5 * 'A fair jaw-cracker dwarf-language must be.' --Samwise Gamgee
7 * [p.285 of _The Lord of the Rings_, II/iii: "The Ring Goes South"]
10 /* This file contains functions for compiling a regular expression. See
11 * also regexec.c which funnily enough, contains functions for executing
12 * a regular expression.
14 * This file is also copied at build time to ext/re/re_comp.c, where
15 * it's built with -DPERL_EXT_RE_BUILD -DPERL_EXT_RE_DEBUG -DPERL_EXT.
16 * This causes the main functions to be compiled under new names and with
17 * debugging support added, which makes "use re 'debug'" work.
20 /* NOTE: this is derived from Henry Spencer's regexp code, and should not
21 * confused with the original package (see point 3 below). Thanks, Henry!
24 /* Additional note: this code is very heavily munged from Henry's version
25 * in places. In some spots I've traded clarity for efficiency, so don't
26 * blame Henry for some of the lack of readability.
29 /* The names of the functions have been changed from regcomp and
30 * regexec to pregcomp and pregexec in order to avoid conflicts
31 * with the POSIX routines of the same names.
34 #ifdef PERL_EXT_RE_BUILD
39 * pregcomp and pregexec -- regsub and regerror are not used in perl
41 * Copyright (c) 1986 by University of Toronto.
42 * Written by Henry Spencer. Not derived from licensed software.
44 * Permission is granted to anyone to use this software for any
45 * purpose on any computer system, and to redistribute it freely,
46 * subject to the following restrictions:
48 * 1. The author is not responsible for the consequences of use of
49 * this software, no matter how awful, even if they arise
52 * 2. The origin of this software must not be misrepresented, either
53 * by explicit claim or by omission.
55 * 3. Altered versions must be plainly marked as such, and must not
56 * be misrepresented as being the original software.
59 **** Alterations to Henry's code are...
61 **** Copyright (C) 1991, 1992, 1993, 1994, 1995, 1996, 1997, 1998, 1999,
62 **** 2000, 2001, 2002, 2003, 2004, 2005, 2006, 2007, 2008
63 **** by Larry Wall and others
65 **** You may distribute under the terms of either the GNU General Public
66 **** License or the Artistic License, as specified in the README file.
69 * Beware that some of this code is subtly aware of the way operator
70 * precedence is structured in regular expressions. Serious changes in
71 * regular-expression syntax might require a total rethink.
74 #define PERL_IN_REGCOMP_C
77 #ifndef PERL_IN_XSUB_RE
82 #ifdef PERL_IN_XSUB_RE
84 EXTERN_C const struct regexp_engine my_reg_engine;
89 #include "dquote_static.c"
90 #include "inline_invlist.c"
91 #include "unicode_constants.h"
93 #define HAS_NONLATIN1_FOLD_CLOSURE(i) \
94 _HAS_NONLATIN1_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
95 #define HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(i) \
96 _HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE_ONLY_FOR_USE_BY_REGCOMP_DOT_C_AND_REGEXEC_DOT_C(i)
97 #define IS_NON_FINAL_FOLD(c) _IS_NON_FINAL_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
98 #define IS_IN_SOME_FOLD_L1(c) _IS_IN_SOME_FOLD_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
101 #define STATIC static
105 #define MIN(a,b) ((a) < (b) ? (a) : (b))
108 /* this is a chain of data about sub patterns we are processing that
109 need to be handled separately/specially in study_chunk. Its so
110 we can simulate recursion without losing state. */
112 typedef struct scan_frame {
113 regnode *last_regnode; /* last node to process in this frame */
114 regnode *next_regnode; /* next node to process when last is reached */
115 U32 prev_recursed_depth;
116 I32 stopparen; /* what stopparen do we use */
117 U32 is_top_frame; /* what flags do we use? */
119 struct scan_frame *this_prev_frame; /* this previous frame */
120 struct scan_frame *prev_frame; /* previous frame */
121 struct scan_frame *next_frame; /* next frame */
124 /* Certain characters are output as a sequence with the first being a
126 #define isBACKSLASHED_PUNCT(c) \
127 ((c) == '-' || (c) == ']' || (c) == '\\' || (c) == '^')
130 struct RExC_state_t {
131 U32 flags; /* RXf_* are we folding, multilining? */
132 U32 pm_flags; /* PMf_* stuff from the calling PMOP */
133 char *precomp; /* uncompiled string. */
134 REGEXP *rx_sv; /* The SV that is the regexp. */
135 regexp *rx; /* perl core regexp structure */
136 regexp_internal *rxi; /* internal data for regexp object
138 char *start; /* Start of input for compile */
139 char *end; /* End of input for compile */
140 char *parse; /* Input-scan pointer. */
141 SSize_t whilem_seen; /* number of WHILEM in this expr */
142 regnode *emit_start; /* Start of emitted-code area */
143 regnode *emit_bound; /* First regnode outside of the
145 regnode *emit; /* Code-emit pointer; if = &emit_dummy,
146 implies compiling, so don't emit */
147 regnode_ssc emit_dummy; /* placeholder for emit to point to;
148 large enough for the largest
149 non-EXACTish node, so can use it as
151 I32 naughty; /* How bad is this pattern? */
152 I32 sawback; /* Did we see \1, ...? */
154 SSize_t size; /* Code size. */
155 I32 npar; /* Capture buffer count, (OPEN) plus
156 one. ("par" 0 is the whole
158 I32 nestroot; /* root parens we are in - used by
162 regnode **open_parens; /* pointers to open parens */
163 regnode **close_parens; /* pointers to close parens */
164 regnode *opend; /* END node in program */
165 I32 utf8; /* whether the pattern is utf8 or not */
166 I32 orig_utf8; /* whether the pattern was originally in utf8 */
167 /* XXX use this for future optimisation of case
168 * where pattern must be upgraded to utf8. */
169 I32 uni_semantics; /* If a d charset modifier should use unicode
170 rules, even if the pattern is not in
172 HV *paren_names; /* Paren names */
174 regnode **recurse; /* Recurse regops */
175 I32 recurse_count; /* Number of recurse regops */
176 U8 *study_chunk_recursed; /* bitmap of which subs we have moved
178 U32 study_chunk_recursed_bytes; /* bytes in bitmap */
182 I32 override_recoding;
184 I32 recode_x_to_native;
186 I32 in_multi_char_class;
187 struct reg_code_block *code_blocks; /* positions of literal (?{})
189 int num_code_blocks; /* size of code_blocks[] */
190 int code_index; /* next code_blocks[] slot */
191 SSize_t maxlen; /* mininum possible number of chars in string to match */
192 scan_frame *frame_head;
193 scan_frame *frame_last;
196 #ifdef ADD_TO_REGEXEC
197 char *starttry; /* -Dr: where regtry was called. */
198 #define RExC_starttry (pRExC_state->starttry)
200 SV *runtime_code_qr; /* qr with the runtime code blocks */
202 const char *lastparse;
204 AV *paren_name_list; /* idx -> name */
205 U32 study_chunk_recursed_count;
208 #define RExC_lastparse (pRExC_state->lastparse)
209 #define RExC_lastnum (pRExC_state->lastnum)
210 #define RExC_paren_name_list (pRExC_state->paren_name_list)
211 #define RExC_study_chunk_recursed_count (pRExC_state->study_chunk_recursed_count)
212 #define RExC_mysv (pRExC_state->mysv1)
213 #define RExC_mysv1 (pRExC_state->mysv1)
214 #define RExC_mysv2 (pRExC_state->mysv2)
219 #define RExC_flags (pRExC_state->flags)
220 #define RExC_pm_flags (pRExC_state->pm_flags)
221 #define RExC_precomp (pRExC_state->precomp)
222 #define RExC_rx_sv (pRExC_state->rx_sv)
223 #define RExC_rx (pRExC_state->rx)
224 #define RExC_rxi (pRExC_state->rxi)
225 #define RExC_start (pRExC_state->start)
226 #define RExC_end (pRExC_state->end)
227 #define RExC_parse (pRExC_state->parse)
228 #define RExC_whilem_seen (pRExC_state->whilem_seen)
229 #ifdef RE_TRACK_PATTERN_OFFSETS
230 #define RExC_offsets (pRExC_state->rxi->u.offsets) /* I am not like the
233 #define RExC_emit (pRExC_state->emit)
234 #define RExC_emit_dummy (pRExC_state->emit_dummy)
235 #define RExC_emit_start (pRExC_state->emit_start)
236 #define RExC_emit_bound (pRExC_state->emit_bound)
237 #define RExC_sawback (pRExC_state->sawback)
238 #define RExC_seen (pRExC_state->seen)
239 #define RExC_size (pRExC_state->size)
240 #define RExC_maxlen (pRExC_state->maxlen)
241 #define RExC_npar (pRExC_state->npar)
242 #define RExC_nestroot (pRExC_state->nestroot)
243 #define RExC_extralen (pRExC_state->extralen)
244 #define RExC_seen_zerolen (pRExC_state->seen_zerolen)
245 #define RExC_utf8 (pRExC_state->utf8)
246 #define RExC_uni_semantics (pRExC_state->uni_semantics)
247 #define RExC_orig_utf8 (pRExC_state->orig_utf8)
248 #define RExC_open_parens (pRExC_state->open_parens)
249 #define RExC_close_parens (pRExC_state->close_parens)
250 #define RExC_opend (pRExC_state->opend)
251 #define RExC_paren_names (pRExC_state->paren_names)
252 #define RExC_recurse (pRExC_state->recurse)
253 #define RExC_recurse_count (pRExC_state->recurse_count)
254 #define RExC_study_chunk_recursed (pRExC_state->study_chunk_recursed)
255 #define RExC_study_chunk_recursed_bytes \
256 (pRExC_state->study_chunk_recursed_bytes)
257 #define RExC_in_lookbehind (pRExC_state->in_lookbehind)
258 #define RExC_contains_locale (pRExC_state->contains_locale)
259 #define RExC_contains_i (pRExC_state->contains_i)
260 #define RExC_override_recoding (pRExC_state->override_recoding)
262 # define RExC_recode_x_to_native (pRExC_state->recode_x_to_native)
264 #define RExC_in_multi_char_class (pRExC_state->in_multi_char_class)
265 #define RExC_frame_head (pRExC_state->frame_head)
266 #define RExC_frame_last (pRExC_state->frame_last)
267 #define RExC_frame_count (pRExC_state->frame_count)
268 #define RExC_strict (pRExC_state->strict)
270 /* Heuristic check on the complexity of the pattern: if TOO_NAUGHTY, we set
271 * a flag to disable back-off on the fixed/floating substrings - if it's
272 * a high complexity pattern we assume the benefit of avoiding a full match
273 * is worth the cost of checking for the substrings even if they rarely help.
275 #define RExC_naughty (pRExC_state->naughty)
276 #define TOO_NAUGHTY (10)
277 #define MARK_NAUGHTY(add) \
278 if (RExC_naughty < TOO_NAUGHTY) \
279 RExC_naughty += (add)
280 #define MARK_NAUGHTY_EXP(exp, add) \
281 if (RExC_naughty < TOO_NAUGHTY) \
282 RExC_naughty += RExC_naughty / (exp) + (add)
284 #define ISMULT1(c) ((c) == '*' || (c) == '+' || (c) == '?')
285 #define ISMULT2(s) ((*s) == '*' || (*s) == '+' || (*s) == '?' || \
286 ((*s) == '{' && regcurly(s)))
289 * Flags to be passed up and down.
291 #define WORST 0 /* Worst case. */
292 #define HASWIDTH 0x01 /* Known to match non-null strings. */
294 /* Simple enough to be STAR/PLUS operand; in an EXACTish node must be a single
295 * character. (There needs to be a case: in the switch statement in regexec.c
296 * for any node marked SIMPLE.) Note that this is not the same thing as
299 #define SPSTART 0x04 /* Starts with * or + */
300 #define POSTPONED 0x08 /* (?1),(?&name), (??{...}) or similar */
301 #define TRYAGAIN 0x10 /* Weeded out a declaration. */
302 #define RESTART_UTF8 0x20 /* Restart, need to calcuate sizes as UTF-8 */
304 #define REG_NODE_NUM(x) ((x) ? (int)((x)-RExC_emit_start) : -1)
306 /* whether trie related optimizations are enabled */
307 #if PERL_ENABLE_EXTENDED_TRIE_OPTIMISATION
308 #define TRIE_STUDY_OPT
309 #define FULL_TRIE_STUDY
315 #define PBYTE(u8str,paren) ((U8*)(u8str))[(paren) >> 3]
316 #define PBITVAL(paren) (1 << ((paren) & 7))
317 #define PAREN_TEST(u8str,paren) ( PBYTE(u8str,paren) & PBITVAL(paren))
318 #define PAREN_SET(u8str,paren) PBYTE(u8str,paren) |= PBITVAL(paren)
319 #define PAREN_UNSET(u8str,paren) PBYTE(u8str,paren) &= (~PBITVAL(paren))
321 #define REQUIRE_UTF8 STMT_START { \
323 *flagp = RESTART_UTF8; \
328 /* This converts the named class defined in regcomp.h to its equivalent class
329 * number defined in handy.h. */
330 #define namedclass_to_classnum(class) ((int) ((class) / 2))
331 #define classnum_to_namedclass(classnum) ((classnum) * 2)
333 #define _invlist_union_complement_2nd(a, b, output) \
334 _invlist_union_maybe_complement_2nd(a, b, TRUE, output)
335 #define _invlist_intersection_complement_2nd(a, b, output) \
336 _invlist_intersection_maybe_complement_2nd(a, b, TRUE, output)
338 /* About scan_data_t.
340 During optimisation we recurse through the regexp program performing
341 various inplace (keyhole style) optimisations. In addition study_chunk
342 and scan_commit populate this data structure with information about
343 what strings MUST appear in the pattern. We look for the longest
344 string that must appear at a fixed location, and we look for the
345 longest string that may appear at a floating location. So for instance
350 Both 'FOO' and 'A' are fixed strings. Both 'B' and 'BAR' are floating
351 strings (because they follow a .* construct). study_chunk will identify
352 both FOO and BAR as being the longest fixed and floating strings respectively.
354 The strings can be composites, for instance
358 will result in a composite fixed substring 'foo'.
360 For each string some basic information is maintained:
362 - offset or min_offset
363 This is the position the string must appear at, or not before.
364 It also implicitly (when combined with minlenp) tells us how many
365 characters must match before the string we are searching for.
366 Likewise when combined with minlenp and the length of the string it
367 tells us how many characters must appear after the string we have
371 Only used for floating strings. This is the rightmost point that
372 the string can appear at. If set to SSize_t_MAX it indicates that the
373 string can occur infinitely far to the right.
376 A pointer to the minimum number of characters of the pattern that the
377 string was found inside. This is important as in the case of positive
378 lookahead or positive lookbehind we can have multiple patterns
383 The minimum length of the pattern overall is 3, the minimum length
384 of the lookahead part is 3, but the minimum length of the part that
385 will actually match is 1. So 'FOO's minimum length is 3, but the
386 minimum length for the F is 1. This is important as the minimum length
387 is used to determine offsets in front of and behind the string being
388 looked for. Since strings can be composites this is the length of the
389 pattern at the time it was committed with a scan_commit. Note that
390 the length is calculated by study_chunk, so that the minimum lengths
391 are not known until the full pattern has been compiled, thus the
392 pointer to the value.
396 In the case of lookbehind the string being searched for can be
397 offset past the start point of the final matching string.
398 If this value was just blithely removed from the min_offset it would
399 invalidate some of the calculations for how many chars must match
400 before or after (as they are derived from min_offset and minlen and
401 the length of the string being searched for).
402 When the final pattern is compiled and the data is moved from the
403 scan_data_t structure into the regexp structure the information
404 about lookbehind is factored in, with the information that would
405 have been lost precalculated in the end_shift field for the
408 The fields pos_min and pos_delta are used to store the minimum offset
409 and the delta to the maximum offset at the current point in the pattern.
413 typedef struct scan_data_t {
414 /*I32 len_min; unused */
415 /*I32 len_delta; unused */
419 SSize_t last_end; /* min value, <0 unless valid. */
420 SSize_t last_start_min;
421 SSize_t last_start_max;
422 SV **longest; /* Either &l_fixed, or &l_float. */
423 SV *longest_fixed; /* longest fixed string found in pattern */
424 SSize_t offset_fixed; /* offset where it starts */
425 SSize_t *minlen_fixed; /* pointer to the minlen relevant to the string */
426 I32 lookbehind_fixed; /* is the position of the string modfied by LB */
427 SV *longest_float; /* longest floating string found in pattern */
428 SSize_t offset_float_min; /* earliest point in string it can appear */
429 SSize_t offset_float_max; /* latest point in string it can appear */
430 SSize_t *minlen_float; /* pointer to the minlen relevant to the string */
431 SSize_t lookbehind_float; /* is the pos of the string modified by LB */
434 SSize_t *last_closep;
435 regnode_ssc *start_class;
439 * Forward declarations for pregcomp()'s friends.
442 static const scan_data_t zero_scan_data =
443 { 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0 ,0};
445 #define SF_BEFORE_EOL (SF_BEFORE_SEOL|SF_BEFORE_MEOL)
446 #define SF_BEFORE_SEOL 0x0001
447 #define SF_BEFORE_MEOL 0x0002
448 #define SF_FIX_BEFORE_EOL (SF_FIX_BEFORE_SEOL|SF_FIX_BEFORE_MEOL)
449 #define SF_FL_BEFORE_EOL (SF_FL_BEFORE_SEOL|SF_FL_BEFORE_MEOL)
451 #define SF_FIX_SHIFT_EOL (+2)
452 #define SF_FL_SHIFT_EOL (+4)
454 #define SF_FIX_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FIX_SHIFT_EOL)
455 #define SF_FIX_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FIX_SHIFT_EOL)
457 #define SF_FL_BEFORE_SEOL (SF_BEFORE_SEOL << SF_FL_SHIFT_EOL)
458 #define SF_FL_BEFORE_MEOL (SF_BEFORE_MEOL << SF_FL_SHIFT_EOL) /* 0x20 */
459 #define SF_IS_INF 0x0040
460 #define SF_HAS_PAR 0x0080
461 #define SF_IN_PAR 0x0100
462 #define SF_HAS_EVAL 0x0200
463 #define SCF_DO_SUBSTR 0x0400
464 #define SCF_DO_STCLASS_AND 0x0800
465 #define SCF_DO_STCLASS_OR 0x1000
466 #define SCF_DO_STCLASS (SCF_DO_STCLASS_AND|SCF_DO_STCLASS_OR)
467 #define SCF_WHILEM_VISITED_POS 0x2000
469 #define SCF_TRIE_RESTUDY 0x4000 /* Do restudy? */
470 #define SCF_SEEN_ACCEPT 0x8000
471 #define SCF_TRIE_DOING_RESTUDY 0x10000
472 #define SCF_IN_DEFINE 0x20000
477 #define UTF cBOOL(RExC_utf8)
479 /* The enums for all these are ordered so things work out correctly */
480 #define LOC (get_regex_charset(RExC_flags) == REGEX_LOCALE_CHARSET)
481 #define DEPENDS_SEMANTICS (get_regex_charset(RExC_flags) \
482 == REGEX_DEPENDS_CHARSET)
483 #define UNI_SEMANTICS (get_regex_charset(RExC_flags) == REGEX_UNICODE_CHARSET)
484 #define AT_LEAST_UNI_SEMANTICS (get_regex_charset(RExC_flags) \
485 >= REGEX_UNICODE_CHARSET)
486 #define ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
487 == REGEX_ASCII_RESTRICTED_CHARSET)
488 #define AT_LEAST_ASCII_RESTRICTED (get_regex_charset(RExC_flags) \
489 >= REGEX_ASCII_RESTRICTED_CHARSET)
490 #define ASCII_FOLD_RESTRICTED (get_regex_charset(RExC_flags) \
491 == REGEX_ASCII_MORE_RESTRICTED_CHARSET)
493 #define FOLD cBOOL(RExC_flags & RXf_PMf_FOLD)
495 /* For programs that want to be strictly Unicode compatible by dying if any
496 * attempt is made to match a non-Unicode code point against a Unicode
498 #define ALWAYS_WARN_SUPER ckDEAD(packWARN(WARN_NON_UNICODE))
500 #define OOB_NAMEDCLASS -1
502 /* There is no code point that is out-of-bounds, so this is problematic. But
503 * its only current use is to initialize a variable that is always set before
505 #define OOB_UNICODE 0xDEADBEEF
507 #define CHR_SVLEN(sv) (UTF ? sv_len_utf8(sv) : SvCUR(sv))
508 #define CHR_DIST(a,b) (UTF ? utf8_distance(a,b) : a - b)
511 /* length of regex to show in messages that don't mark a position within */
512 #define RegexLengthToShowInErrorMessages 127
515 * If MARKER[12] are adjusted, be sure to adjust the constants at the top
516 * of t/op/regmesg.t, the tests in t/op/re_tests, and those in
517 * op/pragma/warn/regcomp.
519 #define MARKER1 "<-- HERE" /* marker as it appears in the description */
520 #define MARKER2 " <-- HERE " /* marker as it appears within the regex */
522 #define REPORT_LOCATION " in regex; marked by " MARKER1 \
523 " in m/%"UTF8f MARKER2 "%"UTF8f"/"
525 #define REPORT_LOCATION_ARGS(offset) \
526 UTF8fARG(UTF, offset, RExC_precomp), \
527 UTF8fARG(UTF, RExC_end - RExC_precomp - offset, RExC_precomp + offset)
529 /* Used to point after bad bytes for an error message, but avoid skipping
530 * past a nul byte. */
531 #define SKIP_IF_CHAR(s) (!*(s) ? 0 : UTF ? UTF8SKIP(s) : 1)
534 * Calls SAVEDESTRUCTOR_X if needed, then calls Perl_croak with the given
535 * arg. Show regex, up to a maximum length. If it's too long, chop and add
538 #define _FAIL(code) STMT_START { \
539 const char *ellipses = ""; \
540 IV len = RExC_end - RExC_precomp; \
543 SAVEFREESV(RExC_rx_sv); \
544 if (len > RegexLengthToShowInErrorMessages) { \
545 /* chop 10 shorter than the max, to ensure meaning of "..." */ \
546 len = RegexLengthToShowInErrorMessages - 10; \
552 #define FAIL(msg) _FAIL( \
553 Perl_croak(aTHX_ "%s in regex m/%"UTF8f"%s/", \
554 msg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
556 #define FAIL2(msg,arg) _FAIL( \
557 Perl_croak(aTHX_ msg " in regex m/%"UTF8f"%s/", \
558 arg, UTF8fARG(UTF, len, RExC_precomp), ellipses))
561 * Simple_vFAIL -- like FAIL, but marks the current location in the scan
563 #define Simple_vFAIL(m) STMT_START { \
565 (RExC_parse > RExC_end ? RExC_end : RExC_parse) - RExC_precomp; \
566 Perl_croak(aTHX_ "%s" REPORT_LOCATION, \
567 m, REPORT_LOCATION_ARGS(offset)); \
571 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL()
573 #define vFAIL(m) STMT_START { \
575 SAVEFREESV(RExC_rx_sv); \
580 * Like Simple_vFAIL(), but accepts two arguments.
582 #define Simple_vFAIL2(m,a1) STMT_START { \
583 const IV offset = RExC_parse - RExC_precomp; \
584 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
585 REPORT_LOCATION_ARGS(offset)); \
589 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL2().
591 #define vFAIL2(m,a1) STMT_START { \
593 SAVEFREESV(RExC_rx_sv); \
594 Simple_vFAIL2(m, a1); \
599 * Like Simple_vFAIL(), but accepts three arguments.
601 #define Simple_vFAIL3(m, a1, a2) STMT_START { \
602 const IV offset = RExC_parse - RExC_precomp; \
603 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, \
604 REPORT_LOCATION_ARGS(offset)); \
608 * Calls SAVEDESTRUCTOR_X if needed, then Simple_vFAIL3().
610 #define vFAIL3(m,a1,a2) STMT_START { \
612 SAVEFREESV(RExC_rx_sv); \
613 Simple_vFAIL3(m, a1, a2); \
617 * Like Simple_vFAIL(), but accepts four arguments.
619 #define Simple_vFAIL4(m, a1, a2, a3) STMT_START { \
620 const IV offset = RExC_parse - RExC_precomp; \
621 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, a2, a3, \
622 REPORT_LOCATION_ARGS(offset)); \
625 #define vFAIL4(m,a1,a2,a3) STMT_START { \
627 SAVEFREESV(RExC_rx_sv); \
628 Simple_vFAIL4(m, a1, a2, a3); \
631 /* A specialized version of vFAIL2 that works with UTF8f */
632 #define vFAIL2utf8f(m, a1) STMT_START { \
633 const IV offset = RExC_parse - RExC_precomp; \
635 SAVEFREESV(RExC_rx_sv); \
636 S_re_croak2(aTHX_ UTF, m, REPORT_LOCATION, a1, \
637 REPORT_LOCATION_ARGS(offset)); \
640 /* These have asserts in them because of [perl #122671] Many warnings in
641 * regcomp.c can occur twice. If they get output in pass1 and later in that
642 * pass, the pattern has to be converted to UTF-8 and the pass restarted, they
643 * would get output again. So they should be output in pass2, and these
644 * asserts make sure new warnings follow that paradigm. */
646 /* m is not necessarily a "literal string", in this macro */
647 #define reg_warn_non_literal_string(loc, m) STMT_START { \
648 const IV offset = loc - RExC_precomp; \
649 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), "%s" REPORT_LOCATION, \
650 m, REPORT_LOCATION_ARGS(offset)); \
653 #define ckWARNreg(loc,m) STMT_START { \
654 const IV offset = loc - RExC_precomp; \
655 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
656 REPORT_LOCATION_ARGS(offset)); \
659 #define vWARN(loc, m) STMT_START { \
660 const IV offset = loc - RExC_precomp; \
661 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
662 REPORT_LOCATION_ARGS(offset)); \
665 #define vWARN_dep(loc, m) STMT_START { \
666 const IV offset = loc - RExC_precomp; \
667 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_DEPRECATED), m REPORT_LOCATION, \
668 REPORT_LOCATION_ARGS(offset)); \
671 #define ckWARNdep(loc,m) STMT_START { \
672 const IV offset = loc - RExC_precomp; \
673 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_DEPRECATED), \
675 REPORT_LOCATION_ARGS(offset)); \
678 #define ckWARNregdep(loc,m) STMT_START { \
679 const IV offset = loc - RExC_precomp; \
680 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN2(WARN_DEPRECATED, WARN_REGEXP), \
682 REPORT_LOCATION_ARGS(offset)); \
685 #define ckWARN2reg_d(loc,m, a1) STMT_START { \
686 const IV offset = loc - RExC_precomp; \
687 __ASSERT_(PASS2) Perl_ck_warner_d(aTHX_ packWARN(WARN_REGEXP), \
689 a1, REPORT_LOCATION_ARGS(offset)); \
692 #define ckWARN2reg(loc, m, a1) STMT_START { \
693 const IV offset = loc - RExC_precomp; \
694 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
695 a1, REPORT_LOCATION_ARGS(offset)); \
698 #define vWARN3(loc, m, a1, a2) STMT_START { \
699 const IV offset = loc - RExC_precomp; \
700 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
701 a1, a2, REPORT_LOCATION_ARGS(offset)); \
704 #define ckWARN3reg(loc, m, a1, a2) STMT_START { \
705 const IV offset = loc - RExC_precomp; \
706 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
707 a1, a2, REPORT_LOCATION_ARGS(offset)); \
710 #define vWARN4(loc, m, a1, a2, a3) STMT_START { \
711 const IV offset = loc - RExC_precomp; \
712 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
713 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
716 #define ckWARN4reg(loc, m, a1, a2, a3) STMT_START { \
717 const IV offset = loc - RExC_precomp; \
718 __ASSERT_(PASS2) Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
719 a1, a2, a3, REPORT_LOCATION_ARGS(offset)); \
722 #define vWARN5(loc, m, a1, a2, a3, a4) STMT_START { \
723 const IV offset = loc - RExC_precomp; \
724 __ASSERT_(PASS2) Perl_warner(aTHX_ packWARN(WARN_REGEXP), m REPORT_LOCATION, \
725 a1, a2, a3, a4, REPORT_LOCATION_ARGS(offset)); \
728 /* Macros for recording node offsets. 20001227 mjd@plover.com
729 * Nodes are numbered 1, 2, 3, 4. Node #n's position is recorded in
730 * element 2*n-1 of the array. Element #2n holds the byte length node #n.
731 * Element 0 holds the number n.
732 * Position is 1 indexed.
734 #ifndef RE_TRACK_PATTERN_OFFSETS
735 #define Set_Node_Offset_To_R(node,byte)
736 #define Set_Node_Offset(node,byte)
737 #define Set_Cur_Node_Offset
738 #define Set_Node_Length_To_R(node,len)
739 #define Set_Node_Length(node,len)
740 #define Set_Node_Cur_Length(node,start)
741 #define Node_Offset(n)
742 #define Node_Length(n)
743 #define Set_Node_Offset_Length(node,offset,len)
744 #define ProgLen(ri) ri->u.proglen
745 #define SetProgLen(ri,x) ri->u.proglen = x
747 #define ProgLen(ri) ri->u.offsets[0]
748 #define SetProgLen(ri,x) ri->u.offsets[0] = x
749 #define Set_Node_Offset_To_R(node,byte) STMT_START { \
751 MJD_OFFSET_DEBUG(("** (%d) offset of node %d is %d.\n", \
752 __LINE__, (int)(node), (int)(byte))); \
754 Perl_croak(aTHX_ "value of node is %d in Offset macro", \
757 RExC_offsets[2*(node)-1] = (byte); \
762 #define Set_Node_Offset(node,byte) \
763 Set_Node_Offset_To_R((node)-RExC_emit_start, (byte)-RExC_start)
764 #define Set_Cur_Node_Offset Set_Node_Offset(RExC_emit, RExC_parse)
766 #define Set_Node_Length_To_R(node,len) STMT_START { \
768 MJD_OFFSET_DEBUG(("** (%d) size of node %d is %d.\n", \
769 __LINE__, (int)(node), (int)(len))); \
771 Perl_croak(aTHX_ "value of node is %d in Length macro", \
774 RExC_offsets[2*(node)] = (len); \
779 #define Set_Node_Length(node,len) \
780 Set_Node_Length_To_R((node)-RExC_emit_start, len)
781 #define Set_Node_Cur_Length(node, start) \
782 Set_Node_Length(node, RExC_parse - start)
784 /* Get offsets and lengths */
785 #define Node_Offset(n) (RExC_offsets[2*((n)-RExC_emit_start)-1])
786 #define Node_Length(n) (RExC_offsets[2*((n)-RExC_emit_start)])
788 #define Set_Node_Offset_Length(node,offset,len) STMT_START { \
789 Set_Node_Offset_To_R((node)-RExC_emit_start, (offset)); \
790 Set_Node_Length_To_R((node)-RExC_emit_start, (len)); \
794 #if PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS
795 #define EXPERIMENTAL_INPLACESCAN
796 #endif /*PERL_ENABLE_EXPERIMENTAL_REGEX_OPTIMISATIONS*/
798 #define DEBUG_RExC_seen() \
799 DEBUG_OPTIMISE_MORE_r({ \
800 PerlIO_printf(Perl_debug_log,"RExC_seen: "); \
802 if (RExC_seen & REG_ZERO_LEN_SEEN) \
803 PerlIO_printf(Perl_debug_log,"REG_ZERO_LEN_SEEN "); \
805 if (RExC_seen & REG_LOOKBEHIND_SEEN) \
806 PerlIO_printf(Perl_debug_log,"REG_LOOKBEHIND_SEEN "); \
808 if (RExC_seen & REG_GPOS_SEEN) \
809 PerlIO_printf(Perl_debug_log,"REG_GPOS_SEEN "); \
811 if (RExC_seen & REG_CANY_SEEN) \
812 PerlIO_printf(Perl_debug_log,"REG_CANY_SEEN "); \
814 if (RExC_seen & REG_RECURSE_SEEN) \
815 PerlIO_printf(Perl_debug_log,"REG_RECURSE_SEEN "); \
817 if (RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN) \
818 PerlIO_printf(Perl_debug_log,"REG_TOP_LEVEL_BRANCHES_SEEN "); \
820 if (RExC_seen & REG_VERBARG_SEEN) \
821 PerlIO_printf(Perl_debug_log,"REG_VERBARG_SEEN "); \
823 if (RExC_seen & REG_CUTGROUP_SEEN) \
824 PerlIO_printf(Perl_debug_log,"REG_CUTGROUP_SEEN "); \
826 if (RExC_seen & REG_RUN_ON_COMMENT_SEEN) \
827 PerlIO_printf(Perl_debug_log,"REG_RUN_ON_COMMENT_SEEN "); \
829 if (RExC_seen & REG_UNFOLDED_MULTI_SEEN) \
830 PerlIO_printf(Perl_debug_log,"REG_UNFOLDED_MULTI_SEEN "); \
832 if (RExC_seen & REG_GOSTART_SEEN) \
833 PerlIO_printf(Perl_debug_log,"REG_GOSTART_SEEN "); \
835 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) \
836 PerlIO_printf(Perl_debug_log,"REG_UNBOUNDED_QUANTIFIER_SEEN "); \
838 PerlIO_printf(Perl_debug_log,"\n"); \
841 #define DEBUG_SHOW_STUDY_FLAG(flags,flag) \
842 if ((flags) & flag) PerlIO_printf(Perl_debug_log, "%s ", #flag)
844 #define DEBUG_SHOW_STUDY_FLAGS(flags,open_str,close_str) \
846 PerlIO_printf(Perl_debug_log, "%s", open_str); \
847 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_SEOL); \
848 DEBUG_SHOW_STUDY_FLAG(flags,SF_FL_BEFORE_MEOL); \
849 DEBUG_SHOW_STUDY_FLAG(flags,SF_IS_INF); \
850 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_PAR); \
851 DEBUG_SHOW_STUDY_FLAG(flags,SF_IN_PAR); \
852 DEBUG_SHOW_STUDY_FLAG(flags,SF_HAS_EVAL); \
853 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_SUBSTR); \
854 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_AND); \
855 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS_OR); \
856 DEBUG_SHOW_STUDY_FLAG(flags,SCF_DO_STCLASS); \
857 DEBUG_SHOW_STUDY_FLAG(flags,SCF_WHILEM_VISITED_POS); \
858 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_RESTUDY); \
859 DEBUG_SHOW_STUDY_FLAG(flags,SCF_SEEN_ACCEPT); \
860 DEBUG_SHOW_STUDY_FLAG(flags,SCF_TRIE_DOING_RESTUDY); \
861 DEBUG_SHOW_STUDY_FLAG(flags,SCF_IN_DEFINE); \
862 PerlIO_printf(Perl_debug_log, "%s", close_str); \
866 #define DEBUG_STUDYDATA(str,data,depth) \
867 DEBUG_OPTIMISE_MORE_r(if(data){ \
868 PerlIO_printf(Perl_debug_log, \
869 "%*s" str "Pos:%"IVdf"/%"IVdf \
871 (int)(depth)*2, "", \
872 (IV)((data)->pos_min), \
873 (IV)((data)->pos_delta), \
874 (UV)((data)->flags) \
876 DEBUG_SHOW_STUDY_FLAGS((data)->flags," [ ","]"); \
877 PerlIO_printf(Perl_debug_log, \
878 " Whilem_c: %"IVdf" Lcp: %"IVdf" %s", \
879 (IV)((data)->whilem_c), \
880 (IV)((data)->last_closep ? *((data)->last_closep) : -1), \
881 is_inf ? "INF " : "" \
883 if ((data)->last_found) \
884 PerlIO_printf(Perl_debug_log, \
885 "Last:'%s' %"IVdf":%"IVdf"/%"IVdf" %sFixed:'%s' @ %"IVdf \
886 " %sFloat: '%s' @ %"IVdf"/%"IVdf"", \
887 SvPVX_const((data)->last_found), \
888 (IV)((data)->last_end), \
889 (IV)((data)->last_start_min), \
890 (IV)((data)->last_start_max), \
891 ((data)->longest && \
892 (data)->longest==&((data)->longest_fixed)) ? "*" : "", \
893 SvPVX_const((data)->longest_fixed), \
894 (IV)((data)->offset_fixed), \
895 ((data)->longest && \
896 (data)->longest==&((data)->longest_float)) ? "*" : "", \
897 SvPVX_const((data)->longest_float), \
898 (IV)((data)->offset_float_min), \
899 (IV)((data)->offset_float_max) \
901 PerlIO_printf(Perl_debug_log,"\n"); \
904 /* is c a control character for which we have a mnemonic? */
905 #define isMNEMONIC_CNTRL(c) _IS_MNEMONIC_CNTRL_ONLY_FOR_USE_BY_REGCOMP_DOT_C(c)
908 S_cntrl_to_mnemonic(const U8 c)
910 /* Returns the mnemonic string that represents character 'c', if one
911 * exists; NULL otherwise. The only ones that exist for the purposes of
912 * this routine are a few control characters */
915 case '\a': return "\\a";
916 case '\b': return "\\b";
917 case ESC_NATIVE: return "\\e";
918 case '\f': return "\\f";
919 case '\n': return "\\n";
920 case '\r': return "\\r";
921 case '\t': return "\\t";
927 /* Mark that we cannot extend a found fixed substring at this point.
928 Update the longest found anchored substring and the longest found
929 floating substrings if needed. */
932 S_scan_commit(pTHX_ const RExC_state_t *pRExC_state, scan_data_t *data,
933 SSize_t *minlenp, int is_inf)
935 const STRLEN l = CHR_SVLEN(data->last_found);
936 const STRLEN old_l = CHR_SVLEN(*data->longest);
937 GET_RE_DEBUG_FLAGS_DECL;
939 PERL_ARGS_ASSERT_SCAN_COMMIT;
941 if ((l >= old_l) && ((l > old_l) || (data->flags & SF_BEFORE_EOL))) {
942 SvSetMagicSV(*data->longest, data->last_found);
943 if (*data->longest == data->longest_fixed) {
944 data->offset_fixed = l ? data->last_start_min : data->pos_min;
945 if (data->flags & SF_BEFORE_EOL)
947 |= ((data->flags & SF_BEFORE_EOL) << SF_FIX_SHIFT_EOL);
949 data->flags &= ~SF_FIX_BEFORE_EOL;
950 data->minlen_fixed=minlenp;
951 data->lookbehind_fixed=0;
953 else { /* *data->longest == data->longest_float */
954 data->offset_float_min = l ? data->last_start_min : data->pos_min;
955 data->offset_float_max = (l
956 ? data->last_start_max
957 : (data->pos_delta > SSize_t_MAX - data->pos_min
959 : data->pos_min + data->pos_delta));
961 || (STRLEN)data->offset_float_max > (STRLEN)SSize_t_MAX)
962 data->offset_float_max = SSize_t_MAX;
963 if (data->flags & SF_BEFORE_EOL)
965 |= ((data->flags & SF_BEFORE_EOL) << SF_FL_SHIFT_EOL);
967 data->flags &= ~SF_FL_BEFORE_EOL;
968 data->minlen_float=minlenp;
969 data->lookbehind_float=0;
972 SvCUR_set(data->last_found, 0);
974 SV * const sv = data->last_found;
975 if (SvUTF8(sv) && SvMAGICAL(sv)) {
976 MAGIC * const mg = mg_find(sv, PERL_MAGIC_utf8);
982 data->flags &= ~SF_BEFORE_EOL;
983 DEBUG_STUDYDATA("commit: ",data,0);
986 /* An SSC is just a regnode_charclass_posix with an extra field: the inversion
987 * list that describes which code points it matches */
990 S_ssc_anything(pTHX_ regnode_ssc *ssc)
992 /* Set the SSC 'ssc' to match an empty string or any code point */
994 PERL_ARGS_ASSERT_SSC_ANYTHING;
996 assert(is_ANYOF_SYNTHETIC(ssc));
998 ssc->invlist = sv_2mortal(_new_invlist(2)); /* mortalize so won't leak */
999 _append_range_to_invlist(ssc->invlist, 0, UV_MAX);
1000 ANYOF_FLAGS(ssc) |= SSC_MATCHES_EMPTY_STRING; /* Plus matches empty */
1004 S_ssc_is_anything(const regnode_ssc *ssc)
1006 /* Returns TRUE if the SSC 'ssc' can match the empty string and any code
1007 * point; FALSE otherwise. Thus, this is used to see if using 'ssc' buys
1008 * us anything: if the function returns TRUE, 'ssc' hasn't been restricted
1009 * in any way, so there's no point in using it */
1014 PERL_ARGS_ASSERT_SSC_IS_ANYTHING;
1016 assert(is_ANYOF_SYNTHETIC(ssc));
1018 if (! (ANYOF_FLAGS(ssc) & SSC_MATCHES_EMPTY_STRING)) {
1022 /* See if the list consists solely of the range 0 - Infinity */
1023 invlist_iterinit(ssc->invlist);
1024 ret = invlist_iternext(ssc->invlist, &start, &end)
1028 invlist_iterfinish(ssc->invlist);
1034 /* If e.g., both \w and \W are set, matches everything */
1035 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1037 for (i = 0; i < ANYOF_POSIXL_MAX; i += 2) {
1038 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i+1)) {
1048 S_ssc_init(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc)
1050 /* Initializes the SSC 'ssc'. This includes setting it to match an empty
1051 * string, any code point, or any posix class under locale */
1053 PERL_ARGS_ASSERT_SSC_INIT;
1055 Zero(ssc, 1, regnode_ssc);
1056 set_ANYOF_SYNTHETIC(ssc);
1057 ARG_SET(ssc, ANYOF_ONLY_HAS_BITMAP);
1060 /* If any portion of the regex is to operate under locale rules that aren't
1061 * fully known at compile time, initialization includes it. The reason
1062 * this isn't done for all regexes is that the optimizer was written under
1063 * the assumption that locale was all-or-nothing. Given the complexity and
1064 * lack of documentation in the optimizer, and that there are inadequate
1065 * test cases for locale, many parts of it may not work properly, it is
1066 * safest to avoid locale unless necessary. */
1067 if (RExC_contains_locale) {
1068 ANYOF_POSIXL_SETALL(ssc);
1071 ANYOF_POSIXL_ZERO(ssc);
1076 S_ssc_is_cp_posixl_init(const RExC_state_t *pRExC_state,
1077 const regnode_ssc *ssc)
1079 /* Returns TRUE if the SSC 'ssc' is in its initial state with regard only
1080 * to the list of code points matched, and locale posix classes; hence does
1081 * not check its flags) */
1086 PERL_ARGS_ASSERT_SSC_IS_CP_POSIXL_INIT;
1088 assert(is_ANYOF_SYNTHETIC(ssc));
1090 invlist_iterinit(ssc->invlist);
1091 ret = invlist_iternext(ssc->invlist, &start, &end)
1095 invlist_iterfinish(ssc->invlist);
1101 if (RExC_contains_locale && ! ANYOF_POSIXL_SSC_TEST_ALL_SET(ssc)) {
1109 S_get_ANYOF_cp_list_for_ssc(pTHX_ const RExC_state_t *pRExC_state,
1110 const regnode_charclass* const node)
1112 /* Returns a mortal inversion list defining which code points are matched
1113 * by 'node', which is of type ANYOF. Handles complementing the result if
1114 * appropriate. If some code points aren't knowable at this time, the
1115 * returned list must, and will, contain every code point that is a
1118 SV* invlist = sv_2mortal(_new_invlist(0));
1119 SV* only_utf8_locale_invlist = NULL;
1121 const U32 n = ARG(node);
1122 bool new_node_has_latin1 = FALSE;
1124 PERL_ARGS_ASSERT_GET_ANYOF_CP_LIST_FOR_SSC;
1126 /* Look at the data structure created by S_set_ANYOF_arg() */
1127 if (n != ANYOF_ONLY_HAS_BITMAP) {
1128 SV * const rv = MUTABLE_SV(RExC_rxi->data->data[n]);
1129 AV * const av = MUTABLE_AV(SvRV(rv));
1130 SV **const ary = AvARRAY(av);
1131 assert(RExC_rxi->data->what[n] == 's');
1133 if (ary[1] && ary[1] != &PL_sv_undef) { /* Has compile-time swash */
1134 invlist = sv_2mortal(invlist_clone(_get_swash_invlist(ary[1])));
1136 else if (ary[0] && ary[0] != &PL_sv_undef) {
1138 /* Here, no compile-time swash, and there are things that won't be
1139 * known until runtime -- we have to assume it could be anything */
1140 return _add_range_to_invlist(invlist, 0, UV_MAX);
1142 else if (ary[3] && ary[3] != &PL_sv_undef) {
1144 /* Here no compile-time swash, and no run-time only data. Use the
1145 * node's inversion list */
1146 invlist = sv_2mortal(invlist_clone(ary[3]));
1149 /* Get the code points valid only under UTF-8 locales */
1150 if ((ANYOF_FLAGS(node) & ANYOF_LOC_FOLD)
1151 && ary[2] && ary[2] != &PL_sv_undef)
1153 only_utf8_locale_invlist = ary[2];
1157 /* An ANYOF node contains a bitmap for the first NUM_ANYOF_CODE_POINTS
1158 * code points, and an inversion list for the others, but if there are code
1159 * points that should match only conditionally on the target string being
1160 * UTF-8, those are placed in the inversion list, and not the bitmap.
1161 * Since there are circumstances under which they could match, they are
1162 * included in the SSC. But if the ANYOF node is to be inverted, we have
1163 * to exclude them here, so that when we invert below, the end result
1164 * actually does include them. (Think about "\xe0" =~ /[^\xc0]/di;). We
1165 * have to do this here before we add the unconditionally matched code
1167 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1168 _invlist_intersection_complement_2nd(invlist,
1173 /* Add in the points from the bit map */
1174 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
1175 if (ANYOF_BITMAP_TEST(node, i)) {
1176 invlist = add_cp_to_invlist(invlist, i);
1177 new_node_has_latin1 = TRUE;
1181 /* If this can match all upper Latin1 code points, have to add them
1183 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
1184 _invlist_union(invlist, PL_UpperLatin1, &invlist);
1187 /* Similarly for these */
1188 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
1189 _invlist_union_complement_2nd(invlist, PL_InBitmap, &invlist);
1192 if (ANYOF_FLAGS(node) & ANYOF_INVERT) {
1193 _invlist_invert(invlist);
1195 else if (new_node_has_latin1 && ANYOF_FLAGS(node) & ANYOF_LOC_FOLD) {
1197 /* Under /li, any 0-255 could fold to any other 0-255, depending on the
1198 * locale. We can skip this if there are no 0-255 at all. */
1199 _invlist_union(invlist, PL_Latin1, &invlist);
1202 /* Similarly add the UTF-8 locale possible matches. These have to be
1203 * deferred until after the non-UTF-8 locale ones are taken care of just
1204 * above, or it leads to wrong results under ANYOF_INVERT */
1205 if (only_utf8_locale_invlist) {
1206 _invlist_union_maybe_complement_2nd(invlist,
1207 only_utf8_locale_invlist,
1208 ANYOF_FLAGS(node) & ANYOF_INVERT,
1215 /* These two functions currently do the exact same thing */
1216 #define ssc_init_zero ssc_init
1218 #define ssc_add_cp(ssc, cp) ssc_add_range((ssc), (cp), (cp))
1219 #define ssc_match_all_cp(ssc) ssc_add_range(ssc, 0, UV_MAX)
1221 /* 'AND' a given class with another one. Can create false positives. 'ssc'
1222 * should not be inverted. 'and_with->flags & ANYOF_MATCHES_POSIXL' should be
1223 * 0 if 'and_with' is a regnode_charclass instead of a regnode_ssc. */
1226 S_ssc_and(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1227 const regnode_charclass *and_with)
1229 /* Accumulate into SSC 'ssc' its 'AND' with 'and_with', which is either
1230 * another SSC or a regular ANYOF class. Can create false positives. */
1235 PERL_ARGS_ASSERT_SSC_AND;
1237 assert(is_ANYOF_SYNTHETIC(ssc));
1239 /* 'and_with' is used as-is if it too is an SSC; otherwise have to extract
1240 * the code point inversion list and just the relevant flags */
1241 if (is_ANYOF_SYNTHETIC(and_with)) {
1242 anded_cp_list = ((regnode_ssc *)and_with)->invlist;
1243 anded_flags = ANYOF_FLAGS(and_with);
1245 /* XXX This is a kludge around what appears to be deficiencies in the
1246 * optimizer. If we make S_ssc_anything() add in the WARN_SUPER flag,
1247 * there are paths through the optimizer where it doesn't get weeded
1248 * out when it should. And if we don't make some extra provision for
1249 * it like the code just below, it doesn't get added when it should.
1250 * This solution is to add it only when AND'ing, which is here, and
1251 * only when what is being AND'ed is the pristine, original node
1252 * matching anything. Thus it is like adding it to ssc_anything() but
1253 * only when the result is to be AND'ed. Probably the same solution
1254 * could be adopted for the same problem we have with /l matching,
1255 * which is solved differently in S_ssc_init(), and that would lead to
1256 * fewer false positives than that solution has. But if this solution
1257 * creates bugs, the consequences are only that a warning isn't raised
1258 * that should be; while the consequences for having /l bugs is
1259 * incorrect matches */
1260 if (ssc_is_anything((regnode_ssc *)and_with)) {
1261 anded_flags |= ANYOF_WARN_SUPER;
1265 anded_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, and_with);
1266 anded_flags = ANYOF_FLAGS(and_with) & ANYOF_COMMON_FLAGS;
1269 ANYOF_FLAGS(ssc) &= anded_flags;
1271 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1272 * C2 is the list of code points in 'and-with'; P2, its posix classes.
1273 * 'and_with' may be inverted. When not inverted, we have the situation of
1275 * (C1 | P1) & (C2 | P2)
1276 * = (C1 & (C2 | P2)) | (P1 & (C2 | P2))
1277 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1278 * <= ((C1 & C2) | P2)) | ( P1 | (P1 & P2))
1279 * <= ((C1 & C2) | P1 | P2)
1280 * Alternatively, the last few steps could be:
1281 * = ((C1 & C2) | (C1 & P2)) | ((P1 & C2) | (P1 & P2))
1282 * <= ((C1 & C2) | C1 ) | ( C2 | (P1 & P2))
1283 * <= (C1 | C2 | (P1 & P2))
1284 * We favor the second approach if either P1 or P2 is non-empty. This is
1285 * because these components are a barrier to doing optimizations, as what
1286 * they match cannot be known until the moment of matching as they are
1287 * dependent on the current locale, 'AND"ing them likely will reduce or
1289 * But we can do better if we know that C1,P1 are in their initial state (a
1290 * frequent occurrence), each matching everything:
1291 * (<everything>) & (C2 | P2) = C2 | P2
1292 * Similarly, if C2,P2 are in their initial state (again a frequent
1293 * occurrence), the result is a no-op
1294 * (C1 | P1) & (<everything>) = C1 | P1
1297 * (C1 | P1) & ~(C2 | P2) = (C1 | P1) & (~C2 & ~P2)
1298 * = (C1 & (~C2 & ~P2)) | (P1 & (~C2 & ~P2))
1299 * <= (C1 & ~C2) | (P1 & ~P2)
1302 if ((ANYOF_FLAGS(and_with) & ANYOF_INVERT)
1303 && ! is_ANYOF_SYNTHETIC(and_with))
1307 ssc_intersection(ssc,
1309 FALSE /* Has already been inverted */
1312 /* If either P1 or P2 is empty, the intersection will be also; can skip
1314 if (! (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL)) {
1315 ANYOF_POSIXL_ZERO(ssc);
1317 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1319 /* Note that the Posix class component P from 'and_with' actually
1321 * P = Pa | Pb | ... | Pn
1322 * where each component is one posix class, such as in [\w\s].
1324 * ~P = ~(Pa | Pb | ... | Pn)
1325 * = ~Pa & ~Pb & ... & ~Pn
1326 * <= ~Pa | ~Pb | ... | ~Pn
1327 * The last is something we can easily calculate, but unfortunately
1328 * is likely to have many false positives. We could do better
1329 * in some (but certainly not all) instances if two classes in
1330 * P have known relationships. For example
1331 * :lower: <= :alpha: <= :alnum: <= \w <= :graph: <= :print:
1333 * :lower: & :print: = :lower:
1334 * And similarly for classes that must be disjoint. For example,
1335 * since \s and \w can have no elements in common based on rules in
1336 * the POSIX standard,
1337 * \w & ^\S = nothing
1338 * Unfortunately, some vendor locales do not meet the Posix
1339 * standard, in particular almost everything by Microsoft.
1340 * The loop below just changes e.g., \w into \W and vice versa */
1342 regnode_charclass_posixl temp;
1343 int add = 1; /* To calculate the index of the complement */
1345 ANYOF_POSIXL_ZERO(&temp);
1346 for (i = 0; i < ANYOF_MAX; i++) {
1348 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)
1349 || ! ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i + 1));
1351 if (ANYOF_POSIXL_TEST((regnode_charclass_posixl*) and_with, i)) {
1352 ANYOF_POSIXL_SET(&temp, i + add);
1354 add = 0 - add; /* 1 goes to -1; -1 goes to 1 */
1356 ANYOF_POSIXL_AND(&temp, ssc);
1358 } /* else ssc already has no posixes */
1359 } /* else: Not inverted. This routine is a no-op if 'and_with' is an SSC
1360 in its initial state */
1361 else if (! is_ANYOF_SYNTHETIC(and_with)
1362 || ! ssc_is_cp_posixl_init(pRExC_state, (regnode_ssc *)and_with))
1364 /* But if 'ssc' is in its initial state, the result is just 'and_with';
1365 * copy it over 'ssc' */
1366 if (ssc_is_cp_posixl_init(pRExC_state, ssc)) {
1367 if (is_ANYOF_SYNTHETIC(and_with)) {
1368 StructCopy(and_with, ssc, regnode_ssc);
1371 ssc->invlist = anded_cp_list;
1372 ANYOF_POSIXL_ZERO(ssc);
1373 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1374 ANYOF_POSIXL_OR((regnode_charclass_posixl*) and_with, ssc);
1378 else if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)
1379 || (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL))
1381 /* One or the other of P1, P2 is non-empty. */
1382 if (ANYOF_FLAGS(and_with) & ANYOF_MATCHES_POSIXL) {
1383 ANYOF_POSIXL_AND((regnode_charclass_posixl*) and_with, ssc);
1385 ssc_union(ssc, anded_cp_list, FALSE);
1387 else { /* P1 = P2 = empty */
1388 ssc_intersection(ssc, anded_cp_list, FALSE);
1394 S_ssc_or(pTHX_ const RExC_state_t *pRExC_state, regnode_ssc *ssc,
1395 const regnode_charclass *or_with)
1397 /* Accumulate into SSC 'ssc' its 'OR' with 'or_with', which is either
1398 * another SSC or a regular ANYOF class. Can create false positives if
1399 * 'or_with' is to be inverted. */
1404 PERL_ARGS_ASSERT_SSC_OR;
1406 assert(is_ANYOF_SYNTHETIC(ssc));
1408 /* 'or_with' is used as-is if it too is an SSC; otherwise have to extract
1409 * the code point inversion list and just the relevant flags */
1410 if (is_ANYOF_SYNTHETIC(or_with)) {
1411 ored_cp_list = ((regnode_ssc*) or_with)->invlist;
1412 ored_flags = ANYOF_FLAGS(or_with);
1415 ored_cp_list = get_ANYOF_cp_list_for_ssc(pRExC_state, or_with);
1416 ored_flags = ANYOF_FLAGS(or_with) & ANYOF_COMMON_FLAGS;
1419 ANYOF_FLAGS(ssc) |= ored_flags;
1421 /* Below, C1 is the list of code points in 'ssc'; P1, its posix classes.
1422 * C2 is the list of code points in 'or-with'; P2, its posix classes.
1423 * 'or_with' may be inverted. When not inverted, we have the simple
1424 * situation of computing:
1425 * (C1 | P1) | (C2 | P2) = (C1 | C2) | (P1 | P2)
1426 * If P1|P2 yields a situation with both a class and its complement are
1427 * set, like having both \w and \W, this matches all code points, and we
1428 * can delete these from the P component of the ssc going forward. XXX We
1429 * might be able to delete all the P components, but I (khw) am not certain
1430 * about this, and it is better to be safe.
1433 * (C1 | P1) | ~(C2 | P2) = (C1 | P1) | (~C2 & ~P2)
1434 * <= (C1 | P1) | ~C2
1435 * <= (C1 | ~C2) | P1
1436 * (which results in actually simpler code than the non-inverted case)
1439 if ((ANYOF_FLAGS(or_with) & ANYOF_INVERT)
1440 && ! is_ANYOF_SYNTHETIC(or_with))
1442 /* We ignore P2, leaving P1 going forward */
1443 } /* else Not inverted */
1444 else if (ANYOF_FLAGS(or_with) & ANYOF_MATCHES_POSIXL) {
1445 ANYOF_POSIXL_OR((regnode_charclass_posixl*)or_with, ssc);
1446 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1448 for (i = 0; i < ANYOF_MAX; i += 2) {
1449 if (ANYOF_POSIXL_TEST(ssc, i) && ANYOF_POSIXL_TEST(ssc, i + 1))
1451 ssc_match_all_cp(ssc);
1452 ANYOF_POSIXL_CLEAR(ssc, i);
1453 ANYOF_POSIXL_CLEAR(ssc, i+1);
1461 FALSE /* Already has been inverted */
1465 PERL_STATIC_INLINE void
1466 S_ssc_union(pTHX_ regnode_ssc *ssc, SV* const invlist, const bool invert2nd)
1468 PERL_ARGS_ASSERT_SSC_UNION;
1470 assert(is_ANYOF_SYNTHETIC(ssc));
1472 _invlist_union_maybe_complement_2nd(ssc->invlist,
1478 PERL_STATIC_INLINE void
1479 S_ssc_intersection(pTHX_ regnode_ssc *ssc,
1481 const bool invert2nd)
1483 PERL_ARGS_ASSERT_SSC_INTERSECTION;
1485 assert(is_ANYOF_SYNTHETIC(ssc));
1487 _invlist_intersection_maybe_complement_2nd(ssc->invlist,
1493 PERL_STATIC_INLINE void
1494 S_ssc_add_range(pTHX_ regnode_ssc *ssc, const UV start, const UV end)
1496 PERL_ARGS_ASSERT_SSC_ADD_RANGE;
1498 assert(is_ANYOF_SYNTHETIC(ssc));
1500 ssc->invlist = _add_range_to_invlist(ssc->invlist, start, end);
1503 PERL_STATIC_INLINE void
1504 S_ssc_cp_and(pTHX_ regnode_ssc *ssc, const UV cp)
1506 /* AND just the single code point 'cp' into the SSC 'ssc' */
1508 SV* cp_list = _new_invlist(2);
1510 PERL_ARGS_ASSERT_SSC_CP_AND;
1512 assert(is_ANYOF_SYNTHETIC(ssc));
1514 cp_list = add_cp_to_invlist(cp_list, cp);
1515 ssc_intersection(ssc, cp_list,
1516 FALSE /* Not inverted */
1518 SvREFCNT_dec_NN(cp_list);
1521 PERL_STATIC_INLINE void
1522 S_ssc_clear_locale(regnode_ssc *ssc)
1524 /* Set the SSC 'ssc' to not match any locale things */
1525 PERL_ARGS_ASSERT_SSC_CLEAR_LOCALE;
1527 assert(is_ANYOF_SYNTHETIC(ssc));
1529 ANYOF_POSIXL_ZERO(ssc);
1530 ANYOF_FLAGS(ssc) &= ~ANYOF_LOCALE_FLAGS;
1533 #define NON_OTHER_COUNT NON_OTHER_COUNT_FOR_USE_ONLY_BY_REGCOMP_DOT_C
1536 S_is_ssc_worth_it(const RExC_state_t * pRExC_state, const regnode_ssc * ssc)
1538 /* The synthetic start class is used to hopefully quickly winnow down
1539 * places where a pattern could start a match in the target string. If it
1540 * doesn't really narrow things down that much, there isn't much point to
1541 * having the overhead of using it. This function uses some very crude
1542 * heuristics to decide if to use the ssc or not.
1544 * It returns TRUE if 'ssc' rules out more than half what it considers to
1545 * be the "likely" possible matches, but of course it doesn't know what the
1546 * actual things being matched are going to be; these are only guesses
1548 * For /l matches, it assumes that the only likely matches are going to be
1549 * in the 0-255 range, uniformly distributed, so half of that is 127
1550 * For /a and /d matches, it assumes that the likely matches will be just
1551 * the ASCII range, so half of that is 63
1552 * For /u and there isn't anything matching above the Latin1 range, it
1553 * assumes that that is the only range likely to be matched, and uses
1554 * half that as the cut-off: 127. If anything matches above Latin1,
1555 * it assumes that all of Unicode could match (uniformly), except for
1556 * non-Unicode code points and things in the General Category "Other"
1557 * (unassigned, private use, surrogates, controls and formats). This
1558 * is a much large number. */
1560 const U32 max_match = (LOC)
1564 : (invlist_highest(ssc->invlist) < 256)
1566 : ((NON_OTHER_COUNT + 1) / 2) - 1;
1567 U32 count = 0; /* Running total of number of code points matched by
1569 UV start, end; /* Start and end points of current range in inversion
1572 PERL_ARGS_ASSERT_IS_SSC_WORTH_IT;
1574 invlist_iterinit(ssc->invlist);
1575 while (invlist_iternext(ssc->invlist, &start, &end)) {
1577 /* /u is the only thing that we expect to match above 255; so if not /u
1578 * and even if there are matches above 255, ignore them. This catches
1579 * things like \d under /d which does match the digits above 255, but
1580 * since the pattern is /d, it is not likely to be expecting them */
1581 if (! UNI_SEMANTICS) {
1585 end = MIN(end, 255);
1587 count += end - start + 1;
1588 if (count > max_match) {
1589 invlist_iterfinish(ssc->invlist);
1599 S_ssc_finalize(pTHX_ RExC_state_t *pRExC_state, regnode_ssc *ssc)
1601 /* The inversion list in the SSC is marked mortal; now we need a more
1602 * permanent copy, which is stored the same way that is done in a regular
1603 * ANYOF node, with the first NUM_ANYOF_CODE_POINTS code points in a bit
1606 SV* invlist = invlist_clone(ssc->invlist);
1608 PERL_ARGS_ASSERT_SSC_FINALIZE;
1610 assert(is_ANYOF_SYNTHETIC(ssc));
1612 /* The code in this file assumes that all but these flags aren't relevant
1613 * to the SSC, except SSC_MATCHES_EMPTY_STRING, which should be cleared
1614 * by the time we reach here */
1615 assert(! (ANYOF_FLAGS(ssc) & ~ANYOF_COMMON_FLAGS));
1617 populate_ANYOF_from_invlist( (regnode *) ssc, &invlist);
1619 set_ANYOF_arg(pRExC_state, (regnode *) ssc, invlist,
1620 NULL, NULL, NULL, FALSE);
1622 /* Make sure is clone-safe */
1623 ssc->invlist = NULL;
1625 if (ANYOF_POSIXL_SSC_TEST_ANY_SET(ssc)) {
1626 ANYOF_FLAGS(ssc) |= ANYOF_MATCHES_POSIXL;
1629 assert(! (ANYOF_FLAGS(ssc) & ANYOF_LOCALE_FLAGS) || RExC_contains_locale);
1632 #define TRIE_LIST_ITEM(state,idx) (trie->states[state].trans.list)[ idx ]
1633 #define TRIE_LIST_CUR(state) ( TRIE_LIST_ITEM( state, 0 ).forid )
1634 #define TRIE_LIST_LEN(state) ( TRIE_LIST_ITEM( state, 0 ).newstate )
1635 #define TRIE_LIST_USED(idx) ( trie->states[state].trans.list \
1636 ? (TRIE_LIST_CUR( idx ) - 1) \
1642 dump_trie(trie,widecharmap,revcharmap)
1643 dump_trie_interim_list(trie,widecharmap,revcharmap,next_alloc)
1644 dump_trie_interim_table(trie,widecharmap,revcharmap,next_alloc)
1646 These routines dump out a trie in a somewhat readable format.
1647 The _interim_ variants are used for debugging the interim
1648 tables that are used to generate the final compressed
1649 representation which is what dump_trie expects.
1651 Part of the reason for their existence is to provide a form
1652 of documentation as to how the different representations function.
1657 Dumps the final compressed table form of the trie to Perl_debug_log.
1658 Used for debugging make_trie().
1662 S_dump_trie(pTHX_ const struct _reg_trie_data *trie, HV *widecharmap,
1663 AV *revcharmap, U32 depth)
1666 SV *sv=sv_newmortal();
1667 int colwidth= widecharmap ? 6 : 4;
1669 GET_RE_DEBUG_FLAGS_DECL;
1671 PERL_ARGS_ASSERT_DUMP_TRIE;
1673 PerlIO_printf( Perl_debug_log, "%*sChar : %-6s%-6s%-4s ",
1674 (int)depth * 2 + 2,"",
1675 "Match","Base","Ofs" );
1677 for( state = 0 ; state < trie->uniquecharcount ; state++ ) {
1678 SV ** const tmp = av_fetch( revcharmap, state, 0);
1680 PerlIO_printf( Perl_debug_log, "%*s",
1682 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1683 PL_colors[0], PL_colors[1],
1684 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1685 PERL_PV_ESCAPE_FIRSTCHAR
1690 PerlIO_printf( Perl_debug_log, "\n%*sState|-----------------------",
1691 (int)depth * 2 + 2,"");
1693 for( state = 0 ; state < trie->uniquecharcount ; state++ )
1694 PerlIO_printf( Perl_debug_log, "%.*s", colwidth, "--------");
1695 PerlIO_printf( Perl_debug_log, "\n");
1697 for( state = 1 ; state < trie->statecount ; state++ ) {
1698 const U32 base = trie->states[ state ].trans.base;
1700 PerlIO_printf( Perl_debug_log, "%*s#%4"UVXf"|",
1701 (int)depth * 2 + 2,"", (UV)state);
1703 if ( trie->states[ state ].wordnum ) {
1704 PerlIO_printf( Perl_debug_log, " W%4X",
1705 trie->states[ state ].wordnum );
1707 PerlIO_printf( Perl_debug_log, "%6s", "" );
1710 PerlIO_printf( Perl_debug_log, " @%4"UVXf" ", (UV)base );
1715 while( ( base + ofs < trie->uniquecharcount ) ||
1716 ( base + ofs - trie->uniquecharcount < trie->lasttrans
1717 && trie->trans[ base + ofs - trie->uniquecharcount ].check
1721 PerlIO_printf( Perl_debug_log, "+%2"UVXf"[ ", (UV)ofs);
1723 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
1724 if ( ( base + ofs >= trie->uniquecharcount )
1725 && ( base + ofs - trie->uniquecharcount
1727 && trie->trans[ base + ofs
1728 - trie->uniquecharcount ].check == state )
1730 PerlIO_printf( Perl_debug_log, "%*"UVXf,
1732 (UV)trie->trans[ base + ofs
1733 - trie->uniquecharcount ].next );
1735 PerlIO_printf( Perl_debug_log, "%*s",colwidth," ." );
1739 PerlIO_printf( Perl_debug_log, "]");
1742 PerlIO_printf( Perl_debug_log, "\n" );
1744 PerlIO_printf(Perl_debug_log, "%*sword_info N:(prev,len)=",
1746 for (word=1; word <= trie->wordcount; word++) {
1747 PerlIO_printf(Perl_debug_log, " %d:(%d,%d)",
1748 (int)word, (int)(trie->wordinfo[word].prev),
1749 (int)(trie->wordinfo[word].len));
1751 PerlIO_printf(Perl_debug_log, "\n" );
1754 Dumps a fully constructed but uncompressed trie in list form.
1755 List tries normally only are used for construction when the number of
1756 possible chars (trie->uniquecharcount) is very high.
1757 Used for debugging make_trie().
1760 S_dump_trie_interim_list(pTHX_ const struct _reg_trie_data *trie,
1761 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1765 SV *sv=sv_newmortal();
1766 int colwidth= widecharmap ? 6 : 4;
1767 GET_RE_DEBUG_FLAGS_DECL;
1769 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_LIST;
1771 /* print out the table precompression. */
1772 PerlIO_printf( Perl_debug_log, "%*sState :Word | Transition Data\n%*s%s",
1773 (int)depth * 2 + 2,"", (int)depth * 2 + 2,"",
1774 "------:-----+-----------------\n" );
1776 for( state=1 ; state < next_alloc ; state ++ ) {
1779 PerlIO_printf( Perl_debug_log, "%*s %4"UVXf" :",
1780 (int)depth * 2 + 2,"", (UV)state );
1781 if ( ! trie->states[ state ].wordnum ) {
1782 PerlIO_printf( Perl_debug_log, "%5s| ","");
1784 PerlIO_printf( Perl_debug_log, "W%4x| ",
1785 trie->states[ state ].wordnum
1788 for( charid = 1 ; charid <= TRIE_LIST_USED( state ) ; charid++ ) {
1789 SV ** const tmp = av_fetch( revcharmap,
1790 TRIE_LIST_ITEM(state,charid).forid, 0);
1792 PerlIO_printf( Perl_debug_log, "%*s:%3X=%4"UVXf" | ",
1794 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp),
1796 PL_colors[0], PL_colors[1],
1797 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0)
1798 | PERL_PV_ESCAPE_FIRSTCHAR
1800 TRIE_LIST_ITEM(state,charid).forid,
1801 (UV)TRIE_LIST_ITEM(state,charid).newstate
1804 PerlIO_printf(Perl_debug_log, "\n%*s| ",
1805 (int)((depth * 2) + 14), "");
1808 PerlIO_printf( Perl_debug_log, "\n");
1813 Dumps a fully constructed but uncompressed trie in table form.
1814 This is the normal DFA style state transition table, with a few
1815 twists to facilitate compression later.
1816 Used for debugging make_trie().
1819 S_dump_trie_interim_table(pTHX_ const struct _reg_trie_data *trie,
1820 HV *widecharmap, AV *revcharmap, U32 next_alloc,
1825 SV *sv=sv_newmortal();
1826 int colwidth= widecharmap ? 6 : 4;
1827 GET_RE_DEBUG_FLAGS_DECL;
1829 PERL_ARGS_ASSERT_DUMP_TRIE_INTERIM_TABLE;
1832 print out the table precompression so that we can do a visual check
1833 that they are identical.
1836 PerlIO_printf( Perl_debug_log, "%*sChar : ",(int)depth * 2 + 2,"" );
1838 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1839 SV ** const tmp = av_fetch( revcharmap, charid, 0);
1841 PerlIO_printf( Perl_debug_log, "%*s",
1843 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), colwidth,
1844 PL_colors[0], PL_colors[1],
1845 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
1846 PERL_PV_ESCAPE_FIRSTCHAR
1852 PerlIO_printf( Perl_debug_log, "\n%*sState+-",(int)depth * 2 + 2,"" );
1854 for( charid=0 ; charid < trie->uniquecharcount ; charid++ ) {
1855 PerlIO_printf( Perl_debug_log, "%.*s", colwidth,"--------");
1858 PerlIO_printf( Perl_debug_log, "\n" );
1860 for( state=1 ; state < next_alloc ; state += trie->uniquecharcount ) {
1862 PerlIO_printf( Perl_debug_log, "%*s%4"UVXf" : ",
1863 (int)depth * 2 + 2,"",
1864 (UV)TRIE_NODENUM( state ) );
1866 for( charid = 0 ; charid < trie->uniquecharcount ; charid++ ) {
1867 UV v=(UV)SAFE_TRIE_NODENUM( trie->trans[ state + charid ].next );
1869 PerlIO_printf( Perl_debug_log, "%*"UVXf, colwidth, v );
1871 PerlIO_printf( Perl_debug_log, "%*s", colwidth, "." );
1873 if ( ! trie->states[ TRIE_NODENUM( state ) ].wordnum ) {
1874 PerlIO_printf( Perl_debug_log, " (%4"UVXf")\n",
1875 (UV)trie->trans[ state ].check );
1877 PerlIO_printf( Perl_debug_log, " (%4"UVXf") W%4X\n",
1878 (UV)trie->trans[ state ].check,
1879 trie->states[ TRIE_NODENUM( state ) ].wordnum );
1887 /* make_trie(startbranch,first,last,tail,word_count,flags,depth)
1888 startbranch: the first branch in the whole branch sequence
1889 first : start branch of sequence of branch-exact nodes.
1890 May be the same as startbranch
1891 last : Thing following the last branch.
1892 May be the same as tail.
1893 tail : item following the branch sequence
1894 count : words in the sequence
1895 flags : currently the OP() type we will be building one of /EXACT(|F|FA|FU|FU_SS|L|FLU8)/
1896 depth : indent depth
1898 Inplace optimizes a sequence of 2 or more Branch-Exact nodes into a TRIE node.
1900 A trie is an N'ary tree where the branches are determined by digital
1901 decomposition of the key. IE, at the root node you look up the 1st character and
1902 follow that branch repeat until you find the end of the branches. Nodes can be
1903 marked as "accepting" meaning they represent a complete word. Eg:
1907 would convert into the following structure. Numbers represent states, letters
1908 following numbers represent valid transitions on the letter from that state, if
1909 the number is in square brackets it represents an accepting state, otherwise it
1910 will be in parenthesis.
