_parse - 1));
9050 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9055 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"", *(RExC_parse - 1));
9057 case ONCE_PAT_MOD: /* 'o' */
9058 case GLOBAL_PAT_MOD: /* 'g' */
9059 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9060 const I32 wflagbit = *RExC_parse == 'o' ? WASTED_O : WASTED_G;
9061 if (! (wastedflags & wflagbit) ) {
9062 wastedflags |= wflagbit;
9065 "Useless (%s%c) - %suse /%c modifier",
9066 flagsp == &negflags ? "?-" : "?",
9068 flagsp == &negflags ? "don't " : "",
9075 case CONTINUE_PAT_MOD: /* 'c' */
9076 if (SIZE_ONLY && ckWARN(WARN_REGEXP)) {
9077 if (! (wastedflags & WASTED_C) ) {
9078 wastedflags |= WASTED_GC;
9081 "Useless (%sc) - %suse /gc modifier",
9082 flagsp == &negflags ? "?-" : "?",
9083 flagsp == &negflags ? "don't " : ""
9088 case KEEPCOPY_PAT_MOD: /* 'p' */
9089 if (flagsp == &negflags) {
9091 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9093 *flagsp |= RXf_PMf_KEEPCOPY;
9097 /* A flag is a default iff it is following a minus, so
9098 * if there is a minus, it means will be trying to
9099 * re-specify a default which is an error */
9100 if (has_use_defaults || flagsp == &negflags) {
9103 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9107 wastedflags = 0; /* reset so (?g-c) warns twice */
9113 RExC_flags |= posflags;
9114 RExC_flags &= ~negflags;
9115 set_regex_charset(&RExC_flags, cs);
9117 oregflags |= posflags;
9118 oregflags &= ~negflags;
9119 set_regex_charset(&oregflags, cs);
9121 nextchar(pRExC_state);
9132 vFAIL3("Sequence (%.*s...) not recognized", RExC_parse-seqstart, seqstart);
9137 }} /* one for the default block, one for the switch */
9144 ret = reganode(pRExC_state, OPEN, parno);
9147 RExC_nestroot = parno;
9148 if (RExC_seen & REG_SEEN_RECURSE
9149 && !RExC_open_parens[parno-1])
9151 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9152 "Setting open paren #%"IVdf" to %d\n",
9153 (IV)parno, REG_NODE_NUM(ret)));
9154 RExC_open_parens[parno-1]= ret;
9157 Set_Node_Length(ret, 1); /* MJD */
9158 Set_Node_Offset(ret, RExC_parse); /* MJD */
9166 /* Pick up the branches, linking them together. */
9167 parse_start = RExC_parse; /* MJD */
9168 br = regbranch(pRExC_state, &flags, 1,depth+1);
9170 /* branch_len = (paren != 0); */
9174 if (*RExC_parse == '|') {
9175 if (!SIZE_ONLY && RExC_extralen) {
9176 reginsert(pRExC_state, BRANCHJ, br, depth+1);
9179 reginsert(pRExC_state, BRANCH, br, depth+1);
9180 Set_Node_Length(br, paren != 0);
9181 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
9185 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
9187 else if (paren == ':') {
9188 *flagp |= flags&SIMPLE;
9190 if (is_open) { /* Starts with OPEN. */
9191 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
9193 else if (paren != '?') /* Not Conditional */
9195 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9197 while (*RExC_parse == '|') {
9198 if (!SIZE_ONLY && RExC_extralen) {
9199 ender = reganode(pRExC_state, LONGJMP,0);
9200 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender); /* Append to the previous. */
9203 RExC_extralen += 2; /* Account for LONGJMP. */
9204 nextchar(pRExC_state);
9206 if (RExC_npar > after_freeze)
9207 after_freeze = RExC_npar;
9208 RExC_npar = freeze_paren;
9210 br = regbranch(pRExC_state, &flags, 0, depth+1);
9214 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
9216 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
9219 if (have_branch || paren != ':') {
9220 /* Make a closing node, and hook it on the end. */
9223 ender = reg_node(pRExC_state, TAIL);
9226 ender = reganode(pRExC_state, CLOSE, parno);
9227 if (!SIZE_ONLY && RExC_seen & REG_SEEN_RECURSE) {
9228 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
9229 "Setting close paren #%"IVdf" to %d\n",
9230 (IV)parno, REG_NODE_NUM(ender)));
9231 RExC_close_parens[parno-1]= ender;
9232 if (RExC_nestroot == parno)
9235 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
9236 Set_Node_Length(ender,1); /* MJD */
9242 *flagp &= ~HASWIDTH;
9245 ender = reg_node(pRExC_state, SUCCEED);
9248 ender = reg_node(pRExC_state, END);
9250 assert(!RExC_opend); /* there can only be one! */
9255 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9256 SV * const mysv_val1=sv_newmortal();
9257 SV * const mysv_val2=sv_newmortal();
9258 DEBUG_PARSE_MSG("lsbr");
9259 regprop(RExC_rx, mysv_val1, lastbr);
9260 regprop(RExC_rx, mysv_val2, ender);
9261 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9262 SvPV_nolen_const(mysv_val1),
9263 (IV)REG_NODE_NUM(lastbr),
9264 SvPV_nolen_const(mysv_val2),
9265 (IV)REG_NODE_NUM(ender),
9266 (IV)(ender - lastbr)
9269 REGTAIL(pRExC_state, lastbr, ender);
9271 if (have_branch && !SIZE_ONLY) {
9274 RExC_seen |= REG_TOP_LEVEL_BRANCHES;
9276 /* Hook the tails of the branches to the closing node. */
9277 for (br = ret; br; br = regnext(br)) {
9278 const U8 op = PL_regkind[OP(br)];
9280 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
9281 if (OP(NEXTOPER(br)) != NOTHING || regnext(NEXTOPER(br)) != ender)
9284 else if (op == BRANCHJ) {
9285 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
9286 /* for now we always disable this optimisation * /
9287 if (OP(NEXTOPER(NEXTOPER(br))) != NOTHING || regnext(NEXTOPER(NEXTOPER(br))) != ender)
9293 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
9294 DEBUG_PARSE_r(if (!SIZE_ONLY) {
9295 SV * const mysv_val1=sv_newmortal();
9296 SV * const mysv_val2=sv_newmortal();
9297 DEBUG_PARSE_MSG("NADA");
9298 regprop(RExC_rx, mysv_val1, ret);
9299 regprop(RExC_rx, mysv_val2, ender);
9300 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
9301 SvPV_nolen_const(mysv_val1),
9302 (IV)REG_NODE_NUM(ret),
9303 SvPV_nolen_const(mysv_val2),
9304 (IV)REG_NODE_NUM(ender),
9309 if (OP(ender) == TAIL) {
9314 for ( opt= br + 1; opt < ender ; opt++ )
9316 NEXT_OFF(br)= ender - br;
9324 static const char parens[] = "=!<,>";
9326 if (paren && (p = strchr(parens, paren))) {
9327 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
9328 int flag = (p - parens) > 1;
9331 node = SUSPEND, flag = 0;
9332 reginsert(pRExC_state, node,ret, depth+1);
9333 Set_Node_Cur_Length(ret);
9334 Set_Node_Offset(ret, parse_start + 1);
9336 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
9340 /* Check for proper termination. */
9342 RExC_flags = oregflags;
9343 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
9344 RExC_parse = oregcomp_parse;
9345 vFAIL("Unmatched (");
9348 else if (!paren && RExC_parse < RExC_end) {
9349 if (*RExC_parse == ')') {
9351 vFAIL("Unmatched )");
9354 FAIL("Junk on end of regexp"); /* "Can't happen". */
9355 assert(0); /* NOTREACHED */
9358 if (RExC_in_lookbehind) {
9359 RExC_in_lookbehind--;
9361 if (after_freeze > RExC_npar)
9362 RExC_npar = after_freeze;
9367 - regbranch - one alternative of an | operator
9369 * Implements the concatenation operator.
9372 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
9376 regnode *chain = NULL;
9378 I32 flags = 0, c = 0;
9379 GET_RE_DEBUG_FLAGS_DECL;
9381 PERL_ARGS_ASSERT_REGBRANCH;
9383 DEBUG_PARSE("brnc");
9388 if (!SIZE_ONLY && RExC_extralen)
9389 ret = reganode(pRExC_state, BRANCHJ,0);
9391 ret = reg_node(pRExC_state, BRANCH);
9392 Set_Node_Length(ret, 1);
9396 if (!first && SIZE_ONLY)
9397 RExC_extralen += 1; /* BRANCHJ */
9399 *flagp = WORST; /* Tentatively. */
9402 nextchar(pRExC_state);
9403 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
9405 latest = regpiece(pRExC_state, &flags,depth+1);
9406 if (latest == NULL) {
9407 if (flags & TRYAGAIN)
9411 else if (ret == NULL)
9413 *flagp |= flags&(HASWIDTH|POSTPONED);
9414 if (chain == NULL) /* First piece. */
9415 *flagp |= flags&SPSTART;
9418 REGTAIL(pRExC_state, chain, latest);
9423 if (chain == NULL) { /* Loop ran zero times. */
9424 chain = reg_node(pRExC_state, NOTHING);
9429 *flagp |= flags&SIMPLE;
9436 - regpiece - something followed by possible [*+?]
9438 * Note that the branching code sequences used for ? and the general cases
9439 * of * and + are somewhat optimized: they use the same NOTHING node as
9440 * both the endmarker for their branch list and the body of the last branch.
9441 * It might seem that this node could be dispensed with entirely, but the
9442 * endmarker role is not redundant.
9445 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
9452 const char * const origparse = RExC_parse;
9454 I32 max = REG_INFTY;
9455 #ifdef RE_TRACK_PATTERN_OFFSETS
9458 const char *maxpos = NULL;
9460 /* Save the original in case we change the emitted regop to a FAIL. */
9461 regnode * const orig_emit = RExC_emit;
9463 GET_RE_DEBUG_FLAGS_DECL;
9465 PERL_ARGS_ASSERT_REGPIECE;
9467 DEBUG_PARSE("piec");
9469 ret = regatom(pRExC_state, &flags,depth+1);
9471 if (flags & TRYAGAIN)
9478 if (op == '{' && regcurly(RExC_parse)) {
9480 #ifdef RE_TRACK_PATTERN_OFFSETS
9481 parse_start = RExC_parse; /* MJD */
9483 next = RExC_parse + 1;
9484 while (isDIGIT(*next) || *next == ',') {
9493 if (*next == '}') { /* got one */
9497 min = atoi(RExC_parse);
9501 maxpos = RExC_parse;
9503 if (!max && *maxpos != '0')
9504 max = REG_INFTY; /* meaning "infinity" */
9505 else if (max >= REG_INFTY)
9506 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
9508 nextchar(pRExC_state);
9509 if (max < min) { /* If can't match, warn and optimize to fail
9512 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
9514 /* We can't back off the size because we have to reserve
9515 * enough space for all the things we are about to throw
9516 * away, but we can shrink it by the ammount we are about
9518 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
9521 RExC_emit = orig_emit;
9523 ret = reg_node(pRExC_state, OPFAIL);
9528 if ((flags&SIMPLE)) {
9529 RExC_naughty += 2 + RExC_naughty / 2;
9530 reginsert(pRExC_state, CURLY, ret, depth+1);
9531 Set_Node_Offset(ret, parse_start+1); /* MJD */
9532 Set_Node_Cur_Length(ret);
9535 regnode * const w = reg_node(pRExC_state, WHILEM);
9538 REGTAIL(pRExC_state, ret, w);
9539 if (!SIZE_ONLY && RExC_extralen) {
9540 reginsert(pRExC_state, LONGJMP,ret, depth+1);
9541 reginsert(pRExC_state, NOTHING,ret, depth+1);
9542 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
9544 reginsert(pRExC_state, CURLYX,ret, depth+1);
9546 Set_Node_Offset(ret, parse_start+1);
9547 Set_Node_Length(ret,
9548 op == '{' ? (RExC_parse - parse_start) : 1);
9550 if (!SIZE_ONLY && RExC_extralen)
9551 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
9552 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
9554 RExC_whilem_seen++, RExC_extralen += 3;
9555 RExC_naughty += 4 + RExC_naughty; /* compound interest */
9564 ARG1_SET(ret, (U16)min);
9565 ARG2_SET(ret, (U16)max);
9577 #if 0 /* Now runtime fix should be reliable. */
9579 /* if this is reinstated, don't forget to put this back into perldiag:
9581 =item Regexp *+ operand could be empty at {#} in regex m/%s/
9583 (F) The part of the regexp subject to either the * or + quantifier
9584 could match an empty string. The {#} shows in the regular
9585 expression about where the problem was discovered.
9589 if (!(flags&HASWIDTH) && op != '?')
9590 vFAIL("Regexp *+ operand could be empty");
9593 #ifdef RE_TRACK_PATTERN_OFFSETS
9594 parse_start = RExC_parse;
9596 nextchar(pRExC_state);
9598 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
9600 if (op == '*' && (flags&SIMPLE)) {
9601 reginsert(pRExC_state, STAR, ret, depth+1);
9605 else if (op == '*') {
9609 else if (op == '+' && (flags&SIMPLE)) {
9610 reginsert(pRExC_state, PLUS, ret, depth+1);
9614 else if (op == '+') {
9618 else if (op == '?') {
9623 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
9624 ckWARN3reg(RExC_parse,
9625 "%.*s matches null string many times",
9626 (int)(RExC_parse >= origparse ? RExC_parse - origparse : 0),
9630 if (RExC_parse < RExC_end && *RExC_parse == '?') {
9631 nextchar(pRExC_state);
9632 reginsert(pRExC_state, MINMOD, ret, depth+1);
9633 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
9635 #ifndef REG_ALLOW_MINMOD_SUSPEND
9638 if (RExC_parse < RExC_end && *RExC_parse == '+') {
9640 nextchar(pRExC_state);
9641 ender = reg_node(pRExC_state, SUCCEED);
9642 REGTAIL(pRExC_state, ret, ender);
9643 reginsert(pRExC_state, SUSPEND, ret, depth+1);
9645 ender = reg_node(pRExC_state, TAIL);
9646 REGTAIL(pRExC_state, ret, ender);
9650 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
9652 vFAIL("Nested quantifiers");
9659 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state, regnode** node_p, UV *valuep, I32 *flagp, U32 depth, bool in_char_class)
9662 /* This is expected to be called by a parser routine that has recognized '\N'
9663 and needs to handle the rest. RExC_parse is expected to point at the first
9664 char following the N at the time of the call. On successful return,
9665 RExC_parse has been updated to point to just after the sequence identified
9666 by this routine, and <*flagp> has been updated.
9668 The \N may be inside (indicated by the boolean <in_char_class>) or outside a
9671 \N may begin either a named sequence, or if outside a character class, mean
9672 to match a non-newline. For non single-quoted regexes, the tokenizer has
9673 attempted to decide which, and in the case of a named sequence, converted it
9674 into one of the forms: \N{} (if the sequence is null), or \N{U+c1.c2...},
9675 where c1... are the characters in the sequence. For single-quoted regexes,
9676 the tokenizer passes the \N sequence through unchanged; this code will not
9677 attempt to determine this nor expand those, instead raising a syntax error.
9678 The net effect is that if the beginning of the passed-in pattern isn't '{U+'
9679 or there is no '}', it signals that this \N occurrence means to match a
9682 Only the \N{U+...} form should occur in a character class, for the same
9683 reason that '.' inside a character class means to just match a period: it
9684 just doesn't make sense.
9686 The function raises an error (via vFAIL), and doesn't return for various
9687 syntax errors. Otherwise it returns TRUE and sets <node_p> or <valuep> on
9688 success; it returns FALSE otherwise.
9690 If <valuep> is non-null, it means the caller can accept an input sequence
9691 consisting of a just a single code point; <*valuep> is set to that value
9692 if the input is such.
9694 If <node_p> is non-null it signifies that the caller can accept any other
9695 legal sequence (i.e., one that isn't just a single code point). <*node_p>
9697 1) \N means not-a-NL: points to a newly created REG_ANY node;
9698 2) \N{}: points to a new NOTHING node;
9699 3) otherwise: points to a new EXACT node containing the resolved
9701 Note that FALSE is returned for single code point sequences if <valuep> is
9705 char * endbrace; /* '}' following the name */
9707 char *endchar; /* Points to '.' or '}' ending cur char in the input
9709 bool has_multiple_chars; /* true if the input stream contains a sequence of
9710 more than one character */
9712 GET_RE_DEBUG_FLAGS_DECL;
9714 PERL_ARGS_ASSERT_GROK_BSLASH_N;
9718 assert(cBOOL(node_p) ^ cBOOL(valuep)); /* Exactly one should be set */
9720 /* The [^\n] meaning of \N ignores spaces and comments under the /x
9721 * modifier. The other meaning does not */
9722 p = (RExC_flags & RXf_PMf_EXTENDED)
9723 ? regwhite( pRExC_state, RExC_parse )
9726 /* Disambiguate between \N meaning a named character versus \N meaning
9727 * [^\n]. The former is assumed when it can't be the latter. */
9728 if (*p != '{' || regcurly(p)) {
9731 /* no bare \N in a charclass */
9732 if (in_char_class) {
9733 vFAIL("\\N in a character class must be a named character: \\N{...}");
9737 nextchar(pRExC_state);
9738 *node_p = reg_node(pRExC_state, REG_ANY);
9739 *flagp |= HASWIDTH|SIMPLE;
9742 Set_Node_Length(*node_p, 1); /* MJD */
9746 /* Here, we have decided it should be a named character or sequence */
9748 /* The test above made sure that the next real character is a '{', but
9749 * under the /x modifier, it could be separated by space (or a comment and
9750 * \n) and this is not allowed (for consistency with \x{...} and the
9751 * tokenizer handling of \N{NAME}). */
9752 if (*RExC_parse != '{') {
9753 vFAIL("Missing braces on \\N{}");
9756 RExC_parse++; /* Skip past the '{' */
9758 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
9759 || ! (endbrace == RExC_parse /* nothing between the {} */
9760 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked below */
9761 && strnEQ(RExC_parse, "U+", 2)))) /* for a better error msg) */
9763 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
9764 vFAIL("\\N{NAME} must be resolved by the lexer");
9767 if (endbrace == RExC_parse) { /* empty: \N{} */
9770 *node_p = reg_node(pRExC_state,NOTHING);
9772 else if (in_char_class) {
9773 if (SIZE_ONLY && in_char_class) {
9774 ckWARNreg(RExC_parse,
9775 "Ignoring zero length \\N{} in character class"
9783 nextchar(pRExC_state);
9787 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
9788 RExC_parse += 2; /* Skip past the 'U+' */
9790 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9792 /* Code points are separated by dots. If none, there is only one code
9793 * point, and is terminated by the brace */
9794 has_multiple_chars = (endchar < endbrace);
9796 if (valuep && (! has_multiple_chars || in_char_class)) {
9797 /* We only pay attention to the first char of
9798 multichar strings being returned in char classes. I kinda wonder
9799 if this makes sense as it does change the behaviour
9800 from earlier versions, OTOH that behaviour was broken
9801 as well. XXX Solution is to recharacterize as
9802 [rest-of-class]|multi1|multi2... */
9804 STRLEN length_of_hex = (STRLEN)(endchar - RExC_parse);
9805 I32 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
9806 | PERL_SCAN_DISALLOW_PREFIX
9807 | (SIZE_ONLY ? PERL_SCAN_SILENT_ILLDIGIT : 0);
9809 *valuep = grok_hex(RExC_parse, &length_of_hex, &grok_hex_flags, NULL);
9811 /* The tokenizer should have guaranteed validity, but it's possible to
9812 * bypass it by using single quoting, so check */
9813 if (length_of_hex == 0
9814 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
9816 RExC_parse += length_of_hex; /* Includes all the valid */
9817 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
9818 ? UTF8SKIP(RExC_parse)
9820 /* Guard against malformed utf8 */
9821 if (RExC_parse >= endchar) {
9822 RExC_parse = endchar;
9824 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9827 if (in_char_class && has_multiple_chars) {
9828 ckWARNreg(endchar, "Using just the first character returned by \\N{} in character class");
9830 RExC_parse = endbrace + 1;
9832 else if (! node_p || ! has_multiple_chars) {
9834 /* Here, the input is legal, but not according to the caller's
9835 * options. We fail without advancing the parse, so that the
9836 * caller can try again */
9842 /* What is done here is to convert this to a sub-pattern of the form
9843 * (?:\x{char1}\x{char2}...)