1912 +-h->+-e->[3]-+-r->(8)-+-s->[9]
1916 (1) +-i->(6)-+-s->[7]
1918 +-s->(3)-+-h->(4)-+-e->[5]
1920 Accept Word Mapping: 3=>1 (he),5=>2 (she), 7=>3 (his), 9=>4 (hers)
1922 This shows that when matching against the string 'hers' we will begin at state 1
1923 read 'h' and move to state 2, read 'e' and move to state 3 which is accepting,
1924 then read 'r' and go to state 8 followed by 's' which takes us to state 9 which
1925 is also accepting. Thus we know that we can match both 'he' and 'hers' with a
1926 single traverse. We store a mapping from accepting to state to which word was
1927 matched, and then when we have multiple possibilities we try to complete the
1928 rest of the regex in the order in which they occurred in the alternation.
1930 The only prior NFA like behaviour that would be changed by the TRIE support is
1931 the silent ignoring of duplicate alternations which are of the form:
1933 / (DUPE|DUPE) X? (?{ ... }) Y /x
1935 Thus EVAL blocks following a trie may be called a different number of times with
1936 and without the optimisation. With the optimisations dupes will be silently
1937 ignored. This inconsistent behaviour of EVAL type nodes is well established as
1938 the following demonstrates:
1940 'words'=~/(word|word|word)(?{ print $1 })[xyz]/
1942 which prints out 'word' three times, but
1944 'words'=~/(word|word|word)(?{ print $1 })S/
1946 which doesnt print it out at all. This is due to other optimisations kicking in.
1948 Example of what happens on a structural level:
1950 The regexp /(ac|ad|ab)+/ will produce the following debug output:
1952 1: CURLYM[1] {1,32767}(18)
1963 This would be optimizable with startbranch=5, first=5, last=16, tail=16
1964 and should turn into:
1966 1: CURLYM[1] {1,32767}(18)
1968 [Words:3 Chars Stored:6 Unique Chars:4 States:5 NCP:1]
1976 Cases where tail != last would be like /(?foo|bar)baz/:
1986 which would be optimizable with startbranch=1, first=1, last=7, tail=8
1987 and would end up looking like:
1990 [Words:2 Chars Stored:6 Unique Chars:5 States:7 NCP:1]
1997 d = uvchr_to_utf8_flags(d, uv, 0);
1999 is the recommended Unicode-aware way of saying
2004 #define TRIE_STORE_REVCHAR(val) \
2007 SV *zlopp = newSV(7); /* XXX: optimize me */ \
2008 unsigned char *flrbbbbb = (unsigned char *) SvPVX(zlopp); \
2009 unsigned const char *const kapow = uvchr_to_utf8(flrbbbbb, val); \
2010 SvCUR_set(zlopp, kapow - flrbbbbb); \
2013 av_push(revcharmap, zlopp); \
2015 char ooooff = (char)val; \
2016 av_push(revcharmap, newSVpvn(&ooooff, 1)); \
2020 /* This gets the next character from the input, folding it if not already
2022 #define TRIE_READ_CHAR STMT_START { \
2025 /* if it is UTF then it is either already folded, or does not need \
2027 uvc = valid_utf8_to_uvchr( (const U8*) uc, &len); \
2029 else if (folder == PL_fold_latin1) { \
2030 /* This folder implies Unicode rules, which in the range expressible \
2031 * by not UTF is the lower case, with the two exceptions, one of \
2032 * which should have been taken care of before calling this */ \
2033 assert(*uc != LATIN_SMALL_LETTER_SHARP_S); \
2034 uvc = toLOWER_L1(*uc); \
2035 if (UNLIKELY(uvc == MICRO_SIGN)) uvc = GREEK_SMALL_LETTER_MU; \
2038 /* raw data, will be folded later if needed */ \
2046 #define TRIE_LIST_PUSH(state,fid,ns) STMT_START { \
2047 if ( TRIE_LIST_CUR( state ) >=TRIE_LIST_LEN( state ) ) { \
2048 U32 ging = TRIE_LIST_LEN( state ) *= 2; \
2049 Renew( trie->states[ state ].trans.list, ging, reg_trie_trans_le ); \
2051 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).forid = fid; \
2052 TRIE_LIST_ITEM( state, TRIE_LIST_CUR( state ) ).newstate = ns; \
2053 TRIE_LIST_CUR( state )++; \
2056 #define TRIE_LIST_NEW(state) STMT_START { \
2057 Newxz( trie->states[ state ].trans.list, \
2058 4, reg_trie_trans_le ); \
2059 TRIE_LIST_CUR( state ) = 1; \
2060 TRIE_LIST_LEN( state ) = 4; \
2063 #define TRIE_HANDLE_WORD(state) STMT_START { \
2064 U16 dupe= trie->states[ state ].wordnum; \
2065 regnode * const noper_next = regnext( noper ); \
2068 /* store the word for dumping */ \
2070 if (OP(noper) != NOTHING) \
2071 tmp = newSVpvn_utf8(STRING(noper), STR_LEN(noper), UTF); \
2073 tmp = newSVpvn_utf8( "", 0, UTF ); \
2074 av_push( trie_words, tmp ); \
2078 trie->wordinfo[curword].prev = 0; \
2079 trie->wordinfo[curword].len = wordlen; \
2080 trie->wordinfo[curword].accept = state; \
2082 if ( noper_next < tail ) { \
2084 trie->jump = (U16 *) PerlMemShared_calloc( word_count + 1, \
2086 trie->jump[curword] = (U16)(noper_next - convert); \
2088 jumper = noper_next; \
2090 nextbranch= regnext(cur); \
2094 /* It's a dupe. Pre-insert into the wordinfo[].prev */\
2095 /* chain, so that when the bits of chain are later */\
2096 /* linked together, the dups appear in the chain */\
2097 trie->wordinfo[curword].prev = trie->wordinfo[dupe].prev; \
2098 trie->wordinfo[dupe].prev = curword; \
2100 /* we haven't inserted this word yet. */ \
2101 trie->states[ state ].wordnum = curword; \
2106 #define TRIE_TRANS_STATE(state,base,ucharcount,charid,special) \
2107 ( ( base + charid >= ucharcount \
2108 && base + charid < ubound \
2109 && state == trie->trans[ base - ucharcount + charid ].check \
2110 && trie->trans[ base - ucharcount + charid ].next ) \
2111 ? trie->trans[ base - ucharcount + charid ].next \
2112 : ( state==1 ? special : 0 ) \
2116 #define MADE_JUMP_TRIE 2
2117 #define MADE_EXACT_TRIE 4
2120 S_make_trie(pTHX_ RExC_state_t *pRExC_state, regnode *startbranch,
2121 regnode *first, regnode *last, regnode *tail,
2122 U32 word_count, U32 flags, U32 depth)
2124 /* first pass, loop through and scan words */
2125 reg_trie_data *trie;
2126 HV *widecharmap = NULL;
2127 AV *revcharmap = newAV();
2133 regnode *jumper = NULL;
2134 regnode *nextbranch = NULL;
2135 regnode *convert = NULL;
2136 U32 *prev_states; /* temp array mapping each state to previous one */
2137 /* we just use folder as a flag in utf8 */
2138 const U8 * folder = NULL;
2141 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tuuu"));
2142 AV *trie_words = NULL;
2143 /* along with revcharmap, this only used during construction but both are
2144 * useful during debugging so we store them in the struct when debugging.
2147 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("tu"));
2148 STRLEN trie_charcount=0;
2150 SV *re_trie_maxbuff;
2151 GET_RE_DEBUG_FLAGS_DECL;
2153 PERL_ARGS_ASSERT_MAKE_TRIE;
2155 PERL_UNUSED_ARG(depth);
2159 case EXACT: case EXACTL: break;
2163 case EXACTFLU8: folder = PL_fold_latin1; break;
2164 case EXACTF: folder = PL_fold; break;
2165 default: Perl_croak( aTHX_ "panic! In trie construction, unknown node type %u %s", (unsigned) flags, PL_reg_name[flags] );
2168 trie = (reg_trie_data *) PerlMemShared_calloc( 1, sizeof(reg_trie_data) );
2170 trie->startstate = 1;
2171 trie->wordcount = word_count;
2172 RExC_rxi->data->data[ data_slot ] = (void*)trie;
2173 trie->charmap = (U16 *) PerlMemShared_calloc( 256, sizeof(U16) );
2174 if (flags == EXACT || flags == EXACTL)
2175 trie->bitmap = (char *) PerlMemShared_calloc( ANYOF_BITMAP_SIZE, 1 );
2176 trie->wordinfo = (reg_trie_wordinfo *) PerlMemShared_calloc(
2177 trie->wordcount+1, sizeof(reg_trie_wordinfo));
2180 trie_words = newAV();
2183 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
2184 assert(re_trie_maxbuff);
2185 if (!SvIOK(re_trie_maxbuff)) {
2186 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
2188 DEBUG_TRIE_COMPILE_r({
2189 PerlIO_printf( Perl_debug_log,
2190 "%*smake_trie start==%d, first==%d, last==%d, tail==%d depth=%d\n",
2191 (int)depth * 2 + 2, "",
2192 REG_NODE_NUM(startbranch),REG_NODE_NUM(first),
2193 REG_NODE_NUM(last), REG_NODE_NUM(tail), (int)depth);
2196 /* Find the node we are going to overwrite */
2197 if ( first == startbranch && OP( last ) != BRANCH ) {
2198 /* whole branch chain */
2201 /* branch sub-chain */
2202 convert = NEXTOPER( first );
2205 /* -- First loop and Setup --
2207 We first traverse the branches and scan each word to determine if it
2208 contains widechars, and how many unique chars there are, this is
2209 important as we have to build a table with at least as many columns as we
2212 We use an array of integers to represent the character codes 0..255
2213 (trie->charmap) and we use a an HV* to store Unicode characters. We use
2214 the native representation of the character value as the key and IV's for
2217 *TODO* If we keep track of how many times each character is used we can
2218 remap the columns so that the table compression later on is more
2219 efficient in terms of memory by ensuring the most common value is in the
2220 middle and the least common are on the outside. IMO this would be better
2221 than a most to least common mapping as theres a decent chance the most
2222 common letter will share a node with the least common, meaning the node
2223 will not be compressible. With a middle is most common approach the worst
2224 case is when we have the least common nodes twice.
2228 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2229 regnode *noper = NEXTOPER( cur );
2230 const U8 *uc = (U8*)STRING( noper );
2231 const U8 *e = uc + STR_LEN( noper );
2233 U32 wordlen = 0; /* required init */
2234 STRLEN minchars = 0;
2235 STRLEN maxchars = 0;
2236 bool set_bit = trie->bitmap ? 1 : 0; /*store the first char in the
2239 if (OP(noper) == NOTHING) {
2240 regnode *noper_next= regnext(noper);
2241 if (noper_next != tail && OP(noper_next) == flags) {
2243 uc= (U8*)STRING(noper);
2244 e= uc + STR_LEN(noper);
2245 trie->minlen= STR_LEN(noper);
2252 if ( set_bit ) { /* bitmap only alloced when !(UTF&&Folding) */
2253 TRIE_BITMAP_SET(trie,*uc); /* store the raw first byte
2254 regardless of encoding */
2255 if (OP( noper ) == EXACTFU_SS) {
2256 /* false positives are ok, so just set this */
2257 TRIE_BITMAP_SET(trie, LATIN_SMALL_LETTER_SHARP_S);
2260 for ( ; uc < e ; uc += len ) { /* Look at each char in the current
2262 TRIE_CHARCOUNT(trie)++;
2265 /* TRIE_READ_CHAR returns the current character, or its fold if /i
2266 * is in effect. Under /i, this character can match itself, or
2267 * anything that folds to it. If not under /i, it can match just
2268 * itself. Most folds are 1-1, for example k, K, and KELVIN SIGN
2269 * all fold to k, and all are single characters. But some folds
2270 * expand to more than one character, so for example LATIN SMALL
2271 * LIGATURE FFI folds to the three character sequence 'ffi'. If
2272 * the string beginning at 'uc' is 'ffi', it could be matched by
2273 * three characters, or just by the one ligature character. (It
2274 * could also be matched by two characters: LATIN SMALL LIGATURE FF
2275 * followed by 'i', or by 'f' followed by LATIN SMALL LIGATURE FI).
2276 * (Of course 'I' and/or 'F' instead of 'i' and 'f' can also
2277 * match.) The trie needs to know the minimum and maximum number
2278 * of characters that could match so that it can use size alone to
2279 * quickly reject many match attempts. The max is simple: it is
2280 * the number of folded characters in this branch (since a fold is
2281 * never shorter than what folds to it. */
2285 /* And the min is equal to the max if not under /i (indicated by
2286 * 'folder' being NULL), or there are no multi-character folds. If
2287 * there is a multi-character fold, the min is incremented just
2288 * once, for the character that folds to the sequence. Each
2289 * character in the sequence needs to be added to the list below of
2290 * characters in the trie, but we count only the first towards the
2291 * min number of characters needed. This is done through the
2292 * variable 'foldlen', which is returned by the macros that look
2293 * for these sequences as the number of bytes the sequence
2294 * occupies. Each time through the loop, we decrement 'foldlen' by
2295 * how many bytes the current char occupies. Only when it reaches
2296 * 0 do we increment 'minchars' or look for another multi-character
2298 if (folder == NULL) {
2301 else if (foldlen > 0) {
2302 foldlen -= (UTF) ? UTF8SKIP(uc) : 1;
2307 /* See if *uc is the beginning of a multi-character fold. If
2308 * so, we decrement the length remaining to look at, to account
2309 * for the current character this iteration. (We can use 'uc'
2310 * instead of the fold returned by TRIE_READ_CHAR because for
2311 * non-UTF, the latin1_safe macro is smart enough to account
2312 * for all the unfolded characters, and because for UTF, the
2313 * string will already have been folded earlier in the
2314 * compilation process */
2316 if ((foldlen = is_MULTI_CHAR_FOLD_utf8_safe(uc, e))) {
2317 foldlen -= UTF8SKIP(uc);
2320 else if ((foldlen = is_MULTI_CHAR_FOLD_latin1_safe(uc, e))) {
2325 /* The current character (and any potential folds) should be added
2326 * to the possible matching characters for this position in this
2330 U8 folded= folder[ (U8) uvc ];
2331 if ( !trie->charmap[ folded ] ) {
2332 trie->charmap[ folded ]=( ++trie->uniquecharcount );
2333 TRIE_STORE_REVCHAR( folded );
2336 if ( !trie->charmap[ uvc ] ) {
2337 trie->charmap[ uvc ]=( ++trie->uniquecharcount );
2338 TRIE_STORE_REVCHAR( uvc );
2341 /* store the codepoint in the bitmap, and its folded
2343 TRIE_BITMAP_SET(trie, uvc);
2345 /* store the folded codepoint */
2346 if ( folder ) TRIE_BITMAP_SET(trie, folder[(U8) uvc ]);
2349 /* store first byte of utf8 representation of
2350 variant codepoints */
2351 if (! UVCHR_IS_INVARIANT(uvc)) {
2352 TRIE_BITMAP_SET(trie, UTF8_TWO_BYTE_HI(uvc));
2355 set_bit = 0; /* We've done our bit :-) */
2359 /* XXX We could come up with the list of code points that fold
2360 * to this using PL_utf8_foldclosures, except not for
2361 * multi-char folds, as there may be multiple combinations
2362 * there that could work, which needs to wait until runtime to
2363 * resolve (The comment about LIGATURE FFI above is such an
2368 widecharmap = newHV();
2370 svpp = hv_fetch( widecharmap, (char*)&uvc, sizeof( UV ), 1 );
2373 Perl_croak( aTHX_ "error creating/fetching widecharmap entry for 0x%"UVXf, uvc );
2375 if ( !SvTRUE( *svpp ) ) {
2376 sv_setiv( *svpp, ++trie->uniquecharcount );
2377 TRIE_STORE_REVCHAR(uvc);
2380 } /* end loop through characters in this branch of the trie */
2382 /* We take the min and max for this branch and combine to find the min
2383 * and max for all branches processed so far */
2384 if( cur == first ) {
2385 trie->minlen = minchars;
2386 trie->maxlen = maxchars;
2387 } else if (minchars < trie->minlen) {
2388 trie->minlen = minchars;
2389 } else if (maxchars > trie->maxlen) {
2390 trie->maxlen = maxchars;
2392 } /* end first pass */
2393 DEBUG_TRIE_COMPILE_r(
2394 PerlIO_printf( Perl_debug_log,
2395 "%*sTRIE(%s): W:%d C:%d Uq:%d Min:%d Max:%d\n",
2396 (int)depth * 2 + 2,"",
2397 ( widecharmap ? "UTF8" : "NATIVE" ), (int)word_count,
2398 (int)TRIE_CHARCOUNT(trie), trie->uniquecharcount,
2399 (int)trie->minlen, (int)trie->maxlen )
2403 We now know what we are dealing with in terms of unique chars and
2404 string sizes so we can calculate how much memory a naive
2405 representation using a flat table will take. If it's over a reasonable
2406 limit (as specified by ${^RE_TRIE_MAXBUF}) we use a more memory
2407 conservative but potentially much slower representation using an array
2410 At the end we convert both representations into the same compressed
2411 form that will be used in regexec.c for matching with. The latter
2412 is a form that cannot be used to construct with but has memory
2413 properties similar to the list form and access properties similar
2414 to the table form making it both suitable for fast searches and
2415 small enough that its feasable to store for the duration of a program.
2417 See the comment in the code where the compressed table is produced
2418 inplace from the flat tabe representation for an explanation of how
2419 the compression works.
2424 Newx(prev_states, TRIE_CHARCOUNT(trie) + 2, U32);
2427 if ( (IV)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount + 1)
2428 > SvIV(re_trie_maxbuff) )
2431 Second Pass -- Array Of Lists Representation
2433 Each state will be represented by a list of charid:state records
2434 (reg_trie_trans_le) the first such element holds the CUR and LEN
2435 points of the allocated array. (See defines above).
2437 We build the initial structure using the lists, and then convert
2438 it into the compressed table form which allows faster lookups
2439 (but cant be modified once converted).
2442 STRLEN transcount = 1;
2444 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2445 "%*sCompiling trie using list compiler\n",
2446 (int)depth * 2 + 2, ""));
2448 trie->states = (reg_trie_state *)
2449 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2450 sizeof(reg_trie_state) );
2454 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2456 regnode *noper = NEXTOPER( cur );
2457 U8 *uc = (U8*)STRING( noper );
2458 const U8 *e = uc + STR_LEN( noper );
2459 U32 state = 1; /* required init */
2460 U16 charid = 0; /* sanity init */
2461 U32 wordlen = 0; /* required init */
2463 if (OP(noper) == NOTHING) {
2464 regnode *noper_next= regnext(noper);
2465 if (noper_next != tail && OP(noper_next) == flags) {
2467 uc= (U8*)STRING(noper);
2468 e= uc + STR_LEN(noper);
2472 if (OP(noper) != NOTHING) {
2473 for ( ; uc < e ; uc += len ) {
2478 charid = trie->charmap[ uvc ];
2480 SV** const svpp = hv_fetch( widecharmap,
2487 charid=(U16)SvIV( *svpp );
2490 /* charid is now 0 if we dont know the char read, or
2491 * nonzero if we do */
2498 if ( !trie->states[ state ].trans.list ) {
2499 TRIE_LIST_NEW( state );
2502 check <= TRIE_LIST_USED( state );
2505 if ( TRIE_LIST_ITEM( state, check ).forid
2508 newstate = TRIE_LIST_ITEM( state, check ).newstate;
2513 newstate = next_alloc++;
2514 prev_states[newstate] = state;
2515 TRIE_LIST_PUSH( state, charid, newstate );
2520 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2524 TRIE_HANDLE_WORD(state);
2526 } /* end second pass */
2528 /* next alloc is the NEXT state to be allocated */
2529 trie->statecount = next_alloc;
2530 trie->states = (reg_trie_state *)
2531 PerlMemShared_realloc( trie->states,
2533 * sizeof(reg_trie_state) );
2535 /* and now dump it out before we compress it */
2536 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_list(trie, widecharmap,
2537 revcharmap, next_alloc,
2541 trie->trans = (reg_trie_trans *)
2542 PerlMemShared_calloc( transcount, sizeof(reg_trie_trans) );
2549 for( state=1 ; state < next_alloc ; state ++ ) {
2553 DEBUG_TRIE_COMPILE_MORE_r(
2554 PerlIO_printf( Perl_debug_log, "tp: %d zp: %d ",tp,zp)
2558 if (trie->states[state].trans.list) {
2559 U16 minid=TRIE_LIST_ITEM( state, 1).forid;
2563 for( idx = 2 ; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2564 const U16 forid = TRIE_LIST_ITEM( state, idx).forid;
2565 if ( forid < minid ) {
2567 } else if ( forid > maxid ) {
2571 if ( transcount < tp + maxid - minid + 1) {
2573 trie->trans = (reg_trie_trans *)
2574 PerlMemShared_realloc( trie->trans,
2576 * sizeof(reg_trie_trans) );
2577 Zero( trie->trans + (transcount / 2),
2581 base = trie->uniquecharcount + tp - minid;
2582 if ( maxid == minid ) {
2584 for ( ; zp < tp ; zp++ ) {
2585 if ( ! trie->trans[ zp ].next ) {
2586 base = trie->uniquecharcount + zp - minid;
2587 trie->trans[ zp ].next = TRIE_LIST_ITEM( state,
2589 trie->trans[ zp ].check = state;
2595 trie->trans[ tp ].next = TRIE_LIST_ITEM( state,
2597 trie->trans[ tp ].check = state;
2602 for ( idx=1; idx <= TRIE_LIST_USED( state ) ; idx++ ) {
2603 const U32 tid = base
2604 - trie->uniquecharcount
2605 + TRIE_LIST_ITEM( state, idx ).forid;
2606 trie->trans[ tid ].next = TRIE_LIST_ITEM( state,
2608 trie->trans[ tid ].check = state;
2610 tp += ( maxid - minid + 1 );
2612 Safefree(trie->states[ state ].trans.list);
2615 DEBUG_TRIE_COMPILE_MORE_r(
2616 PerlIO_printf( Perl_debug_log, " base: %d\n",base);
2619 trie->states[ state ].trans.base=base;
2621 trie->lasttrans = tp + 1;
2625 Second Pass -- Flat Table Representation.
2627 we dont use the 0 slot of either trans[] or states[] so we add 1 to
2628 each. We know that we will need Charcount+1 trans at most to store
2629 the data (one row per char at worst case) So we preallocate both
2630 structures assuming worst case.
2632 We then construct the trie using only the .next slots of the entry
2635 We use the .check field of the first entry of the node temporarily
2636 to make compression both faster and easier by keeping track of how
2637 many non zero fields are in the node.
2639 Since trans are numbered from 1 any 0 pointer in the table is a FAIL
2642 There are two terms at use here: state as a TRIE_NODEIDX() which is
2643 a number representing the first entry of the node, and state as a
2644 TRIE_NODENUM() which is the trans number. state 1 is TRIE_NODEIDX(1)
2645 and TRIE_NODENUM(1), state 2 is TRIE_NODEIDX(2) and TRIE_NODENUM(3)
2646 if there are 2 entrys per node. eg:
2654 The table is internally in the right hand, idx form. However as we
2655 also have to deal with the states array which is indexed by nodenum
2656 we have to use TRIE_NODENUM() to convert.
2659 DEBUG_TRIE_COMPILE_MORE_r( PerlIO_printf( Perl_debug_log,
2660 "%*sCompiling trie using table compiler\n",
2661 (int)depth * 2 + 2, ""));
2663 trie->trans = (reg_trie_trans *)
2664 PerlMemShared_calloc( ( TRIE_CHARCOUNT(trie) + 1 )
2665 * trie->uniquecharcount + 1,
2666 sizeof(reg_trie_trans) );
2667 trie->states = (reg_trie_state *)
2668 PerlMemShared_calloc( TRIE_CHARCOUNT(trie) + 2,
2669 sizeof(reg_trie_state) );
2670 next_alloc = trie->uniquecharcount + 1;
2673 for ( cur = first ; cur < last ; cur = regnext( cur ) ) {
2675 regnode *noper = NEXTOPER( cur );
2676 const U8 *uc = (U8*)STRING( noper );
2677 const U8 *e = uc + STR_LEN( noper );
2679 U32 state = 1; /* required init */
2681 U16 charid = 0; /* sanity init */
2682 U32 accept_state = 0; /* sanity init */
2684 U32 wordlen = 0; /* required init */
2686 if (OP(noper) == NOTHING) {
2687 regnode *noper_next= regnext(noper);
2688 if (noper_next != tail && OP(noper_next) == flags) {
2690 uc= (U8*)STRING(noper);
2691 e= uc + STR_LEN(noper);
2695 if ( OP(noper) != NOTHING ) {
2696 for ( ; uc < e ; uc += len ) {
2701 charid = trie->charmap[ uvc ];
2703 SV* const * const svpp = hv_fetch( widecharmap,
2707 charid = svpp ? (U16)SvIV(*svpp) : 0;
2711 if ( !trie->trans[ state + charid ].next ) {
2712 trie->trans[ state + charid ].next = next_alloc;
2713 trie->trans[ state ].check++;
2714 prev_states[TRIE_NODENUM(next_alloc)]
2715 = TRIE_NODENUM(state);
2716 next_alloc += trie->uniquecharcount;
2718 state = trie->trans[ state + charid ].next;
2720 Perl_croak( aTHX_ "panic! In trie construction, no char mapping for %"IVdf, uvc );
2722 /* charid is now 0 if we dont know the char read, or
2723 * nonzero if we do */
2726 accept_state = TRIE_NODENUM( state );
2727 TRIE_HANDLE_WORD(accept_state);
2729 } /* end second pass */
2731 /* and now dump it out before we compress it */
2732 DEBUG_TRIE_COMPILE_MORE_r(dump_trie_interim_table(trie, widecharmap,
2734 next_alloc, depth+1));
2738 * Inplace compress the table.*
2740 For sparse data sets the table constructed by the trie algorithm will
2741 be mostly 0/FAIL transitions or to put it another way mostly empty.
2742 (Note that leaf nodes will not contain any transitions.)
2744 This algorithm compresses the tables by eliminating most such
2745 transitions, at the cost of a modest bit of extra work during lookup:
2747 - Each states[] entry contains a .base field which indicates the
2748 index in the state[] array wheres its transition data is stored.
2750 - If .base is 0 there are no valid transitions from that node.
2752 - If .base is nonzero then charid is added to it to find an entry in
2755 -If trans[states[state].base+charid].check!=state then the
2756 transition is taken to be a 0/Fail transition. Thus if there are fail
2757 transitions at the front of the node then the .base offset will point
2758 somewhere inside the previous nodes data (or maybe even into a node
2759 even earlier), but the .check field determines if the transition is
2763 The following process inplace converts the table to the compressed
2764 table: We first do not compress the root node 1,and mark all its
2765 .check pointers as 1 and set its .base pointer as 1 as well. This
2766 allows us to do a DFA construction from the compressed table later,
2767 and ensures that any .base pointers we calculate later are greater
2770 - We set 'pos' to indicate the first entry of the second node.
2772 - We then iterate over the columns of the node, finding the first and
2773 last used entry at l and m. We then copy l..m into pos..(pos+m-l),
2774 and set the .check pointers accordingly, and advance pos
2775 appropriately and repreat for the next node. Note that when we copy
2776 the next pointers we have to convert them from the original
2777 NODEIDX form to NODENUM form as the former is not valid post
2780 - If a node has no transitions used we mark its base as 0 and do not
2781 advance the pos pointer.
2783 - If a node only has one transition we use a second pointer into the
2784 structure to fill in allocated fail transitions from other states.
2785 This pointer is independent of the main pointer and scans forward
2786 looking for null transitions that are allocated to a state. When it
2787 finds one it writes the single transition into the "hole". If the
2788 pointer doesnt find one the single transition is appended as normal.
2790 - Once compressed we can Renew/realloc the structures to release the
2793 See "Table-Compression Methods" in sec 3.9 of the Red Dragon,
2794 specifically Fig 3.47 and the associated pseudocode.
2798 const U32 laststate = TRIE_NODENUM( next_alloc );
2801 trie->statecount = laststate;
2803 for ( state = 1 ; state < laststate ; state++ ) {
2805 const U32 stateidx = TRIE_NODEIDX( state );
2806 const U32 o_used = trie->trans[ stateidx ].check;
2807 U32 used = trie->trans[ stateidx ].check;
2808 trie->trans[ stateidx ].check = 0;
2811 used && charid < trie->uniquecharcount;
2814 if ( flag || trie->trans[ stateidx + charid ].next ) {
2815 if ( trie->trans[ stateidx + charid ].next ) {
2817 for ( ; zp < pos ; zp++ ) {
2818 if ( ! trie->trans[ zp ].next ) {
2822 trie->states[ state ].trans.base
2824 + trie->uniquecharcount
2826 trie->trans[ zp ].next
2827 = SAFE_TRIE_NODENUM( trie->trans[ stateidx
2829 trie->trans[ zp ].check = state;
2830 if ( ++zp > pos ) pos = zp;
2837 trie->states[ state ].trans.base
2838 = pos + trie->uniquecharcount - charid ;
2840 trie->trans[ pos ].next
2841 = SAFE_TRIE_NODENUM(
2842 trie->trans[ stateidx + charid ].next );
2843 trie->trans[ pos ].check = state;
2848 trie->lasttrans = pos + 1;
2849 trie->states = (reg_trie_state *)
2850 PerlMemShared_realloc( trie->states, laststate
2851 * sizeof(reg_trie_state) );
2852 DEBUG_TRIE_COMPILE_MORE_r(
2853 PerlIO_printf( Perl_debug_log,
2854 "%*sAlloc: %d Orig: %"IVdf" elements, Final:%"IVdf". Savings of %%%5.2f\n",
2855 (int)depth * 2 + 2,"",
2856 (int)( ( TRIE_CHARCOUNT(trie) + 1 ) * trie->uniquecharcount
2860 ( ( next_alloc - pos ) * 100 ) / (double)next_alloc );
2863 } /* end table compress */
2865 DEBUG_TRIE_COMPILE_MORE_r(
2866 PerlIO_printf(Perl_debug_log,
2867 "%*sStatecount:%"UVxf" Lasttrans:%"UVxf"\n",
2868 (int)depth * 2 + 2, "",
2869 (UV)trie->statecount,
2870 (UV)trie->lasttrans)
2872 /* resize the trans array to remove unused space */
2873 trie->trans = (reg_trie_trans *)
2874 PerlMemShared_realloc( trie->trans, trie->lasttrans
2875 * sizeof(reg_trie_trans) );
2877 { /* Modify the program and insert the new TRIE node */
2878 U8 nodetype =(U8)(flags & 0xFF);
2882 regnode *optimize = NULL;
2883 #ifdef RE_TRACK_PATTERN_OFFSETS
2886 U32 mjd_nodelen = 0;
2887 #endif /* RE_TRACK_PATTERN_OFFSETS */
2888 #endif /* DEBUGGING */
2890 This means we convert either the first branch or the first Exact,
2891 depending on whether the thing following (in 'last') is a branch
2892 or not and whther first is the startbranch (ie is it a sub part of
2893 the alternation or is it the whole thing.)
2894 Assuming its a sub part we convert the EXACT otherwise we convert
2895 the whole branch sequence, including the first.
2897 /* Find the node we are going to overwrite */
2898 if ( first != startbranch || OP( last ) == BRANCH ) {
2899 /* branch sub-chain */
2900 NEXT_OFF( first ) = (U16)(last - first);
2901 #ifdef RE_TRACK_PATTERN_OFFSETS
2903 mjd_offset= Node_Offset((convert));
2904 mjd_nodelen= Node_Length((convert));
2907 /* whole branch chain */
2909 #ifdef RE_TRACK_PATTERN_OFFSETS
2912 const regnode *nop = NEXTOPER( convert );
2913 mjd_offset= Node_Offset((nop));
2914 mjd_nodelen= Node_Length((nop));
2918 PerlIO_printf(Perl_debug_log,
2919 "%*sMJD offset:%"UVuf" MJD length:%"UVuf"\n",
2920 (int)depth * 2 + 2, "",
2921 (UV)mjd_offset, (UV)mjd_nodelen)
2924 /* But first we check to see if there is a common prefix we can
2925 split out as an EXACT and put in front of the TRIE node. */
2926 trie->startstate= 1;
2927 if ( trie->bitmap && !widecharmap && !trie->jump ) {
2929 for ( state = 1 ; state < trie->statecount-1 ; state++ ) {
2933 const U32 base = trie->states[ state ].trans.base;
2935 if ( trie->states[state].wordnum )
2938 for ( ofs = 0 ; ofs < trie->uniquecharcount ; ofs++ ) {
2939 if ( ( base + ofs >= trie->uniquecharcount ) &&
2940 ( base + ofs - trie->uniquecharcount < trie->lasttrans ) &&
2941 trie->trans[ base + ofs - trie->uniquecharcount ].check == state )
2943 if ( ++count > 1 ) {
2944 SV **tmp = av_fetch( revcharmap, ofs, 0);
2945 const U8 *ch = (U8*)SvPV_nolen_const( *tmp );
2946 if ( state == 1 ) break;
2948 Zero(trie->bitmap, ANYOF_BITMAP_SIZE, char);
2950 PerlIO_printf(Perl_debug_log,
2951 "%*sNew Start State=%"UVuf" Class: [",
2952 (int)depth * 2 + 2, "",
2955 SV ** const tmp = av_fetch( revcharmap, idx, 0);
2956 const U8 * const ch = (U8*)SvPV_nolen_const( *tmp );
2958 TRIE_BITMAP_SET(trie,*ch);
2960 TRIE_BITMAP_SET(trie, folder[ *ch ]);
2962 PerlIO_printf(Perl_debug_log, "%s", (char*)ch)
2966 TRIE_BITMAP_SET(trie,*ch);
2968 TRIE_BITMAP_SET(trie,folder[ *ch ]);
2969 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"%s", ch));
2975 SV **tmp = av_fetch( revcharmap, idx, 0);
2977 char *ch = SvPV( *tmp, len );
2979 SV *sv=sv_newmortal();
2980 PerlIO_printf( Perl_debug_log,
2981 "%*sPrefix State: %"UVuf" Idx:%"UVuf" Char='%s'\n",
2982 (int)depth * 2 + 2, "",
2984 pv_pretty(sv, SvPV_nolen_const(*tmp), SvCUR(*tmp), 6,
2985 PL_colors[0], PL_colors[1],
2986 (SvUTF8(*tmp) ? PERL_PV_ESCAPE_UNI : 0) |
2987 PERL_PV_ESCAPE_FIRSTCHAR
2992 OP( convert ) = nodetype;
2993 str=STRING(convert);
2996 STR_LEN(convert) += len;
3002 DEBUG_OPTIMISE_r(PerlIO_printf( Perl_debug_log,"]\n"));
3007 trie->prefixlen = (state-1);
3009 regnode *n = convert+NODE_SZ_STR(convert);
3010 NEXT_OFF(convert) = NODE_SZ_STR(convert);
3011 trie->startstate = state;
3012 trie->minlen -= (state - 1);
3013 trie->maxlen -= (state - 1);
3015 /* At least the UNICOS C compiler choked on this
3016 * being argument to DEBUG_r(), so let's just have
3019 #ifdef PERL_EXT_RE_BUILD
3025 regnode *fix = convert;
3026 U32 word = trie->wordcount;
3028 Set_Node_Offset_Length(convert, mjd_offset, state - 1);
3029 while( ++fix < n ) {
3030 Set_Node_Offset_Length(fix, 0, 0);
3033 SV ** const tmp = av_fetch( trie_words, word, 0 );
3035 if ( STR_LEN(convert) <= SvCUR(*tmp) )
3036 sv_chop(*tmp, SvPV_nolen(*tmp) + STR_LEN(convert));
3038 sv_chop(*tmp, SvPV_nolen(*tmp) + SvCUR(*tmp));
3046 NEXT_OFF(convert) = (U16)(tail - convert);
3047 DEBUG_r(optimize= n);
3053 if ( trie->maxlen ) {
3054 NEXT_OFF( convert ) = (U16)(tail - convert);
3055 ARG_SET( convert, data_slot );
3056 /* Store the offset to the first unabsorbed branch in
3057 jump[0], which is otherwise unused by the jump logic.
3058 We use this when dumping a trie and during optimisation. */
3060 trie->jump[0] = (U16)(nextbranch - convert);
3062 /* If the start state is not accepting (meaning there is no empty string/NOTHING)
3063 * and there is a bitmap
3064 * and the first "jump target" node we found leaves enough room
3065 * then convert the TRIE node into a TRIEC node, with the bitmap
3066 * embedded inline in the opcode - this is hypothetically faster.
3068 if ( !trie->states[trie->startstate].wordnum
3070 && ( (char *)jumper - (char *)convert) >= (int)sizeof(struct regnode_charclass) )
3072 OP( convert ) = TRIEC;
3073 Copy(trie->bitmap, ((struct regnode_charclass *)convert)->bitmap, ANYOF_BITMAP_SIZE, char);
3074 PerlMemShared_free(trie->bitmap);
3077 OP( convert ) = TRIE;
3079 /* store the type in the flags */
3080 convert->flags = nodetype;
3084 + regarglen[ OP( convert ) ];
3086 /* XXX We really should free up the resource in trie now,
3087 as we won't use them - (which resources?) dmq */
3089 /* needed for dumping*/
3090 DEBUG_r(if (optimize) {
3091 regnode *opt = convert;
3093 while ( ++opt < optimize) {
3094 Set_Node_Offset_Length(opt,0,0);
3097 Try to clean up some of the debris left after the
3100 while( optimize < jumper ) {
3101 mjd_nodelen += Node_Length((optimize));
3102 OP( optimize ) = OPTIMIZED;
3103 Set_Node_Offset_Length(optimize,0,0);
3106 Set_Node_Offset_Length(convert,mjd_offset,mjd_nodelen);
3108 } /* end node insert */
3109 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, convert);
3111 /* Finish populating the prev field of the wordinfo array. Walk back
3112 * from each accept state until we find another accept state, and if
3113 * so, point the first word's .prev field at the second word. If the
3114 * second already has a .prev field set, stop now. This will be the
3115 * case either if we've already processed that word's accept state,
3116 * or that state had multiple words, and the overspill words were
3117 * already linked up earlier.
3124 for (word=1; word <= trie->wordcount; word++) {
3126 if (trie->wordinfo[word].prev)
3128 state = trie->wordinfo[word].accept;
3130 state = prev_states[state];
3133 prev = trie->states[state].wordnum;
3137 trie->wordinfo[word].prev = prev;
3139 Safefree(prev_states);
3143 /* and now dump out the compressed format */
3144 DEBUG_TRIE_COMPILE_r(dump_trie(trie, widecharmap, revcharmap, depth+1));
3146 RExC_rxi->data->data[ data_slot + 1 ] = (void*)widecharmap;
3148 RExC_rxi->data->data[ data_slot + TRIE_WORDS_OFFSET ] = (void*)trie_words;
3149 RExC_rxi->data->data[ data_slot + 3 ] = (void*)revcharmap;
3151 SvREFCNT_dec_NN(revcharmap);
3155 : trie->startstate>1
3161 S_construct_ahocorasick_from_trie(pTHX_ RExC_state_t *pRExC_state, regnode *source, U32 depth)
3163 /* The Trie is constructed and compressed now so we can build a fail array if
3166 This is basically the Aho-Corasick algorithm. Its from exercise 3.31 and
3168 "Red Dragon" -- Compilers, principles, techniques, and tools. Aho, Sethi,
3172 We find the fail state for each state in the trie, this state is the longest
3173 proper suffix of the current state's 'word' that is also a proper prefix of
3174 another word in our trie. State 1 represents the word '' and is thus the
3175 default fail state. This allows the DFA not to have to restart after its
3176 tried and failed a word at a given point, it simply continues as though it
3177 had been matching the other word in the first place.
3179 'abcdgu'=~/abcdefg|cdgu/
3180 When we get to 'd' we are still matching the first word, we would encounter
3181 'g' which would fail, which would bring us to the state representing 'd' in
3182 the second word where we would try 'g' and succeed, proceeding to match
3185 /* add a fail transition */
3186 const U32 trie_offset = ARG(source);
3187 reg_trie_data *trie=(reg_trie_data *)RExC_rxi->data->data[trie_offset];
3189 const U32 ucharcount = trie->uniquecharcount;
3190 const U32 numstates = trie->statecount;
3191 const U32 ubound = trie->lasttrans + ucharcount;
3195 U32 base = trie->states[ 1 ].trans.base;
3198 const U32 data_slot = add_data( pRExC_state, STR_WITH_LEN("T"));
3200 GET_RE_DEBUG_FLAGS_DECL;
3202 PERL_ARGS_ASSERT_CONSTRUCT_AHOCORASICK_FROM_TRIE;
3203 PERL_UNUSED_CONTEXT;
3205 PERL_UNUSED_ARG(depth);
3208 if ( OP(source) == TRIE ) {
3209 struct regnode_1 *op = (struct regnode_1 *)
3210 PerlMemShared_calloc(1, sizeof(struct regnode_1));
3211 StructCopy(source,op,struct regnode_1);
3212 stclass = (regnode *)op;
3214 struct regnode_charclass *op = (struct regnode_charclass *)
3215 PerlMemShared_calloc(1, sizeof(struct regnode_charclass));
3216 StructCopy(source,op,struct regnode_charclass);
3217 stclass = (regnode *)op;
3219 OP(stclass)+=2; /* convert the TRIE type to its AHO-CORASICK equivalent */
3221 ARG_SET( stclass, data_slot );
3222 aho = (reg_ac_data *) PerlMemShared_calloc( 1, sizeof(reg_ac_data) );
3223 RExC_rxi->data->data[ data_slot ] = (void*)aho;
3224 aho->trie=trie_offset;
3225 aho->states=(reg_trie_state *)PerlMemShared_malloc( numstates * sizeof(reg_trie_state) );
3226 Copy( trie->states, aho->states, numstates, reg_trie_state );
3227 Newxz( q, numstates, U32);
3228 aho->fail = (U32 *) PerlMemShared_calloc( numstates, sizeof(U32) );
3231 /* initialize fail[0..1] to be 1 so that we always have
3232 a valid final fail state */
3233 fail[ 0 ] = fail[ 1 ] = 1;
3235 for ( charid = 0; charid < ucharcount ; charid++ ) {
3236 const U32 newstate = TRIE_TRANS_STATE( 1, base, ucharcount, charid, 0 );
3238 q[ q_write ] = newstate;
3239 /* set to point at the root */
3240 fail[ q[ q_write++ ] ]=1;
3243 while ( q_read < q_write) {
3244 const U32 cur = q[ q_read++ % numstates ];
3245 base = trie->states[ cur ].trans.base;
3247 for ( charid = 0 ; charid < ucharcount ; charid++ ) {
3248 const U32 ch_state = TRIE_TRANS_STATE( cur, base, ucharcount, charid, 1 );
3250 U32 fail_state = cur;
3253 fail_state = fail[ fail_state ];
3254 fail_base = aho->states[ fail_state ].trans.base;
3255 } while ( !TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 ) );
3257 fail_state = TRIE_TRANS_STATE( fail_state, fail_base, ucharcount, charid, 1 );
3258 fail[ ch_state ] = fail_state;
3259 if ( !aho->states[ ch_state ].wordnum && aho->states[ fail_state ].wordnum )
3261 aho->states[ ch_state ].wordnum = aho->states[ fail_state ].wordnum;
3263 q[ q_write++ % numstates] = ch_state;
3267 /* restore fail[0..1] to 0 so that we "fall out" of the AC loop
3268 when we fail in state 1, this allows us to use the
3269 charclass scan to find a valid start char. This is based on the principle
3270 that theres a good chance the string being searched contains lots of stuff
3271 that cant be a start char.
3273 fail[ 0 ] = fail[ 1 ] = 0;
3274 DEBUG_TRIE_COMPILE_r({
3275 PerlIO_printf(Perl_debug_log,
3276 "%*sStclass Failtable (%"UVuf" states): 0",
3277 (int)(depth * 2), "", (UV)numstates
3279 for( q_read=1; q_read<numstates; q_read++ ) {
3280 PerlIO_printf(Perl_debug_log, ", %"UVuf, (UV)fail[q_read]);
3282 PerlIO_printf(Perl_debug_log, "\n");
3285 /*RExC_seen |= REG_TRIEDFA_SEEN;*/
3290 #define DEBUG_PEEP(str,scan,depth) \
3291 DEBUG_OPTIMISE_r({if (scan){ \
3292 regnode *Next = regnext(scan); \
3293 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state); \
3294 PerlIO_printf(Perl_debug_log, "%*s" str ">%3d: %s (%d)", \
3295 (int)depth*2, "", REG_NODE_NUM(scan), SvPV_nolen_const(RExC_mysv),\
3296 Next ? (REG_NODE_NUM(Next)) : 0 ); \
3297 DEBUG_SHOW_STUDY_FLAGS(flags," [ ","]");\
3298 PerlIO_printf(Perl_debug_log, "\n"); \
3301 /* The below joins as many adjacent EXACTish nodes as possible into a single
3302 * one. The regop may be changed if the node(s) contain certain sequences that
3303 * require special handling. The joining is only done if:
3304 * 1) there is room in the current conglomerated node to entirely contain the
3306 * 2) they are the exact same node type
3308 * The adjacent nodes actually may be separated by NOTHING-kind nodes, and
3309 * these get optimized out
3311 * If a node is to match under /i (folded), the number of characters it matches
3312 * can be different than its character length if it contains a multi-character
3313 * fold. *min_subtract is set to the total delta number of characters of the
3316 * And *unfolded_multi_char is set to indicate whether or not the node contains
3317 * an unfolded multi-char fold. This happens when whether the fold is valid or
3318 * not won't be known until runtime; namely for EXACTF nodes that contain LATIN
3319 * SMALL LETTER SHARP S, as only if the target string being matched against
3320 * turns out to be UTF-8 is that fold valid; and also for EXACTFL nodes whose
3321 * folding rules depend on the locale in force at runtime. (Multi-char folds
3322 * whose components are all above the Latin1 range are not run-time locale
3323 * dependent, and have already been folded by the time this function is
3326 * This is as good a place as any to discuss the design of handling these
3327 * multi-character fold sequences. It's been wrong in Perl for a very long
3328 * time. There are three code points in Unicode whose multi-character folds
3329 * were long ago discovered to mess things up. The previous designs for
3330 * dealing with these involved assigning a special node for them. This
3331 * approach doesn't always work, as evidenced by this example:
3332 * "\xDFs" =~ /s\xDF/ui # Used to fail before these patches
3333 * Both sides fold to "sss", but if the pattern is parsed to create a node that
3334 * would match just the \xDF, it won't be able to handle the case where a
3335 * successful match would have to cross the node's boundary. The new approach
3336 * that hopefully generally solves the problem generates an EXACTFU_SS node
3337 * that is "sss" in this case.
3339 * It turns out that there are problems with all multi-character folds, and not
3340 * just these three. Now the code is general, for all such cases. The
3341 * approach taken is:
3342 * 1) This routine examines each EXACTFish node that could contain multi-
3343 * character folded sequences. Since a single character can fold into
3344 * such a sequence, the minimum match length for this node is less than
3345 * the number of characters in the node. This routine returns in
3346 * *min_subtract how many characters to subtract from the the actual
3347 * length of the string to get a real minimum match length; it is 0 if
3348 * there are no multi-char foldeds. This delta is used by the caller to
3349 * adjust the min length of the match, and the delta between min and max,
3350 * so that the optimizer doesn't reject these possibilities based on size
3352 * 2) For the sequence involving the Sharp s (\xDF), the node type EXACTFU_SS
3353 * is used for an EXACTFU node that contains at least one "ss" sequence in
3354 * it. For non-UTF-8 patterns and strings, this is the only case where
3355 * there is a possible fold length change. That means that a regular
3356 * EXACTFU node without UTF-8 involvement doesn't have to concern itself
3357 * with length changes, and so can be processed faster. regexec.c takes
3358 * advantage of this. Generally, an EXACTFish node that is in UTF-8 is
3359 * pre-folded by regcomp.c (except EXACTFL, some of whose folds aren't
3360 * known until runtime). This saves effort in regex matching. However,
3361 * the pre-folding isn't done for non-UTF8 patterns because the fold of
3362 * the MICRO SIGN requires UTF-8, and we don't want to slow things down by
3363 * forcing the pattern into UTF8 unless necessary. Also what EXACTF (and,
3364 * again, EXACTFL) nodes fold to isn't known until runtime. The fold
3365 * possibilities for the non-UTF8 patterns are quite simple, except for
3366 * the sharp s. All the ones that don't involve a UTF-8 target string are
3367 * members of a fold-pair, and arrays are set up for all of them so that
3368 * the other member of the pair can be found quickly. Code elsewhere in
3369 * this file makes sure that in EXACTFU nodes, the sharp s gets folded to
3370 * 'ss', even if the pattern isn't UTF-8. This avoids the issues
3371 * described in the next item.
3372 * 3) A problem remains for unfolded multi-char folds. (These occur when the
3373 * validity of the fold won't be known until runtime, and so must remain
3374 * unfolded for now. This happens for the sharp s in EXACTF and EXACTFA
3375 * nodes when the pattern isn't in UTF-8. (Note, BTW, that there cannot
3376 * be an EXACTF node with a UTF-8 pattern.) They also occur for various
3377 * folds in EXACTFL nodes, regardless of the UTF-ness of the pattern.)
3378 * The reason this is a problem is that the optimizer part of regexec.c
3379 * (probably unwittingly, in Perl_regexec_flags()) makes an assumption
3380 * that a character in the pattern corresponds to at most a single
3381 * character in the target string. (And I do mean character, and not byte
3382 * here, unlike other parts of the documentation that have never been
3383 * updated to account for multibyte Unicode.) sharp s in EXACTF and
3384 * EXACTFL nodes can match the two character string 'ss'; in EXACTFA nodes
3385 * it can match "\x{17F}\x{17F}". These, along with other ones in EXACTFL
3386 * nodes, violate the assumption, and they are the only instances where it
3387 * is violated. I'm reluctant to try to change the assumption, as the
3388 * code involved is impenetrable to me (khw), so instead the code here
3389 * punts. This routine examines EXACTFL nodes, and (when the pattern
3390 * isn't UTF-8) EXACTF and EXACTFA for such unfolded folds, and returns a
3391 * boolean indicating whether or not the node contains such a fold. When
3392 * it is true, the caller sets a flag that later causes the optimizer in
3393 * this file to not set values for the floating and fixed string lengths,
3394 * and thus avoids the optimizer code in regexec.c that makes the invalid
3395 * assumption. Thus, there is no optimization based on string lengths for
3396 * EXACTFL nodes that contain these few folds, nor for non-UTF8-pattern
3397 * EXACTF and EXACTFA nodes that contain the sharp s. (The reason the
3398 * assumption is wrong only in these cases is that all other non-UTF-8
3399 * folds are 1-1; and, for UTF-8 patterns, we pre-fold all other folds to
3400 * their expanded versions. (Again, we can't prefold sharp s to 'ss' in
3401 * EXACTF nodes because we don't know at compile time if it actually
3402 * matches 'ss' or not. For EXACTF nodes it will match iff the target
3403 * string is in UTF-8. This is in contrast to EXACTFU nodes, where it
3404 * always matches; and EXACTFA where it never does. In an EXACTFA node in
3405 * a UTF-8 pattern, sharp s is folded to "\x{17F}\x{17F}, avoiding the
3406 * problem; but in a non-UTF8 pattern, folding it to that above-Latin1
3407 * string would require the pattern to be forced into UTF-8, the overhead
3408 * of which we want to avoid. Similarly the unfolded multi-char folds in
3409 * EXACTFL nodes will match iff the locale at the time of match is a UTF-8
3412 * Similarly, the code that generates tries doesn't currently handle
3413 * not-already-folded multi-char folds, and it looks like a pain to change
3414 * that. Therefore, trie generation of EXACTFA nodes with the sharp s
3415 * doesn't work. Instead, such an EXACTFA is turned into a new regnode,
3416 * EXACTFA_NO_TRIE, which the trie code knows not to handle. Most people
3417 * using /iaa matching will be doing so almost entirely with ASCII
3418 * strings, so this should rarely be encountered in practice */
3420 #define JOIN_EXACT(scan,min_subtract,unfolded_multi_char, flags) \
3421 if (PL_regkind[OP(scan)] == EXACT) \
3422 join_exact(pRExC_state,(scan),(min_subtract),unfolded_multi_char, (flags),NULL,depth+1)
3425 S_join_exact(pTHX_ RExC_state_t *pRExC_state, regnode *scan,
3426 UV *min_subtract, bool *unfolded_multi_char,
3427 U32 flags,regnode *val, U32 depth)
3429 /* Merge several consecutive EXACTish nodes into one. */
3430 regnode *n = regnext(scan);
3432 regnode *next = scan + NODE_SZ_STR(scan);
3436 regnode *stop = scan;
3437 GET_RE_DEBUG_FLAGS_DECL;
3439 PERL_UNUSED_ARG(depth);
3442 PERL_ARGS_ASSERT_JOIN_EXACT;
3443 #ifndef EXPERIMENTAL_INPLACESCAN
3444 PERL_UNUSED_ARG(flags);
3445 PERL_UNUSED_ARG(val);
3447 DEBUG_PEEP("join",scan,depth);
3449 /* Look through the subsequent nodes in the chain. Skip NOTHING, merge
3450 * EXACT ones that are mergeable to the current one. */
3452 && (PL_regkind[OP(n)] == NOTHING
3453 || (stringok && OP(n) == OP(scan)))
3455 && NEXT_OFF(scan) + NEXT_OFF(n) < I16_MAX)
3458 if (OP(n) == TAIL || n > next)
3460 if (PL_regkind[OP(n)] == NOTHING) {
3461 DEBUG_PEEP("skip:",n,depth);
3462 NEXT_OFF(scan) += NEXT_OFF(n);
3463 next = n + NODE_STEP_REGNODE;
3470 else if (stringok) {
3471 const unsigned int oldl = STR_LEN(scan);
3472 regnode * const nnext = regnext(n);
3474 /* XXX I (khw) kind of doubt that this works on platforms (should
3475 * Perl ever run on one) where U8_MAX is above 255 because of lots
3476 * of other assumptions */
3477 /* Don't join if the sum can't fit into a single node */
3478 if (oldl + STR_LEN(n) > U8_MAX)
3481 DEBUG_PEEP("merg",n,depth);
3484 NEXT_OFF(scan) += NEXT_OFF(n);
3485 STR_LEN(scan) += STR_LEN(n);
3486 next = n + NODE_SZ_STR(n);
3487 /* Now we can overwrite *n : */
3488 Move(STRING(n), STRING(scan) + oldl, STR_LEN(n), char);
3496 #ifdef EXPERIMENTAL_INPLACESCAN
3497 if (flags && !NEXT_OFF(n)) {
3498 DEBUG_PEEP("atch", val, depth);
3499 if (reg_off_by_arg[OP(n)]) {
3500 ARG_SET(n, val - n);
3503 NEXT_OFF(n) = val - n;
3511 *unfolded_multi_char = FALSE;
3513 /* Here, all the adjacent mergeable EXACTish nodes have been merged. We
3514 * can now analyze for sequences of problematic code points. (Prior to
3515 * this final joining, sequences could have been split over boundaries, and
3516 * hence missed). The sequences only happen in folding, hence for any
3517 * non-EXACT EXACTish node */
3518 if (OP(scan) != EXACT && OP(scan) != EXACTL) {
3519 U8* s0 = (U8*) STRING(scan);
3521 U8* s_end = s0 + STR_LEN(scan);
3523 int total_count_delta = 0; /* Total delta number of characters that
3524 multi-char folds expand to */
3526 /* One pass is made over the node's string looking for all the
3527 * possibilities. To avoid some tests in the loop, there are two main
3528 * cases, for UTF-8 patterns (which can't have EXACTF nodes) and
3533 if (OP(scan) == EXACTFL) {
3536 /* An EXACTFL node would already have been changed to another
3537 * node type unless there is at least one character in it that
3538 * is problematic; likely a character whose fold definition
3539 * won't be known until runtime, and so has yet to be folded.
3540 * For all but the UTF-8 locale, folds are 1-1 in length, but
3541 * to handle the UTF-8 case, we need to create a temporary
3542 * folded copy using UTF-8 locale rules in order to analyze it.
3543 * This is because our macros that look to see if a sequence is
3544 * a multi-char fold assume everything is folded (otherwise the
3545 * tests in those macros would be too complicated and slow).
3546 * Note that here, the non-problematic folds will have already
3547 * been done, so we can just copy such characters. We actually
3548 * don't completely fold the EXACTFL string. We skip the
3549 * unfolded multi-char folds, as that would just create work
3550 * below to figure out the size they already are */
3552 Newx(folded, UTF8_MAX_FOLD_CHAR_EXPAND * STR_LEN(scan) + 1, U8);
3555 STRLEN s_len = UTF8SKIP(s);
3556 if (! is_PROBLEMATIC_LOCALE_FOLD_utf8(s)) {
3557 Copy(s, d, s_len, U8);
3560 else if (is_FOLDS_TO_MULTI_utf8(s)) {
3561 *unfolded_multi_char = TRUE;
3562 Copy(s, d, s_len, U8);
3565 else if (isASCII(*s)) {
3566 *(d++) = toFOLD(*s);
3570 _to_utf8_fold_flags(s, d, &len, FOLD_FLAGS_FULL);
3576 /* Point the remainder of the routine to look at our temporary
3580 } /* End of creating folded copy of EXACTFL string */
3582 /* Examine the string for a multi-character fold sequence. UTF-8
3583 * patterns have all characters pre-folded by the time this code is
3585 while (s < s_end - 1) /* Can stop 1 before the end, as minimum
3586 length sequence we are looking for is 2 */
3588 int count = 0; /* How many characters in a multi-char fold */
3589 int len = is_MULTI_CHAR_FOLD_utf8_safe(s, s_end);
3590 if (! len) { /* Not a multi-char fold: get next char */
3595 /* Nodes with 'ss' require special handling, except for
3596 * EXACTFA-ish for which there is no multi-char fold to this */
3597 if (len == 2 && *s == 's' && *(s+1) == 's'
3598 && OP(scan) != EXACTFA
3599 && OP(scan) != EXACTFA_NO_TRIE)
3602 if (OP(scan) != EXACTFL) {
3603 OP(scan) = EXACTFU_SS;
3607 else { /* Here is a generic multi-char fold. */
3608 U8* multi_end = s + len;
3610 /* Count how many characters are in it. In the case of
3611 * /aa, no folds which contain ASCII code points are
3612 * allowed, so check for those, and skip if found. */
3613 if (OP(scan) != EXACTFA && OP(scan) != EXACTFA_NO_TRIE) {
3614 count = utf8_length(s, multi_end);
3618 while (s < multi_end) {
3621 goto next_iteration;
3631 /* The delta is how long the sequence is minus 1 (1 is how long
3632 * the character that folds to the sequence is) */
3633 total_count_delta += count - 1;
3637 /* We created a temporary folded copy of the string in EXACTFL
3638 * nodes. Therefore we need to be sure it doesn't go below zero,
3639 * as the real string could be shorter */
3640 if (OP(scan) == EXACTFL) {
3641 int total_chars = utf8_length((U8*) STRING(scan),
3642 (U8*) STRING(scan) + STR_LEN(scan));
3643 if (total_count_delta > total_chars) {
3644 total_count_delta = total_chars;
3648 *min_subtract += total_count_delta;
3651 else if (OP(scan) == EXACTFA) {
3653 /* Non-UTF-8 pattern, EXACTFA node. There can't be a multi-char
3654 * fold to the ASCII range (and there are no existing ones in the
3655 * upper latin1 range). But, as outlined in the comments preceding
3656 * this function, we need to flag any occurrences of the sharp s.
3657 * This character forbids trie formation (because of added
3660 if (*s == LATIN_SMALL_LETTER_SHARP_S) {
3661 OP(scan) = EXACTFA_NO_TRIE;
3662 *unfolded_multi_char = TRUE;
3671 /* Non-UTF-8 pattern, not EXACTFA node. Look for the multi-char
3672 * folds that are all Latin1. As explained in the comments
3673 * preceding this function, we look also for the sharp s in EXACTF
3674 * and EXACTFL nodes; it can be in the final position. Otherwise
3675 * we can stop looking 1 byte earlier because have to find at least
3676 * two characters for a multi-fold */
3677 const U8* upper = (OP(scan) == EXACTF || OP(scan) == EXACTFL)
3682 int len = is_MULTI_CHAR_FOLD_latin1_safe(s, s_end);
3683 if (! len) { /* Not a multi-char fold. */
3684 if (*s == LATIN_SMALL_LETTER_SHARP_S
3685 && (OP(scan) == EXACTF || OP(scan) == EXACTFL))
3687 *unfolded_multi_char = TRUE;
3694 && isALPHA_FOLD_EQ(*s, 's')
3695 && isALPHA_FOLD_EQ(*(s+1), 's'))
3698 /* EXACTF nodes need to know that the minimum length
3699 * changed so that a sharp s in the string can match this
3700 * ss in the pattern, but they remain EXACTF nodes, as they
3701 * won't match this unless the target string is is UTF-8,
3702 * which we don't know until runtime. EXACTFL nodes can't
3703 * transform into EXACTFU nodes */
3704 if (OP(scan) != EXACTF && OP(scan) != EXACTFL) {
3705 OP(scan) = EXACTFU_SS;
3709 *min_subtract += len - 1;
3716 /* Allow dumping but overwriting the collection of skipped
3717 * ops and/or strings with fake optimized ops */
3718 n = scan + NODE_SZ_STR(scan);
3726 DEBUG_OPTIMISE_r(if (merged){DEBUG_PEEP("finl",scan,depth)});
3730 /* REx optimizer. Converts nodes into quicker variants "in place".
3731 Finds fixed substrings. */
3733 /* Stops at toplevel WHILEM as well as at "last". At end *scanp is set
3734 to the position after last scanned or to NULL. */
3736 #define INIT_AND_WITHP \
3737 assert(!and_withp); \
3738 Newx(and_withp,1, regnode_ssc); \
3739 SAVEFREEPV(and_withp)
3743 S_unwind_scan_frames(pTHX_ const void *p)
3745 scan_frame *f= (scan_frame *)p;
3747 scan_frame *n= f->next_frame;
3755 S_study_chunk(pTHX_ RExC_state_t *pRExC_state, regnode **scanp,
3756 SSize_t *minlenp, SSize_t *deltap,
3761 regnode_ssc *and_withp,
3762 U32 flags, U32 depth)
3763 /* scanp: Start here (read-write). */
3764 /* deltap: Write maxlen-minlen here. */
3765 /* last: Stop before this one. */
3766 /* data: string data about the pattern */
3767 /* stopparen: treat close N as END */
3768 /* recursed: which subroutines have we recursed into */
3769 /* and_withp: Valid if flags & SCF_DO_STCLASS_OR */
3771 /* There must be at least this number of characters to match */
3774 regnode *scan = *scanp, *next;
3776 int is_inf = (flags & SCF_DO_SUBSTR) && (data->flags & SF_IS_INF);
3777 int is_inf_internal = 0; /* The studied chunk is infinite */
3778 I32 is_par = OP(scan) == OPEN ? ARG(scan) : 0;
3779 scan_data_t data_fake;
3780 SV *re_trie_maxbuff = NULL;
3781 regnode *first_non_open = scan;
3782 SSize_t stopmin = SSize_t_MAX;
3783 scan_frame *frame = NULL;
3784 GET_RE_DEBUG_FLAGS_DECL;
3786 PERL_ARGS_ASSERT_STUDY_CHUNK;
3790 while (first_non_open && OP(first_non_open) == OPEN)
3791 first_non_open=regnext(first_non_open);
3797 RExC_study_chunk_recursed_count++;
3799 DEBUG_OPTIMISE_MORE_r(
3801 PerlIO_printf(Perl_debug_log,
3802 "%*sstudy_chunk stopparen=%ld recursed_count=%lu depth=%lu recursed_depth=%lu scan=%p last=%p",
3803 (int)(depth*2), "", (long)stopparen,
3804 (unsigned long)RExC_study_chunk_recursed_count,
3805 (unsigned long)depth, (unsigned long)recursed_depth,
3808 if (recursed_depth) {
3811 for ( j = 0 ; j < recursed_depth ; j++ ) {
3812 for ( i = 0 ; i < (U32)RExC_npar ; i++ ) {
3814 PAREN_TEST(RExC_study_chunk_recursed +
3815 ( j * RExC_study_chunk_recursed_bytes), i )
3818 !PAREN_TEST(RExC_study_chunk_recursed +
3819 (( j - 1 ) * RExC_study_chunk_recursed_bytes), i)
3822 PerlIO_printf(Perl_debug_log," %d",(int)i);
3826 if ( j + 1 < recursed_depth ) {
3827 PerlIO_printf(Perl_debug_log, ",");
3831 PerlIO_printf(Perl_debug_log,"\n");
3834 while ( scan && OP(scan) != END && scan < last ){
3835 UV min_subtract = 0; /* How mmany chars to subtract from the minimum
3836 node length to get a real minimum (because
3837 the folded version may be shorter) */
3838 bool unfolded_multi_char = FALSE;
3839 /* Peephole optimizer: */
3840 DEBUG_STUDYDATA("Peep:", data, depth);
3841 DEBUG_PEEP("Peep", scan, depth);
3844 /* The reason we do this here we need to deal with things like /(?:f)(?:o)(?:o)/
3845 * which cant be dealt with by the normal EXACT parsing code, as each (?:..) is handled
3846 * by a different invocation of reg() -- Yves
3848 JOIN_EXACT(scan,&min_subtract, &unfolded_multi_char, 0);
3850 /* Follow the next-chain of the current node and optimize
3851 away all the NOTHINGs from it. */
3852 if (OP(scan) != CURLYX) {
3853 const int max = (reg_off_by_arg[OP(scan)]
3855 /* I32 may be smaller than U16 on CRAYs! */
3856 : (I32_MAX < U16_MAX ? I32_MAX : U16_MAX));
3857 int off = (reg_off_by_arg[OP(scan)] ? ARG(scan) : NEXT_OFF(scan));
3861 /* Skip NOTHING and LONGJMP. */
3862 while ((n = regnext(n))
3863 && ((PL_regkind[OP(n)] == NOTHING && (noff = NEXT_OFF(n)))
3864 || ((OP(n) == LONGJMP) && (noff = ARG(n))))
3865 && off + noff < max)
3867 if (reg_off_by_arg[OP(scan)])
3870 NEXT_OFF(scan) = off;
3873 /* The principal pseudo-switch. Cannot be a switch, since we
3874 look into several different things. */
3875 if ( OP(scan) == DEFINEP ) {
3877 SSize_t deltanext = 0;
3878 SSize_t fake_last_close = 0;
3879 I32 f = SCF_IN_DEFINE;
3881 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3882 scan = regnext(scan);
3883 assert( OP(scan) == IFTHEN );
3884 DEBUG_PEEP("expect IFTHEN", scan, depth);
3886 data_fake.last_closep= &fake_last_close;
3888 next = regnext(scan);
3889 scan = NEXTOPER(NEXTOPER(scan));
3890 DEBUG_PEEP("scan", scan, depth);
3891 DEBUG_PEEP("next", next, depth);
3893 /* we suppose the run is continuous, last=next...
3894 * NOTE we dont use the return here! */
3895 (void)study_chunk(pRExC_state, &scan, &minlen,
3896 &deltanext, next, &data_fake, stopparen,
3897 recursed_depth, NULL, f, depth+1);
3902 OP(scan) == BRANCH ||
3903 OP(scan) == BRANCHJ ||
3906 next = regnext(scan);
3909 /* The op(next)==code check below is to see if we
3910 * have "BRANCH-BRANCH", "BRANCHJ-BRANCHJ", "IFTHEN-IFTHEN"
3911 * IFTHEN is special as it might not appear in pairs.
3912 * Not sure whether BRANCH-BRANCHJ is possible, regardless
3913 * we dont handle it cleanly. */
3914 if (OP(next) == code || code == IFTHEN) {
3915 /* NOTE - There is similar code to this block below for
3916 * handling TRIE nodes on a re-study. If you change stuff here
3917 * check there too. */
3918 SSize_t max1 = 0, min1 = SSize_t_MAX, num = 0;
3920 regnode * const startbranch=scan;
3922 if (flags & SCF_DO_SUBSTR) {
3923 /* Cannot merge strings after this. */
3924 scan_commit(pRExC_state, data, minlenp, is_inf);
3927 if (flags & SCF_DO_STCLASS)
3928 ssc_init_zero(pRExC_state, &accum);
3930 while (OP(scan) == code) {
3931 SSize_t deltanext, minnext, fake;
3933 regnode_ssc this_class;
3935 DEBUG_PEEP("Branch", scan, depth);
3938 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
3940 data_fake.whilem_c = data->whilem_c;
3941 data_fake.last_closep = data->last_closep;
3944 data_fake.last_closep = &fake;
3946 data_fake.pos_delta = delta;
3947 next = regnext(scan);
3949 scan = NEXTOPER(scan); /* everything */
3950 if (code != BRANCH) /* everything but BRANCH */
3951 scan = NEXTOPER(scan);
3953 if (flags & SCF_DO_STCLASS) {
3954 ssc_init(pRExC_state, &this_class);
3955 data_fake.start_class = &this_class;
3956 f = SCF_DO_STCLASS_AND;
3958 if (flags & SCF_WHILEM_VISITED_POS)
3959 f |= SCF_WHILEM_VISITED_POS;
3961 /* we suppose the run is continuous, last=next...*/
3962 minnext = study_chunk(pRExC_state, &scan, minlenp,
3963 &deltanext, next, &data_fake, stopparen,
3964 recursed_depth, NULL, f,depth+1);
3968 if (deltanext == SSize_t_MAX) {
3969 is_inf = is_inf_internal = 1;
3971 } else if (max1 < minnext + deltanext)
3972 max1 = minnext + deltanext;
3974 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
3976 if (data_fake.flags & SCF_SEEN_ACCEPT) {
3977 if ( stopmin > minnext)
3978 stopmin = min + min1;
3979 flags &= ~SCF_DO_SUBSTR;
3981 data->flags |= SCF_SEEN_ACCEPT;
3984 if (data_fake.flags & SF_HAS_EVAL)
3985 data->flags |= SF_HAS_EVAL;
3986 data->whilem_c = data_fake.whilem_c;
3988 if (flags & SCF_DO_STCLASS)
3989 ssc_or(pRExC_state, &accum, (regnode_charclass*)&this_class);
3991 if (code == IFTHEN && num < 2) /* Empty ELSE branch */
3993 if (flags & SCF_DO_SUBSTR) {
3994 data->pos_min += min1;
3995 if (data->pos_delta >= SSize_t_MAX - (max1 - min1))
3996 data->pos_delta = SSize_t_MAX;
3998 data->pos_delta += max1 - min1;
3999 if (max1 != min1 || is_inf)
4000 data->longest = &(data->longest_float);
4003 if (delta == SSize_t_MAX
4004 || SSize_t_MAX - delta - (max1 - min1) < 0)
4005 delta = SSize_t_MAX;
4007 delta += max1 - min1;
4008 if (flags & SCF_DO_STCLASS_OR) {
4009 ssc_or(pRExC_state, data->start_class, (regnode_charclass*) &accum);
4011 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4012 flags &= ~SCF_DO_STCLASS;
4015 else if (flags & SCF_DO_STCLASS_AND) {
4017 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
4018 flags &= ~SCF_DO_STCLASS;
4021 /* Switch to OR mode: cache the old value of
4022 * data->start_class */
4024 StructCopy(data->start_class, and_withp, regnode_ssc);
4025 flags &= ~SCF_DO_STCLASS_AND;
4026 StructCopy(&accum, data->start_class, regnode_ssc);
4027 flags |= SCF_DO_STCLASS_OR;
4031 if (PERL_ENABLE_TRIE_OPTIMISATION &&
4032 OP( startbranch ) == BRANCH )
4036 Assuming this was/is a branch we are dealing with: 'scan'
4037 now points at the item that follows the branch sequence,
4038 whatever it is. We now start at the beginning of the
4039 sequence and look for subsequences of
4045 which would be constructed from a pattern like
4048 If we can find such a subsequence we need to turn the first
4049 element into a trie and then add the subsequent branch exact
4050 strings to the trie.
4054 1. patterns where the whole set of branches can be
4057 2. patterns where only a subset can be converted.
4059 In case 1 we can replace the whole set with a single regop
4060 for the trie. In case 2 we need to keep the start and end
4063 'BRANCH EXACT; BRANCH EXACT; BRANCH X'
4064 becomes BRANCH TRIE; BRANCH X;
4066 There is an additional case, that being where there is a
4067 common prefix, which gets split out into an EXACT like node
4068 preceding the TRIE node.
4070 If x(1..n)==tail then we can do a simple trie, if not we make
4071 a "jump" trie, such that when we match the appropriate word
4072 we "jump" to the appropriate tail node. Essentially we turn
4073 a nested if into a case structure of sorts.
4078 if (!re_trie_maxbuff) {
4079 re_trie_maxbuff = get_sv(RE_TRIE_MAXBUF_NAME, 1);
4080 if (!SvIOK(re_trie_maxbuff))
4081 sv_setiv(re_trie_maxbuff, RE_TRIE_MAXBUF_INIT);
4083 if ( SvIV(re_trie_maxbuff)>=0 ) {
4085 regnode *first = (regnode *)NULL;
4086 regnode *last = (regnode *)NULL;
4087 regnode *tail = scan;
4091 /* var tail is used because there may be a TAIL
4092 regop in the way. Ie, the exacts will point to the
4093 thing following the TAIL, but the last branch will
4094 point at the TAIL. So we advance tail. If we
4095 have nested (?:) we may have to move through several
4099 while ( OP( tail ) == TAIL ) {
4100 /* this is the TAIL generated by (?:) */
4101 tail = regnext( tail );
4105 DEBUG_TRIE_COMPILE_r({
4106 regprop(RExC_rx, RExC_mysv, tail, NULL, pRExC_state);
4107 PerlIO_printf( Perl_debug_log, "%*s%s%s\n",
4108 (int)depth * 2 + 2, "",
4109 "Looking for TRIE'able sequences. Tail node is: ",
4110 SvPV_nolen_const( RExC_mysv )
4116 Step through the branches
4117 cur represents each branch,
4118 noper is the first thing to be matched as part
4120 noper_next is the regnext() of that node.
4122 We normally handle a case like this
4123 /FOO[xyz]|BAR[pqr]/ via a "jump trie" but we also
4124 support building with NOJUMPTRIE, which restricts
4125 the trie logic to structures like /FOO|BAR/.
4127 If noper is a trieable nodetype then the branch is
4128 a possible optimization target. If we are building
4129 under NOJUMPTRIE then we require that noper_next is
4130 the same as scan (our current position in the regex
4133 Once we have two or more consecutive such branches
4134 we can create a trie of the EXACT's contents and
4135 stitch it in place into the program.
4137 If the sequence represents all of the branches in
4138 the alternation we replace the entire thing with a
4141 Otherwise when it is a subsequence we need to
4142 stitch it in place and replace only the relevant
4143 branches. This means the first branch has to remain
4144 as it is used by the alternation logic, and its
4145 next pointer, and needs to be repointed at the item
4146 on the branch chain following the last branch we
4147 have optimized away.
4149 This could be either a BRANCH, in which case the
4150 subsequence is internal, or it could be the item
4151 following the branch sequence in which case the
4152 subsequence is at the end (which does not
4153 necessarily mean the first node is the start of the
4156 TRIE_TYPE(X) is a define which maps the optype to a
4160 ----------------+-----------
4164 EXACTFU_SS | EXACTFU
4167 EXACTFLU8 | EXACTFLU8
4171 #define TRIE_TYPE(X) ( ( NOTHING == (X) ) \
4173 : ( EXACT == (X) ) \
4175 : ( EXACTFU == (X) || EXACTFU_SS == (X) ) \
4177 : ( EXACTFA == (X) ) \
4179 : ( EXACTL == (X) ) \
4181 : ( EXACTFLU8 == (X) ) \
4185 /* dont use tail as the end marker for this traverse */
4186 for ( cur = startbranch ; cur != scan ; cur = regnext( cur ) ) {
4187 regnode * const noper = NEXTOPER( cur );
4188 U8 noper_type = OP( noper );
4189 U8 noper_trietype = TRIE_TYPE( noper_type );
4190 #if defined(DEBUGGING) || defined(NOJUMPTRIE)
4191 regnode * const noper_next = regnext( noper );
4192 U8 noper_next_type = (noper_next && noper_next != tail) ? OP(noper_next) : 0;
4193 U8 noper_next_trietype = (noper_next && noper_next != tail) ? TRIE_TYPE( noper_next_type ) :0;
4196 DEBUG_TRIE_COMPILE_r({
4197 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4198 PerlIO_printf( Perl_debug_log, "%*s- %s (%d)",
4199 (int)depth * 2 + 2,"", SvPV_nolen_const( RExC_mysv ), REG_NODE_NUM(cur) );
4201 regprop(RExC_rx, RExC_mysv, noper, NULL, pRExC_state);
4202 PerlIO_printf( Perl_debug_log, " -> %s",
4203 SvPV_nolen_const(RExC_mysv));
4206 regprop(RExC_rx, RExC_mysv, noper_next, NULL, pRExC_state);
4207 PerlIO_printf( Perl_debug_log,"\t=> %s\t",
4208 SvPV_nolen_const(RExC_mysv));
4210 PerlIO_printf( Perl_debug_log, "(First==%d,Last==%d,Cur==%d,tt==%s,nt==%s,nnt==%s)\n",
4211 REG_NODE_NUM(first), REG_NODE_NUM(last), REG_NODE_NUM(cur),
4212 PL_reg_name[trietype], PL_reg_name[noper_trietype], PL_reg_name[noper_next_trietype]
4216 /* Is noper a trieable nodetype that can be merged
4217 * with the current trie (if there is one)? */
4221 ( noper_trietype == NOTHING)
4222 || ( trietype == NOTHING )
4223 || ( trietype == noper_trietype )
4226 && noper_next == tail
4230 /* Handle mergable triable node Either we are
4231 * the first node in a new trieable sequence,
4232 * in which case we do some bookkeeping,
4233 * otherwise we update the end pointer. */
4236 if ( noper_trietype == NOTHING ) {
4237 #if !defined(DEBUGGING) && !defined(NOJUMPTRIE)
4238 regnode * const noper_next = regnext( noper );
4239 U8 noper_next_type = (noper_next && noper_next!=tail) ? OP(noper_next) : 0;
4240 U8 noper_next_trietype = noper_next_type ? TRIE_TYPE( noper_next_type ) :0;
4243 if ( noper_next_trietype ) {
4244 trietype = noper_next_trietype;
4245 } else if (noper_next_type) {
4246 /* a NOTHING regop is 1 regop wide.
4247 * We need at least two for a trie
4248 * so we can't merge this in */
4252 trietype = noper_trietype;
4255 if ( trietype == NOTHING )
4256 trietype = noper_trietype;
4261 } /* end handle mergable triable node */
4263 /* handle unmergable node -
4264 * noper may either be a triable node which can
4265 * not be tried together with the current trie,
4266 * or a non triable node */
4268 /* If last is set and trietype is not
4269 * NOTHING then we have found at least two
4270 * triable branch sequences in a row of a
4271 * similar trietype so we can turn them
4272 * into a trie. If/when we allow NOTHING to
4273 * start a trie sequence this condition
4274 * will be required, and it isn't expensive
4275 * so we leave it in for now. */
4276 if ( trietype && trietype != NOTHING )
4277 make_trie( pRExC_state,
4278 startbranch, first, cur, tail,
4279 count, trietype, depth+1 );
4280 last = NULL; /* note: we clear/update
4281 first, trietype etc below,
4282 so we dont do it here */
4286 && noper_next == tail
4289 /* noper is triable, so we can start a new
4293 trietype = noper_trietype;
4295 /* if we already saw a first but the
4296 * current node is not triable then we have
4297 * to reset the first information. */
4302 } /* end handle unmergable node */
4303 } /* loop over branches */
4304 DEBUG_TRIE_COMPILE_r({
4305 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4306 PerlIO_printf( Perl_debug_log,
4307 "%*s- %s (%d) <SCAN FINISHED>\n",
4309 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4312 if ( last && trietype ) {
4313 if ( trietype != NOTHING ) {
4314 /* the last branch of the sequence was part of
4315 * a trie, so we have to construct it here
4316 * outside of the loop */
4317 made= make_trie( pRExC_state, startbranch,
4318 first, scan, tail, count,
4319 trietype, depth+1 );
4320 #ifdef TRIE_STUDY_OPT
4321 if ( ((made == MADE_EXACT_TRIE &&
4322 startbranch == first)
4323 || ( first_non_open == first )) &&
4325 flags |= SCF_TRIE_RESTUDY;
4326 if ( startbranch == first
4329 RExC_seen &=~REG_TOP_LEVEL_BRANCHES_SEEN;
4334 /* at this point we know whatever we have is a
4335 * NOTHING sequence/branch AND if 'startbranch'
4336 * is 'first' then we can turn the whole thing
4339 if ( startbranch == first ) {
4341 /* the entire thing is a NOTHING sequence,
4342 * something like this: (?:|) So we can
4343 * turn it into a plain NOTHING op. */
4344 DEBUG_TRIE_COMPILE_r({
4345 regprop(RExC_rx, RExC_mysv, cur, NULL, pRExC_state);
4346 PerlIO_printf( Perl_debug_log,
4347 "%*s- %s (%d) <NOTHING BRANCH SEQUENCE>\n", (int)depth * 2 + 2,
4348 "", SvPV_nolen_const( RExC_mysv ),REG_NODE_NUM(cur));
4351 OP(startbranch)= NOTHING;
4352 NEXT_OFF(startbranch)= tail - startbranch;
4353 for ( opt= startbranch + 1; opt < tail ; opt++ )
4357 } /* end if ( last) */
4358 } /* TRIE_MAXBUF is non zero */
4363 else if ( code == BRANCHJ ) { /* single branch is optimized. */
4364 scan = NEXTOPER(NEXTOPER(scan));
4365 } else /* single branch is optimized. */
4366 scan = NEXTOPER(scan);
4368 } else if (OP(scan) == SUSPEND || OP(scan) == GOSUB || OP(scan) == GOSTART) {
4370 regnode *start = NULL;
4371 regnode *end = NULL;
4372 U32 my_recursed_depth= recursed_depth;
4375 if (OP(scan) != SUSPEND) { /* GOSUB/GOSTART */
4376 /* Do setup, note this code has side effects beyond
4377 * the rest of this block. Specifically setting
4378 * RExC_recurse[] must happen at least once during
4380 if (OP(scan) == GOSUB) {
4382 RExC_recurse[ARG2L(scan)] = scan;
4383 start = RExC_open_parens[paren-1];
4384 end = RExC_close_parens[paren-1];
4386 start = RExC_rxi->program + 1;
4389 /* NOTE we MUST always execute the above code, even
4390 * if we do nothing with a GOSUB/GOSTART */
4392 ( flags & SCF_IN_DEFINE )
4395 (is_inf_internal || is_inf || (data && data->flags & SF_IS_INF))
4397 ( (flags & (SCF_DO_STCLASS | SCF_DO_SUBSTR)) == 0 )
4400 /* no need to do anything here if we are in a define. */
4401 /* or we are after some kind of infinite construct
4402 * so we can skip recursing into this item.
4403 * Since it is infinite we will not change the maxlen
4404 * or delta, and if we miss something that might raise
4405 * the minlen it will merely pessimise a little.
4407 * Iow /(?(DEFINE)(?<foo>foo|food))a+(?&foo)/
4408 * might result in a minlen of 1 and not of 4,
4409 * but this doesn't make us mismatch, just try a bit
4410 * harder than we should.
4412 scan= regnext(scan);
4419 !PAREN_TEST(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes), paren)
4421 /* it is quite possible that there are more efficient ways
4422 * to do this. We maintain a bitmap per level of recursion
4423 * of which patterns we have entered so we can detect if a
4424 * pattern creates a possible infinite loop. When we
4425 * recurse down a level we copy the previous levels bitmap
4426 * down. When we are at recursion level 0 we zero the top
4427 * level bitmap. It would be nice to implement a different
4428 * more efficient way of doing this. In particular the top
4429 * level bitmap may be unnecessary.
4431 if (!recursed_depth) {
4432 Zero(RExC_study_chunk_recursed, RExC_study_chunk_recursed_bytes, U8);
4434 Copy(RExC_study_chunk_recursed + ((recursed_depth-1) * RExC_study_chunk_recursed_bytes),
4435 RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes),
4436 RExC_study_chunk_recursed_bytes, U8);
4438 /* we havent recursed into this paren yet, so recurse into it */
4439 DEBUG_STUDYDATA("set:", data,depth);
4440 PAREN_SET(RExC_study_chunk_recursed + (recursed_depth * RExC_study_chunk_recursed_bytes), paren);
4441 my_recursed_depth= recursed_depth + 1;
4443 DEBUG_STUDYDATA("inf:", data,depth);
4444 /* some form of infinite recursion, assume infinite length
4446 if (flags & SCF_DO_SUBSTR) {
4447 scan_commit(pRExC_state, data, minlenp, is_inf);
4448 data->longest = &(data->longest_float);
4450 is_inf = is_inf_internal = 1;
4451 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
4452 ssc_anything(data->start_class);
4453 flags &= ~SCF_DO_STCLASS;
4455 start= NULL; /* reset start so we dont recurse later on. */
4460 end = regnext(scan);
4463 scan_frame *newframe;
4465 if (!RExC_frame_last) {
4466 Newxz(newframe, 1, scan_frame);
4467 SAVEDESTRUCTOR_X(S_unwind_scan_frames, newframe);
4468 RExC_frame_head= newframe;
4470 } else if (!RExC_frame_last->next_frame) {
4471 Newxz(newframe,1,scan_frame);
4472 RExC_frame_last->next_frame= newframe;
4473 newframe->prev_frame= RExC_frame_last;
4476 newframe= RExC_frame_last->next_frame;
4478 RExC_frame_last= newframe;
4480 newframe->next_regnode = regnext(scan);
4481 newframe->last_regnode = last;
4482 newframe->stopparen = stopparen;
4483 newframe->prev_recursed_depth = recursed_depth;
4484 newframe->this_prev_frame= frame;
4486 DEBUG_STUDYDATA("frame-new:",data,depth);
4487 DEBUG_PEEP("fnew", scan, depth);
4494 recursed_depth= my_recursed_depth;
4499 else if (OP(scan) == EXACT || OP(scan) == EXACTL) {
4500 SSize_t l = STR_LEN(scan);
4503 const U8 * const s = (U8*)STRING(scan);
4504 uc = utf8_to_uvchr_buf(s, s + l, NULL);
4505 l = utf8_length(s, s + l);
4507 uc = *((U8*)STRING(scan));
4510 if (flags & SCF_DO_SUBSTR) { /* Update longest substr. */
4511 /* The code below prefers earlier match for fixed
4512 offset, later match for variable offset. */
4513 if (data->last_end == -1) { /* Update the start info. */
4514 data->last_start_min = data->pos_min;
4515 data->last_start_max = is_inf
4516 ? SSize_t_MAX : data->pos_min + data->pos_delta;
4518 sv_catpvn(data->last_found, STRING(scan), STR_LEN(scan));
4520 SvUTF8_on(data->last_found);
4522 SV * const sv = data->last_found;
4523 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4524 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4525 if (mg && mg->mg_len >= 0)
4526 mg->mg_len += utf8_length((U8*)STRING(scan),
4527 (U8*)STRING(scan)+STR_LEN(scan));
4529 data->last_end = data->pos_min + l;
4530 data->pos_min += l; /* As in the first entry. */
4531 data->flags &= ~SF_BEFORE_EOL;
4534 /* ANDing the code point leaves at most it, and not in locale, and
4535 * can't match null string */
4536 if (flags & SCF_DO_STCLASS_AND) {
4537 ssc_cp_and(data->start_class, uc);
4538 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4539 ssc_clear_locale(data->start_class);
4541 else if (flags & SCF_DO_STCLASS_OR) {
4542 ssc_add_cp(data->start_class, uc);
4543 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4545 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4546 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4548 flags &= ~SCF_DO_STCLASS;
4550 else if (PL_regkind[OP(scan)] == EXACT) {
4551 /* But OP != EXACT!, so is EXACTFish */
4552 SSize_t l = STR_LEN(scan);
4553 const U8 * s = (U8*)STRING(scan);
4555 /* Search for fixed substrings supports EXACT only. */
4556 if (flags & SCF_DO_SUBSTR) {
4558 scan_commit(pRExC_state, data, minlenp, is_inf);
4561 l = utf8_length(s, s + l);
4563 if (unfolded_multi_char) {
4564 RExC_seen |= REG_UNFOLDED_MULTI_SEEN;
4566 min += l - min_subtract;
4568 delta += min_subtract;
4569 if (flags & SCF_DO_SUBSTR) {
4570 data->pos_min += l - min_subtract;
4571 if (data->pos_min < 0) {
4574 data->pos_delta += min_subtract;
4576 data->longest = &(data->longest_float);
4580 if (flags & SCF_DO_STCLASS) {
4581 SV* EXACTF_invlist = _make_exactf_invlist(pRExC_state, scan);
4583 assert(EXACTF_invlist);
4584 if (flags & SCF_DO_STCLASS_AND) {
4585 if (OP(scan) != EXACTFL)
4586 ssc_clear_locale(data->start_class);
4587 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4588 ANYOF_POSIXL_ZERO(data->start_class);
4589 ssc_intersection(data->start_class, EXACTF_invlist, FALSE);
4591 else { /* SCF_DO_STCLASS_OR */
4592 ssc_union(data->start_class, EXACTF_invlist, FALSE);
4593 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4595 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
4596 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
4598 flags &= ~SCF_DO_STCLASS;
4599 SvREFCNT_dec(EXACTF_invlist);
4602 else if (REGNODE_VARIES(OP(scan))) {
4603 SSize_t mincount, maxcount, minnext, deltanext, pos_before = 0;
4604 I32 fl = 0, f = flags;
4605 regnode * const oscan = scan;
4606 regnode_ssc this_class;
4607 regnode_ssc *oclass = NULL;
4608 I32 next_is_eval = 0;
4610 switch (PL_regkind[OP(scan)]) {
4611 case WHILEM: /* End of (?:...)* . */
4612 scan = NEXTOPER(scan);
4615 if (flags & (SCF_DO_SUBSTR | SCF_DO_STCLASS)) {
4616 next = NEXTOPER(scan);
4617 if (OP(next) == EXACT
4618 || OP(next) == EXACTL
4619 || (flags & SCF_DO_STCLASS))
4622 maxcount = REG_INFTY;
4623 next = regnext(scan);
4624 scan = NEXTOPER(scan);
4628 if (flags & SCF_DO_SUBSTR)
4633 if (flags & SCF_DO_STCLASS) {
4635 maxcount = REG_INFTY;
4636 next = regnext(scan);
4637 scan = NEXTOPER(scan);
4640 if (flags & SCF_DO_SUBSTR) {
4641 scan_commit(pRExC_state, data, minlenp, is_inf);
4642 /* Cannot extend fixed substrings */
4643 data->longest = &(data->longest_float);
4645 is_inf = is_inf_internal = 1;
4646 scan = regnext(scan);
4647 goto optimize_curly_tail;
4649 if (stopparen>0 && (OP(scan)==CURLYN || OP(scan)==CURLYM)
4650 && (scan->flags == stopparen))
4655 mincount = ARG1(scan);
4656 maxcount = ARG2(scan);
4658 next = regnext(scan);
4659 if (OP(scan) == CURLYX) {
4660 I32 lp = (data ? *(data->last_closep) : 0);
4661 scan->flags = ((lp <= (I32)U8_MAX) ? (U8)lp : U8_MAX);
4663 scan = NEXTOPER(scan) + EXTRA_STEP_2ARGS;
4664 next_is_eval = (OP(scan) == EVAL);
4666 if (flags & SCF_DO_SUBSTR) {
4668 scan_commit(pRExC_state, data, minlenp, is_inf);
4669 /* Cannot extend fixed substrings */
4670 pos_before = data->pos_min;
4674 data->flags &= ~(SF_HAS_PAR|SF_IN_PAR|SF_HAS_EVAL);
4676 data->flags |= SF_IS_INF;
4678 if (flags & SCF_DO_STCLASS) {
4679 ssc_init(pRExC_state, &this_class);
4680 oclass = data->start_class;
4681 data->start_class = &this_class;
4682 f |= SCF_DO_STCLASS_AND;
4683 f &= ~SCF_DO_STCLASS_OR;
4685 /* Exclude from super-linear cache processing any {n,m}
4686 regops for which the combination of input pos and regex
4687 pos is not enough information to determine if a match
4690 For example, in the regex /foo(bar\s*){4,8}baz/ with the
4691 regex pos at the \s*, the prospects for a match depend not
4692 only on the input position but also on how many (bar\s*)
4693 repeats into the {4,8} we are. */
4694 if ((mincount > 1) || (maxcount > 1 && maxcount != REG_INFTY))
4695 f &= ~SCF_WHILEM_VISITED_POS;
4697 /* This will finish on WHILEM, setting scan, or on NULL: */
4698 minnext = study_chunk(pRExC_state, &scan, minlenp, &deltanext,
4699 last, data, stopparen, recursed_depth, NULL,
4701 ? (f & ~SCF_DO_SUBSTR)
4705 if (flags & SCF_DO_STCLASS)
4706 data->start_class = oclass;
4707 if (mincount == 0 || minnext == 0) {
4708 if (flags & SCF_DO_STCLASS_OR) {
4709 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4711 else if (flags & SCF_DO_STCLASS_AND) {
4712 /* Switch to OR mode: cache the old value of
4713 * data->start_class */
4715 StructCopy(data->start_class, and_withp, regnode_ssc);
4716 flags &= ~SCF_DO_STCLASS_AND;
4717 StructCopy(&this_class, data->start_class, regnode_ssc);
4718 flags |= SCF_DO_STCLASS_OR;
4719 ANYOF_FLAGS(data->start_class)
4720 |= SSC_MATCHES_EMPTY_STRING;
4722 } else { /* Non-zero len */
4723 if (flags & SCF_DO_STCLASS_OR) {
4724 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4725 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
4727 else if (flags & SCF_DO_STCLASS_AND)
4728 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &this_class);
4729 flags &= ~SCF_DO_STCLASS;
4731 if (!scan) /* It was not CURLYX, but CURLY. */
4733 if (!(flags & SCF_TRIE_DOING_RESTUDY)
4734 /* ? quantifier ok, except for (?{ ... }) */
4735 && (next_is_eval || !(mincount == 0 && maxcount == 1))
4736 && (minnext == 0) && (deltanext == 0)
4737 && data && !(data->flags & (SF_HAS_PAR|SF_IN_PAR))
4738 && maxcount <= REG_INFTY/3) /* Complement check for big
4741 /* Fatal warnings may leak the regexp without this: */
4742 SAVEFREESV(RExC_rx_sv);
4743 Perl_ck_warner(aTHX_ packWARN(WARN_REGEXP),
4744 "Quantifier unexpected on zero-length expression "
4745 "in regex m/%"UTF8f"/",
4746 UTF8fARG(UTF, RExC_end - RExC_precomp,
4748 (void)ReREFCNT_inc(RExC_rx_sv);
4751 min += minnext * mincount;
4752 is_inf_internal |= deltanext == SSize_t_MAX
4753 || (maxcount == REG_INFTY && minnext + deltanext > 0);
4754 is_inf |= is_inf_internal;
4756 delta = SSize_t_MAX;
4758 delta += (minnext + deltanext) * maxcount
4759 - minnext * mincount;
4761 /* Try powerful optimization CURLYX => CURLYN. */
4762 if ( OP(oscan) == CURLYX && data
4763 && data->flags & SF_IN_PAR
4764 && !(data->flags & SF_HAS_EVAL)
4765 && !deltanext && minnext == 1 ) {
4766 /* Try to optimize to CURLYN. */
4767 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS;
4768 regnode * const nxt1 = nxt;
4775 if (!REGNODE_SIMPLE(OP(nxt))
4776 && !(PL_regkind[OP(nxt)] == EXACT
4777 && STR_LEN(nxt) == 1))
4783 if (OP(nxt) != CLOSE)
4785 if (RExC_open_parens) {
4786 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4787 RExC_close_parens[ARG(nxt1)-1]=nxt+2; /*close->while*/
4789 /* Now we know that nxt2 is the only contents: */
4790 oscan->flags = (U8)ARG(nxt);
4792 OP(nxt1) = NOTHING; /* was OPEN. */
4795 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4796 NEXT_OFF(nxt1+ 1) = 0; /* just for consistency. */
4797 NEXT_OFF(nxt2) = 0; /* just for consistency with CURLY. */
4798 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4799 OP(nxt + 1) = OPTIMIZED; /* was count. */
4800 NEXT_OFF(nxt+ 1) = 0; /* just for consistency. */
4805 /* Try optimization CURLYX => CURLYM. */
4806 if ( OP(oscan) == CURLYX && data
4807 && !(data->flags & SF_HAS_PAR)
4808 && !(data->flags & SF_HAS_EVAL)
4809 && !deltanext /* atom is fixed width */
4810 && minnext != 0 /* CURLYM can't handle zero width */
4812 /* Nor characters whose fold at run-time may be
4813 * multi-character */
4814 && ! (RExC_seen & REG_UNFOLDED_MULTI_SEEN)
4816 /* XXXX How to optimize if data == 0? */
4817 /* Optimize to a simpler form. */
4818 regnode *nxt = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN */
4822 while ( (nxt2 = regnext(nxt)) /* skip over embedded stuff*/
4823 && (OP(nxt2) != WHILEM))
4825 OP(nxt2) = SUCCEED; /* Whas WHILEM */
4826 /* Need to optimize away parenths. */
4827 if ((data->flags & SF_IN_PAR) && OP(nxt) == CLOSE) {
4828 /* Set the parenth number. */
4829 regnode *nxt1 = NEXTOPER(oscan) + EXTRA_STEP_2ARGS; /* OPEN*/
4831 oscan->flags = (U8)ARG(nxt);
4832 if (RExC_open_parens) {
4833 RExC_open_parens[ARG(nxt1)-1]=oscan; /*open->CURLYM*/
4834 RExC_close_parens[ARG(nxt1)-1]=nxt2+1; /*close->NOTHING*/
4836 OP(nxt1) = OPTIMIZED; /* was OPEN. */
4837 OP(nxt) = OPTIMIZED; /* was CLOSE. */
4840 OP(nxt1 + 1) = OPTIMIZED; /* was count. */
4841 OP(nxt + 1) = OPTIMIZED; /* was count. */
4842 NEXT_OFF(nxt1 + 1) = 0; /* just for consistency. */
4843 NEXT_OFF(nxt + 1) = 0; /* just for consistency. */
4846 while ( nxt1 && (OP(nxt1) != WHILEM)) {
4847 regnode *nnxt = regnext(nxt1);
4849 if (reg_off_by_arg[OP(nxt1)])
4850 ARG_SET(nxt1, nxt2 - nxt1);
4851 else if (nxt2 - nxt1 < U16_MAX)
4852 NEXT_OFF(nxt1) = nxt2 - nxt1;
4854 OP(nxt) = NOTHING; /* Cannot beautify */
4859 /* Optimize again: */
4860 study_chunk(pRExC_state, &nxt1, minlenp, &deltanext, nxt,
4861 NULL, stopparen, recursed_depth, NULL, 0,depth+1);
4866 else if ((OP(oscan) == CURLYX)
4867 && (flags & SCF_WHILEM_VISITED_POS)
4868 /* See the comment on a similar expression above.
4869 However, this time it's not a subexpression
4870 we care about, but the expression itself. */
4871 && (maxcount == REG_INFTY)
4872 && data && ++data->whilem_c < 16) {
4873 /* This stays as CURLYX, we can put the count/of pair. */
4874 /* Find WHILEM (as in regexec.c) */
4875 regnode *nxt = oscan + NEXT_OFF(oscan);
4877 if (OP(PREVOPER(nxt)) == NOTHING) /* LONGJMP */
4879 PREVOPER(nxt)->flags = (U8)(data->whilem_c
4880 | (RExC_whilem_seen << 4)); /* On WHILEM */
4882 if (data && fl & (SF_HAS_PAR|SF_IN_PAR))
4884 if (flags & SCF_DO_SUBSTR) {
4885 SV *last_str = NULL;
4886 STRLEN last_chrs = 0;
4887 int counted = mincount != 0;
4889 if (data->last_end > 0 && mincount != 0) { /* Ends with a
4891 SSize_t b = pos_before >= data->last_start_min
4892 ? pos_before : data->last_start_min;
4894 const char * const s = SvPV_const(data->last_found, l);
4895 SSize_t old = b - data->last_start_min;
4898 old = utf8_hop((U8*)s, old) - (U8*)s;
4900 /* Get the added string: */
4901 last_str = newSVpvn_utf8(s + old, l, UTF);
4902 last_chrs = UTF ? utf8_length((U8*)(s + old),
4903 (U8*)(s + old + l)) : l;
4904 if (deltanext == 0 && pos_before == b) {
4905 /* What was added is a constant string */
4908 SvGROW(last_str, (mincount * l) + 1);
4909 repeatcpy(SvPVX(last_str) + l,
4910 SvPVX_const(last_str), l,
4912 SvCUR_set(last_str, SvCUR(last_str) * mincount);
4913 /* Add additional parts. */
4914 SvCUR_set(data->last_found,
4915 SvCUR(data->last_found) - l);
4916 sv_catsv(data->last_found, last_str);
4918 SV * sv = data->last_found;
4920 SvUTF8(sv) && SvMAGICAL(sv) ?
4921 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4922 if (mg && mg->mg_len >= 0)
4923 mg->mg_len += last_chrs * (mincount-1);
4925 last_chrs *= mincount;
4926 data->last_end += l * (mincount - 1);
4929 /* start offset must point into the last copy */
4930 data->last_start_min += minnext * (mincount - 1);
4931 data->last_start_max =
4934 : data->last_start_max +
4935 (maxcount - 1) * (minnext + data->pos_delta);
4938 /* It is counted once already... */
4939 data->pos_min += minnext * (mincount - counted);
4941 PerlIO_printf(Perl_debug_log, "counted=%"UVuf" deltanext=%"UVuf
4942 " SSize_t_MAX=%"UVuf" minnext=%"UVuf
4943 " maxcount=%"UVuf" mincount=%"UVuf"\n",
4944 (UV)counted, (UV)deltanext, (UV)SSize_t_MAX, (UV)minnext, (UV)maxcount,
4946 if (deltanext != SSize_t_MAX)
4947 PerlIO_printf(Perl_debug_log, "LHS=%"UVuf" RHS=%"UVuf"\n",
4948 (UV)(-counted * deltanext + (minnext + deltanext) * maxcount
4949 - minnext * mincount), (UV)(SSize_t_MAX - data->pos_delta));
4951 if (deltanext == SSize_t_MAX
4952 || -counted * deltanext + (minnext + deltanext) * maxcount - minnext * mincount >= SSize_t_MAX - data->pos_delta)
4953 data->pos_delta = SSize_t_MAX;
4955 data->pos_delta += - counted * deltanext +
4956 (minnext + deltanext) * maxcount - minnext * mincount;
4957 if (mincount != maxcount) {
4958 /* Cannot extend fixed substrings found inside
4960 scan_commit(pRExC_state, data, minlenp, is_inf);
4961 if (mincount && last_str) {
4962 SV * const sv = data->last_found;
4963 MAGIC * const mg = SvUTF8(sv) && SvMAGICAL(sv) ?
4964 mg_find(sv, PERL_MAGIC_utf8) : NULL;
4968 sv_setsv(sv, last_str);
4969 data->last_end = data->pos_min;
4970 data->last_start_min = data->pos_min - last_chrs;
4971 data->last_start_max = is_inf
4973 : data->pos_min + data->pos_delta - last_chrs;
4975 data->longest = &(data->longest_float);
4977 SvREFCNT_dec(last_str);
4979 if (data && (fl & SF_HAS_EVAL))
4980 data->flags |= SF_HAS_EVAL;
4981 optimize_curly_tail:
4982 if (OP(oscan) != CURLYX) {
4983 while (PL_regkind[OP(next = regnext(oscan))] == NOTHING
4985 NEXT_OFF(oscan) += NEXT_OFF(next);
4991 Perl_croak(aTHX_ "panic: unexpected varying REx opcode %d",
4996 if (flags & SCF_DO_SUBSTR) {
4997 /* Cannot expect anything... */
4998 scan_commit(pRExC_state, data, minlenp, is_inf);
4999 data->longest = &(data->longest_float);
5001 is_inf = is_inf_internal = 1;
5002 if (flags & SCF_DO_STCLASS_OR) {
5003 if (OP(scan) == CLUMP) {
5004 /* Actually is any start char, but very few code points
5005 * aren't start characters */
5006 ssc_match_all_cp(data->start_class);
5009 ssc_anything(data->start_class);
5012 flags &= ~SCF_DO_STCLASS;
5016 else if (OP(scan) == LNBREAK) {
5017 if (flags & SCF_DO_STCLASS) {
5018 if (flags & SCF_DO_STCLASS_AND) {
5019 ssc_intersection(data->start_class,
5020 PL_XPosix_ptrs[_CC_VERTSPACE], FALSE);
5021 ssc_clear_locale(data->start_class);
5022 ANYOF_FLAGS(data->start_class)
5023 &= ~SSC_MATCHES_EMPTY_STRING;
5025 else if (flags & SCF_DO_STCLASS_OR) {
5026 ssc_union(data->start_class,
5027 PL_XPosix_ptrs[_CC_VERTSPACE],
5029 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5031 /* See commit msg for
5032 * 749e076fceedeb708a624933726e7989f2302f6a */
5033 ANYOF_FLAGS(data->start_class)
5034 &= ~SSC_MATCHES_EMPTY_STRING;
5036 flags &= ~SCF_DO_STCLASS;
5039 if (delta != SSize_t_MAX)
5040 delta++; /* Because of the 2 char string cr-lf */
5041 if (flags & SCF_DO_SUBSTR) {
5042 /* Cannot expect anything... */
5043 scan_commit(pRExC_state, data, minlenp, is_inf);
5045 data->pos_delta += 1;
5046 data->longest = &(data->longest_float);
5049 else if (REGNODE_SIMPLE(OP(scan))) {
5051 if (flags & SCF_DO_SUBSTR) {
5052 scan_commit(pRExC_state, data, minlenp, is_inf);
5056 if (flags & SCF_DO_STCLASS) {
5058 SV* my_invlist = NULL;
5061 /* See commit msg 749e076fceedeb708a624933726e7989f2302f6a */
5062 ANYOF_FLAGS(data->start_class) &= ~SSC_MATCHES_EMPTY_STRING;
5064 /* Some of the logic below assumes that switching
5065 locale on will only add false positives. */
5070 Perl_croak(aTHX_ "panic: unexpected simple REx opcode %d",
5075 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5076 ssc_match_all_cp(data->start_class);
5081 SV* REG_ANY_invlist = _new_invlist(2);
5082 REG_ANY_invlist = add_cp_to_invlist(REG_ANY_invlist,
5084 if (flags & SCF_DO_STCLASS_OR) {
5085 ssc_union(data->start_class,
5087 TRUE /* TRUE => invert, hence all but \n
5091 else if (flags & SCF_DO_STCLASS_AND) {
5092 ssc_intersection(data->start_class,
5094 TRUE /* TRUE => invert */
5096 ssc_clear_locale(data->start_class);
5098 SvREFCNT_dec_NN(REG_ANY_invlist);
5104 if (flags & SCF_DO_STCLASS_AND)
5105 ssc_and(pRExC_state, data->start_class,
5106 (regnode_charclass *) scan);
5108 ssc_or(pRExC_state, data->start_class,
5109 (regnode_charclass *) scan);
5117 namedclass = classnum_to_namedclass(FLAGS(scan)) + invert;
5118 if (flags & SCF_DO_STCLASS_AND) {
5119 bool was_there = cBOOL(
5120 ANYOF_POSIXL_TEST(data->start_class,
5122 ANYOF_POSIXL_ZERO(data->start_class);
5123 if (was_there) { /* Do an AND */
5124 ANYOF_POSIXL_SET(data->start_class, namedclass);
5126 /* No individual code points can now match */
5127 data->start_class->invlist
5128 = sv_2mortal(_new_invlist(0));
5131 int complement = namedclass + ((invert) ? -1 : 1);
5133 assert(flags & SCF_DO_STCLASS_OR);
5135 /* If the complement of this class was already there,
5136 * the result is that they match all code points,
5137 * (\d + \D == everything). Remove the classes from
5138 * future consideration. Locale is not relevant in
5140 if (ANYOF_POSIXL_TEST(data->start_class, complement)) {
5141 ssc_match_all_cp(data->start_class);
5142 ANYOF_POSIXL_CLEAR(data->start_class, namedclass);
5143 ANYOF_POSIXL_CLEAR(data->start_class, complement);
5145 else { /* The usual case; just add this class to the
5147 ANYOF_POSIXL_SET(data->start_class, namedclass);
5152 case NPOSIXA: /* For these, we always know the exact set of
5157 if (FLAGS(scan) == _CC_ASCII) {
5158 my_invlist = invlist_clone(PL_XPosix_ptrs[_CC_ASCII]);
5161 _invlist_intersection(PL_XPosix_ptrs[FLAGS(scan)],
5162 PL_XPosix_ptrs[_CC_ASCII],
5173 my_invlist = invlist_clone(PL_XPosix_ptrs[FLAGS(scan)]);
5175 /* NPOSIXD matches all upper Latin1 code points unless the
5176 * target string being matched is UTF-8, which is
5177 * unknowable until match time. Since we are going to
5178 * invert, we want to get rid of all of them so that the
5179 * inversion will match all */
5180 if (OP(scan) == NPOSIXD) {
5181 _invlist_subtract(my_invlist, PL_UpperLatin1,
5187 if (flags & SCF_DO_STCLASS_AND) {
5188 ssc_intersection(data->start_class, my_invlist, invert);
5189 ssc_clear_locale(data->start_class);
5192 assert(flags & SCF_DO_STCLASS_OR);
5193 ssc_union(data->start_class, my_invlist, invert);
5195 SvREFCNT_dec(my_invlist);
5197 if (flags & SCF_DO_STCLASS_OR)
5198 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5199 flags &= ~SCF_DO_STCLASS;
5202 else if (PL_regkind[OP(scan)] == EOL && flags & SCF_DO_SUBSTR) {
5203 data->flags |= (OP(scan) == MEOL
5206 scan_commit(pRExC_state, data, minlenp, is_inf);
5209 else if ( PL_regkind[OP(scan)] == BRANCHJ
5210 /* Lookbehind, or need to calculate parens/evals/stclass: */
5211 && (scan->flags || data || (flags & SCF_DO_STCLASS))
5212 && (OP(scan) == IFMATCH || OP(scan) == UNLESSM))
5214 if ( !PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5215 || OP(scan) == UNLESSM )
5217 /* Negative Lookahead/lookbehind
5218 In this case we can't do fixed string optimisation.
5221 SSize_t deltanext, minnext, fake = 0;
5226 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5228 data_fake.whilem_c = data->whilem_c;
5229 data_fake.last_closep = data->last_closep;
5232 data_fake.last_closep = &fake;
5233 data_fake.pos_delta = delta;
5234 if ( flags & SCF_DO_STCLASS && !scan->flags
5235 && OP(scan) == IFMATCH ) { /* Lookahead */
5236 ssc_init(pRExC_state, &intrnl);
5237 data_fake.start_class = &intrnl;
5238 f |= SCF_DO_STCLASS_AND;
5240 if (flags & SCF_WHILEM_VISITED_POS)
5241 f |= SCF_WHILEM_VISITED_POS;
5242 next = regnext(scan);
5243 nscan = NEXTOPER(NEXTOPER(scan));
5244 minnext = study_chunk(pRExC_state, &nscan, minlenp, &deltanext,
5245 last, &data_fake, stopparen,
5246 recursed_depth, NULL, f, depth+1);
5249 FAIL("Variable length lookbehind not implemented");
5251 else if (minnext > (I32)U8_MAX) {
5252 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5255 scan->flags = (U8)minnext;
5258 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5260 if (data_fake.flags & SF_HAS_EVAL)
5261 data->flags |= SF_HAS_EVAL;
5262 data->whilem_c = data_fake.whilem_c;
5264 if (f & SCF_DO_STCLASS_AND) {
5265 if (flags & SCF_DO_STCLASS_OR) {
5266 /* OR before, AND after: ideally we would recurse with
5267 * data_fake to get the AND applied by study of the
5268 * remainder of the pattern, and then derecurse;
5269 * *** HACK *** for now just treat as "no information".
5270 * See [perl #56690].
5272 ssc_init(pRExC_state, data->start_class);
5274 /* AND before and after: combine and continue. These
5275 * assertions are zero-length, so can match an EMPTY
5277 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5278 ANYOF_FLAGS(data->start_class)
5279 |= SSC_MATCHES_EMPTY_STRING;
5283 #if PERL_ENABLE_POSITIVE_ASSERTION_STUDY
5285 /* Positive Lookahead/lookbehind
5286 In this case we can do fixed string optimisation,
5287 but we must be careful about it. Note in the case of
5288 lookbehind the positions will be offset by the minimum
5289 length of the pattern, something we won't know about
5290 until after the recurse.
5292 SSize_t deltanext, fake = 0;
5296 /* We use SAVEFREEPV so that when the full compile
5297 is finished perl will clean up the allocated
5298 minlens when it's all done. This way we don't
5299 have to worry about freeing them when we know
5300 they wont be used, which would be a pain.
5303 Newx( minnextp, 1, SSize_t );
5304 SAVEFREEPV(minnextp);
5307 StructCopy(data, &data_fake, scan_data_t);
5308 if ((flags & SCF_DO_SUBSTR) && data->last_found) {
5311 scan_commit(pRExC_state, &data_fake, minlenp, is_inf);
5312 data_fake.last_found=newSVsv(data->last_found);
5316 data_fake.last_closep = &fake;
5317 data_fake.flags = 0;
5318 data_fake.pos_delta = delta;
5320 data_fake.flags |= SF_IS_INF;
5321 if ( flags & SCF_DO_STCLASS && !scan->flags
5322 && OP(scan) == IFMATCH ) { /* Lookahead */
5323 ssc_init(pRExC_state, &intrnl);
5324 data_fake.start_class = &intrnl;
5325 f |= SCF_DO_STCLASS_AND;
5327 if (flags & SCF_WHILEM_VISITED_POS)
5328 f |= SCF_WHILEM_VISITED_POS;
5329 next = regnext(scan);
5330 nscan = NEXTOPER(NEXTOPER(scan));
5332 *minnextp = study_chunk(pRExC_state, &nscan, minnextp,
5333 &deltanext, last, &data_fake,
5334 stopparen, recursed_depth, NULL,
5338 FAIL("Variable length lookbehind not implemented");
5340 else if (*minnextp > (I32)U8_MAX) {
5341 FAIL2("Lookbehind longer than %"UVuf" not implemented",
5344 scan->flags = (U8)*minnextp;
5349 if (f & SCF_DO_STCLASS_AND) {
5350 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &intrnl);
5351 ANYOF_FLAGS(data->start_class) |= SSC_MATCHES_EMPTY_STRING;
5354 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5356 if (data_fake.flags & SF_HAS_EVAL)
5357 data->flags |= SF_HAS_EVAL;
5358 data->whilem_c = data_fake.whilem_c;
5359 if ((flags & SCF_DO_SUBSTR) && data_fake.last_found) {
5360 if (RExC_rx->minlen<*minnextp)
5361 RExC_rx->minlen=*minnextp;
5362 scan_commit(pRExC_state, &data_fake, minnextp, is_inf);
5363 SvREFCNT_dec_NN(data_fake.last_found);
5365 if ( data_fake.minlen_fixed != minlenp )
5367 data->offset_fixed= data_fake.offset_fixed;
5368 data->minlen_fixed= data_fake.minlen_fixed;
5369 data->lookbehind_fixed+= scan->flags;
5371 if ( data_fake.minlen_float != minlenp )
5373 data->minlen_float= data_fake.minlen_float;
5374 data->offset_float_min=data_fake.offset_float_min;
5375 data->offset_float_max=data_fake.offset_float_max;
5376 data->lookbehind_float+= scan->flags;
5383 else if (OP(scan) == OPEN) {
5384 if (stopparen != (I32)ARG(scan))
5387 else if (OP(scan) == CLOSE) {
5388 if (stopparen == (I32)ARG(scan)) {
5391 if ((I32)ARG(scan) == is_par) {
5392 next = regnext(scan);
5394 if ( next && (OP(next) != WHILEM) && next < last)
5395 is_par = 0; /* Disable optimization */
5398 *(data->last_closep) = ARG(scan);
5400 else if (OP(scan) == EVAL) {
5402 data->flags |= SF_HAS_EVAL;
5404 else if ( PL_regkind[OP(scan)] == ENDLIKE ) {
5405 if (flags & SCF_DO_SUBSTR) {
5406 scan_commit(pRExC_state, data, minlenp, is_inf);
5407 flags &= ~SCF_DO_SUBSTR;
5409 if (data && OP(scan)==ACCEPT) {
5410 data->flags |= SCF_SEEN_ACCEPT;
5415 else if (OP(scan) == LOGICAL && scan->flags == 2) /* Embedded follows */
5417 if (flags & SCF_DO_SUBSTR) {
5418 scan_commit(pRExC_state, data, minlenp, is_inf);
5419 data->longest = &(data->longest_float);
5421 is_inf = is_inf_internal = 1;
5422 if (flags & SCF_DO_STCLASS_OR) /* Allow everything */
5423 ssc_anything(data->start_class);
5424 flags &= ~SCF_DO_STCLASS;
5426 else if (OP(scan) == GPOS) {
5427 if (!(RExC_rx->intflags & PREGf_GPOS_FLOAT) &&
5428 !(delta || is_inf || (data && data->pos_delta)))
5430 if (!(RExC_rx->intflags & PREGf_ANCH) && (flags & SCF_DO_SUBSTR))
5431 RExC_rx->intflags |= PREGf_ANCH_GPOS;
5432 if (RExC_rx->gofs < (STRLEN)min)
5433 RExC_rx->gofs = min;
5435 RExC_rx->intflags |= PREGf_GPOS_FLOAT;
5439 #ifdef TRIE_STUDY_OPT
5440 #ifdef FULL_TRIE_STUDY
5441 else if (PL_regkind[OP(scan)] == TRIE) {
5442 /* NOTE - There is similar code to this block above for handling
5443 BRANCH nodes on the initial study. If you change stuff here
5445 regnode *trie_node= scan;
5446 regnode *tail= regnext(scan);
5447 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5448 SSize_t max1 = 0, min1 = SSize_t_MAX;
5451 if (flags & SCF_DO_SUBSTR) { /* XXXX Add !SUSPEND? */
5452 /* Cannot merge strings after this. */
5453 scan_commit(pRExC_state, data, minlenp, is_inf);
5455 if (flags & SCF_DO_STCLASS)
5456 ssc_init_zero(pRExC_state, &accum);
5462 const regnode *nextbranch= NULL;
5465 for ( word=1 ; word <= trie->wordcount ; word++)
5467 SSize_t deltanext=0, minnext=0, f = 0, fake;
5468 regnode_ssc this_class;
5470 StructCopy(&zero_scan_data, &data_fake, scan_data_t);
5472 data_fake.whilem_c = data->whilem_c;
5473 data_fake.last_closep = data->last_closep;
5476 data_fake.last_closep = &fake;
5477 data_fake.pos_delta = delta;
5478 if (flags & SCF_DO_STCLASS) {
5479 ssc_init(pRExC_state, &this_class);
5480 data_fake.start_class = &this_class;
5481 f = SCF_DO_STCLASS_AND;
5483 if (flags & SCF_WHILEM_VISITED_POS)
5484 f |= SCF_WHILEM_VISITED_POS;
5486 if (trie->jump[word]) {
5488 nextbranch = trie_node + trie->jump[0];
5489 scan= trie_node + trie->jump[word];
5490 /* We go from the jump point to the branch that follows
5491 it. Note this means we need the vestigal unused
5492 branches even though they arent otherwise used. */
5493 minnext = study_chunk(pRExC_state, &scan, minlenp,
5494 &deltanext, (regnode *)nextbranch, &data_fake,
5495 stopparen, recursed_depth, NULL, f,depth+1);
5497 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
5498 nextbranch= regnext((regnode*)nextbranch);
5500 if (min1 > (SSize_t)(minnext + trie->minlen))
5501 min1 = minnext + trie->minlen;
5502 if (deltanext == SSize_t_MAX) {
5503 is_inf = is_inf_internal = 1;
5505 } else if (max1 < (SSize_t)(minnext + deltanext + trie->maxlen))
5506 max1 = minnext + deltanext + trie->maxlen;
5508 if (data_fake.flags & (SF_HAS_PAR|SF_IN_PAR))
5510 if (data_fake.flags & SCF_SEEN_ACCEPT) {
5511 if ( stopmin > min + min1)
5512 stopmin = min + min1;
5513 flags &= ~SCF_DO_SUBSTR;
5515 data->flags |= SCF_SEEN_ACCEPT;
5518 if (data_fake.flags & SF_HAS_EVAL)
5519 data->flags |= SF_HAS_EVAL;
5520 data->whilem_c = data_fake.whilem_c;
5522 if (flags & SCF_DO_STCLASS)
5523 ssc_or(pRExC_state, &accum, (regnode_charclass *) &this_class);
5526 if (flags & SCF_DO_SUBSTR) {
5527 data->pos_min += min1;
5528 data->pos_delta += max1 - min1;
5529 if (max1 != min1 || is_inf)
5530 data->longest = &(data->longest_float);
5533 if (delta != SSize_t_MAX)
5534 delta += max1 - min1;
5535 if (flags & SCF_DO_STCLASS_OR) {
5536 ssc_or(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5538 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5539 flags &= ~SCF_DO_STCLASS;
5542 else if (flags & SCF_DO_STCLASS_AND) {
5544 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) &accum);
5545 flags &= ~SCF_DO_STCLASS;
5548 /* Switch to OR mode: cache the old value of
5549 * data->start_class */
5551 StructCopy(data->start_class, and_withp, regnode_ssc);
5552 flags &= ~SCF_DO_STCLASS_AND;
5553 StructCopy(&accum, data->start_class, regnode_ssc);
5554 flags |= SCF_DO_STCLASS_OR;
5561 else if (PL_regkind[OP(scan)] == TRIE) {
5562 reg_trie_data *trie = (reg_trie_data*)RExC_rxi->data->data[ ARG(scan) ];
5565 min += trie->minlen;
5566 delta += (trie->maxlen - trie->minlen);
5567 flags &= ~SCF_DO_STCLASS; /* xxx */
5568 if (flags & SCF_DO_SUBSTR) {
5569 /* Cannot expect anything... */
5570 scan_commit(pRExC_state, data, minlenp, is_inf);
5571 data->pos_min += trie->minlen;
5572 data->pos_delta += (trie->maxlen - trie->minlen);
5573 if (trie->maxlen != trie->minlen)
5574 data->longest = &(data->longest_float);
5576 if (trie->jump) /* no more substrings -- for now /grr*/
5577 flags &= ~SCF_DO_SUBSTR;
5579 #endif /* old or new */
5580 #endif /* TRIE_STUDY_OPT */
5582 /* Else: zero-length, ignore. */
5583 scan = regnext(scan);
5585 /* If we are exiting a recursion we can unset its recursed bit
5586 * and allow ourselves to enter it again - no danger of an
5587 * infinite loop there.
5588 if (stopparen > -1 && recursed) {
5589 DEBUG_STUDYDATA("unset:", data,depth);
5590 PAREN_UNSET( recursed, stopparen);
5596 DEBUG_STUDYDATA("frame-end:",data,depth);
5597 DEBUG_PEEP("fend", scan, depth);
5599 /* restore previous context */
5600 last = frame->last_regnode;
5601 scan = frame->next_regnode;
5602 stopparen = frame->stopparen;
5603 recursed_depth = frame->prev_recursed_depth;
5605 RExC_frame_last = frame->prev_frame;
5606 frame = frame->this_prev_frame;
5607 goto fake_study_recurse;
5612 DEBUG_STUDYDATA("pre-fin:",data,depth);
5615 *deltap = is_inf_internal ? SSize_t_MAX : delta;
5617 if (flags & SCF_DO_SUBSTR && is_inf)
5618 data->pos_delta = SSize_t_MAX - data->pos_min;
5619 if (is_par > (I32)U8_MAX)
5621 if (is_par && pars==1 && data) {
5622 data->flags |= SF_IN_PAR;
5623 data->flags &= ~SF_HAS_PAR;
5625 else if (pars && data) {
5626 data->flags |= SF_HAS_PAR;
5627 data->flags &= ~SF_IN_PAR;
5629 if (flags & SCF_DO_STCLASS_OR)
5630 ssc_and(pRExC_state, data->start_class, (regnode_charclass *) and_withp);
5631 if (flags & SCF_TRIE_RESTUDY)
5632 data->flags |= SCF_TRIE_RESTUDY;
5634 DEBUG_STUDYDATA("post-fin:",data,depth);
5637 SSize_t final_minlen= min < stopmin ? min : stopmin;
5639 if (!(RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN)) {
5640 if (final_minlen > SSize_t_MAX - delta)
5641 RExC_maxlen = SSize_t_MAX;
5642 else if (RExC_maxlen < final_minlen + delta)
5643 RExC_maxlen = final_minlen + delta;
5645 return final_minlen;
5647 NOT_REACHED; /* NOTREACHED */
5651 S_add_data(RExC_state_t* const pRExC_state, const char* const s, const U32 n)
5653 U32 count = RExC_rxi->data ? RExC_rxi->data->count : 0;
5655 PERL_ARGS_ASSERT_ADD_DATA;
5657 Renewc(RExC_rxi->data,
5658 sizeof(*RExC_rxi->data) + sizeof(void*) * (count + n - 1),
5659 char, struct reg_data);
5661 Renew(RExC_rxi->data->what, count + n, U8);
5663 Newx(RExC_rxi->data->what, n, U8);
5664 RExC_rxi->data->count = count + n;
5665 Copy(s, RExC_rxi->data->what + count, n, U8);
5669 /*XXX: todo make this not included in a non debugging perl, but appears to be
5670 * used anyway there, in 'use re' */
5671 #ifndef PERL_IN_XSUB_RE
5673 Perl_reginitcolors(pTHX)
5675 const char * const s = PerlEnv_getenv("PERL_RE_COLORS");
5677 char *t = savepv(s);
5681 t = strchr(t, '\t');
5687 PL_colors[i] = t = (char *)"";
5692 PL_colors[i++] = (char *)"";
5699 #ifdef TRIE_STUDY_OPT
5700 #define CHECK_RESTUDY_GOTO_butfirst(dOsomething) \
5703 (data.flags & SCF_TRIE_RESTUDY) \
5711 #define CHECK_RESTUDY_GOTO_butfirst
5715 * pregcomp - compile a regular expression into internal code
5717 * Decides which engine's compiler to call based on the hint currently in
5721 #ifndef PERL_IN_XSUB_RE
5723 /* return the currently in-scope regex engine (or the default if none) */
5725 regexp_engine const *
5726 Perl_current_re_engine(pTHX)
5728 if (IN_PERL_COMPILETIME) {
5729 HV * const table = GvHV(PL_hintgv);
5732 if (!table || !(PL_hints & HINT_LOCALIZE_HH))
5733 return &reh_regexp_engine;
5734 ptr = hv_fetchs(table, "regcomp", FALSE);
5735 if ( !(ptr && SvIOK(*ptr) && SvIV(*ptr)))
5736 return &reh_regexp_engine;
5737 return INT2PTR(regexp_engine*,SvIV(*ptr));
5741 if (!PL_curcop->cop_hints_hash)
5742 return &reh_regexp_engine;
5743 ptr = cop_hints_fetch_pvs(PL_curcop, "regcomp", 0);
5744 if ( !(ptr && SvIOK(ptr) && SvIV(ptr)))
5745 return &reh_regexp_engine;
5746 return INT2PTR(regexp_engine*,SvIV(ptr));
5752 Perl_pregcomp(pTHX_ SV * const pattern, const U32 flags)
5754 regexp_engine const *eng = current_re_engine();
5755 GET_RE_DEBUG_FLAGS_DECL;
5757 PERL_ARGS_ASSERT_PREGCOMP;
5759 /* Dispatch a request to compile a regexp to correct regexp engine. */
5761 PerlIO_printf(Perl_debug_log, "Using engine %"UVxf"\n",
5764 return CALLREGCOMP_ENG(eng, pattern, flags);
5768 /* public(ish) entry point for the perl core's own regex compiling code.
5769 * It's actually a wrapper for Perl_re_op_compile that only takes an SV
5770 * pattern rather than a list of OPs, and uses the internal engine rather
5771 * than the current one */
5774 Perl_re_compile(pTHX_ SV * const pattern, U32 rx_flags)
5776 SV *pat = pattern; /* defeat constness! */
5777 PERL_ARGS_ASSERT_RE_COMPILE;
5778 return Perl_re_op_compile(aTHX_ &pat, 1, NULL,
5779 #ifdef PERL_IN_XSUB_RE
5784 NULL, NULL, rx_flags, 0);
5788 /* upgrade pattern pat_p of length plen_p to UTF8, and if there are code
5789 * blocks, recalculate the indices. Update pat_p and plen_p in-place to
5790 * point to the realloced string and length.
5792 * This is essentially a copy of Perl_bytes_to_utf8() with the code index
5796 S_pat_upgrade_to_utf8(pTHX_ RExC_state_t * const pRExC_state,
5797 char **pat_p, STRLEN *plen_p, int num_code_blocks)
5799 U8 *const src = (U8*)*pat_p;
5804 GET_RE_DEBUG_FLAGS_DECL;
5806 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
5807 "UTF8 mismatch! Converting to utf8 for resizing and compile\n"));
5809 Newx(dst, *plen_p * 2 + 1, U8);
5812 while (s < *plen_p) {
5813 append_utf8_from_native_byte(src[s], &d);
5814 if (n < num_code_blocks) {
5815 if (!do_end && pRExC_state->code_blocks[n].start == s) {
5816 pRExC_state->code_blocks[n].start = d - dst - 1;
5817 assert(*(d - 1) == '(');
5820 else if (do_end && pRExC_state->code_blocks[n].end == s) {
5821 pRExC_state->code_blocks[n].end = d - dst - 1;
5822 assert(*(d - 1) == ')');
5831 *pat_p = (char*) dst;
5833 RExC_orig_utf8 = RExC_utf8 = 1;
5838 /* S_concat_pat(): concatenate a list of args to the pattern string pat,
5839 * while recording any code block indices, and handling overloading,
5840 * nested qr// objects etc. If pat is null, it will allocate a new
5841 * string, or just return the first arg, if there's only one.
5843 * Returns the malloced/updated pat.
5844 * patternp and pat_count is the array of SVs to be concatted;
5845 * oplist is the optional list of ops that generated the SVs;
5846 * recompile_p is a pointer to a boolean that will be set if
5847 * the regex will need to be recompiled.
5848 * delim, if non-null is an SV that will be inserted between each element
5852 S_concat_pat(pTHX_ RExC_state_t * const pRExC_state,
5853 SV *pat, SV ** const patternp, int pat_count,
5854 OP *oplist, bool *recompile_p, SV *delim)
5858 bool use_delim = FALSE;
5859 bool alloced = FALSE;
5861 /* if we know we have at least two args, create an empty string,
5862 * then concatenate args to that. For no args, return an empty string */
5863 if (!pat && pat_count != 1) {
5869 for (svp = patternp; svp < patternp + pat_count; svp++) {
5872 STRLEN orig_patlen = 0;
5874 SV *msv = use_delim ? delim : *svp;
5875 if (!msv) msv = &PL_sv_undef;
5877 /* if we've got a delimiter, we go round the loop twice for each
5878 * svp slot (except the last), using the delimiter the second
5887 if (SvTYPE(msv) == SVt_PVAV) {
5888 /* we've encountered an interpolated array within
5889 * the pattern, e.g. /...@a..../. Expand the list of elements,
5890 * then recursively append elements.
5891 * The code in this block is based on S_pushav() */
5893 AV *const av = (AV*)msv;
5894 const SSize_t maxarg = AvFILL(av) + 1;
5898 assert(oplist->op_type == OP_PADAV
5899 || oplist->op_type == OP_RV2AV);
5900 oplist = OpSIBLING(oplist);
5903 if (SvRMAGICAL(av)) {
5906 Newx(array, maxarg, SV*);
5908 for (i=0; i < maxarg; i++) {
5909 SV ** const svp = av_fetch(av, i, FALSE);
5910 array[i] = svp ? *svp : &PL_sv_undef;
5914 array = AvARRAY(av);
5916 pat = S_concat_pat(aTHX_ pRExC_state, pat,
5917 array, maxarg, NULL, recompile_p,
5919 GvSV((gv_fetchpvs("\"", GV_ADDMULTI, SVt_PV))));
5925 /* we make the assumption here that each op in the list of
5926 * op_siblings maps to one SV pushed onto the stack,
5927 * except for code blocks, with have both an OP_NULL and
5929 * This allows us to match up the list of SVs against the
5930 * list of OPs to find the next code block.
5932 * Note that PUSHMARK PADSV PADSV ..
5934 * PADRANGE PADSV PADSV ..
5935 * so the alignment still works. */
5938 if (oplist->op_type == OP_NULL
5939 && (oplist->op_flags & OPf_SPECIAL))
5941 assert(n < pRExC_state->num_code_blocks);
5942 pRExC_state->code_blocks[n].start = pat ? SvCUR(pat) : 0;
5943 pRExC_state->code_blocks[n].block = oplist;
5944 pRExC_state->code_blocks[n].src_regex = NULL;
5947 oplist = OpSIBLING(oplist); /* skip CONST */
5950 oplist = OpSIBLING(oplist);;
5953 /* apply magic and QR overloading to arg */
5956 if (SvROK(msv) && SvAMAGIC(msv)) {
5957 SV *sv = AMG_CALLunary(msv, regexp_amg);
5961 if (SvTYPE(sv) != SVt_REGEXP)
5962 Perl_croak(aTHX_ "Overloaded qr did not return a REGEXP");
5967 /* try concatenation overload ... */
5968 if (pat && (SvAMAGIC(pat) || SvAMAGIC(msv)) &&
5969 (sv = amagic_call(pat, msv, concat_amg, AMGf_assign)))
5972 /* overloading involved: all bets are off over literal
5973 * code. Pretend we haven't seen it */
5974 pRExC_state->num_code_blocks -= n;
5978 /* ... or failing that, try "" overload */
5979 while (SvAMAGIC(msv)
5980 && (sv = AMG_CALLunary(msv, string_amg))
5984 && SvRV(msv) == SvRV(sv))
5989 if (SvROK(msv) && SvTYPE(SvRV(msv)) == SVt_REGEXP)
5993 /* this is a partially unrolled
5994 * sv_catsv_nomg(pat, msv);
5995 * that allows us to adjust code block indices if
5998 char *dst = SvPV_force_nomg(pat, dlen);
6000 if (SvUTF8(msv) && !SvUTF8(pat)) {
6001 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &dst, &dlen, n);
6002 sv_setpvn(pat, dst, dlen);
6005 sv_catsv_nomg(pat, msv);
6012 pRExC_state->code_blocks[n-1].end = SvCUR(pat)-1;
6015 /* extract any code blocks within any embedded qr//'s */
6016 if (rx && SvTYPE(rx) == SVt_REGEXP
6017 && RX_ENGINE((REGEXP*)rx)->op_comp)
6020 RXi_GET_DECL(ReANY((REGEXP *)rx), ri);
6021 if (ri->num_code_blocks) {
6023 /* the presence of an embedded qr// with code means
6024 * we should always recompile: the text of the
6025 * qr// may not have changed, but it may be a
6026 * different closure than last time */
6028 Renew(pRExC_state->code_blocks,
6029 pRExC_state->num_code_blocks + ri->num_code_blocks,
6030 struct reg_code_block);
6031 pRExC_state->num_code_blocks += ri->num_code_blocks;
6033 for (i=0; i < ri->num_code_blocks; i++) {
6034 struct reg_code_block *src, *dst;
6035 STRLEN offset = orig_patlen
6036 + ReANY((REGEXP *)rx)->pre_prefix;
6037 assert(n < pRExC_state->num_code_blocks);
6038 src = &ri->code_blocks[i];
6039 dst = &pRExC_state->code_blocks[n];
6040 dst->start = src->start + offset;
6041 dst->end = src->end + offset;
6042 dst->block = src->block;
6043 dst->src_regex = (REGEXP*) SvREFCNT_inc( (SV*)
6052 /* avoid calling magic multiple times on a single element e.g. =~ $qr */
6061 /* see if there are any run-time code blocks in the pattern.
6062 * False positives are allowed */
6065 S_has_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6066 char *pat, STRLEN plen)
6071 PERL_UNUSED_CONTEXT;
6073 for (s = 0; s < plen; s++) {
6074 if (n < pRExC_state->num_code_blocks
6075 && s == pRExC_state->code_blocks[n].start)
6077 s = pRExC_state->code_blocks[n].end;
6081 /* TODO ideally should handle [..], (#..), /#.../x to reduce false
6083 if (pat[s] == '(' && s+2 <= plen && pat[s+1] == '?' &&
6085 || (s + 2 <= plen && pat[s+2] == '?' && pat[s+3] == '{'))
6092 /* Handle run-time code blocks. We will already have compiled any direct
6093 * or indirect literal code blocks. Now, take the pattern 'pat' and make a
6094 * copy of it, but with any literal code blocks blanked out and
6095 * appropriate chars escaped; then feed it into
6097 * eval "qr'modified_pattern'"
6101 * a\bc(?{"this was literal"})def'ghi\\jkl(?{"this is runtime"})mno
6105 * qr'a\\bc_______________________def\'ghi\\\\jkl(?{"this is runtime"})mno'
6107 * After eval_sv()-ing that, grab any new code blocks from the returned qr
6108 * and merge them with any code blocks of the original regexp.
6110 * If the pat is non-UTF8, while the evalled qr is UTF8, don't merge;
6111 * instead, just save the qr and return FALSE; this tells our caller that
6112 * the original pattern needs upgrading to utf8.
6116 S_compile_runtime_code(pTHX_ RExC_state_t * const pRExC_state,
6117 char *pat, STRLEN plen)
6121 GET_RE_DEBUG_FLAGS_DECL;
6123 if (pRExC_state->runtime_code_qr) {
6124 /* this is the second time we've been called; this should
6125 * only happen if the main pattern got upgraded to utf8
6126 * during compilation; re-use the qr we compiled first time
6127 * round (which should be utf8 too)
6129 qr = pRExC_state->runtime_code_qr;
6130 pRExC_state->runtime_code_qr = NULL;
6131 assert(RExC_utf8 && SvUTF8(qr));
6137 int newlen = plen + 6; /* allow for "qr''x\0" extra chars */
6141 /* determine how many extra chars we need for ' and \ escaping */
6142 for (s = 0; s < plen; s++) {
6143 if (pat[s] == '\'' || pat[s] == '\\')
6147 Newx(newpat, newlen, char);
6149 *p++ = 'q'; *p++ = 'r'; *p++ = '\'';
6151 for (s = 0; s < plen; s++) {
6152 if (n < pRExC_state->num_code_blocks
6153 && s == pRExC_state->code_blocks[n].start)
6155 /* blank out literal code block */
6156 assert(pat[s] == '(');
6157 while (s <= pRExC_state->code_blocks[n].end) {
6165 if (pat[s] == '\'' || pat[s] == '\\')
6170 if (pRExC_state->pm_flags & RXf_PMf_EXTENDED)
6174 PerlIO_printf(Perl_debug_log,
6175 "%sre-parsing pattern for runtime code:%s %s\n",
6176 PL_colors[4],PL_colors[5],newpat);
6179 sv = newSVpvn_flags(newpat, p-newpat-1, RExC_utf8 ? SVf_UTF8 : 0);
6185 PUSHSTACKi(PERLSI_REQUIRE);
6186 /* G_RE_REPARSING causes the toker to collapse \\ into \ when
6187 * parsing qr''; normally only q'' does this. It also alters
6189 eval_sv(sv, G_SCALAR|G_RE_REPARSING);
6190 SvREFCNT_dec_NN(sv);
6195 SV * const errsv = ERRSV;
6196 if (SvTRUE_NN(errsv))
6198 Safefree(pRExC_state->code_blocks);
6199 /* use croak_sv ? */
6200 Perl_croak_nocontext("%"SVf, SVfARG(errsv));
6203 assert(SvROK(qr_ref));
6205 assert(SvTYPE(qr) == SVt_REGEXP && RX_ENGINE((REGEXP*)qr)->op_comp);
6206 /* the leaving below frees the tmp qr_ref.
6207 * Give qr a life of its own */
6215 if (!RExC_utf8 && SvUTF8(qr)) {
6216 /* first time through; the pattern got upgraded; save the
6217 * qr for the next time through */
6218 assert(!pRExC_state->runtime_code_qr);
6219 pRExC_state->runtime_code_qr = qr;
6224 /* extract any code blocks within the returned qr// */
6227 /* merge the main (r1) and run-time (r2) code blocks into one */
6229 RXi_GET_DECL(ReANY((REGEXP *)qr), r2);
6230 struct reg_code_block *new_block, *dst;
6231 RExC_state_t * const r1 = pRExC_state; /* convenient alias */
6234 if (!r2->num_code_blocks) /* we guessed wrong */
6236 SvREFCNT_dec_NN(qr);
6241 r1->num_code_blocks + r2->num_code_blocks,
6242 struct reg_code_block);
6245 while ( i1 < r1->num_code_blocks
6246 || i2 < r2->num_code_blocks)
6248 struct reg_code_block *src;
6251 if (i1 == r1->num_code_blocks) {
6252 src = &r2->code_blocks[i2++];
6255 else if (i2 == r2->num_code_blocks)
6256 src = &r1->code_blocks[i1++];
6257 else if ( r1->code_blocks[i1].start
6258 < r2->code_blocks[i2].start)
6260 src = &r1->code_blocks[i1++];
6261 assert(src->end < r2->code_blocks[i2].start);
6264 assert( r1->code_blocks[i1].start
6265 > r2->code_blocks[i2].start);
6266 src = &r2->code_blocks[i2++];
6268 assert(src->end < r1->code_blocks[i1].start);
6271 assert(pat[src->start] == '(');
6272 assert(pat[src->end] == ')');
6273 dst->start = src->start;
6274 dst->end = src->end;
6275 dst->block = src->block;
6276 dst->src_regex = is_qr ? (REGEXP*) SvREFCNT_inc( (SV*) qr)
6280 r1->num_code_blocks += r2->num_code_blocks;
6281 Safefree(r1->code_blocks);
6282 r1->code_blocks = new_block;
6285 SvREFCNT_dec_NN(qr);
6291 S_setup_longest(pTHX_ RExC_state_t *pRExC_state, SV* sv_longest,
6292 SV** rx_utf8, SV** rx_substr, SSize_t* rx_end_shift,
6293 SSize_t lookbehind, SSize_t offset, SSize_t *minlen,
6294 STRLEN longest_length, bool eol, bool meol)
6296 /* This is the common code for setting up the floating and fixed length
6297 * string data extracted from Perl_re_op_compile() below. Returns a boolean
6298 * as to whether succeeded or not */
6303 if (! (longest_length
6304 || (eol /* Can't have SEOL and MULTI */
6305 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)))
6307 /* See comments for join_exact for why REG_UNFOLDED_MULTI_SEEN */
6308 || (RExC_seen & REG_UNFOLDED_MULTI_SEEN))
6313 /* copy the information about the longest from the reg_scan_data
6314 over to the program. */
6315 if (SvUTF8(sv_longest)) {
6316 *rx_utf8 = sv_longest;
6319 *rx_substr = sv_longest;
6322 /* end_shift is how many chars that must be matched that
6323 follow this item. We calculate it ahead of time as once the
6324 lookbehind offset is added in we lose the ability to correctly
6326 ml = minlen ? *(minlen) : (SSize_t)longest_length;
6327 *rx_end_shift = ml - offset
6328 - longest_length + (SvTAIL(sv_longest) != 0)
6331 t = (eol/* Can't have SEOL and MULTI */
6332 && (! meol || (RExC_flags & RXf_PMf_MULTILINE)));
6333 fbm_compile(sv_longest, t ? FBMcf_TAIL : 0);
6339 * Perl_re_op_compile - the perl internal RE engine's function to compile a
6340 * regular expression into internal code.
6341 * The pattern may be passed either as:
6342 * a list of SVs (patternp plus pat_count)
6343 * a list of OPs (expr)
6344 * If both are passed, the SV list is used, but the OP list indicates
6345 * which SVs are actually pre-compiled code blocks
6347 * The SVs in the list have magic and qr overloading applied to them (and
6348 * the list may be modified in-place with replacement SVs in the latter
6351 * If the pattern hasn't changed from old_re, then old_re will be
6354 * eng is the current engine. If that engine has an op_comp method, then
6355 * handle directly (i.e. we assume that op_comp was us); otherwise, just
6356 * do the initial concatenation of arguments and pass on to the external
6359 * If is_bare_re is not null, set it to a boolean indicating whether the
6360 * arg list reduced (after overloading) to a single bare regex which has
6361 * been returned (i.e. /$qr/).
6363 * orig_rx_flags contains RXf_* flags. See perlreapi.pod for more details.
6365 * pm_flags contains the PMf_* flags, typically based on those from the
6366 * pm_flags field of the related PMOP. Currently we're only interested in
6367 * PMf_HAS_CV, PMf_IS_QR, PMf_USE_RE_EVAL.