9844 * and then call reg recursively. That way, it retains its atomicness,
9845 * while not having to worry about special handling that some code
9846 * points may have. toke.c has converted the original Unicode values
9847 * to native, so that we can just pass on the hex values unchanged. We
9848 * do have to set a flag to keep recoding from happening in the
9851 SV * substitute_parse = newSVpvn_flags("?:", 2, SVf_UTF8|SVs_TEMP);
9853 char *orig_end = RExC_end;
9856 while (RExC_parse < endbrace) {
9858 /* Convert to notation the rest of the code understands */
9859 sv_catpv(substitute_parse, "\\x{");
9860 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
9861 sv_catpv(substitute_parse, "}");
9863 /* Point to the beginning of the next character in the sequence. */
9864 RExC_parse = endchar + 1;
9865 endchar = RExC_parse + strcspn(RExC_parse, ".}");
9867 sv_catpv(substitute_parse, ")");
9869 RExC_parse = SvPV(substitute_parse, len);
9871 /* Don't allow empty number */
9873 vFAIL("Invalid hexadecimal number in \\N{U+...}");
9875 RExC_end = RExC_parse + len;
9877 /* The values are Unicode, and therefore not subject to recoding */
9878 RExC_override_recoding = 1;
9880 *node_p = reg(pRExC_state, 1, &flags, depth+1);
9881 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
9883 RExC_parse = endbrace;
9884 RExC_end = orig_end;
9885 RExC_override_recoding = 0;
9887 nextchar(pRExC_state);
9897 * It returns the code point in utf8 for the value in *encp.
9898 * value: a code value in the source encoding
9899 * encp: a pointer to an Encode object
9901 * If the result from Encode is not a single character,
9902 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
9905 S_reg_recode(pTHX_ const char value, SV **encp)
9908 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
9909 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
9910 const STRLEN newlen = SvCUR(sv);
9911 UV uv = UNICODE_REPLACEMENT;
9913 PERL_ARGS_ASSERT_REG_RECODE;
9917 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
9920 if (!newlen || numlen != newlen) {
9921 uv = UNICODE_REPLACEMENT;
9927 PERL_STATIC_INLINE U8
9928 S_compute_EXACTish(pTHX_ RExC_state_t *pRExC_state)
9932 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
9938 op = get_regex_charset(RExC_flags);
9939 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
9940 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
9941 been, so there is no hole */
9947 PERL_STATIC_INLINE void
9948 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state, regnode *node, I32* flagp, STRLEN len, UV code_point)
9950 /* This knows the details about sizing an EXACTish node, setting flags for
9951 * it (by setting <*flagp>, and potentially populating it with a single
9954 * If <len> is non-zero, this function assumes that the node has already
9955 * been populated, and just does the sizing. In this case <code_point>
9956 * should be the final code point that has already been placed into the
9957 * node. This value will be ignored except that under some circumstances
9958 * <*flagp> is set based on it.
9960 * If <len is zero, the function assumes that the node is to contain only
9961 * the single character given by <code_point> and calculates what <len>
9962 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
9963 * additionally will populate the node's STRING with <code_point>, if <len>
9964 * is 0. In both cases <*flagp> is appropriately set
9966 * It knows that under FOLD, UTF characters and the Latin Sharp S must be
9967 * folded (the latter only when the rules indicate it can match 'ss') */
9969 bool len_passed_in = cBOOL(len != 0);
9970 U8 character[UTF8_MAXBYTES_CASE+1];
9972 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
9974 if (! len_passed_in) {
9977 to_uni_fold(NATIVE_TO_UNI(code_point), character, &len);
9980 uvchr_to_utf8( character, code_point);
9981 len = UTF8SKIP(character);
9985 || code_point != LATIN_SMALL_LETTER_SHARP_S
9986 || ASCII_FOLD_RESTRICTED
9987 || ! AT_LEAST_UNI_SEMANTICS)
9989 *character = (U8) code_point;
9994 *(character + 1) = 's';
10000 RExC_size += STR_SZ(len);
10003 RExC_emit += STR_SZ(len);
10004 STR_LEN(node) = len;
10005 if (! len_passed_in) {
10006 Copy((char *) character, STRING(node), len, char);
10010 *flagp |= HASWIDTH;
10011 if (len == 1 && UNI_IS_INVARIANT(code_point))
10016 - regatom - the lowest level
10018 Try to identify anything special at the start of the pattern. If there
10019 is, then handle it as required. This may involve generating a single regop,
10020 such as for an assertion; or it may involve recursing, such as to
10021 handle a () structure.
10023 If the string doesn't start with something special then we gobble up
10024 as much literal text as we can.
10026 Once we have been able to handle whatever type of thing started the
10027 sequence, we return.
10029 Note: we have to be careful with escapes, as they can be both literal
10030 and special, and in the case of \10 and friends, context determines which.
10032 A summary of the code structure is:
10034 switch (first_byte) {
10035 cases for each special:
10036 handle this special;
10039 switch (2nd byte) {
10040 cases for each unambiguous special:
10041 handle this special;
10043 cases for each ambigous special/literal:
10045 if (special) handle here
10047 default: // unambiguously literal:
10050 default: // is a literal char
10053 create EXACTish node for literal;
10054 while (more input and node isn't full) {
10055 switch (input_byte) {
10056 cases for each special;
10057 make sure parse pointer is set so that the next call to
10058 regatom will see this special first
10059 goto loopdone; // EXACTish node terminated by prev. char
10061 append char to EXACTISH node;
10063 get next input byte;
10067 return the generated node;
10069 Specifically there are two separate switches for handling
10070 escape sequences, with the one for handling literal escapes requiring
10071 a dummy entry for all of the special escapes that are actually handled
10076 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10079 regnode *ret = NULL;
10081 char *parse_start = RExC_parse;
10083 GET_RE_DEBUG_FLAGS_DECL;
10084 DEBUG_PARSE("atom");
10085 *flagp = WORST; /* Tentatively. */
10087 PERL_ARGS_ASSERT_REGATOM;
10090 switch ((U8)*RExC_parse) {
10092 RExC_seen_zerolen++;
10093 nextchar(pRExC_state);
10094 if (RExC_flags & RXf_PMf_MULTILINE)
10095 ret = reg_node(pRExC_state, MBOL);
10096 else if (RExC_flags & RXf_PMf_SINGLELINE)
10097 ret = reg_node(pRExC_state, SBOL);
10099 ret = reg_node(pRExC_state, BOL);
10100 Set_Node_Length(ret, 1); /* MJD */
10103 nextchar(pRExC_state);
10105 RExC_seen_zerolen++;
10106 if (RExC_flags & RXf_PMf_MULTILINE)
10107 ret = reg_node(pRExC_state, MEOL);
10108 else if (RExC_flags & RXf_PMf_SINGLELINE)
10109 ret = reg_node(pRExC_state, SEOL);
10111 ret = reg_node(pRExC_state, EOL);
10112 Set_Node_Length(ret, 1); /* MJD */
10115 nextchar(pRExC_state);
10116 if (RExC_flags & RXf_PMf_SINGLELINE)
10117 ret = reg_node(pRExC_state, SANY);
10119 ret = reg_node(pRExC_state, REG_ANY);
10120 *flagp |= HASWIDTH|SIMPLE;
10122 Set_Node_Length(ret, 1); /* MJD */
10126 char * const oregcomp_parse = ++RExC_parse;
10127 ret = regclass(pRExC_state, flagp,depth+1);
10128 if (*RExC_parse != ']') {
10129 RExC_parse = oregcomp_parse;
10130 vFAIL("Unmatched [");
10132 nextchar(pRExC_state);
10133 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
10137 nextchar(pRExC_state);
10138 ret = reg(pRExC_state, 1, &flags,depth+1);
10140 if (flags & TRYAGAIN) {
10141 if (RExC_parse == RExC_end) {
10142 /* Make parent create an empty node if needed. */
10143 *flagp |= TRYAGAIN;
10150 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
10154 if (flags & TRYAGAIN) {
10155 *flagp |= TRYAGAIN;
10158 vFAIL("Internal urp");
10159 /* Supposed to be caught earlier. */
10165 vFAIL("Quantifier follows nothing");
10170 This switch handles escape sequences that resolve to some kind
10171 of special regop and not to literal text. Escape sequnces that
10172 resolve to literal text are handled below in the switch marked
10175 Every entry in this switch *must* have a corresponding entry
10176 in the literal escape switch. However, the opposite is not
10177 required, as the default for this switch is to jump to the
10178 literal text handling code.
10180 switch ((U8)*++RExC_parse) {
10181 /* Special Escapes */
10183 RExC_seen_zerolen++;
10184 ret = reg_node(pRExC_state, SBOL);
10186 goto finish_meta_pat;
10188 ret = reg_node(pRExC_state, GPOS);
10189 RExC_seen |= REG_SEEN_GPOS;
10191 goto finish_meta_pat;
10193 RExC_seen_zerolen++;
10194 ret = reg_node(pRExC_state, KEEPS);
10196 /* XXX:dmq : disabling in-place substitution seems to
10197 * be necessary here to avoid cases of memory corruption, as
10198 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
10200 RExC_seen |= REG_SEEN_LOOKBEHIND;
10201 goto finish_meta_pat;
10203 ret = reg_node(pRExC_state, SEOL);
10205 RExC_seen_zerolen++; /* Do not optimize RE away */
10206 goto finish_meta_pat;
10208 ret = reg_node(pRExC_state, EOS);
10210 RExC_seen_zerolen++; /* Do not optimize RE away */
10211 goto finish_meta_pat;
10213 ret = reg_node(pRExC_state, CANY);
10214 RExC_seen |= REG_SEEN_CANY;
10215 *flagp |= HASWIDTH|SIMPLE;
10216 goto finish_meta_pat;
10218 ret = reg_node(pRExC_state, CLUMP);
10219 *flagp |= HASWIDTH;
10220 goto finish_meta_pat;
10222 op = ALNUM + get_regex_charset(RExC_flags);
10223 if (op > ALNUMA) { /* /aa is same as /a */
10226 ret = reg_node(pRExC_state, op);
10227 *flagp |= HASWIDTH|SIMPLE;
10228 goto finish_meta_pat;
10230 op = NALNUM + get_regex_charset(RExC_flags);
10231 if (op > NALNUMA) { /* /aa is same as /a */
10234 ret = reg_node(pRExC_state, op);
10235 *flagp |= HASWIDTH|SIMPLE;
10236 goto finish_meta_pat;
10238 RExC_seen_zerolen++;
10239 RExC_seen |= REG_SEEN_LOOKBEHIND;
10240 op = BOUND + get_regex_charset(RExC_flags);
10241 if (op > BOUNDA) { /* /aa is same as /a */
10244 ret = reg_node(pRExC_state, op);
10245 FLAGS(ret) = get_regex_charset(RExC_flags);
10247 goto finish_meta_pat;
10249 RExC_seen_zerolen++;
10250 RExC_seen |= REG_SEEN_LOOKBEHIND;
10251 op = NBOUND + get_regex_charset(RExC_flags);
10252 if (op > NBOUNDA) { /* /aa is same as /a */
10255 ret = reg_node(pRExC_state, op);
10256 FLAGS(ret) = get_regex_charset(RExC_flags);
10258 goto finish_meta_pat;
10260 op = SPACE + get_regex_charset(RExC_flags);
10261 if (op > SPACEA) { /* /aa is same as /a */
10264 ret = reg_node(pRExC_state, op);
10265 *flagp |= HASWIDTH|SIMPLE;
10266 goto finish_meta_pat;
10268 op = NSPACE + get_regex_charset(RExC_flags);
10269 if (op > NSPACEA) { /* /aa is same as /a */
10272 ret = reg_node(pRExC_state, op);
10273 *flagp |= HASWIDTH|SIMPLE;
10274 goto finish_meta_pat;
10282 U8 offset = get_regex_charset(RExC_flags);
10283 if (offset == REGEX_UNICODE_CHARSET) {
10284 offset = REGEX_DEPENDS_CHARSET;
10286 else if (offset == REGEX_ASCII_MORE_RESTRICTED_CHARSET) {
10287 offset = REGEX_ASCII_RESTRICTED_CHARSET;
10291 ret = reg_node(pRExC_state, op);
10292 *flagp |= HASWIDTH|SIMPLE;
10293 goto finish_meta_pat;
10295 ret = reg_node(pRExC_state, LNBREAK);
10296 *flagp |= HASWIDTH|SIMPLE;
10297 goto finish_meta_pat;
10299 ret = reg_node(pRExC_state, HORIZWS);
10300 *flagp |= HASWIDTH|SIMPLE;
10301 goto finish_meta_pat;
10303 ret = reg_node(pRExC_state, NHORIZWS);
10304 *flagp |= HASWIDTH|SIMPLE;
10305 goto finish_meta_pat;
10307 ret = reg_node(pRExC_state, VERTWS);
10308 *flagp |= HASWIDTH|SIMPLE;
10309 goto finish_meta_pat;
10311 ret = reg_node(pRExC_state, NVERTWS);
10312 *flagp |= HASWIDTH|SIMPLE;
10314 nextchar(pRExC_state);
10315 Set_Node_Length(ret, 2); /* MJD */
10320 char* const oldregxend = RExC_end;
10322 char* parse_start = RExC_parse - 2;
10325 if (RExC_parse[1] == '{') {
10326 /* a lovely hack--pretend we saw [\pX] instead */
10327 RExC_end = strchr(RExC_parse, '}');
10329 const U8 c = (U8)*RExC_parse;
10331 RExC_end = oldregxend;
10332 vFAIL2("Missing right brace on \\%c{}", c);
10337 RExC_end = RExC_parse + 2;
10338 if (RExC_end > oldregxend)
10339 RExC_end = oldregxend;
10343 ret = regclass(pRExC_state, flagp,depth+1);
10345 RExC_end = oldregxend;
10348 Set_Node_Offset(ret, parse_start + 2);
10349 Set_Node_Cur_Length(ret);
10350 nextchar(pRExC_state);
10354 /* Handle \N and \N{NAME} with multiple code points here and not
10355 * below because it can be multicharacter. join_exact() will join
10356 * them up later on. Also this makes sure that things like
10357 * /\N{BLAH}+/ and \N{BLAH} being multi char Just Happen. dmq.
10358 * The options to the grok function call causes it to fail if the
10359 * sequence is just a single code point. We then go treat it as
10360 * just another character in the current EXACT node, and hence it
10361 * gets uniform treatment with all the other characters. The
10362 * special treatment for quantifiers is not needed for such single
10363 * character sequences */
10365 if (! grok_bslash_N(pRExC_state, &ret, NULL, flagp, depth, FALSE)) {
10370 case 'k': /* Handle \k<NAME> and \k'NAME' */
10373 char ch= RExC_parse[1];
10374 if (ch != '<' && ch != '\'' && ch != '{') {
10376 vFAIL2("Sequence %.2s... not terminated",parse_start);
10378 /* this pretty much dupes the code for (?P=...) in reg(), if
10379 you change this make sure you change that */
10380 char* name_start = (RExC_parse += 2);
10382 SV *sv_dat = reg_scan_name(pRExC_state,
10383 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10384 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
10385 if (RExC_parse == name_start || *RExC_parse != ch)
10386 vFAIL2("Sequence %.3s... not terminated",parse_start);
10389 num = add_data( pRExC_state, 1, "S" );
10390 RExC_rxi->data->data[num]=(void*)sv_dat;
10391 SvREFCNT_inc_simple_void(sv_dat);
10395 ret = reganode(pRExC_state,
10398 : (ASCII_FOLD_RESTRICTED)
10400 : (AT_LEAST_UNI_SEMANTICS)
10406 *flagp |= HASWIDTH;
10408 /* override incorrect value set in reganode MJD */
10409 Set_Node_Offset(ret, parse_start+1);
10410 Set_Node_Cur_Length(ret); /* MJD */
10411 nextchar(pRExC_state);
10417 case '1': case '2': case '3': case '4':
10418 case '5': case '6': case '7': case '8': case '9':
10421 bool isg = *RExC_parse == 'g';
10426 if (*RExC_parse == '{') {
10430 if (*RExC_parse == '-') {
10434 if (hasbrace && !isDIGIT(*RExC_parse)) {
10435 if (isrel) RExC_parse--;
10437 goto parse_named_seq;
10439 num = atoi(RExC_parse);
10440 if (isg && num == 0)
10441 vFAIL("Reference to invalid group 0");
10443 num = RExC_npar - num;
10445 vFAIL("Reference to nonexistent or unclosed group");
10447 if (!isg && num > 9 && num >= RExC_npar)
10448 /* Probably a character specified in octal, e.g. \35 */
10451 char * const parse_start = RExC_parse - 1; /* MJD */
10452 while (isDIGIT(*RExC_parse))
10454 if (parse_start == RExC_parse - 1)
10455 vFAIL("Unterminated \\g... pattern");
10457 if (*RExC_parse != '}')
10458 vFAIL("Unterminated \\g{...} pattern");
10462 if (num > (I32)RExC_rx->nparens)
10463 vFAIL("Reference to nonexistent group");
10466 ret = reganode(pRExC_state,
10469 : (ASCII_FOLD_RESTRICTED)
10471 : (AT_LEAST_UNI_SEMANTICS)
10477 *flagp |= HASWIDTH;
10479 /* override incorrect value set in reganode MJD */
10480 Set_Node_Offset(ret, parse_start+1);
10481 Set_Node_Cur_Length(ret); /* MJD */
10483 nextchar(pRExC_state);
10488 if (RExC_parse >= RExC_end)
10489 FAIL("Trailing \\");
10492 /* Do not generate "unrecognized" warnings here, we fall
10493 back into the quick-grab loop below */
10500 if (RExC_flags & RXf_PMf_EXTENDED) {
10501 if ( reg_skipcomment( pRExC_state ) )
10508 parse_start = RExC_parse - 1;
10517 #define MAX_NODE_STRING_SIZE 127
10518 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
10520 U8 upper_parse = MAX_NODE_STRING_SIZE;
10523 bool next_is_quantifier;
10524 char * oldp = NULL;
10527 node_type = compute_EXACTish(pRExC_state);
10528 ret = reg_node(pRExC_state, node_type);
10530 /* In pass1, folded, we use a temporary buffer instead of the
10531 * actual node, as the node doesn't exist yet */
10532 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
10538 /* XXX The node can hold up to 255 bytes, yet this only goes to
10539 * 127. I (khw) do not know why. Keeping it somewhat less than
10540 * 255 allows us to not have to worry about overflow due to
10541 * converting to utf8 and fold expansion, but that value is
10542 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
10543 * split up by this limit into a single one using the real max of
10544 * 255. Even at 127, this breaks under rare circumstances. If
10545 * folding, we do not want to split a node at a character that is a
10546 * non-final in a multi-char fold, as an input string could just
10547 * happen to want to match across the node boundary. The join
10548 * would solve that problem if the join actually happens. But a
10549 * series of more than two nodes in a row each of 127 would cause
10550 * the first join to succeed to get to 254, but then there wouldn't
10551 * be room for the next one, which could at be one of those split
10552 * multi-char folds. I don't know of any fool-proof solution. One
10553 * could back off to end with only a code point that isn't such a
10554 * non-final, but it is possible for there not to be any in the
10556 for (p = RExC_parse - 1;
10557 len < upper_parse && p < RExC_end;
10562 if (RExC_flags & RXf_PMf_EXTENDED)
10563 p = regwhite( pRExC_state, p );
10574 /* Literal Escapes Switch
10576 This switch is meant to handle escape sequences that
10577 resolve to a literal character.