6369 * We can't allocate space until we know how big the compiled form will be,
6370 * but we can't compile it (and thus know how big it is) until we've got a
6371 * place to put the code. So we cheat: we compile it twice, once with code
6372 * generation turned off and size counting turned on, and once "for real".
6373 * This also means that we don't allocate space until we are sure that the
6374 * thing really will compile successfully, and we never have to move the
6375 * code and thus invalidate pointers into it. (Note that it has to be in
6376 * one piece because free() must be able to free it all.) [NB: not true in perl]
6378 * Beware that the optimization-preparation code in here knows about some
6379 * of the structure of the compiled regexp. [I'll say.]
6383 Perl_re_op_compile(pTHX_ SV ** const patternp, int pat_count,
6384 OP *expr, const regexp_engine* eng, REGEXP *old_re,
6385 bool *is_bare_re, U32 orig_rx_flags, U32 pm_flags)
6389 regexp_internal *ri;
6397 SV *code_blocksv = NULL;
6398 SV** new_patternp = patternp;
6400 /* these are all flags - maybe they should be turned
6401 * into a single int with different bit masks */
6402 I32 sawlookahead = 0;
6407 regex_charset initial_charset = get_regex_charset(orig_rx_flags);
6409 bool runtime_code = 0;
6411 RExC_state_t RExC_state;
6412 RExC_state_t * const pRExC_state = &RExC_state;
6413 #ifdef TRIE_STUDY_OPT
6415 RExC_state_t copyRExC_state;
6417 GET_RE_DEBUG_FLAGS_DECL;
6419 PERL_ARGS_ASSERT_RE_OP_COMPILE;
6421 DEBUG_r(if (!PL_colorset) reginitcolors());
6423 /* Initialize these here instead of as-needed, as is quick and avoids
6424 * having to test them each time otherwise */
6425 if (! PL_AboveLatin1) {
6426 PL_AboveLatin1 = _new_invlist_C_array(AboveLatin1_invlist);
6427 PL_Latin1 = _new_invlist_C_array(Latin1_invlist);
6428 PL_UpperLatin1 = _new_invlist_C_array(UpperLatin1_invlist);
6429 PL_utf8_foldable = _new_invlist_C_array(_Perl_Any_Folds_invlist);
6430 PL_HasMultiCharFold =
6431 _new_invlist_C_array(_Perl_Folds_To_Multi_Char_invlist);
6433 /* This is calculated here, because the Perl program that generates the
6434 * static global ones doesn't currently have access to
6435 * NUM_ANYOF_CODE_POINTS */
6436 PL_InBitmap = _new_invlist(2);
6437 PL_InBitmap = _add_range_to_invlist(PL_InBitmap, 0,
6438 NUM_ANYOF_CODE_POINTS - 1);
6441 pRExC_state->code_blocks = NULL;
6442 pRExC_state->num_code_blocks = 0;
6445 *is_bare_re = FALSE;
6447 if (expr && (expr->op_type == OP_LIST ||
6448 (expr->op_type == OP_NULL && expr->op_targ == OP_LIST))) {
6449 /* allocate code_blocks if needed */
6453 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o))
6454 if (o->op_type == OP_NULL && (o->op_flags & OPf_SPECIAL))
6455 ncode++; /* count of DO blocks */
6457 pRExC_state->num_code_blocks = ncode;
6458 Newx(pRExC_state->code_blocks, ncode, struct reg_code_block);
6463 /* compile-time pattern with just OP_CONSTs and DO blocks */
6468 /* find how many CONSTs there are */
6471 if (expr->op_type == OP_CONST)
6474 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6475 if (o->op_type == OP_CONST)
6479 /* fake up an SV array */
6481 assert(!new_patternp);
6482 Newx(new_patternp, n, SV*);
6483 SAVEFREEPV(new_patternp);
6487 if (expr->op_type == OP_CONST)
6488 new_patternp[n] = cSVOPx_sv(expr);
6490 for (o = cLISTOPx(expr)->op_first; o; o = OpSIBLING(o)) {
6491 if (o->op_type == OP_CONST)
6492 new_patternp[n++] = cSVOPo_sv;
6497 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6498 "Assembling pattern from %d elements%s\n", pat_count,
6499 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6501 /* set expr to the first arg op */
6503 if (pRExC_state->num_code_blocks
6504 && expr->op_type != OP_CONST)
6506 expr = cLISTOPx(expr)->op_first;
6507 assert( expr->op_type == OP_PUSHMARK
6508 || (expr->op_type == OP_NULL && expr->op_targ == OP_PUSHMARK)
6509 || expr->op_type == OP_PADRANGE);
6510 expr = OpSIBLING(expr);
6513 pat = S_concat_pat(aTHX_ pRExC_state, NULL, new_patternp, pat_count,
6514 expr, &recompile, NULL);
6516 /* handle bare (possibly after overloading) regex: foo =~ $re */
6521 if (SvTYPE(re) == SVt_REGEXP) {
6525 Safefree(pRExC_state->code_blocks);
6526 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log,
6527 "Precompiled pattern%s\n",
6528 orig_rx_flags & RXf_SPLIT ? " for split" : ""));
6534 exp = SvPV_nomg(pat, plen);
6536 if (!eng->op_comp) {
6537 if ((SvUTF8(pat) && IN_BYTES)
6538 || SvGMAGICAL(pat) || SvAMAGIC(pat))
6540 /* make a temporary copy; either to convert to bytes,
6541 * or to avoid repeating get-magic / overloaded stringify */
6542 pat = newSVpvn_flags(exp, plen, SVs_TEMP |
6543 (IN_BYTES ? 0 : SvUTF8(pat)));
6545 Safefree(pRExC_state->code_blocks);
6546 return CALLREGCOMP_ENG(eng, pat, orig_rx_flags);
6549 /* ignore the utf8ness if the pattern is 0 length */
6550 RExC_utf8 = RExC_orig_utf8 = (plen == 0 || IN_BYTES) ? 0 : SvUTF8(pat);
6551 RExC_uni_semantics = 0;
6552 RExC_contains_locale = 0;
6553 RExC_contains_i = 0;
6554 RExC_strict = cBOOL(pm_flags & RXf_PMf_STRICT);
6555 pRExC_state->runtime_code_qr = NULL;
6556 RExC_frame_head= NULL;
6557 RExC_frame_last= NULL;
6558 RExC_frame_count= 0;
6561 RExC_mysv1= sv_newmortal();
6562 RExC_mysv2= sv_newmortal();
6565 SV *dsv= sv_newmortal();
6566 RE_PV_QUOTED_DECL(s, RExC_utf8, dsv, exp, plen, 60);
6567 PerlIO_printf(Perl_debug_log, "%sCompiling REx%s %s\n",
6568 PL_colors[4],PL_colors[5],s);
6572 /* we jump here if we upgrade the pattern to utf8 and have to
6575 if ((pm_flags & PMf_USE_RE_EVAL)
6576 /* this second condition covers the non-regex literal case,
6577 * i.e. $foo =~ '(?{})'. */
6578 || (IN_PERL_COMPILETIME && (PL_hints & HINT_RE_EVAL))
6580 runtime_code = S_has_runtime_code(aTHX_ pRExC_state, exp, plen);
6582 /* return old regex if pattern hasn't changed */
6583 /* XXX: note in the below we have to check the flags as well as the
6586 * Things get a touch tricky as we have to compare the utf8 flag
6587 * independently from the compile flags. */
6591 && !!RX_UTF8(old_re) == !!RExC_utf8
6592 && ( RX_COMPFLAGS(old_re) == ( orig_rx_flags & RXf_PMf_FLAGCOPYMASK ) )
6593 && RX_PRECOMP(old_re)
6594 && RX_PRELEN(old_re) == plen
6595 && memEQ(RX_PRECOMP(old_re), exp, plen)
6596 && !runtime_code /* with runtime code, always recompile */ )
6598 Safefree(pRExC_state->code_blocks);
6602 rx_flags = orig_rx_flags;
6604 if (rx_flags & PMf_FOLD) {
6605 RExC_contains_i = 1;
6607 if (RExC_utf8 && initial_charset == REGEX_DEPENDS_CHARSET) {
6609 /* Set to use unicode semantics if the pattern is in utf8 and has the
6610 * 'depends' charset specified, as it means unicode when utf8 */
6611 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6615 RExC_flags = rx_flags;
6616 RExC_pm_flags = pm_flags;
6619 if (TAINTING_get && TAINT_get)
6620 Perl_croak(aTHX_ "Eval-group in insecure regular expression");
6622 if (!S_compile_runtime_code(aTHX_ pRExC_state, exp, plen)) {
6623 /* whoops, we have a non-utf8 pattern, whilst run-time code
6624 * got compiled as utf8. Try again with a utf8 pattern */
6625 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6626 pRExC_state->num_code_blocks);
6627 goto redo_first_pass;
6630 assert(!pRExC_state->runtime_code_qr);
6636 RExC_in_lookbehind = 0;
6637 RExC_seen_zerolen = *exp == '^' ? -1 : 0;
6639 RExC_override_recoding = 0;
6641 RExC_recode_x_to_native = 0;
6643 RExC_in_multi_char_class = 0;
6645 /* First pass: determine size, legality. */
6648 RExC_end = exp + plen;
6653 RExC_emit = (regnode *) &RExC_emit_dummy;
6654 RExC_whilem_seen = 0;
6655 RExC_open_parens = NULL;
6656 RExC_close_parens = NULL;
6658 RExC_paren_names = NULL;
6660 RExC_paren_name_list = NULL;
6662 RExC_recurse = NULL;
6663 RExC_study_chunk_recursed = NULL;
6664 RExC_study_chunk_recursed_bytes= 0;
6665 RExC_recurse_count = 0;
6666 pRExC_state->code_index = 0;
6669 PerlIO_printf(Perl_debug_log, "Starting first pass (sizing)\n");
6671 RExC_lastparse=NULL;
6673 /* reg may croak on us, not giving us a chance to free
6674 pRExC_state->code_blocks. We cannot SAVEFREEPV it now, as we may
6675 need it to survive as long as the regexp (qr/(?{})/).
6676 We must check that code_blocksv is not already set, because we may
6677 have jumped back to restart the sizing pass. */
6678 if (pRExC_state->code_blocks && !code_blocksv) {
6679 code_blocksv = newSV_type(SVt_PV);
6680 SAVEFREESV(code_blocksv);
6681 SvPV_set(code_blocksv, (char *)pRExC_state->code_blocks);
6682 SvLEN_set(code_blocksv, 1); /*sufficient to make sv_clear free it*/
6684 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6685 /* It's possible to write a regexp in ascii that represents Unicode
6686 codepoints outside of the byte range, such as via \x{100}. If we
6687 detect such a sequence we have to convert the entire pattern to utf8
6688 and then recompile, as our sizing calculation will have been based
6689 on 1 byte == 1 character, but we will need to use utf8 to encode
6690 at least some part of the pattern, and therefore must convert the whole
6693 if (flags & RESTART_UTF8) {
6694 S_pat_upgrade_to_utf8(aTHX_ pRExC_state, &exp, &plen,
6695 pRExC_state->num_code_blocks);
6696 goto redo_first_pass;
6698 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for sizing pass, flags=%#"UVxf"", (UV) flags);
6701 SvLEN_set(code_blocksv,0); /* no you can't have it, sv_clear */
6704 PerlIO_printf(Perl_debug_log,
6705 "Required size %"IVdf" nodes\n"
6706 "Starting second pass (creation)\n",
6709 RExC_lastparse=NULL;
6712 /* The first pass could have found things that force Unicode semantics */
6713 if ((RExC_utf8 || RExC_uni_semantics)
6714 && get_regex_charset(rx_flags) == REGEX_DEPENDS_CHARSET)
6716 set_regex_charset(&rx_flags, REGEX_UNICODE_CHARSET);
6719 /* Small enough for pointer-storage convention?
6720 If extralen==0, this means that we will not need long jumps. */
6721 if (RExC_size >= 0x10000L && RExC_extralen)
6722 RExC_size += RExC_extralen;
6725 if (RExC_whilem_seen > 15)
6726 RExC_whilem_seen = 15;
6728 /* Allocate space and zero-initialize. Note, the two step process
6729 of zeroing when in debug mode, thus anything assigned has to
6730 happen after that */
6731 rx = (REGEXP*) newSV_type(SVt_REGEXP);
6733 Newxc(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6734 char, regexp_internal);
6735 if ( r == NULL || ri == NULL )
6736 FAIL("Regexp out of space");
6738 /* avoid reading uninitialized memory in DEBUGGING code in study_chunk() */
6739 Zero(ri, sizeof(regexp_internal) + (unsigned)RExC_size * sizeof(regnode),
6742 /* bulk initialize base fields with 0. */
6743 Zero(ri, sizeof(regexp_internal), char);
6746 /* non-zero initialization begins here */
6749 r->extflags = rx_flags;
6750 RXp_COMPFLAGS(r) = orig_rx_flags & RXf_PMf_FLAGCOPYMASK;
6752 if (pm_flags & PMf_IS_QR) {
6753 ri->code_blocks = pRExC_state->code_blocks;
6754 ri->num_code_blocks = pRExC_state->num_code_blocks;
6759 for (n = 0; n < pRExC_state->num_code_blocks; n++)
6760 if (pRExC_state->code_blocks[n].src_regex)
6761 SAVEFREESV(pRExC_state->code_blocks[n].src_regex);
6762 SAVEFREEPV(pRExC_state->code_blocks);
6766 bool has_p = ((r->extflags & RXf_PMf_KEEPCOPY) == RXf_PMf_KEEPCOPY);
6767 bool has_charset = (get_regex_charset(r->extflags)
6768 != REGEX_DEPENDS_CHARSET);
6770 /* The caret is output if there are any defaults: if not all the STD
6771 * flags are set, or if no character set specifier is needed */
6773 (((r->extflags & RXf_PMf_STD_PMMOD) != RXf_PMf_STD_PMMOD)
6775 bool has_runon = ((RExC_seen & REG_RUN_ON_COMMENT_SEEN)
6776 == REG_RUN_ON_COMMENT_SEEN);
6777 U16 reganch = (U16)((r->extflags & RXf_PMf_STD_PMMOD)
6778 >> RXf_PMf_STD_PMMOD_SHIFT);
6779 const char *fptr = STD_PAT_MODS; /*"msixn"*/
6781 /* Allocate for the worst case, which is all the std flags are turned
6782 * on. If more precision is desired, we could do a population count of
6783 * the flags set. This could be done with a small lookup table, or by
6784 * shifting, masking and adding, or even, when available, assembly
6785 * language for a machine-language population count.
6786 * We never output a minus, as all those are defaults, so are
6787 * covered by the caret */
6788 const STRLEN wraplen = plen + has_p + has_runon
6789 + has_default /* If needs a caret */
6791 /* If needs a character set specifier */
6792 + ((has_charset) ? MAX_CHARSET_NAME_LENGTH : 0)
6793 + (sizeof(STD_PAT_MODS) - 1)
6794 + (sizeof("(?:)") - 1);
6796 Newx(p, wraplen + 1, char); /* +1 for the ending NUL */
6797 r->xpv_len_u.xpvlenu_pv = p;
6799 SvFLAGS(rx) |= SVf_UTF8;
6802 /* If a default, cover it using the caret */
6804 *p++= DEFAULT_PAT_MOD;
6808 const char* const name = get_regex_charset_name(r->extflags, &len);
6809 Copy(name, p, len, char);
6813 *p++ = KEEPCOPY_PAT_MOD; /*'p'*/
6816 while((ch = *fptr++)) {
6824 Copy(RExC_precomp, p, plen, char);
6825 assert ((RX_WRAPPED(rx) - p) < 16);
6826 r->pre_prefix = p - RX_WRAPPED(rx);
6832 SvCUR_set(rx, p - RX_WRAPPED(rx));
6836 r->nparens = RExC_npar - 1; /* set early to validate backrefs */
6838 /* setup various meta data about recursion, this all requires
6839 * RExC_npar to be correctly set, and a bit later on we clear it */
6840 if (RExC_seen & REG_RECURSE_SEEN) {
6841 Newxz(RExC_open_parens, RExC_npar,regnode *);
6842 SAVEFREEPV(RExC_open_parens);
6843 Newxz(RExC_close_parens,RExC_npar,regnode *);
6844 SAVEFREEPV(RExC_close_parens);
6846 if (RExC_seen & (REG_RECURSE_SEEN | REG_GOSTART_SEEN)) {
6847 /* Note, RExC_npar is 1 + the number of parens in a pattern.
6848 * So its 1 if there are no parens. */
6849 RExC_study_chunk_recursed_bytes= (RExC_npar >> 3) +
6850 ((RExC_npar & 0x07) != 0);
6851 Newx(RExC_study_chunk_recursed,
6852 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6853 SAVEFREEPV(RExC_study_chunk_recursed);
6856 /* Useful during FAIL. */
6857 #ifdef RE_TRACK_PATTERN_OFFSETS
6858 Newxz(ri->u.offsets, 2*RExC_size+1, U32); /* MJD 20001228 */
6859 DEBUG_OFFSETS_r(PerlIO_printf(Perl_debug_log,
6860 "%s %"UVuf" bytes for offset annotations.\n",
6861 ri->u.offsets ? "Got" : "Couldn't get",
6862 (UV)((2*RExC_size+1) * sizeof(U32))));
6864 SetProgLen(ri,RExC_size);
6868 REH_CALL_COMP_BEGIN_HOOK(pRExC_state->rx);
6870 /* Second pass: emit code. */
6871 RExC_flags = rx_flags; /* don't let top level (?i) bleed */
6872 RExC_pm_flags = pm_flags;
6874 RExC_end = exp + plen;
6877 RExC_emit_start = ri->program;
6878 RExC_emit = ri->program;
6879 RExC_emit_bound = ri->program + RExC_size + 1;
6880 pRExC_state->code_index = 0;
6882 *((char*) RExC_emit++) = (char) REG_MAGIC;
6883 if (reg(pRExC_state, 0, &flags,1) == NULL) {
6885 Perl_croak(aTHX_ "panic: reg returned NULL to re_op_compile for generation pass, flags=%#"UVxf"", (UV) flags);
6887 /* XXXX To minimize changes to RE engine we always allocate
6888 3-units-long substrs field. */
6889 Newx(r->substrs, 1, struct reg_substr_data);
6890 if (RExC_recurse_count) {
6891 Newxz(RExC_recurse,RExC_recurse_count,regnode *);
6892 SAVEFREEPV(RExC_recurse);
6896 r->minlen = minlen = sawlookahead = sawplus = sawopen = sawminmod = 0;
6898 RExC_study_chunk_recursed_count= 0;
6900 Zero(r->substrs, 1, struct reg_substr_data);
6901 if (RExC_study_chunk_recursed) {
6902 Zero(RExC_study_chunk_recursed,
6903 RExC_study_chunk_recursed_bytes * RExC_npar, U8);
6907 #ifdef TRIE_STUDY_OPT
6909 StructCopy(&zero_scan_data, &data, scan_data_t);
6910 copyRExC_state = RExC_state;
6913 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log,"Restudying\n"));
6915 RExC_state = copyRExC_state;
6916 if (seen & REG_TOP_LEVEL_BRANCHES_SEEN)
6917 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
6919 RExC_seen &= ~REG_TOP_LEVEL_BRANCHES_SEEN;
6920 StructCopy(&zero_scan_data, &data, scan_data_t);
6923 StructCopy(&zero_scan_data, &data, scan_data_t);
6926 /* Dig out information for optimizations. */
6927 r->extflags = RExC_flags; /* was pm_op */
6928 /*dmq: removed as part of de-PMOP: pm->op_pmflags = RExC_flags; */
6931 SvUTF8_on(rx); /* Unicode in it? */
6932 ri->regstclass = NULL;
6933 if (RExC_naughty >= TOO_NAUGHTY) /* Probably an expensive pattern. */
6934 r->intflags |= PREGf_NAUGHTY;
6935 scan = ri->program + 1; /* First BRANCH. */
6937 /* testing for BRANCH here tells us whether there is "must appear"
6938 data in the pattern. If there is then we can use it for optimisations */
6939 if (!(RExC_seen & REG_TOP_LEVEL_BRANCHES_SEEN)) { /* Only one top-level choice.
6942 STRLEN longest_float_length, longest_fixed_length;
6943 regnode_ssc ch_class; /* pointed to by data */
6945 SSize_t last_close = 0; /* pointed to by data */
6946 regnode *first= scan;
6947 regnode *first_next= regnext(first);
6949 * Skip introductions and multiplicators >= 1
6950 * so that we can extract the 'meat' of the pattern that must
6951 * match in the large if() sequence following.
6952 * NOTE that EXACT is NOT covered here, as it is normally
6953 * picked up by the optimiser separately.
6955 * This is unfortunate as the optimiser isnt handling lookahead
6956 * properly currently.
6959 while ((OP(first) == OPEN && (sawopen = 1)) ||
6960 /* An OR of *one* alternative - should not happen now. */
6961 (OP(first) == BRANCH && OP(first_next) != BRANCH) ||
6962 /* for now we can't handle lookbehind IFMATCH*/
6963 (OP(first) == IFMATCH && !first->flags && (sawlookahead = 1)) ||
6964 (OP(first) == PLUS) ||
6965 (OP(first) == MINMOD) ||
6966 /* An {n,m} with n>0 */
6967 (PL_regkind[OP(first)] == CURLY && ARG1(first) > 0) ||
6968 (OP(first) == NOTHING && PL_regkind[OP(first_next)] != END ))
6971 * the only op that could be a regnode is PLUS, all the rest
6972 * will be regnode_1 or regnode_2.
6974 * (yves doesn't think this is true)
6976 if (OP(first) == PLUS)
6979 if (OP(first) == MINMOD)
6981 first += regarglen[OP(first)];
6983 first = NEXTOPER(first);
6984 first_next= regnext(first);
6987 /* Starting-point info. */
6989 DEBUG_PEEP("first:",first,0);
6990 /* Ignore EXACT as we deal with it later. */
6991 if (PL_regkind[OP(first)] == EXACT) {
6992 if (OP(first) == EXACT || OP(first) == EXACTL)
6993 NOOP; /* Empty, get anchored substr later. */
6995 ri->regstclass = first;
6998 else if (PL_regkind[OP(first)] == TRIE &&
6999 ((reg_trie_data *)ri->data->data[ ARG(first) ])->minlen>0)
7001 /* this can happen only on restudy */
7002 ri->regstclass = construct_ahocorasick_from_trie(pRExC_state, (regnode *)first, 0);
7005 else if (REGNODE_SIMPLE(OP(first)))
7006 ri->regstclass = first;
7007 else if (PL_regkind[OP(first)] == BOUND ||
7008 PL_regkind[OP(first)] == NBOUND)
7009 ri->regstclass = first;
7010 else if (PL_regkind[OP(first)] == BOL) {
7011 r->intflags |= (OP(first) == MBOL
7014 first = NEXTOPER(first);
7017 else if (OP(first) == GPOS) {
7018 r->intflags |= PREGf_ANCH_GPOS;
7019 first = NEXTOPER(first);
7022 else if ((!sawopen || !RExC_sawback) &&
7024 (OP(first) == STAR &&
7025 PL_regkind[OP(NEXTOPER(first))] == REG_ANY) &&
7026 !(r->intflags & PREGf_ANCH) && !pRExC_state->num_code_blocks)
7028 /* turn .* into ^.* with an implied $*=1 */
7030 (OP(NEXTOPER(first)) == REG_ANY)
7033 r->intflags |= (type | PREGf_IMPLICIT);
7034 first = NEXTOPER(first);
7037 if (sawplus && !sawminmod && !sawlookahead
7038 && (!sawopen || !RExC_sawback)
7039 && !pRExC_state->num_code_blocks) /* May examine pos and $& */
7040 /* x+ must match at the 1st pos of run of x's */
7041 r->intflags |= PREGf_SKIP;
7043 /* Scan is after the zeroth branch, first is atomic matcher. */
7044 #ifdef TRIE_STUDY_OPT
7047 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7048 (IV)(first - scan + 1))
7052 PerlIO_printf(Perl_debug_log, "first at %"IVdf"\n",
7053 (IV)(first - scan + 1))
7059 * If there's something expensive in the r.e., find the
7060 * longest literal string that must appear and make it the
7061 * regmust. Resolve ties in favor of later strings, since
7062 * the regstart check works with the beginning of the r.e.
7063 * and avoiding duplication strengthens checking. Not a
7064 * strong reason, but sufficient in the absence of others.
7065 * [Now we resolve ties in favor of the earlier string if
7066 * it happens that c_offset_min has been invalidated, since the
7067 * earlier string may buy us something the later one won't.]
7070 data.longest_fixed = newSVpvs("");
7071 data.longest_float = newSVpvs("");
7072 data.last_found = newSVpvs("");
7073 data.longest = &(data.longest_fixed);
7074 ENTER_with_name("study_chunk");
7075 SAVEFREESV(data.longest_fixed);
7076 SAVEFREESV(data.longest_float);
7077 SAVEFREESV(data.last_found);
7079 if (!ri->regstclass) {
7080 ssc_init(pRExC_state, &ch_class);
7081 data.start_class = &ch_class;
7082 stclass_flag = SCF_DO_STCLASS_AND;
7083 } else /* XXXX Check for BOUND? */
7085 data.last_closep = &last_close;
7088 minlen = study_chunk(pRExC_state, &first, &minlen, &fake,
7089 scan + RExC_size, /* Up to end */
7091 SCF_DO_SUBSTR | SCF_WHILEM_VISITED_POS | stclass_flag
7092 | (restudied ? SCF_TRIE_DOING_RESTUDY : 0),
7096 CHECK_RESTUDY_GOTO_butfirst(LEAVE_with_name("study_chunk"));
7099 if ( RExC_npar == 1 && data.longest == &(data.longest_fixed)
7100 && data.last_start_min == 0 && data.last_end > 0
7101 && !RExC_seen_zerolen
7102 && !(RExC_seen & REG_VERBARG_SEEN)
7103 && !(RExC_seen & REG_GPOS_SEEN)
7105 r->extflags |= RXf_CHECK_ALL;
7107 scan_commit(pRExC_state, &data,&minlen,0);
7109 longest_float_length = CHR_SVLEN(data.longest_float);
7111 if (! ((SvCUR(data.longest_fixed) /* ok to leave SvCUR */
7112 && data.offset_fixed == data.offset_float_min
7113 && SvCUR(data.longest_fixed) == SvCUR(data.longest_float)))
7114 && S_setup_longest (aTHX_ pRExC_state,
7118 &(r->float_end_shift),
7119 data.lookbehind_float,
7120 data.offset_float_min,
7122 longest_float_length,
7123 cBOOL(data.flags & SF_FL_BEFORE_EOL),
7124 cBOOL(data.flags & SF_FL_BEFORE_MEOL)))
7126 r->float_min_offset = data.offset_float_min - data.lookbehind_float;
7127 r->float_max_offset = data.offset_float_max;
7128 if (data.offset_float_max < SSize_t_MAX) /* Don't offset infinity */
7129 r->float_max_offset -= data.lookbehind_float;
7130 SvREFCNT_inc_simple_void_NN(data.longest_float);
7133 r->float_substr = r->float_utf8 = NULL;
7134 longest_float_length = 0;
7137 longest_fixed_length = CHR_SVLEN(data.longest_fixed);
7139 if (S_setup_longest (aTHX_ pRExC_state,
7141 &(r->anchored_utf8),
7142 &(r->anchored_substr),
7143 &(r->anchored_end_shift),
7144 data.lookbehind_fixed,
7147 longest_fixed_length,
7148 cBOOL(data.flags & SF_FIX_BEFORE_EOL),
7149 cBOOL(data.flags & SF_FIX_BEFORE_MEOL)))
7151 r->anchored_offset = data.offset_fixed - data.lookbehind_fixed;
7152 SvREFCNT_inc_simple_void_NN(data.longest_fixed);
7155 r->anchored_substr = r->anchored_utf8 = NULL;
7156 longest_fixed_length = 0;
7158 LEAVE_with_name("study_chunk");
7161 && (OP(ri->regstclass) == REG_ANY || OP(ri->regstclass) == SANY))
7162 ri->regstclass = NULL;
7164 if ((!(r->anchored_substr || r->anchored_utf8) || r->anchored_offset)
7166 && ! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7167 && is_ssc_worth_it(pRExC_state, data.start_class))
7169 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7171 ssc_finalize(pRExC_state, data.start_class);
7173 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7174 StructCopy(data.start_class,
7175 (regnode_ssc*)RExC_rxi->data->data[n],
7177 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7178 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7179 DEBUG_COMPILE_r({ SV *sv = sv_newmortal();
7180 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7181 PerlIO_printf(Perl_debug_log,
7182 "synthetic stclass \"%s\".\n",
7183 SvPVX_const(sv));});
7184 data.start_class = NULL;
7187 /* A temporary algorithm prefers floated substr to fixed one to dig
7189 if (longest_fixed_length > longest_float_length) {
7190 r->substrs->check_ix = 0;
7191 r->check_end_shift = r->anchored_end_shift;
7192 r->check_substr = r->anchored_substr;
7193 r->check_utf8 = r->anchored_utf8;
7194 r->check_offset_min = r->check_offset_max = r->anchored_offset;
7195 if (r->intflags & (PREGf_ANCH_SBOL|PREGf_ANCH_GPOS))
7196 r->intflags |= PREGf_NOSCAN;
7199 r->substrs->check_ix = 1;
7200 r->check_end_shift = r->float_end_shift;
7201 r->check_substr = r->float_substr;
7202 r->check_utf8 = r->float_utf8;
7203 r->check_offset_min = r->float_min_offset;
7204 r->check_offset_max = r->float_max_offset;
7206 if ((r->check_substr || r->check_utf8) ) {
7207 r->extflags |= RXf_USE_INTUIT;
7208 if (SvTAIL(r->check_substr ? r->check_substr : r->check_utf8))
7209 r->extflags |= RXf_INTUIT_TAIL;
7211 r->substrs->data[0].max_offset = r->substrs->data[0].min_offset;
7213 /* XXX Unneeded? dmq (shouldn't as this is handled elsewhere)
7214 if ( (STRLEN)minlen < longest_float_length )
7215 minlen= longest_float_length;
7216 if ( (STRLEN)minlen < longest_fixed_length )
7217 minlen= longest_fixed_length;
7221 /* Several toplevels. Best we can is to set minlen. */
7223 regnode_ssc ch_class;
7224 SSize_t last_close = 0;
7226 DEBUG_PARSE_r(PerlIO_printf(Perl_debug_log, "\nMulti Top Level\n"));
7228 scan = ri->program + 1;
7229 ssc_init(pRExC_state, &ch_class);
7230 data.start_class = &ch_class;
7231 data.last_closep = &last_close;
7234 minlen = study_chunk(pRExC_state,
7235 &scan, &minlen, &fake, scan + RExC_size, &data, -1, 0, NULL,
7236 SCF_DO_STCLASS_AND|SCF_WHILEM_VISITED_POS|(restudied
7237 ? SCF_TRIE_DOING_RESTUDY
7241 CHECK_RESTUDY_GOTO_butfirst(NOOP);
7243 r->check_substr = r->check_utf8 = r->anchored_substr = r->anchored_utf8
7244 = r->float_substr = r->float_utf8 = NULL;
7246 if (! (ANYOF_FLAGS(data.start_class) & SSC_MATCHES_EMPTY_STRING)
7247 && is_ssc_worth_it(pRExC_state, data.start_class))
7249 const U32 n = add_data(pRExC_state, STR_WITH_LEN("f"));
7251 ssc_finalize(pRExC_state, data.start_class);
7253 Newx(RExC_rxi->data->data[n], 1, regnode_ssc);
7254 StructCopy(data.start_class,
7255 (regnode_ssc*)RExC_rxi->data->data[n],
7257 ri->regstclass = (regnode*)RExC_rxi->data->data[n];
7258 r->intflags &= ~PREGf_SKIP; /* Used in find_byclass(). */
7259 DEBUG_COMPILE_r({ SV* sv = sv_newmortal();
7260 regprop(r, sv, (regnode*)data.start_class, NULL, pRExC_state);
7261 PerlIO_printf(Perl_debug_log,
7262 "synthetic stclass \"%s\".\n",
7263 SvPVX_const(sv));});
7264 data.start_class = NULL;
7268 if (RExC_seen & REG_UNBOUNDED_QUANTIFIER_SEEN) {
7269 r->extflags |= RXf_UNBOUNDED_QUANTIFIER_SEEN;
7270 r->maxlen = REG_INFTY;
7273 r->maxlen = RExC_maxlen;
7276 /* Guard against an embedded (?=) or (?<=) with a longer minlen than
7277 the "real" pattern. */
7279 PerlIO_printf(Perl_debug_log,"minlen: %"IVdf" r->minlen:%"IVdf" maxlen:%"IVdf"\n",
7280 (IV)minlen, (IV)r->minlen, (IV)RExC_maxlen);
7282 r->minlenret = minlen;
7283 if (r->minlen < minlen)
7286 if (RExC_seen & REG_GPOS_SEEN)
7287 r->intflags |= PREGf_GPOS_SEEN;
7288 if (RExC_seen & REG_LOOKBEHIND_SEEN)
7289 r->extflags |= RXf_NO_INPLACE_SUBST; /* inplace might break the
7291 if (pRExC_state->num_code_blocks)
7292 r->extflags |= RXf_EVAL_SEEN;
7293 if (RExC_seen & REG_CANY_SEEN)
7294 r->intflags |= PREGf_CANY_SEEN;
7295 if (RExC_seen & REG_VERBARG_SEEN)
7297 r->intflags |= PREGf_VERBARG_SEEN;
7298 r->extflags |= RXf_NO_INPLACE_SUBST; /* don't understand this! Yves */
7300 if (RExC_seen & REG_CUTGROUP_SEEN)
7301 r->intflags |= PREGf_CUTGROUP_SEEN;
7302 if (pm_flags & PMf_USE_RE_EVAL)
7303 r->intflags |= PREGf_USE_RE_EVAL;
7304 if (RExC_paren_names)
7305 RXp_PAREN_NAMES(r) = MUTABLE_HV(SvREFCNT_inc(RExC_paren_names));
7307 RXp_PAREN_NAMES(r) = NULL;
7309 /* If we have seen an anchor in our pattern then we set the extflag RXf_IS_ANCHORED
7310 * so it can be used in pp.c */
7311 if (r->intflags & PREGf_ANCH)
7312 r->extflags |= RXf_IS_ANCHORED;
7316 /* this is used to identify "special" patterns that might result
7317 * in Perl NOT calling the regex engine and instead doing the match "itself",
7318 * particularly special cases in split//. By having the regex compiler
7319 * do this pattern matching at a regop level (instead of by inspecting the pattern)
7320 * we avoid weird issues with equivalent patterns resulting in different behavior,
7321 * AND we allow non Perl engines to get the same optimizations by the setting the
7322 * flags appropriately - Yves */
7323 regnode *first = ri->program + 1;
7325 regnode *next = regnext(first);
7328 if (PL_regkind[fop] == NOTHING && nop == END)
7329 r->extflags |= RXf_NULL;
7330 else if ((fop == MBOL || (fop == SBOL && !first->flags)) && nop == END)
7331 /* when fop is SBOL first->flags will be true only when it was
7332 * produced by parsing /\A/, and not when parsing /^/. This is
7333 * very important for the split code as there we want to
7334 * treat /^/ as /^/m, but we do not want to treat /\A/ as /^/m.
7335 * See rt #122761 for more details. -- Yves */
7336 r->extflags |= RXf_START_ONLY;
7337 else if (fop == PLUS
7338 && PL_regkind[nop] == POSIXD && FLAGS(next) == _CC_SPACE
7340 r->extflags |= RXf_WHITE;
7341 else if ( r->extflags & RXf_SPLIT
7342 && (fop == EXACT || fop == EXACTL)
7343 && STR_LEN(first) == 1
7344 && *(STRING(first)) == ' '
7346 r->extflags |= (RXf_SKIPWHITE|RXf_WHITE);
7350 if (RExC_contains_locale) {
7351 RXp_EXTFLAGS(r) |= RXf_TAINTED;
7355 if (RExC_paren_names) {
7356 ri->name_list_idx = add_data( pRExC_state, STR_WITH_LEN("a"));
7357 ri->data->data[ri->name_list_idx]
7358 = (void*)SvREFCNT_inc(RExC_paren_name_list);
7361 ri->name_list_idx = 0;
7363 if (RExC_recurse_count) {
7364 for ( ; RExC_recurse_count ; RExC_recurse_count-- ) {
7365 const regnode *scan = RExC_recurse[RExC_recurse_count-1];
7366 ARG2L_SET( scan, RExC_open_parens[ARG(scan)-1] - scan );
7369 Newxz(r->offs, RExC_npar, regexp_paren_pair);
7370 /* assume we don't need to swap parens around before we match */
7372 PerlIO_printf(Perl_debug_log,"study_chunk_recursed_count: %lu\n",
7373 (unsigned long)RExC_study_chunk_recursed_count);
7377 PerlIO_printf(Perl_debug_log,"Final program:\n");
7380 #ifdef RE_TRACK_PATTERN_OFFSETS
7381 DEBUG_OFFSETS_r(if (ri->u.offsets) {
7382 const STRLEN len = ri->u.offsets[0];
7384 GET_RE_DEBUG_FLAGS_DECL;
7385 PerlIO_printf(Perl_debug_log,
7386 "Offsets: [%"UVuf"]\n\t", (UV)ri->u.offsets[0]);
7387 for (i = 1; i <= len; i++) {
7388 if (ri->u.offsets[i*2-1] || ri->u.offsets[i*2])
7389 PerlIO_printf(Perl_debug_log, "%"UVuf":%"UVuf"[%"UVuf"] ",
7390 (UV)i, (UV)ri->u.offsets[i*2-1], (UV)ri->u.offsets[i*2]);
7392 PerlIO_printf(Perl_debug_log, "\n");
7397 /* under ithreads the ?pat? PMf_USED flag on the pmop is simulated
7398 * by setting the regexp SV to readonly-only instead. If the
7399 * pattern's been recompiled, the USEDness should remain. */
7400 if (old_re && SvREADONLY(old_re))
7408 Perl_reg_named_buff(pTHX_ REGEXP * const rx, SV * const key, SV * const value,
7411 PERL_ARGS_ASSERT_REG_NAMED_BUFF;
7413 PERL_UNUSED_ARG(value);
7415 if (flags & RXapif_FETCH) {
7416 return reg_named_buff_fetch(rx, key, flags);
7417 } else if (flags & (RXapif_STORE | RXapif_DELETE | RXapif_CLEAR)) {
7418 Perl_croak_no_modify();
7420 } else if (flags & RXapif_EXISTS) {
7421 return reg_named_buff_exists(rx, key, flags)
7424 } else if (flags & RXapif_REGNAMES) {
7425 return reg_named_buff_all(rx, flags);
7426 } else if (flags & (RXapif_SCALAR | RXapif_REGNAMES_COUNT)) {
7427 return reg_named_buff_scalar(rx, flags);
7429 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff", (int)flags);
7435 Perl_reg_named_buff_iter(pTHX_ REGEXP * const rx, const SV * const lastkey,
7438 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ITER;
7439 PERL_UNUSED_ARG(lastkey);
7441 if (flags & RXapif_FIRSTKEY)
7442 return reg_named_buff_firstkey(rx, flags);
7443 else if (flags & RXapif_NEXTKEY)
7444 return reg_named_buff_nextkey(rx, flags);
7446 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_iter",
7453 Perl_reg_named_buff_fetch(pTHX_ REGEXP * const r, SV * const namesv,
7456 AV *retarray = NULL;
7458 struct regexp *const rx = ReANY(r);
7460 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FETCH;
7462 if (flags & RXapif_ALL)
7465 if (rx && RXp_PAREN_NAMES(rx)) {
7466 HE *he_str = hv_fetch_ent( RXp_PAREN_NAMES(rx), namesv, 0, 0 );
7469 SV* sv_dat=HeVAL(he_str);
7470 I32 *nums=(I32*)SvPVX(sv_dat);
7471 for ( i=0; i<SvIVX(sv_dat); i++ ) {
7472 if ((I32)(rx->nparens) >= nums[i]
7473 && rx->offs[nums[i]].start != -1
7474 && rx->offs[nums[i]].end != -1)
7477 CALLREG_NUMBUF_FETCH(r,nums[i],ret);
7482 ret = newSVsv(&PL_sv_undef);
7485 av_push(retarray, ret);
7488 return newRV_noinc(MUTABLE_SV(retarray));
7495 Perl_reg_named_buff_exists(pTHX_ REGEXP * const r, SV * const key,
7498 struct regexp *const rx = ReANY(r);
7500 PERL_ARGS_ASSERT_REG_NAMED_BUFF_EXISTS;
7502 if (rx && RXp_PAREN_NAMES(rx)) {
7503 if (flags & RXapif_ALL) {
7504 return hv_exists_ent(RXp_PAREN_NAMES(rx), key, 0);
7506 SV *sv = CALLREG_NAMED_BUFF_FETCH(r, key, flags);
7508 SvREFCNT_dec_NN(sv);
7520 Perl_reg_named_buff_firstkey(pTHX_ REGEXP * const r, const U32 flags)
7522 struct regexp *const rx = ReANY(r);
7524 PERL_ARGS_ASSERT_REG_NAMED_BUFF_FIRSTKEY;
7526 if ( rx && RXp_PAREN_NAMES(rx) ) {
7527 (void)hv_iterinit(RXp_PAREN_NAMES(rx));
7529 return CALLREG_NAMED_BUFF_NEXTKEY(r, NULL, flags & ~RXapif_FIRSTKEY);
7536 Perl_reg_named_buff_nextkey(pTHX_ REGEXP * const r, const U32 flags)
7538 struct regexp *const rx = ReANY(r);
7539 GET_RE_DEBUG_FLAGS_DECL;
7541 PERL_ARGS_ASSERT_REG_NAMED_BUFF_NEXTKEY;
7543 if (rx && RXp_PAREN_NAMES(rx)) {
7544 HV *hv = RXp_PAREN_NAMES(rx);
7546 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7549 SV* sv_dat = HeVAL(temphe);
7550 I32 *nums = (I32*)SvPVX(sv_dat);
7551 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7552 if ((I32)(rx->lastparen) >= nums[i] &&
7553 rx->offs[nums[i]].start != -1 &&
7554 rx->offs[nums[i]].end != -1)
7560 if (parno || flags & RXapif_ALL) {
7561 return newSVhek(HeKEY_hek(temphe));
7569 Perl_reg_named_buff_scalar(pTHX_ REGEXP * const r, const U32 flags)
7574 struct regexp *const rx = ReANY(r);
7576 PERL_ARGS_ASSERT_REG_NAMED_BUFF_SCALAR;
7578 if (rx && RXp_PAREN_NAMES(rx)) {
7579 if (flags & (RXapif_ALL | RXapif_REGNAMES_COUNT)) {
7580 return newSViv(HvTOTALKEYS(RXp_PAREN_NAMES(rx)));
7581 } else if (flags & RXapif_ONE) {
7582 ret = CALLREG_NAMED_BUFF_ALL(r, (flags | RXapif_REGNAMES));
7583 av = MUTABLE_AV(SvRV(ret));
7584 length = av_tindex(av);
7585 SvREFCNT_dec_NN(ret);
7586 return newSViv(length + 1);
7588 Perl_croak(aTHX_ "panic: Unknown flags %d in named_buff_scalar",
7593 return &PL_sv_undef;
7597 Perl_reg_named_buff_all(pTHX_ REGEXP * const r, const U32 flags)
7599 struct regexp *const rx = ReANY(r);
7602 PERL_ARGS_ASSERT_REG_NAMED_BUFF_ALL;
7604 if (rx && RXp_PAREN_NAMES(rx)) {
7605 HV *hv= RXp_PAREN_NAMES(rx);
7607 (void)hv_iterinit(hv);
7608 while ( (temphe = hv_iternext_flags(hv,0)) ) {
7611 SV* sv_dat = HeVAL(temphe);
7612 I32 *nums = (I32*)SvPVX(sv_dat);
7613 for ( i = 0; i < SvIVX(sv_dat); i++ ) {
7614 if ((I32)(rx->lastparen) >= nums[i] &&
7615 rx->offs[nums[i]].start != -1 &&
7616 rx->offs[nums[i]].end != -1)
7622 if (parno || flags & RXapif_ALL) {
7623 av_push(av, newSVhek(HeKEY_hek(temphe)));
7628 return newRV_noinc(MUTABLE_SV(av));
7632 Perl_reg_numbered_buff_fetch(pTHX_ REGEXP * const r, const I32 paren,
7635 struct regexp *const rx = ReANY(r);
7641 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_FETCH;
7643 if ( n == RX_BUFF_IDX_CARET_PREMATCH
7644 || n == RX_BUFF_IDX_CARET_FULLMATCH
7645 || n == RX_BUFF_IDX_CARET_POSTMATCH
7648 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7650 /* on something like
7653 * the KEEPCOPY is set on the PMOP rather than the regex */
7654 if (PL_curpm && r == PM_GETRE(PL_curpm))
7655 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7664 if (n == RX_BUFF_IDX_CARET_FULLMATCH)
7665 /* no need to distinguish between them any more */
7666 n = RX_BUFF_IDX_FULLMATCH;
7668 if ((n == RX_BUFF_IDX_PREMATCH || n == RX_BUFF_IDX_CARET_PREMATCH)
7669 && rx->offs[0].start != -1)
7671 /* $`, ${^PREMATCH} */
7672 i = rx->offs[0].start;
7676 if ((n == RX_BUFF_IDX_POSTMATCH || n == RX_BUFF_IDX_CARET_POSTMATCH)
7677 && rx->offs[0].end != -1)
7679 /* $', ${^POSTMATCH} */
7680 s = rx->subbeg - rx->suboffset + rx->offs[0].end;
7681 i = rx->sublen + rx->suboffset - rx->offs[0].end;
7684 if ( 0 <= n && n <= (I32)rx->nparens &&
7685 (s1 = rx->offs[n].start) != -1 &&
7686 (t1 = rx->offs[n].end) != -1)
7688 /* $&, ${^MATCH}, $1 ... */
7690 s = rx->subbeg + s1 - rx->suboffset;
7695 assert(s >= rx->subbeg);
7696 assert((STRLEN)rx->sublen >= (STRLEN)((s - rx->subbeg) + i) );
7698 #ifdef NO_TAINT_SUPPORT
7699 sv_setpvn(sv, s, i);
7701 const int oldtainted = TAINT_get;
7703 sv_setpvn(sv, s, i);
7704 TAINT_set(oldtainted);
7706 if ( (rx->intflags & PREGf_CANY_SEEN)
7707 ? (RXp_MATCH_UTF8(rx)
7708 && (!i || is_utf8_string((U8*)s, i)))
7709 : (RXp_MATCH_UTF8(rx)) )
7716 if (RXp_MATCH_TAINTED(rx)) {
7717 if (SvTYPE(sv) >= SVt_PVMG) {
7718 MAGIC* const mg = SvMAGIC(sv);
7721 SvMAGIC_set(sv, mg->mg_moremagic);
7723 if ((mgt = SvMAGIC(sv))) {
7724 mg->mg_moremagic = mgt;
7725 SvMAGIC_set(sv, mg);
7736 sv_setsv(sv,&PL_sv_undef);
7742 Perl_reg_numbered_buff_store(pTHX_ REGEXP * const rx, const I32 paren,
7743 SV const * const value)
7745 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_STORE;
7747 PERL_UNUSED_ARG(rx);
7748 PERL_UNUSED_ARG(paren);
7749 PERL_UNUSED_ARG(value);
7752 Perl_croak_no_modify();
7756 Perl_reg_numbered_buff_length(pTHX_ REGEXP * const r, const SV * const sv,
7759 struct regexp *const rx = ReANY(r);
7763 PERL_ARGS_ASSERT_REG_NUMBERED_BUFF_LENGTH;
7765 if ( paren == RX_BUFF_IDX_CARET_PREMATCH
7766 || paren == RX_BUFF_IDX_CARET_FULLMATCH
7767 || paren == RX_BUFF_IDX_CARET_POSTMATCH
7770 bool keepcopy = cBOOL(rx->extflags & RXf_PMf_KEEPCOPY);
7772 /* on something like
7775 * the KEEPCOPY is set on the PMOP rather than the regex */
7776 if (PL_curpm && r == PM_GETRE(PL_curpm))
7777 keepcopy = cBOOL(PL_curpm->op_pmflags & PMf_KEEPCOPY);
7783 /* Some of this code was originally in C<Perl_magic_len> in F<mg.c> */
7785 case RX_BUFF_IDX_CARET_PREMATCH: /* ${^PREMATCH} */
7786 case RX_BUFF_IDX_PREMATCH: /* $` */
7787 if (rx->offs[0].start != -1) {
7788 i = rx->offs[0].start;
7797 case RX_BUFF_IDX_CARET_POSTMATCH: /* ${^POSTMATCH} */
7798 case RX_BUFF_IDX_POSTMATCH: /* $' */
7799 if (rx->offs[0].end != -1) {
7800 i = rx->sublen - rx->offs[0].end;
7802 s1 = rx->offs[0].end;
7809 default: /* $& / ${^MATCH}, $1, $2, ... */
7810 if (paren <= (I32)rx->nparens &&
7811 (s1 = rx->offs[paren].start) != -1 &&
7812 (t1 = rx->offs[paren].end) != -1)
7818 if (ckWARN(WARN_UNINITIALIZED))
7819 report_uninit((const SV *)sv);
7824 if (i > 0 && RXp_MATCH_UTF8(rx)) {
7825 const char * const s = rx->subbeg - rx->suboffset + s1;
7830 if (is_utf8_string_loclen((U8*)s, i, &ep, &el))
7837 Perl_reg_qr_package(pTHX_ REGEXP * const rx)
7839 PERL_ARGS_ASSERT_REG_QR_PACKAGE;
7840 PERL_UNUSED_ARG(rx);
7844 return newSVpvs("Regexp");
7847 /* Scans the name of a named buffer from the pattern.
7848 * If flags is REG_RSN_RETURN_NULL returns null.
7849 * If flags is REG_RSN_RETURN_NAME returns an SV* containing the name
7850 * If flags is REG_RSN_RETURN_DATA returns the data SV* corresponding
7851 * to the parsed name as looked up in the RExC_paren_names hash.
7852 * If there is an error throws a vFAIL().. type exception.
7855 #define REG_RSN_RETURN_NULL 0
7856 #define REG_RSN_RETURN_NAME 1
7857 #define REG_RSN_RETURN_DATA 2
7860 S_reg_scan_name(pTHX_ RExC_state_t *pRExC_state, U32 flags)
7862 char *name_start = RExC_parse;
7864 PERL_ARGS_ASSERT_REG_SCAN_NAME;
7866 assert (RExC_parse <= RExC_end);
7867 if (RExC_parse == RExC_end) NOOP;
7868 else if (isIDFIRST_lazy_if(RExC_parse, UTF)) {
7869 /* skip IDFIRST by using do...while */
7872 RExC_parse += UTF8SKIP(RExC_parse);
7873 } while (isWORDCHAR_utf8((U8*)RExC_parse));
7877 } while (isWORDCHAR(*RExC_parse));
7879 RExC_parse++; /* so the <- from the vFAIL is after the offending
7881 vFAIL("Group name must start with a non-digit word character");
7885 = newSVpvn_flags(name_start, (int)(RExC_parse - name_start),
7886 SVs_TEMP | (UTF ? SVf_UTF8 : 0));
7887 if ( flags == REG_RSN_RETURN_NAME)
7889 else if (flags==REG_RSN_RETURN_DATA) {
7892 if ( ! sv_name ) /* should not happen*/
7893 Perl_croak(aTHX_ "panic: no svname in reg_scan_name");
7894 if (RExC_paren_names)
7895 he_str = hv_fetch_ent( RExC_paren_names, sv_name, 0, 0 );
7897 sv_dat = HeVAL(he_str);
7899 vFAIL("Reference to nonexistent named group");
7903 Perl_croak(aTHX_ "panic: bad flag %lx in reg_scan_name",
7904 (unsigned long) flags);
7906 NOT_REACHED; /* NOTREACHED */
7911 #define DEBUG_PARSE_MSG(funcname) DEBUG_PARSE_r({ \
7913 if (RExC_lastparse!=RExC_parse) { \
7914 PerlIO_printf(Perl_debug_log, "%s", \
7915 Perl_pv_pretty(aTHX_ RExC_mysv1, RExC_parse, \
7916 RExC_end - RExC_parse, 16, \
7918 PERL_PV_ESCAPE_UNI_DETECT | \
7919 PERL_PV_PRETTY_ELLIPSES | \
7920 PERL_PV_PRETTY_LTGT | \
7921 PERL_PV_ESCAPE_RE | \
7922 PERL_PV_PRETTY_EXACTSIZE \
7926 PerlIO_printf(Perl_debug_log,"%16s",""); \
7929 num = RExC_size + 1; \
7931 num=REG_NODE_NUM(RExC_emit); \
7932 if (RExC_lastnum!=num) \
7933 PerlIO_printf(Perl_debug_log,"|%4d",num); \
7935 PerlIO_printf(Perl_debug_log,"|%4s",""); \
7936 PerlIO_printf(Perl_debug_log,"|%*s%-4s", \
7937 (int)((depth*2)), "", \
7941 RExC_lastparse=RExC_parse; \
7946 #define DEBUG_PARSE(funcname) DEBUG_PARSE_r({ \
7947 DEBUG_PARSE_MSG((funcname)); \
7948 PerlIO_printf(Perl_debug_log,"%4s","\n"); \
7950 #define DEBUG_PARSE_FMT(funcname,fmt,args) DEBUG_PARSE_r({ \
7951 DEBUG_PARSE_MSG((funcname)); \
7952 PerlIO_printf(Perl_debug_log,fmt "\n",args); \
7955 /* This section of code defines the inversion list object and its methods. The
7956 * interfaces are highly subject to change, so as much as possible is static to
7957 * this file. An inversion list is here implemented as a malloc'd C UV array
7958 * as an SVt_INVLIST scalar.
7960 * An inversion list for Unicode is an array of code points, sorted by ordinal
7961 * number. The zeroth element is the first code point in the list. The 1th
7962 * element is the first element beyond that not in the list. In other words,
7963 * the first range is
7964 * invlist[0]..(invlist[1]-1)
7965 * The other ranges follow. Thus every element whose index is divisible by two
7966 * marks the beginning of a range that is in the list, and every element not
7967 * divisible by two marks the beginning of a range not in the list. A single
7968 * element inversion list that contains the single code point N generally
7969 * consists of two elements
7972 * (The exception is when N is the highest representable value on the
7973 * machine, in which case the list containing just it would be a single
7974 * element, itself. By extension, if the last range in the list extends to
7975 * infinity, then the first element of that range will be in the inversion list
7976 * at a position that is divisible by two, and is the final element in the
7978 * Taking the complement (inverting) an inversion list is quite simple, if the
7979 * first element is 0, remove it; otherwise add a 0 element at the beginning.
7980 * This implementation reserves an element at the beginning of each inversion
7981 * list to always contain 0; there is an additional flag in the header which
7982 * indicates if the list begins at the 0, or is offset to begin at the next
7985 * More about inversion lists can be found in "Unicode Demystified"
7986 * Chapter 13 by Richard Gillam, published by Addison-Wesley.
7987 * More will be coming when functionality is added later.
7989 * The inversion list data structure is currently implemented as an SV pointing
7990 * to an array of UVs that the SV thinks are bytes. This allows us to have an
7991 * array of UV whose memory management is automatically handled by the existing
7992 * facilities for SV's.
7994 * Some of the methods should always be private to the implementation, and some
7995 * should eventually be made public */
7997 /* The header definitions are in F<inline_invlist.c> */
7999 PERL_STATIC_INLINE UV*
8000 S__invlist_array_init(SV* const invlist, const bool will_have_0)
8002 /* Returns a pointer to the first element in the inversion list's array.
8003 * This is called upon initialization of an inversion list. Where the
8004 * array begins depends on whether the list has the code point U+0000 in it
8005 * or not. The other parameter tells it whether the code that follows this
8006 * call is about to put a 0 in the inversion list or not. The first
8007 * element is either the element reserved for 0, if TRUE, or the element
8008 * after it, if FALSE */
8010 bool* offset = get_invlist_offset_addr(invlist);
8011 UV* zero_addr = (UV *) SvPVX(invlist);
8013 PERL_ARGS_ASSERT__INVLIST_ARRAY_INIT;
8016 assert(! _invlist_len(invlist));
8020 /* 1^1 = 0; 1^0 = 1 */
8021 *offset = 1 ^ will_have_0;
8022 return zero_addr + *offset;
8025 PERL_STATIC_INLINE void
8026 S_invlist_set_len(pTHX_ SV* const invlist, const UV len, const bool offset)
8028 /* Sets the current number of elements stored in the inversion list.
8029 * Updates SvCUR correspondingly */
8030 PERL_UNUSED_CONTEXT;
8031 PERL_ARGS_ASSERT_INVLIST_SET_LEN;
8033 assert(SvTYPE(invlist) == SVt_INVLIST);
8038 : TO_INTERNAL_SIZE(len + offset));
8039 assert(SvLEN(invlist) == 0 || SvCUR(invlist) <= SvLEN(invlist));
8042 #ifndef PERL_IN_XSUB_RE
8044 PERL_STATIC_INLINE IV*
8045 S_get_invlist_previous_index_addr(SV* invlist)
8047 /* Return the address of the IV that is reserved to hold the cached index
8049 PERL_ARGS_ASSERT_GET_INVLIST_PREVIOUS_INDEX_ADDR;
8051 assert(SvTYPE(invlist) == SVt_INVLIST);
8053 return &(((XINVLIST*) SvANY(invlist))->prev_index);
8056 PERL_STATIC_INLINE IV
8057 S_invlist_previous_index(SV* const invlist)
8059 /* Returns cached index of previous search */
8061 PERL_ARGS_ASSERT_INVLIST_PREVIOUS_INDEX;
8063 return *get_invlist_previous_index_addr(invlist);
8066 PERL_STATIC_INLINE void
8067 S_invlist_set_previous_index(SV* const invlist, const IV index)
8069 /* Caches <index> for later retrieval */
8071 PERL_ARGS_ASSERT_INVLIST_SET_PREVIOUS_INDEX;
8073 assert(index == 0 || index < (int) _invlist_len(invlist));
8075 *get_invlist_previous_index_addr(invlist) = index;
8078 PERL_STATIC_INLINE void
8079 S_invlist_trim(SV* const invlist)
8081 PERL_ARGS_ASSERT_INVLIST_TRIM;
8083 assert(SvTYPE(invlist) == SVt_INVLIST);
8085 /* Change the length of the inversion list to how many entries it currently
8087 SvPV_shrink_to_cur((SV *) invlist);
8090 PERL_STATIC_INLINE bool
8091 S_invlist_is_iterating(SV* const invlist)
8093 PERL_ARGS_ASSERT_INVLIST_IS_ITERATING;
8095 return *(get_invlist_iter_addr(invlist)) < (STRLEN) UV_MAX;
8098 #endif /* ifndef PERL_IN_XSUB_RE */
8100 PERL_STATIC_INLINE UV
8101 S_invlist_max(SV* const invlist)
8103 /* Returns the maximum number of elements storable in the inversion list's
8104 * array, without having to realloc() */
8106 PERL_ARGS_ASSERT_INVLIST_MAX;
8108 assert(SvTYPE(invlist) == SVt_INVLIST);
8110 /* Assumes worst case, in which the 0 element is not counted in the
8111 * inversion list, so subtracts 1 for that */
8112 return SvLEN(invlist) == 0 /* This happens under _new_invlist_C_array */
8113 ? FROM_INTERNAL_SIZE(SvCUR(invlist)) - 1
8114 : FROM_INTERNAL_SIZE(SvLEN(invlist)) - 1;
8117 #ifndef PERL_IN_XSUB_RE
8119 Perl__new_invlist(pTHX_ IV initial_size)
8122 /* Return a pointer to a newly constructed inversion list, with enough
8123 * space to store 'initial_size' elements. If that number is negative, a
8124 * system default is used instead */
8128 if (initial_size < 0) {
8132 /* Allocate the initial space */
8133 new_list = newSV_type(SVt_INVLIST);
8135 /* First 1 is in case the zero element isn't in the list; second 1 is for
8137 SvGROW(new_list, TO_INTERNAL_SIZE(initial_size + 1) + 1);
8138 invlist_set_len(new_list, 0, 0);
8140 /* Force iterinit() to be used to get iteration to work */
8141 *get_invlist_iter_addr(new_list) = (STRLEN) UV_MAX;
8143 *get_invlist_previous_index_addr(new_list) = 0;
8149 Perl__new_invlist_C_array(pTHX_ const UV* const list)
8151 /* Return a pointer to a newly constructed inversion list, initialized to
8152 * point to <list>, which has to be in the exact correct inversion list
8153 * form, including internal fields. Thus this is a dangerous routine that
8154 * should not be used in the wrong hands. The passed in 'list' contains
8155 * several header fields at the beginning that are not part of the
8156 * inversion list body proper */
8158 const STRLEN length = (STRLEN) list[0];
8159 const UV version_id = list[1];
8160 const bool offset = cBOOL(list[2]);
8161 #define HEADER_LENGTH 3
8162 /* If any of the above changes in any way, you must change HEADER_LENGTH
8163 * (if appropriate) and regenerate INVLIST_VERSION_ID by running
8164 * perl -E 'say int(rand 2**31-1)'
8166 #define INVLIST_VERSION_ID 148565664 /* This is a combination of a version and
8167 data structure type, so that one being
8168 passed in can be validated to be an
8169 inversion list of the correct vintage.
8172 SV* invlist = newSV_type(SVt_INVLIST);
8174 PERL_ARGS_ASSERT__NEW_INVLIST_C_ARRAY;
8176 if (version_id != INVLIST_VERSION_ID) {
8177 Perl_croak(aTHX_ "panic: Incorrect version for previously generated inversion list");
8180 /* The generated array passed in includes header elements that aren't part
8181 * of the list proper, so start it just after them */
8182 SvPV_set(invlist, (char *) (list + HEADER_LENGTH));
8184 SvLEN_set(invlist, 0); /* Means we own the contents, and the system
8185 shouldn't touch it */
8187 *(get_invlist_offset_addr(invlist)) = offset;
8189 /* The 'length' passed to us is the physical number of elements in the
8190 * inversion list. But if there is an offset the logical number is one
8192 invlist_set_len(invlist, length - offset, offset);
8194 invlist_set_previous_index(invlist, 0);
8196 /* Initialize the iteration pointer. */
8197 invlist_iterfinish(invlist);
8199 SvREADONLY_on(invlist);
8203 #endif /* ifndef PERL_IN_XSUB_RE */
8206 S_invlist_extend(pTHX_ SV* const invlist, const UV new_max)
8208 /* Grow the maximum size of an inversion list */
8210 PERL_ARGS_ASSERT_INVLIST_EXTEND;
8212 assert(SvTYPE(invlist) == SVt_INVLIST);
8214 /* Add one to account for the zero element at the beginning which may not
8215 * be counted by the calling parameters */
8216 SvGROW((SV *)invlist, TO_INTERNAL_SIZE(new_max + 1));
8220 S__append_range_to_invlist(pTHX_ SV* const invlist,
8221 const UV start, const UV end)
8223 /* Subject to change or removal. Append the range from 'start' to 'end' at
8224 * the end of the inversion list. The range must be above any existing
8228 UV max = invlist_max(invlist);
8229 UV len = _invlist_len(invlist);
8232 PERL_ARGS_ASSERT__APPEND_RANGE_TO_INVLIST;
8234 if (len == 0) { /* Empty lists must be initialized */
8235 offset = start != 0;
8236 array = _invlist_array_init(invlist, ! offset);
8239 /* Here, the existing list is non-empty. The current max entry in the
8240 * list is generally the first value not in the set, except when the
8241 * set extends to the end of permissible values, in which case it is
8242 * the first entry in that final set, and so this call is an attempt to
8243 * append out-of-order */
8245 UV final_element = len - 1;
8246 array = invlist_array(invlist);
8247 if (array[final_element] > start
8248 || ELEMENT_RANGE_MATCHES_INVLIST(final_element))
8250 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",
8251 array[final_element], start,
8252 ELEMENT_RANGE_MATCHES_INVLIST(final_element) ? 't' : 'f');
8255 /* Here, it is a legal append. If the new range begins with the first
8256 * value not in the set, it is extending the set, so the new first
8257 * value not in the set is one greater than the newly extended range.
8259 offset = *get_invlist_offset_addr(invlist);
8260 if (array[final_element] == start) {
8261 if (end != UV_MAX) {
8262 array[final_element] = end + 1;
8265 /* But if the end is the maximum representable on the machine,
8266 * just let the range that this would extend to have no end */
8267 invlist_set_len(invlist, len - 1, offset);
8273 /* Here the new range doesn't extend any existing set. Add it */
8275 len += 2; /* Includes an element each for the start and end of range */
8277 /* If wll overflow the existing space, extend, which may cause the array to
8280 invlist_extend(invlist, len);
8282 /* Have to set len here to avoid assert failure in invlist_array() */
8283 invlist_set_len(invlist, len, offset);
8285 array = invlist_array(invlist);
8288 invlist_set_len(invlist, len, offset);
8291 /* The next item on the list starts the range, the one after that is
8292 * one past the new range. */
8293 array[len - 2] = start;
8294 if (end != UV_MAX) {
8295 array[len - 1] = end + 1;
8298 /* But if the end is the maximum representable on the machine, just let
8299 * the range have no end */
8300 invlist_set_len(invlist, len - 1, offset);
8304 #ifndef PERL_IN_XSUB_RE
8307 Perl__invlist_search(SV* const invlist, const UV cp)
8309 /* Searches the inversion list for the entry that contains the input code
8310 * point <cp>. If <cp> is not in the list, -1 is returned. Otherwise, the
8311 * return value is the index into the list's array of the range that
8316 IV high = _invlist_len(invlist);
8317 const IV highest_element = high - 1;
8320 PERL_ARGS_ASSERT__INVLIST_SEARCH;
8322 /* If list is empty, return failure. */
8327 /* (We can't get the array unless we know the list is non-empty) */
8328 array = invlist_array(invlist);
8330 mid = invlist_previous_index(invlist);
8331 assert(mid >=0 && mid <= highest_element);
8333 /* <mid> contains the cache of the result of the previous call to this
8334 * function (0 the first time). See if this call is for the same result,
8335 * or if it is for mid-1. This is under the theory that calls to this
8336 * function will often be for related code points that are near each other.
8337 * And benchmarks show that caching gives better results. We also test
8338 * here if the code point is within the bounds of the list. These tests
8339 * replace others that would have had to be made anyway to make sure that
8340 * the array bounds were not exceeded, and these give us extra information
8341 * at the same time */
8342 if (cp >= array[mid]) {
8343 if (cp >= array[highest_element]) {
8344 return highest_element;
8347 /* Here, array[mid] <= cp < array[highest_element]. This means that
8348 * the final element is not the answer, so can exclude it; it also
8349 * means that <mid> is not the final element, so can refer to 'mid + 1'
8351 if (cp < array[mid + 1]) {
8357 else { /* cp < aray[mid] */
8358 if (cp < array[0]) { /* Fail if outside the array */
8362 if (cp >= array[mid - 1]) {
8367 /* Binary search. What we are looking for is <i> such that
8368 * array[i] <= cp < array[i+1]
8369 * The loop below converges on the i+1. Note that there may not be an
8370 * (i+1)th element in the array, and things work nonetheless */
8371 while (low < high) {
8372 mid = (low + high) / 2;
8373 assert(mid <= highest_element);
8374 if (array[mid] <= cp) { /* cp >= array[mid] */
8377 /* We could do this extra test to exit the loop early.
8378 if (cp < array[low]) {
8383 else { /* cp < array[mid] */
8390 invlist_set_previous_index(invlist, high);
8395 Perl__invlist_populate_swatch(SV* const invlist,
8396 const UV start, const UV end, U8* swatch)
8398 /* populates a swatch of a swash the same way swatch_get() does in utf8.c,
8399 * but is used when the swash has an inversion list. This makes this much
8400 * faster, as it uses a binary search instead of a linear one. This is
8401 * intimately tied to that function, and perhaps should be in utf8.c,
8402 * except it is intimately tied to inversion lists as well. It assumes
8403 * that <swatch> is all 0's on input */
8406 const IV len = _invlist_len(invlist);
8410 PERL_ARGS_ASSERT__INVLIST_POPULATE_SWATCH;
8412 if (len == 0) { /* Empty inversion list */
8416 array = invlist_array(invlist);
8418 /* Find which element it is */
8419 i = _invlist_search(invlist, start);
8421 /* We populate from <start> to <end> */
8422 while (current < end) {
8425 /* The inversion list gives the results for every possible code point
8426 * after the first one in the list. Only those ranges whose index is
8427 * even are ones that the inversion list matches. For the odd ones,
8428 * and if the initial code point is not in the list, we have to skip
8429 * forward to the next element */
8430 if (i == -1 || ! ELEMENT_RANGE_MATCHES_INVLIST(i)) {
8432 if (i >= len) { /* Finished if beyond the end of the array */
8436 if (current >= end) { /* Finished if beyond the end of what we
8438 if (LIKELY(end < UV_MAX)) {
8442 /* We get here when the upper bound is the maximum
8443 * representable on the machine, and we are looking for just
8444 * that code point. Have to special case it */
8446 goto join_end_of_list;
8449 assert(current >= start);
8451 /* The current range ends one below the next one, except don't go past
8454 upper = (i < len && array[i] < end) ? array[i] : end;
8456 /* Here we are in a range that matches. Populate a bit in the 3-bit U8
8457 * for each code point in it */
8458 for (; current < upper; current++) {
8459 const STRLEN offset = (STRLEN)(current - start);
8460 swatch[offset >> 3] |= 1 << (offset & 7);
8465 /* Quit if at the end of the list */
8468 /* But first, have to deal with the highest possible code point on
8469 * the platform. The previous code assumes that <end> is one
8470 * beyond where we want to populate, but that is impossible at the
8471 * platform's infinity, so have to handle it specially */
8472 if (UNLIKELY(end == UV_MAX && ELEMENT_RANGE_MATCHES_INVLIST(len-1)))
8474 const STRLEN offset = (STRLEN)(end - start);
8475 swatch[offset >> 3] |= 1 << (offset & 7);
8480 /* Advance to the next range, which will be for code points not in the
8489 Perl__invlist_union_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8490 const bool complement_b, SV** output)
8492 /* Take the union of two inversion lists and point <output> to it. *output
8493 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8494 * the reference count to that list will be decremented if not already a
8495 * temporary (mortal); otherwise *output will be made correspondingly
8496 * mortal. The first list, <a>, may be NULL, in which case a copy of the
8497 * second list is returned. If <complement_b> is TRUE, the union is taken
8498 * of the complement (inversion) of <b> instead of b itself.
8500 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8501 * Richard Gillam, published by Addison-Wesley, and explained at some
8502 * length there. The preface says to incorporate its examples into your
8503 * code at your own risk.
8505 * The algorithm is like a merge sort.
8507 * XXX A potential performance improvement is to keep track as we go along
8508 * if only one of the inputs contributes to the result, meaning the other
8509 * is a subset of that one. In that case, we can skip the final copy and
8510 * return the larger of the input lists, but then outside code might need
8511 * to keep track of whether to free the input list or not */
8513 const UV* array_a; /* a's array */
8515 UV len_a; /* length of a's array */
8518 SV* u; /* the resulting union */
8522 UV i_a = 0; /* current index into a's array */
8526 /* running count, as explained in the algorithm source book; items are
8527 * stopped accumulating and are output when the count changes to/from 0.
8528 * The count is incremented when we start a range that's in the set, and
8529 * decremented when we start a range that's not in the set. So its range
8530 * is 0 to 2. Only when the count is zero is something not in the set.
8534 PERL_ARGS_ASSERT__INVLIST_UNION_MAYBE_COMPLEMENT_2ND;
8537 /* If either one is empty, the union is the other one */
8538 if (a == NULL || ((len_a = _invlist_len(a)) == 0)) {
8539 bool make_temp = FALSE; /* Should we mortalize the result? */
8543 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8549 *output = invlist_clone(b);
8551 _invlist_invert(*output);
8553 } /* else *output already = b; */
8556 sv_2mortal(*output);
8560 else if ((len_b = _invlist_len(b)) == 0) {
8561 bool make_temp = FALSE;
8563 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8568 /* The complement of an empty list is a list that has everything in it,
8569 * so the union with <a> includes everything too */
8572 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8576 *output = _new_invlist(1);
8577 _append_range_to_invlist(*output, 0, UV_MAX);
8579 else if (*output != a) {
8580 *output = invlist_clone(a);
8582 /* else *output already = a; */
8585 sv_2mortal(*output);
8590 /* Here both lists exist and are non-empty */
8591 array_a = invlist_array(a);
8592 array_b = invlist_array(b);
8594 /* If are to take the union of 'a' with the complement of b, set it
8595 * up so are looking at b's complement. */
8598 /* To complement, we invert: if the first element is 0, remove it. To
8599 * do this, we just pretend the array starts one later */
8600 if (array_b[0] == 0) {
8606 /* But if the first element is not zero, we pretend the list starts
8607 * at the 0 that is always stored immediately before the array. */
8613 /* Size the union for the worst case: that the sets are completely
8615 u = _new_invlist(len_a + len_b);
8617 /* Will contain U+0000 if either component does */
8618 array_u = _invlist_array_init(u, (len_a > 0 && array_a[0] == 0)
8619 || (len_b > 0 && array_b[0] == 0));
8621 /* Go through each list item by item, stopping when exhausted one of
8623 while (i_a < len_a && i_b < len_b) {
8624 UV cp; /* The element to potentially add to the union's array */
8625 bool cp_in_set; /* is it in the the input list's set or not */
8627 /* We need to take one or the other of the two inputs for the union.