10579 Every escape sequence that represents something
10580 else, like an assertion or a char class, is handled
10581 in the switch marked 'Special Escapes' above in this
10582 routine, but also has an entry here as anything that
10583 isn't explicitly mentioned here will be treated as
10584 an unescaped equivalent literal.
10587 switch ((U8)*++p) {
10588 /* These are all the special escapes. */
10589 case 'A': /* Start assertion */
10590 case 'b': case 'B': /* Word-boundary assertion*/
10591 case 'C': /* Single char !DANGEROUS! */
10592 case 'd': case 'D': /* digit class */
10593 case 'g': case 'G': /* generic-backref, pos assertion */
10594 case 'h': case 'H': /* HORIZWS */
10595 case 'k': case 'K': /* named backref, keep marker */
10596 case 'p': case 'P': /* Unicode property */
10597 case 'R': /* LNBREAK */
10598 case 's': case 'S': /* space class */
10599 case 'v': case 'V': /* VERTWS */
10600 case 'w': case 'W': /* word class */
10601 case 'X': /* eXtended Unicode "combining character sequence" */
10602 case 'z': case 'Z': /* End of line/string assertion */
10606 /* Anything after here is an escape that resolves to a
10607 literal. (Except digits, which may or may not)
10613 case 'N': /* Handle a single-code point named character. */
10614 /* The options cause it to fail if a multiple code
10615 * point sequence. Handle those in the switch() above
10617 RExC_parse = p + 1;
10618 if (! grok_bslash_N(pRExC_state, NULL, &ender,
10619 flagp, depth, FALSE))
10621 RExC_parse = p = oldp;
10625 if (ender > 0xff) {
10642 ender = ASCII_TO_NATIVE('\033');
10646 ender = ASCII_TO_NATIVE('\007');
10651 STRLEN brace_len = len;
10653 const char* error_msg;
10655 bool valid = grok_bslash_o(p,
10662 RExC_parse = p; /* going to die anyway; point
10663 to exact spot of failure */
10670 if (PL_encoding && ender < 0x100) {
10671 goto recode_encoding;
10673 if (ender > 0xff) {
10680 STRLEN brace_len = len;
10682 const char* error_msg;
10684 bool valid = grok_bslash_x(p,
10691 RExC_parse = p; /* going to die anyway; point
10692 to exact spot of failure */
10698 if (PL_encoding && ender < 0x100) {
10699 goto recode_encoding;
10701 if (ender > 0xff) {
10708 ender = grok_bslash_c(*p++, UTF, SIZE_ONLY);
10710 case '0': case '1': case '2': case '3':case '4':
10711 case '5': case '6': case '7':
10713 (isDIGIT(p[1]) && atoi(p) >= RExC_npar))
10715 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
10717 ender = grok_oct(p, &numlen, &flags, NULL);
10718 if (ender > 0xff) {
10727 if (PL_encoding && ender < 0x100)
10728 goto recode_encoding;
10731 if (! RExC_override_recoding) {
10732 SV* enc = PL_encoding;
10733 ender = reg_recode((const char)(U8)ender, &enc);
10734 if (!enc && SIZE_ONLY)
10735 ckWARNreg(p, "Invalid escape in the specified encoding");
10741 FAIL("Trailing \\");
10744 if (!SIZE_ONLY&& isALNUMC(*p)) {
10745 ckWARN2reg(p + 1, "Unrecognized escape \\%.1s passed through", p);
10747 goto normal_default;
10751 /* Currently we don't warn when the lbrace is at the start
10752 * of a construct. This catches it in the middle of a
10753 * literal string, or when its the first thing after
10754 * something like "\b" */
10756 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
10758 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
10763 if (UTF8_IS_START(*p) && UTF) {
10765 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
10766 &numlen, UTF8_ALLOW_DEFAULT);
10772 } /* End of switch on the literal */
10774 /* Here, have looked at the literal character and <ender>
10775 * contains its ordinal, <p> points to the character after it
10778 if ( RExC_flags & RXf_PMf_EXTENDED)
10779 p = regwhite( pRExC_state, p );
10781 /* If the next thing is a quantifier, it applies to this
10782 * character only, which means that this character has to be in
10783 * its own node and can't just be appended to the string in an
10784 * existing node, so if there are already other characters in
10785 * the node, close the node with just them, and set up to do
10786 * this character again next time through, when it will be the
10787 * only thing in its new node */
10788 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
10796 /* See comments for join_exact() as to why we fold
10797 * this non-UTF at compile time */
10798 || (node_type == EXACTFU
10799 && ender == LATIN_SMALL_LETTER_SHARP_S))
10803 /* Prime the casefolded buffer. Locale rules, which
10804 * apply only to code points < 256, aren't known until
10805 * execution, so for them, just output the original
10806 * character using utf8. If we start to fold non-UTF
10807 * patterns, be sure to update join_exact() */
10808 if (LOC && ender < 256) {
10809 if (UNI_IS_INVARIANT(ender)) {
10813 *s = UTF8_TWO_BYTE_HI(ender);
10814 *(s + 1) = UTF8_TWO_BYTE_LO(ender);
10819 ender = _to_uni_fold_flags(ender, (U8 *) s, &foldlen,
10821 | ((LOC) ? FOLD_FLAGS_LOCALE
10822 : (ASCII_FOLD_RESTRICTED)
10823 ? FOLD_FLAGS_NOMIX_ASCII
10829 /* The loop increments <len> each time, as all but this
10830 * path (and the one just below for UTF) through it add
10831 * a single byte to the EXACTish node. But this one
10832 * has changed len to be the correct final value, so
10833 * subtract one to cancel out the increment that
10835 len += foldlen - 1;
10842 const STRLEN unilen = reguni(pRExC_state, ender, s);
10848 /* See comment just above for - 1 */
10852 REGC((char)ender, s++);
10855 if (next_is_quantifier) {
10857 /* Here, the next input is a quantifier, and to get here,
10858 * the current character is the only one in the node.
10859 * Also, here <len> doesn't include the final byte for this
10865 } /* End of loop through literal characters */
10867 /* Here we have either exhausted the input or ran out of room in
10868 * the node. (If we encountered a character that can't be in the
10869 * node, transfer is made directly to <loopdone>, and so we
10870 * wouldn't have fallen off the end of the loop.) In the latter
10871 * case, we artificially have to split the node into two, because
10872 * we just don't have enough space to hold everything. This
10873 * creates a problem if the final character participates in a
10874 * multi-character fold in the non-final position, as a match that
10875 * should have occurred won't, due to the way nodes are matched,
10876 * and our artificial boundary. So back off until we find a non-
10877 * problematic character -- one that isn't at the beginning or
10878 * middle of such a fold. (Either it doesn't participate in any
10879 * folds, or appears only in the final position of all the folds it
10880 * does participate in.) A better solution with far fewer false
10881 * positives, and that would fill the nodes more completely, would
10882 * be to actually have available all the multi-character folds to
10883 * test against, and to back-off only far enough to be sure that
10884 * this node isn't ending with a partial one. <upper_parse> is set
10885 * further below (if we need to reparse the node) to include just
10886 * up through that final non-problematic character that this code
10887 * identifies, so when it is set to less than the full node, we can
10888 * skip the rest of this */
10889 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
10891 const STRLEN full_len = len;
10893 assert(len >= MAX_NODE_STRING_SIZE);
10895 /* Here, <s> points to the final byte of the final character.
10896 * Look backwards through the string until find a non-
10897 * problematic character */
10901 /* These two have no multi-char folds to non-UTF characters
10903 if (ASCII_FOLD_RESTRICTED || LOC) {
10907 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
10911 if (! PL_NonL1NonFinalFold) {
10912 PL_NonL1NonFinalFold = _new_invlist_C_array(
10913 NonL1_Perl_Non_Final_Folds_invlist);
10916 /* Point to the first byte of the final character */
10917 s = (char *) utf8_hop((U8 *) s, -1);
10919 while (s >= s0) { /* Search backwards until find
10920 non-problematic char */
10921 if (UTF8_IS_INVARIANT(*s)) {
10923 /* There are no ascii characters that participate
10924 * in multi-char folds under /aa. In EBCDIC, the
10925 * non-ascii invariants are all control characters,
10926 * so don't ever participate in any folds. */
10927 if (ASCII_FOLD_RESTRICTED
10928 || ! IS_NON_FINAL_FOLD(*s))
10933 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
10935 /* No Latin1 characters participate in multi-char
10936 * folds under /l */
10938 || ! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_UNI(
10944 else if (! _invlist_contains_cp(
10945 PL_NonL1NonFinalFold,
10946 valid_utf8_to_uvchr((U8 *) s, NULL)))
10951 /* Here, the current character is problematic in that
10952 * it does occur in the non-final position of some
10953 * fold, so try the character before it, but have to
10954 * special case the very first byte in the string, so
10955 * we don't read outside the string */
10956 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
10957 } /* End of loop backwards through the string */
10959 /* If there were only problematic characters in the string,
10960 * <s> will point to before s0, in which case the length
10961 * should be 0, otherwise include the length of the
10962 * non-problematic character just found */
10963 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
10966 /* Here, have found the final character, if any, that is
10967 * non-problematic as far as ending the node without splitting
10968 * it across a potential multi-char fold. <len> contains the
10969 * number of bytes in the node up-to and including that
10970 * character, or is 0 if there is no such character, meaning
10971 * the whole node contains only problematic characters. In
10972 * this case, give up and just take the node as-is. We can't
10978 /* Here, the node does contain some characters that aren't
10979 * problematic. If one such is the final character in the
10980 * node, we are done */
10981 if (len == full_len) {
10984 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
10986 /* If the final character is problematic, but the
10987 * penultimate is not, back-off that last character to
10988 * later start a new node with it */
10993 /* Here, the final non-problematic character is earlier
10994 * in the input than the penultimate character. What we do
10995 * is reparse from the beginning, going up only as far as
10996 * this final ok one, thus guaranteeing that the node ends
10997 * in an acceptable character. The reason we reparse is
10998 * that we know how far in the character is, but we don't
10999 * know how to correlate its position with the input parse.
11000 * An alternate implementation would be to build that
11001 * correlation as we go along during the original parse,
11002 * but that would entail extra work for every node, whereas
11003 * this code gets executed only when the string is too
11004 * large for the node, and the final two characters are
11005 * problematic, an infrequent occurrence. Yet another
11006 * possible strategy would be to save the tail of the
11007 * string, and the next time regatom is called, initialize
11008 * with that. The problem with this is that unless you
11009 * back off one more character, you won't be guaranteed
11010 * regatom will get called again, unless regbranch,
11011 * regpiece ... are also changed. If you do back off that
11012 * extra character, so that there is input guaranteed to
11013 * force calling regatom, you can't handle the case where
11014 * just the first character in the node is acceptable. I
11015 * (khw) decided to try this method which doesn't have that
11016 * pitfall; if performance issues are found, we can do a
11017 * combination of the current approach plus that one */
11023 } /* End of verifying node ends with an appropriate char */
11025 loopdone: /* Jumped to when encounters something that shouldn't be in
11028 /* I (khw) don't know if you can get here with zero length, but the
11029 * old code handled this situation by creating a zero-length EXACT
11030 * node. Might as well be NOTHING instead */
11035 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender);
11038 RExC_parse = p - 1;
11039 Set_Node_Cur_Length(ret); /* MJD */
11040 nextchar(pRExC_state);
11042 /* len is STRLEN which is unsigned, need to copy to signed */
11045 vFAIL("Internal disaster");
11048 } /* End of label 'defchar:' */
11050 } /* End of giant switch on input character */
11056 S_regwhite( RExC_state_t *pRExC_state, char *p )
11058 const char *e = RExC_end;
11060 PERL_ARGS_ASSERT_REGWHITE;
11065 else if (*p == '#') {
11068 if (*p++ == '\n') {
11074 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
11082 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
11083 Character classes ([:foo:]) can also be negated ([:^foo:]).
11084 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
11085 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
11086 but trigger failures because they are currently unimplemented. */
11088 #define POSIXCC_DONE(c) ((c) == ':')
11089 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
11090 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
11093 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value)
11096 I32 namedclass = OOB_NAMEDCLASS;
11098 PERL_ARGS_ASSERT_REGPPOSIXCC;
11100 if (value == '[' && RExC_parse + 1 < RExC_end &&
11101 /* I smell either [: or [= or [. -- POSIX has been here, right? */
11102 POSIXCC(UCHARAT(RExC_parse))) {
11103 const char c = UCHARAT(RExC_parse);
11104 char* const s = RExC_parse++;
11106 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
11108 if (RExC_parse == RExC_end)
11109 /* Grandfather lone [:, [=, [. */
11112 const char* const t = RExC_parse++; /* skip over the c */
11115 if (UCHARAT(RExC_parse) == ']') {
11116 const char *posixcc = s + 1;
11117 RExC_parse++; /* skip over the ending ] */
11120 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
11121 const I32 skip = t - posixcc;
11123 /* Initially switch on the length of the name. */
11126 if (memEQ(posixcc, "word", 4)) /* this is not POSIX, this is the Perl \w */
11127 namedclass = ANYOF_ALNUM;
11130 /* Names all of length 5. */
11131 /* alnum alpha ascii blank cntrl digit graph lower
11132 print punct space upper */
11133 /* Offset 4 gives the best switch position. */
11134 switch (posixcc[4]) {
11136 if (memEQ(posixcc, "alph", 4)) /* alpha */
11137 namedclass = ANYOF_ALPHA;
11140 if (memEQ(posixcc, "spac", 4)) /* space */
11141 namedclass = ANYOF_PSXSPC;
11144 if (memEQ(posixcc, "grap", 4)) /* graph */
11145 namedclass = ANYOF_GRAPH;
11148 if (memEQ(posixcc, "asci", 4)) /* ascii */
11149 namedclass = ANYOF_ASCII;
11152 if (memEQ(posixcc, "blan", 4)) /* blank */
11153 namedclass = ANYOF_BLANK;
11156 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
11157 namedclass = ANYOF_CNTRL;
11160 if (memEQ(posixcc, "alnu", 4)) /* alnum */
11161 namedclass = ANYOF_ALNUMC;
11164 if (memEQ(posixcc, "lowe", 4)) /* lower */
11165 namedclass = ANYOF_LOWER;
11166 else if (memEQ(posixcc, "uppe", 4)) /* upper */
11167 namedclass = ANYOF_UPPER;
11170 if (memEQ(posixcc, "digi", 4)) /* digit */
11171 namedclass = ANYOF_DIGIT;
11172 else if (memEQ(posixcc, "prin", 4)) /* print */
11173 namedclass = ANYOF_PRINT;
11174 else if (memEQ(posixcc, "punc", 4)) /* punct */
11175 namedclass = ANYOF_PUNCT;
11180 if (memEQ(posixcc, "xdigit", 6))
11181 namedclass = ANYOF_XDIGIT;
11185 if (namedclass == OOB_NAMEDCLASS)
11186 Simple_vFAIL3("POSIX class [:%.*s:] unknown",
11189 /* The #defines are structured so each complement is +1 to
11190 * the normal one */
11194 assert (posixcc[skip] == ':');
11195 assert (posixcc[skip+1] == ']');
11196 } else if (!SIZE_ONLY) {
11197 /* [[=foo=]] and [[.foo.]] are still future. */
11199 /* adjust RExC_parse so the warning shows after
11200 the class closes */
11201 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
11203 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11206 /* Maternal grandfather:
11207 * "[:" ending in ":" but not in ":]" */
11217 S_checkposixcc(pTHX_ RExC_state_t *pRExC_state)
11221 PERL_ARGS_ASSERT_CHECKPOSIXCC;
11223 if (POSIXCC(UCHARAT(RExC_parse))) {
11224 const char *s = RExC_parse;
11225 const char c = *s++;
11227 while (isALNUM(*s))
11229 if (*s && c == *s && s[1] == ']') {
11231 "POSIX syntax [%c %c] belongs inside character classes",
11234 /* [[=foo=]] and [[.foo.]] are still future. */
11235 if (POSIXCC_NOTYET(c)) {
11236 /* adjust RExC_parse so the error shows after
11237 the class closes */
11238 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse++) != ']')
11240 Simple_vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
11246 /* Generate the code to add a full posix character <class> to the bracketed
11247 * character class given by <node>. (<node> is needed only under locale rules)
11248 * destlist is the inversion list for non-locale rules that this class is
11250 * sourcelist is the ASCII-range inversion list to add under /a rules
11251 * Xsourcelist is the full Unicode range list to use otherwise. */
11252 #define DO_POSIX(node, class, destlist, sourcelist, Xsourcelist) \
11254 SV* scratch_list = NULL; \
11256 /* Set this class in the node for runtime matching */ \
11257 ANYOF_CLASS_SET(node, class); \
11259 /* For above Latin1 code points, we use the full Unicode range */ \
11260 _invlist_intersection(PL_AboveLatin1, \
11263 /* And set the output to it, adding instead if there already is an \
11264 * output. Checking if <destlist> is NULL first saves an extra \
11265 * clone. Its reference count will be decremented at the next \
11266 * union, etc, or if this is the only instance, at the end of the \
11268 if (! destlist) { \
11269 destlist = scratch_list; \
11272 _invlist_union(destlist, scratch_list, &destlist); \
11273 SvREFCNT_dec(scratch_list); \
11277 /* For non-locale, just add it to any existing list */ \
11278 _invlist_union(destlist, \
11279 (AT_LEAST_ASCII_RESTRICTED) \
11285 /* Like DO_POSIX, but matches the complement of <sourcelist> and <Xsourcelist>.
11287 #define DO_N_POSIX(node, class, destlist, sourcelist, Xsourcelist) \
11289 SV* scratch_list = NULL; \
11290 ANYOF_CLASS_SET(node, class); \
11291 _invlist_subtract(PL_AboveLatin1, Xsourcelist, &scratch_list); \
11292 if (! destlist) { \
11293 destlist = scratch_list; \
11296 _invlist_union(destlist, scratch_list, &destlist); \
11297 SvREFCNT_dec(scratch_list); \
11301 _invlist_union_complement_2nd(destlist, \
11302 (AT_LEAST_ASCII_RESTRICTED) \
11306 /* Under /d, everything in the upper half of the Latin1 range \
11307 * matches this complement */ \
11308 if (DEPENDS_SEMANTICS) { \
11309 ANYOF_FLAGS(node) |= ANYOF_NON_UTF8_LATIN1_ALL; \
11313 /* Generate the code to add a posix character <class> to the bracketed
11314 * character class given by <node>. (<node> is needed only under locale rules)
11315 * destlist is the inversion list for non-locale rules that this class is
11317 * sourcelist is the ASCII-range inversion list to add under /a rules
11318 * l1_sourcelist is the Latin1 range list to use otherwise.
11319 * Xpropertyname is the name to add to <run_time_list> of the property to
11320 * specify the code points above Latin1 that will have to be
11321 * determined at run-time
11322 * run_time_list is a SV* that contains text names of properties that are to
11323 * be computed at run time. This concatenates <Xpropertyname>
11324 * to it, appropriately
11325 * This is essentially DO_POSIX, but we know only the Latin1 values at compile
11327 #define DO_POSIX_LATIN1_ONLY_KNOWN(node, class, destlist, sourcelist, \
11328 l1_sourcelist, Xpropertyname, run_time_list) \
11329 /* First, resolve whether to use the ASCII-only list or the L1 \
11331 DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(node, class, destlist, \
11332 ((AT_LEAST_ASCII_RESTRICTED) ? sourcelist : l1_sourcelist),\
11333 Xpropertyname, run_time_list)
11335 #define DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(node, class, destlist, sourcelist, \
11336 Xpropertyname, run_time_list) \
11337 /* If not /a matching, there are going to be code points we will have \
11338 * to defer to runtime to look-up */ \
11339 if (! AT_LEAST_ASCII_RESTRICTED) { \
11340 Perl_sv_catpvf(aTHX_ run_time_list, "+utf8::%s\n", Xpropertyname); \
11343 ANYOF_CLASS_SET(node, class); \
11346 _invlist_union(destlist, sourcelist, &destlist); \
11349 /* Like DO_POSIX_LATIN1_ONLY_KNOWN, but for the complement. A combination of
11350 * this and DO_N_POSIX. Sets <matches_above_unicode> only if it can; unchanged
11352 #define DO_N_POSIX_LATIN1_ONLY_KNOWN(node, class, destlist, sourcelist, \
11353 l1_sourcelist, Xpropertyname, run_time_list, matches_above_unicode) \
11354 if (AT_LEAST_ASCII_RESTRICTED) { \
11355 _invlist_union_complement_2nd(destlist, sourcelist, &destlist); \
11358 Perl_sv_catpvf(aTHX_ run_time_list, "!utf8::%s\n", Xpropertyname); \
11359 matches_above_unicode = TRUE; \
11361 ANYOF_CLASS_SET(node, namedclass); \
11364 SV* scratch_list = NULL; \
11365 _invlist_subtract(PL_Latin1, l1_sourcelist, &scratch_list); \
11366 if (! destlist) { \
11367 destlist = scratch_list; \
11370 _invlist_union(destlist, scratch_list, &destlist); \
11371 SvREFCNT_dec(scratch_list); \
11373 if (DEPENDS_SEMANTICS) { \
11374 ANYOF_FLAGS(node) |= ANYOF_NON_UTF8_LATIN1_ALL; \
11380 S_add_alternate(pTHX_ AV** alternate_ptr, U8* string, STRLEN len)
11382 /* Adds input 'string' with length 'len' to the ANYOF node's unicode
11383 * alternate list, pointed to by 'alternate_ptr'. This is an array of
11384 * the multi-character folds of characters in the node */
11387 PERL_ARGS_ASSERT_ADD_ALTERNATE;
11389 if (! *alternate_ptr) {
11390 *alternate_ptr = newAV();
11392 sv = newSVpvn_utf8((char*)string, len, TRUE);
11393 av_push(*alternate_ptr, sv);
11397 /* The names of properties whose definitions are not known at compile time are
11398 * stored in this SV, after a constant heading. So if the length has been
11399 * changed since initialization, then there is a run-time definition. */
11400 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION (SvCUR(listsv) != initial_listsv_len)
11402 /* This converts the named class defined in regcomp.h to its equivalent class
11403 * number defined in handy.h. */
11404 #define namedclass_to_classnum(class) ((class) / 2)
11407 parse a class specification and produce either an ANYOF node that
11408 matches the pattern or perhaps will be optimized into an EXACTish node
11409 instead. The node contains a bit map for the first 256 characters, with the
11410 corresponding bit set if that character is in the list. For characters
11411 above 255, a range list is used */
11414 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11418 UV prevvalue = OOB_UNICODE;
11423 IV namedclass = OOB_NAMEDCLASS;
11424 char *rangebegin = NULL;
11425 bool need_class = 0;
11426 bool allow_full_fold = TRUE; /* Assume wants multi-char folding */
11428 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
11429 than just initialized. */
11430 SV* properties = NULL; /* Code points that match \p{} \P{} */
11431 SV* posixes = NULL; /* Code points that match classes like, [:word:],
11432 extended beyond the Latin1 range */
11433 UV element_count = 0; /* Number of distinct elements in the class.
11434 Optimizations may be possible if this is tiny */
11437 /* Unicode properties are stored in a swash; this holds the current one
11438 * being parsed. If this swash is the only above-latin1 component of the
11439 * character class, an optimization is to pass it directly on to the
11440 * execution engine. Otherwise, it is set to NULL to indicate that there
11441 * are other things in the class that have to be dealt with at execution
11443 SV* swash = NULL; /* Code points that match \p{} \P{} */
11445 /* Set if a component of this character class is user-defined; just passed
11446 * on to the engine */
11447 bool has_user_defined_property = FALSE;
11449 /* inversion list of code points this node matches only when the target
11450 * string is in UTF-8. (Because is under /d) */
11451 SV* depends_list = NULL;
11453 /* inversion list of code points this node matches. For much of the
11454 * function, it includes only those that match regardless of the utf8ness
11455 * of the target string */
11456 SV* cp_list = NULL;
11458 /* List of multi-character folds that are matched by this node */
11459 AV* unicode_alternate = NULL;
11461 /* In a range, counts how many 0-2 of the ends of it came from literals,
11462 * not escapes. Thus we can tell if 'A' was input vs \x{C1} */
11463 UV literal_endpoint = 0;
11465 bool invert = FALSE; /* Is this class to be complemented */
11467 /* Is there any thing like \W or [:^digit:] that matches above the legal
11468 * Unicode range? */
11469 bool runtime_posix_matches_above_Unicode = FALSE;
11471 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
11472 case we need to change the emitted regop to an EXACT. */
11473 const char * orig_parse = RExC_parse;
11474 const I32 orig_size = RExC_size;
11475 GET_RE_DEBUG_FLAGS_DECL;
11477 PERL_ARGS_ASSERT_REGCLASS;
11479 PERL_UNUSED_ARG(depth);
11482 DEBUG_PARSE("clas");
11484 /* Assume we are going to generate an ANYOF node. */
11485 ret = reganode(pRExC_state, ANYOF, 0);
11489 ANYOF_FLAGS(ret) = 0;
11492 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
11497 /* We have decided to not allow multi-char folds in inverted character
11498 * classes, due to the confusion that can happen, especially with
11499 * classes that are designed for a non-Unicode world: You have the
11500 * peculiar case that:
11501 "s s" =~ /^[^\xDF]+$/i => Y
11502 "ss" =~ /^[^\xDF]+$/i => N
11504 * See [perl #89750] */
11505 allow_full_fold = FALSE;
11509 RExC_size += ANYOF_SKIP;
11510 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
11513 RExC_emit += ANYOF_SKIP;
11515 ANYOF_FLAGS(ret) |= ANYOF_LOCALE;
11517 listsv = newSVpvs("# comment\n");
11518 initial_listsv_len = SvCUR(listsv);
11521 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
11523 if (!SIZE_ONLY && POSIXCC(nextvalue))
11524 checkposixcc(pRExC_state);
11526 /* allow 1st char to be ] (allowing it to be - is dealt with later) */
11527 if (UCHARAT(RExC_parse) == ']')
11528 goto charclassloop;
11531 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
11535 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
11538 rangebegin = RExC_parse;
11542 value = utf8n_to_uvchr((U8*)RExC_parse,
11543 RExC_end - RExC_parse,
11544 &numlen, UTF8_ALLOW_DEFAULT);
11545 RExC_parse += numlen;
11548 value = UCHARAT(RExC_parse++);
11550 nextvalue = RExC_parse < RExC_end ? UCHARAT(RExC_parse) : 0;
11551 if (value == '[' && POSIXCC(nextvalue))
11552 namedclass = regpposixcc(pRExC_state, value);
11553 else if (value == '\\') {
11555 value = utf8n_to_uvchr((U8*)RExC_parse,
11556 RExC_end - RExC_parse,
11557 &numlen, UTF8_ALLOW_DEFAULT);
11558 RExC_parse += numlen;
11561 value = UCHARAT(RExC_parse++);
11562 /* Some compilers cannot handle switching on 64-bit integer
11563 * values, therefore value cannot be an UV. Yes, this will
11564 * be a problem later if we want switch on Unicode.
11565 * A similar issue a little bit later when switching on
11566 * namedclass. --jhi */
11567 switch ((I32)value) {
11568 case 'w': namedclass = ANYOF_ALNUM; break;
11569 case 'W': namedclass = ANYOF_NALNUM; break;
11570 case 's': namedclass = ANYOF_SPACE; break;
11571 case 'S': namedclass = ANYOF_NSPACE; break;
11572 case 'd': namedclass = ANYOF_DIGIT; break;
11573 case 'D': namedclass = ANYOF_NDIGIT; break;
11574 case 'v': namedclass = ANYOF_VERTWS; break;
11575 case 'V': namedclass = ANYOF_NVERTWS; break;
11576 case 'h': namedclass = ANYOF_HORIZWS; break;
11577 case 'H': namedclass = ANYOF_NHORIZWS; break;
11578 case 'N': /* Handle \N{NAME} in class */
11580 /* We only pay attention to the first char of
11581 multichar strings being returned. I kinda wonder
11582 if this makes sense as it does change the behaviour
11583 from earlier versions, OTOH that behaviour was broken
11585 if (! grok_bslash_N(pRExC_state, NULL, &value, flagp, depth,
11586 TRUE /* => charclass */))
11597 /* This routine will handle any undefined properties */
11598 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF;
11600 if (RExC_parse >= RExC_end)
11601 vFAIL2("Empty \\%c{}", (U8)value);
11602 if (*RExC_parse == '{') {
11603 const U8 c = (U8)value;
11604 e = strchr(RExC_parse++, '}');
11606 vFAIL2("Missing right brace on \\%c{}", c);
11607 while (isSPACE(UCHARAT(RExC_parse)))
11609 if (e == RExC_parse)
11610 vFAIL2("Empty \\%c{}", c);
11611 n = e - RExC_parse;
11612 while (isSPACE(UCHARAT(RExC_parse + n - 1)))
11623 if (UCHARAT(RExC_parse) == '^') {
11626 value = value == 'p' ? 'P' : 'p'; /* toggle */
11627 while (isSPACE(UCHARAT(RExC_parse))) {
11632 /* Try to get the definition of the property into
11633 * <invlist>. If /i is in effect, the effective property
11634 * will have its name be <__NAME_i>. The design is
11635 * discussed in commit
11636 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
11637 Newx(name, n + sizeof("_i__\n"), char);
11639 sprintf(name, "%s%.*s%s\n",
11640 (FOLD) ? "__" : "",
11646 /* Look up the property name, and get its swash and
11647 * inversion list, if the property is found */
11649 SvREFCNT_dec(swash);
11651 swash = _core_swash_init("utf8", name, &PL_sv_undef,
11654 NULL, /* No inversion list */
11657 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
11659 SvREFCNT_dec(swash);
11663 /* Here didn't find it. It could be a user-defined
11664 * property that will be available at run-time. Add it
11665 * to the list to look up then */
11666 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%s\n",
11667 (value == 'p' ? '+' : '!'),
11669 has_user_defined_property = TRUE;
11671 /* We don't know yet, so have to assume that the
11672 * property could match something in the Latin1 range,
11673 * hence something that isn't utf8. Note that this
11674 * would cause things in <depends_list> to match
11675 * inappropriately, except that any \p{}, including
11676 * this one forces Unicode semantics, which means there
11677 * is <no depends_list> */
11678 ANYOF_FLAGS(ret) |= ANYOF_NONBITMAP_NON_UTF8;
11682 /* Here, did get the swash and its inversion list. If
11683 * the swash is from a user-defined property, then this
11684 * whole character class should be regarded as such */
11685 has_user_defined_property =
11687 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY);
11689 /* Invert if asking for the complement */
11690 if (value == 'P') {
11691 _invlist_union_complement_2nd(properties,
11695 /* The swash can't be used as-is, because we've
11696 * inverted things; delay removing it to here after
11697 * have copied its invlist above */
11698 SvREFCNT_dec(swash);
11702 _invlist_union(properties, invlist, &properties);
11707 RExC_parse = e + 1;
11708 namedclass = ANYOF_MAX; /* no official name, but it's named */
11710 /* \p means they want Unicode semantics */
11711 RExC_uni_semantics = 1;
11714 case 'n': value = '\n'; break;
11715 case 'r': value = '\r'; break;
11716 case 't': value = '\t'; break;
11717 case 'f': value = '\f'; break;
11718 case 'b': value = '\b'; break;
11719 case 'e': value = ASCII_TO_NATIVE('\033');break;
11720 case 'a': value = ASCII_TO_NATIVE('\007');break;
11722 RExC_parse--; /* function expects to be pointed at the 'o' */
11724 const char* error_msg;
11725 bool valid = grok_bslash_o(RExC_parse,
11730 RExC_parse += numlen;
11735 if (PL_encoding && value < 0x100) {
11736 goto recode_encoding;
11740 RExC_parse--; /* function expects to be pointed at the 'x' */
11742 const char* error_msg;
11743 bool valid = grok_bslash_x(RExC_parse,
11748 RExC_parse += numlen;
11753 if (PL_encoding && value < 0x100)
11754 goto recode_encoding;
11757 value = grok_bslash_c(*RExC_parse++, UTF, SIZE_ONLY);
11759 case '0': case '1': case '2': case '3': case '4':
11760 case '5': case '6': case '7':
11762 /* Take 1-3 octal digits */
11763 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
11765 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
11766 RExC_parse += numlen;
11767 if (PL_encoding && value < 0x100)
11768 goto recode_encoding;
11772 if (! RExC_override_recoding) {
11773 SV* enc = PL_encoding;
11774 value = reg_recode((const char)(U8)value, &enc);
11775 if (!enc && SIZE_ONLY)
11776 ckWARNreg(RExC_parse,
11777 "Invalid escape in the specified encoding");
11781 /* Allow \_ to not give an error */
11782 if (!SIZE_ONLY && isALNUM(value) && value != '_') {
11783 ckWARN2reg(RExC_parse,
11784 "Unrecognized escape \\%c in character class passed through",
11789 } /* end of \blah */
11792 literal_endpoint++;
11795 /* What matches in a locale is not known until runtime. This
11796 * includes what the Posix classes (like \w, [:space:]) match.
11797 * Room must be reserved (one time per class) to store such
11798 * classes, either if Perl is compiled so that locale nodes always
11799 * should have this space, or if there is such class info to be
11800 * stored. The space will contain a bit for each named class that
11801 * is to be matched against. This isn't needed for \p{} and
11802 * pseudo-classes, as they are not affected by locale, and hence
11803 * are dealt with separately */
11806 && (ANYOF_LOCALE == ANYOF_CLASS
11807 || (namedclass > OOB_NAMEDCLASS && namedclass < ANYOF_MAX)))
11811 RExC_size += ANYOF_CLASS_SKIP - ANYOF_SKIP;
11814 RExC_emit += ANYOF_CLASS_SKIP - ANYOF_SKIP;
11815 ANYOF_CLASS_ZERO(ret);
11817 ANYOF_FLAGS(ret) |= ANYOF_CLASS;
11820 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
11822 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
11823 * literal, as is the character that began the false range, i.e.