8628 * Since we are merging two sorted lists, we take the smaller of the
8629 * next items. In case of a tie, we take the one that is in its set
8630 * first. If we took one not in the set first, it would decrement the
8631 * count, possibly to 0 which would cause it to be output as ending the
8632 * range, and the next time through we would take the same number, and
8633 * output it again as beginning the next range. By doing it the
8634 * opposite way, there is no possibility that the count will be
8635 * momentarily decremented to 0, and thus the two adjoining ranges will
8636 * be seamlessly merged. (In a tie and both are in the set or both not
8637 * in the set, it doesn't matter which we take first.) */
8638 if (array_a[i_a] < array_b[i_b]
8639 || (array_a[i_a] == array_b[i_b]
8640 && ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8642 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8646 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8647 cp = array_b[i_b++];
8650 /* Here, have chosen which of the two inputs to look at. Only output
8651 * if the running count changes to/from 0, which marks the
8652 * beginning/end of a range in that's in the set */
8655 array_u[i_u++] = cp;
8662 array_u[i_u++] = cp;
8667 /* Here, we are finished going through at least one of the lists, which
8668 * means there is something remaining in at most one. We check if the list
8669 * that hasn't been exhausted is positioned such that we are in the middle
8670 * of a range in its set or not. (i_a and i_b point to the element beyond
8671 * the one we care about.) If in the set, we decrement 'count'; if 0, there
8672 * is potentially more to output.
8673 * There are four cases:
8674 * 1) Both weren't in their sets, count is 0, and remains 0. What's left
8675 * in the union is entirely from the non-exhausted set.
8676 * 2) Both were in their sets, count is 2. Nothing further should
8677 * be output, as everything that remains will be in the exhausted
8678 * list's set, hence in the union; decrementing to 1 but not 0 insures
8680 * 3) the exhausted was in its set, non-exhausted isn't, count is 1.
8681 * Nothing further should be output because the union includes
8682 * everything from the exhausted set. Not decrementing ensures that.
8683 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1;
8684 * decrementing to 0 insures that we look at the remainder of the
8685 * non-exhausted set */
8686 if ((i_a != len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8687 || (i_b != len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8692 /* The final length is what we've output so far, plus what else is about to
8693 * be output. (If 'count' is non-zero, then the input list we exhausted
8694 * has everything remaining up to the machine's limit in its set, and hence
8695 * in the union, so there will be no further output. */
8698 /* At most one of the subexpressions will be non-zero */
8699 len_u += (len_a - i_a) + (len_b - i_b);
8702 /* Set result to final length, which can change the pointer to array_u, so
8704 if (len_u != _invlist_len(u)) {
8705 invlist_set_len(u, len_u, *get_invlist_offset_addr(u));
8707 array_u = invlist_array(u);
8710 /* When 'count' is 0, the list that was exhausted (if one was shorter than
8711 * the other) ended with everything above it not in its set. That means
8712 * that the remaining part of the union is precisely the same as the
8713 * non-exhausted list, so can just copy it unchanged. (If both list were
8714 * exhausted at the same time, then the operations below will be both 0.)
8717 IV copy_count; /* At most one will have a non-zero copy count */
8718 if ((copy_count = len_a - i_a) > 0) {
8719 Copy(array_a + i_a, array_u + i_u, copy_count, UV);
8721 else if ((copy_count = len_b - i_b) > 0) {
8722 Copy(array_b + i_b, array_u + i_u, copy_count, UV);
8726 /* We may be removing a reference to one of the inputs. If so, the output
8727 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8728 * count decremented) */
8729 if (a == *output || b == *output) {
8730 assert(! invlist_is_iterating(*output));
8731 if ((SvTEMP(*output))) {
8735 SvREFCNT_dec_NN(*output);
8745 Perl__invlist_intersection_maybe_complement_2nd(pTHX_ SV* const a, SV* const b,
8746 const bool complement_b, SV** i)
8748 /* Take the intersection of two inversion lists and point <i> to it. *i
8749 * SHOULD BE DEFINED upon input, and if it points to one of the two lists,
8750 * the reference count to that list will be decremented if not already a
8751 * temporary (mortal); otherwise *i will be made correspondingly mortal.
8752 * The first list, <a>, may be NULL, in which case an empty list is
8753 * returned. If <complement_b> is TRUE, the result will be the
8754 * intersection of <a> and the complement (or inversion) of <b> instead of
8757 * The basis for this comes from "Unicode Demystified" Chapter 13 by
8758 * Richard Gillam, published by Addison-Wesley, and explained at some
8759 * length there. The preface says to incorporate its examples into your
8760 * code at your own risk. In fact, it had bugs
8762 * The algorithm is like a merge sort, and is essentially the same as the
8766 const UV* array_a; /* a's array */
8768 UV len_a; /* length of a's array */
8771 SV* r; /* the resulting intersection */
8775 UV i_a = 0; /* current index into a's array */
8779 /* running count, as explained in the algorithm source book; items are
8780 * stopped accumulating and are output when the count changes to/from 2.
8781 * The count is incremented when we start a range that's in the set, and
8782 * decremented when we start a range that's not in the set. So its range
8783 * is 0 to 2. Only when the count is 2 is something in the intersection.
8787 PERL_ARGS_ASSERT__INVLIST_INTERSECTION_MAYBE_COMPLEMENT_2ND;
8790 /* Special case if either one is empty */
8791 len_a = (a == NULL) ? 0 : _invlist_len(a);
8792 if ((len_a == 0) || ((len_b = _invlist_len(b)) == 0)) {
8793 bool make_temp = FALSE;
8795 if (len_a != 0 && complement_b) {
8797 /* Here, 'a' is not empty, therefore from the above 'if', 'b' must
8798 * be empty. Here, also we are using 'b's complement, which hence
8799 * must be every possible code point. Thus the intersection is
8803 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8808 *i = invlist_clone(a);
8810 /* else *i is already 'a' */
8818 /* Here, 'a' or 'b' is empty and not using the complement of 'b'. The
8819 * intersection must be empty */
8821 if (! (make_temp = cBOOL(SvTEMP(a)))) {
8826 if (! (make_temp = cBOOL(SvTEMP(b)))) {
8830 *i = _new_invlist(0);
8838 /* Here both lists exist and are non-empty */
8839 array_a = invlist_array(a);
8840 array_b = invlist_array(b);
8842 /* If are to take the intersection of 'a' with the complement of b, set it
8843 * up so are looking at b's complement. */
8846 /* To complement, we invert: if the first element is 0, remove it. To
8847 * do this, we just pretend the array starts one later */
8848 if (array_b[0] == 0) {
8854 /* But if the first element is not zero, we pretend the list starts
8855 * at the 0 that is always stored immediately before the array. */
8861 /* Size the intersection for the worst case: that the intersection ends up
8862 * fragmenting everything to be completely disjoint */
8863 r= _new_invlist(len_a + len_b);
8865 /* Will contain U+0000 iff both components do */
8866 array_r = _invlist_array_init(r, len_a > 0 && array_a[0] == 0
8867 && len_b > 0 && array_b[0] == 0);
8869 /* Go through each list item by item, stopping when exhausted one of
8871 while (i_a < len_a && i_b < len_b) {
8872 UV cp; /* The element to potentially add to the intersection's
8874 bool cp_in_set; /* Is it in the input list's set or not */
8876 /* We need to take one or the other of the two inputs for the
8877 * intersection. Since we are merging two sorted lists, we take the
8878 * smaller of the next items. In case of a tie, we take the one that
8879 * is not in its set first (a difference from the union algorithm). If
8880 * we took one in the set first, it would increment the count, possibly
8881 * to 2 which would cause it to be output as starting a range in the
8882 * intersection, and the next time through we would take that same
8883 * number, and output it again as ending the set. By doing it the
8884 * opposite of this, there is no possibility that the count will be
8885 * momentarily incremented to 2. (In a tie and both are in the set or
8886 * both not in the set, it doesn't matter which we take first.) */
8887 if (array_a[i_a] < array_b[i_b]
8888 || (array_a[i_a] == array_b[i_b]
8889 && ! ELEMENT_RANGE_MATCHES_INVLIST(i_a)))
8891 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_a);
8895 cp_in_set = ELEMENT_RANGE_MATCHES_INVLIST(i_b);
8899 /* Here, have chosen which of the two inputs to look at. Only output
8900 * if the running count changes to/from 2, which marks the
8901 * beginning/end of a range that's in the intersection */
8905 array_r[i_r++] = cp;
8910 array_r[i_r++] = cp;
8916 /* Here, we are finished going through at least one of the lists, which
8917 * means there is something remaining in at most one. We check if the list
8918 * that has been exhausted is positioned such that we are in the middle
8919 * of a range in its set or not. (i_a and i_b point to elements 1 beyond
8920 * the ones we care about.) There are four cases:
8921 * 1) Both weren't in their sets, count is 0, and remains 0. There's
8922 * nothing left in the intersection.
8923 * 2) Both were in their sets, count is 2 and perhaps is incremented to
8924 * above 2. What should be output is exactly that which is in the
8925 * non-exhausted set, as everything it has is also in the intersection
8926 * set, and everything it doesn't have can't be in the intersection
8927 * 3) The exhausted was in its set, non-exhausted isn't, count is 1, and
8928 * gets incremented to 2. Like the previous case, the intersection is
8929 * everything that remains in the non-exhausted set.
8930 * 4) the exhausted wasn't in its set, non-exhausted is, count is 1, and
8931 * remains 1. And the intersection has nothing more. */
8932 if ((i_a == len_a && PREV_RANGE_MATCHES_INVLIST(i_a))
8933 || (i_b == len_b && PREV_RANGE_MATCHES_INVLIST(i_b)))
8938 /* The final length is what we've output so far plus what else is in the
8939 * intersection. At most one of the subexpressions below will be non-zero
8943 len_r += (len_a - i_a) + (len_b - i_b);
8946 /* Set result to final length, which can change the pointer to array_r, so
8948 if (len_r != _invlist_len(r)) {
8949 invlist_set_len(r, len_r, *get_invlist_offset_addr(r));
8951 array_r = invlist_array(r);
8954 /* Finish outputting any remaining */
8955 if (count >= 2) { /* At most one will have a non-zero copy count */
8957 if ((copy_count = len_a - i_a) > 0) {
8958 Copy(array_a + i_a, array_r + i_r, copy_count, UV);
8960 else if ((copy_count = len_b - i_b) > 0) {
8961 Copy(array_b + i_b, array_r + i_r, copy_count, UV);
8965 /* We may be removing a reference to one of the inputs. If so, the output
8966 * is made mortal if the input was. (Mortal SVs shouldn't have their ref
8967 * count decremented) */
8968 if (a == *i || b == *i) {
8969 assert(! invlist_is_iterating(*i));
8974 SvREFCNT_dec_NN(*i);
8984 Perl__add_range_to_invlist(pTHX_ SV* invlist, const UV start, const UV end)
8986 /* Add the range from 'start' to 'end' inclusive to the inversion list's
8987 * set. A pointer to the inversion list is returned. This may actually be
8988 * a new list, in which case the passed in one has been destroyed. The
8989 * passed-in inversion list can be NULL, in which case a new one is created
8990 * with just the one range in it */
8995 if (invlist == NULL) {
8996 invlist = _new_invlist(2);
9000 len = _invlist_len(invlist);
9003 /* If comes after the final entry actually in the list, can just append it
9006 || (! ELEMENT_RANGE_MATCHES_INVLIST(len - 1)
9007 && start >= invlist_array(invlist)[len - 1]))
9009 _append_range_to_invlist(invlist, start, end);
9013 /* Here, can't just append things, create and return a new inversion list
9014 * which is the union of this range and the existing inversion list */
9015 range_invlist = _new_invlist(2);
9016 _append_range_to_invlist(range_invlist, start, end);
9018 _invlist_union(invlist, range_invlist, &invlist);
9020 /* The temporary can be freed */
9021 SvREFCNT_dec_NN(range_invlist);
9027 Perl__setup_canned_invlist(pTHX_ const STRLEN size, const UV element0,
9028 UV** other_elements_ptr)
9030 /* Create and return an inversion list whose contents are to be populated
9031 * by the caller. The caller gives the number of elements (in 'size') and
9032 * the very first element ('element0'). This function will set
9033 * '*other_elements_ptr' to an array of UVs, where the remaining elements
9036 * Obviously there is some trust involved that the caller will properly
9037 * fill in the other elements of the array.
9039 * (The first element needs to be passed in, as the underlying code does
9040 * things differently depending on whether it is zero or non-zero) */
9042 SV* invlist = _new_invlist(size);
9045 PERL_ARGS_ASSERT__SETUP_CANNED_INVLIST;
9047 _append_range_to_invlist(invlist, element0, element0);
9048 offset = *get_invlist_offset_addr(invlist);
9050 invlist_set_len(invlist, size, offset);
9051 *other_elements_ptr = invlist_array(invlist) + 1;
9057 PERL_STATIC_INLINE SV*
9058 S_add_cp_to_invlist(pTHX_ SV* invlist, const UV cp) {
9059 return _add_range_to_invlist(invlist, cp, cp);
9062 #ifndef PERL_IN_XSUB_RE
9064 Perl__invlist_invert(pTHX_ SV* const invlist)
9066 /* Complement the input inversion list. This adds a 0 if the list didn't
9067 * have a zero; removes it otherwise. As described above, the data
9068 * structure is set up so that this is very efficient */
9070 PERL_ARGS_ASSERT__INVLIST_INVERT;
9072 assert(! invlist_is_iterating(invlist));
9074 /* The inverse of matching nothing is matching everything */
9075 if (_invlist_len(invlist) == 0) {
9076 _append_range_to_invlist(invlist, 0, UV_MAX);
9080 *get_invlist_offset_addr(invlist) = ! *get_invlist_offset_addr(invlist);
9085 PERL_STATIC_INLINE SV*
9086 S_invlist_clone(pTHX_ SV* const invlist)
9089 /* Return a new inversion list that is a copy of the input one, which is
9090 * unchanged. The new list will not be mortal even if the old one was. */
9092 /* Need to allocate extra space to accommodate Perl's addition of a
9093 * trailing NUL to SvPV's, since it thinks they are always strings */
9094 SV* new_invlist = _new_invlist(_invlist_len(invlist) + 1);
9095 STRLEN physical_length = SvCUR(invlist);
9096 bool offset = *(get_invlist_offset_addr(invlist));
9098 PERL_ARGS_ASSERT_INVLIST_CLONE;
9100 *(get_invlist_offset_addr(new_invlist)) = offset;
9101 invlist_set_len(new_invlist, _invlist_len(invlist), offset);
9102 Copy(SvPVX(invlist), SvPVX(new_invlist), physical_length, char);
9107 PERL_STATIC_INLINE STRLEN*
9108 S_get_invlist_iter_addr(SV* invlist)
9110 /* Return the address of the UV that contains the current iteration
9113 PERL_ARGS_ASSERT_GET_INVLIST_ITER_ADDR;
9115 assert(SvTYPE(invlist) == SVt_INVLIST);
9117 return &(((XINVLIST*) SvANY(invlist))->iterator);
9120 PERL_STATIC_INLINE void
9121 S_invlist_iterinit(SV* invlist) /* Initialize iterator for invlist */
9123 PERL_ARGS_ASSERT_INVLIST_ITERINIT;
9125 *get_invlist_iter_addr(invlist) = 0;
9128 PERL_STATIC_INLINE void
9129 S_invlist_iterfinish(SV* invlist)
9131 /* Terminate iterator for invlist. This is to catch development errors.
9132 * Any iteration that is interrupted before completed should call this
9133 * function. Functions that add code points anywhere else but to the end
9134 * of an inversion list assert that they are not in the middle of an
9135 * iteration. If they were, the addition would make the iteration
9136 * problematical: if the iteration hadn't reached the place where things
9137 * were being added, it would be ok */
9139 PERL_ARGS_ASSERT_INVLIST_ITERFINISH;
9141 *get_invlist_iter_addr(invlist) = (STRLEN) UV_MAX;
9145 S_invlist_iternext(SV* invlist, UV* start, UV* end)
9147 /* An C<invlist_iterinit> call on <invlist> must be used to set this up.
9148 * This call sets in <*start> and <*end>, the next range in <invlist>.
9149 * Returns <TRUE> if successful and the next call will return the next
9150 * range; <FALSE> if was already at the end of the list. If the latter,
9151 * <*start> and <*end> are unchanged, and the next call to this function
9152 * will start over at the beginning of the list */
9154 STRLEN* pos = get_invlist_iter_addr(invlist);
9155 UV len = _invlist_len(invlist);
9158 PERL_ARGS_ASSERT_INVLIST_ITERNEXT;
9161 *pos = (STRLEN) UV_MAX; /* Force iterinit() to be required next time */
9165 array = invlist_array(invlist);
9167 *start = array[(*pos)++];
9173 *end = array[(*pos)++] - 1;
9179 PERL_STATIC_INLINE UV
9180 S_invlist_highest(SV* const invlist)
9182 /* Returns the highest code point that matches an inversion list. This API
9183 * has an ambiguity, as it returns 0 under either the highest is actually
9184 * 0, or if the list is empty. If this distinction matters to you, check
9185 * for emptiness before calling this function */
9187 UV len = _invlist_len(invlist);
9190 PERL_ARGS_ASSERT_INVLIST_HIGHEST;
9196 array = invlist_array(invlist);
9198 /* The last element in the array in the inversion list always starts a
9199 * range that goes to infinity. That range may be for code points that are
9200 * matched in the inversion list, or it may be for ones that aren't
9201 * matched. In the latter case, the highest code point in the set is one
9202 * less than the beginning of this range; otherwise it is the final element
9203 * of this range: infinity */
9204 return (ELEMENT_RANGE_MATCHES_INVLIST(len - 1))
9206 : array[len - 1] - 1;
9209 #ifndef PERL_IN_XSUB_RE
9211 Perl__invlist_contents(pTHX_ SV* const invlist)
9213 /* Get the contents of an inversion list into a string SV so that they can
9214 * be printed out. It uses the format traditionally done for debug tracing
9218 SV* output = newSVpvs("\n");
9220 PERL_ARGS_ASSERT__INVLIST_CONTENTS;
9222 assert(! invlist_is_iterating(invlist));
9224 invlist_iterinit(invlist);
9225 while (invlist_iternext(invlist, &start, &end)) {
9226 if (end == UV_MAX) {
9227 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\tINFINITY\n", start);
9229 else if (end != start) {
9230 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\t%04"UVXf"\n",
9234 Perl_sv_catpvf(aTHX_ output, "%04"UVXf"\n", start);
9242 #ifndef PERL_IN_XSUB_RE
9244 Perl__invlist_dump(pTHX_ PerlIO *file, I32 level,
9245 const char * const indent, SV* const invlist)
9247 /* Designed to be called only by do_sv_dump(). Dumps out the ranges of the
9248 * inversion list 'invlist' to 'file' at 'level' Each line is prefixed by
9249 * the string 'indent'. The output looks like this:
9250 [0] 0x000A .. 0x000D
9252 [4] 0x2028 .. 0x2029
9253 [6] 0x3104 .. INFINITY
9254 * This means that the first range of code points matched by the list are
9255 * 0xA through 0xD; the second range contains only the single code point
9256 * 0x85, etc. An inversion list is an array of UVs. Two array elements
9257 * are used to define each range (except if the final range extends to
9258 * infinity, only a single element is needed). The array index of the
9259 * first element for the corresponding range is given in brackets. */
9264 PERL_ARGS_ASSERT__INVLIST_DUMP;
9266 if (invlist_is_iterating(invlist)) {
9267 Perl_dump_indent(aTHX_ level, file,
9268 "%sCan't dump inversion list because is in middle of iterating\n",
9273 invlist_iterinit(invlist);
9274 while (invlist_iternext(invlist, &start, &end)) {
9275 if (end == UV_MAX) {
9276 Perl_dump_indent(aTHX_ level, file,
9277 "%s[%"UVuf"] 0x%04"UVXf" .. INFINITY\n",
9278 indent, (UV)count, start);
9280 else if (end != start) {
9281 Perl_dump_indent(aTHX_ level, file,
9282 "%s[%"UVuf"] 0x%04"UVXf" .. 0x%04"UVXf"\n",
9283 indent, (UV)count, start, end);
9286 Perl_dump_indent(aTHX_ level, file, "%s[%"UVuf"] 0x%04"UVXf"\n",
9287 indent, (UV)count, start);
9294 Perl__load_PL_utf8_foldclosures (pTHX)
9296 assert(! PL_utf8_foldclosures);
9298 /* If the folds haven't been read in, call a fold function
9300 if (! PL_utf8_tofold) {
9301 U8 dummy[UTF8_MAXBYTES_CASE+1];
9303 /* This string is just a short named one above \xff */
9304 to_utf8_fold((U8*) HYPHEN_UTF8, dummy, NULL);
9305 assert(PL_utf8_tofold); /* Verify that worked */
9307 PL_utf8_foldclosures = _swash_inversion_hash(PL_utf8_tofold);
9311 #ifdef PERL_ARGS_ASSERT__INVLISTEQ
9313 S__invlistEQ(pTHX_ SV* const a, SV* const b, const bool complement_b)
9315 /* Return a boolean as to if the two passed in inversion lists are
9316 * identical. The final argument, if TRUE, says to take the complement of
9317 * the second inversion list before doing the comparison */
9319 const UV* array_a = invlist_array(a);
9320 const UV* array_b = invlist_array(b);
9321 UV len_a = _invlist_len(a);
9322 UV len_b = _invlist_len(b);
9324 UV i = 0; /* current index into the arrays */
9325 bool retval = TRUE; /* Assume are identical until proven otherwise */
9327 PERL_ARGS_ASSERT__INVLISTEQ;
9329 /* If are to compare 'a' with the complement of b, set it
9330 * up so are looking at b's complement. */
9333 /* The complement of nothing is everything, so <a> would have to have
9334 * just one element, starting at zero (ending at infinity) */
9336 return (len_a == 1 && array_a[0] == 0);
9338 else if (array_b[0] == 0) {
9340 /* Otherwise, to complement, we invert. Here, the first element is
9341 * 0, just remove it. To do this, we just pretend the array starts
9349 /* But if the first element is not zero, we pretend the list starts
9350 * at the 0 that is always stored immediately before the array. */
9356 /* Make sure that the lengths are the same, as well as the final element
9357 * before looping through the remainder. (Thus we test the length, final,
9358 * and first elements right off the bat) */
9359 if (len_a != len_b || array_a[len_a-1] != array_b[len_a-1]) {
9362 else for (i = 0; i < len_a - 1; i++) {
9363 if (array_a[i] != array_b[i]) {
9374 * As best we can, determine the characters that can match the start of
9375 * the given EXACTF-ish node.
9377 * Returns the invlist as a new SV*; it is the caller's responsibility to
9378 * call SvREFCNT_dec() when done with it.
9381 S__make_exactf_invlist(pTHX_ RExC_state_t *pRExC_state, regnode *node)
9383 const U8 * s = (U8*)STRING(node);
9384 SSize_t bytelen = STR_LEN(node);
9386 /* Start out big enough for 2 separate code points */
9387 SV* invlist = _new_invlist(4);
9389 PERL_ARGS_ASSERT__MAKE_EXACTF_INVLIST;
9394 /* We punt and assume can match anything if the node begins
9395 * with a multi-character fold. Things are complicated. For
9396 * example, /ffi/i could match any of:
9397 * "\N{LATIN SMALL LIGATURE FFI}"
9398 * "\N{LATIN SMALL LIGATURE FF}I"
9399 * "F\N{LATIN SMALL LIGATURE FI}"
9400 * plus several other things; and making sure we have all the
9401 * possibilities is hard. */
9402 if (is_MULTI_CHAR_FOLD_latin1_safe(s, s + bytelen)) {
9403 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9406 /* Any Latin1 range character can potentially match any
9407 * other depending on the locale */
9408 if (OP(node) == EXACTFL) {
9409 _invlist_union(invlist, PL_Latin1, &invlist);
9412 /* But otherwise, it matches at least itself. We can
9413 * quickly tell if it has a distinct fold, and if so,
9414 * it matches that as well */
9415 invlist = add_cp_to_invlist(invlist, uc);
9416 if (IS_IN_SOME_FOLD_L1(uc))
9417 invlist = add_cp_to_invlist(invlist, PL_fold_latin1[uc]);
9420 /* Some characters match above-Latin1 ones under /i. This
9421 * is true of EXACTFL ones when the locale is UTF-8 */
9422 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(uc)
9423 && (! isASCII(uc) || (OP(node) != EXACTFA
9424 && OP(node) != EXACTFA_NO_TRIE)))
9426 add_above_Latin1_folds(pRExC_state, (U8) uc, &invlist);
9430 else { /* Pattern is UTF-8 */
9431 U8 folded[UTF8_MAX_FOLD_CHAR_EXPAND * UTF8_MAXBYTES_CASE + 1] = { '\0' };
9432 STRLEN foldlen = UTF8SKIP(s);
9433 const U8* e = s + bytelen;
9436 uc = utf8_to_uvchr_buf(s, s + bytelen, NULL);
9438 /* The only code points that aren't folded in a UTF EXACTFish
9439 * node are are the problematic ones in EXACTFL nodes */
9440 if (OP(node) == EXACTFL && is_PROBLEMATIC_LOCALE_FOLDEDS_START_cp(uc)) {
9441 /* We need to check for the possibility that this EXACTFL
9442 * node begins with a multi-char fold. Therefore we fold
9443 * the first few characters of it so that we can make that
9448 for (i = 0; i < UTF8_MAX_FOLD_CHAR_EXPAND && s < e; i++) {
9450 *(d++) = (U8) toFOLD(*s);
9455 to_utf8_fold(s, d, &len);
9461 /* And set up so the code below that looks in this folded
9462 * buffer instead of the node's string */
9464 foldlen = UTF8SKIP(folded);
9468 /* When we reach here 's' points to the fold of the first
9469 * character(s) of the node; and 'e' points to far enough along
9470 * the folded string to be just past any possible multi-char
9471 * fold. 'foldlen' is the length in bytes of the first
9474 * Unlike the non-UTF-8 case, the macro for determining if a
9475 * string is a multi-char fold requires all the characters to
9476 * already be folded. This is because of all the complications
9477 * if not. Note that they are folded anyway, except in EXACTFL
9478 * nodes. Like the non-UTF case above, we punt if the node
9479 * begins with a multi-char fold */
9481 if (is_MULTI_CHAR_FOLD_utf8_safe(s, e)) {
9482 invlist = _add_range_to_invlist(invlist, 0, UV_MAX);
9484 else { /* Single char fold */
9486 /* It matches all the things that fold to it, which are
9487 * found in PL_utf8_foldclosures (including itself) */
9488 invlist = add_cp_to_invlist(invlist, uc);
9489 if (! PL_utf8_foldclosures)
9490 _load_PL_utf8_foldclosures();
9491 if ((listp = hv_fetch(PL_utf8_foldclosures,
9492 (char *) s, foldlen, FALSE)))
9494 AV* list = (AV*) *listp;
9496 for (k = 0; k <= av_tindex(list); k++) {
9497 SV** c_p = av_fetch(list, k, FALSE);
9503 /* /aa doesn't allow folds between ASCII and non- */
9504 if ((OP(node) == EXACTFA || OP(node) == EXACTFA_NO_TRIE)
9505 && isASCII(c) != isASCII(uc))
9510 invlist = add_cp_to_invlist(invlist, c);
9519 #undef HEADER_LENGTH
9520 #undef TO_INTERNAL_SIZE
9521 #undef FROM_INTERNAL_SIZE
9522 #undef INVLIST_VERSION_ID
9524 /* End of inversion list object */
9527 S_parse_lparen_question_flags(pTHX_ RExC_state_t *pRExC_state)
9529 /* This parses the flags that are in either the '(?foo)' or '(?foo:bar)'
9530 * constructs, and updates RExC_flags with them. On input, RExC_parse
9531 * should point to the first flag; it is updated on output to point to the
9532 * final ')' or ':'. There needs to be at least one flag, or this will
9535 /* for (?g), (?gc), and (?o) warnings; warning
9536 about (?c) will warn about (?g) -- japhy */
9538 #define WASTED_O 0x01
9539 #define WASTED_G 0x02
9540 #define WASTED_C 0x04
9541 #define WASTED_GC (WASTED_G|WASTED_C)
9542 I32 wastedflags = 0x00;
9543 U32 posflags = 0, negflags = 0;
9544 U32 *flagsp = &posflags;
9545 char has_charset_modifier = '\0';
9547 bool has_use_defaults = FALSE;
9548 const char* const seqstart = RExC_parse - 1; /* Point to the '?' */
9549 int x_mod_count = 0;
9551 PERL_ARGS_ASSERT_PARSE_LPAREN_QUESTION_FLAGS;
9553 /* '^' as an initial flag sets certain defaults */
9554 if (UCHARAT(RExC_parse) == '^') {
9556 has_use_defaults = TRUE;
9557 STD_PMMOD_FLAGS_CLEAR(&RExC_flags);
9558 set_regex_charset(&RExC_flags, (RExC_utf8 || RExC_uni_semantics)
9559 ? REGEX_UNICODE_CHARSET
9560 : REGEX_DEPENDS_CHARSET);
9563 cs = get_regex_charset(RExC_flags);
9564 if (cs == REGEX_DEPENDS_CHARSET
9565 && (RExC_utf8 || RExC_uni_semantics))
9567 cs = REGEX_UNICODE_CHARSET;
9570 while (*RExC_parse) {
9571 /* && strchr("iogcmsx", *RExC_parse) */
9572 /* (?g), (?gc) and (?o) are useless here
9573 and must be globally applied -- japhy */
9574 switch (*RExC_parse) {
9576 /* Code for the imsxn flags */
9577 CASE_STD_PMMOD_FLAGS_PARSE_SET(flagsp, x_mod_count);
9579 case LOCALE_PAT_MOD:
9580 if (has_charset_modifier) {
9581 goto excess_modifier;
9583 else if (flagsp == &negflags) {
9586 cs = REGEX_LOCALE_CHARSET;
9587 has_charset_modifier = LOCALE_PAT_MOD;
9589 case UNICODE_PAT_MOD:
9590 if (has_charset_modifier) {
9591 goto excess_modifier;
9593 else if (flagsp == &negflags) {
9596 cs = REGEX_UNICODE_CHARSET;
9597 has_charset_modifier = UNICODE_PAT_MOD;
9599 case ASCII_RESTRICT_PAT_MOD:
9600 if (flagsp == &negflags) {
9603 if (has_charset_modifier) {
9604 if (cs != REGEX_ASCII_RESTRICTED_CHARSET) {
9605 goto excess_modifier;
9607 /* Doubled modifier implies more restricted */
9608 cs = REGEX_ASCII_MORE_RESTRICTED_CHARSET;
9611 cs = REGEX_ASCII_RESTRICTED_CHARSET;
9613 has_charset_modifier = ASCII_RESTRICT_PAT_MOD;
9615 case DEPENDS_PAT_MOD:
9616 if (has_use_defaults) {
9617 goto fail_modifiers;
9619 else if (flagsp == &negflags) {
9622 else if (has_charset_modifier) {
9623 goto excess_modifier;
9626 /* The dual charset means unicode semantics if the
9627 * pattern (or target, not known until runtime) are
9628 * utf8, or something in the pattern indicates unicode
9630 cs = (RExC_utf8 || RExC_uni_semantics)
9631 ? REGEX_UNICODE_CHARSET
9632 : REGEX_DEPENDS_CHARSET;
9633 has_charset_modifier = DEPENDS_PAT_MOD;
9637 if (has_charset_modifier == ASCII_RESTRICT_PAT_MOD) {
9638 vFAIL2("Regexp modifier \"%c\" may appear a maximum of twice", ASCII_RESTRICT_PAT_MOD);
9640 else if (has_charset_modifier == *(RExC_parse - 1)) {
9641 vFAIL2("Regexp modifier \"%c\" may not appear twice",
9645 vFAIL3("Regexp modifiers \"%c\" and \"%c\" are mutually exclusive", has_charset_modifier, *(RExC_parse - 1));
9647 NOT_REACHED; /*NOTREACHED*/
9650 vFAIL2("Regexp modifier \"%c\" may not appear after the \"-\"",
9652 NOT_REACHED; /*NOTREACHED*/
9653 case ONCE_PAT_MOD: /* 'o' */
9654 case GLOBAL_PAT_MOD: /* 'g' */
9655 if (PASS2 && ckWARN(WARN_REGEXP)) {
9656 const I32 wflagbit = *RExC_parse == 'o'
9659 if (! (wastedflags & wflagbit) ) {
9660 wastedflags |= wflagbit;
9661 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9664 "Useless (%s%c) - %suse /%c modifier",
9665 flagsp == &negflags ? "?-" : "?",
9667 flagsp == &negflags ? "don't " : "",
9674 case CONTINUE_PAT_MOD: /* 'c' */
9675 if (PASS2 && ckWARN(WARN_REGEXP)) {
9676 if (! (wastedflags & WASTED_C) ) {
9677 wastedflags |= WASTED_GC;
9678 /* diag_listed_as: Useless (?-%s) - don't use /%s modifier in regex; marked by <-- HERE in m/%s/ */
9681 "Useless (%sc) - %suse /gc modifier",
9682 flagsp == &negflags ? "?-" : "?",
9683 flagsp == &negflags ? "don't " : ""
9688 case KEEPCOPY_PAT_MOD: /* 'p' */
9689 if (flagsp == &negflags) {
9691 ckWARNreg(RExC_parse + 1,"Useless use of (?-p)");
9693 *flagsp |= RXf_PMf_KEEPCOPY;
9697 /* A flag is a default iff it is following a minus, so
9698 * if there is a minus, it means will be trying to
9699 * re-specify a default which is an error */
9700 if (has_use_defaults || flagsp == &negflags) {
9701 goto fail_modifiers;
9704 wastedflags = 0; /* reset so (?g-c) warns twice */
9708 RExC_flags |= posflags;
9709 RExC_flags &= ~negflags;
9710 set_regex_charset(&RExC_flags, cs);
9711 if (RExC_flags & RXf_PMf_FOLD) {
9712 RExC_contains_i = 1;
9715 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9721 RExC_parse += SKIP_IF_CHAR(RExC_parse);
9722 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9723 vFAIL2utf8f("Sequence (%"UTF8f"...) not recognized",
9724 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
9725 NOT_REACHED; /*NOTREACHED*/
9732 STD_PMMOD_FLAGS_PARSE_X_WARN(x_mod_count);
9737 - reg - regular expression, i.e. main body or parenthesized thing
9739 * Caller must absorb opening parenthesis.
9741 * Combining parenthesis handling with the base level of regular expression
9742 * is a trifle forced, but the need to tie the tails of the branches to what
9743 * follows makes it hard to avoid.
9745 #define REGTAIL(x,y,z) regtail((x),(y),(z),depth+1)
9747 #define REGTAIL_STUDY(x,y,z) regtail_study((x),(y),(z),depth+1)
9749 #define REGTAIL_STUDY(x,y,z) regtail((x),(y),(z),depth+1)
9752 /* Returns NULL, setting *flagp to TRYAGAIN at the end of (?) that only sets
9753 flags. Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan
9754 needs to be restarted.
9755 Otherwise would only return NULL if regbranch() returns NULL, which
9758 S_reg(pTHX_ RExC_state_t *pRExC_state, I32 paren, I32 *flagp,U32 depth)
9759 /* paren: Parenthesized? 0=top; 1,2=inside '(': changed to letter.
9760 * 2 is like 1, but indicates that nextchar() has been called to advance
9761 * RExC_parse beyond the '('. Things like '(?' are indivisible tokens, and
9762 * this flag alerts us to the need to check for that */
9764 regnode *ret; /* Will be the head of the group. */
9767 regnode *ender = NULL;
9770 U32 oregflags = RExC_flags;
9771 bool have_branch = 0;
9773 I32 freeze_paren = 0;
9774 I32 after_freeze = 0;
9775 I32 num; /* numeric backreferences */
9777 char * parse_start = RExC_parse; /* MJD */
9778 char * const oregcomp_parse = RExC_parse;
9780 GET_RE_DEBUG_FLAGS_DECL;
9782 PERL_ARGS_ASSERT_REG;
9783 DEBUG_PARSE("reg ");
9785 *flagp = 0; /* Tentatively. */
9788 /* Make an OPEN node, if parenthesized. */
9791 /* Under /x, space and comments can be gobbled up between the '(' and
9792 * here (if paren ==2). The forms '(*VERB' and '(?...' disallow such
9793 * intervening space, as the sequence is a token, and a token should be
9795 bool has_intervening_patws = paren == 2 && *(RExC_parse - 1) != '(';
9797 if ( *RExC_parse == '*') { /* (*VERB:ARG) */
9798 char *start_verb = RExC_parse;
9799 STRLEN verb_len = 0;
9800 char *start_arg = NULL;
9801 unsigned char op = 0;
9803 int internal_argval = 0; /* internal_argval is only useful if
9806 if (has_intervening_patws) {
9808 vFAIL("In '(*VERB...)', the '(' and '*' must be adjacent");
9810 while ( *RExC_parse && *RExC_parse != ')' ) {
9811 if ( *RExC_parse == ':' ) {
9812 start_arg = RExC_parse + 1;
9818 verb_len = RExC_parse - start_verb;
9821 while ( *RExC_parse && *RExC_parse != ')' )
9823 if ( *RExC_parse != ')' )
9824 vFAIL("Unterminated verb pattern argument");
9825 if ( RExC_parse == start_arg )
9828 if ( *RExC_parse != ')' )
9829 vFAIL("Unterminated verb pattern");
9832 switch ( *start_verb ) {
9833 case 'A': /* (*ACCEPT) */
9834 if ( memEQs(start_verb,verb_len,"ACCEPT") ) {
9836 internal_argval = RExC_nestroot;
9839 case 'C': /* (*COMMIT) */
9840 if ( memEQs(start_verb,verb_len,"COMMIT") )
9843 case 'F': /* (*FAIL) */
9844 if ( verb_len==1 || memEQs(start_verb,verb_len,"FAIL") ) {
9849 case ':': /* (*:NAME) */
9850 case 'M': /* (*MARK:NAME) */
9851 if ( verb_len==0 || memEQs(start_verb,verb_len,"MARK") ) {
9856 case 'P': /* (*PRUNE) */
9857 if ( memEQs(start_verb,verb_len,"PRUNE") )
9860 case 'S': /* (*SKIP) */
9861 if ( memEQs(start_verb,verb_len,"SKIP") )
9864 case 'T': /* (*THEN) */
9865 /* [19:06] <TimToady> :: is then */
9866 if ( memEQs(start_verb,verb_len,"THEN") ) {
9868 RExC_seen |= REG_CUTGROUP_SEEN;
9873 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
9875 "Unknown verb pattern '%"UTF8f"'",
9876 UTF8fARG(UTF, verb_len, start_verb));
9879 if ( start_arg && internal_argval ) {
9880 vFAIL3("Verb pattern '%.*s' may not have an argument",
9881 verb_len, start_verb);
9882 } else if ( argok < 0 && !start_arg ) {
9883 vFAIL3("Verb pattern '%.*s' has a mandatory argument",
9884 verb_len, start_verb);
9886 ret = reganode(pRExC_state, op, internal_argval);
9887 if ( ! internal_argval && ! SIZE_ONLY ) {
9889 SV *sv = newSVpvn( start_arg,
9890 RExC_parse - start_arg);
9891 ARG(ret) = add_data( pRExC_state,
9893 RExC_rxi->data->data[ARG(ret)]=(void*)sv;
9900 if (!internal_argval)
9901 RExC_seen |= REG_VERBARG_SEEN;
9902 } else if ( start_arg ) {
9903 vFAIL3("Verb pattern '%.*s' may not have an argument",
9904 verb_len, start_verb);
9906 ret = reg_node(pRExC_state, op);
9908 nextchar(pRExC_state);
9911 else if (*RExC_parse == '?') { /* (?...) */
9912 bool is_logical = 0;
9913 const char * const seqstart = RExC_parse;
9914 const char * endptr;
9915 if (has_intervening_patws) {
9917 vFAIL("In '(?...)', the '(' and '?' must be adjacent");
9921 paren = *RExC_parse++;
9922 ret = NULL; /* For look-ahead/behind. */
9925 case 'P': /* (?P...) variants for those used to PCRE/Python */
9926 paren = *RExC_parse++;
9927 if ( paren == '<') /* (?P<...>) named capture */
9929 else if (paren == '>') { /* (?P>name) named recursion */
9930 goto named_recursion;
9932 else if (paren == '=') { /* (?P=...) named backref */
9933 /* this pretty much dupes the code for \k<NAME> in
9934 * regatom(), if you change this make sure you change that
9936 char* name_start = RExC_parse;
9938 SV *sv_dat = reg_scan_name(pRExC_state,
9939 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
9940 if (RExC_parse == name_start || *RExC_parse != ')')
9941 /* diag_listed_as: Sequence ?P=... not terminated in regex; marked by <-- HERE in m/%s/ */
9942 vFAIL2("Sequence %.3s... not terminated",parse_start);
9945 num = add_data( pRExC_state, STR_WITH_LEN("S"));
9946 RExC_rxi->data->data[num]=(void*)sv_dat;
9947 SvREFCNT_inc_simple_void(sv_dat);
9950 ret = reganode(pRExC_state,
9953 : (ASCII_FOLD_RESTRICTED)
9955 : (AT_LEAST_UNI_SEMANTICS)
9963 Set_Node_Offset(ret, parse_start+1);
9964 Set_Node_Cur_Length(ret, parse_start);
9966 nextchar(pRExC_state);
9970 RExC_parse += SKIP_IF_CHAR(RExC_parse);
9971 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
9972 vFAIL3("Sequence (%.*s...) not recognized",
9973 RExC_parse-seqstart, seqstart);
9974 NOT_REACHED; /*NOTREACHED*/
9975 case '<': /* (?<...) */
9976 if (*RExC_parse == '!')
9978 else if (*RExC_parse != '=')
9984 case '\'': /* (?'...') */
9985 name_start= RExC_parse;
9986 svname = reg_scan_name(pRExC_state,
9987 SIZE_ONLY /* reverse test from the others */
9988 ? REG_RSN_RETURN_NAME
9989 : REG_RSN_RETURN_NULL);
9990 if (RExC_parse == name_start || *RExC_parse != paren)
9991 vFAIL2("Sequence (?%c... not terminated",
9992 paren=='>' ? '<' : paren);
9996 if (!svname) /* shouldn't happen */
9998 "panic: reg_scan_name returned NULL");
9999 if (!RExC_paren_names) {
10000 RExC_paren_names= newHV();
10001 sv_2mortal(MUTABLE_SV(RExC_paren_names));
10003 RExC_paren_name_list= newAV();
10004 sv_2mortal(MUTABLE_SV(RExC_paren_name_list));
10007 he_str = hv_fetch_ent( RExC_paren_names, svname, 1, 0 );
10009 sv_dat = HeVAL(he_str);
10011 /* croak baby croak */
10013 "panic: paren_name hash element allocation failed");
10014 } else if ( SvPOK(sv_dat) ) {
10015 /* (?|...) can mean we have dupes so scan to check
10016 its already been stored. Maybe a flag indicating
10017 we are inside such a construct would be useful,
10018 but the arrays are likely to be quite small, so
10019 for now we punt -- dmq */
10020 IV count = SvIV(sv_dat);
10021 I32 *pv = (I32*)SvPVX(sv_dat);
10023 for ( i = 0 ; i < count ; i++ ) {
10024 if ( pv[i] == RExC_npar ) {
10030 pv = (I32*)SvGROW(sv_dat,
10031 SvCUR(sv_dat) + sizeof(I32)+1);
10032 SvCUR_set(sv_dat, SvCUR(sv_dat) + sizeof(I32));
10033 pv[count] = RExC_npar;
10034 SvIV_set(sv_dat, SvIVX(sv_dat) + 1);
10037 (void)SvUPGRADE(sv_dat,SVt_PVNV);
10038 sv_setpvn(sv_dat, (char *)&(RExC_npar),
10041 SvIV_set(sv_dat, 1);
10044 /* Yes this does cause a memory leak in debugging Perls
10046 if (!av_store(RExC_paren_name_list,
10047 RExC_npar, SvREFCNT_inc(svname)))
10048 SvREFCNT_dec_NN(svname);
10051 /*sv_dump(sv_dat);*/
10053 nextchar(pRExC_state);
10055 goto capturing_parens;
10057 RExC_seen |= REG_LOOKBEHIND_SEEN;
10058 RExC_in_lookbehind++;
10061 case '=': /* (?=...) */
10062 RExC_seen_zerolen++;
10064 case '!': /* (?!...) */
10065 RExC_seen_zerolen++;
10066 /* check if we're really just a "FAIL" assertion */
10068 nextchar(pRExC_state);
10069 if (*RExC_parse == ')') {
10070 ret=reg_node(pRExC_state, OPFAIL);
10071 nextchar(pRExC_state);
10075 case '|': /* (?|...) */
10076 /* branch reset, behave like a (?:...) except that
10077 buffers in alternations share the same numbers */
10079 after_freeze = freeze_paren = RExC_npar;
10081 case ':': /* (?:...) */
10082 case '>': /* (?>...) */
10084 case '$': /* (?$...) */
10085 case '@': /* (?@...) */
10086 vFAIL2("Sequence (?%c...) not implemented", (int)paren);
10088 case '0' : /* (?0) */
10089 case 'R' : /* (?R) */
10090 if (*RExC_parse != ')')
10091 FAIL("Sequence (?R) not terminated");
10092 ret = reg_node(pRExC_state, GOSTART);
10093 RExC_seen |= REG_GOSTART_SEEN;
10094 *flagp |= POSTPONED;
10095 nextchar(pRExC_state);
10098 /* named and numeric backreferences */
10099 case '&': /* (?&NAME) */
10100 parse_start = RExC_parse - 1;
10103 SV *sv_dat = reg_scan_name(pRExC_state,
10104 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10105 num = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10107 if (RExC_parse == RExC_end || *RExC_parse != ')')
10108 vFAIL("Sequence (?&... not terminated");
10109 goto gen_recurse_regop;
10112 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10114 vFAIL("Illegal pattern");
10116 goto parse_recursion;
10118 case '-': /* (?-1) */
10119 if (!(RExC_parse[0] >= '1' && RExC_parse[0] <= '9')) {
10120 RExC_parse--; /* rewind to let it be handled later */
10124 case '1': case '2': case '3': case '4': /* (?1) */
10125 case '5': case '6': case '7': case '8': case '9':
10129 bool is_neg = FALSE;
10131 parse_start = RExC_parse - 1; /* MJD */
10132 if (*RExC_parse == '-') {
10136 if (grok_atoUV(RExC_parse, &unum, &endptr)
10140 RExC_parse = (char*)endptr;
10144 /* Some limit for num? */
10148 if (*RExC_parse!=')')
10149 vFAIL("Expecting close bracket");
10152 if ( paren == '-' ) {
10154 Diagram of capture buffer numbering.
10155 Top line is the normal capture buffer numbers
10156 Bottom line is the negative indexing as from
10160 /(a(x)y)(a(b(c(?-2)d)e)f)(g(h))/
10164 num = RExC_npar + num;
10167 vFAIL("Reference to nonexistent group");
10169 } else if ( paren == '+' ) {
10170 num = RExC_npar + num - 1;
10173 ret = reg2Lanode(pRExC_state, GOSUB, num, RExC_recurse_count);
10175 if (num > (I32)RExC_rx->nparens) {
10177 vFAIL("Reference to nonexistent group");
10179 RExC_recurse_count++;
10180 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10181 "%*s%*s Recurse #%"UVuf" to %"IVdf"\n",
10182 22, "| |", (int)(depth * 2 + 1), "",
10183 (UV)ARG(ret), (IV)ARG2L(ret)));
10185 RExC_seen |= REG_RECURSE_SEEN;
10186 Set_Node_Length(ret, 1 + regarglen[OP(ret)]); /* MJD */
10187 Set_Node_Offset(ret, parse_start); /* MJD */
10189 *flagp |= POSTPONED;
10190 nextchar(pRExC_state);
10195 case '?': /* (??...) */
10197 if (*RExC_parse != '{') {
10198 RExC_parse += SKIP_IF_CHAR(RExC_parse);
10199 /* diag_listed_as: Sequence (?%s...) not recognized in regex; marked by <-- HERE in m/%s/ */
10201 "Sequence (%"UTF8f"...) not recognized",
10202 UTF8fARG(UTF, RExC_parse-seqstart, seqstart));
10203 NOT_REACHED; /*NOTREACHED*/
10205 *flagp |= POSTPONED;
10206 paren = *RExC_parse++;
10208 case '{': /* (?{...}) */
10211 struct reg_code_block *cb;
10213 RExC_seen_zerolen++;
10215 if ( !pRExC_state->num_code_blocks
10216 || pRExC_state->code_index >= pRExC_state->num_code_blocks
10217 || pRExC_state->code_blocks[pRExC_state->code_index].start
10218 != (STRLEN)((RExC_parse -3 - (is_logical ? 1 : 0))
10221 if (RExC_pm_flags & PMf_USE_RE_EVAL)
10222 FAIL("panic: Sequence (?{...}): no code block found\n");
10223 FAIL("Eval-group not allowed at runtime, use re 'eval'");
10225 /* this is a pre-compiled code block (?{...}) */
10226 cb = &pRExC_state->code_blocks[pRExC_state->code_index];
10227 RExC_parse = RExC_start + cb->end;
10230 if (cb->src_regex) {
10231 n = add_data(pRExC_state, STR_WITH_LEN("rl"));
10232 RExC_rxi->data->data[n] =
10233 (void*)SvREFCNT_inc((SV*)cb->src_regex);
10234 RExC_rxi->data->data[n+1] = (void*)o;
10237 n = add_data(pRExC_state,
10238 (RExC_pm_flags & PMf_HAS_CV) ? "L" : "l", 1);
10239 RExC_rxi->data->data[n] = (void*)o;
10242 pRExC_state->code_index++;
10243 nextchar(pRExC_state);
10247 ret = reg_node(pRExC_state, LOGICAL);
10249 eval = reg2Lanode(pRExC_state, EVAL,
10252 /* for later propagation into (??{})
10254 RExC_flags & RXf_PMf_COMPILETIME
10259 REGTAIL(pRExC_state, ret, eval);
10260 /* deal with the length of this later - MJD */
10263 ret = reg2Lanode(pRExC_state, EVAL, n, 0);
10264 Set_Node_Length(ret, RExC_parse - parse_start + 1);
10265 Set_Node_Offset(ret, parse_start);
10268 case '(': /* (?(?{...})...) and (?(?=...)...) */
10271 const int DEFINE_len = sizeof("DEFINE") - 1;
10272 if (RExC_parse[0] == '?') { /* (?(?...)) */
10273 if (RExC_parse[1] == '=' || RExC_parse[1] == '!'
10274 || RExC_parse[1] == '<'
10275 || RExC_parse[1] == '{') { /* Lookahead or eval. */
10279 ret = reg_node(pRExC_state, LOGICAL);
10283 tail = reg(pRExC_state, 1, &flag, depth+1);
10284 if (flag & RESTART_UTF8) {
10285 *flagp = RESTART_UTF8;
10288 REGTAIL(pRExC_state, ret, tail);
10291 /* Fall through to ‘Unknown switch condition’ at the
10292 end of the if/else chain. */
10294 else if ( RExC_parse[0] == '<' /* (?(<NAME>)...) */
10295 || RExC_parse[0] == '\'' ) /* (?('NAME')...) */
10297 char ch = RExC_parse[0] == '<' ? '>' : '\'';
10298 char *name_start= RExC_parse++;
10300 SV *sv_dat=reg_scan_name(pRExC_state,
10301 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
10302 if (RExC_parse == name_start || *RExC_parse != ch)
10303 vFAIL2("Sequence (?(%c... not terminated",
10304 (ch == '>' ? '<' : ch));
10307 num = add_data( pRExC_state, STR_WITH_LEN("S"));
10308 RExC_rxi->data->data[num]=(void*)sv_dat;
10309 SvREFCNT_inc_simple_void(sv_dat);
10311 ret = reganode(pRExC_state,NGROUPP,num);
10312 goto insert_if_check_paren;
10314 else if (RExC_end - RExC_parse >= DEFINE_len
10315 && strnEQ(RExC_parse, "DEFINE", DEFINE_len))
10317 ret = reganode(pRExC_state,DEFINEP,0);
10318 RExC_parse += DEFINE_len;
10320 goto insert_if_check_paren;
10322 else if (RExC_parse[0] == 'R') {
10325 if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10327 if (grok_atoUV(RExC_parse, &uv, &endptr)
10331 RExC_parse = (char*)endptr;
10333 /* else "Switch condition not recognized" below */
10334 } else if (RExC_parse[0] == '&') {
10337 sv_dat = reg_scan_name(pRExC_state,
10339 ? REG_RSN_RETURN_NULL
10340 : REG_RSN_RETURN_DATA);
10341 parno = sv_dat ? *((I32 *)SvPVX(sv_dat)) : 0;
10343 ret = reganode(pRExC_state,INSUBP,parno);
10344 goto insert_if_check_paren;
10346 else if (RExC_parse[0] >= '1' && RExC_parse[0] <= '9' ) {
10351 if (grok_atoUV(RExC_parse, &uv, &endptr)
10355 RExC_parse = (char*)endptr;
10357 /* XXX else what? */
10358 ret = reganode(pRExC_state, GROUPP, parno);
10360 insert_if_check_paren:
10361 if (*(tmp = nextchar(pRExC_state)) != ')') {
10362 /* nextchar also skips comments, so undo its work
10363 * and skip over the the next character.
10366 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10367 vFAIL("Switch condition not recognized");
10370 REGTAIL(pRExC_state, ret, reganode(pRExC_state, IFTHEN, 0));
10371 br = regbranch(pRExC_state, &flags, 1,depth+1);
10373 if (flags & RESTART_UTF8) {
10374 *flagp = RESTART_UTF8;
10377 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10380 REGTAIL(pRExC_state, br, reganode(pRExC_state,
10382 c = *nextchar(pRExC_state);
10383 if (flags&HASWIDTH)
10384 *flagp |= HASWIDTH;
10387 vFAIL("(?(DEFINE)....) does not allow branches");
10389 /* Fake one for optimizer. */
10390 lastbr = reganode(pRExC_state, IFTHEN, 0);
10392 if (!regbranch(pRExC_state, &flags, 1,depth+1)) {
10393 if (flags & RESTART_UTF8) {
10394 *flagp = RESTART_UTF8;
10397 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"",
10400 REGTAIL(pRExC_state, ret, lastbr);
10401 if (flags&HASWIDTH)
10402 *flagp |= HASWIDTH;
10403 c = *nextchar(pRExC_state);
10408 if (RExC_parse>RExC_end)
10409 vFAIL("Switch (?(condition)... not terminated");
10411 vFAIL("Switch (?(condition)... contains too many branches");
10413 ender = reg_node(pRExC_state, TAIL);
10414 REGTAIL(pRExC_state, br, ender);
10416 REGTAIL(pRExC_state, lastbr, ender);
10417 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10420 REGTAIL(pRExC_state, ret, ender);
10421 RExC_size++; /* XXX WHY do we need this?!!
10422 For large programs it seems to be required
10423 but I can't figure out why. -- dmq*/
10426 RExC_parse += UTF ? UTF8SKIP(RExC_parse) : 1;
10427 vFAIL("Unknown switch condition (?(...))");
10429 case '[': /* (?[ ... ]) */
10430 return handle_regex_sets(pRExC_state, NULL, flagp, depth,
10433 RExC_parse--; /* for vFAIL to print correctly */
10434 vFAIL("Sequence (? incomplete");
10436 default: /* e.g., (?i) */
10439 parse_lparen_question_flags(pRExC_state);
10440 if (UCHARAT(RExC_parse) != ':') {
10442 nextchar(pRExC_state);
10447 nextchar(pRExC_state);
10452 else if (!(RExC_flags & RXf_PMf_NOCAPTURE)) { /* (...) */
10457 ret = reganode(pRExC_state, OPEN, parno);
10459 if (!RExC_nestroot)
10460 RExC_nestroot = parno;
10461 if (RExC_seen & REG_RECURSE_SEEN
10462 && !RExC_open_parens[parno-1])
10464 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10465 "%*s%*s Setting open paren #%"IVdf" to %d\n",
10466 22, "| |", (int)(depth * 2 + 1), "",
10467 (IV)parno, REG_NODE_NUM(ret)));
10468 RExC_open_parens[parno-1]= ret;
10471 Set_Node_Length(ret, 1); /* MJD */
10472 Set_Node_Offset(ret, RExC_parse); /* MJD */
10475 /* with RXf_PMf_NOCAPTURE treat (...) as (?:...) */
10484 /* Pick up the branches, linking them together. */
10485 parse_start = RExC_parse; /* MJD */
10486 br = regbranch(pRExC_state, &flags, 1,depth+1);
10488 /* branch_len = (paren != 0); */
10491 if (flags & RESTART_UTF8) {
10492 *flagp = RESTART_UTF8;
10495 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10497 if (*RExC_parse == '|') {
10498 if (!SIZE_ONLY && RExC_extralen) {
10499 reginsert(pRExC_state, BRANCHJ, br, depth+1);
10502 reginsert(pRExC_state, BRANCH, br, depth+1);
10503 Set_Node_Length(br, paren != 0);
10504 Set_Node_Offset_To_R(br-RExC_emit_start, parse_start-RExC_start);
10508 RExC_extralen += 1; /* For BRANCHJ-BRANCH. */
10510 else if (paren == ':') {
10511 *flagp |= flags&SIMPLE;
10513 if (is_open) { /* Starts with OPEN. */
10514 REGTAIL(pRExC_state, ret, br); /* OPEN -> first. */
10516 else if (paren != '?') /* Not Conditional */
10518 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10520 while (*RExC_parse == '|') {
10521 if (!SIZE_ONLY && RExC_extralen) {
10522 ender = reganode(pRExC_state, LONGJMP,0);
10524 /* Append to the previous. */
10525 REGTAIL(pRExC_state, NEXTOPER(NEXTOPER(lastbr)), ender);
10528 RExC_extralen += 2; /* Account for LONGJMP. */
10529 nextchar(pRExC_state);
10530 if (freeze_paren) {
10531 if (RExC_npar > after_freeze)
10532 after_freeze = RExC_npar;
10533 RExC_npar = freeze_paren;
10535 br = regbranch(pRExC_state, &flags, 0, depth+1);
10538 if (flags & RESTART_UTF8) {
10539 *flagp = RESTART_UTF8;
10542 FAIL2("panic: regbranch returned NULL, flags=%#"UVxf"", (UV) flags);
10544 REGTAIL(pRExC_state, lastbr, br); /* BRANCH -> BRANCH. */
10546 *flagp |= flags & (SPSTART | HASWIDTH | POSTPONED);
10549 if (have_branch || paren != ':') {
10550 /* Make a closing node, and hook it on the end. */
10553 ender = reg_node(pRExC_state, TAIL);
10556 ender = reganode(pRExC_state, CLOSE, parno);
10557 if (!SIZE_ONLY && RExC_seen & REG_RECURSE_SEEN) {
10558 DEBUG_OPTIMISE_MORE_r(PerlIO_printf(Perl_debug_log,
10559 "%*s%*s Setting close paren #%"IVdf" to %d\n",
10560 22, "| |", (int)(depth * 2 + 1), "", (IV)parno, REG_NODE_NUM(ender)));
10561 RExC_close_parens[parno-1]= ender;
10562 if (RExC_nestroot == parno)
10565 Set_Node_Offset(ender,RExC_parse+1); /* MJD */
10566 Set_Node_Length(ender,1); /* MJD */
10572 *flagp &= ~HASWIDTH;
10575 ender = reg_node(pRExC_state, SUCCEED);
10578 ender = reg_node(pRExC_state, END);
10580 assert(!RExC_opend); /* there can only be one! */
10581 RExC_opend = ender;
10585 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10586 DEBUG_PARSE_MSG("lsbr");
10587 regprop(RExC_rx, RExC_mysv1, lastbr, NULL, pRExC_state);
10588 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10589 PerlIO_printf(Perl_debug_log, "~ tying lastbr %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10590 SvPV_nolen_const(RExC_mysv1),
10591 (IV)REG_NODE_NUM(lastbr),
10592 SvPV_nolen_const(RExC_mysv2),
10593 (IV)REG_NODE_NUM(ender),
10594 (IV)(ender - lastbr)
10597 REGTAIL(pRExC_state, lastbr, ender);
10599 if (have_branch && !SIZE_ONLY) {
10600 char is_nothing= 1;
10602 RExC_seen |= REG_TOP_LEVEL_BRANCHES_SEEN;
10604 /* Hook the tails of the branches to the closing node. */
10605 for (br = ret; br; br = regnext(br)) {
10606 const U8 op = PL_regkind[OP(br)];
10607 if (op == BRANCH) {
10608 REGTAIL_STUDY(pRExC_state, NEXTOPER(br), ender);
10609 if ( OP(NEXTOPER(br)) != NOTHING
10610 || regnext(NEXTOPER(br)) != ender)
10613 else if (op == BRANCHJ) {
10614 REGTAIL_STUDY(pRExC_state, NEXTOPER(NEXTOPER(br)), ender);
10615 /* for now we always disable this optimisation * /
10616 if ( OP(NEXTOPER(NEXTOPER(br))) != NOTHING
10617 || regnext(NEXTOPER(NEXTOPER(br))) != ender)
10623 br= PL_regkind[OP(ret)] != BRANCH ? regnext(ret) : ret;
10624 DEBUG_PARSE_r(if (!SIZE_ONLY) {
10625 DEBUG_PARSE_MSG("NADA");
10626 regprop(RExC_rx, RExC_mysv1, ret, NULL, pRExC_state);
10627 regprop(RExC_rx, RExC_mysv2, ender, NULL, pRExC_state);
10628 PerlIO_printf(Perl_debug_log, "~ converting ret %s (%"IVdf") to ender %s (%"IVdf") offset %"IVdf"\n",
10629 SvPV_nolen_const(RExC_mysv1),
10630 (IV)REG_NODE_NUM(ret),
10631 SvPV_nolen_const(RExC_mysv2),
10632 (IV)REG_NODE_NUM(ender),
10637 if (OP(ender) == TAIL) {
10642 for ( opt= br + 1; opt < ender ; opt++ )
10643 OP(opt)= OPTIMIZED;
10644 NEXT_OFF(br)= ender - br;
10652 static const char parens[] = "=!<,>";
10654 if (paren && (p = strchr(parens, paren))) {
10655 U8 node = ((p - parens) % 2) ? UNLESSM : IFMATCH;
10656 int flag = (p - parens) > 1;
10659 node = SUSPEND, flag = 0;
10660 reginsert(pRExC_state, node,ret, depth+1);
10661 Set_Node_Cur_Length(ret, parse_start);
10662 Set_Node_Offset(ret, parse_start + 1);
10664 REGTAIL_STUDY(pRExC_state, ret, reg_node(pRExC_state, TAIL));
10668 /* Check for proper termination. */
10670 /* restore original flags, but keep (?p) */
10671 RExC_flags = oregflags | (RExC_flags & RXf_PMf_KEEPCOPY);
10672 if (RExC_parse >= RExC_end || *nextchar(pRExC_state) != ')') {
10673 RExC_parse = oregcomp_parse;
10674 vFAIL("Unmatched (");
10677 else if (!paren && RExC_parse < RExC_end) {
10678 if (*RExC_parse == ')') {
10680 vFAIL("Unmatched )");
10683 FAIL("Junk on end of regexp"); /* "Can't happen". */
10684 NOT_REACHED; /* NOTREACHED */
10687 if (RExC_in_lookbehind) {
10688 RExC_in_lookbehind--;
10690 if (after_freeze > RExC_npar)
10691 RExC_npar = after_freeze;
10696 - regbranch - one alternative of an | operator
10698 * Implements the concatenation operator.
10700 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10704 S_regbranch(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, I32 first, U32 depth)
10707 regnode *chain = NULL;
10709 I32 flags = 0, c = 0;
10710 GET_RE_DEBUG_FLAGS_DECL;
10712 PERL_ARGS_ASSERT_REGBRANCH;
10714 DEBUG_PARSE("brnc");
10719 if (!SIZE_ONLY && RExC_extralen)
10720 ret = reganode(pRExC_state, BRANCHJ,0);
10722 ret = reg_node(pRExC_state, BRANCH);
10723 Set_Node_Length(ret, 1);
10727 if (!first && SIZE_ONLY)
10728 RExC_extralen += 1; /* BRANCHJ */
10730 *flagp = WORST; /* Tentatively. */
10733 nextchar(pRExC_state);
10734 while (RExC_parse < RExC_end && *RExC_parse != '|' && *RExC_parse != ')') {
10735 flags &= ~TRYAGAIN;
10736 latest = regpiece(pRExC_state, &flags,depth+1);
10737 if (latest == NULL) {
10738 if (flags & TRYAGAIN)
10740 if (flags & RESTART_UTF8) {
10741 *flagp = RESTART_UTF8;
10744 FAIL2("panic: regpiece returned NULL, flags=%#"UVxf"", (UV) flags);
10746 else if (ret == NULL)
10748 *flagp |= flags&(HASWIDTH|POSTPONED);
10749 if (chain == NULL) /* First piece. */
10750 *flagp |= flags&SPSTART;
10752 /* FIXME adding one for every branch after the first is probably
10753 * excessive now we have TRIE support. (hv) */
10755 REGTAIL(pRExC_state, chain, latest);
10760 if (chain == NULL) { /* Loop ran zero times. */
10761 chain = reg_node(pRExC_state, NOTHING);
10766 *flagp |= flags&SIMPLE;
10773 - regpiece - something followed by possible [*+?]
10775 * Note that the branching code sequences used for ? and the general cases
10776 * of * and + are somewhat optimized: they use the same NOTHING node as
10777 * both the endmarker for their branch list and the body of the last branch.
10778 * It might seem that this node could be dispensed with entirely, but the
10779 * endmarker role is not redundant.
10781 * Returns NULL, setting *flagp to TRYAGAIN if regatom() returns NULL with
10783 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
10787 S_regpiece(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
10793 const char * const origparse = RExC_parse;
10795 I32 max = REG_INFTY;
10796 #ifdef RE_TRACK_PATTERN_OFFSETS
10799 const char *maxpos = NULL;
10802 /* Save the original in case we change the emitted regop to a FAIL. */
10803 regnode * const orig_emit = RExC_emit;
10805 GET_RE_DEBUG_FLAGS_DECL;
10807 PERL_ARGS_ASSERT_REGPIECE;
10809 DEBUG_PARSE("piec");
10811 ret = regatom(pRExC_state, &flags,depth+1);
10813 if (flags & (TRYAGAIN|RESTART_UTF8))
10814 *flagp |= flags & (TRYAGAIN|RESTART_UTF8);
10816 FAIL2("panic: regatom returned NULL, flags=%#"UVxf"", (UV) flags);
10822 if (op == '{' && regcurly(RExC_parse)) {
10824 #ifdef RE_TRACK_PATTERN_OFFSETS
10825 parse_start = RExC_parse; /* MJD */
10827 next = RExC_parse + 1;
10828 while (isDIGIT(*next) || *next == ',') {
10829 if (*next == ',') {
10837 if (*next == '}') { /* got one */
10838 const char* endptr;
10842 if (isDIGIT(*RExC_parse)) {
10843 if (!grok_atoUV(RExC_parse, &uv, &endptr))
10844 vFAIL("Invalid quantifier in {,}");
10845 if (uv >= REG_INFTY)
10846 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10851 if (*maxpos == ',')
10854 maxpos = RExC_parse;
10855 if (isDIGIT(*maxpos)) {
10856 if (!grok_atoUV(maxpos, &uv, &endptr))
10857 vFAIL("Invalid quantifier in {,}");
10858 if (uv >= REG_INFTY)
10859 vFAIL2("Quantifier in {,} bigger than %d", REG_INFTY - 1);
10862 max = REG_INFTY; /* meaning "infinity" */
10865 nextchar(pRExC_state);
10866 if (max < min) { /* If can't match, warn and optimize to fail
10870 /* We can't back off the size because we have to reserve
10871 * enough space for all the things we are about to throw
10872 * away, but we can shrink it by the ammount we are about
10873 * to re-use here */
10874 RExC_size = PREVOPER(RExC_size) - regarglen[(U8)OPFAIL];
10877 ckWARNreg(RExC_parse, "Quantifier {n,m} with n > m can't match");
10878 RExC_emit = orig_emit;
10880 ret = reg_node(pRExC_state, OPFAIL);
10883 else if (min == max
10884 && RExC_parse < RExC_end
10885 && (*RExC_parse == '?' || *RExC_parse == '+'))
10888 ckWARN2reg(RExC_parse + 1,
10889 "Useless use of greediness modifier '%c'",
10892 /* Absorb the modifier, so later code doesn't see nor use
10894 nextchar(pRExC_state);
10898 if ((flags&SIMPLE)) {
10899 MARK_NAUGHTY_EXP(2, 2);
10900 reginsert(pRExC_state, CURLY, ret, depth+1);
10901 Set_Node_Offset(ret, parse_start+1); /* MJD */
10902 Set_Node_Cur_Length(ret, parse_start);
10905 regnode * const w = reg_node(pRExC_state, WHILEM);
10908 REGTAIL(pRExC_state, ret, w);
10909 if (!SIZE_ONLY && RExC_extralen) {
10910 reginsert(pRExC_state, LONGJMP,ret, depth+1);
10911 reginsert(pRExC_state, NOTHING,ret, depth+1);
10912 NEXT_OFF(ret) = 3; /* Go over LONGJMP. */
10914 reginsert(pRExC_state, CURLYX,ret, depth+1);
10916 Set_Node_Offset(ret, parse_start+1);
10917 Set_Node_Length(ret,
10918 op == '{' ? (RExC_parse - parse_start) : 1);
10920 if (!SIZE_ONLY && RExC_extralen)
10921 NEXT_OFF(ret) = 3; /* Go over NOTHING to LONGJMP. */
10922 REGTAIL(pRExC_state, ret, reg_node(pRExC_state, NOTHING));
10924 RExC_whilem_seen++, RExC_extralen += 3;
10925 MARK_NAUGHTY_EXP(1, 4); /* compound interest */
10932 *flagp |= HASWIDTH;
10934 ARG1_SET(ret, (U16)min);
10935 ARG2_SET(ret, (U16)max);
10937 if (max == REG_INFTY)
10938 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10944 if (!ISMULT1(op)) {
10949 #if 0 /* Now runtime fix should be reliable. */
10951 /* if this is reinstated, don't forget to put this back into perldiag:
10953 =item Regexp *+ operand could be empty at {#} in regex m/%s/
10955 (F) The part of the regexp subject to either the * or + quantifier
10956 could match an empty string. The {#} shows in the regular
10957 expression about where the problem was discovered.
10961 if (!(flags&HASWIDTH) && op != '?')
10962 vFAIL("Regexp *+ operand could be empty");
10965 #ifdef RE_TRACK_PATTERN_OFFSETS
10966 parse_start = RExC_parse;
10968 nextchar(pRExC_state);
10970 *flagp = (op != '+') ? (WORST|SPSTART|HASWIDTH) : (WORST|HASWIDTH);
10972 if (op == '*' && (flags&SIMPLE)) {
10973 reginsert(pRExC_state, STAR, ret, depth+1);
10976 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10978 else if (op == '*') {
10982 else if (op == '+' && (flags&SIMPLE)) {
10983 reginsert(pRExC_state, PLUS, ret, depth+1);
10986 RExC_seen |= REG_UNBOUNDED_QUANTIFIER_SEEN;
10988 else if (op == '+') {
10992 else if (op == '?') {
10997 if (!SIZE_ONLY && !(flags&(HASWIDTH|POSTPONED)) && max > REG_INFTY/3) {
10998 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
10999 ckWARN2reg(RExC_parse,
11000 "%"UTF8f" matches null string many times",
11001 UTF8fARG(UTF, (RExC_parse >= origparse
11002 ? RExC_parse - origparse
11005 (void)ReREFCNT_inc(RExC_rx_sv);
11008 if (RExC_parse < RExC_end && *RExC_parse == '?') {
11009 nextchar(pRExC_state);
11010 reginsert(pRExC_state, MINMOD, ret, depth+1);
11011 REGTAIL(pRExC_state, ret, ret + NODE_STEP_REGNODE);
11014 if (RExC_parse < RExC_end && *RExC_parse == '+') {
11016 nextchar(pRExC_state);
11017 ender = reg_node(pRExC_state, SUCCEED);
11018 REGTAIL(pRExC_state, ret, ender);
11019 reginsert(pRExC_state, SUSPEND, ret, depth+1);
11021 ender = reg_node(pRExC_state, TAIL);
11022 REGTAIL(pRExC_state, ret, ender);
11025 if (RExC_parse < RExC_end && ISMULT2(RExC_parse)) {
11027 vFAIL("Nested quantifiers");
11034 S_grok_bslash_N(pTHX_ RExC_state_t *pRExC_state,
11042 /* This routine teases apart the various meanings of \N and returns
11043 * accordingly. The input parameters constrain which meaning(s) is/are valid
11044 * in the current context.