11824 * the 'a' in the examples */
11828 RExC_parse >= rangebegin ?
11829 RExC_parse - rangebegin : 0;
11830 ckWARN4reg(RExC_parse,
11831 "False [] range \"%*.*s\"",
11833 cp_list = add_cp_to_invlist(cp_list, '-');
11834 cp_list = add_cp_to_invlist(cp_list, prevvalue);
11837 range = 0; /* this was not a true range */
11838 element_count += 2; /* So counts for three values */
11842 switch ((I32)namedclass) {
11844 case ANYOF_ALNUMC: /* C's alnum, in contrast to \w */
11845 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11846 PL_PosixAlnum, PL_L1PosixAlnum, "XPosixAlnum", listsv);
11848 case ANYOF_NALNUMC:
11849 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11850 PL_PosixAlnum, PL_L1PosixAlnum, "XPosixAlnum", listsv,
11851 runtime_posix_matches_above_Unicode);
11854 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11855 PL_PosixAlpha, PL_L1PosixAlpha, "XPosixAlpha", listsv);
11858 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11859 PL_PosixAlpha, PL_L1PosixAlpha, "XPosixAlpha", listsv,
11860 runtime_posix_matches_above_Unicode);
11864 ANYOF_CLASS_SET(ret, namedclass);
11867 _invlist_union(posixes, PL_ASCII, &posixes);
11872 ANYOF_CLASS_SET(ret, namedclass);
11875 _invlist_union_complement_2nd(posixes,
11876 PL_ASCII, &posixes);
11877 if (DEPENDS_SEMANTICS) {
11878 ANYOF_FLAGS(ret) |= ANYOF_NON_UTF8_LATIN1_ALL;
11883 DO_POSIX(ret, namedclass, posixes,
11884 PL_PosixBlank, PL_XPosixBlank);
11887 DO_N_POSIX(ret, namedclass, posixes,
11888 PL_PosixBlank, PL_XPosixBlank);
11891 DO_POSIX(ret, namedclass, posixes,
11892 PL_PosixCntrl, PL_XPosixCntrl);
11895 DO_N_POSIX(ret, namedclass, posixes,
11896 PL_PosixCntrl, PL_XPosixCntrl);
11899 /* There are no digits in the Latin1 range outside of
11900 * ASCII, so call the macro that doesn't have to resolve
11902 DO_POSIX_LATIN1_ONLY_KNOWN_L1_RESOLVED(ret, namedclass, posixes,
11903 PL_PosixDigit, "XPosixDigit", listsv);
11906 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11907 PL_PosixDigit, PL_PosixDigit, "XPosixDigit", listsv,
11908 runtime_posix_matches_above_Unicode);
11911 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11912 PL_PosixGraph, PL_L1PosixGraph, "XPosixGraph", listsv);
11915 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11916 PL_PosixGraph, PL_L1PosixGraph, "XPosixGraph", listsv,
11917 runtime_posix_matches_above_Unicode);
11919 case ANYOF_HORIZWS:
11920 /* For these, we use the cp_list, as /d doesn't make a
11921 * difference in what these match. There would be problems
11922 * if these characters had folds other than themselves, as
11923 * cp_list is subject to folding. It turns out that \h
11924 * is just a synonym for XPosixBlank */
11925 _invlist_union(cp_list, PL_XPosixBlank, &cp_list);
11927 case ANYOF_NHORIZWS:
11928 _invlist_union_complement_2nd(cp_list,
11929 PL_XPosixBlank, &cp_list);
11933 { /* These require special handling, as they differ under
11934 folding, matching Cased there (which in the ASCII range
11935 is the same as Alpha */
11941 if (FOLD && ! LOC) {
11942 ascii_source = PL_PosixAlpha;
11943 l1_source = PL_L1Cased;
11947 ascii_source = PL_PosixLower;
11948 l1_source = PL_L1PosixLower;
11949 Xname = "XPosixLower";
11951 if (namedclass == ANYOF_LOWER) {
11952 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11953 ascii_source, l1_source, Xname, listsv);
11956 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass,
11957 posixes, ascii_source, l1_source, Xname, listsv,
11958 runtime_posix_matches_above_Unicode);
11963 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11964 PL_PosixPrint, PL_L1PosixPrint, "XPosixPrint", listsv);
11967 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11968 PL_PosixPrint, PL_L1PosixPrint, "XPosixPrint", listsv,
11969 runtime_posix_matches_above_Unicode);
11972 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11973 PL_PosixPunct, PL_L1PosixPunct, "XPosixPunct", listsv);
11976 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
11977 PL_PosixPunct, PL_L1PosixPunct, "XPosixPunct", listsv,
11978 runtime_posix_matches_above_Unicode);
11981 DO_POSIX(ret, namedclass, posixes,
11982 PL_PosixSpace, PL_XPosixSpace);
11984 case ANYOF_NPSXSPC:
11985 DO_N_POSIX(ret, namedclass, posixes,
11986 PL_PosixSpace, PL_XPosixSpace);
11989 DO_POSIX(ret, namedclass, posixes,
11990 PL_PerlSpace, PL_XPerlSpace);
11993 DO_N_POSIX(ret, namedclass, posixes,
11994 PL_PerlSpace, PL_XPerlSpace);
11996 case ANYOF_UPPER: /* Same as LOWER, above */
12003 if (FOLD && ! LOC) {
12004 ascii_source = PL_PosixAlpha;
12005 l1_source = PL_L1Cased;
12009 ascii_source = PL_PosixUpper;
12010 l1_source = PL_L1PosixUpper;
12011 Xname = "XPosixUpper";
12013 if (namedclass == ANYOF_UPPER) {
12014 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
12015 ascii_source, l1_source, Xname, listsv);
12018 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass,
12019 posixes, ascii_source, l1_source, Xname, listsv,
12020 runtime_posix_matches_above_Unicode);
12024 case ANYOF_ALNUM: /* Really is 'Word' */
12025 DO_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
12026 PL_PosixWord, PL_L1PosixWord, "XPosixWord", listsv);
12029 DO_N_POSIX_LATIN1_ONLY_KNOWN(ret, namedclass, posixes,
12030 PL_PosixWord, PL_L1PosixWord, "XPosixWord", listsv,
12031 runtime_posix_matches_above_Unicode);
12034 /* For these, we use the cp_list, as /d doesn't make a
12035 * difference in what these match. There would be problems
12036 * if these characters had folds other than themselves, as
12037 * cp_list is subject to folding */
12038 _invlist_union(cp_list, PL_VertSpace, &cp_list);
12040 case ANYOF_NVERTWS:
12041 _invlist_union_complement_2nd(cp_list,
12042 PL_VertSpace, &cp_list);
12045 DO_POSIX(ret, namedclass, posixes,
12046 PL_PosixXDigit, PL_XPosixXDigit);
12048 case ANYOF_NXDIGIT:
12049 DO_N_POSIX(ret, namedclass, posixes,
12050 PL_PosixXDigit, PL_XPosixXDigit);
12053 /* this is to handle \p and \P */
12056 vFAIL("Invalid [::] class");
12060 continue; /* Go get next character */
12062 } /* end of namedclass \blah */
12065 if (prevvalue > value) /* b-a */ {
12066 const int w = RExC_parse - rangebegin;
12067 Simple_vFAIL4("Invalid [] range \"%*.*s\"", w, w, rangebegin);
12068 range = 0; /* not a valid range */
12072 prevvalue = value; /* save the beginning of the potential range */
12073 if (RExC_parse+1 < RExC_end
12074 && *RExC_parse == '-'
12075 && RExC_parse[1] != ']')
12079 /* a bad range like \w-, [:word:]- ? */
12080 if (namedclass > OOB_NAMEDCLASS) {
12081 if (ckWARN(WARN_REGEXP)) {
12083 RExC_parse >= rangebegin ?
12084 RExC_parse - rangebegin : 0;
12086 "False [] range \"%*.*s\"",
12090 cp_list = add_cp_to_invlist(cp_list, '-');
12094 range = 1; /* yeah, it's a range! */
12095 continue; /* but do it the next time */
12099 /* Here, <prevvalue> is the beginning of the range, if any; or <value>
12102 /* non-Latin1 code point implies unicode semantics. Must be set in
12103 * pass1 so is there for the whole of pass 2 */
12105 RExC_uni_semantics = 1;
12108 /* Ready to process either the single value, or the completed range */
12111 cp_list = _add_range_to_invlist(cp_list, prevvalue, value);
12113 UV* this_range = _new_invlist(1);
12114 _append_range_to_invlist(this_range, prevvalue, value);
12116 /* In EBCDIC, the ranges 'A-Z' and 'a-z' are each not contiguous.
12117 * If this range was specified using something like 'i-j', we want
12118 * to include only the 'i' and the 'j', and not anything in
12119 * between, so exclude non-ASCII, non-alphabetics from it.
12120 * However, if the range was specified with something like
12121 * [\x89-\x91] or [\x89-j], all code points within it should be
12122 * included. literal_endpoint==2 means both ends of the range used
12123 * a literal character, not \x{foo} */
12124 if (literal_endpoint == 2
12125 && (prevvalue >= 'a' && value <= 'z')
12126 || (prevvalue >= 'A' && value <= 'Z'))
12128 _invlist_intersection(this_range, PL_ASCII, &this_range, );
12129 _invlist_intersection(this_range, PL_Alpha, &this_range, );
12131 _invlist_union(cp_list, this_range, &cp_list);
12132 literal_endpoint = 0;
12136 range = 0; /* this range (if it was one) is done now */
12137 } /* End of loop through all the text within the brackets */
12139 /* If the character class contains only a single element, it may be
12140 * optimizable into another node type which is smaller and runs faster.
12141 * Check if this is the case for this class */
12142 if (element_count == 1) {
12146 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like \w or
12147 [:digit:] or \p{foo} */
12149 /* Certain named classes have equivalents that can appear outside a
12150 * character class, e.g. \w, \H. We use these instead of a
12151 * character class. */
12152 switch ((I32)namedclass) {
12155 /* The first group is for node types that depend on the charset
12156 * modifier to the regex. We first calculate the base node
12157 * type, and if it should be inverted */
12164 goto join_charset_classes;
12171 goto join_charset_classes;
12179 join_charset_classes:
12181 /* Now that we have the base node type, we take advantage
12182 * of the enum ordering of the charset modifiers to get the
12183 * exact node type, For example the base SPACE also has
12184 * SPACEL, SPACEU, and SPACEA */
12186 offset = get_regex_charset(RExC_flags);
12188 /* /aa is the same as /a for these */
12189 if (offset == REGEX_ASCII_MORE_RESTRICTED_CHARSET) {
12190 offset = REGEX_ASCII_RESTRICTED_CHARSET;
12192 else if (op == DIGIT && offset == REGEX_UNICODE_CHARSET) {
12193 offset = REGEX_DEPENDS_CHARSET; /* There is no DIGITU */
12198 /* The number of varieties of each of these is the same,
12199 * hence, so is the delta between the normal and
12200 * complemented nodes */
12202 op += NALNUM - ALNUM;
12204 *flagp |= HASWIDTH|SIMPLE;
12207 /* The second group doesn't depend of the charset modifiers.
12208 * We just have normal and complemented */
12209 case ANYOF_NHORIZWS:
12212 case ANYOF_HORIZWS:
12214 op = (invert) ? NHORIZWS : HORIZWS;
12215 *flagp |= HASWIDTH|SIMPLE;
12218 case ANYOF_NVERTWS:
12222 op = (invert) ? NVERTWS : VERTWS;
12223 *flagp |= HASWIDTH|SIMPLE;
12233 if (AT_LEAST_UNI_SEMANTICS && ! AT_LEAST_ASCII_RESTRICTED) {
12238 /* A generic posix class. All the /a ones can be handled
12239 * by the POSIXA opcode. And all are closed under folding
12240 * in the ASCII range, so FOLD doesn't matter */
12241 if (AT_LEAST_ASCII_RESTRICTED
12242 || (! LOC && namedclass == ANYOF_ASCII))
12244 /* The odd numbered ones are the complements of the
12245 * next-lower even number one */
12246 if (namedclass % 2 == 1) {
12250 arg = namedclass_to_classnum(namedclass);
12251 op = (invert) ? NPOSIXA : POSIXA;
12256 else if (value == prevvalue) {
12258 /* Here, the class consists of just a single code point */
12261 if (! LOC && value == '\n') {
12262 op = REG_ANY; /* Optimize [^\n] */
12263 *flagp |= HASWIDTH|SIMPLE;
12267 else if (value < 256 || UTF) {
12269 /* Optimize a single value into an EXACTish node, but not if it
12270 * would require converting the pattern to UTF-8. */
12271 op = compute_EXACTish(pRExC_state);
12273 } /* Otherwise is a range */
12274 else if (! LOC) { /* locale could vary these */
12275 if (prevvalue == '0') {
12276 if (value == '9') {
12277 op = (invert) ? NDIGITA : DIGITA;
12278 *flagp |= HASWIDTH|SIMPLE;
12283 /* Here, we have changed <op> away from its initial value iff we found
12284 * an optimization */
12287 /* Throw away this ANYOF regnode, and emit the calculated one,
12288 * which should correspond to the beginning, not current, state of
12290 const char * cur_parse = RExC_parse;
12291 RExC_parse = (char *)orig_parse;
12295 /* To get locale nodes to not use the full ANYOF size would
12296 * require moving the code above that writes the portions
12297 * of it that aren't in other nodes to after this point.
12298 * e.g. ANYOF_CLASS_SET */
12299 RExC_size = orig_size;
12303 RExC_emit = (regnode *)orig_emit;
12306 ret = reg_node(pRExC_state, op);
12308 if (PL_regkind[op] == POSIXD) {
12312 *flagp |= HASWIDTH|SIMPLE;
12314 else if (PL_regkind[op] == EXACT) {
12315 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
12318 RExC_parse = (char *) cur_parse;
12320 SvREFCNT_dec(listsv);
12327 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
12329 /* If folding, we calculate all characters that could fold to or from the
12330 * ones already on the list */
12331 if (FOLD && cp_list) {
12332 UV start, end; /* End points of code point ranges */
12334 SV* fold_intersection = NULL;
12336 /* In the Latin1 range, the characters that can be folded-to or -from
12337 * are precisely the alphabetic characters. If the highest code point
12338 * is within Latin1, we can use the compiled-in list, and not have to
12339 * go out to disk. */
12340 if (invlist_highest(cp_list) < 256) {
12341 _invlist_intersection(PL_L1PosixAlpha, cp_list, &fold_intersection);
12345 /* Here, there are non-Latin1 code points, so we will have to go
12346 * fetch the list of all the characters that participate in folds
12348 if (! PL_utf8_foldable) {
12349 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
12350 &PL_sv_undef, 1, 0);
12351 PL_utf8_foldable = _get_swash_invlist(swash);
12352 SvREFCNT_dec(swash);
12355 /* This is a hash that for a particular fold gives all characters
12356 * that are involved in it */
12357 if (! PL_utf8_foldclosures) {
12359 /* If we were unable to find any folds, then we likely won't be
12360 * able to find the closures. So just create an empty list.
12361 * Folding will effectively be restricted to the non-Unicode
12362 * rules hard-coded into Perl. (This case happens legitimately
12363 * during compilation of Perl itself before the Unicode tables
12364 * are generated) */
12365 if (_invlist_len(PL_utf8_foldable) == 0) {
12366 PL_utf8_foldclosures = newHV();
12369 /* If the folds haven't been read in, call a fold function
12371 if (! PL_utf8_tofold) {
12372 U8 dummy[UTF8_MAXBYTES+1];
12375 /* This string is just a short named one above \xff */
12376 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, &dummy_len);
12377 assert(PL_utf8_tofold); /* Verify that worked */
12379 PL_utf8_foldclosures =
12380 _swash_inversion_hash(PL_utf8_tofold);
12384 /* Only the characters in this class that participate in folds need
12385 * be checked. Get the intersection of this class and all the
12386 * possible characters that are foldable. This can quickly narrow
12387 * down a large class */
12388 _invlist_intersection(PL_utf8_foldable, cp_list,
12389 &fold_intersection);
12392 /* Now look at the foldable characters in this class individually */
12393 invlist_iterinit(fold_intersection);
12394 while (invlist_iternext(fold_intersection, &start, &end)) {
12397 /* Locale folding for Latin1 characters is deferred until runtime */
12398 if (LOC && start < 256) {
12402 /* Look at every character in the range */
12403 for (j = start; j <= end; j++) {
12405 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
12411 /* We have the latin1 folding rules hard-coded here so that
12412 * an innocent-looking character class, like /[ks]/i won't
12413 * have to go out to disk to find the possible matches.
12414 * XXX It would be better to generate these via regen, in
12415 * case a new version of the Unicode standard adds new
12416 * mappings, though that is not really likely, and may be
12417 * caught by the default: case of the switch below. */
12419 if (PL_fold_latin1[j] != j) {
12421 /* ASCII is always matched; non-ASCII is matched only
12422 * under Unicode rules */
12423 if (isASCII(j) || AT_LEAST_UNI_SEMANTICS) {
12425 add_cp_to_invlist(cp_list, PL_fold_latin1[j]);
12429 add_cp_to_invlist(depends_list, PL_fold_latin1[j]);
12433 if (HAS_NONLATIN1_FOLD_CLOSURE(j)
12434 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
12436 /* Certain Latin1 characters have matches outside
12437 * Latin1, or are multi-character. To get here, 'j' is
12438 * one of those characters. None of these matches is
12439 * valid for ASCII characters under /aa, which is why
12440 * the 'if' just above excludes those. The matches
12441 * fall into three categories:
12442 * 1) They are singly folded-to or -from an above 255
12443 * character, e.g., LATIN SMALL LETTER Y WITH
12444 * DIAERESIS and LATIN CAPITAL LETTER Y WITH
12446 * 2) They are part of a multi-char fold with another
12447 * latin1 character; only LATIN SMALL LETTER
12448 * SHARP S => "ss" fits this;
12449 * 3) They are part of a multi-char fold with a
12450 * character outside of Latin1, such as various
12452 * We aren't dealing fully with multi-char folds, except
12453 * we do deal with the pattern containing a character
12454 * that has a multi-char fold (not so much the inverse).
12455 * For types 1) and 3), the matches only happen when the
12456 * target string is utf8; that's not true for 2), and we
12457 * set a flag for it.