11046 * Exactly one of <node_p> and <code_point_p> must be non-NULL.
11048 * If <code_point_p> is not NULL, the context is expecting the result to be a
11049 * single code point. If this \N instance turns out to a single code point,
11050 * the function returns TRUE and sets *code_point_p to that code point.
11052 * If <node_p> is not NULL, the context is expecting the result to be one of
11053 * the things representable by a regnode. If this \N instance turns out to be
11054 * one such, the function generates the regnode, returns TRUE and sets *node_p
11055 * to point to that regnode.
11057 * If this instance of \N isn't legal in any context, this function will
11058 * generate a fatal error and not return.
11060 * On input, RExC_parse should point to the first char following the \N at the
11061 * time of the call. On successful return, RExC_parse will have been updated
11062 * to point to just after the sequence identified by this routine. Also
11063 * *flagp has been updated as needed.
11065 * When there is some problem with the current context and this \N instance,
11066 * the function returns FALSE, without advancing RExC_parse, nor setting
11067 * *node_p, nor *code_point_p, nor *flagp.
11069 * If <cp_count> is not NULL, the caller wants to know the length (in code
11070 * points) that this \N sequence matches. This is set even if the function
11071 * returns FALSE, as detailed below.
11073 * There are 5 possibilities here, as detailed in the next 5 paragraphs.
11075 * Probably the most common case is for the \N to specify a single code point.
11076 * *cp_count will be set to 1, and *code_point_p will be set to that code
11079 * Another possibility is for the input to be an empty \N{}, which for
11080 * backwards compatibility we accept. *cp_count will be set to 0. *node_p
11081 * will be set to a generated NOTHING node.
11083 * Still another possibility is for the \N to mean [^\n]. *cp_count will be
11084 * set to 0. *node_p will be set to a generated REG_ANY node.
11086 * The fourth possibility is that \N resolves to a sequence of more than one
11087 * code points. *cp_count will be set to the number of code points in the
11088 * sequence. *node_p * will be set to a generated node returned by this
11089 * function calling S_reg().
11091 * The final possibility, which happens only when the fourth one would
11092 * otherwise be in effect, is that one of those code points requires the
11093 * pattern to be recompiled as UTF-8. The function returns FALSE, and sets
11094 * the RESTART_UTF8 flag in *flagp. When this happens, the caller needs to
11095 * desist from continuing parsing, and return this information to its caller.
11096 * This is not set for when there is only one code point, as this can be
11097 * called as part of an ANYOF node, and they can store above-Latin1 code
11098 * points without the pattern having to be in UTF-8.
11100 * For non-single-quoted regexes, the tokenizer has resolved character and
11101 * sequence names inside \N{...} into their Unicode values, normalizing the
11102 * result into what we should see here: '\N{U+c1.c2...}', where c1... are the
11103 * hex-represented code points in the sequence. This is done there because
11104 * the names can vary based on what charnames pragma is in scope at the time,
11105 * so we need a way to take a snapshot of what they resolve to at the time of
11106 * the original parse. [perl #56444].
11108 * That parsing is skipped for single-quoted regexes, so we may here get
11109 * '\N{NAME}'. This is a fatal error. These names have to be resolved by the
11110 * parser. But if the single-quoted regex is something like '\N{U+41}', that
11111 * is legal and handled here. The code point is Unicode, and has to be
11112 * translated into the native character set for non-ASCII platforms.
11113 * the tokenizer passes the \N sequence through unchanged; this code will not
11114 * attempt to determine this nor expand those, instead raising a syntax error.
11117 char * endbrace; /* points to '}' following the name */
11118 char *endchar; /* Points to '.' or '}' ending cur char in the input
11120 char* p; /* Temporary */
11122 GET_RE_DEBUG_FLAGS_DECL;
11124 PERL_ARGS_ASSERT_GROK_BSLASH_N;
11126 GET_RE_DEBUG_FLAGS;
11128 assert(cBOOL(node_p) ^ cBOOL(code_point_p)); /* Exactly one should be set */
11129 assert(! (node_p && cp_count)); /* At most 1 should be set */
11131 if (cp_count) { /* Initialize return for the most common case */
11135 /* The [^\n] meaning of \N ignores spaces and comments under the /x
11136 * modifier. The other meanings do not, so use a temporary until we find
11137 * out which we are being called with */
11138 p = (RExC_flags & RXf_PMf_EXTENDED)
11139 ? regpatws(pRExC_state, RExC_parse,
11140 TRUE) /* means recognize comments */
11143 /* Disambiguate between \N meaning a named character versus \N meaning
11144 * [^\n]. The latter is assumed when the {...} following the \N is a legal
11145 * quantifier, or there is no a '{' at all */
11146 if (*p != '{' || regcurly(p)) {
11155 RExC_parse--; /* Need to back off so nextchar() doesn't skip the
11157 nextchar(pRExC_state);
11158 *node_p = reg_node(pRExC_state, REG_ANY);
11159 *flagp |= HASWIDTH|SIMPLE;
11161 Set_Node_Length(*node_p, 1); /* MJD */
11165 /* Here, we have decided it should be a named character or sequence */
11167 /* The test above made sure that the next real character is a '{', but
11168 * under the /x modifier, it could be separated by space (or a comment and
11169 * \n) and this is not allowed (for consistency with \x{...} and the
11170 * tokenizer handling of \N{NAME}). */
11171 if (*RExC_parse != '{') {
11172 vFAIL("Missing braces on \\N{}");
11175 RExC_parse++; /* Skip past the '{' */
11177 if (! (endbrace = strchr(RExC_parse, '}')) /* no trailing brace */
11178 || ! (endbrace == RExC_parse /* nothing between the {} */
11179 || (endbrace - RExC_parse >= 2 /* U+ (bad hex is checked... */
11180 && strnEQ(RExC_parse, "U+", 2)))) /* ... below for a better
11183 if (endbrace) RExC_parse = endbrace; /* position msg's '<--HERE' */
11184 vFAIL("\\N{NAME} must be resolved by the lexer");
11187 RExC_uni_semantics = 1; /* Unicode named chars imply Unicode semantics */
11189 if (endbrace == RExC_parse) { /* empty: \N{} */
11193 nextchar(pRExC_state);
11198 *node_p = reg_node(pRExC_state,NOTHING);
11202 RExC_parse += 2; /* Skip past the 'U+' */
11204 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11206 /* Code points are separated by dots. If none, there is only one code
11207 * point, and is terminated by the brace */
11209 if (endchar >= endbrace) {
11210 STRLEN length_of_hex;
11211 I32 grok_hex_flags;
11213 /* Here, exactly one code point. If that isn't what is wanted, fail */
11214 if (! code_point_p) {
11219 /* Convert code point from hex */
11220 length_of_hex = (STRLEN)(endchar - RExC_parse);
11221 grok_hex_flags = PERL_SCAN_ALLOW_UNDERSCORES
11222 | PERL_SCAN_DISALLOW_PREFIX
11224 /* No errors in the first pass (See [perl
11225 * #122671].) We let the code below find the
11226 * errors when there are multiple chars. */
11228 ? PERL_SCAN_SILENT_ILLDIGIT
11231 /* This routine is the one place where both single- and double-quotish
11232 * \N{U+xxxx} are evaluated. The value is a Unicode code point which
11233 * must be converted to native. */
11234 *code_point_p = UNI_TO_NATIVE(grok_hex(RExC_parse,
11239 /* The tokenizer should have guaranteed validity, but it's possible to
11240 * bypass it by using single quoting, so check. Don't do the check
11241 * here when there are multiple chars; we do it below anyway. */
11242 if (length_of_hex == 0
11243 || length_of_hex != (STRLEN)(endchar - RExC_parse) )
11245 RExC_parse += length_of_hex; /* Includes all the valid */
11246 RExC_parse += (RExC_orig_utf8) /* point to after 1st invalid */
11247 ? UTF8SKIP(RExC_parse)
11249 /* Guard against malformed utf8 */
11250 if (RExC_parse >= endchar) {
11251 RExC_parse = endchar;
11253 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11256 RExC_parse = endbrace + 1;
11259 else { /* Is a multiple character sequence */
11260 SV * substitute_parse;
11262 char *orig_end = RExC_end;
11265 /* Count the code points, if desired, in the sequence */
11268 while (RExC_parse < endbrace) {
11269 /* Point to the beginning of the next character in the sequence. */
11270 RExC_parse = endchar + 1;
11271 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11276 /* Fail if caller doesn't want to handle a multi-code-point sequence.
11277 * But don't backup up the pointer if the caller want to know how many
11278 * code points there are (they can then handle things) */
11286 /* What is done here is to convert this to a sub-pattern of the form
11287 * \x{char1}\x{char2}... and then call reg recursively to parse it
11288 * (enclosing in "(?: ... )" ). That way, it retains its atomicness,
11289 * while not having to worry about special handling that some code
11290 * points may have. */
11292 substitute_parse = newSVpvs("?:");
11294 while (RExC_parse < endbrace) {
11296 /* Convert to notation the rest of the code understands */
11297 sv_catpv(substitute_parse, "\\x{");
11298 sv_catpvn(substitute_parse, RExC_parse, endchar - RExC_parse);
11299 sv_catpv(substitute_parse, "}");
11301 /* Point to the beginning of the next character in the sequence. */
11302 RExC_parse = endchar + 1;
11303 endchar = RExC_parse + strcspn(RExC_parse, ".}");
11306 sv_catpv(substitute_parse, ")");
11308 RExC_parse = SvPV(substitute_parse, len);
11310 /* Don't allow empty number */
11311 if (len < (STRLEN) 8) {
11312 RExC_parse = endbrace;
11313 vFAIL("Invalid hexadecimal number in \\N{U+...}");
11315 RExC_end = RExC_parse + len;
11317 /* The values are Unicode, and therefore not subject to recoding, but
11318 * have to be converted to native on a non-Unicode (meaning non-ASCII)
11320 RExC_override_recoding = 1;
11322 RExC_recode_x_to_native = 1;
11326 if (!(*node_p = reg(pRExC_state, 1, &flags, depth+1))) {
11327 if (flags & RESTART_UTF8) {
11328 *flagp = RESTART_UTF8;
11331 FAIL2("panic: reg returned NULL to grok_bslash_N, flags=%#"UVxf"",
11334 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11337 /* Restore the saved values */
11338 RExC_parse = endbrace;
11339 RExC_end = orig_end;
11340 RExC_override_recoding = 0;
11342 RExC_recode_x_to_native = 0;
11345 SvREFCNT_dec_NN(substitute_parse);
11346 nextchar(pRExC_state);
11356 * It returns the code point in utf8 for the value in *encp.
11357 * value: a code value in the source encoding
11358 * encp: a pointer to an Encode object
11360 * If the result from Encode is not a single character,
11361 * it returns U+FFFD (Replacement character) and sets *encp to NULL.
11364 S_reg_recode(pTHX_ const char value, SV **encp)
11367 SV * const sv = newSVpvn_flags(&value, numlen, SVs_TEMP);
11368 const char * const s = *encp ? sv_recode_to_utf8(sv, *encp) : SvPVX(sv);
11369 const STRLEN newlen = SvCUR(sv);
11370 UV uv = UNICODE_REPLACEMENT;
11372 PERL_ARGS_ASSERT_REG_RECODE;
11376 ? utf8n_to_uvchr((U8*)s, newlen, &numlen, UTF8_ALLOW_DEFAULT)
11379 if (!newlen || numlen != newlen) {
11380 uv = UNICODE_REPLACEMENT;
11386 PERL_STATIC_INLINE U8
11387 S_compute_EXACTish(RExC_state_t *pRExC_state)
11391 PERL_ARGS_ASSERT_COMPUTE_EXACTISH;
11399 op = get_regex_charset(RExC_flags);
11400 if (op >= REGEX_ASCII_RESTRICTED_CHARSET) {
11401 op--; /* /a is same as /u, and map /aa's offset to what /a's would have
11402 been, so there is no hole */
11405 return op + EXACTF;
11408 PERL_STATIC_INLINE void
11409 S_alloc_maybe_populate_EXACT(pTHX_ RExC_state_t *pRExC_state,
11410 regnode *node, I32* flagp, STRLEN len, UV code_point,
11413 /* This knows the details about sizing an EXACTish node, setting flags for
11414 * it (by setting <*flagp>, and potentially populating it with a single
11417 * If <len> (the length in bytes) is non-zero, this function assumes that
11418 * the node has already been populated, and just does the sizing. In this
11419 * case <code_point> should be the final code point that has already been
11420 * placed into the node. This value will be ignored except that under some
11421 * circumstances <*flagp> is set based on it.
11423 * If <len> is zero, the function assumes that the node is to contain only
11424 * the single character given by <code_point> and calculates what <len>
11425 * should be. In pass 1, it sizes the node appropriately. In pass 2, it
11426 * additionally will populate the node's STRING with <code_point> or its
11429 * In both cases <*flagp> is appropriately set
11431 * It knows that under FOLD, the Latin Sharp S and UTF characters above
11432 * 255, must be folded (the former only when the rules indicate it can
11435 * When it does the populating, it looks at the flag 'downgradable'. If
11436 * true with a node that folds, it checks if the single code point
11437 * participates in a fold, and if not downgrades the node to an EXACT.
11438 * This helps the optimizer */
11440 bool len_passed_in = cBOOL(len != 0);
11441 U8 character[UTF8_MAXBYTES_CASE+1];
11443 PERL_ARGS_ASSERT_ALLOC_MAYBE_POPULATE_EXACT;
11445 /* Don't bother to check for downgrading in PASS1, as it doesn't make any
11446 * sizing difference, and is extra work that is thrown away */
11447 if (downgradable && ! PASS2) {
11448 downgradable = FALSE;
11451 if (! len_passed_in) {
11453 if (UVCHR_IS_INVARIANT(code_point)) {
11454 if (LOC || ! FOLD) { /* /l defers folding until runtime */
11455 *character = (U8) code_point;
11457 else { /* Here is /i and not /l. (toFOLD() is defined on just
11458 ASCII, which isn't the same thing as INVARIANT on
11459 EBCDIC, but it works there, as the extra invariants
11460 fold to themselves) */
11461 *character = toFOLD((U8) code_point);
11463 /* We can downgrade to an EXACT node if this character
11464 * isn't a folding one. Note that this assumes that
11465 * nothing above Latin1 folds to some other invariant than
11466 * one of these alphabetics; otherwise we would also have
11468 * && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11469 * || ASCII_FOLD_RESTRICTED))
11471 if (downgradable && PL_fold[code_point] == code_point) {
11477 else if (FOLD && (! LOC
11478 || ! is_PROBLEMATIC_LOCALE_FOLD_cp(code_point)))
11479 { /* Folding, and ok to do so now */
11480 UV folded = _to_uni_fold_flags(
11484 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
11485 ? FOLD_FLAGS_NOMIX_ASCII
11488 && folded == code_point /* This quickly rules out many
11489 cases, avoiding the
11490 _invlist_contains_cp() overhead
11492 && ! _invlist_contains_cp(PL_utf8_foldable, code_point))
11499 else if (code_point <= MAX_UTF8_TWO_BYTE) {
11501 /* Not folding this cp, and can output it directly */
11502 *character = UTF8_TWO_BYTE_HI(code_point);
11503 *(character + 1) = UTF8_TWO_BYTE_LO(code_point);
11507 uvchr_to_utf8( character, code_point);
11508 len = UTF8SKIP(character);
11510 } /* Else pattern isn't UTF8. */
11512 *character = (U8) code_point;
11514 } /* Else is folded non-UTF8 */
11515 else if (LIKELY(code_point != LATIN_SMALL_LETTER_SHARP_S)) {
11517 /* We don't fold any non-UTF8 except possibly the Sharp s (see
11518 * comments at join_exact()); */
11519 *character = (U8) code_point;
11522 /* Can turn into an EXACT node if we know the fold at compile time,
11523 * and it folds to itself and doesn't particpate in other folds */
11526 && PL_fold_latin1[code_point] == code_point
11527 && (! HAS_NONLATIN1_FOLD_CLOSURE(code_point)
11528 || (isASCII(code_point) && ASCII_FOLD_RESTRICTED)))
11532 } /* else is Sharp s. May need to fold it */
11533 else if (AT_LEAST_UNI_SEMANTICS && ! ASCII_FOLD_RESTRICTED) {
11535 *(character + 1) = 's';
11539 *character = LATIN_SMALL_LETTER_SHARP_S;
11545 RExC_size += STR_SZ(len);
11548 RExC_emit += STR_SZ(len);
11549 STR_LEN(node) = len;
11550 if (! len_passed_in) {
11551 Copy((char *) character, STRING(node), len, char);
11555 *flagp |= HASWIDTH;
11557 /* A single character node is SIMPLE, except for the special-cased SHARP S
11559 if ((len == 1 || (UTF && len == UNISKIP(code_point)))
11560 && (code_point != LATIN_SMALL_LETTER_SHARP_S
11561 || ! FOLD || ! DEPENDS_SEMANTICS))
11566 /* The OP may not be well defined in PASS1 */
11567 if (PASS2 && OP(node) == EXACTFL) {
11568 RExC_contains_locale = 1;
11573 /* Parse backref decimal value, unless it's too big to sensibly be a backref,
11574 * in which case return I32_MAX (rather than possibly 32-bit wrapping) */
11577 S_backref_value(char *p)
11579 const char* endptr;
11581 if (grok_atoUV(p, &val, &endptr) && val <= I32_MAX)
11588 - regatom - the lowest level
11590 Try to identify anything special at the start of the pattern. If there
11591 is, then handle it as required. This may involve generating a single regop,
11592 such as for an assertion; or it may involve recursing, such as to
11593 handle a () structure.
11595 If the string doesn't start with something special then we gobble up
11596 as much literal text as we can.
11598 Once we have been able to handle whatever type of thing started the
11599 sequence, we return.
11601 Note: we have to be careful with escapes, as they can be both literal
11602 and special, and in the case of \10 and friends, context determines which.
11604 A summary of the code structure is:
11606 switch (first_byte) {
11607 cases for each special:
11608 handle this special;
11611 switch (2nd byte) {
11612 cases for each unambiguous special:
11613 handle this special;
11615 cases for each ambigous special/literal:
11617 if (special) handle here
11619 default: // unambiguously literal:
11622 default: // is a literal char
11625 create EXACTish node for literal;
11626 while (more input and node isn't full) {
11627 switch (input_byte) {
11628 cases for each special;
11629 make sure parse pointer is set so that the next call to
11630 regatom will see this special first
11631 goto loopdone; // EXACTish node terminated by prev. char
11633 append char to EXACTISH node;
11635 get next input byte;
11639 return the generated node;
11641 Specifically there are two separate switches for handling
11642 escape sequences, with the one for handling literal escapes requiring
11643 a dummy entry for all of the special escapes that are actually handled
11646 Returns NULL, setting *flagp to TRYAGAIN if reg() returns NULL with
11648 Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs to be
11650 Otherwise does not return NULL.
11654 S_regatom(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth)
11656 regnode *ret = NULL;
11658 char *parse_start = RExC_parse;
11663 GET_RE_DEBUG_FLAGS_DECL;
11665 *flagp = WORST; /* Tentatively. */
11667 DEBUG_PARSE("atom");
11669 PERL_ARGS_ASSERT_REGATOM;
11672 switch ((U8)*RExC_parse) {
11674 RExC_seen_zerolen++;
11675 nextchar(pRExC_state);
11676 if (RExC_flags & RXf_PMf_MULTILINE)
11677 ret = reg_node(pRExC_state, MBOL);
11679 ret = reg_node(pRExC_state, SBOL);
11680 Set_Node_Length(ret, 1); /* MJD */
11683 nextchar(pRExC_state);
11685 RExC_seen_zerolen++;
11686 if (RExC_flags & RXf_PMf_MULTILINE)
11687 ret = reg_node(pRExC_state, MEOL);
11689 ret = reg_node(pRExC_state, SEOL);
11690 Set_Node_Length(ret, 1); /* MJD */
11693 nextchar(pRExC_state);
11694 if (RExC_flags & RXf_PMf_SINGLELINE)
11695 ret = reg_node(pRExC_state, SANY);
11697 ret = reg_node(pRExC_state, REG_ANY);
11698 *flagp |= HASWIDTH|SIMPLE;
11700 Set_Node_Length(ret, 1); /* MJD */
11704 char * const oregcomp_parse = ++RExC_parse;
11705 ret = regclass(pRExC_state, flagp,depth+1,
11706 FALSE, /* means parse the whole char class */
11707 TRUE, /* allow multi-char folds */
11708 FALSE, /* don't silence non-portable warnings. */
11709 (bool) RExC_strict,
11711 if (*RExC_parse != ']') {
11712 RExC_parse = oregcomp_parse;
11713 vFAIL("Unmatched [");
11716 if (*flagp & RESTART_UTF8)
11718 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
11721 nextchar(pRExC_state);
11722 Set_Node_Length(ret, RExC_parse - oregcomp_parse + 1); /* MJD */
11726 nextchar(pRExC_state);
11727 ret = reg(pRExC_state, 2, &flags,depth+1);
11729 if (flags & TRYAGAIN) {
11730 if (RExC_parse == RExC_end) {
11731 /* Make parent create an empty node if needed. */
11732 *flagp |= TRYAGAIN;
11737 if (flags & RESTART_UTF8) {
11738 *flagp = RESTART_UTF8;
11741 FAIL2("panic: reg returned NULL to regatom, flags=%#"UVxf"",
11744 *flagp |= flags&(HASWIDTH|SPSTART|SIMPLE|POSTPONED);
11748 if (flags & TRYAGAIN) {
11749 *flagp |= TRYAGAIN;
11752 vFAIL("Internal urp");
11753 /* Supposed to be caught earlier. */
11759 vFAIL("Quantifier follows nothing");
11764 This switch handles escape sequences that resolve to some kind
11765 of special regop and not to literal text. Escape sequnces that
11766 resolve to literal text are handled below in the switch marked
11769 Every entry in this switch *must* have a corresponding entry
11770 in the literal escape switch. However, the opposite is not
11771 required, as the default for this switch is to jump to the
11772 literal text handling code.
11774 switch ((U8)*++RExC_parse) {
11775 /* Special Escapes */
11777 RExC_seen_zerolen++;
11778 ret = reg_node(pRExC_state, SBOL);
11779 /* SBOL is shared with /^/ so we set the flags so we can tell
11780 * /\A/ from /^/ in split. We check ret because first pass we
11781 * have no regop struct to set the flags on. */
11785 goto finish_meta_pat;
11787 ret = reg_node(pRExC_state, GPOS);
11788 RExC_seen |= REG_GPOS_SEEN;
11790 goto finish_meta_pat;
11792 RExC_seen_zerolen++;
11793 ret = reg_node(pRExC_state, KEEPS);
11795 /* XXX:dmq : disabling in-place substitution seems to
11796 * be necessary here to avoid cases of memory corruption, as
11797 * with: C<$_="x" x 80; s/x\K/y/> -- rgs
11799 RExC_seen |= REG_LOOKBEHIND_SEEN;
11800 goto finish_meta_pat;
11802 ret = reg_node(pRExC_state, SEOL);
11804 RExC_seen_zerolen++; /* Do not optimize RE away */
11805 goto finish_meta_pat;
11807 ret = reg_node(pRExC_state, EOS);
11809 RExC_seen_zerolen++; /* Do not optimize RE away */
11810 goto finish_meta_pat;
11812 ret = reg_node(pRExC_state, CANY);
11813 RExC_seen |= REG_CANY_SEEN;
11814 *flagp |= HASWIDTH|SIMPLE;
11816 ckWARNdep(RExC_parse+1, "\\C is deprecated");
11818 goto finish_meta_pat;
11820 ret = reg_node(pRExC_state, CLUMP);
11821 *flagp |= HASWIDTH;
11822 goto finish_meta_pat;
11828 arg = ANYOF_WORDCHAR;
11836 regex_charset charset = get_regex_charset(RExC_flags);
11838 RExC_seen_zerolen++;
11839 RExC_seen |= REG_LOOKBEHIND_SEEN;
11840 op = BOUND + charset;
11842 if (op == BOUNDL) {
11843 RExC_contains_locale = 1;
11846 ret = reg_node(pRExC_state, op);
11848 if (*(RExC_parse + 1) != '{') {
11849 FLAGS(ret) = TRADITIONAL_BOUND;
11850 if (PASS2 && op > BOUNDA) { /* /aa is same as /a */
11856 char name = *RExC_parse;
11859 endbrace = strchr(RExC_parse, '}');
11862 vFAIL2("Missing right brace on \\%c{}", name);
11864 /* XXX Need to decide whether to take spaces or not. Should be
11865 * consistent with \p{}, but that currently is SPACE, which
11866 * means vertical too, which seems wrong
11867 * while (isBLANK(*RExC_parse)) {
11870 if (endbrace == RExC_parse) {
11871 RExC_parse++; /* After the '}' */
11872 vFAIL2("Empty \\%c{}", name);
11874 length = endbrace - RExC_parse;
11875 /*while (isBLANK(*(RExC_parse + length - 1))) {
11878 switch (*RExC_parse) {
11881 && (length != 3 || strnNE(RExC_parse + 1, "cb", 2)))
11883 goto bad_bound_type;
11885 FLAGS(ret) = GCB_BOUND;
11888 if (length != 2 || *(RExC_parse + 1) != 'b') {
11889 goto bad_bound_type;
11891 FLAGS(ret) = SB_BOUND;
11894 if (length != 2 || *(RExC_parse + 1) != 'b') {
11895 goto bad_bound_type;
11897 FLAGS(ret) = WB_BOUND;
11901 RExC_parse = endbrace;
11903 "'%"UTF8f"' is an unknown bound type",
11904 UTF8fARG(UTF, length, endbrace - length));
11905 NOT_REACHED; /*NOTREACHED*/
11907 RExC_parse = endbrace;
11908 RExC_uni_semantics = 1;
11910 if (PASS2 && op >= BOUNDA) { /* /aa is same as /a */
11914 /* Don't have to worry about UTF-8, in this message because
11915 * to get here the contents of the \b must be ASCII */
11916 ckWARN4reg(RExC_parse + 1, /* Include the '}' in msg */
11917 "Using /u for '%.*s' instead of /%s",
11919 endbrace - length + 1,
11920 (charset == REGEX_ASCII_RESTRICTED_CHARSET)
11921 ? ASCII_RESTRICT_PAT_MODS
11922 : ASCII_MORE_RESTRICT_PAT_MODS);
11926 if (PASS2 && invert) {
11927 OP(ret) += NBOUND - BOUND;
11929 goto finish_meta_pat;
11937 if (! DEPENDS_SEMANTICS) {
11941 /* \d doesn't have any matches in the upper Latin1 range, hence /d
11942 * is equivalent to /u. Changing to /u saves some branches at
11945 goto join_posix_op_known;
11948 ret = reg_node(pRExC_state, LNBREAK);
11949 *flagp |= HASWIDTH|SIMPLE;
11950 goto finish_meta_pat;
11958 goto join_posix_op_known;
11964 arg = ANYOF_VERTWS;
11966 goto join_posix_op_known;
11976 op = POSIXD + get_regex_charset(RExC_flags);
11977 if (op > POSIXA) { /* /aa is same as /a */
11980 else if (op == POSIXL) {
11981 RExC_contains_locale = 1;
11984 join_posix_op_known:
11987 op += NPOSIXD - POSIXD;
11990 ret = reg_node(pRExC_state, op);
11992 FLAGS(ret) = namedclass_to_classnum(arg);
11995 *flagp |= HASWIDTH|SIMPLE;
11999 nextchar(pRExC_state);
12000 Set_Node_Length(ret, 2); /* MJD */
12006 char* parse_start = RExC_parse - 2;
12011 ret = regclass(pRExC_state, flagp,depth+1,
12012 TRUE, /* means just parse this element */
12013 FALSE, /* don't allow multi-char folds */
12014 FALSE, /* don't silence non-portable warnings.
12015 It would be a bug if these returned
12017 (bool) RExC_strict,
12019 /* regclass() can only return RESTART_UTF8 if multi-char folds
12022 FAIL2("panic: regclass returned NULL to regatom, flags=%#"UVxf"",
12027 Set_Node_Offset(ret, parse_start + 2);
12028 Set_Node_Cur_Length(ret, parse_start);
12029 nextchar(pRExC_state);
12033 /* Handle \N, \N{} and \N{NAMED SEQUENCE} (the latter meaning the
12034 * \N{...} evaluates to a sequence of more than one code points).
12035 * The function call below returns a regnode, which is our result.
12036 * The parameters cause it to fail if the \N{} evaluates to a
12037 * single code point; we handle those like any other literal. The
12038 * reason that the multicharacter case is handled here and not as
12039 * part of the EXACtish code is because of quantifiers. In
12040 * /\N{BLAH}+/, the '+' applies to the whole thing, and doing it
12041 * this way makes that Just Happen. dmq.
12042 * join_exact() will join this up with adjacent EXACTish nodes
12043 * later on, if appropriate. */
12045 if (grok_bslash_N(pRExC_state,
12046 &ret, /* Want a regnode returned */
12047 NULL, /* Fail if evaluates to a single code
12049 NULL, /* Don't need a count of how many code
12057 if (*flagp & RESTART_UTF8)
12062 case 'k': /* Handle \k<NAME> and \k'NAME' */
12065 char ch= RExC_parse[1];
12066 if (ch != '<' && ch != '\'' && ch != '{') {
12068 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12069 vFAIL2("Sequence %.2s... not terminated",parse_start);
12071 /* this pretty much dupes the code for (?P=...) in reg(), if
12072 you change this make sure you change that */
12073 char* name_start = (RExC_parse += 2);
12075 SV *sv_dat = reg_scan_name(pRExC_state,
12076 SIZE_ONLY ? REG_RSN_RETURN_NULL : REG_RSN_RETURN_DATA);
12077 ch= (ch == '<') ? '>' : (ch == '{') ? '}' : '\'';
12078 if (RExC_parse == name_start || *RExC_parse != ch)
12079 /* diag_listed_as: Sequence \%s... not terminated in regex; marked by <-- HERE in m/%s/ */
12080 vFAIL2("Sequence %.3s... not terminated",parse_start);
12083 num = add_data( pRExC_state, STR_WITH_LEN("S"));
12084 RExC_rxi->data->data[num]=(void*)sv_dat;
12085 SvREFCNT_inc_simple_void(sv_dat);
12089 ret = reganode(pRExC_state,
12092 : (ASCII_FOLD_RESTRICTED)
12094 : (AT_LEAST_UNI_SEMANTICS)
12100 *flagp |= HASWIDTH;
12102 /* override incorrect value set in reganode MJD */
12103 Set_Node_Offset(ret, parse_start+1);
12104 Set_Node_Cur_Length(ret, parse_start);
12105 nextchar(pRExC_state);
12111 case '1': case '2': case '3': case '4':
12112 case '5': case '6': case '7': case '8': case '9':
12117 if (*RExC_parse == 'g') {
12121 if (*RExC_parse == '{') {
12125 if (*RExC_parse == '-') {
12129 if (hasbrace && !isDIGIT(*RExC_parse)) {
12130 if (isrel) RExC_parse--;
12132 goto parse_named_seq;
12135 num = S_backref_value(RExC_parse);
12137 vFAIL("Reference to invalid group 0");
12138 else if (num == I32_MAX) {
12139 if (isDIGIT(*RExC_parse))
12140 vFAIL("Reference to nonexistent group");
12142 vFAIL("Unterminated \\g... pattern");
12146 num = RExC_npar - num;
12148 vFAIL("Reference to nonexistent or unclosed group");
12152 num = S_backref_value(RExC_parse);
12153 /* bare \NNN might be backref or octal - if it is larger
12154 * than or equal RExC_npar then it is assumed to be an
12155 * octal escape. Note RExC_npar is +1 from the actual
12156 * number of parens. */
12157 /* Note we do NOT check if num == I32_MAX here, as that is
12158 * handled by the RExC_npar check */
12161 /* any numeric escape < 10 is always a backref */
12163 /* any numeric escape < RExC_npar is a backref */
12164 && num >= RExC_npar
12165 /* cannot be an octal escape if it starts with 8 */
12166 && *RExC_parse != '8'
12167 /* cannot be an octal escape it it starts with 9 */
12168 && *RExC_parse != '9'
12171 /* Probably not a backref, instead likely to be an
12172 * octal character escape, e.g. \35 or \777.
12173 * The above logic should make it obvious why using
12174 * octal escapes in patterns is problematic. - Yves */
12179 /* At this point RExC_parse points at a numeric escape like
12180 * \12 or \88 or something similar, which we should NOT treat
12181 * as an octal escape. It may or may not be a valid backref
12182 * escape. For instance \88888888 is unlikely to be a valid
12185 #ifdef RE_TRACK_PATTERN_OFFSETS
12186 char * const parse_start = RExC_parse - 1; /* MJD */
12188 while (isDIGIT(*RExC_parse))
12191 if (*RExC_parse != '}')
12192 vFAIL("Unterminated \\g{...} pattern");
12196 if (num > (I32)RExC_rx->nparens)
12197 vFAIL("Reference to nonexistent group");
12200 ret = reganode(pRExC_state,
12203 : (ASCII_FOLD_RESTRICTED)
12205 : (AT_LEAST_UNI_SEMANTICS)
12211 *flagp |= HASWIDTH;
12213 /* override incorrect value set in reganode MJD */
12214 Set_Node_Offset(ret, parse_start+1);
12215 Set_Node_Cur_Length(ret, parse_start);
12217 nextchar(pRExC_state);
12222 if (RExC_parse >= RExC_end)
12223 FAIL("Trailing \\");
12226 /* Do not generate "unrecognized" warnings here, we fall
12227 back into the quick-grab loop below */
12234 if (RExC_flags & RXf_PMf_EXTENDED) {
12235 RExC_parse = reg_skipcomment( pRExC_state, RExC_parse );
12236 if (RExC_parse < RExC_end)
12243 parse_start = RExC_parse - 1;
12252 #define MAX_NODE_STRING_SIZE 127
12253 char foldbuf[MAX_NODE_STRING_SIZE+UTF8_MAXBYTES_CASE];
12255 U8 upper_parse = MAX_NODE_STRING_SIZE;
12256 U8 node_type = compute_EXACTish(pRExC_state);
12257 bool next_is_quantifier;
12258 char * oldp = NULL;
12260 /* We can convert EXACTF nodes to EXACTFU if they contain only
12261 * characters that match identically regardless of the target
12262 * string's UTF8ness. The reason to do this is that EXACTF is not
12263 * trie-able, EXACTFU is.
12265 * Similarly, we can convert EXACTFL nodes to EXACTFU if they
12266 * contain only above-Latin1 characters (hence must be in UTF8),
12267 * which don't participate in folds with Latin1-range characters,
12268 * as the latter's folds aren't known until runtime. (We don't
12269 * need to figure this out until pass 2) */
12270 bool maybe_exactfu = PASS2
12271 && (node_type == EXACTF || node_type == EXACTFL);
12273 /* If a folding node contains only code points that don't
12274 * participate in folds, it can be changed into an EXACT node,
12275 * which allows the optimizer more things to look for */
12278 ret = reg_node(pRExC_state, node_type);
12280 /* In pass1, folded, we use a temporary buffer instead of the
12281 * actual node, as the node doesn't exist yet */
12282 s = (SIZE_ONLY && FOLD) ? foldbuf : STRING(ret);
12288 /* We do the EXACTFish to EXACT node only if folding. (And we
12289 * don't need to figure this out until pass 2) */
12290 maybe_exact = FOLD && PASS2;
12292 /* XXX The node can hold up to 255 bytes, yet this only goes to
12293 * 127. I (khw) do not know why. Keeping it somewhat less than
12294 * 255 allows us to not have to worry about overflow due to
12295 * converting to utf8 and fold expansion, but that value is
12296 * 255-UTF8_MAXBYTES_CASE. join_exact() may join adjacent nodes
12297 * split up by this limit into a single one using the real max of
12298 * 255. Even at 127, this breaks under rare circumstances. If
12299 * folding, we do not want to split a node at a character that is a
12300 * non-final in a multi-char fold, as an input string could just
12301 * happen to want to match across the node boundary. The join
12302 * would solve that problem if the join actually happens. But a
12303 * series of more than two nodes in a row each of 127 would cause
12304 * the first join to succeed to get to 254, but then there wouldn't
12305 * be room for the next one, which could at be one of those split
12306 * multi-char folds. I don't know of any fool-proof solution. One
12307 * could back off to end with only a code point that isn't such a
12308 * non-final, but it is possible for there not to be any in the
12310 for (p = RExC_parse - 1;
12311 len < upper_parse && p < RExC_end;
12316 if (RExC_flags & RXf_PMf_EXTENDED)
12317 p = regpatws(pRExC_state, p,
12318 TRUE); /* means recognize comments */
12329 /* Literal Escapes Switch
12331 This switch is meant to handle escape sequences that
12332 resolve to a literal character.
12334 Every escape sequence that represents something
12335 else, like an assertion or a char class, is handled
12336 in the switch marked 'Special Escapes' above in this
12337 routine, but also has an entry here as anything that
12338 isn't explicitly mentioned here will be treated as
12339 an unescaped equivalent literal.
12342 switch ((U8)*++p) {
12343 /* These are all the special escapes. */
12344 case 'A': /* Start assertion */
12345 case 'b': case 'B': /* Word-boundary assertion*/
12346 case 'C': /* Single char !DANGEROUS! */
12347 case 'd': case 'D': /* digit class */
12348 case 'g': case 'G': /* generic-backref, pos assertion */
12349 case 'h': case 'H': /* HORIZWS */
12350 case 'k': case 'K': /* named backref, keep marker */
12351 case 'p': case 'P': /* Unicode property */
12352 case 'R': /* LNBREAK */
12353 case 's': case 'S': /* space class */
12354 case 'v': case 'V': /* VERTWS */
12355 case 'w': case 'W': /* word class */
12356 case 'X': /* eXtended Unicode "combining
12357 character sequence" */
12358 case 'z': case 'Z': /* End of line/string assertion */
12362 /* Anything after here is an escape that resolves to a
12363 literal. (Except digits, which may or may not)
12369 case 'N': /* Handle a single-code point named character. */
12370 RExC_parse = p + 1;
12371 if (! grok_bslash_N(pRExC_state,
12372 NULL, /* Fail if evaluates to
12373 anything other than a
12374 single code point */
12375 &ender, /* The returned single code
12377 NULL, /* Don't need a count of
12378 how many code points */
12382 if (*flagp & RESTART_UTF8)
12383 FAIL("panic: grok_bslash_N set RESTART_UTF8");
12385 /* Here, it wasn't a single code point. Go close
12386 * up this EXACTish node. The switch() prior to
12387 * this switch handles the other cases */
12388 RExC_parse = p = oldp;
12392 if (ender > 0xff) {
12409 ender = ESC_NATIVE;
12419 const char* error_msg;
12421 bool valid = grok_bslash_o(&p,
12424 PASS2, /* out warnings */
12425 (bool) RExC_strict,
12426 TRUE, /* Output warnings
12431 RExC_parse = p; /* going to die anyway; point
12432 to exact spot of failure */
12436 if (IN_ENCODING && ender < 0x100) {
12437 goto recode_encoding;
12439 if (ender > 0xff) {
12446 UV result = UV_MAX; /* initialize to erroneous
12448 const char* error_msg;
12450 bool valid = grok_bslash_x(&p,
12453 PASS2, /* out warnings */
12454 (bool) RExC_strict,
12455 TRUE, /* Silence warnings
12460 RExC_parse = p; /* going to die anyway; point
12461 to exact spot of failure */
12466 if (ender < 0x100) {
12468 if (RExC_recode_x_to_native) {
12469 ender = LATIN1_TO_NATIVE(ender);
12474 goto recode_encoding;
12484 ender = grok_bslash_c(*p++, PASS2);
12486 case '8': case '9': /* must be a backreference */
12488 /* we have an escape like \8 which cannot be an octal escape
12489 * so we exit the loop, and let the outer loop handle this
12490 * escape which may or may not be a legitimate backref. */
12492 case '1': case '2': case '3':case '4':
12493 case '5': case '6': case '7':
12494 /* When we parse backslash escapes there is ambiguity
12495 * between backreferences and octal escapes. Any escape
12496 * from \1 - \9 is a backreference, any multi-digit
12497 * escape which does not start with 0 and which when
12498 * evaluated as decimal could refer to an already
12499 * parsed capture buffer is a back reference. Anything
12502 * Note this implies that \118 could be interpreted as
12503 * 118 OR as "\11" . "8" depending on whether there
12504 * were 118 capture buffers defined already in the
12507 /* NOTE, RExC_npar is 1 more than the actual number of
12508 * parens we have seen so far, hence the < RExC_npar below. */
12510 if ( !isDIGIT(p[1]) || S_backref_value(p) < RExC_npar)
12511 { /* Not to be treated as an octal constant, go
12519 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
12521 ender = grok_oct(p, &numlen, &flags, NULL);
12522 if (ender > 0xff) {
12526 if (PASS2 /* like \08, \178 */
12529 && isDIGIT(*p) && ckWARN(WARN_REGEXP))
12531 reg_warn_non_literal_string(
12533 form_short_octal_warning(p, numlen));
12536 if (IN_ENCODING && ender < 0x100)
12537 goto recode_encoding;
12540 if (! RExC_override_recoding) {
12541 SV* enc = _get_encoding();
12542 ender = reg_recode((const char)(U8)ender, &enc);
12544 ckWARNreg(p, "Invalid escape in the specified encoding");
12550 FAIL("Trailing \\");
12553 if (!SIZE_ONLY&& isALPHANUMERIC(*p)) {
12554 /* Include any { following the alpha to emphasize
12555 * that it could be part of an escape at some point
12557 int len = (isALPHA(*p) && *(p + 1) == '{') ? 2 : 1;
12558 ckWARN3reg(p + len, "Unrecognized escape \\%.*s passed through", len, p);
12560 goto normal_default;
12561 } /* End of switch on '\' */
12564 /* Currently we don't warn when the lbrace is at the start
12565 * of a construct. This catches it in the middle of a
12566 * literal string, or when its the first thing after
12567 * something like "\b" */
12569 && (len || (p > RExC_start && isALPHA_A(*(p -1)))))
12571 ckWARNregdep(p + 1, "Unescaped left brace in regex is deprecated, passed through");
12574 default: /* A literal character */
12576 if (UTF8_IS_START(*p) && UTF) {
12578 ender = utf8n_to_uvchr((U8*)p, RExC_end - p,
12579 &numlen, UTF8_ALLOW_DEFAULT);
12585 } /* End of switch on the literal */
12587 /* Here, have looked at the literal character and <ender>
12588 * contains its ordinal, <p> points to the character after it
12591 if ( RExC_flags & RXf_PMf_EXTENDED)
12592 p = regpatws(pRExC_state, p,
12593 TRUE); /* means recognize comments */
12595 /* If the next thing is a quantifier, it applies to this
12596 * character only, which means that this character has to be in
12597 * its own node and can't just be appended to the string in an
12598 * existing node, so if there are already other characters in
12599 * the node, close the node with just them, and set up to do
12600 * this character again next time through, when it will be the
12601 * only thing in its new node */
12602 if ((next_is_quantifier = (p < RExC_end && ISMULT2(p))) && len)
12608 if (! FOLD) { /* The simple case, just append the literal */
12610 /* In the sizing pass, we need only the size of the
12611 * character we are appending, hence we can delay getting
12612 * its representation until PASS2. */
12615 const STRLEN unilen = UNISKIP(ender);
12618 /* We have to subtract 1 just below (and again in
12619 * the corresponding PASS2 code) because the loop
12620 * increments <len> each time, as all but this path
12621 * (and one other) through it add a single byte to
12622 * the EXACTish node. But these paths would change
12623 * len to be the correct final value, so cancel out
12624 * the increment that follows */
12630 } else { /* PASS2 */
12633 U8 * new_s = uvchr_to_utf8((U8*)s, ender);
12634 len += (char *) new_s - s - 1;
12635 s = (char *) new_s;
12638 *(s++) = (char) ender;
12642 else if (LOC && is_PROBLEMATIC_LOCALE_FOLD_cp(ender)) {
12644 /* Here are folding under /l, and the code point is
12645 * problematic. First, we know we can't simplify things */
12646 maybe_exact = FALSE;
12647 maybe_exactfu = FALSE;
12649 /* A problematic code point in this context means that its
12650 * fold isn't known until runtime, so we can't fold it now.
12651 * (The non-problematic code points are the above-Latin1
12652 * ones that fold to also all above-Latin1. Their folds
12653 * don't vary no matter what the locale is.) But here we
12654 * have characters whose fold depends on the locale.
12655 * Unlike the non-folding case above, we have to keep track
12656 * of these in the sizing pass, so that we can make sure we
12657 * don't split too-long nodes in the middle of a potential
12658 * multi-char fold. And unlike the regular fold case
12659 * handled in the else clauses below, we don't actually
12660 * fold and don't have special cases to consider. What we
12661 * do for both passes is the PASS2 code for non-folding */
12662 goto not_fold_common;
12664 else /* A regular FOLD code point */
12666 /* See comments for join_exact() as to why we fold this
12667 * non-UTF at compile time */
12668 || (node_type == EXACTFU
12669 && ender == LATIN_SMALL_LETTER_SHARP_S)))
12671 /* Here, are folding and are not UTF-8 encoded; therefore
12672 * the character must be in the range 0-255, and is not /l
12673 * (Not /l because we already handled these under /l in
12674 * is_PROBLEMATIC_LOCALE_FOLD_cp) */
12675 if (IS_IN_SOME_FOLD_L1(ender)) {
12676 maybe_exact = FALSE;
12678 /* See if the character's fold differs between /d and
12679 * /u. This includes the multi-char fold SHARP S to
12682 && (PL_fold[ender] != PL_fold_latin1[ender]
12683 || ender == LATIN_SMALL_LETTER_SHARP_S
12685 && isALPHA_FOLD_EQ(ender, 's')
12686 && isALPHA_FOLD_EQ(*(s-1), 's'))))
12688 maybe_exactfu = FALSE;
12692 /* Even when folding, we store just the input character, as
12693 * we have an array that finds its fold quickly */
12694 *(s++) = (char) ender;
12696 else { /* FOLD and UTF */
12697 /* Unlike the non-fold case, we do actually have to
12698 * calculate the results here in pass 1. This is for two
12699 * reasons, the folded length may be longer than the
12700 * unfolded, and we have to calculate how many EXACTish
12701 * nodes it will take; and we may run out of room in a node
12702 * in the middle of a potential multi-char fold, and have
12703 * to back off accordingly. */
12706 if (isASCII_uni(ender)) {
12707 folded = toFOLD(ender);
12708 *(s)++ = (U8) folded;
12713 folded = _to_uni_fold_flags(
12717 FOLD_FLAGS_FULL | ((ASCII_FOLD_RESTRICTED)
12718 ? FOLD_FLAGS_NOMIX_ASCII
12722 /* The loop increments <len> each time, as all but this
12723 * path (and one other) through it add a single byte to
12724 * the EXACTish node. But this one has changed len to
12725 * be the correct final value, so subtract one to
12726 * cancel out the increment that follows */
12727 len += foldlen - 1;
12729 /* If this node only contains non-folding code points so
12730 * far, see if this new one is also non-folding */
12732 if (folded != ender) {
12733 maybe_exact = FALSE;
12736 /* Here the fold is the original; we have to check
12737 * further to see if anything folds to it */
12738 if (_invlist_contains_cp(PL_utf8_foldable,
12741 maybe_exact = FALSE;
12748 if (next_is_quantifier) {
12750 /* Here, the next input is a quantifier, and to get here,
12751 * the current character is the only one in the node.
12752 * Also, here <len> doesn't include the final byte for this
12758 } /* End of loop through literal characters */
12760 /* Here we have either exhausted the input or ran out of room in
12761 * the node. (If we encountered a character that can't be in the
12762 * node, transfer is made directly to <loopdone>, and so we
12763 * wouldn't have fallen off the end of the loop.) In the latter
12764 * case, we artificially have to split the node into two, because
12765 * we just don't have enough space to hold everything. This
12766 * creates a problem if the final character participates in a
12767 * multi-character fold in the non-final position, as a match that
12768 * should have occurred won't, due to the way nodes are matched,
12769 * and our artificial boundary. So back off until we find a non-
12770 * problematic character -- one that isn't at the beginning or
12771 * middle of such a fold. (Either it doesn't participate in any
12772 * folds, or appears only in the final position of all the folds it
12773 * does participate in.) A better solution with far fewer false
12774 * positives, and that would fill the nodes more completely, would
12775 * be to actually have available all the multi-character folds to
12776 * test against, and to back-off only far enough to be sure that
12777 * this node isn't ending with a partial one. <upper_parse> is set
12778 * further below (if we need to reparse the node) to include just
12779 * up through that final non-problematic character that this code
12780 * identifies, so when it is set to less than the full node, we can
12781 * skip the rest of this */
12782 if (FOLD && p < RExC_end && upper_parse == MAX_NODE_STRING_SIZE) {
12784 const STRLEN full_len = len;
12786 assert(len >= MAX_NODE_STRING_SIZE);
12788 /* Here, <s> points to the final byte of the final character.
12789 * Look backwards through the string until find a non-
12790 * problematic character */
12794 /* This has no multi-char folds to non-UTF characters */
12795 if (ASCII_FOLD_RESTRICTED) {
12799 while (--s >= s0 && IS_NON_FINAL_FOLD(*s)) { }
12803 if (! PL_NonL1NonFinalFold) {
12804 PL_NonL1NonFinalFold = _new_invlist_C_array(
12805 NonL1_Perl_Non_Final_Folds_invlist);
12808 /* Point to the first byte of the final character */
12809 s = (char *) utf8_hop((U8 *) s, -1);
12811 while (s >= s0) { /* Search backwards until find
12812 non-problematic char */
12813 if (UTF8_IS_INVARIANT(*s)) {
12815 /* There are no ascii characters that participate
12816 * in multi-char folds under /aa. In EBCDIC, the
12817 * non-ascii invariants are all control characters,
12818 * so don't ever participate in any folds. */
12819 if (ASCII_FOLD_RESTRICTED
12820 || ! IS_NON_FINAL_FOLD(*s))
12825 else if (UTF8_IS_DOWNGRADEABLE_START(*s)) {
12826 if (! IS_NON_FINAL_FOLD(TWO_BYTE_UTF8_TO_NATIVE(
12832 else if (! _invlist_contains_cp(
12833 PL_NonL1NonFinalFold,
12834 valid_utf8_to_uvchr((U8 *) s, NULL)))
12839 /* Here, the current character is problematic in that
12840 * it does occur in the non-final position of some
12841 * fold, so try the character before it, but have to
12842 * special case the very first byte in the string, so
12843 * we don't read outside the string */
12844 s = (s == s0) ? s -1 : (char *) utf8_hop((U8 *) s, -1);
12845 } /* End of loop backwards through the string */
12847 /* If there were only problematic characters in the string,
12848 * <s> will point to before s0, in which case the length
12849 * should be 0, otherwise include the length of the
12850 * non-problematic character just found */
12851 len = (s < s0) ? 0 : s - s0 + UTF8SKIP(s);
12854 /* Here, have found the final character, if any, that is
12855 * non-problematic as far as ending the node without splitting
12856 * it across a potential multi-char fold. <len> contains the
12857 * number of bytes in the node up-to and including that
12858 * character, or is 0 if there is no such character, meaning
12859 * the whole node contains only problematic characters. In
12860 * this case, give up and just take the node as-is. We can't
12865 /* If the node ends in an 's' we make sure it stays EXACTF,
12866 * as if it turns into an EXACTFU, it could later get
12867 * joined with another 's' that would then wrongly match
12869 if (maybe_exactfu && isALPHA_FOLD_EQ(ender, 's'))
12871 maybe_exactfu = FALSE;
12875 /* Here, the node does contain some characters that aren't
12876 * problematic. If one such is the final character in the
12877 * node, we are done */
12878 if (len == full_len) {
12881 else if (len + ((UTF) ? UTF8SKIP(s) : 1) == full_len) {
12883 /* If the final character is problematic, but the
12884 * penultimate is not, back-off that last character to
12885 * later start a new node with it */
12890 /* Here, the final non-problematic character is earlier
12891 * in the input than the penultimate character. What we do
12892 * is reparse from the beginning, going up only as far as
12893 * this final ok one, thus guaranteeing that the node ends
12894 * in an acceptable character. The reason we reparse is
12895 * that we know how far in the character is, but we don't
12896 * know how to correlate its position with the input parse.
12897 * An alternate implementation would be to build that
12898 * correlation as we go along during the original parse,
12899 * but that would entail extra work for every node, whereas
12900 * this code gets executed only when the string is too
12901 * large for the node, and the final two characters are
12902 * problematic, an infrequent occurrence. Yet another
12903 * possible strategy would be to save the tail of the
12904 * string, and the next time regatom is called, initialize
12905 * with that. The problem with this is that unless you
12906 * back off one more character, you won't be guaranteed
12907 * regatom will get called again, unless regbranch,
12908 * regpiece ... are also changed. If you do back off that
12909 * extra character, so that there is input guaranteed to
12910 * force calling regatom, you can't handle the case where
12911 * just the first character in the node is acceptable. I
12912 * (khw) decided to try this method which doesn't have that
12913 * pitfall; if performance issues are found, we can do a
12914 * combination of the current approach plus that one */
12920 } /* End of verifying node ends with an appropriate char */
12922 loopdone: /* Jumped to when encounters something that shouldn't be
12925 /* I (khw) don't know if you can get here with zero length, but the
12926 * old code handled this situation by creating a zero-length EXACT
12927 * node. Might as well be NOTHING instead */
12933 /* If 'maybe_exact' is still set here, means there are no
12934 * code points in the node that participate in folds;
12935 * similarly for 'maybe_exactfu' and code points that match
12936 * differently depending on UTF8ness of the target string
12937 * (for /u), or depending on locale for /l */
12943 else if (maybe_exactfu) {
12949 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, len, ender,
12950 FALSE /* Don't look to see if could
12951 be turned into an EXACT
12952 node, as we have already
12957 RExC_parse = p - 1;
12958 Set_Node_Cur_Length(ret, parse_start);
12959 nextchar(pRExC_state);
12961 /* len is STRLEN which is unsigned, need to copy to signed */
12964 vFAIL("Internal disaster");
12967 } /* End of label 'defchar:' */
12969 } /* End of giant switch on input character */
12975 S_regpatws(RExC_state_t *pRExC_state, char *p , const bool recognize_comment )
12977 /* Returns the next non-pattern-white space, non-comment character (the
12978 * latter only if 'recognize_comment is true) in the string p, which is
12979 * ended by RExC_end. See also reg_skipcomment */
12980 const char *e = RExC_end;
12982 PERL_ARGS_ASSERT_REGPATWS;
12986 if ((len = is_PATWS_safe(p, e, UTF))) {
12989 else if (recognize_comment && *p == '#') {
12990 p = reg_skipcomment(pRExC_state, p);
12999 S_populate_ANYOF_from_invlist(pTHX_ regnode *node, SV** invlist_ptr)
13001 /* Uses the inversion list '*invlist_ptr' to populate the ANYOF 'node'. It
13002 * sets up the bitmap and any flags, removing those code points from the
13003 * inversion list, setting it to NULL should it become completely empty */
13005 PERL_ARGS_ASSERT_POPULATE_ANYOF_FROM_INVLIST;
13006 assert(PL_regkind[OP(node)] == ANYOF);
13008 ANYOF_BITMAP_ZERO(node);
13009 if (*invlist_ptr) {
13011 /* This gets set if we actually need to modify things */
13012 bool change_invlist = FALSE;
13016 /* Start looking through *invlist_ptr */
13017 invlist_iterinit(*invlist_ptr);
13018 while (invlist_iternext(*invlist_ptr, &start, &end)) {
13022 if (end == UV_MAX && start <= NUM_ANYOF_CODE_POINTS) {
13023 ANYOF_FLAGS(node) |= ANYOF_MATCHES_ALL_ABOVE_BITMAP;
13025 else if (end >= NUM_ANYOF_CODE_POINTS) {
13026 ANYOF_FLAGS(node) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
13029 /* Quit if are above what we should change */
13030 if (start >= NUM_ANYOF_CODE_POINTS) {
13034 change_invlist = TRUE;
13036 /* Set all the bits in the range, up to the max that we are doing */
13037 high = (end < NUM_ANYOF_CODE_POINTS - 1)
13039 : NUM_ANYOF_CODE_POINTS - 1;
13040 for (i = start; i <= (int) high; i++) {
13041 if (! ANYOF_BITMAP_TEST(node, i)) {
13042 ANYOF_BITMAP_SET(node, i);
13046 invlist_iterfinish(*invlist_ptr);
13048 /* Done with loop; remove any code points that are in the bitmap from
13049 * *invlist_ptr; similarly for code points above the bitmap if we have
13050 * a flag to match all of them anyways */
13051 if (change_invlist) {
13052 _invlist_subtract(*invlist_ptr, PL_InBitmap, invlist_ptr);
13054 if (ANYOF_FLAGS(node) & ANYOF_MATCHES_ALL_ABOVE_BITMAP) {
13055 _invlist_intersection(*invlist_ptr, PL_InBitmap, invlist_ptr);
13058 /* If have completely emptied it, remove it completely */
13059 if (_invlist_len(*invlist_ptr) == 0) {
13060 SvREFCNT_dec_NN(*invlist_ptr);
13061 *invlist_ptr = NULL;
13066 /* Parse POSIX character classes: [[:foo:]], [[=foo=]], [[.foo.]].
13067 Character classes ([:foo:]) can also be negated ([:^foo:]).
13068 Returns a named class id (ANYOF_XXX) if successful, -1 otherwise.
13069 Equivalence classes ([=foo=]) and composites ([.foo.]) are parsed,
13070 but trigger failures because they are currently unimplemented. */
13072 #define POSIXCC_DONE(c) ((c) == ':')
13073 #define POSIXCC_NOTYET(c) ((c) == '=' || (c) == '.')
13074 #define POSIXCC(c) (POSIXCC_DONE(c) || POSIXCC_NOTYET(c))
13076 PERL_STATIC_INLINE I32
13077 S_regpposixcc(pTHX_ RExC_state_t *pRExC_state, I32 value, const bool strict)
13079 I32 namedclass = OOB_NAMEDCLASS;
13081 PERL_ARGS_ASSERT_REGPPOSIXCC;
13083 if (value == '[' && RExC_parse + 1 < RExC_end &&
13084 /* I smell either [: or [= or [. -- POSIX has been here, right? */
13085 POSIXCC(UCHARAT(RExC_parse)))
13087 const char c = UCHARAT(RExC_parse);
13088 char* const s = RExC_parse++;
13090 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != c)
13092 if (RExC_parse == RExC_end) {
13095 /* Try to give a better location for the error (than the end of
13096 * the string) by looking for the matching ']' */
13098 while (RExC_parse < RExC_end && UCHARAT(RExC_parse) != ']') {
13101 vFAIL2("Unmatched '%c' in POSIX class", c);
13103 /* Grandfather lone [:, [=, [. */
13107 const char* const t = RExC_parse++; /* skip over the c */
13110 if (UCHARAT(RExC_parse) == ']') {
13111 const char *posixcc = s + 1;
13112 RExC_parse++; /* skip over the ending ] */
13115 const I32 complement = *posixcc == '^' ? *posixcc++ : 0;
13116 const I32 skip = t - posixcc;
13118 /* Initially switch on the length of the name. */
13121 if (memEQ(posixcc, "word", 4)) /* this is not POSIX,
13122 this is the Perl \w
13124 namedclass = ANYOF_WORDCHAR;
13127 /* Names all of length 5. */
13128 /* alnum alpha ascii blank cntrl digit graph lower
13129 print punct space upper */
13130 /* Offset 4 gives the best switch position. */
13131 switch (posixcc[4]) {
13133 if (memEQ(posixcc, "alph", 4)) /* alpha */
13134 namedclass = ANYOF_ALPHA;
13137 if (memEQ(posixcc, "spac", 4)) /* space */
13138 namedclass = ANYOF_SPACE;
13141 if (memEQ(posixcc, "grap", 4)) /* graph */
13142 namedclass = ANYOF_GRAPH;
13145 if (memEQ(posixcc, "asci", 4)) /* ascii */
13146 namedclass = ANYOF_ASCII;
13149 if (memEQ(posixcc, "blan", 4)) /* blank */
13150 namedclass = ANYOF_BLANK;
13153 if (memEQ(posixcc, "cntr", 4)) /* cntrl */
13154 namedclass = ANYOF_CNTRL;
13157 if (memEQ(posixcc, "alnu", 4)) /* alnum */
13158 namedclass = ANYOF_ALPHANUMERIC;
13161 if (memEQ(posixcc, "lowe", 4)) /* lower */
13162 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_LOWER;
13163 else if (memEQ(posixcc, "uppe", 4)) /* upper */
13164 namedclass = (FOLD) ? ANYOF_CASED : ANYOF_UPPER;
13167 if (memEQ(posixcc, "digi", 4)) /* digit */
13168 namedclass = ANYOF_DIGIT;
13169 else if (memEQ(posixcc, "prin", 4)) /* print */
13170 namedclass = ANYOF_PRINT;
13171 else if (memEQ(posixcc, "punc", 4)) /* punct */
13172 namedclass = ANYOF_PUNCT;
13177 if (memEQ(posixcc, "xdigit", 6))
13178 namedclass = ANYOF_XDIGIT;
13182 if (namedclass == OOB_NAMEDCLASS)
13184 "POSIX class [:%"UTF8f":] unknown",
13185 UTF8fARG(UTF, t - s - 1, s + 1));
13187 /* The #defines are structured so each complement is +1 to
13188 * the normal one */
13192 assert (posixcc[skip] == ':');
13193 assert (posixcc[skip+1] == ']');
13194 } else if (!SIZE_ONLY) {
13195 /* [[=foo=]] and [[.foo.]] are still future. */
13197 /* adjust RExC_parse so the warning shows after
13198 the class closes */
13199 while (UCHARAT(RExC_parse) && UCHARAT(RExC_parse) != ']')
13201 vFAIL3("POSIX syntax [%c %c] is reserved for future extensions", c, c);
13204 /* Maternal grandfather:
13205 * "[:" ending in ":" but not in ":]" */
13207 vFAIL("Unmatched '[' in POSIX class");
13210 /* Grandfather lone [:, [=, [. */
13220 S_could_it_be_a_POSIX_class(RExC_state_t *pRExC_state)
13222 /* This applies some heuristics at the current parse position (which should
13223 * be at a '[') to see if what follows might be intended to be a [:posix:]
13224 * class. It returns true if it really is a posix class, of course, but it
13225 * also can return true if it thinks that what was intended was a posix
13226 * class that didn't quite make it.
13228 * It will return true for
13230 * [:alphanumerics] (as long as the ] isn't followed immediately by a
13231 * ')' indicating the end of the (?[
13232 * [:any garbage including %^&$ punctuation:]
13234 * This is designed to be called only from S_handle_regex_sets; it could be
13235 * easily adapted to be called from the spot at the beginning of regclass()
13236 * that checks to see in a normal bracketed class if the surrounding []
13237 * have been omitted ([:word:] instead of [[:word:]]). But doing so would
13238 * change long-standing behavior, so I (khw) didn't do that */
13239 char* p = RExC_parse + 1;
13240 char first_char = *p;
13242 PERL_ARGS_ASSERT_COULD_IT_BE_A_POSIX_CLASS;
13244 assert(*(p - 1) == '[');
13246 if (! POSIXCC(first_char)) {
13251 while (p < RExC_end && isWORDCHAR(*p)) p++;
13253 if (p >= RExC_end) {
13257 if (p - RExC_parse > 2 /* Got at least 1 word character */
13258 && (*p == first_char
13259 || (*p == ']' && p + 1 < RExC_end && *(p + 1) != ')')))
13264 p = (char *) memchr(RExC_parse, ']', RExC_end - RExC_parse);
13267 && p - RExC_parse > 2 /* [:] evaluates to colon;
13268 [::] is a bad posix class. */
13269 && first_char == *(p - 1));
13272 STATIC unsigned int
13273 S_regex_set_precedence(const U8 my_operator) {
13275 /* Returns the precedence in the (?[...]) construct of the input operator,
13276 * specified by its character representation. The precedence follows
13277 * general Perl rules, but it extends this so that ')' and ']' have (low)
13278 * precedence even though they aren't really operators */
13280 switch (my_operator) {
13296 NOT_REACHED; /* NOTREACHED */
13297 return 0; /* Silence compiler warning */
13301 S_handle_regex_sets(pTHX_ RExC_state_t *pRExC_state, SV** return_invlist,
13302 I32 *flagp, U32 depth,
13303 char * const oregcomp_parse)
13305 /* Handle the (?[...]) construct to do set operations */
13307 U8 curchar; /* Current character being parsed */
13308 UV start, end; /* End points of code point ranges */
13309 SV* final = NULL; /* The end result inversion list */
13310 SV* result_string; /* 'final' stringified */
13311 AV* stack; /* stack of operators and operands not yet
13313 AV* fence_stack = NULL; /* A stack containing the positions in
13314 'stack' of where the undealt-with left
13315 parens would be if they were actually
13317 IV fence = 0; /* Position of where most recent undealt-
13318 with left paren in stack is; -1 if none.
13320 STRLEN len; /* Temporary */
13321 regnode* node; /* Temporary, and final regnode returned by
13323 const bool save_fold = FOLD; /* Temporary */
13324 char *save_end, *save_parse; /* Temporaries */
13326 GET_RE_DEBUG_FLAGS_DECL;
13328 PERL_ARGS_ASSERT_HANDLE_REGEX_SETS;
13330 if (LOC) { /* XXX could make valid in UTF-8 locales */
13331 vFAIL("(?[...]) not valid in locale");
13333 RExC_uni_semantics = 1; /* The use of this operator implies /u. This
13334 is required so that the compile time values
13335 are valid in all runtime cases */
13337 /* This will return only an ANYOF regnode, or (unlikely) something smaller
13338 * (such as EXACT). Thus we can skip most everything if just sizing. We
13339 * call regclass to handle '[]' so as to not have to reinvent its parsing
13340 * rules here (throwing away the size it computes each time). And, we exit
13341 * upon an unescaped ']' that isn't one ending a regclass. To do both
13342 * these things, we need to realize that something preceded by a backslash
13343 * is escaped, so we have to keep track of backslashes */
13345 UV depth = 0; /* how many nested (?[...]) constructs */
13347 while (RExC_parse < RExC_end) {
13348 SV* current = NULL;
13349 RExC_parse = regpatws(pRExC_state, RExC_parse,
13350 TRUE); /* means recognize comments */
13351 switch (*RExC_parse) {
13353 if (RExC_parse[1] == '[') depth++, RExC_parse++;
13358 /* Skip the next byte (which could cause us to end up in
13359 * the middle of a UTF-8 character, but since none of those
13360 * are confusable with anything we currently handle in this
13361 * switch (invariants all), it's safe. We'll just hit the
13362 * default: case next time and keep on incrementing until
13363 * we find one of the invariants we do handle. */
13368 /* If this looks like it is a [:posix:] class, leave the
13369 * parse pointer at the '[' to fool regclass() into
13370 * thinking it is part of a '[[:posix:]]'. That function
13371 * will use strict checking to force a syntax error if it
13372 * doesn't work out to a legitimate class */
13373 bool is_posix_class
13374 = could_it_be_a_POSIX_class(pRExC_state);
13375 if (! is_posix_class) {
13379 /* regclass() can only return RESTART_UTF8 if multi-char
13380 folds are allowed. */
13381 if (!regclass(pRExC_state, flagp,depth+1,
13382 is_posix_class, /* parse the whole char
13383 class only if not a
13385 FALSE, /* don't allow multi-char folds */
13386 TRUE, /* silence non-portable warnings. */
13390 FAIL2("panic: regclass returned NULL to handle_sets, "
13391 "flags=%#"UVxf"", (UV) *flagp);
13393 /* function call leaves parse pointing to the ']', except
13394 * if we faked it */
13395 if (is_posix_class) {
13399 SvREFCNT_dec(current); /* In case it returned something */
13404 if (depth--) break;
13406 if (RExC_parse < RExC_end
13407 && *RExC_parse == ')')
13409 node = reganode(pRExC_state, ANYOF, 0);
13410 RExC_size += ANYOF_SKIP;
13411 nextchar(pRExC_state);
13412 Set_Node_Length(node,
13413 RExC_parse - oregcomp_parse + 1); /* MJD */
13422 FAIL("Syntax error in (?[...])");
13425 /* Pass 2 only after this. */
13426 Perl_ck_warner_d(aTHX_
13427 packWARN(WARN_EXPERIMENTAL__REGEX_SETS),
13428 "The regex_sets feature is experimental" REPORT_LOCATION,
13429 UTF8fARG(UTF, (RExC_parse - RExC_precomp), RExC_precomp),
13431 RExC_end - RExC_start - (RExC_parse - RExC_precomp),
13432 RExC_precomp + (RExC_parse - RExC_precomp)));
13434 /* Everything in this construct is a metacharacter. Operands begin with
13435 * either a '\' (for an escape sequence), or a '[' for a bracketed
13436 * character class. Any other character should be an operator, or
13437 * parenthesis for grouping. Both types of operands are handled by calling
13438 * regclass() to parse them. It is called with a parameter to indicate to
13439 * return the computed inversion list. The parsing here is implemented via
13440 * a stack. Each entry on the stack is a single character representing one
13441 * of the operators; or else a pointer to an operand inversion list. */
13443 #define IS_OPERAND(a) (! SvIOK(a))
13445 /* The stack is kept in Łukasiewicz order. (That's pronounced similar
13446 * to luke-a-shave-itch (or -itz), but people who didn't want to bother
13447 * with prounouncing it called it Reverse Polish instead, but now that YOU
13448 * know how to prounounce it you can use the correct term, thus giving due
13449 * credit to the person who invented it, and impressing your geek friends.
13450 * Wikipedia says that the pronounciation of "Ł" has been changing so that
13451 * it is now more like an English initial W (as in wonk) than an L.)
13453 * This means that, for example, 'a | b & c' is stored on the stack as
13461 * where the numbers in brackets give the stack [array] element number.
13462 * In this implementation, parentheses are not stored on the stack.
13463 * Instead a '(' creates a "fence" so that the part of the stack below the
13464 * fence is invisible except to the corresponding ')' (this allows us to
13465 * replace testing for parens, by using instead subtraction of the fence
13466 * position). As new operands are processed they are pushed onto the stack
13467 * (except as noted in the next paragraph). New operators of higher
13468 * precedence than the current final one are inserted on the stack before
13469 * the lhs operand (so that when the rhs is pushed next, everything will be
13470 * in the correct positions shown above. When an operator of equal or
13471 * lower precedence is encountered in parsing, all the stacked operations
13472 * of equal or higher precedence are evaluated, leaving the result as the
13473 * top entry on the stack. This makes higher precedence operations
13474 * evaluate before lower precedence ones, and causes operations of equal
13475 * precedence to left associate.
13477 * The only unary operator '!' is immediately pushed onto the stack when
13478 * encountered. When an operand is encountered, if the top of the stack is
13479 * a '!", the complement is immediately performed, and the '!' popped. The
13480 * resulting value is treated as a new operand, and the logic in the
13481 * previous paragraph is executed. Thus in the expression
13483 * the stack looks like
13489 * as 'b' gets parsed, the latter gets evaluated to '!b', and the stack
13496 * A ')' is treated as an operator with lower precedence than all the
13497 * aforementioned ones, which causes all operations on the stack above the
13498 * corresponding '(' to be evaluated down to a single resultant operand.
13499 * Then the fence for the '(' is removed, and the operand goes through the
13500 * algorithm above, without the fence.
13502 * A separate stack is kept of the fence positions, so that the position of
13503 * the latest so-far unbalanced '(' is at the top of it.