12459 * The code below adds the single fold closures for 'j'
12460 * to the inversion list. */
12465 add_cp_to_invlist(cp_list, KELVIN_SIGN);
12469 cp_list = add_cp_to_invlist(cp_list,
12470 LATIN_SMALL_LETTER_LONG_S);
12473 cp_list = add_cp_to_invlist(cp_list,
12474 GREEK_CAPITAL_LETTER_MU);
12475 cp_list = add_cp_to_invlist(cp_list,
12476 GREEK_SMALL_LETTER_MU);
12478 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
12479 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
12481 add_cp_to_invlist(cp_list, ANGSTROM_SIGN);
12483 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
12484 cp_list = add_cp_to_invlist(cp_list,
12485 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
12487 case LATIN_SMALL_LETTER_SHARP_S:
12488 cp_list = add_cp_to_invlist(cp_list,
12489 LATIN_CAPITAL_LETTER_SHARP_S);
12491 /* Under /a, /d, and /u, this can match the two
12493 if (! ASCII_FOLD_RESTRICTED) {
12494 add_alternate(&unicode_alternate,
12497 /* And under /u or /a, it can match even if
12498 * the target is not utf8 */
12499 if (AT_LEAST_UNI_SEMANTICS) {
12500 ANYOF_FLAGS(ret) |=
12501 ANYOF_NONBITMAP_NON_UTF8;
12505 case 'F': case 'f':
12506 case 'I': case 'i':
12507 case 'L': case 'l':
12508 case 'T': case 't':
12509 case 'A': case 'a':
12510 case 'H': case 'h':
12511 case 'J': case 'j':
12512 case 'N': case 'n':
12513 case 'W': case 'w':
12514 case 'Y': case 'y':
12515 /* These all are targets of multi-character
12516 * folds from code points that require UTF8 to
12517 * express, so they can't match unless the
12518 * target string is in UTF-8, so no action here
12519 * is necessary, as regexec.c properly handles
12520 * the general case for UTF-8 matching */
12523 /* Use deprecated warning to increase the
12524 * chances of this being output */
12525 ckWARN2regdep(RExC_parse, "Perl folding rules are not up-to-date for 0x%"UVXf"; please use the perlbug utility to report;", j);
12532 /* Here is an above Latin1 character. We don't have the rules
12533 * hard-coded for it. First, get its fold */
12534 f = _to_uni_fold_flags(j, foldbuf, &foldlen,
12535 ((allow_full_fold) ? FOLD_FLAGS_FULL : 0)
12537 ? FOLD_FLAGS_LOCALE
12538 : (ASCII_FOLD_RESTRICTED)
12539 ? FOLD_FLAGS_NOMIX_ASCII
12542 if (foldlen > (STRLEN)UNISKIP(f)) {
12544 /* Any multicharacter foldings (disallowed in lookbehind
12545 * patterns) require the following transform: [ABCDEF] ->
12546 * (?:[ABCabcDEFd]|pq|rst) where E folds into "pq" and F
12547 * folds into "rst", all other characters fold to single
12548 * characters. We save away these multicharacter foldings,
12549 * to be later saved as part of the additional "s" data. */
12550 if (! RExC_in_lookbehind) {
12552 U8* e = foldbuf + foldlen;
12554 /* If any of the folded characters of this are in the
12555 * Latin1 range, tell the regex engine that this can
12556 * match a non-utf8 target string. */
12558 if (UTF8_IS_INVARIANT(*loc)
12559 || UTF8_IS_DOWNGRADEABLE_START(*loc))
12562 |= ANYOF_NONBITMAP_NON_UTF8;
12565 loc += UTF8SKIP(loc);
12568 add_alternate(&unicode_alternate, foldbuf, foldlen);
12572 /* Single character fold of above Latin1. Add everything
12573 * in its fold closure to the list that this node should
12577 /* The fold closures data structure is a hash with the keys
12578 * being every character that is folded to, like 'k', and
12579 * the values each an array of everything that folds to its
12580 * key. e.g. [ 'k', 'K', KELVIN_SIGN ] */
12581 if ((listp = hv_fetch(PL_utf8_foldclosures,
12582 (char *) foldbuf, foldlen, FALSE)))
12584 AV* list = (AV*) *listp;
12586 for (k = 0; k <= av_len(list); k++) {
12587 SV** c_p = av_fetch(list, k, FALSE);
12590 Perl_croak(aTHX_ "panic: invalid PL_utf8_foldclosures structure");
12594 /* /aa doesn't allow folds between ASCII and non-;
12595 * /l doesn't allow them between above and below
12597 if ((ASCII_FOLD_RESTRICTED
12598 && (isASCII(c) != isASCII(j)))
12599 || (LOC && ((c < 256) != (j < 256))))
12604 /* Folds involving non-ascii Latin1 characters
12605 * under /d are added to a separate list */
12606 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
12608 cp_list = add_cp_to_invlist(cp_list, c);
12611 depends_list = add_cp_to_invlist(depends_list, c);
12618 SvREFCNT_dec(fold_intersection);
12621 /* And combine the result (if any) with any inversion list from posix
12622 * classes. The lists are kept separate up to now because we don't want to
12623 * fold the classes (folding of those is automatically handled by the swash
12624 * fetching code) */
12626 if (! DEPENDS_SEMANTICS) {
12628 _invlist_union(cp_list, posixes, &cp_list);
12629 SvREFCNT_dec(posixes);
12636 /* Under /d, we put into a separate list the Latin1 things that
12637 * match only when the target string is utf8 */
12638 SV* nonascii_but_latin1_properties = NULL;
12639 _invlist_intersection(posixes, PL_Latin1,
12640 &nonascii_but_latin1_properties);
12641 _invlist_subtract(nonascii_but_latin1_properties, PL_ASCII,
12642 &nonascii_but_latin1_properties);
12643 _invlist_subtract(posixes, nonascii_but_latin1_properties,
12646 _invlist_union(cp_list, posixes, &cp_list);
12647 SvREFCNT_dec(posixes);
12653 if (depends_list) {
12654 _invlist_union(depends_list, nonascii_but_latin1_properties,
12656 SvREFCNT_dec(nonascii_but_latin1_properties);
12659 depends_list = nonascii_but_latin1_properties;
12664 /* And combine the result (if any) with any inversion list from properties.
12665 * The lists are kept separate up to now so that we can distinguish the two
12666 * in regards to matching above-Unicode. A run-time warning is generated
12667 * if a Unicode property is matched against a non-Unicode code point. But,
12668 * we allow user-defined properties to match anything, without any warning,
12669 * and we also suppress the warning if there is a portion of the character
12670 * class that isn't a Unicode property, and which matches above Unicode, \W
12671 * or [\x{110000}] for example.
12672 * (Note that in this case, unlike the Posix one above, there is no
12673 * <depends_list>, because having a Unicode property forces Unicode
12676 bool warn_super = ! has_user_defined_property;
12679 /* If it matters to the final outcome, see if a non-property
12680 * component of the class matches above Unicode. If so, the
12681 * warning gets suppressed. This is true even if just a single
12682 * such code point is specified, as though not strictly correct if
12683 * another such code point is matched against, the fact that they
12684 * are using above-Unicode code points indicates they should know
12685 * the issues involved */
12687 bool non_prop_matches_above_Unicode =
12688 runtime_posix_matches_above_Unicode
12689 | (invlist_highest(cp_list) > PERL_UNICODE_MAX);
12691 non_prop_matches_above_Unicode =
12692 ! non_prop_matches_above_Unicode;
12694 warn_super = ! non_prop_matches_above_Unicode;
12697 _invlist_union(properties, cp_list, &cp_list);
12698 SvREFCNT_dec(properties);
12701 cp_list = properties;
12705 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
12709 /* Here, we have calculated what code points should be in the character
12712 * Now we can see about various optimizations. Fold calculation (which we
12713 * did above) needs to take place before inversion. Otherwise /[^k]/i
12714 * would invert to include K, which under /i would match k, which it
12715 * shouldn't. Therefore we can't invert folded locale now, as it won't be
12716 * folded until runtime */
12718 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
12719 * at compile time. Besides not inverting folded locale now, we can't invert
12720 * if there are things such as \w, which aren't known until runtime */
12722 && ! (LOC && (FOLD || (ANYOF_FLAGS(ret) & ANYOF_CLASS)))
12724 && ! unicode_alternate
12725 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12727 _invlist_invert(cp_list);
12729 /* Any swash can't be used as-is, because we've inverted things */
12731 SvREFCNT_dec(swash);
12735 /* Clear the invert flag since have just done it here */
12739 /* If we didn't do folding, it's because some information isn't available
12740 * until runtime; set the run-time fold flag for these. (We don't have to
12741 * worry about properties folding, as that is taken care of by the swash
12743 if (FOLD && (LOC || unicode_alternate))
12745 ANYOF_FLAGS(ret) |= ANYOF_LOC_NONBITMAP_FOLD;
12748 /* Some character classes are equivalent to other nodes. Such nodes take
12749 * up less room and generally fewer operations to execute than ANYOF nodes.
12750 * Above, we checked for and optimized into some such equivalents for
12751 * certain common classes that are easy to test. Getting to this point in
12752 * the code means that the class didn't get optimized there. Since this
12753 * code is only executed in Pass 2, it is too late to save space--it has
12754 * been allocated in Pass 1, and currently isn't given back. But turning
12755 * things into an EXACTish node can allow the optimizer to join it to any
12756 * adjacent such nodes. And if the class is equivalent to things like /./,
12757 * expensive run-time swashes can be avoided. Now that we have more
12758 * complete information, we can find things necessarily missed by the
12759 * earlier code. I (khw) am not sure how much to look for here. It would
12760 * be easy, but perhaps too slow, to check any candidates against all the
12761 * node types they could possibly match using _invlistEQ(). */
12764 && ! unicode_alternate
12767 && ! (ANYOF_FLAGS(ret) & ANYOF_CLASS)
12768 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12771 U8 op = END; /* The optimzation node-type */
12772 const char * cur_parse= RExC_parse;
12774 invlist_iterinit(cp_list);
12775 if (! invlist_iternext(cp_list, &start, &end)) {
12777 /* Here, the list is empty. This happens, for example, when a
12778 * Unicode property is the only thing in the character class, and
12779 * it doesn't match anything. (perluniprops.pod notes such
12782 *flagp |= HASWIDTH|SIMPLE;
12784 else if (start == end) { /* The range is a single code point */
12785 if (! invlist_iternext(cp_list, &start, &end)
12787 /* Don't do this optimization if it would require changing
12788 * the pattern to UTF-8 */
12789 && (start < 256 || UTF))
12791 /* Here, the list contains a single code point. Can optimize
12792 * into an EXACT node */
12801 /* A locale node under folding with one code point can be
12802 * an EXACTFL, as its fold won't be calculated until
12808 /* Here, we are generally folding, but there is only one
12809 * code point to match. If we have to, we use an EXACT
12810 * node, but it would be better for joining with adjacent
12811 * nodes in the optimization pass if we used the same
12812 * EXACTFish node that any such are likely to be. We can
12813 * do this iff the code point doesn't participate in any
12814 * folds. For example, an EXACTF of a colon is the same as
12815 * an EXACT one, since nothing folds to or from a colon.
12816 * In the Latin1 range, being an alpha means that the
12817 * character participates in a fold (except for the
12818 * feminine and masculine ordinals, which I (khw) don't
12819 * think are worrying about optimizing for). */
12821 if (isALPHA_L1(value)) {
12826 if (! PL_utf8_foldable) {
12827 SV* swash = swash_init("utf8", "_Perl_Any_Folds",
12828 &PL_sv_undef, 1, 0);
12829 PL_utf8_foldable = _get_swash_invlist(swash);
12830 SvREFCNT_dec(swash);
12832 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
12837 /* If we haven't found the node type, above, it means we
12838 * can use the prevailing one */
12840 op = compute_EXACTish(pRExC_state);
12845 else if (start == 0) {
12846 if (end == UV_MAX) {
12848 *flagp |= HASWIDTH|SIMPLE;
12851 else if (end == '\n' - 1
12852 && invlist_iternext(cp_list, &start, &end)
12853 && start == '\n' + 1 && end == UV_MAX)
12856 *flagp |= HASWIDTH|SIMPLE;
12862 RExC_parse = (char *)orig_parse;
12863 RExC_emit = (regnode *)orig_emit;
12865 ret = reg_node(pRExC_state, op);
12867 RExC_parse = (char *)cur_parse;
12869 if (PL_regkind[op] == EXACT) {
12870 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value);
12873 SvREFCNT_dec(listsv);
12878 /* Here, <cp_list> contains all the code points we can determine at
12879 * compile time that match under all conditions. Go through it, and
12880 * for things that belong in the bitmap, put them there, and delete from
12881 * <cp_list>. While we are at it, see if everything above 255 is in the
12882 * list, and if so, set a flag to speed up execution */
12883 ANYOF_BITMAP_ZERO(ret);
12886 /* This gets set if we actually need to modify things */
12887 bool change_invlist = FALSE;
12891 /* Start looking through <cp_list> */
12892 invlist_iterinit(cp_list);
12893 while (invlist_iternext(cp_list, &start, &end)) {
12897 if (end == UV_MAX && start <= 256) {
12898 ANYOF_FLAGS(ret) |= ANYOF_UNICODE_ALL;
12901 /* Quit if are above what we should change */
12906 change_invlist = TRUE;
12908 /* Set all the bits in the range, up to the max that we are doing */
12909 high = (end < 255) ? end : 255;
12910 for (i = start; i <= (int) high; i++) {
12911 if (! ANYOF_BITMAP_TEST(ret, i)) {
12912 ANYOF_BITMAP_SET(ret, i);
12919 /* Done with loop; remove any code points that are in the bitmap from
12921 if (change_invlist) {
12922 _invlist_subtract(cp_list, PL_Latin1, &cp_list);
12925 /* If have completely emptied it, remove it completely */
12926 if (_invlist_len(cp_list) == 0) {
12927 SvREFCNT_dec(cp_list);
12933 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
12936 /* Here, the bitmap has been populated with all the Latin1 code points that
12937 * always match. Can now add to the overall list those that match only
12938 * when the target string is UTF-8 (<depends_list>). */
12939 if (depends_list) {
12941 _invlist_union(cp_list, depends_list, &cp_list);
12942 SvREFCNT_dec(depends_list);
12945 cp_list = depends_list;
12949 /* If there is a swash and more than one element, we can't use the swash in
12950 * the optimization below. */
12951 if (swash && element_count > 1) {
12952 SvREFCNT_dec(swash);
12957 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
12958 && ! unicode_alternate)
12960 ARG_SET(ret, ANYOF_NONBITMAP_EMPTY);
12961 SvREFCNT_dec(listsv);
12962 SvREFCNT_dec(unicode_alternate);
12965 /* av[0] stores the character class description in its textual form:
12966 * used later (regexec.c:Perl_regclass_swash()) to initialize the
12967 * appropriate swash, and is also useful for dumping the regnode.
12968 * av[1] if NULL, is a placeholder to later contain the swash computed
12969 * from av[0]. But if no further computation need be done, the
12970 * swash is stored there now.
12971 * av[2] stores the multicharacter foldings, used later in
12972 * regexec.c:S_reginclass().
12973 * av[3] stores the cp_list inversion list for use in addition or
12974 * instead of av[0]; used only if av[1] is NULL
12975 * av[4] is set if any component of the class is from a user-defined
12976 * property; used only if av[1] is NULL */
12977 AV * const av = newAV();
12980 av_store(av, 0, (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
12984 av_store(av, 1, swash);
12985 SvREFCNT_dec(cp_list);
12988 av_store(av, 1, NULL);
12990 av_store(av, 3, cp_list);
12991 av_store(av, 4, newSVuv(has_user_defined_property));
12995 /* Store any computed multi-char folds only if we are allowing
12997 if (allow_full_fold) {
12998 av_store(av, 2, MUTABLE_SV(unicode_alternate));
12999 if (unicode_alternate) { /* This node is variable length */
13004 av_store(av, 2, NULL);
13006 rv = newRV_noinc(MUTABLE_SV(av));
13007 n = add_data(pRExC_state, 1, "s");
13008 RExC_rxi->data->data[n] = (void*)rv;
13012 *flagp |= HASWIDTH|SIMPLE;
13015 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
13018 /* reg_skipcomment()
13020 Absorbs an /x style # comments from the input stream.
13021 Returns true if there is more text remaining in the stream.
13022 Will set the REG_SEEN_RUN_ON_COMMENT flag if the comment
13023 terminates the pattern without including a newline.
13025 Note its the callers responsibility to ensure that we are
13026 actually in /x mode
13031 S_reg_skipcomment(pTHX_ RExC_state_t *pRExC_state)
13035 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
13037 while (RExC_parse < RExC_end)
13038 if (*RExC_parse++ == '\n') {
13043 /* we ran off the end of the pattern without ending
13044 the comment, so we have to add an \n when wrapping */
13045 RExC_seen |= REG_SEEN_RUN_ON_COMMENT;
13053 Advances the parse position, and optionally absorbs
13054 "whitespace" from the inputstream.
13056 Without /x "whitespace" means (?#...) style comments only,
13057 with /x this means (?#...) and # comments and whitespace proper.
13059 Returns the RExC_parse point from BEFORE the scan occurs.
13061 This is the /x friendly way of saying RExC_parse++.
13065 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
13067 char* const retval = RExC_parse++;
13069 PERL_ARGS_ASSERT_NEXTCHAR;
13072 if (RExC_end - RExC_parse >= 3
13073 && *RExC_parse == '('
13074 && RExC_parse[1] == '?'
13075 && RExC_parse[2] == '#')
13077 while (*RExC_parse != ')') {
13078 if (RExC_parse == RExC_end)
13079 FAIL("Sequence (?#... not terminated");
13085 if (RExC_flags & RXf_PMf_EXTENDED) {
13086 if (isSPACE(*RExC_parse)) {
13090 else if (*RExC_parse == '#') {
13091 if ( reg_skipcomment( pRExC_state ) )
13100 - reg_node - emit a node
13102 STATIC regnode * /* Location. */
13103 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
13107 regnode * const ret = RExC_emit;
13108 GET_RE_DEBUG_FLAGS_DECL;
13110 PERL_ARGS_ASSERT_REG_NODE;
13113 SIZE_ALIGN(RExC_size);
13117 if (RExC_emit >= RExC_emit_bound)
13118 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13119 op, RExC_emit, RExC_emit_bound);
13121 NODE_ALIGN_FILL(ret);
13123 FILL_ADVANCE_NODE(ptr, op);
13124 #ifdef RE_TRACK_PATTERN_OFFSETS
13125 if (RExC_offsets) { /* MJD */
13126 MJD_OFFSET_DEBUG(("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
13127 "reg_node", __LINE__,
13129 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
13130 ? "Overwriting end of array!\n" : "OK",
13131 (UV)(RExC_emit - RExC_emit_start),
13132 (UV)(RExC_parse - RExC_start),
13133 (UV)RExC_offsets[0]));
13134 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
13142 - reganode - emit a node with an argument
13144 STATIC regnode * /* Location. */
13145 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
13149 regnode * const ret = RExC_emit;
13150 GET_RE_DEBUG_FLAGS_DECL;
13152 PERL_ARGS_ASSERT_REGANODE;
13155 SIZE_ALIGN(RExC_size);
13160 assert(2==regarglen[op]+1);
13162 Anything larger than this has to allocate the extra amount.
13163 If we changed this to be:
13165 RExC_size += (1 + regarglen[op]);
13167 then it wouldn't matter. Its not clear what side effect
13168 might come from that so its not done so far.
13173 if (RExC_emit >= RExC_emit_bound)
13174 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
13175 op, RExC_emit, RExC_emit_bound);
13177 NODE_ALIGN_FILL(ret);
13179 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
13180 #ifdef RE_TRACK_PATTERN_OFFSETS
13181 if (RExC_offsets) { /* MJD */
13182 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
13186 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0] ?
13187 "Overwriting end of array!\n" : "OK",
13188 (UV)(RExC_emit - RExC_emit_start),
13189 (UV)(RExC_parse - RExC_start),
13190 (UV)RExC_offsets[0]));
13191 Set_Cur_Node_Offset;
13199 - reguni - emit (if appropriate) a Unicode character
13202 S_reguni(pTHX_ const RExC_state_t *pRExC_state, UV uv, char* s)
13206 PERL_ARGS_ASSERT_REGUNI;
13208 return SIZE_ONLY ? UNISKIP(uv) : (uvchr_to_utf8((U8*)s, uv) - (U8*)s);
13212 - reginsert - insert an operator in front of already-emitted operand
13214 * Means relocating the operand.