13505 * The ']' ending the construct is treated as the lowest operator of all,
13506 * so that everything gets evaluated down to a single operand, which is the
13509 sv_2mortal((SV *)(stack = newAV()));
13510 sv_2mortal((SV *)(fence_stack = newAV()));
13512 while (RExC_parse < RExC_end) {
13513 I32 top_index; /* Index of top-most element in 'stack' */
13514 SV** top_ptr; /* Pointer to top 'stack' element */
13515 SV* current = NULL; /* To contain the current inversion list
13517 SV* only_to_avoid_leaks;
13519 /* Skip white space */
13520 RExC_parse = regpatws(pRExC_state, RExC_parse,
13521 TRUE /* means recognize comments */ );
13522 if (RExC_parse >= RExC_end) {
13523 Perl_croak(aTHX_ "panic: Read past end of '(?[ ])'");
13526 curchar = UCHARAT(RExC_parse);
13530 top_index = av_tindex(stack);
13533 SV** stacked_ptr; /* Ptr to something already on 'stack' */
13534 char stacked_operator; /* The topmost operator on the 'stack'. */
13535 SV* lhs; /* Operand to the left of the operator */
13536 SV* rhs; /* Operand to the right of the operator */
13537 SV* fence_ptr; /* Pointer to top element of the fence
13542 if (RExC_parse < RExC_end && (UCHARAT(RExC_parse + 1) == '?'))
13544 /* If is a '(?', could be an embedded '(?flags:(?[...])'.
13545 * This happens when we have some thing like
13547 * my $thai_or_lao = qr/(?[ \p{Thai} + \p{Lao} ])/;
13549 * qr/(?[ \p{Digit} & $thai_or_lao ])/;
13551 * Here we would be handling the interpolated
13552 * '$thai_or_lao'. We handle this by a recursive call to
13553 * ourselves which returns the inversion list the
13554 * interpolated expression evaluates to. We use the flags
13555 * from the interpolated pattern. */
13556 U32 save_flags = RExC_flags;
13557 const char * save_parse;
13559 RExC_parse += 2; /* Skip past the '(?' */
13560 save_parse = RExC_parse;
13562 /* Parse any flags for the '(?' */
13563 parse_lparen_question_flags(pRExC_state);
13565 if (RExC_parse == save_parse /* Makes sure there was at
13566 least one flag (or else
13567 this embedding wasn't
13569 || RExC_parse >= RExC_end - 4
13570 || UCHARAT(RExC_parse) != ':'
13571 || UCHARAT(++RExC_parse) != '('
13572 || UCHARAT(++RExC_parse) != '?'
13573 || UCHARAT(++RExC_parse) != '[')
13576 /* In combination with the above, this moves the
13577 * pointer to the point just after the first erroneous
13578 * character (or if there are no flags, to where they
13579 * should have been) */
13580 if (RExC_parse >= RExC_end - 4) {
13581 RExC_parse = RExC_end;
13583 else if (RExC_parse != save_parse) {
13584 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13586 vFAIL("Expecting '(?flags:(?[...'");
13589 /* Recurse, with the meat of the embedded expression */
13591 (void) handle_regex_sets(pRExC_state, ¤t, flagp,
13592 depth+1, oregcomp_parse);
13594 /* Here, 'current' contains the embedded expression's
13595 * inversion list, and RExC_parse points to the trailing
13596 * ']'; the next character should be the ')' */
13598 assert(RExC_parse < RExC_end && UCHARAT(RExC_parse) == ')');
13600 /* Then the ')' matching the original '(' handled by this
13601 * case: statement */
13603 assert(RExC_parse < RExC_end && UCHARAT(RExC_parse) == ')');
13606 RExC_flags = save_flags;
13607 goto handle_operand;
13610 /* A regular '('. Look behind for illegal syntax */
13611 if (top_index - fence >= 0) {
13612 /* If the top entry on the stack is an operator, it had
13613 * better be a '!', otherwise the entry below the top
13614 * operand should be an operator */
13615 if ( ! (top_ptr = av_fetch(stack, top_index, FALSE))
13616 || (! IS_OPERAND(*top_ptr) && SvUV(*top_ptr) != '!')
13617 || top_index - fence < 1
13618 || ! (stacked_ptr = av_fetch(stack,
13621 || IS_OPERAND(*stacked_ptr))
13624 vFAIL("Unexpected '(' with no preceding operator");
13628 /* Stack the position of this undealt-with left paren */
13629 fence = top_index + 1;
13630 av_push(fence_stack, newSViv(fence));
13634 /* regclass() can only return RESTART_UTF8 if multi-char
13635 folds are allowed. */
13636 if (!regclass(pRExC_state, flagp,depth+1,
13637 TRUE, /* means parse just the next thing */
13638 FALSE, /* don't allow multi-char folds */
13639 FALSE, /* don't silence non-portable warnings. */
13643 FAIL2("panic: regclass returned NULL to handle_sets, "
13644 "flags=%#"UVxf"", (UV) *flagp);
13647 /* regclass() will return with parsing just the \ sequence,
13648 * leaving the parse pointer at the next thing to parse */
13650 goto handle_operand;
13652 case '[': /* Is a bracketed character class */
13654 bool is_posix_class = could_it_be_a_POSIX_class(pRExC_state);
13656 if (! is_posix_class) {
13660 /* regclass() can only return RESTART_UTF8 if multi-char
13661 folds are allowed. */
13662 if(!regclass(pRExC_state, flagp,depth+1,
13663 is_posix_class, /* parse the whole char class
13664 only if not a posix class */
13665 FALSE, /* don't allow multi-char folds */
13666 FALSE, /* don't silence non-portable warnings. */
13671 FAIL2("panic: regclass returned NULL to handle_sets, "
13672 "flags=%#"UVxf"", (UV) *flagp);
13675 /* function call leaves parse pointing to the ']', except if we
13677 if (is_posix_class) {
13681 goto handle_operand;
13685 if (top_index >= 1) {
13686 goto join_operators;
13689 /* Only a single operand on the stack: are done */
13693 if (av_tindex(fence_stack) < 0) {
13695 vFAIL("Unexpected ')'");
13698 /* If at least two thing on the stack, treat this as an
13700 if (top_index - fence >= 1) {
13701 goto join_operators;
13704 /* Here only a single thing on the fenced stack, and there is a
13705 * fence. Get rid of it */
13706 fence_ptr = av_pop(fence_stack);
13708 fence = SvIV(fence_ptr) - 1;
13709 SvREFCNT_dec_NN(fence_ptr);
13716 /* Having gotten rid of the fence, we pop the operand at the
13717 * stack top and process it as a newly encountered operand */
13718 current = av_pop(stack);
13719 assert(IS_OPERAND(current));
13720 goto handle_operand;
13728 /* These binary operators should have a left operand already
13730 if ( top_index - fence < 0
13731 || top_index - fence == 1
13732 || ( ! (top_ptr = av_fetch(stack, top_index, FALSE)))
13733 || ! IS_OPERAND(*top_ptr))
13735 goto unexpected_binary;
13738 /* If only the one operand is on the part of the stack visible
13739 * to us, we just place this operator in the proper position */
13740 if (top_index - fence < 2) {
13742 /* Place the operator before the operand */
13744 SV* lhs = av_pop(stack);
13745 av_push(stack, newSVuv(curchar));
13746 av_push(stack, lhs);
13750 /* But if there is something else on the stack, we need to
13751 * process it before this new operator if and only if the
13752 * stacked operation has equal or higher precedence than the
13757 /* The operator on the stack is supposed to be below both its
13759 if ( ! (stacked_ptr = av_fetch(stack, top_index - 2, FALSE))
13760 || IS_OPERAND(*stacked_ptr))
13762 /* But if not, it's legal and indicates we are completely
13763 * done if and only if we're currently processing a ']',
13764 * which should be the final thing in the expression */
13765 if (curchar == ']') {
13771 vFAIL2("Unexpected binary operator '%c' with no "
13772 "preceding operand", curchar);
13774 stacked_operator = (char) SvUV(*stacked_ptr);
13776 if (regex_set_precedence(curchar)
13777 > regex_set_precedence(stacked_operator))
13779 /* Here, the new operator has higher precedence than the
13780 * stacked one. This means we need to add the new one to
13781 * the stack to await its rhs operand (and maybe more
13782 * stuff). We put it before the lhs operand, leaving
13783 * untouched the stacked operator and everything below it
13785 lhs = av_pop(stack);
13786 assert(IS_OPERAND(lhs));
13788 av_push(stack, newSVuv(curchar));
13789 av_push(stack, lhs);
13793 /* Here, the new operator has equal or lower precedence than
13794 * what's already there. This means the operation already
13795 * there should be performed now, before the new one. */
13796 rhs = av_pop(stack);
13797 lhs = av_pop(stack);
13799 assert(IS_OPERAND(rhs));
13800 assert(IS_OPERAND(lhs));
13802 switch (stacked_operator) {
13804 _invlist_intersection(lhs, rhs, &rhs);
13809 _invlist_union(lhs, rhs, &rhs);
13813 _invlist_subtract(lhs, rhs, &rhs);
13816 case '^': /* The union minus the intersection */
13822 _invlist_union(lhs, rhs, &u);
13823 _invlist_intersection(lhs, rhs, &i);
13824 /* _invlist_subtract will overwrite rhs
13825 without freeing what it already contains */
13827 _invlist_subtract(u, i, &rhs);
13828 SvREFCNT_dec_NN(i);
13829 SvREFCNT_dec_NN(u);
13830 SvREFCNT_dec_NN(element);
13836 /* Here, the higher precedence operation has been done, and the
13837 * result is in 'rhs'. We overwrite the stacked operator with
13838 * the result. Then we redo this code to either push the new
13839 * operator onto the stack or perform any higher precedence
13840 * stacked operation */
13841 only_to_avoid_leaks = av_pop(stack);
13842 SvREFCNT_dec(only_to_avoid_leaks);
13843 av_push(stack, rhs);
13846 case '!': /* Highest priority, right associative, so just push
13848 av_push(stack, newSVuv(curchar));
13852 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13853 vFAIL("Unexpected character");
13857 /* Here 'current' is the operand. If something is already on the
13858 * stack, we have to check if it is a !. */
13859 top_index = av_tindex(stack); /* Code above may have altered the
13860 * stack in the time since we
13861 * earlier set 'top_index'. */
13862 if (top_index - fence >= 0) {
13863 /* If the top entry on the stack is an operator, it had better
13864 * be a '!', otherwise the entry below the top operand should
13865 * be an operator */
13866 top_ptr = av_fetch(stack, top_index, FALSE);
13868 if (! IS_OPERAND(*top_ptr)) {
13870 /* The only permissible operator at the top of the stack is
13871 * '!', which is applied immediately to this operand. */
13872 curchar = (char) SvUV(*top_ptr);
13873 if (curchar != '!') {
13874 SvREFCNT_dec(current);
13875 vFAIL2("Unexpected binary operator '%c' with no "
13876 "preceding operand", curchar);
13879 _invlist_invert(current);
13881 only_to_avoid_leaks = av_pop(stack);
13882 SvREFCNT_dec(only_to_avoid_leaks);
13883 top_index = av_tindex(stack);
13885 /* And we redo with the inverted operand. This allows
13886 * handling multiple ! in a row */
13887 goto handle_operand;
13889 /* Single operand is ok only for the non-binary ')'
13891 else if ((top_index - fence == 0 && curchar != ')')
13892 || (top_index - fence > 0
13893 && (! (stacked_ptr = av_fetch(stack,
13896 || IS_OPERAND(*stacked_ptr))))
13898 SvREFCNT_dec(current);
13899 vFAIL("Operand with no preceding operator");
13903 /* Here there was nothing on the stack or the top element was
13904 * another operand. Just add this new one */
13905 av_push(stack, current);
13907 } /* End of switch on next parse token */
13909 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
13910 } /* End of loop parsing through the construct */
13913 if (av_tindex(fence_stack) >= 0) {
13914 vFAIL("Unmatched (");
13917 if (av_tindex(stack) < 0 /* Was empty */
13918 || ((final = av_pop(stack)) == NULL)
13919 || ! IS_OPERAND(final)
13920 || av_tindex(stack) >= 0) /* More left on stack */
13922 SvREFCNT_dec(final);
13923 vFAIL("Incomplete expression within '(?[ ])'");
13926 /* Here, 'final' is the resultant inversion list from evaluating the
13927 * expression. Return it if so requested */
13928 if (return_invlist) {
13929 *return_invlist = final;
13933 /* Otherwise generate a resultant node, based on 'final'. regclass() is
13934 * expecting a string of ranges and individual code points */
13935 invlist_iterinit(final);
13936 result_string = newSVpvs("");
13937 while (invlist_iternext(final, &start, &end)) {
13938 if (start == end) {
13939 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}", start);
13942 Perl_sv_catpvf(aTHX_ result_string, "\\x{%"UVXf"}-\\x{%"UVXf"}",
13947 /* About to generate an ANYOF (or similar) node from the inversion list we
13948 * have calculated */
13949 save_parse = RExC_parse;
13950 RExC_parse = SvPV(result_string, len);
13951 save_end = RExC_end;
13952 RExC_end = RExC_parse + len;
13954 /* We turn off folding around the call, as the class we have constructed
13955 * already has all folding taken into consideration, and we don't want
13956 * regclass() to add to that */
13957 RExC_flags &= ~RXf_PMf_FOLD;
13958 /* regclass() can only return RESTART_UTF8 if multi-char folds are allowed.
13960 node = regclass(pRExC_state, flagp,depth+1,
13961 FALSE, /* means parse the whole char class */
13962 FALSE, /* don't allow multi-char folds */
13963 TRUE, /* silence non-portable warnings. The above may very
13964 well have generated non-portable code points, but
13965 they're valid on this machine */
13966 FALSE, /* similarly, no need for strict */
13970 FAIL2("panic: regclass returned NULL to handle_sets, flags=%#"UVxf,
13973 RExC_flags |= RXf_PMf_FOLD;
13975 RExC_parse = save_parse + 1;
13976 RExC_end = save_end;
13977 SvREFCNT_dec_NN(final);
13978 SvREFCNT_dec_NN(result_string);
13980 nextchar(pRExC_state);
13981 Set_Node_Length(node, RExC_parse - oregcomp_parse + 1); /* MJD */
13987 S_add_above_Latin1_folds(pTHX_ RExC_state_t *pRExC_state, const U8 cp, SV** invlist)
13989 /* This hard-codes the Latin1/above-Latin1 folding rules, so that an
13990 * innocent-looking character class, like /[ks]/i won't have to go out to
13991 * disk to find the possible matches.
13993 * This should be called only for a Latin1-range code points, cp, which is
13994 * known to be involved in a simple fold with other code points above
13995 * Latin1. It would give false results if /aa has been specified.
13996 * Multi-char folds are outside the scope of this, and must be handled
13999 * XXX It would be better to generate these via regen, in case a new
14000 * version of the Unicode standard adds new mappings, though that is not
14001 * really likely, and may be caught by the default: case of the switch
14004 PERL_ARGS_ASSERT_ADD_ABOVE_LATIN1_FOLDS;
14006 assert(HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(cp));
14012 add_cp_to_invlist(*invlist, KELVIN_SIGN);
14016 *invlist = add_cp_to_invlist(*invlist, LATIN_SMALL_LETTER_LONG_S);
14019 *invlist = add_cp_to_invlist(*invlist, GREEK_CAPITAL_LETTER_MU);
14020 *invlist = add_cp_to_invlist(*invlist, GREEK_SMALL_LETTER_MU);
14022 case LATIN_CAPITAL_LETTER_A_WITH_RING_ABOVE:
14023 case LATIN_SMALL_LETTER_A_WITH_RING_ABOVE:
14024 *invlist = add_cp_to_invlist(*invlist, ANGSTROM_SIGN);
14026 case LATIN_SMALL_LETTER_Y_WITH_DIAERESIS:
14027 *invlist = add_cp_to_invlist(*invlist,
14028 LATIN_CAPITAL_LETTER_Y_WITH_DIAERESIS);
14030 case LATIN_SMALL_LETTER_SHARP_S:
14031 *invlist = add_cp_to_invlist(*invlist, LATIN_CAPITAL_LETTER_SHARP_S);
14034 /* Use deprecated warning to increase the chances of this being
14037 ckWARN2reg_d(RExC_parse, "Perl folding rules are not up-to-date for 0x%02X; please use the perlbug utility to report;", cp);
14044 S_add_multi_match(pTHX_ AV* multi_char_matches, SV* multi_string, const STRLEN cp_count)
14046 /* This adds the string scalar <multi_string> to the array
14047 * <multi_char_matches>. <multi_string> is known to have exactly
14048 * <cp_count> code points in it. This is used when constructing a
14049 * bracketed character class and we find something that needs to match more
14050 * than a single character.
14052 * <multi_char_matches> is actually an array of arrays. Each top-level
14053 * element is an array that contains all the strings known so far that are
14054 * the same length. And that length (in number of code points) is the same
14055 * as the index of the top-level array. Hence, the [2] element is an
14056 * array, each element thereof is a string containing TWO code points;
14057 * while element [3] is for strings of THREE characters, and so on. Since
14058 * this is for multi-char strings there can never be a [0] nor [1] element.
14060 * When we rewrite the character class below, we will do so such that the
14061 * longest strings are written first, so that it prefers the longest
14062 * matching strings first. This is done even if it turns out that any
14063 * quantifier is non-greedy, out of this programmer's (khw) laziness. Tom
14064 * Christiansen has agreed that this is ok. This makes the test for the
14065 * ligature 'ffi' come before the test for 'ff', for example */
14068 AV** this_array_ptr;
14070 PERL_ARGS_ASSERT_ADD_MULTI_MATCH;
14072 if (! multi_char_matches) {
14073 multi_char_matches = newAV();
14076 if (av_exists(multi_char_matches, cp_count)) {
14077 this_array_ptr = (AV**) av_fetch(multi_char_matches, cp_count, FALSE);
14078 this_array = *this_array_ptr;
14081 this_array = newAV();
14082 av_store(multi_char_matches, cp_count,
14085 av_push(this_array, multi_string);
14087 return multi_char_matches;
14090 /* The names of properties whose definitions are not known at compile time are
14091 * stored in this SV, after a constant heading. So if the length has been
14092 * changed since initialization, then there is a run-time definition. */
14093 #define HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION \
14094 (SvCUR(listsv) != initial_listsv_len)
14097 S_regclass(pTHX_ RExC_state_t *pRExC_state, I32 *flagp, U32 depth,
14098 const bool stop_at_1, /* Just parse the next thing, don't
14099 look for a full character class */
14100 bool allow_multi_folds,
14101 const bool silence_non_portable, /* Don't output warnings
14105 SV** ret_invlist /* Return an inversion list, not a node */
14108 /* parse a bracketed class specification. Most of these will produce an
14109 * ANYOF node; but something like [a] will produce an EXACT node; [aA], an
14110 * EXACTFish node; [[:ascii:]], a POSIXA node; etc. It is more complex
14111 * under /i with multi-character folds: it will be rewritten following the
14112 * paradigm of this example, where the <multi-fold>s are characters which
14113 * fold to multiple character sequences:
14114 * /[abc\x{multi-fold1}def\x{multi-fold2}ghi]/i
14115 * gets effectively rewritten as:
14116 * /(?:\x{multi-fold1}|\x{multi-fold2}|[abcdefghi]/i
14117 * reg() gets called (recursively) on the rewritten version, and this
14118 * function will return what it constructs. (Actually the <multi-fold>s
14119 * aren't physically removed from the [abcdefghi], it's just that they are
14120 * ignored in the recursion by means of a flag:
14121 * <RExC_in_multi_char_class>.)
14123 * ANYOF nodes contain a bit map for the first NUM_ANYOF_CODE_POINTS
14124 * characters, with the corresponding bit set if that character is in the
14125 * list. For characters above this, a range list or swash is used. There
14126 * are extra bits for \w, etc. in locale ANYOFs, as what these match is not
14127 * determinable at compile time
14129 * Returns NULL, setting *flagp to RESTART_UTF8 if the sizing scan needs
14130 * to be restarted. This can only happen if ret_invlist is non-NULL.
14133 UV prevvalue = OOB_UNICODE, save_prevvalue = OOB_UNICODE;
14135 UV value = OOB_UNICODE, save_value = OOB_UNICODE;
14138 IV namedclass = OOB_NAMEDCLASS;
14139 char *rangebegin = NULL;
14140 bool need_class = 0;
14142 STRLEN initial_listsv_len = 0; /* Kind of a kludge to see if it is more
14143 than just initialized. */
14144 SV* properties = NULL; /* Code points that match \p{} \P{} */
14145 SV* posixes = NULL; /* Code points that match classes like [:word:],
14146 extended beyond the Latin1 range. These have to
14147 be kept separate from other code points for much
14148 of this function because their handling is
14149 different under /i, and for most classes under
14151 SV* nposixes = NULL; /* Similarly for [:^word:]. These are kept
14152 separate for a while from the non-complemented
14153 versions because of complications with /d
14155 SV* simple_posixes = NULL; /* But under some conditions, the classes can be
14156 treated more simply than the general case,
14157 leading to less compilation and execution
14159 UV element_count = 0; /* Number of distinct elements in the class.
14160 Optimizations may be possible if this is tiny */
14161 AV * multi_char_matches = NULL; /* Code points that fold to more than one
14162 character; used under /i */
14164 char * stop_ptr = RExC_end; /* where to stop parsing */
14165 const bool skip_white = cBOOL(ret_invlist); /* ignore unescaped white
14168 /* Unicode properties are stored in a swash; this holds the current one
14169 * being parsed. If this swash is the only above-latin1 component of the
14170 * character class, an optimization is to pass it directly on to the
14171 * execution engine. Otherwise, it is set to NULL to indicate that there
14172 * are other things in the class that have to be dealt with at execution
14174 SV* swash = NULL; /* Code points that match \p{} \P{} */
14176 /* Set if a component of this character class is user-defined; just passed
14177 * on to the engine */
14178 bool has_user_defined_property = FALSE;
14180 /* inversion list of code points this node matches only when the target
14181 * string is in UTF-8. (Because is under /d) */
14182 SV* depends_list = NULL;
14184 /* Inversion list of code points this node matches regardless of things
14185 * like locale, folding, utf8ness of the target string */
14186 SV* cp_list = NULL;
14188 /* Like cp_list, but code points on this list need to be checked for things
14189 * that fold to/from them under /i */
14190 SV* cp_foldable_list = NULL;
14192 /* Like cp_list, but code points on this list are valid only when the
14193 * runtime locale is UTF-8 */
14194 SV* only_utf8_locale_list = NULL;
14196 /* In a range, if one of the endpoints is non-character-set portable,
14197 * meaning that it hard-codes a code point that may mean a different
14198 * charactger in ASCII vs. EBCDIC, as opposed to, say, a literal 'A' or a
14199 * mnemonic '\t' which each mean the same character no matter which
14200 * character set the platform is on. */
14201 unsigned int non_portable_endpoint = 0;
14203 /* Is the range unicode? which means on a platform that isn't 1-1 native
14204 * to Unicode (i.e. non-ASCII), each code point in it should be considered
14205 * to be a Unicode value. */
14206 bool unicode_range = FALSE;
14207 bool invert = FALSE; /* Is this class to be complemented */
14209 bool warn_super = ALWAYS_WARN_SUPER;
14211 regnode * const orig_emit = RExC_emit; /* Save the original RExC_emit in
14212 case we need to change the emitted regop to an EXACT. */
14213 const char * orig_parse = RExC_parse;
14214 const SSize_t orig_size = RExC_size;
14215 bool posixl_matches_all = FALSE; /* Does /l class have both e.g. \W,\w ? */
14216 GET_RE_DEBUG_FLAGS_DECL;
14218 PERL_ARGS_ASSERT_REGCLASS;
14220 PERL_UNUSED_ARG(depth);
14223 DEBUG_PARSE("clas");
14225 /* Assume we are going to generate an ANYOF node. */
14226 ret = reganode(pRExC_state,
14233 RExC_size += ANYOF_SKIP;
14234 listsv = &PL_sv_undef; /* For code scanners: listsv always non-NULL. */
14237 ANYOF_FLAGS(ret) = 0;
14239 RExC_emit += ANYOF_SKIP;
14240 listsv = newSVpvs_flags("# comment\n", SVs_TEMP);
14241 initial_listsv_len = SvCUR(listsv);
14242 SvTEMP_off(listsv); /* Grr, TEMPs and mortals are conflated. */
14246 RExC_parse = regpatws(pRExC_state, RExC_parse,
14247 FALSE /* means don't recognize comments */ );
14250 if (UCHARAT(RExC_parse) == '^') { /* Complement of range. */
14253 allow_multi_folds = FALSE;
14256 RExC_parse = regpatws(pRExC_state, RExC_parse,
14257 FALSE /* means don't recognize comments */ );
14261 /* Check that they didn't say [:posix:] instead of [[:posix:]] */
14262 if (!SIZE_ONLY && RExC_parse < RExC_end && POSIXCC(UCHARAT(RExC_parse))) {
14263 const char *s = RExC_parse;
14264 const char c = *s++;
14269 while (isWORDCHAR(*s))
14271 if (*s && c == *s && s[1] == ']') {
14272 SAVEFREESV(RExC_rx_sv);
14274 "POSIX syntax [%c %c] belongs inside character classes",
14276 (void)ReREFCNT_inc(RExC_rx_sv);
14280 /* If the caller wants us to just parse a single element, accomplish this
14281 * by faking the loop ending condition */
14282 if (stop_at_1 && RExC_end > RExC_parse) {
14283 stop_ptr = RExC_parse + 1;
14286 /* allow 1st char to be ']' (allowing it to be '-' is dealt with later) */
14287 if (UCHARAT(RExC_parse) == ']')
14288 goto charclassloop;
14291 if (RExC_parse >= stop_ptr) {
14296 RExC_parse = regpatws(pRExC_state, RExC_parse,
14297 FALSE /* means don't recognize comments */ );
14300 if (UCHARAT(RExC_parse) == ']') {
14306 namedclass = OOB_NAMEDCLASS; /* initialize as illegal */
14307 save_value = value;
14308 save_prevvalue = prevvalue;
14311 rangebegin = RExC_parse;
14313 non_portable_endpoint = 0;
14316 value = utf8n_to_uvchr((U8*)RExC_parse,
14317 RExC_end - RExC_parse,
14318 &numlen, UTF8_ALLOW_DEFAULT);
14319 RExC_parse += numlen;
14322 value = UCHARAT(RExC_parse++);
14325 && RExC_parse < RExC_end
14326 && POSIXCC(UCHARAT(RExC_parse)))
14328 namedclass = regpposixcc(pRExC_state, value, strict);
14330 else if (value == '\\') {
14331 /* Is a backslash; get the code point of the char after it */
14332 if (UTF && ! UTF8_IS_INVARIANT(UCHARAT(RExC_parse))) {
14333 value = utf8n_to_uvchr((U8*)RExC_parse,
14334 RExC_end - RExC_parse,
14335 &numlen, UTF8_ALLOW_DEFAULT);
14336 RExC_parse += numlen;
14339 value = UCHARAT(RExC_parse++);
14341 /* Some compilers cannot handle switching on 64-bit integer
14342 * values, therefore value cannot be an UV. Yes, this will
14343 * be a problem later if we want switch on Unicode.
14344 * A similar issue a little bit later when switching on
14345 * namedclass. --jhi */
14347 /* If the \ is escaping white space when white space is being
14348 * skipped, it means that that white space is wanted literally, and
14349 * is already in 'value'. Otherwise, need to translate the escape
14350 * into what it signifies. */
14351 if (! skip_white || ! is_PATWS_cp(value)) switch ((I32)value) {
14353 case 'w': namedclass = ANYOF_WORDCHAR; break;
14354 case 'W': namedclass = ANYOF_NWORDCHAR; break;
14355 case 's': namedclass = ANYOF_SPACE; break;
14356 case 'S': namedclass = ANYOF_NSPACE; break;
14357 case 'd': namedclass = ANYOF_DIGIT; break;
14358 case 'D': namedclass = ANYOF_NDIGIT; break;
14359 case 'v': namedclass = ANYOF_VERTWS; break;
14360 case 'V': namedclass = ANYOF_NVERTWS; break;
14361 case 'h': namedclass = ANYOF_HORIZWS; break;
14362 case 'H': namedclass = ANYOF_NHORIZWS; break;
14363 case 'N': /* Handle \N{NAME} in class */
14365 const char * const backslash_N_beg = RExC_parse - 2;
14368 if (! grok_bslash_N(pRExC_state,
14369 NULL, /* No regnode */
14370 &value, /* Yes single value */
14371 &cp_count, /* Multiple code pt count */
14376 if (*flagp & RESTART_UTF8)
14377 FAIL("panic: grok_bslash_N set RESTART_UTF8");
14379 if (cp_count < 0) {
14380 vFAIL("\\N in a character class must be a named character: \\N{...}");
14382 else if (cp_count == 0) {
14384 RExC_parse++; /* Position after the "}" */
14385 vFAIL("Zero length \\N{}");
14388 ckWARNreg(RExC_parse,
14389 "Ignoring zero length \\N{} in character class");
14392 else { /* cp_count > 1 */
14393 if (! RExC_in_multi_char_class) {
14394 if (invert || range || *RExC_parse == '-') {
14397 vFAIL("\\N{} in inverted character class or as a range end-point is restricted to one character");
14400 ckWARNreg(RExC_parse, "Using just the first character returned by \\N{} in character class");
14402 break; /* <value> contains the first code
14403 point. Drop out of the switch to
14407 SV * multi_char_N = newSVpvn(backslash_N_beg,
14408 RExC_parse - backslash_N_beg);
14410 = add_multi_match(multi_char_matches,
14415 } /* End of cp_count != 1 */
14417 /* This element should not be processed further in this
14420 value = save_value;
14421 prevvalue = save_prevvalue;
14422 continue; /* Back to top of loop to get next char */
14425 /* Here, is a single code point, and <value> contains it */
14426 unicode_range = TRUE; /* \N{} are Unicode */
14434 /* We will handle any undefined properties ourselves */
14435 U8 swash_init_flags = _CORE_SWASH_INIT_RETURN_IF_UNDEF
14436 /* And we actually would prefer to get
14437 * the straight inversion list of the
14438 * swash, since we will be accessing it
14439 * anyway, to save a little time */
14440 |_CORE_SWASH_INIT_ACCEPT_INVLIST;
14442 if (RExC_parse >= RExC_end)
14443 vFAIL2("Empty \\%c{}", (U8)value);
14444 if (*RExC_parse == '{') {
14445 const U8 c = (U8)value;
14446 e = strchr(RExC_parse++, '}');
14448 vFAIL2("Missing right brace on \\%c{}", c);
14449 while (isSPACE(*RExC_parse))
14451 if (e == RExC_parse)
14452 vFAIL2("Empty \\%c{}", c);
14453 n = e - RExC_parse;
14454 while (isSPACE(*(RExC_parse + n - 1)))
14465 if (UCHARAT(RExC_parse) == '^') {
14468 /* toggle. (The rhs xor gets the single bit that
14469 * differs between P and p; the other xor inverts just
14471 value ^= 'P' ^ 'p';
14473 while (isSPACE(*RExC_parse)) {
14478 /* Try to get the definition of the property into
14479 * <invlist>. If /i is in effect, the effective property
14480 * will have its name be <__NAME_i>. The design is
14481 * discussed in commit
14482 * 2f833f5208e26b208886e51e09e2c072b5eabb46 */
14483 name = savepv(Perl_form(aTHX_
14485 (FOLD) ? "__" : "",
14491 /* Look up the property name, and get its swash and
14492 * inversion list, if the property is found */
14494 SvREFCNT_dec_NN(swash);
14496 swash = _core_swash_init("utf8", name, &PL_sv_undef,
14499 NULL, /* No inversion list */
14502 if (! swash || ! (invlist = _get_swash_invlist(swash))) {
14503 HV* curpkg = (IN_PERL_COMPILETIME)
14505 : CopSTASH(PL_curcop);
14507 SvREFCNT_dec_NN(swash);
14511 /* Here didn't find it. It could be a user-defined
14512 * property that will be available at run-time. If we
14513 * accept only compile-time properties, is an error;
14514 * otherwise add it to the list for run-time look up */
14516 RExC_parse = e + 1;
14518 "Property '%"UTF8f"' is unknown",
14519 UTF8fARG(UTF, n, name));
14522 /* If the property name doesn't already have a package
14523 * name, add the current one to it so that it can be
14524 * referred to outside it. [perl #121777] */
14525 if (curpkg && ! instr(name, "::")) {
14526 char* pkgname = HvNAME(curpkg);
14527 if (strNE(pkgname, "main")) {
14528 char* full_name = Perl_form(aTHX_
14532 n = strlen(full_name);
14534 name = savepvn(full_name, n);
14537 Perl_sv_catpvf(aTHX_ listsv, "%cutf8::%"UTF8f"\n",
14538 (value == 'p' ? '+' : '!'),
14539 UTF8fARG(UTF, n, name));
14540 has_user_defined_property = TRUE;
14542 /* We don't know yet, so have to assume that the
14543 * property could match something in the Latin1 range,
14544 * hence something that isn't utf8. Note that this
14545 * would cause things in <depends_list> to match
14546 * inappropriately, except that any \p{}, including
14547 * this one forces Unicode semantics, which means there
14548 * is no <depends_list> */
14550 |= ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES;
14554 /* Here, did get the swash and its inversion list. If
14555 * the swash is from a user-defined property, then this
14556 * whole character class should be regarded as such */
14557 if (swash_init_flags
14558 & _CORE_SWASH_INIT_USER_DEFINED_PROPERTY)
14560 has_user_defined_property = TRUE;
14563 /* We warn on matching an above-Unicode code point
14564 * if the match would return true, except don't
14565 * warn for \p{All}, which has exactly one element
14567 (_invlist_contains_cp(invlist, 0x110000)
14568 && (! (_invlist_len(invlist) == 1
14569 && *invlist_array(invlist) == 0)))
14575 /* Invert if asking for the complement */
14576 if (value == 'P') {
14577 _invlist_union_complement_2nd(properties,
14581 /* The swash can't be used as-is, because we've
14582 * inverted things; delay removing it to here after
14583 * have copied its invlist above */
14584 SvREFCNT_dec_NN(swash);
14588 _invlist_union(properties, invlist, &properties);
14593 RExC_parse = e + 1;
14594 namedclass = ANYOF_UNIPROP; /* no official name, but it's
14597 /* \p means they want Unicode semantics */
14598 RExC_uni_semantics = 1;
14601 case 'n': value = '\n'; break;
14602 case 'r': value = '\r'; break;
14603 case 't': value = '\t'; break;
14604 case 'f': value = '\f'; break;
14605 case 'b': value = '\b'; break;
14606 case 'e': value = ESC_NATIVE; break;
14607 case 'a': value = '\a'; break;
14609 RExC_parse--; /* function expects to be pointed at the 'o' */
14611 const char* error_msg;
14612 bool valid = grok_bslash_o(&RExC_parse,
14615 PASS2, /* warnings only in
14618 silence_non_portable,
14624 non_portable_endpoint++;
14625 if (IN_ENCODING && value < 0x100) {
14626 goto recode_encoding;
14630 RExC_parse--; /* function expects to be pointed at the 'x' */
14632 const char* error_msg;
14633 bool valid = grok_bslash_x(&RExC_parse,
14636 PASS2, /* Output warnings */
14638 silence_non_portable,
14644 non_portable_endpoint++;
14645 if (IN_ENCODING && value < 0x100)
14646 goto recode_encoding;
14649 value = grok_bslash_c(*RExC_parse++, PASS2);
14650 non_portable_endpoint++;
14652 case '0': case '1': case '2': case '3': case '4':
14653 case '5': case '6': case '7':
14655 /* Take 1-3 octal digits */
14656 I32 flags = PERL_SCAN_SILENT_ILLDIGIT;
14657 numlen = (strict) ? 4 : 3;
14658 value = grok_oct(--RExC_parse, &numlen, &flags, NULL);
14659 RExC_parse += numlen;
14662 RExC_parse += (UTF) ? UTF8SKIP(RExC_parse) : 1;
14663 vFAIL("Need exactly 3 octal digits");
14665 else if (! SIZE_ONLY /* like \08, \178 */
14667 && RExC_parse < RExC_end
14668 && isDIGIT(*RExC_parse)
14669 && ckWARN(WARN_REGEXP))
14671 SAVEFREESV(RExC_rx_sv);
14672 reg_warn_non_literal_string(
14674 form_short_octal_warning(RExC_parse, numlen));
14675 (void)ReREFCNT_inc(RExC_rx_sv);
14678 non_portable_endpoint++;
14679 if (IN_ENCODING && value < 0x100)
14680 goto recode_encoding;
14684 if (! RExC_override_recoding) {
14685 SV* enc = _get_encoding();
14686 value = reg_recode((const char)(U8)value, &enc);
14689 vFAIL("Invalid escape in the specified encoding");
14692 ckWARNreg(RExC_parse,
14693 "Invalid escape in the specified encoding");
14699 /* Allow \_ to not give an error */
14700 if (!SIZE_ONLY && isWORDCHAR(value) && value != '_') {
14702 vFAIL2("Unrecognized escape \\%c in character class",
14706 SAVEFREESV(RExC_rx_sv);
14707 ckWARN2reg(RExC_parse,
14708 "Unrecognized escape \\%c in character class passed through",
14710 (void)ReREFCNT_inc(RExC_rx_sv);
14714 } /* End of switch on char following backslash */
14715 } /* end of handling backslash escape sequences */
14717 /* Here, we have the current token in 'value' */
14719 if (namedclass > OOB_NAMEDCLASS) { /* this is a named class \blah */
14722 /* a bad range like a-\d, a-[:digit:]. The '-' is taken as a
14723 * literal, as is the character that began the false range, i.e.
14724 * the 'a' in the examples */
14727 const int w = (RExC_parse >= rangebegin)
14728 ? RExC_parse - rangebegin
14732 "False [] range \"%"UTF8f"\"",
14733 UTF8fARG(UTF, w, rangebegin));
14736 SAVEFREESV(RExC_rx_sv); /* in case of fatal warnings */
14737 ckWARN2reg(RExC_parse,
14738 "False [] range \"%"UTF8f"\"",
14739 UTF8fARG(UTF, w, rangebegin));
14740 (void)ReREFCNT_inc(RExC_rx_sv);
14741 cp_list = add_cp_to_invlist(cp_list, '-');
14742 cp_foldable_list = add_cp_to_invlist(cp_foldable_list,
14747 range = 0; /* this was not a true range */
14748 element_count += 2; /* So counts for three values */
14751 classnum = namedclass_to_classnum(namedclass);
14753 if (LOC && namedclass < ANYOF_POSIXL_MAX
14754 #ifndef HAS_ISASCII
14755 && classnum != _CC_ASCII
14758 /* What the Posix classes (like \w, [:space:]) match in locale
14759 * isn't knowable under locale until actual match time. Room
14760 * must be reserved (one time per outer bracketed class) to
14761 * store such classes. The space will contain a bit for each
14762 * named class that is to be matched against. This isn't
14763 * needed for \p{} and pseudo-classes, as they are not affected
14764 * by locale, and hence are dealt with separately */
14765 if (! need_class) {
14768 RExC_size += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14771 RExC_emit += ANYOF_POSIXL_SKIP - ANYOF_SKIP;
14773 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_POSIXL;
14774 ANYOF_POSIXL_ZERO(ret);
14777 /* Coverity thinks it is possible for this to be negative; both
14778 * jhi and khw think it's not, but be safer */
14779 assert(! (ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14780 || (namedclass + ((namedclass % 2) ? -1 : 1)) >= 0);
14782 /* See if it already matches the complement of this POSIX
14784 if ((ANYOF_FLAGS(ret) & ANYOF_MATCHES_POSIXL)
14785 && ANYOF_POSIXL_TEST(ret, namedclass + ((namedclass % 2)
14789 posixl_matches_all = TRUE;
14790 break; /* No need to continue. Since it matches both
14791 e.g., \w and \W, it matches everything, and the
14792 bracketed class can be optimized into qr/./s */
14795 /* Add this class to those that should be checked at runtime */
14796 ANYOF_POSIXL_SET(ret, namedclass);
14798 /* The above-Latin1 characters are not subject to locale rules.
14799 * Just add them, in the second pass, to the
14800 * unconditionally-matched list */
14802 SV* scratch_list = NULL;
14804 /* Get the list of the above-Latin1 code points this
14806 _invlist_intersection_maybe_complement_2nd(PL_AboveLatin1,
14807 PL_XPosix_ptrs[classnum],
14809 /* Odd numbers are complements, like
14810 * NDIGIT, NASCII, ... */
14811 namedclass % 2 != 0,
14813 /* Checking if 'cp_list' is NULL first saves an extra
14814 * clone. Its reference count will be decremented at the
14815 * next union, etc, or if this is the only instance, at the
14816 * end of the routine */
14818 cp_list = scratch_list;
14821 _invlist_union(cp_list, scratch_list, &cp_list);
14822 SvREFCNT_dec_NN(scratch_list);
14824 continue; /* Go get next character */
14827 else if (! SIZE_ONLY) {
14829 /* Here, not in pass1 (in that pass we skip calculating the
14830 * contents of this class), and is /l, or is a POSIX class for
14831 * which /l doesn't matter (or is a Unicode property, which is
14832 * skipped here). */
14833 if (namedclass >= ANYOF_POSIXL_MAX) { /* If a special class */
14834 if (namedclass != ANYOF_UNIPROP) { /* UNIPROP = \p and \P */
14836 /* Here, should be \h, \H, \v, or \V. None of /d, /i
14837 * nor /l make a difference in what these match,
14838 * therefore we just add what they match to cp_list. */
14839 if (classnum != _CC_VERTSPACE) {
14840 assert( namedclass == ANYOF_HORIZWS
14841 || namedclass == ANYOF_NHORIZWS);
14843 /* It turns out that \h is just a synonym for
14845 classnum = _CC_BLANK;
14848 _invlist_union_maybe_complement_2nd(
14850 PL_XPosix_ptrs[classnum],
14851 namedclass % 2 != 0, /* Complement if odd
14852 (NHORIZWS, NVERTWS)
14857 else if (UNI_SEMANTICS
14858 || classnum == _CC_ASCII
14859 || (DEPENDS_SEMANTICS && (classnum == _CC_DIGIT
14860 || classnum == _CC_XDIGIT)))
14862 /* We usually have to worry about /d and /a affecting what
14863 * POSIX classes match, with special code needed for /d
14864 * because we won't know until runtime what all matches.
14865 * But there is no extra work needed under /u, and
14866 * [:ascii:] is unaffected by /a and /d; and :digit: and
14867 * :xdigit: don't have runtime differences under /d. So we
14868 * can special case these, and avoid some extra work below,
14869 * and at runtime. */
14870 _invlist_union_maybe_complement_2nd(
14872 PL_XPosix_ptrs[classnum],
14873 namedclass % 2 != 0,
14876 else { /* Garden variety class. If is NUPPER, NALPHA, ...
14877 complement and use nposixes */
14878 SV** posixes_ptr = namedclass % 2 == 0
14881 _invlist_union_maybe_complement_2nd(
14883 PL_XPosix_ptrs[classnum],
14884 namedclass % 2 != 0,
14888 } /* end of namedclass \blah */
14891 RExC_parse = regpatws(pRExC_state, RExC_parse,
14892 FALSE /* means don't recognize comments */ );
14895 /* If 'range' is set, 'value' is the ending of a range--check its
14896 * validity. (If value isn't a single code point in the case of a
14897 * range, we should have figured that out above in the code that
14898 * catches false ranges). Later, we will handle each individual code
14899 * point in the range. If 'range' isn't set, this could be the
14900 * beginning of a range, so check for that by looking ahead to see if
14901 * the next real character to be processed is the range indicator--the
14906 /* For unicode ranges, we have to test that the Unicode as opposed
14907 * to the native values are not decreasing. (Above 255, there is
14908 * no difference between native and Unicode) */
14909 if (unicode_range && prevvalue < 255 && value < 255) {
14910 if (NATIVE_TO_LATIN1(prevvalue) > NATIVE_TO_LATIN1(value)) {
14911 goto backwards_range;
14916 if (prevvalue > value) /* b-a */ {
14921 w = RExC_parse - rangebegin;
14923 "Invalid [] range \"%"UTF8f"\"",
14924 UTF8fARG(UTF, w, rangebegin));
14925 NOT_REACHED; /* NOTREACHED */
14929 prevvalue = value; /* save the beginning of the potential range */
14930 if (! stop_at_1 /* Can't be a range if parsing just one thing */
14931 && *RExC_parse == '-')
14933 char* next_char_ptr = RExC_parse + 1;
14934 if (skip_white) { /* Get the next real char after the '-' */
14935 next_char_ptr = regpatws(pRExC_state,
14937 FALSE); /* means don't recognize
14941 /* If the '-' is at the end of the class (just before the ']',
14942 * it is a literal minus; otherwise it is a range */
14943 if (next_char_ptr < RExC_end && *next_char_ptr != ']') {
14944 RExC_parse = next_char_ptr;
14946 /* a bad range like \w-, [:word:]- ? */
14947 if (namedclass > OOB_NAMEDCLASS) {
14948 if (strict || (PASS2 && ckWARN(WARN_REGEXP))) {
14949 const int w = RExC_parse >= rangebegin
14950 ? RExC_parse - rangebegin
14953 vFAIL4("False [] range \"%*.*s\"",
14958 "False [] range \"%*.*s\"",
14963 cp_list = add_cp_to_invlist(cp_list, '-');
14967 range = 1; /* yeah, it's a range! */
14968 continue; /* but do it the next time */
14973 if (namedclass > OOB_NAMEDCLASS) {
14977 /* Here, we have a single value this time through the loop, and
14978 * <prevvalue> is the beginning of the range, if any; or <value> if
14981 /* non-Latin1 code point implies unicode semantics. Must be set in
14982 * pass1 so is there for the whole of pass 2 */
14984 RExC_uni_semantics = 1;
14987 /* Ready to process either the single value, or the completed range.
14988 * For single-valued non-inverted ranges, we consider the possibility
14989 * of multi-char folds. (We made a conscious decision to not do this
14990 * for the other cases because it can often lead to non-intuitive
14991 * results. For example, you have the peculiar case that:
14992 * "s s" =~ /^[^\xDF]+$/i => Y
14993 * "ss" =~ /^[^\xDF]+$/i => N
14995 * See [perl #89750] */
14996 if (FOLD && allow_multi_folds && value == prevvalue) {
14997 if (value == LATIN_SMALL_LETTER_SHARP_S
14998 || (value > 255 && _invlist_contains_cp(PL_HasMultiCharFold,
15001 /* Here <value> is indeed a multi-char fold. Get what it is */
15003 U8 foldbuf[UTF8_MAXBYTES_CASE];
15006 UV folded = _to_uni_fold_flags(
15010 FOLD_FLAGS_FULL | (ASCII_FOLD_RESTRICTED
15011 ? FOLD_FLAGS_NOMIX_ASCII
15015 /* Here, <folded> should be the first character of the
15016 * multi-char fold of <value>, with <foldbuf> containing the
15017 * whole thing. But, if this fold is not allowed (because of
15018 * the flags), <fold> will be the same as <value>, and should
15019 * be processed like any other character, so skip the special
15021 if (folded != value) {
15023 /* Skip if we are recursed, currently parsing the class
15024 * again. Otherwise add this character to the list of
15025 * multi-char folds. */
15026 if (! RExC_in_multi_char_class) {
15027 STRLEN cp_count = utf8_length(foldbuf,
15028 foldbuf + foldlen);
15029 SV* multi_fold = sv_2mortal(newSVpvs(""));
15031 Perl_sv_catpvf(aTHX_ multi_fold, "\\x{%"UVXf"}", value);
15034 = add_multi_match(multi_char_matches,
15040 /* This element should not be processed further in this
15043 value = save_value;
15044 prevvalue = save_prevvalue;
15050 if (strict && PASS2 && ckWARN(WARN_REGEXP)) {
15053 /* If the range starts above 255, everything is portable and
15054 * likely to be so for any forseeable character set, so don't
15056 if (unicode_range && non_portable_endpoint && prevvalue < 256) {
15057 vWARN(RExC_parse, "Both or neither range ends should be Unicode");
15059 else if (prevvalue != value) {
15061 /* Under strict, ranges that stop and/or end in an ASCII
15062 * printable should have each end point be a portable value
15063 * for it (preferably like 'A', but we don't warn if it is
15064 * a (portable) Unicode name or code point), and the range
15065 * must be be all digits or all letters of the same case.
15066 * Otherwise, the range is non-portable and unclear as to
15067 * what it contains */
15068 if ((isPRINT_A(prevvalue) || isPRINT_A(value))
15069 && (non_portable_endpoint
15070 || ! ((isDIGIT_A(prevvalue) && isDIGIT_A(value))
15071 || (isLOWER_A(prevvalue) && isLOWER_A(value))
15072 || (isUPPER_A(prevvalue) && isUPPER_A(value)))))
15074 vWARN(RExC_parse, "Ranges of ASCII printables should be some subset of \"0-9\", \"A-Z\", or \"a-z\"");
15076 else if (prevvalue >= 0x660) { /* ARABIC_INDIC_DIGIT_ZERO */
15078 /* But the nature of Unicode and languages mean we
15079 * can't do the same checks for above-ASCII ranges,
15080 * except in the case of digit ones. These should
15081 * contain only digits from the same group of 10. The
15082 * ASCII case is handled just above. 0x660 is the
15083 * first digit character beyond ASCII. Hence here, the
15084 * range could be a range of digits. Find out. */
15085 IV index_start = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
15087 IV index_final = _invlist_search(PL_XPosix_ptrs[_CC_DIGIT],
15090 /* If the range start and final points are in the same
15091 * inversion list element, it means that either both
15092 * are not digits, or both are digits in a consecutive
15093 * sequence of digits. (So far, Unicode has kept all
15094 * such sequences as distinct groups of 10, but assert
15095 * to make sure). If the end points are not in the
15096 * same element, neither should be a digit. */
15097 if (index_start == index_final) {
15098 assert(! ELEMENT_RANGE_MATCHES_INVLIST(index_start)
15099 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
15100 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15102 /* But actually Unicode did have one group of 11
15103 * 'digits' in 5.2, so in case we are operating
15104 * on that version, let that pass */
15105 || (invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start+1]
15106 - invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15108 && invlist_array(PL_XPosix_ptrs[_CC_DIGIT])[index_start]
15112 else if ((index_start >= 0
15113 && ELEMENT_RANGE_MATCHES_INVLIST(index_start))
15114 || (index_final >= 0
15115 && ELEMENT_RANGE_MATCHES_INVLIST(index_final)))
15117 vWARN(RExC_parse, "Ranges of digits should be from the same group of 10");
15122 if ((! range || prevvalue == value) && non_portable_endpoint) {
15123 if (isPRINT_A(value)) {
15126 if (isBACKSLASHED_PUNCT(value)) {
15127 literal[d++] = '\\';
15129 literal[d++] = (char) value;
15130 literal[d++] = '\0';
15133 "\"%.*s\" is more clearly written simply as \"%s\"",
15134 (int) (RExC_parse - rangebegin),
15139 else if isMNEMONIC_CNTRL(value) {
15141 "\"%.*s\" is more clearly written simply as \"%s\"",
15142 (int) (RExC_parse - rangebegin),
15144 cntrl_to_mnemonic((char) value)
15150 /* Deal with this element of the class */
15154 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15157 /* On non-ASCII platforms, for ranges that span all of 0..255, and
15158 * ones that don't require special handling, we can just add the
15159 * range like we do for ASCII platforms */
15160 if ((UNLIKELY(prevvalue == 0) && value >= 255)
15161 || ! (prevvalue < 256
15163 || (! non_portable_endpoint
15164 && ((isLOWER_A(prevvalue) && isLOWER_A(value))
15165 || (isUPPER_A(prevvalue)
15166 && isUPPER_A(value)))))))
15168 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15172 /* Here, requires special handling. This can be because it is
15173 * a range whose code points are considered to be Unicode, and
15174 * so must be individually translated into native, or because
15175 * its a subrange of 'A-Z' or 'a-z' which each aren't
15176 * contiguous in EBCDIC, but we have defined them to include
15177 * only the "expected" upper or lower case ASCII alphabetics.
15178 * Subranges above 255 are the same in native and Unicode, so
15179 * can be added as a range */
15180 U8 start = NATIVE_TO_LATIN1(prevvalue);
15182 U8 end = (value < 256) ? NATIVE_TO_LATIN1(value) : 255;
15183 for (j = start; j <= end; j++) {
15184 cp_foldable_list = add_cp_to_invlist(cp_foldable_list, LATIN1_TO_NATIVE(j));
15187 cp_foldable_list = _add_range_to_invlist(cp_foldable_list,
15194 range = 0; /* this range (if it was one) is done now */
15195 } /* End of loop through all the text within the brackets */
15197 /* If anything in the class expands to more than one character, we have to
15198 * deal with them by building up a substitute parse string, and recursively
15199 * calling reg() on it, instead of proceeding */
15200 if (multi_char_matches) {
15201 SV * substitute_parse = newSVpvn_flags("?:", 2, SVs_TEMP);
15204 char *save_end = RExC_end;
15205 char *save_parse = RExC_parse;
15206 bool first_time = TRUE; /* First multi-char occurrence doesn't get
15211 #if 0 /* Have decided not to deal with multi-char folds in inverted classes,
15212 because too confusing */
15214 sv_catpv(substitute_parse, "(?:");
15218 /* Look at the longest folds first */
15219 for (cp_count = av_tindex(multi_char_matches); cp_count > 0; cp_count--) {
15221 if (av_exists(multi_char_matches, cp_count)) {
15222 AV** this_array_ptr;
15225 this_array_ptr = (AV**) av_fetch(multi_char_matches,
15227 while ((this_sequence = av_pop(*this_array_ptr)) !=
15230 if (! first_time) {
15231 sv_catpv(substitute_parse, "|");
15233 first_time = FALSE;
15235 sv_catpv(substitute_parse, SvPVX(this_sequence));
15240 /* If the character class contains anything else besides these
15241 * multi-character folds, have to include it in recursive parsing */
15242 if (element_count) {
15243 sv_catpv(substitute_parse, "|[");
15244 sv_catpvn(substitute_parse, orig_parse, RExC_parse - orig_parse);
15245 sv_catpv(substitute_parse, "]");
15248 sv_catpv(substitute_parse, ")");
15251 /* This is a way to get the parse to skip forward a whole named
15252 * sequence instead of matching the 2nd character when it fails the
15254 sv_catpv(substitute_parse, "(*THEN)(*SKIP)(*FAIL)|.)");
15258 RExC_parse = SvPV(substitute_parse, len);
15259 RExC_end = RExC_parse + len;
15260 RExC_in_multi_char_class = 1;
15261 RExC_override_recoding = 1;
15262 RExC_emit = (regnode *)orig_emit;
15264 ret = reg(pRExC_state, 1, ®_flags, depth+1);
15266 *flagp |= reg_flags&(HASWIDTH|SIMPLE|SPSTART|POSTPONED|RESTART_UTF8);
15268 RExC_parse = save_parse;
15269 RExC_end = save_end;
15270 RExC_in_multi_char_class = 0;
15271 RExC_override_recoding = 0;
15272 SvREFCNT_dec_NN(multi_char_matches);
15276 /* Here, we've gone through the entire class and dealt with multi-char
15277 * folds. We are now in a position that we can do some checks to see if we
15278 * can optimize this ANYOF node into a simpler one, even in Pass 1.
15279 * Currently we only do two checks:
15280 * 1) is in the unlikely event that the user has specified both, eg. \w and
15281 * \W under /l, then the class matches everything. (This optimization
15282 * is done only to make the optimizer code run later work.)
15283 * 2) if the character class contains only a single element (including a
15284 * single range), we see if there is an equivalent node for it.
15285 * Other checks are possible */
15286 if (! ret_invlist /* Can't optimize if returning the constructed
15288 && (UNLIKELY(posixl_matches_all) || element_count == 1))
15293 if (UNLIKELY(posixl_matches_all)) {
15296 else if (namedclass > OOB_NAMEDCLASS) { /* this is a named class, like
15297 \w or [:digit:] or \p{foo}
15300 /* All named classes are mapped into POSIXish nodes, with its FLAG
15301 * argument giving which class it is */
15302 switch ((I32)namedclass) {
15303 case ANYOF_UNIPROP:
15306 /* These don't depend on the charset modifiers. They always
15307 * match under /u rules */
15308 case ANYOF_NHORIZWS:
15309 case ANYOF_HORIZWS:
15310 namedclass = ANYOF_BLANK + namedclass - ANYOF_HORIZWS;
15313 case ANYOF_NVERTWS:
15318 /* The actual POSIXish node for all the rest depends on the
15319 * charset modifier. The ones in the first set depend only on
15320 * ASCII or, if available on this platform, also locale */
15324 op = (LOC) ? POSIXL : POSIXA;
15330 /* The following don't have any matches in the upper Latin1
15331 * range, hence /d is equivalent to /u for them. Making it /u
15332 * saves some branches at runtime */
15336 case ANYOF_NXDIGIT:
15337 if (! DEPENDS_SEMANTICS) {
15338 goto treat_as_default;
15344 /* The following change to CASED under /i */
15350 namedclass = ANYOF_CASED + (namedclass % 2);
15354 /* The rest have more possibilities depending on the charset.
15355 * We take advantage of the enum ordering of the charset
15356 * modifiers to get the exact node type, */
15359 op = POSIXD + get_regex_charset(RExC_flags);
15360 if (op > POSIXA) { /* /aa is same as /a */
15365 /* The odd numbered ones are the complements of the
15366 * next-lower even number one */
15367 if (namedclass % 2 == 1) {
15371 arg = namedclass_to_classnum(namedclass);
15375 else if (value == prevvalue) {
15377 /* Here, the class consists of just a single code point */
15380 if (! LOC && value == '\n') {
15381 op = REG_ANY; /* Optimize [^\n] */
15382 *flagp |= HASWIDTH|SIMPLE;
15386 else if (value < 256 || UTF) {
15388 /* Optimize a single value into an EXACTish node, but not if it
15389 * would require converting the pattern to UTF-8. */
15390 op = compute_EXACTish(pRExC_state);
15392 } /* Otherwise is a range */
15393 else if (! LOC) { /* locale could vary these */
15394 if (prevvalue == '0') {
15395 if (value == '9') {
15400 else if (! FOLD || ASCII_FOLD_RESTRICTED) {
15401 /* We can optimize A-Z or a-z, but not if they could match
15402 * something like the KELVIN SIGN under /i. */
15403 if (prevvalue == 'A') {
15406 && ! non_portable_endpoint
15409 arg = (FOLD) ? _CC_ALPHA : _CC_UPPER;
15413 else if (prevvalue == 'a') {
15416 && ! non_portable_endpoint
15419 arg = (FOLD) ? _CC_ALPHA : _CC_LOWER;
15426 /* Here, we have changed <op> away from its initial value iff we found
15427 * an optimization */
15430 /* Throw away this ANYOF regnode, and emit the calculated one,
15431 * which should correspond to the beginning, not current, state of
15433 const char * cur_parse = RExC_parse;
15434 RExC_parse = (char *)orig_parse;
15438 /* To get locale nodes to not use the full ANYOF size would
15439 * require moving the code above that writes the portions
15440 * of it that aren't in other nodes to after this point.
15441 * e.g. ANYOF_POSIXL_SET */
15442 RExC_size = orig_size;
15446 RExC_emit = (regnode *)orig_emit;
15447 if (PL_regkind[op] == POSIXD) {
15448 if (op == POSIXL) {
15449 RExC_contains_locale = 1;
15452 op += NPOSIXD - POSIXD;
15457 ret = reg_node(pRExC_state, op);
15459 if (PL_regkind[op] == POSIXD || PL_regkind[op] == NPOSIXD) {
15463 *flagp |= HASWIDTH|SIMPLE;
15465 else if (PL_regkind[op] == EXACT) {
15466 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15467 TRUE /* downgradable to EXACT */
15471 RExC_parse = (char *) cur_parse;
15473 SvREFCNT_dec(posixes);
15474 SvREFCNT_dec(nposixes);
15475 SvREFCNT_dec(simple_posixes);
15476 SvREFCNT_dec(cp_list);
15477 SvREFCNT_dec(cp_foldable_list);
15484 /****** !SIZE_ONLY (Pass 2) AFTER HERE *********/
15486 /* If folding, we calculate all characters that could fold to or from the
15487 * ones already on the list */
15488 if (cp_foldable_list) {
15490 UV start, end; /* End points of code point ranges */
15492 SV* fold_intersection = NULL;
15495 /* Our calculated list will be for Unicode rules. For locale
15496 * matching, we have to keep a separate list that is consulted at
15497 * runtime only when the locale indicates Unicode rules. For
15498 * non-locale, we just use to the general list */
15500 use_list = &only_utf8_locale_list;
15503 use_list = &cp_list;
15506 /* Only the characters in this class that participate in folds need
15507 * be checked. Get the intersection of this class and all the
15508 * possible characters that are foldable. This can quickly narrow
15509 * down a large class */
15510 _invlist_intersection(PL_utf8_foldable, cp_foldable_list,
15511 &fold_intersection);
15513 /* The folds for all the Latin1 characters are hard-coded into this
15514 * program, but we have to go out to disk to get the others. */
15515 if (invlist_highest(cp_foldable_list) >= 256) {
15517 /* This is a hash that for a particular fold gives all
15518 * characters that are involved in it */
15519 if (! PL_utf8_foldclosures) {
15520 _load_PL_utf8_foldclosures();
15524 /* Now look at the foldable characters in this class individually */
15525 invlist_iterinit(fold_intersection);
15526 while (invlist_iternext(fold_intersection, &start, &end)) {
15529 /* Look at every character in the range */
15530 for (j = start; j <= end; j++) {
15531 U8 foldbuf[UTF8_MAXBYTES_CASE+1];
15537 if (IS_IN_SOME_FOLD_L1(j)) {
15539 /* ASCII is always matched; non-ASCII is matched
15540 * only under Unicode rules (which could happen
15541 * under /l if the locale is a UTF-8 one */
15542 if (isASCII(j) || ! DEPENDS_SEMANTICS) {
15543 *use_list = add_cp_to_invlist(*use_list,
15544 PL_fold_latin1[j]);
15548 add_cp_to_invlist(depends_list,
15549 PL_fold_latin1[j]);
15553 if (HAS_NONLATIN1_SIMPLE_FOLD_CLOSURE(j)
15554 && (! isASCII(j) || ! ASCII_FOLD_RESTRICTED))
15556 add_above_Latin1_folds(pRExC_state,
15563 /* Here is an above Latin1 character. We don't have the
15564 * rules hard-coded for it. First, get its fold. This is
15565 * the simple fold, as the multi-character folds have been
15566 * handled earlier and separated out */
15567 _to_uni_fold_flags(j, foldbuf, &foldlen,
15568 (ASCII_FOLD_RESTRICTED)
15569 ? FOLD_FLAGS_NOMIX_ASCII
15572 /* Single character fold of above Latin1. Add everything in
15573 * its fold closure to the list that this node should match.
15574 * The fold closures data structure is a hash with the keys
15575 * being the UTF-8 of every character that is folded to, like
15576 * 'k', and the values each an array of all code points that
15577 * fold to its key. e.g. [ 'k', 'K', KELVIN_SIGN ].
15578 * Multi-character folds are not included */
15579 if ((listp = hv_fetch(PL_utf8_foldclosures,
15580 (char *) foldbuf, foldlen, FALSE)))
15582 AV* list = (AV*) *listp;
15584 for (k = 0; k <= av_tindex(list); k++) {
15585 SV** c_p = av_fetch(list, k, FALSE);
15591 /* /aa doesn't allow folds between ASCII and non- */
15592 if ((ASCII_FOLD_RESTRICTED
15593 && (isASCII(c) != isASCII(j))))
15598 /* Folds under /l which cross the 255/256 boundary
15599 * are added to a separate list. (These are valid
15600 * only when the locale is UTF-8.) */
15601 if (c < 256 && LOC) {
15602 *use_list = add_cp_to_invlist(*use_list, c);
15606 if (isASCII(c) || c > 255 || AT_LEAST_UNI_SEMANTICS)
15608 cp_list = add_cp_to_invlist(cp_list, c);
15611 /* Similarly folds involving non-ascii Latin1
15612 * characters under /d are added to their list */
15613 depends_list = add_cp_to_invlist(depends_list,
15620 SvREFCNT_dec_NN(fold_intersection);
15623 /* Now that we have finished adding all the folds, there is no reason
15624 * to keep the foldable list separate */
15625 _invlist_union(cp_list, cp_foldable_list, &cp_list);
15626 SvREFCNT_dec_NN(cp_foldable_list);
15629 /* And combine the result (if any) with any inversion list from posix
15630 * classes. The lists are kept separate up to now because we don't want to
15631 * fold the classes (folding of those is automatically handled by the swash
15632 * fetching code) */
15633 if (simple_posixes) {
15634 _invlist_union(cp_list, simple_posixes, &cp_list);
15635 SvREFCNT_dec_NN(simple_posixes);
15637 if (posixes || nposixes) {
15638 if (posixes && AT_LEAST_ASCII_RESTRICTED) {
15639 /* Under /a and /aa, nothing above ASCII matches these */
15640 _invlist_intersection(posixes,
15641 PL_XPosix_ptrs[_CC_ASCII],
15645 if (DEPENDS_SEMANTICS) {
15646 /* Under /d, everything in the upper half of the Latin1 range
15647 * matches these complements */
15648 ANYOF_FLAGS(ret) |= ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII;
15650 else if (AT_LEAST_ASCII_RESTRICTED) {
15651 /* Under /a and /aa, everything above ASCII matches these
15653 _invlist_union_complement_2nd(nposixes,
15654 PL_XPosix_ptrs[_CC_ASCII],
15658 _invlist_union(posixes, nposixes, &posixes);
15659 SvREFCNT_dec_NN(nposixes);
15662 posixes = nposixes;
15665 if (! DEPENDS_SEMANTICS) {
15667 _invlist_union(cp_list, posixes, &cp_list);
15668 SvREFCNT_dec_NN(posixes);
15675 /* Under /d, we put into a separate list the Latin1 things that
15676 * match only when the target string is utf8 */
15677 SV* nonascii_but_latin1_properties = NULL;
15678 _invlist_intersection(posixes, PL_UpperLatin1,
15679 &nonascii_but_latin1_properties);
15680 _invlist_subtract(posixes, nonascii_but_latin1_properties,
15683 _invlist_union(cp_list, posixes, &cp_list);
15684 SvREFCNT_dec_NN(posixes);
15690 if (depends_list) {
15691 _invlist_union(depends_list, nonascii_but_latin1_properties,
15693 SvREFCNT_dec_NN(nonascii_but_latin1_properties);
15696 depends_list = nonascii_but_latin1_properties;
15701 /* And combine the result (if any) with any inversion list from properties.
15702 * The lists are kept separate up to now so that we can distinguish the two
15703 * in regards to matching above-Unicode. A run-time warning is generated
15704 * if a Unicode property is matched against a non-Unicode code point. But,
15705 * we allow user-defined properties to match anything, without any warning,
15706 * and we also suppress the warning if there is a portion of the character
15707 * class that isn't a Unicode property, and which matches above Unicode, \W
15708 * or [\x{110000}] for example.
15709 * (Note that in this case, unlike the Posix one above, there is no
15710 * <depends_list>, because having a Unicode property forces Unicode
15715 /* If it matters to the final outcome, see if a non-property
15716 * component of the class matches above Unicode. If so, the
15717 * warning gets suppressed. This is true even if just a single
15718 * such code point is specified, as though not strictly correct if
15719 * another such code point is matched against, the fact that they
15720 * are using above-Unicode code points indicates they should know
15721 * the issues involved */
15723 warn_super = ! (invert
15724 ^ (invlist_highest(cp_list) > PERL_UNICODE_MAX));
15727 _invlist_union(properties, cp_list, &cp_list);
15728 SvREFCNT_dec_NN(properties);
15731 cp_list = properties;
15735 ANYOF_FLAGS(ret) |= ANYOF_WARN_SUPER;
15739 /* Here, we have calculated what code points should be in the character
15742 * Now we can see about various optimizations. Fold calculation (which we
15743 * did above) needs to take place before inversion. Otherwise /[^k]/i
15744 * would invert to include K, which under /i would match k, which it
15745 * shouldn't. Therefore we can't invert folded locale now, as it won't be
15746 * folded until runtime */
15748 /* If we didn't do folding, it's because some information isn't available
15749 * until runtime; set the run-time fold flag for these. (We don't have to
15750 * worry about properties folding, as that is taken care of by the swash
15751 * fetching). We know to set the flag if we have a non-NULL list for UTF-8
15752 * locales, or the class matches at least one 0-255 range code point */
15754 if (only_utf8_locale_list) {
15755 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15757 else if (cp_list) { /* Look to see if there a 0-255 code point is in
15760 invlist_iterinit(cp_list);
15761 if (invlist_iternext(cp_list, &start, &end) && start < 256) {
15762 ANYOF_FLAGS(ret) |= ANYOF_LOC_FOLD;
15764 invlist_iterfinish(cp_list);
15768 /* Optimize inverted simple patterns (e.g. [^a-z]) when everything is known
15769 * at compile time. Besides not inverting folded locale now, we can't
15770 * invert if there are things such as \w, which aren't known until runtime
15774 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15776 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15778 _invlist_invert(cp_list);
15780 /* Any swash can't be used as-is, because we've inverted things */
15782 SvREFCNT_dec_NN(swash);
15786 /* Clear the invert flag since have just done it here */
15793 *ret_invlist = cp_list;
15794 SvREFCNT_dec(swash);
15796 /* Discard the generated node */
15798 RExC_size = orig_size;
15801 RExC_emit = orig_emit;
15806 /* Some character classes are equivalent to other nodes. Such nodes take
15807 * up less room and generally fewer operations to execute than ANYOF nodes.
15808 * Above, we checked for and optimized into some such equivalents for
15809 * certain common classes that are easy to test. Getting to this point in
15810 * the code means that the class didn't get optimized there. Since this
15811 * code is only executed in Pass 2, it is too late to save space--it has
15812 * been allocated in Pass 1, and currently isn't given back. But turning
15813 * things into an EXACTish node can allow the optimizer to join it to any
15814 * adjacent such nodes. And if the class is equivalent to things like /./,
15815 * expensive run-time swashes can be avoided. Now that we have more
15816 * complete information, we can find things necessarily missed by the
15817 * earlier code. I (khw) am not sure how much to look for here. It would
15818 * be easy, but perhaps too slow, to check any candidates against all the
15819 * node types they could possibly match using _invlistEQ(). */
15824 && ! (ANYOF_FLAGS(ret) & (ANYOF_LOCALE_FLAGS))
15825 && ! HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15827 /* We don't optimize if we are supposed to make sure all non-Unicode
15828 * code points raise a warning, as only ANYOF nodes have this check.
15830 && ! ((ANYOF_FLAGS(ret) & ANYOF_WARN_SUPER) && ALWAYS_WARN_SUPER))
15833 U8 op = END; /* The optimzation node-type */
15834 const char * cur_parse= RExC_parse;
15836 invlist_iterinit(cp_list);
15837 if (! invlist_iternext(cp_list, &start, &end)) {
15839 /* Here, the list is empty. This happens, for example, when a
15840 * Unicode property is the only thing in the character class, and
15841 * it doesn't match anything. (perluniprops.pod notes such
15844 *flagp |= HASWIDTH|SIMPLE;
15846 else if (start == end) { /* The range is a single code point */
15847 if (! invlist_iternext(cp_list, &start, &end)
15849 /* Don't do this optimization if it would require changing
15850 * the pattern to UTF-8 */
15851 && (start < 256 || UTF))
15853 /* Here, the list contains a single code point. Can optimize
15854 * into an EXACTish node */
15865 /* A locale node under folding with one code point can be
15866 * an EXACTFL, as its fold won't be calculated until
15872 /* Here, we are generally folding, but there is only one
15873 * code point to match. If we have to, we use an EXACT
15874 * node, but it would be better for joining with adjacent
15875 * nodes in the optimization pass if we used the same
15876 * EXACTFish node that any such are likely to be. We can
15877 * do this iff the code point doesn't participate in any
15878 * folds. For example, an EXACTF of a colon is the same as
15879 * an EXACT one, since nothing folds to or from a colon. */
15881 if (IS_IN_SOME_FOLD_L1(value)) {
15886 if (_invlist_contains_cp(PL_utf8_foldable, value)) {
15891 /* If we haven't found the node type, above, it means we
15892 * can use the prevailing one */
15894 op = compute_EXACTish(pRExC_state);
15899 else if (start == 0) {
15900 if (end == UV_MAX) {
15902 *flagp |= HASWIDTH|SIMPLE;
15905 else if (end == '\n' - 1
15906 && invlist_iternext(cp_list, &start, &end)
15907 && start == '\n' + 1 && end == UV_MAX)
15910 *flagp |= HASWIDTH|SIMPLE;
15914 invlist_iterfinish(cp_list);
15917 RExC_parse = (char *)orig_parse;
15918 RExC_emit = (regnode *)orig_emit;
15920 ret = reg_node(pRExC_state, op);
15922 RExC_parse = (char *)cur_parse;
15924 if (PL_regkind[op] == EXACT) {
15925 alloc_maybe_populate_EXACT(pRExC_state, ret, flagp, 0, value,
15926 TRUE /* downgradable to EXACT */
15930 SvREFCNT_dec_NN(cp_list);
15935 /* Here, <cp_list> contains all the code points we can determine at
15936 * compile time that match under all conditions. Go through it, and
15937 * for things that belong in the bitmap, put them there, and delete from
15938 * <cp_list>. While we are at it, see if everything above 255 is in the
15939 * list, and if so, set a flag to speed up execution */
15941 populate_ANYOF_from_invlist(ret, &cp_list);
15944 ANYOF_FLAGS(ret) |= ANYOF_INVERT;
15947 /* Here, the bitmap has been populated with all the Latin1 code points that
15948 * always match. Can now add to the overall list those that match only
15949 * when the target string is UTF-8 (<depends_list>). */
15950 if (depends_list) {
15952 _invlist_union(cp_list, depends_list, &cp_list);
15953 SvREFCNT_dec_NN(depends_list);
15956 cp_list = depends_list;
15958 ANYOF_FLAGS(ret) |= ANYOF_HAS_UTF8_NONBITMAP_MATCHES;
15961 /* If there is a swash and more than one element, we can't use the swash in
15962 * the optimization below. */
15963 if (swash && element_count > 1) {
15964 SvREFCNT_dec_NN(swash);
15968 /* Note that the optimization of using 'swash' if it is the only thing in
15969 * the class doesn't have us change swash at all, so it can include things
15970 * that are also in the bitmap; otherwise we have purposely deleted that
15971 * duplicate information */
15972 set_ANYOF_arg(pRExC_state, ret, cp_list,
15973 (HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION)
15975 only_utf8_locale_list,
15976 swash, has_user_defined_property);
15978 *flagp |= HASWIDTH|SIMPLE;
15980 if (ANYOF_FLAGS(ret) & ANYOF_LOCALE_FLAGS) {
15981 RExC_contains_locale = 1;
15987 #undef HAS_NONLOCALE_RUNTIME_PROPERTY_DEFINITION
15990 S_set_ANYOF_arg(pTHX_ RExC_state_t* const pRExC_state,
15991 regnode* const node,
15993 SV* const runtime_defns,
15994 SV* const only_utf8_locale_list,
15996 const bool has_user_defined_property)
15998 /* Sets the arg field of an ANYOF-type node 'node', using information about
15999 * the node passed-in. If there is nothing outside the node's bitmap, the
16000 * arg is set to ANYOF_ONLY_HAS_BITMAP. Otherwise, it sets the argument to
16001 * the count returned by add_data(), having allocated and stored an array,
16002 * av, that that count references, as follows:
16003 * av[0] stores the character class description in its textual form.