13217 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
13223 const int offset = regarglen[(U8)op];
13224 const int size = NODE_STEP_REGNODE + offset;
13225 GET_RE_DEBUG_FLAGS_DECL;
13227 PERL_ARGS_ASSERT_REGINSERT;
13228 PERL_UNUSED_ARG(depth);
13229 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
13230 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
13239 if (RExC_open_parens) {
13241 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
13242 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
13243 if ( RExC_open_parens[paren] >= opnd ) {
13244 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
13245 RExC_open_parens[paren] += size;
13247 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
13249 if ( RExC_close_parens[paren] >= opnd ) {
13250 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
13251 RExC_close_parens[paren] += size;
13253 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
13258 while (src > opnd) {
13259 StructCopy(--src, --dst, regnode);
13260 #ifdef RE_TRACK_PATTERN_OFFSETS
13261 if (RExC_offsets) { /* MJD 20010112 */
13262 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
13266 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
13267 ? "Overwriting end of array!\n" : "OK",
13268 (UV)(src - RExC_emit_start),
13269 (UV)(dst - RExC_emit_start),
13270 (UV)RExC_offsets[0]));
13271 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
13272 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
13278 place = opnd; /* Op node, where operand used to be. */
13279 #ifdef RE_TRACK_PATTERN_OFFSETS
13280 if (RExC_offsets) { /* MJD */
13281 MJD_OFFSET_DEBUG(("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
13285 (UV)(place - RExC_emit_start) > RExC_offsets[0]
13286 ? "Overwriting end of array!\n" : "OK",
13287 (UV)(place - RExC_emit_start),
13288 (UV)(RExC_parse - RExC_start),
13289 (UV)RExC_offsets[0]));
13290 Set_Node_Offset(place, RExC_parse);
13291 Set_Node_Length(place, 1);
13294 src = NEXTOPER(place);
13295 FILL_ADVANCE_NODE(place, op);
13296 Zero(src, offset, regnode);
13300 - regtail - set the next-pointer at the end of a node chain of p to val.
13301 - SEE ALSO: regtail_study
13303 /* TODO: All three parms should be const */
13305 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
13309 GET_RE_DEBUG_FLAGS_DECL;
13311 PERL_ARGS_ASSERT_REGTAIL;
13313 PERL_UNUSED_ARG(depth);
13319 /* Find last node. */
13322 regnode * const temp = regnext(scan);
13324 SV * const mysv=sv_newmortal();
13325 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
13326 regprop(RExC_rx, mysv, scan);
13327 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
13328 SvPV_nolen_const(mysv), REG_NODE_NUM(scan),
13329 (temp == NULL ? "->" : ""),
13330 (temp == NULL ? PL_reg_name[OP(val)] : "")
13338 if (reg_off_by_arg[OP(scan)]) {
13339 ARG_SET(scan, val - scan);
13342 NEXT_OFF(scan) = val - scan;
13348 - regtail_study - set the next-pointer at the end of a node chain of p to val.
13349 - Look for optimizable sequences at the same time.
13350 - currently only looks for EXACT chains.
13352 This is experimental code. The idea is to use this routine to perform
13353 in place optimizations on branches and groups as they are constructed,
13354 with the long term intention of removing optimization from study_chunk so
13355 that it is purely analytical.
13357 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
13358 to control which is which.
13361 /* TODO: All four parms should be const */
13364 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p, const regnode *val,U32 depth)
13369 #ifdef EXPERIMENTAL_INPLACESCAN
13372 GET_RE_DEBUG_FLAGS_DECL;
13374 PERL_ARGS_ASSERT_REGTAIL_STUDY;
13380 /* Find last node. */
13384 regnode * const temp = regnext(scan);
13385 #ifdef EXPERIMENTAL_INPLACESCAN
13386 if (PL_regkind[OP(scan)] == EXACT) {
13387 bool has_exactf_sharp_s; /* Unexamined in this routine */
13388 if (join_exact(pRExC_state,scan,&min, &has_exactf_sharp_s, 1,val,depth+1))
13393 switch (OP(scan)) {
13399 case EXACTFU_TRICKYFOLD:
13401 if( exact == PSEUDO )
13403 else if ( exact != OP(scan) )
13412 SV * const mysv=sv_newmortal();
13413 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
13414 regprop(RExC_rx, mysv, scan);
13415 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
13416 SvPV_nolen_const(mysv),
13417 REG_NODE_NUM(scan),
13418 PL_reg_name[exact]);
13425 SV * const mysv_val=sv_newmortal();
13426 DEBUG_PARSE_MSG("");
13427 regprop(RExC_rx, mysv_val, val);
13428 PerlIO_printf(Perl_debug_log, "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
13429 SvPV_nolen_const(mysv_val),
13430 (IV)REG_NODE_NUM(val),
13434 if (reg_off_by_arg[OP(scan)]) {
13435 ARG_SET(scan, val - scan);
13438 NEXT_OFF(scan) = val - scan;
13446 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
13450 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
13456 for (bit=0; bit<32; bit++) {
13457 if (flags & (1<<bit)) {
13458 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
13461 if (!set++ && lead)
13462 PerlIO_printf(Perl_debug_log, "%s",lead);
13463 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
13466 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
13467 if (!set++ && lead) {
13468 PerlIO_printf(Perl_debug_log, "%s",lead);
13471 case REGEX_UNICODE_CHARSET:
13472 PerlIO_printf(Perl_debug_log, "UNICODE");
13474 case REGEX_LOCALE_CHARSET:
13475 PerlIO_printf(Perl_debug_log, "LOCALE");
13477 case REGEX_ASCII_RESTRICTED_CHARSET:
13478 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
13480 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
13481 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
13484 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
13490 PerlIO_printf(Perl_debug_log, "\n");
13492 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
13498 Perl_regdump(pTHX_ const regexp *r)
13502 SV * const sv = sv_newmortal();
13503 SV *dsv= sv_newmortal();
13504 RXi_GET_DECL(r,ri);
13505 GET_RE_DEBUG_FLAGS_DECL;
13507 PERL_ARGS_ASSERT_REGDUMP;
13509 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
13511 /* Header fields of interest. */
13512 if (r->anchored_substr) {
13513 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
13514 RE_SV_DUMPLEN(r->anchored_substr), 30);
13515 PerlIO_printf(Perl_debug_log,
13516 "anchored %s%s at %"IVdf" ",
13517 s, RE_SV_TAIL(r->anchored_substr),
13518 (IV)r->anchored_offset);
13519 } else if (r->anchored_utf8) {
13520 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
13521 RE_SV_DUMPLEN(r->anchored_utf8), 30);
13522 PerlIO_printf(Perl_debug_log,
13523 "anchored utf8 %s%s at %"IVdf" ",
13524 s, RE_SV_TAIL(r->anchored_utf8),
13525 (IV)r->anchored_offset);
13527 if (r->float_substr) {
13528 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
13529 RE_SV_DUMPLEN(r->float_substr), 30);
13530 PerlIO_printf(Perl_debug_log,
13531 "floating %s%s at %"IVdf"..%"UVuf" ",
13532 s, RE_SV_TAIL(r->float_substr),
13533 (IV)r->float_min_offset, (UV)r->float_max_offset);
13534 } else if (r->float_utf8) {
13535 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
13536 RE_SV_DUMPLEN(r->float_utf8), 30);
13537 PerlIO_printf(Perl_debug_log,
13538 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
13539 s, RE_SV_TAIL(r->float_utf8),
13540 (IV)r->float_min_offset, (UV)r->float_max_offset);
13542 if (r->check_substr || r->check_utf8)
13543 PerlIO_printf(Perl_debug_log,
13545 (r->check_substr == r->float_substr
13546 && r->check_utf8 == r->float_utf8
13547 ? "(checking floating" : "(checking anchored"));
13548 if (r->extflags & RXf_NOSCAN)
13549 PerlIO_printf(Perl_debug_log, " noscan");
13550 if (r->extflags & RXf_CHECK_ALL)
13551 PerlIO_printf(Perl_debug_log, " isall");
13552 if (r->check_substr || r->check_utf8)
13553 PerlIO_printf(Perl_debug_log, ") ");
13555 if (ri->regstclass) {
13556 regprop(r, sv, ri->regstclass);
13557 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
13559 if (r->extflags & RXf_ANCH) {
13560 PerlIO_printf(Perl_debug_log, "anchored");
13561 if (r->extflags & RXf_ANCH_BOL)
13562 PerlIO_printf(Perl_debug_log, "(BOL)");
13563 if (r->extflags & RXf_ANCH_MBOL)
13564 PerlIO_printf(Perl_debug_log, "(MBOL)");
13565 if (r->extflags & RXf_ANCH_SBOL)
13566 PerlIO_printf(Perl_debug_log, "(SBOL)");
13567 if (r->extflags & RXf_ANCH_GPOS)
13568 PerlIO_printf(Perl_debug_log, "(GPOS)");
13569 PerlIO_putc(Perl_debug_log, ' ');
13571 if (r->extflags & RXf_GPOS_SEEN)
13572 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
13573 if (r->intflags & PREGf_SKIP)
13574 PerlIO_printf(Perl_debug_log, "plus ");
13575 if (r->intflags & PREGf_IMPLICIT)
13576 PerlIO_printf(Perl_debug_log, "implicit ");
13577 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
13578 if (r->extflags & RXf_EVAL_SEEN)
13579 PerlIO_printf(Perl_debug_log, "with eval ");
13580 PerlIO_printf(Perl_debug_log, "\n");
13581 DEBUG_FLAGS_r(regdump_extflags("r->extflags: ",r->extflags));
13583 PERL_ARGS_ASSERT_REGDUMP;
13584 PERL_UNUSED_CONTEXT;
13585 PERL_UNUSED_ARG(r);
13586 #endif /* DEBUGGING */
13590 - regprop - printable representation of opcode
13592 #define EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags) \
13595 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]); \
13596 if (flags & ANYOF_INVERT) \
13597 /*make sure the invert info is in each */ \
13598 sv_catpvs(sv, "^"); \
13604 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o)
13610 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
13611 static const char * const anyofs[] = {
13643 RXi_GET_DECL(prog,progi);
13644 GET_RE_DEBUG_FLAGS_DECL;
13646 PERL_ARGS_ASSERT_REGPROP;
13650 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
13651 /* It would be nice to FAIL() here, but this may be called from
13652 regexec.c, and it would be hard to supply pRExC_state. */
13653 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(o), (int)REGNODE_MAX);
13654 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
13656 k = PL_regkind[OP(o)];
13659 sv_catpvs(sv, " ");
13660 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
13661 * is a crude hack but it may be the best for now since
13662 * we have no flag "this EXACTish node was UTF-8"
13664 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
13665 PERL_PV_ESCAPE_UNI_DETECT |
13666 PERL_PV_ESCAPE_NONASCII |
13667 PERL_PV_PRETTY_ELLIPSES |
13668 PERL_PV_PRETTY_LTGT |
13669 PERL_PV_PRETTY_NOCLEAR
13671 } else if (k == TRIE) {
13672 /* print the details of the trie in dumpuntil instead, as
13673 * progi->data isn't available here */
13674 const char op = OP(o);
13675 const U32 n = ARG(o);
13676 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
13677 (reg_ac_data *)progi->data->data[n] :
13679 const reg_trie_data * const trie
13680 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
13682 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
13683 DEBUG_TRIE_COMPILE_r(
13684 Perl_sv_catpvf(aTHX_ sv,
13685 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
13686 (UV)trie->startstate,
13687 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
13688 (UV)trie->wordcount,
13691 (UV)TRIE_CHARCOUNT(trie),
13692 (UV)trie->uniquecharcount
13695 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
13697 int rangestart = -1;
13698 U8* bitmap = IS_ANYOF_TRIE(op) ? (U8*)ANYOF_BITMAP(o) : (U8*)TRIE_BITMAP(trie);
13699 sv_catpvs(sv, "[");
13700 for (i = 0; i <= 256; i++) {
13701 if (i < 256 && BITMAP_TEST(bitmap,i)) {
13702 if (rangestart == -1)
13704 } else if (rangestart != -1) {
13705 if (i <= rangestart + 3)
13706 for (; rangestart < i; rangestart++)
13707 put_byte(sv, rangestart);
13709 put_byte(sv, rangestart);
13710 sv_catpvs(sv, "-");
13711 put_byte(sv, i - 1);
13716 sv_catpvs(sv, "]");
13719 } else if (k == CURLY) {
13720 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
13721 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
13722 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
13724 else if (k == WHILEM && o->flags) /* Ordinal/of */
13725 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
13726 else if (k == REF || k == OPEN || k == CLOSE || k == GROUPP || OP(o)==ACCEPT) {
13727 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
13728 if ( RXp_PAREN_NAMES(prog) ) {
13729 if ( k != REF || (OP(o) < NREF)) {
13730 AV *list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
13731 SV **name= av_fetch(list, ARG(o), 0 );
13733 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
13736 AV *list= MUTABLE_AV(progi->data->data[ progi->name_list_idx ]);
13737 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
13738 I32 *nums=(I32*)SvPVX(sv_dat);
13739 SV **name= av_fetch(list, nums[0], 0 );
13742 for ( n=0; n<SvIVX(sv_dat); n++ ) {
13743 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
13744 (n ? "," : ""), (IV)nums[n]);
13746 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
13750 } else if (k == GOSUB)
13751 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o)); /* Paren and offset */
13752 else if (k == VERB) {
13754 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
13755 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
13756 } else if (k == LOGICAL)
13757 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* 2: embedded, otherwise 1 */
13758 else if (k == ANYOF) {
13759 int i, rangestart = -1;
13760 const U8 flags = ANYOF_FLAGS(o);
13764 if (flags & ANYOF_LOCALE)
13765 sv_catpvs(sv, "{loc}");
13766 if (flags & ANYOF_LOC_NONBITMAP_FOLD)
13767 sv_catpvs(sv, "{i}");
13768 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
13769 if (flags & ANYOF_INVERT)
13770 sv_catpvs(sv, "^");
13772 /* output what the standard cp 0-255 bitmap matches */
13773 for (i = 0; i <= 256; i++) {
13774 if (i < 256 && ANYOF_BITMAP_TEST(o,i)) {
13775 if (rangestart == -1)
13777 } else if (rangestart != -1) {
13778 if (i <= rangestart + 3)
13779 for (; rangestart < i; rangestart++)
13780 put_byte(sv, rangestart);
13782 put_byte(sv, rangestart);
13783 sv_catpvs(sv, "-");
13784 put_byte(sv, i - 1);
13791 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
13792 /* output any special charclass tests (used entirely under use locale) */
13793 if (ANYOF_CLASS_TEST_ANY_SET(o))
13794 for (i = 0; i < (int)(sizeof(anyofs)/sizeof(char*)); i++)
13795 if (ANYOF_CLASS_TEST(o,i)) {
13796 sv_catpv(sv, anyofs[i]);
13800 EMIT_ANYOF_TEST_SEPARATOR(do_sep,sv,flags);
13802 if (flags & ANYOF_NON_UTF8_LATIN1_ALL) {
13803 sv_catpvs(sv, "{non-utf8-latin1-all}");
13806 /* output information about the unicode matching */
13807 if (flags & ANYOF_UNICODE_ALL)
13808 sv_catpvs(sv, "{unicode_all}");
13809 else if (ANYOF_NONBITMAP(o))
13810 sv_catpvs(sv, "{unicode}");
13811 if (flags & ANYOF_NONBITMAP_NON_UTF8)
13812 sv_catpvs(sv, "{outside bitmap}");
13814 if (ANYOF_NONBITMAP(o)) {
13815 SV *lv; /* Set if there is something outside the bit map */
13816 SV * const sw = regclass_swash(prog, o, FALSE, &lv, 0);
13817 bool byte_output = FALSE; /* If something in the bitmap has been
13820 if (lv && lv != &PL_sv_undef) {
13822 U8 s[UTF8_MAXBYTES_CASE+1];
13824 for (i = 0; i <= 256; i++) { /* Look at chars in bitmap */
13825 uvchr_to_utf8(s, i);
13828 && ! ANYOF_BITMAP_TEST(o, i) /* Don't duplicate
13832 && swash_fetch(sw, s, TRUE))
13834 if (rangestart == -1)
13836 } else if (rangestart != -1) {
13837 byte_output = TRUE;
13838 if (i <= rangestart + 3)
13839 for (; rangestart < i; rangestart++) {
13840 put_byte(sv, rangestart);
13843 put_byte(sv, rangestart);
13844 sv_catpvs(sv, "-");
13853 char *s = savesvpv(lv);
13854 char * const origs = s;
13856 while (*s && *s != '\n')
13860 const char * const t = ++s;
13863 sv_catpvs(sv, " ");
13869 /* Truncate very long output */
13870 if (s - origs > 256) {
13871 Perl_sv_catpvf(aTHX_ sv,
13873 (int) (s - origs - 1),
13879 else if (*s == '\t') {
13898 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
13900 else if (k == POSIXD) {
13901 U8 index = FLAGS(o) * 2;
13902 if (index > (sizeof(anyofs) / sizeof(anyofs[0]))) {
13903 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
13906 sv_catpv(sv, anyofs[index]);
13909 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
13910 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
13912 PERL_UNUSED_CONTEXT;
13913 PERL_UNUSED_ARG(sv);
13914 PERL_UNUSED_ARG(o);
13915 PERL_UNUSED_ARG(prog);
13916 #endif /* DEBUGGING */
13920 Perl_re_intuit_string(pTHX_ REGEXP * const r)
13921 { /* Assume that RE_INTUIT is set */
13923 struct regexp *const prog = (struct regexp *)SvANY(r);
13924 GET_RE_DEBUG_FLAGS_DECL;
13926 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
13927 PERL_UNUSED_CONTEXT;
13931 const char * const s = SvPV_nolen_const(prog->check_substr
13932 ? prog->check_substr : prog->check_utf8);
13934 if (!PL_colorset) reginitcolors();
13935 PerlIO_printf(Perl_debug_log,
13936 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
13938 prog->check_substr ? "" : "utf8 ",
13939 PL_colors[5],PL_colors[0],
13942 (strlen(s) > 60 ? "..." : ""));
13945 return prog->check_substr ? prog->check_substr : prog->check_utf8;
13951 handles refcounting and freeing the perl core regexp structure. When
13952 it is necessary to actually free the structure the first thing it
13953 does is call the 'free' method of the regexp_engine associated to
13954 the regexp, allowing the handling of the void *pprivate; member
13955 first. (This routine is not overridable by extensions, which is why
13956 the extensions free is called first.)
13958 See regdupe and regdupe_internal if you change anything here.
13960 #ifndef PERL_IN_XSUB_RE
13962 Perl_pregfree(pTHX_ REGEXP *r)
13968 Perl_pregfree2(pTHX_ REGEXP *rx)
13971 struct regexp *const r = (struct regexp *)SvANY(rx);
13972 GET_RE_DEBUG_FLAGS_DECL;
13974 PERL_ARGS_ASSERT_PREGFREE2;
13976 if (r->mother_re) {
13977 ReREFCNT_dec(r->mother_re);
13979 CALLREGFREE_PVT(rx); /* free the private data */
13980 SvREFCNT_dec(RXp_PAREN_NAMES(r));
13983 SvREFCNT_dec(r->anchored_substr);
13984 SvREFCNT_dec(r->anchored_utf8);
13985 SvREFCNT_dec(r->float_substr);
13986 SvREFCNT_dec(r->float_utf8);
13987 Safefree(r->substrs);
13989 RX_MATCH_COPY_FREE(rx);
13990 #ifdef PERL_OLD_COPY_ON_WRITE
13991 SvREFCNT_dec(r->saved_copy);
13994 SvREFCNT_dec(r->qr_anoncv);
13999 This is a hacky workaround to the structural issue of match results
14000 being stored in the regexp structure which is in turn stored in
14001 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
14002 could be PL_curpm in multiple contexts, and could require multiple
14003 result sets being associated with the pattern simultaneously, such
14004 as when doing a recursive match with (??{$qr})
14006 The solution is to make a lightweight copy of the regexp structure
14007 when a qr// is returned from the code executed by (??{$qr}) this
14008 lightweight copy doesn't actually own any of its data except for
14009 the starp/end and the actual regexp structure itself.
14015 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
14017 struct regexp *ret;
14018 struct regexp *const r = (struct regexp *)SvANY(rx);
14020 PERL_ARGS_ASSERT_REG_TEMP_COPY;
14023 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
14024 ret = (struct regexp *)SvANY(ret_x);
14026 (void)ReREFCNT_inc(rx);
14027 /* We can take advantage of the existing "copied buffer" mechanism in SVs
14028 by pointing directly at the buffer, but flagging that the allocated
14029 space in the copy is zero. As we've just done a struct copy, it's now
14030 a case of zero-ing that, rather than copying the current length. */
14031 SvPV_set(ret_x, RX_WRAPPED(rx));
14032 SvFLAGS(ret_x) |= SvFLAGS(rx) & (SVf_POK|SVp_POK|SVf_UTF8);
14033 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
14034 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
14035 SvLEN_set(ret_x, 0);
14036 SvSTASH_set(ret_x, NULL);
14037 SvMAGIC_set(ret_x, NULL);
14039 const I32 npar = r->nparens+1;
14040 Newx(ret->offs, npar, regexp_paren_pair);
14041 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14044 Newx(ret->substrs, 1, struct reg_substr_data);
14045 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14047 SvREFCNT_inc_void(ret->anchored_substr);
14048 SvREFCNT_inc_void(ret->anchored_utf8);
14049 SvREFCNT_inc_void(ret->float_substr);
14050 SvREFCNT_inc_void(ret->float_utf8);
14052 /* check_substr and check_utf8, if non-NULL, point to either their
14053 anchored or float namesakes, and don't hold a second reference. */
14055 RX_MATCH_COPIED_off(ret_x);
14056 #ifdef PERL_OLD_COPY_ON_WRITE
14057 ret->saved_copy = NULL;
14059 ret->mother_re = rx;
14060 SvREFCNT_inc_void(ret->qr_anoncv);
14066 /* regfree_internal()
14068 Free the private data in a regexp. This is overloadable by
14069 extensions. Perl takes care of the regexp structure in pregfree(),
14070 this covers the *pprivate pointer which technically perl doesn't
14071 know about, however of course we have to handle the
14072 regexp_internal structure when no extension is in use.
14074 Note this is called before freeing anything in the regexp
14079 Perl_regfree_internal(pTHX_ REGEXP * const rx)
14082 struct regexp *const r = (struct regexp *)SvANY(rx);
14083 RXi_GET_DECL(r,ri);
14084 GET_RE_DEBUG_FLAGS_DECL;
14086 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
14092 SV *dsv= sv_newmortal();
14093 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
14094 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
14095 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
14096 PL_colors[4],PL_colors[5],s);
14099 #ifdef RE_TRACK_PATTERN_OFFSETS
14101 Safefree(ri->u.offsets); /* 20010421 MJD */
14103 if (ri->code_blocks) {
14105 for (n = 0; n < ri->num_code_blocks; n++)
14106 SvREFCNT_dec(ri->code_blocks[n].src_regex);
14107 Safefree(ri->code_blocks);
14111 int n = ri->data->count;
14114 /* If you add a ->what type here, update the comment in regcomp.h */
14115 switch (ri->data->what[n]) {
14121 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
14124 Safefree(ri->data->data[n]);
14130 { /* Aho Corasick add-on structure for a trie node.
14131 Used in stclass optimization only */
14133 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
14135 refcount = --aho->refcount;
14138 PerlMemShared_free(aho->states);
14139 PerlMemShared_free(aho->fail);
14140 /* do this last!!!! */
14141 PerlMemShared_free(ri->data->data[n]);
14142 PerlMemShared_free(ri->regstclass);
14148 /* trie structure. */
14150 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
14152 refcount = --trie->refcount;
14155 PerlMemShared_free(trie->charmap);
14156 PerlMemShared_free(trie->states);
14157 PerlMemShared_free(trie->trans);
14159 PerlMemShared_free(trie->bitmap);
14161 PerlMemShared_free(trie->jump);
14162 PerlMemShared_free(trie->wordinfo);
14163 /* do this last!!!! */
14164 PerlMemShared_free(ri->data->data[n]);
14169 Perl_croak(aTHX_ "panic: regfree data code '%c'", ri->data->what[n]);
14172 Safefree(ri->data->what);
14173 Safefree(ri->data);
14179 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
14180 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
14181 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
14184 re_dup - duplicate a regexp.
14186 This routine is expected to clone a given regexp structure. It is only
14187 compiled under USE_ITHREADS.
14189 After all of the core data stored in struct regexp is duplicated
14190 the regexp_engine.dupe method is used to copy any private data
14191 stored in the *pprivate pointer. This allows extensions to handle
14192 any duplication it needs to do.
14194 See pregfree() and regfree_internal() if you change anything here.
14196 #if defined(USE_ITHREADS)
14197 #ifndef PERL_IN_XSUB_RE
14199 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
14203 const struct regexp *r = (const struct regexp *)SvANY(sstr);
14204 struct regexp *ret = (struct regexp *)SvANY(dstr);
14206 PERL_ARGS_ASSERT_RE_DUP_GUTS;
14208 npar = r->nparens+1;
14209 Newx(ret->offs, npar, regexp_paren_pair);
14210 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
14212 /* no need to copy these */
14213 Newx(ret->swap, npar, regexp_paren_pair);
14216 if (ret->substrs) {
14217 /* Do it this way to avoid reading from *r after the StructCopy().
14218 That way, if any of the sv_dup_inc()s dislodge *r from the L1
14219 cache, it doesn't matter. */
14220 const bool anchored = r->check_substr
14221 ? r->check_substr == r->anchored_substr
14222 : r->check_utf8 == r->anchored_utf8;
14223 Newx(ret->substrs, 1, struct reg_substr_data);
14224 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
14226 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
14227 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
14228 ret->float_substr = sv_dup_inc(ret->float_substr, param);
14229 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
14231 /* check_substr and check_utf8, if non-NULL, point to either their
14232 anchored or float namesakes, and don't hold a second reference. */
14234 if (ret->check_substr) {
14236 assert(r->check_utf8 == r->anchored_utf8);
14237 ret->check_substr = ret->anchored_substr;
14238 ret->check_utf8 = ret->anchored_utf8;
14240 assert(r->check_substr == r->float_substr);
14241 assert(r->check_utf8 == r->float_utf8);
14242 ret->check_substr = ret->float_substr;
14243 ret->check_utf8 = ret->float_utf8;
14245 } else if (ret->check_utf8) {
14247 ret->check_utf8 = ret->anchored_utf8;
14249 ret->check_utf8 = ret->float_utf8;
14254 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
14255 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
14258 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
14260 if (RX_MATCH_COPIED(dstr))
14261 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
14263 ret->subbeg = NULL;
14264 #ifdef PERL_OLD_COPY_ON_WRITE
14265 ret->saved_copy = NULL;
14268 if (ret->mother_re) {
14269 if (SvPVX_const(dstr) == SvPVX_const(ret->mother_re)) {
14270 /* Our storage points directly to our mother regexp, but that's
14271 1: a buffer in a different thread
14272 2: something we no longer hold a reference on
14273 so we need to copy it locally. */
14274 /* Note we need to use SvCUR(), rather than
14275 SvLEN(), on our mother_re, because it, in
14276 turn, may well be pointing to its own mother_re. */
14277 SvPV_set(dstr, SAVEPVN(SvPVX_const(ret->mother_re),
14278 SvCUR(ret->mother_re)+1));
14279 SvLEN_set(dstr, SvCUR(ret->mother_re)+1);
14281 ret->mother_re = NULL;
14285 #endif /* PERL_IN_XSUB_RE */
14290 This is the internal complement to regdupe() which is used to copy
14291 the structure pointed to by the *pprivate pointer in the regexp.
14292 This is the core version of the extension overridable cloning hook.
14293 The regexp structure being duplicated will be copied by perl prior
14294 to this and will be provided as the regexp *r argument, however
14295 with the /old/ structures pprivate pointer value. Thus this routine
14296 may override any copying normally done by perl.
14298 It returns a pointer to the new regexp_internal structure.
14302 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
14305 struct regexp *const r = (struct regexp *)SvANY(rx);
14306 regexp_internal *reti;
14308 RXi_GET_DECL(r,ri);
14310 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
14314 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode), char, regexp_internal);
14315 Copy(ri->program, reti->program, len+1, regnode);
14317 reti->num_code_blocks = ri->num_code_blocks;
14318 if (ri->code_blocks) {
14320 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
14321 struct reg_code_block);
14322 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
14323 struct reg_code_block);
14324 for (n = 0; n < ri->num_code_blocks; n++)
14325 reti->code_blocks[n].src_regex = (REGEXP*)
14326 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
14329 reti->code_blocks = NULL;
14331 reti->regstclass = NULL;
14334 struct reg_data *d;
14335 const int count = ri->data->count;
14338 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
14339 char, struct reg_data);
14340 Newx(d->what, count, U8);
14343 for (i = 0; i < count; i++) {
14344 d->what[i] = ri->data->what[i];
14345 switch (d->what[i]) {
14346 /* see also regcomp.h and regfree_internal() */
14347 case 'a': /* actually an AV, but the dup function is identical. */
14351 case 'u': /* actually an HV, but the dup function is identical. */
14352 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
14355 /* This is cheating. */
14356 Newx(d->data[i], 1, struct regnode_charclass_class);
14357 StructCopy(ri->data->data[i], d->data[i],
14358 struct regnode_charclass_class);
14359 reti->regstclass = (regnode*)d->data[i];
14362 /* Trie stclasses are readonly and can thus be shared
14363 * without duplication. We free the stclass in pregfree
14364 * when the corresponding reg_ac_data struct is freed.
14366 reti->regstclass= ri->regstclass;
14370 ((reg_trie_data*)ri->data->data[i])->refcount++;
14375 d->data[i] = ri->data->data[i];
14378 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'", ri->data->what[i]);
14387 reti->name_list_idx = ri->name_list_idx;
14389 #ifdef RE_TRACK_PATTERN_OFFSETS
14390 if (ri->u.offsets) {
14391 Newx(reti->u.offsets, 2*len+1, U32);
14392 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
14395 SetProgLen(reti,len);
14398 return (void*)reti;
14401 #endif /* USE_ITHREADS */
14403 #ifndef PERL_IN_XSUB_RE
14406 - regnext - dig the "next" pointer out of a node
14409 Perl_regnext(pTHX_ register regnode *p)
14417 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
14418 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d", (int)OP(p), (int)REGNODE_MAX);
14421 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
14430 S_re_croak2(pTHX_ const char* pat1,const char* pat2,...)
14433 STRLEN l1 = strlen(pat1);
14434 STRLEN l2 = strlen(pat2);
14437 const char *message;
14439 PERL_ARGS_ASSERT_RE_CROAK2;
14445 Copy(pat1, buf, l1 , char);
14446 Copy(pat2, buf + l1, l2 , char);
14447 buf[l1 + l2] = '\n';
14448 buf[l1 + l2 + 1] = '\0';
14450 /* ANSI variant takes additional second argument */
14451 va_start(args, pat2);
14455 msv = vmess(buf, &args);
14457 message = SvPV_const(msv,l1);
14460 Copy(message, buf, l1 , char);
14461 buf[l1-1] = '\0'; /* Overwrite \n */
14462 Perl_croak(aTHX_ "%s", buf);
14465 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
14467 #ifndef PERL_IN_XSUB_RE
14469 Perl_save_re_context(pTHX)
14473 struct re_save_state *state;
14475 SAVEVPTR(PL_curcop);
14476 SSGROW(SAVESTACK_ALLOC_FOR_RE_SAVE_STATE + 1);
14478 state = (struct re_save_state *)(PL_savestack + PL_savestack_ix);
14479 PL_savestack_ix += SAVESTACK_ALLOC_FOR_RE_SAVE_STATE;
14480 SSPUSHUV(SAVEt_RE_STATE);
14482 Copy(&PL_reg_state, state, 1, struct re_save_state);
14484 PL_reg_oldsaved = NULL;
14485 PL_reg_oldsavedlen = 0;
14486 PL_reg_oldsavedoffset = 0;
14487 PL_reg_oldsavedcoffset = 0;
14488 PL_reg_maxiter = 0;
14489 PL_reg_leftiter = 0;
14490 PL_reg_poscache = NULL;
14491 PL_reg_poscache_size = 0;
14492 #ifdef PERL_OLD_COPY_ON_WRITE
14496 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
14498 const REGEXP * const rx = PM_GETRE(PL_curpm);
14501 for (i = 1; i <= RX_NPARENS(rx); i++) {
14502 char digits[TYPE_CHARS(long)];
14503 const STRLEN len = my_snprintf(digits, sizeof(digits), "%lu", (long)i);
14504 GV *const *const gvp
14505 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
14508 GV * const gv = *gvp;
14509 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
14519 clear_re(pTHX_ void *r)
14522 ReREFCNT_dec((REGEXP *)r);
14528 S_put_byte(pTHX_ SV *sv, int c)
14530 PERL_ARGS_ASSERT_PUT_BYTE;
14532 /* Our definition of isPRINT() ignores locales, so only bytes that are
14533 not part of UTF-8 are considered printable. I assume that the same
14534 holds for UTF-EBCDIC.
14535 Also, code point 255 is not printable in either (it's E0 in EBCDIC,
14536 which Wikipedia says:
14538 EO, or Eight Ones, is an 8-bit EBCDIC character code represented as all
14539 ones (binary 1111 1111, hexadecimal FF). It is similar, but not
14540 identical, to the ASCII delete (DEL) or rubout control character.
14541 ) So the old condition can be simplified to !isPRINT(c) */
14544 Perl_sv_catpvf(aTHX_ sv, "\\x%02x", c);
14547 Perl_sv_catpvf(aTHX_ sv, "\\x{%x}", c);
14551 const char string = c;
14552 if (c == '-' || c == ']' || c == '\\' || c == '^')
14553 sv_catpvs(sv, "\\");
14554 sv_catpvn(sv, &string, 1);
14559 #define CLEAR_OPTSTART \
14560 if (optstart) STMT_START { \
14561 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
14565 #define DUMPUNTIL(b,e) CLEAR_OPTSTART; node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
14567 STATIC const regnode *
14568 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
14569 const regnode *last, const regnode *plast,
14570 SV* sv, I32 indent, U32 depth)
14573 U8 op = PSEUDO; /* Arbitrary non-END op. */
14574 const regnode *next;
14575 const regnode *optstart= NULL;
14577 RXi_GET_DECL(r,ri);
14578 GET_RE_DEBUG_FLAGS_DECL;
14580 PERL_ARGS_ASSERT_DUMPUNTIL;
14582 #ifdef DEBUG_DUMPUNTIL
14583 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
14584 last ? last-start : 0,plast ? plast-start : 0);
14587 if (plast && plast < last)
14590 while (PL_regkind[op] != END && (!last || node < last)) {
14591 /* While that wasn't END last time... */
14594 if (op == CLOSE || op == WHILEM)
14596 next = regnext((regnode *)node);
14599 if (OP(node) == OPTIMIZED) {
14600 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
14607 regprop(r, sv, node);
14608 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
14609 (int)(2*indent + 1), "", SvPVX_const(sv));
14611 if (OP(node) != OPTIMIZED) {
14612 if (next == NULL) /* Next ptr. */
14613 PerlIO_printf(Perl_debug_log, " (0)");
14614 else if (PL_regkind[(U8)op] == BRANCH && PL_regkind[OP(next)] != BRANCH )
14615 PerlIO_printf(Perl_debug_log, " (FAIL)");
14617 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
14618 (void)PerlIO_putc(Perl_debug_log, '\n');
14622 if (PL_regkind[(U8)op] == BRANCHJ) {
14625 const regnode *nnode = (OP(next) == LONGJMP
14626 ? regnext((regnode *)next)
14628 if (last && nnode > last)
14630 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
14633 else if (PL_regkind[(U8)op] == BRANCH) {
14635 DUMPUNTIL(NEXTOPER(node), next);
14637 else if ( PL_regkind[(U8)op] == TRIE ) {
14638 const regnode *this_trie = node;
14639 const char op = OP(node);
14640 const U32 n = ARG(node);
14641 const reg_ac_data * const ac = op>=AHOCORASICK ?
14642 (reg_ac_data *)ri->data->data[n] :
14644 const reg_trie_data * const trie =
14645 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
14647 AV *const trie_words = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
14649 const regnode *nextbranch= NULL;
14652 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
14653 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
14655 PerlIO_printf(Perl_debug_log, "%*s%s ",
14656 (int)(2*(indent+3)), "",
14657 elem_ptr ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr), SvCUR(*elem_ptr), 60,
14658 PL_colors[0], PL_colors[1],
14659 (SvUTF8(*elem_ptr) ? PERL_PV_ESCAPE_UNI : 0) |
14660 PERL_PV_PRETTY_ELLIPSES |
14661 PERL_PV_PRETTY_LTGT
14666 U16 dist= trie->jump[word_idx+1];
14667 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
14668 (UV)((dist ? this_trie + dist : next) - start));
14671 nextbranch= this_trie + trie->jump[0];
14672 DUMPUNTIL(this_trie + dist, nextbranch);
14674 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
14675 nextbranch= regnext((regnode *)nextbranch);
14677 PerlIO_printf(Perl_debug_log, "\n");
14680 if (last && next > last)
14685 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
14686 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
14687 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
14689 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
14691 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
14693 else if ( op == PLUS || op == STAR) {
14694 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
14696 else if (PL_regkind[(U8)op] == ANYOF) {
14697 /* arglen 1 + class block */
14698 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_CLASS)
14699 ? ANYOF_CLASS_SKIP : ANYOF_SKIP);
14700 node = NEXTOPER(node);
14702 else if (PL_regkind[(U8)op] == EXACT) {
14703 /* Literal string, where present. */
14704 node += NODE_SZ_STR(node) - 1;
14705 node = NEXTOPER(node);
14708 node = NEXTOPER(node);
14709 node += regarglen[(U8)op];
14711 if (op == CURLYX || op == OPEN)
14715 #ifdef DEBUG_DUMPUNTIL
14716 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
14721 #endif /* DEBUGGING */
14725 * c-indentation-style: bsd
14726 * c-basic-offset: 4
14727 * indent-tabs-mode: nil
14730 * ex: set ts=8 sts=4 sw=4 et:
https://perl5.git.perl.org / perl5.git / blob