16004 * This is used later (regexec.c:Perl_regclass_swash()) to
16005 * initialize the appropriate swash, and is also useful for dumping
16006 * the regnode. This is set to &PL_sv_undef if the textual
16007 * description is not needed at run-time (as happens if the other
16008 * elements completely define the class)
16009 * av[1] if &PL_sv_undef, is a placeholder to later contain the swash
16010 * computed from av[0]. But if no further computation need be done,
16011 * the swash is stored here now (and av[0] is &PL_sv_undef).
16012 * av[2] stores the inversion list of code points that match only if the
16013 * current locale is UTF-8
16014 * av[3] stores the cp_list inversion list for use in addition or instead
16015 * of av[0]; used only if cp_list exists and av[1] is &PL_sv_undef.
16016 * (Otherwise everything needed is already in av[0] and av[1])
16017 * av[4] is set if any component of the class is from a user-defined
16018 * property; used only if av[3] exists */
16022 PERL_ARGS_ASSERT_SET_ANYOF_ARG;
16024 if (! cp_list && ! runtime_defns && ! only_utf8_locale_list) {
16025 assert(! (ANYOF_FLAGS(node)
16026 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16027 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES)));
16028 ARG_SET(node, ANYOF_ONLY_HAS_BITMAP);
16031 AV * const av = newAV();
16034 assert(ANYOF_FLAGS(node)
16035 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16036 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
16038 av_store(av, 0, (runtime_defns)
16039 ? SvREFCNT_inc(runtime_defns) : &PL_sv_undef);
16042 av_store(av, 1, swash);
16043 SvREFCNT_dec_NN(cp_list);
16046 av_store(av, 1, &PL_sv_undef);
16048 av_store(av, 3, cp_list);
16049 av_store(av, 4, newSVuv(has_user_defined_property));
16053 if (only_utf8_locale_list) {
16054 av_store(av, 2, only_utf8_locale_list);
16057 av_store(av, 2, &PL_sv_undef);
16060 rv = newRV_noinc(MUTABLE_SV(av));
16061 n = add_data(pRExC_state, STR_WITH_LEN("s"));
16062 RExC_rxi->data->data[n] = (void*)rv;
16067 #if !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION)
16069 Perl__get_regclass_nonbitmap_data(pTHX_ const regexp *prog,
16070 const regnode* node,
16073 SV** only_utf8_locale_ptr,
16077 /* For internal core use only.
16078 * Returns the swash for the input 'node' in the regex 'prog'.
16079 * If <doinit> is 'true', will attempt to create the swash if not already
16081 * If <listsvp> is non-null, will return the printable contents of the
16082 * swash. This can be used to get debugging information even before the
16083 * swash exists, by calling this function with 'doinit' set to false, in
16084 * which case the components that will be used to eventually create the
16085 * swash are returned (in a printable form).
16086 * If <exclude_list> is not NULL, it is an inversion list of things to
16087 * exclude from what's returned in <listsvp>.
16088 * Tied intimately to how S_set_ANYOF_arg sets up the data structure. Note
16089 * that, in spite of this function's name, the swash it returns may include
16090 * the bitmap data as well */
16093 SV *si = NULL; /* Input swash initialization string */
16094 SV* invlist = NULL;
16096 RXi_GET_DECL(prog,progi);
16097 const struct reg_data * const data = prog ? progi->data : NULL;
16099 PERL_ARGS_ASSERT__GET_REGCLASS_NONBITMAP_DATA;
16101 assert(ANYOF_FLAGS(node)
16102 & (ANYOF_HAS_UTF8_NONBITMAP_MATCHES
16103 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES|ANYOF_LOC_FOLD));
16105 if (data && data->count) {
16106 const U32 n = ARG(node);
16108 if (data->what[n] == 's') {
16109 SV * const rv = MUTABLE_SV(data->data[n]);
16110 AV * const av = MUTABLE_AV(SvRV(rv));
16111 SV **const ary = AvARRAY(av);
16112 U8 swash_init_flags = _CORE_SWASH_INIT_ACCEPT_INVLIST;
16114 si = *ary; /* ary[0] = the string to initialize the swash with */
16116 /* Elements 3 and 4 are either both present or both absent. [3] is
16117 * any inversion list generated at compile time; [4] indicates if
16118 * that inversion list has any user-defined properties in it. */
16119 if (av_tindex(av) >= 2) {
16120 if (only_utf8_locale_ptr
16122 && ary[2] != &PL_sv_undef)
16124 *only_utf8_locale_ptr = ary[2];
16127 assert(only_utf8_locale_ptr);
16128 *only_utf8_locale_ptr = NULL;
16131 if (av_tindex(av) >= 3) {
16133 if (SvUV(ary[4])) {
16134 swash_init_flags |= _CORE_SWASH_INIT_USER_DEFINED_PROPERTY;
16142 /* Element [1] is reserved for the set-up swash. If already there,
16143 * return it; if not, create it and store it there */
16144 if (ary[1] && SvROK(ary[1])) {
16147 else if (doinit && ((si && si != &PL_sv_undef)
16148 || (invlist && invlist != &PL_sv_undef))) {
16150 sw = _core_swash_init("utf8", /* the utf8 package */
16154 0, /* not from tr/// */
16156 &swash_init_flags);
16157 (void)av_store(av, 1, sw);
16162 /* If requested, return a printable version of what this swash matches */
16164 SV* matches_string = newSVpvs("");
16166 /* The swash should be used, if possible, to get the data, as it
16167 * contains the resolved data. But this function can be called at
16168 * compile-time, before everything gets resolved, in which case we
16169 * return the currently best available information, which is the string
16170 * that will eventually be used to do that resolving, 'si' */
16171 if ((! sw || (invlist = _get_swash_invlist(sw)) == NULL)
16172 && (si && si != &PL_sv_undef))
16174 sv_catsv(matches_string, si);
16177 /* Add the inversion list to whatever we have. This may have come from
16178 * the swash, or from an input parameter */
16180 if (exclude_list) {
16181 SV* clone = invlist_clone(invlist);
16182 _invlist_subtract(clone, exclude_list, &clone);
16183 sv_catsv(matches_string, _invlist_contents(clone));
16184 SvREFCNT_dec_NN(clone);
16187 sv_catsv(matches_string, _invlist_contents(invlist));
16190 *listsvp = matches_string;
16195 #endif /* !defined(PERL_IN_XSUB_RE) || defined(PLUGGABLE_RE_EXTENSION) */
16197 /* reg_skipcomment()
16199 Absorbs an /x style # comment from the input stream,
16200 returning a pointer to the first character beyond the comment, or if the
16201 comment terminates the pattern without anything following it, this returns
16202 one past the final character of the pattern (in other words, RExC_end) and
16203 sets the REG_RUN_ON_COMMENT_SEEN flag.
16205 Note it's the callers responsibility to ensure that we are
16206 actually in /x mode
16210 PERL_STATIC_INLINE char*
16211 S_reg_skipcomment(RExC_state_t *pRExC_state, char* p)
16213 PERL_ARGS_ASSERT_REG_SKIPCOMMENT;
16217 while (p < RExC_end) {
16218 if (*(++p) == '\n') {
16223 /* we ran off the end of the pattern without ending the comment, so we have
16224 * to add an \n when wrapping */
16225 RExC_seen |= REG_RUN_ON_COMMENT_SEEN;
16231 Advances the parse position, and optionally absorbs
16232 "whitespace" from the inputstream.
16234 Without /x "whitespace" means (?#...) style comments only,
16235 with /x this means (?#...) and # comments and whitespace proper.
16237 Returns the RExC_parse point from BEFORE the scan occurs.
16239 This is the /x friendly way of saying RExC_parse++.
16243 S_nextchar(pTHX_ RExC_state_t *pRExC_state)
16245 char* const retval = RExC_parse++;
16247 PERL_ARGS_ASSERT_NEXTCHAR;
16250 if (RExC_end - RExC_parse >= 3
16251 && *RExC_parse == '('
16252 && RExC_parse[1] == '?'
16253 && RExC_parse[2] == '#')
16255 while (*RExC_parse != ')') {
16256 if (RExC_parse == RExC_end)
16257 FAIL("Sequence (?#... not terminated");
16263 if (RExC_flags & RXf_PMf_EXTENDED) {
16264 char * p = regpatws(pRExC_state, RExC_parse,
16265 TRUE); /* means recognize comments */
16266 if (p != RExC_parse) {
16276 S_regnode_guts(pTHX_ RExC_state_t *pRExC_state, const U8 op, const STRLEN extra_size, const char* const name)
16278 /* Allocate a regnode for 'op' and returns it, with 'extra_size' extra
16279 * space. In pass1, it aligns and increments RExC_size; in pass2,
16282 regnode * const ret = RExC_emit;
16283 GET_RE_DEBUG_FLAGS_DECL;
16285 PERL_ARGS_ASSERT_REGNODE_GUTS;
16287 assert(extra_size >= regarglen[op]);
16290 SIZE_ALIGN(RExC_size);
16291 RExC_size += 1 + extra_size;
16294 if (RExC_emit >= RExC_emit_bound)
16295 Perl_croak(aTHX_ "panic: reg_node overrun trying to emit %d, %p>=%p",
16296 op, (void*)RExC_emit, (void*)RExC_emit_bound);
16298 NODE_ALIGN_FILL(ret);
16299 #ifndef RE_TRACK_PATTERN_OFFSETS
16300 PERL_UNUSED_ARG(name);
16302 if (RExC_offsets) { /* MJD */
16304 ("%s:%d: (op %s) %s %"UVuf" (len %"UVuf") (max %"UVuf").\n",
16307 (UV)(RExC_emit - RExC_emit_start) > RExC_offsets[0]
16308 ? "Overwriting end of array!\n" : "OK",
16309 (UV)(RExC_emit - RExC_emit_start),
16310 (UV)(RExC_parse - RExC_start),
16311 (UV)RExC_offsets[0]));
16312 Set_Node_Offset(RExC_emit, RExC_parse + (op == END));
16319 - reg_node - emit a node
16321 STATIC regnode * /* Location. */
16322 S_reg_node(pTHX_ RExC_state_t *pRExC_state, U8 op)
16324 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg_node");
16326 PERL_ARGS_ASSERT_REG_NODE;
16328 assert(regarglen[op] == 0);
16331 regnode *ptr = ret;
16332 FILL_ADVANCE_NODE(ptr, op);
16333 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 1);
16340 - reganode - emit a node with an argument
16342 STATIC regnode * /* Location. */
16343 S_reganode(pTHX_ RExC_state_t *pRExC_state, U8 op, U32 arg)
16345 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reganode");
16347 PERL_ARGS_ASSERT_REGANODE;
16349 assert(regarglen[op] == 1);
16352 regnode *ptr = ret;
16353 FILL_ADVANCE_NODE_ARG(ptr, op, arg);
16354 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (ptr) - 2);
16361 S_reg2Lanode(pTHX_ RExC_state_t *pRExC_state, const U8 op, const U32 arg1, const I32 arg2)
16363 /* emit a node with U32 and I32 arguments */
16365 regnode * const ret = regnode_guts(pRExC_state, op, regarglen[op], "reg2Lanode");
16367 PERL_ARGS_ASSERT_REG2LANODE;
16369 assert(regarglen[op] == 2);
16372 regnode *ptr = ret;
16373 FILL_ADVANCE_NODE_2L_ARG(ptr, op, arg1, arg2);
16380 - reginsert - insert an operator in front of already-emitted operand
16382 * Means relocating the operand.
16385 S_reginsert(pTHX_ RExC_state_t *pRExC_state, U8 op, regnode *opnd, U32 depth)
16390 const int offset = regarglen[(U8)op];
16391 const int size = NODE_STEP_REGNODE + offset;
16392 GET_RE_DEBUG_FLAGS_DECL;
16394 PERL_ARGS_ASSERT_REGINSERT;
16395 PERL_UNUSED_CONTEXT;
16396 PERL_UNUSED_ARG(depth);
16397 /* (PL_regkind[(U8)op] == CURLY ? EXTRA_STEP_2ARGS : 0); */
16398 DEBUG_PARSE_FMT("inst"," - %s",PL_reg_name[op]);
16407 if (RExC_open_parens) {
16409 /*DEBUG_PARSE_FMT("inst"," - %"IVdf, (IV)RExC_npar);*/
16410 for ( paren=0 ; paren < RExC_npar ; paren++ ) {
16411 if ( RExC_open_parens[paren] >= opnd ) {
16412 /*DEBUG_PARSE_FMT("open"," - %d",size);*/
16413 RExC_open_parens[paren] += size;
16415 /*DEBUG_PARSE_FMT("open"," - %s","ok");*/
16417 if ( RExC_close_parens[paren] >= opnd ) {
16418 /*DEBUG_PARSE_FMT("close"," - %d",size);*/
16419 RExC_close_parens[paren] += size;
16421 /*DEBUG_PARSE_FMT("close"," - %s","ok");*/
16426 while (src > opnd) {
16427 StructCopy(--src, --dst, regnode);
16428 #ifdef RE_TRACK_PATTERN_OFFSETS
16429 if (RExC_offsets) { /* MJD 20010112 */
16431 ("%s(%d): (op %s) %s copy %"UVuf" -> %"UVuf" (max %"UVuf").\n",
16435 (UV)(dst - RExC_emit_start) > RExC_offsets[0]
16436 ? "Overwriting end of array!\n" : "OK",
16437 (UV)(src - RExC_emit_start),
16438 (UV)(dst - RExC_emit_start),
16439 (UV)RExC_offsets[0]));
16440 Set_Node_Offset_To_R(dst-RExC_emit_start, Node_Offset(src));
16441 Set_Node_Length_To_R(dst-RExC_emit_start, Node_Length(src));
16447 place = opnd; /* Op node, where operand used to be. */
16448 #ifdef RE_TRACK_PATTERN_OFFSETS
16449 if (RExC_offsets) { /* MJD */
16451 ("%s(%d): (op %s) %s %"UVuf" <- %"UVuf" (max %"UVuf").\n",
16455 (UV)(place - RExC_emit_start) > RExC_offsets[0]
16456 ? "Overwriting end of array!\n" : "OK",
16457 (UV)(place - RExC_emit_start),
16458 (UV)(RExC_parse - RExC_start),
16459 (UV)RExC_offsets[0]));
16460 Set_Node_Offset(place, RExC_parse);
16461 Set_Node_Length(place, 1);
16464 src = NEXTOPER(place);
16465 FILL_ADVANCE_NODE(place, op);
16466 REH_CALL_COMP_NODE_HOOK(pRExC_state->rx, (place) - 1);
16467 Zero(src, offset, regnode);
16471 - regtail - set the next-pointer at the end of a node chain of p to val.
16472 - SEE ALSO: regtail_study
16474 /* TODO: All three parms should be const */
16476 S_regtail(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16477 const regnode *val,U32 depth)
16480 GET_RE_DEBUG_FLAGS_DECL;
16482 PERL_ARGS_ASSERT_REGTAIL;
16484 PERL_UNUSED_ARG(depth);
16490 /* Find last node. */
16493 regnode * const temp = regnext(scan);
16495 DEBUG_PARSE_MSG((scan==p ? "tail" : ""));
16496 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16497 PerlIO_printf(Perl_debug_log, "~ %s (%d) %s %s\n",
16498 SvPV_nolen_const(RExC_mysv), REG_NODE_NUM(scan),
16499 (temp == NULL ? "->" : ""),
16500 (temp == NULL ? PL_reg_name[OP(val)] : "")
16508 if (reg_off_by_arg[OP(scan)]) {
16509 ARG_SET(scan, val - scan);
16512 NEXT_OFF(scan) = val - scan;
16518 - regtail_study - set the next-pointer at the end of a node chain of p to val.
16519 - Look for optimizable sequences at the same time.
16520 - currently only looks for EXACT chains.
16522 This is experimental code. The idea is to use this routine to perform
16523 in place optimizations on branches and groups as they are constructed,
16524 with the long term intention of removing optimization from study_chunk so
16525 that it is purely analytical.
16527 Currently only used when in DEBUG mode. The macro REGTAIL_STUDY() is used
16528 to control which is which.
16531 /* TODO: All four parms should be const */
16534 S_regtail_study(pTHX_ RExC_state_t *pRExC_state, regnode *p,
16535 const regnode *val,U32 depth)
16539 #ifdef EXPERIMENTAL_INPLACESCAN
16542 GET_RE_DEBUG_FLAGS_DECL;
16544 PERL_ARGS_ASSERT_REGTAIL_STUDY;
16550 /* Find last node. */
16554 regnode * const temp = regnext(scan);
16555 #ifdef EXPERIMENTAL_INPLACESCAN
16556 if (PL_regkind[OP(scan)] == EXACT) {
16557 bool unfolded_multi_char; /* Unexamined in this routine */
16558 if (join_exact(pRExC_state, scan, &min,
16559 &unfolded_multi_char, 1, val, depth+1))
16564 switch (OP(scan)) {
16568 case EXACTFA_NO_TRIE:
16574 if( exact == PSEUDO )
16576 else if ( exact != OP(scan) )
16585 DEBUG_PARSE_MSG((scan==p ? "tsdy" : ""));
16586 regprop(RExC_rx, RExC_mysv, scan, NULL, pRExC_state);
16587 PerlIO_printf(Perl_debug_log, "~ %s (%d) -> %s\n",
16588 SvPV_nolen_const(RExC_mysv),
16589 REG_NODE_NUM(scan),
16590 PL_reg_name[exact]);
16597 DEBUG_PARSE_MSG("");
16598 regprop(RExC_rx, RExC_mysv, val, NULL, pRExC_state);
16599 PerlIO_printf(Perl_debug_log,
16600 "~ attach to %s (%"IVdf") offset to %"IVdf"\n",
16601 SvPV_nolen_const(RExC_mysv),
16602 (IV)REG_NODE_NUM(val),
16606 if (reg_off_by_arg[OP(scan)]) {
16607 ARG_SET(scan, val - scan);
16610 NEXT_OFF(scan) = val - scan;
16618 - regdump - dump a regexp onto Perl_debug_log in vaguely comprehensible form
16623 S_regdump_intflags(pTHX_ const char *lead, const U32 flags)
16628 ASSUME(REG_INTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16630 for (bit=0; bit<REG_INTFLAGS_NAME_SIZE; bit++) {
16631 if (flags & (1<<bit)) {
16632 if (!set++ && lead)
16633 PerlIO_printf(Perl_debug_log, "%s",lead);
16634 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_intflags_name[bit]);
16639 PerlIO_printf(Perl_debug_log, "\n");
16641 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16646 S_regdump_extflags(pTHX_ const char *lead, const U32 flags)
16652 ASSUME(REG_EXTFLAGS_NAME_SIZE <= sizeof(flags)*8);
16654 for (bit=0; bit<REG_EXTFLAGS_NAME_SIZE; bit++) {
16655 if (flags & (1<<bit)) {
16656 if ((1<<bit) & RXf_PMf_CHARSET) { /* Output separately, below */
16659 if (!set++ && lead)
16660 PerlIO_printf(Perl_debug_log, "%s",lead);
16661 PerlIO_printf(Perl_debug_log, "%s ",PL_reg_extflags_name[bit]);
16664 if ((cs = get_regex_charset(flags)) != REGEX_DEPENDS_CHARSET) {
16665 if (!set++ && lead) {
16666 PerlIO_printf(Perl_debug_log, "%s",lead);
16669 case REGEX_UNICODE_CHARSET:
16670 PerlIO_printf(Perl_debug_log, "UNICODE");
16672 case REGEX_LOCALE_CHARSET:
16673 PerlIO_printf(Perl_debug_log, "LOCALE");
16675 case REGEX_ASCII_RESTRICTED_CHARSET:
16676 PerlIO_printf(Perl_debug_log, "ASCII-RESTRICTED");
16678 case REGEX_ASCII_MORE_RESTRICTED_CHARSET:
16679 PerlIO_printf(Perl_debug_log, "ASCII-MORE_RESTRICTED");
16682 PerlIO_printf(Perl_debug_log, "UNKNOWN CHARACTER SET");
16688 PerlIO_printf(Perl_debug_log, "\n");
16690 PerlIO_printf(Perl_debug_log, "%s[none-set]\n",lead);
16696 Perl_regdump(pTHX_ const regexp *r)
16699 SV * const sv = sv_newmortal();
16700 SV *dsv= sv_newmortal();
16701 RXi_GET_DECL(r,ri);
16702 GET_RE_DEBUG_FLAGS_DECL;
16704 PERL_ARGS_ASSERT_REGDUMP;
16706 (void)dumpuntil(r, ri->program, ri->program + 1, NULL, NULL, sv, 0, 0);
16708 /* Header fields of interest. */
16709 if (r->anchored_substr) {
16710 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->anchored_substr),
16711 RE_SV_DUMPLEN(r->anchored_substr), 30);
16712 PerlIO_printf(Perl_debug_log,
16713 "anchored %s%s at %"IVdf" ",
16714 s, RE_SV_TAIL(r->anchored_substr),
16715 (IV)r->anchored_offset);
16716 } else if (r->anchored_utf8) {
16717 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->anchored_utf8),
16718 RE_SV_DUMPLEN(r->anchored_utf8), 30);
16719 PerlIO_printf(Perl_debug_log,
16720 "anchored utf8 %s%s at %"IVdf" ",
16721 s, RE_SV_TAIL(r->anchored_utf8),
16722 (IV)r->anchored_offset);
16724 if (r->float_substr) {
16725 RE_PV_QUOTED_DECL(s, 0, dsv, SvPVX_const(r->float_substr),
16726 RE_SV_DUMPLEN(r->float_substr), 30);
16727 PerlIO_printf(Perl_debug_log,
16728 "floating %s%s at %"IVdf"..%"UVuf" ",
16729 s, RE_SV_TAIL(r->float_substr),
16730 (IV)r->float_min_offset, (UV)r->float_max_offset);
16731 } else if (r->float_utf8) {
16732 RE_PV_QUOTED_DECL(s, 1, dsv, SvPVX_const(r->float_utf8),
16733 RE_SV_DUMPLEN(r->float_utf8), 30);
16734 PerlIO_printf(Perl_debug_log,
16735 "floating utf8 %s%s at %"IVdf"..%"UVuf" ",
16736 s, RE_SV_TAIL(r->float_utf8),
16737 (IV)r->float_min_offset, (UV)r->float_max_offset);
16739 if (r->check_substr || r->check_utf8)
16740 PerlIO_printf(Perl_debug_log,
16742 (r->check_substr == r->float_substr
16743 && r->check_utf8 == r->float_utf8
16744 ? "(checking floating" : "(checking anchored"));
16745 if (r->intflags & PREGf_NOSCAN)
16746 PerlIO_printf(Perl_debug_log, " noscan");
16747 if (r->extflags & RXf_CHECK_ALL)
16748 PerlIO_printf(Perl_debug_log, " isall");
16749 if (r->check_substr || r->check_utf8)
16750 PerlIO_printf(Perl_debug_log, ") ");
16752 if (ri->regstclass) {
16753 regprop(r, sv, ri->regstclass, NULL, NULL);
16754 PerlIO_printf(Perl_debug_log, "stclass %s ", SvPVX_const(sv));
16756 if (r->intflags & PREGf_ANCH) {
16757 PerlIO_printf(Perl_debug_log, "anchored");
16758 if (r->intflags & PREGf_ANCH_MBOL)
16759 PerlIO_printf(Perl_debug_log, "(MBOL)");
16760 if (r->intflags & PREGf_ANCH_SBOL)
16761 PerlIO_printf(Perl_debug_log, "(SBOL)");
16762 if (r->intflags & PREGf_ANCH_GPOS)
16763 PerlIO_printf(Perl_debug_log, "(GPOS)");
16764 PerlIO_putc(Perl_debug_log, ' ');
16766 if (r->intflags & PREGf_GPOS_SEEN)
16767 PerlIO_printf(Perl_debug_log, "GPOS:%"UVuf" ", (UV)r->gofs);
16768 if (r->intflags & PREGf_SKIP)
16769 PerlIO_printf(Perl_debug_log, "plus ");
16770 if (r->intflags & PREGf_IMPLICIT)
16771 PerlIO_printf(Perl_debug_log, "implicit ");
16772 PerlIO_printf(Perl_debug_log, "minlen %"IVdf" ", (IV)r->minlen);
16773 if (r->extflags & RXf_EVAL_SEEN)
16774 PerlIO_printf(Perl_debug_log, "with eval ");
16775 PerlIO_printf(Perl_debug_log, "\n");
16777 regdump_extflags("r->extflags: ",r->extflags);
16778 regdump_intflags("r->intflags: ",r->intflags);
16781 PERL_ARGS_ASSERT_REGDUMP;
16782 PERL_UNUSED_CONTEXT;
16783 PERL_UNUSED_ARG(r);
16784 #endif /* DEBUGGING */
16788 - regprop - printable representation of opcode, with run time support
16792 Perl_regprop(pTHX_ const regexp *prog, SV *sv, const regnode *o, const regmatch_info *reginfo, const RExC_state_t *pRExC_state)
16797 /* Should be synchronized with * ANYOF_ #xdefines in regcomp.h */
16798 static const char * const anyofs[] = {
16799 #if _CC_WORDCHAR != 0 || _CC_DIGIT != 1 || _CC_ALPHA != 2 || _CC_LOWER != 3 \
16800 || _CC_UPPER != 4 || _CC_PUNCT != 5 || _CC_PRINT != 6 \
16801 || _CC_ALPHANUMERIC != 7 || _CC_GRAPH != 8 || _CC_CASED != 9 \
16802 || _CC_SPACE != 10 || _CC_BLANK != 11 || _CC_XDIGIT != 12 \
16803 || _CC_CNTRL != 13 || _CC_ASCII != 14 || _CC_VERTSPACE != 15
16804 #error Need to adjust order of anyofs[]
16839 RXi_GET_DECL(prog,progi);
16840 GET_RE_DEBUG_FLAGS_DECL;
16842 PERL_ARGS_ASSERT_REGPROP;
16844 sv_setpvn(sv, "", 0);
16846 if (OP(o) > REGNODE_MAX) /* regnode.type is unsigned */
16847 /* It would be nice to FAIL() here, but this may be called from
16848 regexec.c, and it would be hard to supply pRExC_state. */
16849 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
16850 (int)OP(o), (int)REGNODE_MAX);
16851 sv_catpv(sv, PL_reg_name[OP(o)]); /* Take off const! */
16853 k = PL_regkind[OP(o)];
16856 sv_catpvs(sv, " ");
16857 /* Using is_utf8_string() (via PERL_PV_UNI_DETECT)
16858 * is a crude hack but it may be the best for now since
16859 * we have no flag "this EXACTish node was UTF-8"
16861 pv_pretty(sv, STRING(o), STR_LEN(o), 60, PL_colors[0], PL_colors[1],
16862 PERL_PV_ESCAPE_UNI_DETECT |
16863 PERL_PV_ESCAPE_NONASCII |
16864 PERL_PV_PRETTY_ELLIPSES |
16865 PERL_PV_PRETTY_LTGT |
16866 PERL_PV_PRETTY_NOCLEAR
16868 } else if (k == TRIE) {
16869 /* print the details of the trie in dumpuntil instead, as
16870 * progi->data isn't available here */
16871 const char op = OP(o);
16872 const U32 n = ARG(o);
16873 const reg_ac_data * const ac = IS_TRIE_AC(op) ?
16874 (reg_ac_data *)progi->data->data[n] :
16876 const reg_trie_data * const trie
16877 = (reg_trie_data*)progi->data->data[!IS_TRIE_AC(op) ? n : ac->trie];
16879 Perl_sv_catpvf(aTHX_ sv, "-%s",PL_reg_name[o->flags]);
16880 DEBUG_TRIE_COMPILE_r(
16881 Perl_sv_catpvf(aTHX_ sv,
16882 "<S:%"UVuf"/%"IVdf" W:%"UVuf" L:%"UVuf"/%"UVuf" C:%"UVuf"/%"UVuf">",
16883 (UV)trie->startstate,
16884 (IV)trie->statecount-1, /* -1 because of the unused 0 element */
16885 (UV)trie->wordcount,
16888 (UV)TRIE_CHARCOUNT(trie),
16889 (UV)trie->uniquecharcount
16892 if ( IS_ANYOF_TRIE(op) || trie->bitmap ) {
16893 sv_catpvs(sv, "[");
16894 (void) put_charclass_bitmap_innards(sv,
16895 (IS_ANYOF_TRIE(op))
16897 : TRIE_BITMAP(trie),
16899 sv_catpvs(sv, "]");
16902 } else if (k == CURLY) {
16903 if (OP(o) == CURLYM || OP(o) == CURLYN || OP(o) == CURLYX)
16904 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags); /* Parenth number */
16905 Perl_sv_catpvf(aTHX_ sv, " {%d,%d}", ARG1(o), ARG2(o));
16907 else if (k == WHILEM && o->flags) /* Ordinal/of */
16908 Perl_sv_catpvf(aTHX_ sv, "[%d/%d]", o->flags & 0xf, o->flags>>4);
16909 else if (k == REF || k == OPEN || k == CLOSE
16910 || k == GROUPP || OP(o)==ACCEPT)
16912 AV *name_list= NULL;
16913 Perl_sv_catpvf(aTHX_ sv, "%d", (int)ARG(o)); /* Parenth number */
16914 if ( RXp_PAREN_NAMES(prog) ) {
16915 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16916 } else if ( pRExC_state ) {
16917 name_list= RExC_paren_name_list;
16920 if ( k != REF || (OP(o) < NREF)) {
16921 SV **name= av_fetch(name_list, ARG(o), 0 );
16923 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16926 SV *sv_dat= MUTABLE_SV(progi->data->data[ ARG( o ) ]);
16927 I32 *nums=(I32*)SvPVX(sv_dat);
16928 SV **name= av_fetch(name_list, nums[0], 0 );
16931 for ( n=0; n<SvIVX(sv_dat); n++ ) {
16932 Perl_sv_catpvf(aTHX_ sv, "%s%"IVdf,
16933 (n ? "," : ""), (IV)nums[n]);
16935 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16939 if ( k == REF && reginfo) {
16940 U32 n = ARG(o); /* which paren pair */
16941 I32 ln = prog->offs[n].start;
16942 if (prog->lastparen < n || ln == -1)
16943 Perl_sv_catpvf(aTHX_ sv, ": FAIL");
16944 else if (ln == prog->offs[n].end)
16945 Perl_sv_catpvf(aTHX_ sv, ": ACCEPT - EMPTY STRING");
16947 const char *s = reginfo->strbeg + ln;
16948 Perl_sv_catpvf(aTHX_ sv, ": ");
16949 Perl_pv_pretty( aTHX_ sv, s, prog->offs[n].end - prog->offs[n].start, 32, 0, 0,
16950 PERL_PV_ESCAPE_UNI_DETECT|PERL_PV_PRETTY_NOCLEAR|PERL_PV_PRETTY_ELLIPSES|PERL_PV_PRETTY_QUOTE );
16953 } else if (k == GOSUB) {
16954 AV *name_list= NULL;
16955 if ( RXp_PAREN_NAMES(prog) ) {
16956 name_list= MUTABLE_AV(progi->data->data[progi->name_list_idx]);
16957 } else if ( pRExC_state ) {
16958 name_list= RExC_paren_name_list;
16961 /* Paren and offset */
16962 Perl_sv_catpvf(aTHX_ sv, "%d[%+d]", (int)ARG(o),(int)ARG2L(o));
16964 SV **name= av_fetch(name_list, ARG(o), 0 );
16966 Perl_sv_catpvf(aTHX_ sv, " '%"SVf"'", SVfARG(*name));
16969 else if (k == VERB) {
16971 Perl_sv_catpvf(aTHX_ sv, ":%"SVf,
16972 SVfARG((MUTABLE_SV(progi->data->data[ ARG( o ) ]))));
16973 } else if (k == LOGICAL)
16974 /* 2: embedded, otherwise 1 */
16975 Perl_sv_catpvf(aTHX_ sv, "[%d]", o->flags);
16976 else if (k == ANYOF) {
16977 const U8 flags = ANYOF_FLAGS(o);
16979 SV* bitmap_invlist; /* Will hold what the bit map contains */
16982 if (OP(o) == ANYOFL)
16983 sv_catpvs(sv, "{loc}");
16984 if (flags & ANYOF_LOC_FOLD)
16985 sv_catpvs(sv, "{i}");
16986 Perl_sv_catpvf(aTHX_ sv, "[%s", PL_colors[0]);
16987 if (flags & ANYOF_INVERT)
16988 sv_catpvs(sv, "^");
16990 /* output what the standard cp 0-NUM_ANYOF_CODE_POINTS-1 bitmap matches
16992 do_sep = put_charclass_bitmap_innards(sv, ANYOF_BITMAP(o),
16995 /* output any special charclass tests (used entirely under use
16997 if (ANYOF_POSIXL_TEST_ANY_SET(o)) {
16999 for (i = 0; i < ANYOF_POSIXL_MAX; i++) {
17000 if (ANYOF_POSIXL_TEST(o,i)) {
17001 sv_catpv(sv, anyofs[i]);
17007 if ((flags & (ANYOF_MATCHES_ALL_ABOVE_BITMAP
17008 |ANYOF_HAS_UTF8_NONBITMAP_MATCHES
17009 |ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES
17013 Perl_sv_catpvf(aTHX_ sv,"%s][%s",PL_colors[1],PL_colors[0]);
17014 if (flags & ANYOF_INVERT)
17015 /*make sure the invert info is in each */
17016 sv_catpvs(sv, "^");
17019 if (flags & ANYOF_MATCHES_ALL_NON_UTF8_NON_ASCII) {
17020 sv_catpvs(sv, "{non-utf8-latin1-all}");
17023 if (flags & ANYOF_MATCHES_ALL_ABOVE_BITMAP)
17024 sv_catpvs(sv, "{above_bitmap_all}");
17026 if (ARG(o) != ANYOF_ONLY_HAS_BITMAP) {
17027 SV *lv; /* Set if there is something outside the bit map. */
17028 bool byte_output = FALSE; /* If something has been output */
17029 SV *only_utf8_locale;
17031 /* Get the stuff that wasn't in the bitmap. 'bitmap_invlist'
17032 * is used to guarantee that nothing in the bitmap gets
17034 (void) _get_regclass_nonbitmap_data(prog, o, FALSE,
17035 &lv, &only_utf8_locale,
17037 if (lv && lv != &PL_sv_undef) {
17038 char *s = savesvpv(lv);
17039 char * const origs = s;
17041 while (*s && *s != '\n')
17045 const char * const t = ++s;
17047 if (flags & ANYOF_HAS_NONBITMAP_NON_UTF8_MATCHES) {
17048 sv_catpvs(sv, "{outside bitmap}");
17051 sv_catpvs(sv, "{utf8}");
17055 sv_catpvs(sv, " ");
17061 /* Truncate very long output */
17062 if (s - origs > 256) {
17063 Perl_sv_catpvf(aTHX_ sv,
17065 (int) (s - origs - 1),
17071 else if (*s == '\t') {
17085 SvREFCNT_dec_NN(lv);
17088 if ((flags & ANYOF_LOC_FOLD)
17089 && only_utf8_locale
17090 && only_utf8_locale != &PL_sv_undef)
17093 int max_entries = 256;
17095 sv_catpvs(sv, "{utf8 locale}");
17096 invlist_iterinit(only_utf8_locale);
17097 while (invlist_iternext(only_utf8_locale,
17099 put_range(sv, start, end, FALSE);
17101 if (max_entries < 0) {
17102 sv_catpvs(sv, "...");
17106 invlist_iterfinish(only_utf8_locale);
17110 SvREFCNT_dec(bitmap_invlist);
17113 Perl_sv_catpvf(aTHX_ sv, "%s]", PL_colors[1]);
17115 else if (k == POSIXD || k == NPOSIXD) {
17116 U8 index = FLAGS(o) * 2;
17117 if (index < C_ARRAY_LENGTH(anyofs)) {
17118 if (*anyofs[index] != '[') {
17121 sv_catpv(sv, anyofs[index]);
17122 if (*anyofs[index] != '[') {
17127 Perl_sv_catpvf(aTHX_ sv, "[illegal type=%d])", index);
17130 else if (k == BOUND || k == NBOUND) {
17131 /* Must be synced with order of 'bound_type' in regcomp.h */
17132 const char * const bounds[] = {
17133 "", /* Traditional */
17138 sv_catpv(sv, bounds[FLAGS(o)]);
17140 else if (k == BRANCHJ && (OP(o) == UNLESSM || OP(o) == IFMATCH))
17141 Perl_sv_catpvf(aTHX_ sv, "[%d]", -(o->flags));
17142 else if (OP(o) == SBOL)
17143 Perl_sv_catpvf(aTHX_ sv, " /%s/", o->flags ? "\\A" : "^");
17145 PERL_UNUSED_CONTEXT;
17146 PERL_UNUSED_ARG(sv);
17147 PERL_UNUSED_ARG(o);
17148 PERL_UNUSED_ARG(prog);
17149 PERL_UNUSED_ARG(reginfo);
17150 PERL_UNUSED_ARG(pRExC_state);
17151 #endif /* DEBUGGING */
17157 Perl_re_intuit_string(pTHX_ REGEXP * const r)
17158 { /* Assume that RE_INTUIT is set */
17159 struct regexp *const prog = ReANY(r);
17160 GET_RE_DEBUG_FLAGS_DECL;
17162 PERL_ARGS_ASSERT_RE_INTUIT_STRING;
17163 PERL_UNUSED_CONTEXT;
17167 const char * const s = SvPV_nolen_const(RX_UTF8(r)
17168 ? prog->check_utf8 : prog->check_substr);
17170 if (!PL_colorset) reginitcolors();
17171 PerlIO_printf(Perl_debug_log,
17172 "%sUsing REx %ssubstr:%s \"%s%.60s%s%s\"\n",
17174 RX_UTF8(r) ? "utf8 " : "",
17175 PL_colors[5],PL_colors[0],
17178 (strlen(s) > 60 ? "..." : ""));
17181 /* use UTF8 check substring if regexp pattern itself is in UTF8 */
17182 return RX_UTF8(r) ? prog->check_utf8 : prog->check_substr;
17188 handles refcounting and freeing the perl core regexp structure. When
17189 it is necessary to actually free the structure the first thing it
17190 does is call the 'free' method of the regexp_engine associated to
17191 the regexp, allowing the handling of the void *pprivate; member
17192 first. (This routine is not overridable by extensions, which is why
17193 the extensions free is called first.)
17195 See regdupe and regdupe_internal if you change anything here.
17197 #ifndef PERL_IN_XSUB_RE
17199 Perl_pregfree(pTHX_ REGEXP *r)
17205 Perl_pregfree2(pTHX_ REGEXP *rx)
17207 struct regexp *const r = ReANY(rx);
17208 GET_RE_DEBUG_FLAGS_DECL;
17210 PERL_ARGS_ASSERT_PREGFREE2;
17212 if (r->mother_re) {
17213 ReREFCNT_dec(r->mother_re);
17215 CALLREGFREE_PVT(rx); /* free the private data */
17216 SvREFCNT_dec(RXp_PAREN_NAMES(r));
17217 Safefree(r->xpv_len_u.xpvlenu_pv);
17220 SvREFCNT_dec(r->anchored_substr);
17221 SvREFCNT_dec(r->anchored_utf8);
17222 SvREFCNT_dec(r->float_substr);
17223 SvREFCNT_dec(r->float_utf8);
17224 Safefree(r->substrs);
17226 RX_MATCH_COPY_FREE(rx);
17227 #ifdef PERL_ANY_COW
17228 SvREFCNT_dec(r->saved_copy);
17231 SvREFCNT_dec(r->qr_anoncv);
17232 rx->sv_u.svu_rx = 0;
17237 This is a hacky workaround to the structural issue of match results
17238 being stored in the regexp structure which is in turn stored in
17239 PL_curpm/PL_reg_curpm. The problem is that due to qr// the pattern
17240 could be PL_curpm in multiple contexts, and could require multiple
17241 result sets being associated with the pattern simultaneously, such
17242 as when doing a recursive match with (??{$qr})
17244 The solution is to make a lightweight copy of the regexp structure
17245 when a qr// is returned from the code executed by (??{$qr}) this
17246 lightweight copy doesn't actually own any of its data except for
17247 the starp/end and the actual regexp structure itself.
17253 Perl_reg_temp_copy (pTHX_ REGEXP *ret_x, REGEXP *rx)
17255 struct regexp *ret;
17256 struct regexp *const r = ReANY(rx);
17257 const bool islv = ret_x && SvTYPE(ret_x) == SVt_PVLV;
17259 PERL_ARGS_ASSERT_REG_TEMP_COPY;
17262 ret_x = (REGEXP*) newSV_type(SVt_REGEXP);
17264 SvOK_off((SV *)ret_x);
17266 /* For PVLVs, SvANY points to the xpvlv body while sv_u points
17267 to the regexp. (For SVt_REGEXPs, sv_upgrade has already
17268 made both spots point to the same regexp body.) */
17269 REGEXP *temp = (REGEXP *)newSV_type(SVt_REGEXP);
17270 assert(!SvPVX(ret_x));
17271 ret_x->sv_u.svu_rx = temp->sv_any;
17272 temp->sv_any = NULL;
17273 SvFLAGS(temp) = (SvFLAGS(temp) & ~SVTYPEMASK) | SVt_NULL;
17274 SvREFCNT_dec_NN(temp);
17275 /* SvCUR still resides in the xpvlv struct, so the regexp copy-
17276 ing below will not set it. */
17277 SvCUR_set(ret_x, SvCUR(rx));
17280 /* This ensures that SvTHINKFIRST(sv) is true, and hence that
17281 sv_force_normal(sv) is called. */
17283 ret = ReANY(ret_x);
17285 SvFLAGS(ret_x) |= SvUTF8(rx);
17286 /* We share the same string buffer as the original regexp, on which we
17287 hold a reference count, incremented when mother_re is set below.
17288 The string pointer is copied here, being part of the regexp struct.
17290 memcpy(&(ret->xpv_cur), &(r->xpv_cur),
17291 sizeof(regexp) - STRUCT_OFFSET(regexp, xpv_cur));
17293 const I32 npar = r->nparens+1;
17294 Newx(ret->offs, npar, regexp_paren_pair);
17295 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17298 Newx(ret->substrs, 1, struct reg_substr_data);
17299 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17301 SvREFCNT_inc_void(ret->anchored_substr);
17302 SvREFCNT_inc_void(ret->anchored_utf8);
17303 SvREFCNT_inc_void(ret->float_substr);
17304 SvREFCNT_inc_void(ret->float_utf8);
17306 /* check_substr and check_utf8, if non-NULL, point to either their
17307 anchored or float namesakes, and don't hold a second reference. */
17309 RX_MATCH_COPIED_off(ret_x);
17310 #ifdef PERL_ANY_COW
17311 ret->saved_copy = NULL;
17313 ret->mother_re = ReREFCNT_inc(r->mother_re ? r->mother_re : rx);
17314 SvREFCNT_inc_void(ret->qr_anoncv);
17320 /* regfree_internal()
17322 Free the private data in a regexp. This is overloadable by
17323 extensions. Perl takes care of the regexp structure in pregfree(),
17324 this covers the *pprivate pointer which technically perl doesn't
17325 know about, however of course we have to handle the
17326 regexp_internal structure when no extension is in use.
17328 Note this is called before freeing anything in the regexp
17333 Perl_regfree_internal(pTHX_ REGEXP * const rx)
17335 struct regexp *const r = ReANY(rx);
17336 RXi_GET_DECL(r,ri);
17337 GET_RE_DEBUG_FLAGS_DECL;
17339 PERL_ARGS_ASSERT_REGFREE_INTERNAL;
17345 SV *dsv= sv_newmortal();
17346 RE_PV_QUOTED_DECL(s, RX_UTF8(rx),
17347 dsv, RX_PRECOMP(rx), RX_PRELEN(rx), 60);
17348 PerlIO_printf(Perl_debug_log,"%sFreeing REx:%s %s\n",
17349 PL_colors[4],PL_colors[5],s);
17352 #ifdef RE_TRACK_PATTERN_OFFSETS
17354 Safefree(ri->u.offsets); /* 20010421 MJD */
17356 if (ri->code_blocks) {
17358 for (n = 0; n < ri->num_code_blocks; n++)
17359 SvREFCNT_dec(ri->code_blocks[n].src_regex);
17360 Safefree(ri->code_blocks);
17364 int n = ri->data->count;
17367 /* If you add a ->what type here, update the comment in regcomp.h */
17368 switch (ri->data->what[n]) {
17374 SvREFCNT_dec(MUTABLE_SV(ri->data->data[n]));
17377 Safefree(ri->data->data[n]);
17383 { /* Aho Corasick add-on structure for a trie node.
17384 Used in stclass optimization only */
17386 reg_ac_data *aho=(reg_ac_data*)ri->data->data[n];
17387 #ifdef USE_ITHREADS
17391 refcount = --aho->refcount;
17394 PerlMemShared_free(aho->states);
17395 PerlMemShared_free(aho->fail);
17396 /* do this last!!!! */
17397 PerlMemShared_free(ri->data->data[n]);
17398 /* we should only ever get called once, so
17399 * assert as much, and also guard the free
17400 * which /might/ happen twice. At the least
17401 * it will make code anlyzers happy and it
17402 * doesn't cost much. - Yves */
17403 assert(ri->regstclass);
17404 if (ri->regstclass) {
17405 PerlMemShared_free(ri->regstclass);
17406 ri->regstclass = 0;
17413 /* trie structure. */
17415 reg_trie_data *trie=(reg_trie_data*)ri->data->data[n];
17416 #ifdef USE_ITHREADS
17420 refcount = --trie->refcount;
17423 PerlMemShared_free(trie->charmap);
17424 PerlMemShared_free(trie->states);
17425 PerlMemShared_free(trie->trans);
17427 PerlMemShared_free(trie->bitmap);
17429 PerlMemShared_free(trie->jump);
17430 PerlMemShared_free(trie->wordinfo);
17431 /* do this last!!!! */
17432 PerlMemShared_free(ri->data->data[n]);
17437 Perl_croak(aTHX_ "panic: regfree data code '%c'",
17438 ri->data->what[n]);
17441 Safefree(ri->data->what);
17442 Safefree(ri->data);
17448 #define av_dup_inc(s,t) MUTABLE_AV(sv_dup_inc((const SV *)s,t))
17449 #define hv_dup_inc(s,t) MUTABLE_HV(sv_dup_inc((const SV *)s,t))
17450 #define SAVEPVN(p,n) ((p) ? savepvn(p,n) : NULL)
17453 re_dup - duplicate a regexp.
17455 This routine is expected to clone a given regexp structure. It is only
17456 compiled under USE_ITHREADS.
17458 After all of the core data stored in struct regexp is duplicated
17459 the regexp_engine.dupe method is used to copy any private data
17460 stored in the *pprivate pointer. This allows extensions to handle
17461 any duplication it needs to do.
17463 See pregfree() and regfree_internal() if you change anything here.
17465 #if defined(USE_ITHREADS)
17466 #ifndef PERL_IN_XSUB_RE
17468 Perl_re_dup_guts(pTHX_ const REGEXP *sstr, REGEXP *dstr, CLONE_PARAMS *param)
17472 const struct regexp *r = ReANY(sstr);
17473 struct regexp *ret = ReANY(dstr);
17475 PERL_ARGS_ASSERT_RE_DUP_GUTS;
17477 npar = r->nparens+1;
17478 Newx(ret->offs, npar, regexp_paren_pair);
17479 Copy(r->offs, ret->offs, npar, regexp_paren_pair);
17481 if (ret->substrs) {
17482 /* Do it this way to avoid reading from *r after the StructCopy().
17483 That way, if any of the sv_dup_inc()s dislodge *r from the L1
17484 cache, it doesn't matter. */
17485 const bool anchored = r->check_substr
17486 ? r->check_substr == r->anchored_substr
17487 : r->check_utf8 == r->anchored_utf8;
17488 Newx(ret->substrs, 1, struct reg_substr_data);
17489 StructCopy(r->substrs, ret->substrs, struct reg_substr_data);
17491 ret->anchored_substr = sv_dup_inc(ret->anchored_substr, param);
17492 ret->anchored_utf8 = sv_dup_inc(ret->anchored_utf8, param);
17493 ret->float_substr = sv_dup_inc(ret->float_substr, param);
17494 ret->float_utf8 = sv_dup_inc(ret->float_utf8, param);
17496 /* check_substr and check_utf8, if non-NULL, point to either their
17497 anchored or float namesakes, and don't hold a second reference. */
17499 if (ret->check_substr) {
17501 assert(r->check_utf8 == r->anchored_utf8);
17502 ret->check_substr = ret->anchored_substr;
17503 ret->check_utf8 = ret->anchored_utf8;
17505 assert(r->check_substr == r->float_substr);
17506 assert(r->check_utf8 == r->float_utf8);
17507 ret->check_substr = ret->float_substr;
17508 ret->check_utf8 = ret->float_utf8;
17510 } else if (ret->check_utf8) {
17512 ret->check_utf8 = ret->anchored_utf8;
17514 ret->check_utf8 = ret->float_utf8;
17519 RXp_PAREN_NAMES(ret) = hv_dup_inc(RXp_PAREN_NAMES(ret), param);
17520 ret->qr_anoncv = MUTABLE_CV(sv_dup_inc((const SV *)ret->qr_anoncv, param));
17523 RXi_SET(ret,CALLREGDUPE_PVT(dstr,param));
17525 if (RX_MATCH_COPIED(dstr))
17526 ret->subbeg = SAVEPVN(ret->subbeg, ret->sublen);
17528 ret->subbeg = NULL;
17529 #ifdef PERL_ANY_COW
17530 ret->saved_copy = NULL;
17533 /* Whether mother_re be set or no, we need to copy the string. We
17534 cannot refrain from copying it when the storage points directly to
17535 our mother regexp, because that's
17536 1: a buffer in a different thread
17537 2: something we no longer hold a reference on
17538 so we need to copy it locally. */
17539 RX_WRAPPED(dstr) = SAVEPVN(RX_WRAPPED(sstr), SvCUR(sstr)+1);
17540 ret->mother_re = NULL;
17542 #endif /* PERL_IN_XSUB_RE */
17547 This is the internal complement to regdupe() which is used to copy
17548 the structure pointed to by the *pprivate pointer in the regexp.
17549 This is the core version of the extension overridable cloning hook.
17550 The regexp structure being duplicated will be copied by perl prior
17551 to this and will be provided as the regexp *r argument, however
17552 with the /old/ structures pprivate pointer value. Thus this routine
17553 may override any copying normally done by perl.
17555 It returns a pointer to the new regexp_internal structure.
17559 Perl_regdupe_internal(pTHX_ REGEXP * const rx, CLONE_PARAMS *param)
17562 struct regexp *const r = ReANY(rx);
17563 regexp_internal *reti;
17565 RXi_GET_DECL(r,ri);
17567 PERL_ARGS_ASSERT_REGDUPE_INTERNAL;
17571 Newxc(reti, sizeof(regexp_internal) + len*sizeof(regnode),
17572 char, regexp_internal);
17573 Copy(ri->program, reti->program, len+1, regnode);
17575 reti->num_code_blocks = ri->num_code_blocks;
17576 if (ri->code_blocks) {
17578 Newxc(reti->code_blocks, ri->num_code_blocks, struct reg_code_block,
17579 struct reg_code_block);
17580 Copy(ri->code_blocks, reti->code_blocks, ri->num_code_blocks,
17581 struct reg_code_block);
17582 for (n = 0; n < ri->num_code_blocks; n++)
17583 reti->code_blocks[n].src_regex = (REGEXP*)
17584 sv_dup_inc((SV*)(ri->code_blocks[n].src_regex), param);
17587 reti->code_blocks = NULL;
17589 reti->regstclass = NULL;
17592 struct reg_data *d;
17593 const int count = ri->data->count;
17596 Newxc(d, sizeof(struct reg_data) + count*sizeof(void *),
17597 char, struct reg_data);
17598 Newx(d->what, count, U8);
17601 for (i = 0; i < count; i++) {
17602 d->what[i] = ri->data->what[i];
17603 switch (d->what[i]) {
17604 /* see also regcomp.h and regfree_internal() */
17605 case 'a': /* actually an AV, but the dup function is identical. */
17609 case 'u': /* actually an HV, but the dup function is identical. */
17610 d->data[i] = sv_dup_inc((const SV *)ri->data->data[i], param);
17613 /* This is cheating. */
17614 Newx(d->data[i], 1, regnode_ssc);
17615 StructCopy(ri->data->data[i], d->data[i], regnode_ssc);
17616 reti->regstclass = (regnode*)d->data[i];
17619 /* Trie stclasses are readonly and can thus be shared
17620 * without duplication. We free the stclass in pregfree
17621 * when the corresponding reg_ac_data struct is freed.
17623 reti->regstclass= ri->regstclass;
17627 ((reg_trie_data*)ri->data->data[i])->refcount++;
17632 d->data[i] = ri->data->data[i];
17635 Perl_croak(aTHX_ "panic: re_dup unknown data code '%c'",
17636 ri->data->what[i]);
17645 reti->name_list_idx = ri->name_list_idx;
17647 #ifdef RE_TRACK_PATTERN_OFFSETS
17648 if (ri->u.offsets) {
17649 Newx(reti->u.offsets, 2*len+1, U32);
17650 Copy(ri->u.offsets, reti->u.offsets, 2*len+1, U32);
17653 SetProgLen(reti,len);
17656 return (void*)reti;
17659 #endif /* USE_ITHREADS */
17661 #ifndef PERL_IN_XSUB_RE
17664 - regnext - dig the "next" pointer out of a node
17667 Perl_regnext(pTHX_ regnode *p)
17674 if (OP(p) > REGNODE_MAX) { /* regnode.type is unsigned */
17675 Perl_croak(aTHX_ "Corrupted regexp opcode %d > %d",
17676 (int)OP(p), (int)REGNODE_MAX);
17679 offset = (reg_off_by_arg[OP(p)] ? ARG(p) : NEXT_OFF(p));
17688 S_re_croak2(pTHX_ bool utf8, const char* pat1,const char* pat2,...)
17691 STRLEN l1 = strlen(pat1);
17692 STRLEN l2 = strlen(pat2);
17695 const char *message;
17697 PERL_ARGS_ASSERT_RE_CROAK2;
17703 Copy(pat1, buf, l1 , char);
17704 Copy(pat2, buf + l1, l2 , char);
17705 buf[l1 + l2] = '\n';
17706 buf[l1 + l2 + 1] = '\0';
17707 va_start(args, pat2);
17708 msv = vmess(buf, &args);
17710 message = SvPV_const(msv,l1);
17713 Copy(message, buf, l1 , char);
17714 /* l1-1 to avoid \n */
17715 Perl_croak(aTHX_ "%"UTF8f, UTF8fARG(utf8, l1-1, buf));
17718 /* XXX Here's a total kludge. But we need to re-enter for swash routines. */
17720 #ifndef PERL_IN_XSUB_RE
17722 Perl_save_re_context(pTHX)
17727 /* Save $1..$n (#18107: UTF-8 s/(\w+)/uc($1)/e); AMS 20021106. */
17730 const REGEXP * const rx = PM_GETRE(PL_curpm);
17732 nparens = RX_NPARENS(rx);
17735 /* RT #124109. This is a complete hack; in the SWASHNEW case we know
17736 * that PL_curpm will be null, but that utf8.pm and the modules it
17737 * loads will only use $1..$3.
17738 * The t/porting/re_context.t test file checks this assumption.
17743 for (i = 1; i <= nparens; i++) {
17744 char digits[TYPE_CHARS(long)];
17745 const STRLEN len = my_snprintf(digits, sizeof(digits),
17747 GV *const *const gvp
17748 = (GV**)hv_fetch(PL_defstash, digits, len, 0);
17751 GV * const gv = *gvp;
17752 if (SvTYPE(gv) == SVt_PVGV && GvSV(gv))
17762 S_put_code_point(pTHX_ SV *sv, UV c)
17764 PERL_ARGS_ASSERT_PUT_CODE_POINT;
17767 Perl_sv_catpvf(aTHX_ sv, "\\x{%04"UVXf"}", c);
17769 else if (isPRINT(c)) {
17770 const char string = (char) c;
17771 if (isBACKSLASHED_PUNCT(c))
17772 sv_catpvs(sv, "\\");
17773 sv_catpvn(sv, &string, 1);
17776 const char * const mnemonic = cntrl_to_mnemonic((char) c);
17778 Perl_sv_catpvf(aTHX_ sv, "%s", mnemonic);
17781 Perl_sv_catpvf(aTHX_ sv, "\\x{%02X}", (U8) c);
17786 #define MAX_PRINT_A MAX_PRINT_A_FOR_USE_ONLY_BY_REGCOMP_DOT_C
17789 S_put_range(pTHX_ SV *sv, UV start, const UV end, const bool allow_literals)
17791 /* Appends to 'sv' a displayable version of the range of code points from
17792 * 'start' to 'end'. It assumes that only ASCII printables are displayable
17793 * as-is (though some of these will be escaped by put_code_point()). */
17795 const unsigned int min_range_count = 3;
17797 assert(start <= end);
17799 PERL_ARGS_ASSERT_PUT_RANGE;
17801 while (start <= end) {
17803 const char * format;
17805 if (end - start < min_range_count) {
17807 /* Individual chars in short ranges */
17808 for (; start <= end; start++) {
17809 put_code_point(sv, start);
17814 /* If permitted by the input options, and there is a possibility that
17815 * this range contains a printable literal, look to see if there is
17817 if (allow_literals && start <= MAX_PRINT_A) {
17819 /* If the range begin isn't an ASCII printable, effectively split
17820 * the range into two parts:
17821 * 1) the portion before the first such printable,
17823 * and output them separately. */
17824 if (! isPRINT_A(start)) {
17825 UV temp_end = start + 1;
17827 /* There is no point looking beyond the final possible
17828 * printable, in MAX_PRINT_A */
17829 UV max = MIN(end, MAX_PRINT_A);
17831 while (temp_end <= max && ! isPRINT_A(temp_end)) {
17835 /* Here, temp_end points to one beyond the first printable if
17836 * found, or to one beyond 'max' if not. If none found, make
17837 * sure that we use the entire range */
17838 if (temp_end > MAX_PRINT_A) {
17839 temp_end = end + 1;
17842 /* Output the first part of the split range, the part that
17843 * doesn't have printables, with no looking for literals
17844 * (otherwise we would infinitely recurse) */
17845 put_range(sv, start, temp_end - 1, FALSE);
17847 /* The 2nd part of the range (if any) starts here. */
17850 /* We continue instead of dropping down because even if the 2nd
17851 * part is non-empty, it could be so short that we want to
17852 * output it specially, as tested for at the top of this loop.
17857 /* Here, 'start' is a printable ASCII. If it is an alphanumeric,
17858 * output a sub-range of just the digits or letters, then process
17859 * the remaining portion as usual. */
17860 if (isALPHANUMERIC_A(start)) {
17861 UV mask = (isDIGIT_A(start))
17866 UV temp_end = start + 1;
17868 /* Find the end of the sub-range that includes just the
17869 * characters in the same class as the first character in it */
17870 while (temp_end <= end && _generic_isCC_A(temp_end, mask)) {
17875 /* For short ranges, don't duplicate the code above to output
17876 * them; just call recursively */
17877 if (temp_end - start < min_range_count) {
17878 put_range(sv, start, temp_end, FALSE);
17880 else { /* Output as a range */
17881 put_code_point(sv, start);
17882 sv_catpvs(sv, "-");
17883 put_code_point(sv, temp_end);
17885 start = temp_end + 1;
17889 /* We output any other printables as individual characters */
17890 if (isPUNCT_A(start) || isSPACE_A(start)) {
17891 while (start <= end && (isPUNCT_A(start)
17892 || isSPACE_A(start)))
17894 put_code_point(sv, start);
17899 } /* End of looking for literals */
17901 /* Here is not to output as a literal. Some control characters have
17902 * mnemonic names. Split off any of those at the beginning and end of
17903 * the range to print mnemonically. It isn't possible for many of
17904 * these to be in a row, so this won't overwhelm with output */
17905 while (isMNEMONIC_CNTRL(start) && start <= end) {
17906 put_code_point(sv, start);
17909 if (start < end && isMNEMONIC_CNTRL(end)) {
17911 /* Here, the final character in the range has a mnemonic name.
17912 * Work backwards from the end to find the final non-mnemonic */
17913 UV temp_end = end - 1;
17914 while (isMNEMONIC_CNTRL(temp_end)) {
17918 /* And separately output the range that doesn't have mnemonics */
17919 put_range(sv, start, temp_end, FALSE);
17921 /* Then output the mnemonic trailing controls */
17922 start = temp_end + 1;
17923 while (start <= end) {
17924 put_code_point(sv, start);
17930 /* As a final resort, output the range or subrange as hex. */
17932 this_end = (end < NUM_ANYOF_CODE_POINTS)
17934 : NUM_ANYOF_CODE_POINTS - 1;
17935 format = (this_end < 256)
17936 ? "\\x{%02"UVXf"}-\\x{%02"UVXf"}"
17937 : "\\x{%04"UVXf"}-\\x{%04"UVXf"}";
17938 GCC_DIAG_IGNORE(-Wformat-nonliteral);
17939 Perl_sv_catpvf(aTHX_ sv, format, start, this_end);
17946 S_put_charclass_bitmap_innards(pTHX_ SV *sv, char *bitmap, SV** bitmap_invlist)
17948 /* Appends to 'sv' a displayable version of the innards of the bracketed
17949 * character class whose bitmap is 'bitmap'; Returns 'TRUE' if it actually
17950 * output anything, and bitmap_invlist, if not NULL, will point to an
17951 * inversion list of what is in the bit map */
17955 unsigned int punct_count = 0;
17956 SV* invlist = NULL;
17957 SV** invlist_ptr; /* Temporary, in case bitmap_invlist is NULL */
17958 bool allow_literals = TRUE;
17960 PERL_ARGS_ASSERT_PUT_CHARCLASS_BITMAP_INNARDS;
17962 invlist_ptr = (bitmap_invlist) ? bitmap_invlist : &invlist;
17964 /* Worst case is exactly every-other code point is in the list */
17965 *invlist_ptr = _new_invlist(NUM_ANYOF_CODE_POINTS / 2);
17967 /* Convert the bit map to an inversion list, keeping track of how many
17968 * ASCII puncts are set, including an extra amount for the backslashed
17970 for (i = 0; i < NUM_ANYOF_CODE_POINTS; i++) {
17971 if (BITMAP_TEST(bitmap, i)) {
17972 *invlist_ptr = add_cp_to_invlist(*invlist_ptr, i);
17973 if (isPUNCT_A(i)) {
17975 if isBACKSLASHED_PUNCT(i) {
17982 /* Nothing to output */
17983 if (_invlist_len(*invlist_ptr) == 0) {
17984 SvREFCNT_dec(invlist);
17988 /* Generally, it is more readable if printable characters are output as
17989 * literals, but if a range (nearly) spans all of them, it's best to output
17990 * it as a single range. This code will use a single range if all but 2
17991 * printables are in it */
17992 invlist_iterinit(*invlist_ptr);
17993 while (invlist_iternext(*invlist_ptr, &start, &end)) {
17995 /* If range starts beyond final printable, it doesn't have any in it */
17996 if (start > MAX_PRINT_A) {
18000 /* In both ASCII and EBCDIC, a SPACE is the lowest printable. To span
18001 * all but two, the range must start and end no later than 2 from
18003 if (start < ' ' + 2 && end > MAX_PRINT_A - 2) {
18004 if (end > MAX_PRINT_A) {
18010 if (end - start >= MAX_PRINT_A - ' ' - 2) {
18011 allow_literals = FALSE;
18016 invlist_iterfinish(*invlist_ptr);
18018 /* The legibility of the output depends mostly on how many punctuation
18019 * characters are output. There are 32 possible ASCII ones, and some have
18020 * an additional backslash, bringing it to currently 36, so if any more
18021 * than 18 are to be output, we can instead output it as its complement,
18022 * yielding fewer puncts, and making it more legible. But give some weight
18023 * to the fact that outputting it as a complement is less legible than a
18024 * straight output, so don't complement unless we are somewhat over the 18
18026 if (allow_literals && punct_count > 22) {
18027 sv_catpvs(sv, "^");
18029 /* Add everything remaining to the list, so when we invert it just
18030 * below, it will be excluded */
18031 _invlist_union_complement_2nd(*invlist_ptr, PL_InBitmap, invlist_ptr);
18032 _invlist_invert(*invlist_ptr);
18035 /* Here we have figured things out. Output each range */
18036 invlist_iterinit(*invlist_ptr);
18037 while (invlist_iternext(*invlist_ptr, &start, &end)) {
18038 if (start >= NUM_ANYOF_CODE_POINTS) {
18041 put_range(sv, start, end, allow_literals);
18043 invlist_iterfinish(*invlist_ptr);
18048 #define CLEAR_OPTSTART \
18049 if (optstart) STMT_START { \
18050 DEBUG_OPTIMISE_r(PerlIO_printf(Perl_debug_log, \
18051 " (%"IVdf" nodes)\n", (IV)(node - optstart))); \
18055 #define DUMPUNTIL(b,e) \
18057 node=dumpuntil(r,start,(b),(e),last,sv,indent+1,depth+1);
18059 STATIC const regnode *
18060 S_dumpuntil(pTHX_ const regexp *r, const regnode *start, const regnode *node,
18061 const regnode *last, const regnode *plast,
18062 SV* sv, I32 indent, U32 depth)
18064 U8 op = PSEUDO; /* Arbitrary non-END op. */
18065 const regnode *next;
18066 const regnode *optstart= NULL;
18068 RXi_GET_DECL(r,ri);
18069 GET_RE_DEBUG_FLAGS_DECL;
18071 PERL_ARGS_ASSERT_DUMPUNTIL;
18073 #ifdef DEBUG_DUMPUNTIL
18074 PerlIO_printf(Perl_debug_log, "--- %d : %d - %d - %d\n",indent,node-start,
18075 last ? last-start : 0,plast ? plast-start : 0);
18078 if (plast && plast < last)
18081 while (PL_regkind[op] != END && (!last || node < last)) {
18083 /* While that wasn't END last time... */
18086 if (op == CLOSE || op == WHILEM)
18088 next = regnext((regnode *)node);
18091 if (OP(node) == OPTIMIZED) {
18092 if (!optstart && RE_DEBUG_FLAG(RE_DEBUG_COMPILE_OPTIMISE))
18099 regprop(r, sv, node, NULL, NULL);
18100 PerlIO_printf(Perl_debug_log, "%4"IVdf":%*s%s", (IV)(node - start),
18101 (int)(2*indent + 1), "", SvPVX_const(sv));
18103 if (OP(node) != OPTIMIZED) {
18104 if (next == NULL) /* Next ptr. */
18105 PerlIO_printf(Perl_debug_log, " (0)");
18106 else if (PL_regkind[(U8)op] == BRANCH
18107 && PL_regkind[OP(next)] != BRANCH )
18108 PerlIO_printf(Perl_debug_log, " (FAIL)");
18110 PerlIO_printf(Perl_debug_log, " (%"IVdf")", (IV)(next - start));
18111 (void)PerlIO_putc(Perl_debug_log, '\n');
18115 if (PL_regkind[(U8)op] == BRANCHJ) {
18118 const regnode *nnode = (OP(next) == LONGJMP
18119 ? regnext((regnode *)next)
18121 if (last && nnode > last)
18123 DUMPUNTIL(NEXTOPER(NEXTOPER(node)), nnode);
18126 else if (PL_regkind[(U8)op] == BRANCH) {
18128 DUMPUNTIL(NEXTOPER(node), next);
18130 else if ( PL_regkind[(U8)op] == TRIE ) {
18131 const regnode *this_trie = node;
18132 const char op = OP(node);
18133 const U32 n = ARG(node);
18134 const reg_ac_data * const ac = op>=AHOCORASICK ?
18135 (reg_ac_data *)ri->data->data[n] :
18137 const reg_trie_data * const trie =
18138 (reg_trie_data*)ri->data->data[op<AHOCORASICK ? n : ac->trie];
18140 AV *const trie_words
18141 = MUTABLE_AV(ri->data->data[n + TRIE_WORDS_OFFSET]);
18143 const regnode *nextbranch= NULL;
18146 for (word_idx= 0; word_idx < (I32)trie->wordcount; word_idx++) {
18147 SV ** const elem_ptr = av_fetch(trie_words,word_idx,0);
18149 PerlIO_printf(Perl_debug_log, "%*s%s ",
18150 (int)(2*(indent+3)), "",
18152 ? pv_pretty(sv, SvPV_nolen_const(*elem_ptr),
18153 SvCUR(*elem_ptr), 60,
18154 PL_colors[0], PL_colors[1],
18156 ? PERL_PV_ESCAPE_UNI
18158 | PERL_PV_PRETTY_ELLIPSES
18159 | PERL_PV_PRETTY_LTGT
18164 U16 dist= trie->jump[word_idx+1];
18165 PerlIO_printf(Perl_debug_log, "(%"UVuf")\n",
18166 (UV)((dist ? this_trie + dist : next) - start));
18169 nextbranch= this_trie + trie->jump[0];
18170 DUMPUNTIL(this_trie + dist, nextbranch);
18172 if (nextbranch && PL_regkind[OP(nextbranch)]==BRANCH)
18173 nextbranch= regnext((regnode *)nextbranch);
18175 PerlIO_printf(Perl_debug_log, "\n");
18178 if (last && next > last)
18183 else if ( op == CURLY ) { /* "next" might be very big: optimizer */
18184 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS,
18185 NEXTOPER(node) + EXTRA_STEP_2ARGS + 1);
18187 else if (PL_regkind[(U8)op] == CURLY && op != CURLYX) {
18189 DUMPUNTIL(NEXTOPER(node) + EXTRA_STEP_2ARGS, next);
18191 else if ( op == PLUS || op == STAR) {
18192 DUMPUNTIL(NEXTOPER(node), NEXTOPER(node) + 1);
18194 else if (PL_regkind[(U8)op] == ANYOF) {
18195 /* arglen 1 + class block */
18196 node += 1 + ((ANYOF_FLAGS(node) & ANYOF_MATCHES_POSIXL)
18197 ? ANYOF_POSIXL_SKIP
18199 node = NEXTOPER(node);
18201 else if (PL_regkind[(U8)op] == EXACT) {
18202 /* Literal string, where present. */
18203 node += NODE_SZ_STR(node) - 1;
18204 node = NEXTOPER(node);
18207 node = NEXTOPER(node);
18208 node += regarglen[(U8)op];
18210 if (op == CURLYX || op == OPEN)
18214 #ifdef DEBUG_DUMPUNTIL
18215 PerlIO_printf(Perl_debug_log, "--- %d\n", (int)indent);
18220 #endif /* DEBUGGING */
18223 * ex: set ts=8 sts=4 sw=4 et: